TEXT stringlengths 0 3.96M | YEAR stringclasses 6 values | TITLE stringlengths 7 331 | DOI stringlengths 14 25 | AUTHORS list | LICENSE stringclasses 1 value |
|---|---|---|---|---|---|
Title
1
Autophagy promotes organelle clearance and organized cell separation of living root cap
2
cells in Arabidopsis thaliana
3
4
Running title
5
Role of autophagy in root cap
6
7
Authors
8
Tatsuaki Goh1,§,*, Kaoru Sakamoto1,§, Pengfei Wang2, Saki Kozono1, Koki Ueno1,
9
Shunsuke Miyashima1, Koichi Toyokura3, Hidehiro Fukaki3, Byung-Ho Kang2, Keiji
10
Nakajima1,*
11
12
Affiliations
13
1Graduate School of Science and Technology, Nara Institute of Science and Technology,
14
8916-5 Takayama, Ikoma, Nara 630-0192, Japan.
15
2School of Life Sciences, Centre for Cell & Developmental Biology and State Key
16
Laboratory of Agrobiotechnology, The Chinese University of Hong Kong, Shatin, New
17
Territories, Hong Kong, China.
18
3Department of Biology, Graduate School of Science, Kobe University, Rokkodai, Kobe
19
657-8501, Japan
20
§These authors contributed equally.
21
22
1
*Corresponding authors:
23
Tatsuaki Goh <goh@bs.naist.jp> and Keiji Nakajima <k-nakaji@bs.naist.jp>
24
25
Keywords
26
Arabidopsis thaliana, amyloplast, autophagy, cell separation, root cap
27
28
Summary statement
29
Time-lapse microscope imaging revealed spatiotemporal dynamics of intracellular
30
reorganization associated with functional transition and cell separation in the Arabidopsis
31
root cap and the roles of autophagy in this process.
32
33
34
2
Abstract
35
The root cap is a multi-layered tissue covering the tip of a plant root that directs root
36
growth through its unique functions such as gravity-sensing and rhizosphere interaction.
37
To prevent damages from the soil environment, cells in the root cap continuously turn
38
over through balanced cell division and cell detachment at the inner and the outer cell
39
layers, respectively. Upon displacement toward the outermost layer, columella cells at
40
the central root cap domain functionally transition from gravity-sensing cells to secretory
41
cells, but the mechanisms underlying this drastic cell fate transition are largely unknown.
42
By using live-cell tracking microscopy, we here show that organelles in the outermost
43
cell layer undergo dramatic rearrangements, and at least a part of this rearrangement
44
depends on spatiotemporally regulated activation of autophagy. Notably, this root cap
45
autophagy does not lead to immediate cell death, but rather is necessary for organized
46
separation of living root cap cells, highlighting a previously undescribed role of
47
developmentally regulated autophagy in plants.
48
3
Introduction
49
50
The root cap is a cap-like tissue covering the tip of a plant root. The root cap protects the
51
root meristem where rapid cell division takes place to promote root elongation (Arnaud
52
et al., 2010; Kumpf and Nowack, 2015). The root cap is also responsible for a number of
53
physiological functions, such as gravity-sensing to redirect the root growth axis (Strohm
54
et al., 2012), and metabolite secretion for lubrication and rhizosphere interaction
55
(Cannesan et al., 2012; Driouich et al., 2013; Hawes et al., 2016; Maeda et al., 2019). In
56
addition to its unique functions, the root cap exhibits a striking developmental feature,
57
namely continuous turnover of its constituent cells (Fig. 1A) (Kamiya et al., 2016). This
58
cell turnover is enabled by concerted production and detachment of cells at the inner stem
59
cells layer and the outer mature cell layer, respectively. Notably, the outermost root cap
60
cells detach from the root tip and disperse into the rhizosphere, creating a unique
61
environment at the border between the root and the soil. For this, detaching root cap cells
62
are called "border cells" (Hawes and Lin, 1990). Cell turnover is commonly seen in
63
animals but rarely found in plants where morphogenesis relies not only on the production
64
of new cells but also on the accumulation of mature and sometimes dead cells. Thus, the
65
root cap serves as a unique experimental material to study how plant cells dynamically
66
change their morphology and functions during tissue maintenance.
67
In the model angiosperm Arabidopsis thaliana (Arabidopsis), the root cap is
68
composed of two radially organized domains, the central columella and the surrounding
69
lateral root cap (LRC) that together constitute five to six cell layers along the root
70
4
proximodistal axis (Fig. 1) (Dolan et al., 1993). In Arabidopsis, the outermost root cap
71
cells do not detach individually, but rather separate as a cell layer (Fig. 1) (Driouich et al.,
72
2007; Kamiya et al., 2016; Vicre et al., 2005). Previous studies revealed that detachment
73
of the Arabidopsis root cap cells is initiated by localized activation of programmed cell
74
death (PCD) at the proximal LRC region, and requires the functions of the NAC-type
75
transcription factor SOMBRERO (SMB), a master regulator of root cap cell maturation
76
(Bennett et al., 2010; Fendrych et al., 2014; Willemsen et al., 2008; Xuan et al., 2016).
77
While SMB is expressed in all root cap cells and acts as a master regulator of cell
78
maturation in the root cap, two related NAC-type transcription factors, BEARSKIN1
79
(BRN1) and BRN2, are specifically expressed in the outer two cell layers of the root cap
80
(Bennett et al., 2010; Kamiya et al., 2016). BRN1 and BRN2 share high sequence
81
similarities and redundantly promote the separation of central columella cells. Cell
82
separation in plants requires partial degradation of cell walls. Indeed, ROOT CAP
83
POLYGLACTUROSE (RCPG) gene encoding a putative pectin-degrading enzymes acts
84
downstream of BRN1 and BRN2, and at least BRN1 can directly bind to the RCPG
85
promoter (Kamiya et al., 2016). CELLULASE5 (CEL5) gene encoding a putative
86
cellulose-degrading enzyme is also implicated in cell separation in the root cap (Bennett
87
et al., 2010; del Campillo et al., 2004).
88
Previous electron microscopic studies reported profound differences in the
89
intracellular organization between the inner and the outer root cap cells of Arabidopsis
90
(Maeda et al., 2019; Sack and Kiss, 1989). As expected from their gravity-sensing
91
functions, columella cells in the inner layers accumulate large amyloplasts. Amyloplasts
92
5
are specialized plastids containing starch granules and known to act as statoliths in the
93
gravity-sensing cells (statocytes) in both roots and shoots (Gilroy and Swanson, 2014).
94
In contrast, columella cells constituting the outermost root cap layer do not contain large
95
amyloplasts, and instead accumulate secretory vesicles (Maeda et al., 2019; Poulsen et
96
al., 2008). Thus, the observed difference in subcellular structures correlates well with the
97
functional transition of columella cells from gravity-sensing cells to the secretory cells
98
(Blancaflor et al., 1998; Maeda et al., 2019; Vicre et al., 2005). Before detachment, the
99
outermost root cap cells contain a large central vacuole, likely for the storage of various
100
metabolites (Baetz and Martinoia, 2014). In addition, a novel role of cell death promotion
101
has been proposed for the large central vacuole in the LRC cells (Fendrych et al., 2014).
102
In eukaryotes, dispensable or damaged proteins and organelles are degraded by
103
a self-digestion process called autophagy (Mizushima and Komatsu, 2011). Autophagy
104
initiates with expansion of isolated membranes, which subsequently form spherical
105
structures called the autophagosomes and engulf target components. In later steps,
106
autophagosomes fuse with vacuoles, and the content of autophagosomes degraded by
107
hydrolytic enzymes stored in the vacuole. When eukaryotic cells are subjected to stress
108
conditions such as nutrient starvation, autophagy is activated to recycle nutrients and
109
maintain intracellular environments in order to sustain the life of cells and/or individuals
110
(Mizushima and Komatsu, 2011). Autophagy plays an important role not only in stress
111
response but also in development and differentiation, as autophagy-deficient mutants are
112
lethal in a variety of model organisms including yeast, nematode, fruit fly, and mouse
113
(Mizushima and Levine, 2010). Genes encoding central components of autophagy, the
114
6
core ATG genes, are conserved in the Arabidopsis genome (Hanaoka et al., 2002; Liu and
115
Bassham, 2012). However, under normal growth conditions, autophagy-deficient
116
Arabidopsis mutants grow normally except for early senescence (Hanaoka et al., 2002;
117
Yoshimoto et al., 2009). Thus roles of autophagy in plant growth and development remain
118
largely unknown.
119
In this study, we revealed morphological and temporal dynamics of
120
intracellular rearrangement that enable the functional transition of the root cap cells in
121
Arabidopsis by using motion-tracking time-lapse imaging. We also found that the
122
autophagy-deficient Arabidopsis mutants are defective in cell clearance and vacuolization
123
of the outermost root cap cells. Unexpectedly, the autophagy-deficient mutants are
124
impaired in the organized separation of the outermost root cap layer. Thus our study
125
revealed a novel role of developmentally regulated autophagy in the root cap
126
differentiation and functions.
127
128
129
Results
130
131
Outermost columella cells undergo rapid organelle rearrangement before cell
132
detachment
133
While previous electron microscopic studies have revealed profound differences in
134
intracellular structures between the inner and the outer root cap cells (Maeda et al., 2019;
135
Poulsen et al., 2008; Sack and Kiss, 1989), spatiotemporal dynamics of subcellular
136
7
reorganization in the root cap cells has not been analyzed, due to a difficulty in performing
137
prolonged time-lapse imaging of the root tip that quickly relocates as the root elongates.
138
To overcome this problem, we developed a motion-tracking microscope system with a
139
horizontal optical axis and a spinning disc confocal unit. A similar system has been
140
reported by another group (von Wangenheim et al., 2017). Our microscope system
141
enabled high-magnification time-lapse confocal imaging of the tip of vertically growing
142
roots for up to six days, allowing visualization of cellular and subcellular dynamics of
143
root cap cells during three consecutive detachment events (Supplementary Fig. S1).
144
Under our experimental conditions, the outermost root cap layer of wild-type
145
Arabidopsis sloughed off with a largely fixed interval of about 38 hours (h)
146
(Supplementary Fig. S1F). This periodicity is comparable to that reported for roots
147
growing on agar plates (Shi et al., 2018), indicating that our microscope system does not
148
affect the cell turnover rate of the root cap. Bright-field observation revealed that the cell
149
detachment initiates in the proximal LRC region and extends toward the central columella
150
region (Fig. 1 and Fig. S1A-S1D). In concert with the periodic detachment of the
151
outermost layer, subcellular structures of the neighboring inner cell layer (hereafter called
152
the second outermost layer) rearranged dynamically (Fig. 2A and Supplementary Movie
153
S1). Before the detachment of the outermost layer, columella cells in the inner three to
154
four cell layers contained large amyloplasts that sedimented toward the distal (bottom)
155
side of the cell (Fig. 2A, -4 h, light blue arrowheads), whereas those in the outermost
156
layer were localized in the middle region of the cell (Fig. 2A, -4 h, dark blue arrowhead).
157
A few hours after the outermost layer started to detach at the proximal LRC region, the
158
8
amyloplasts in the second outermost layer relocated toward the middle region of the cell,
159
resulting in a similar localization pattern to those of the outermost layer (Fig. 2A, 0.5 h,
160
dark blue arrowheads). Toward the completion of the cell separation, rapid vacuolization
161
and shrinkage of amyloplasts took place in the outermost layer (Fig. 2A, 18 h, green
162
arrowhead).
163
By using plants expressing nuclear-localized red fluorescent proteins
164
(DR5v2:H2B-tdTomato), we could also visualize dynamic relocation of nuclei, as well as
165
its temporal relationship with amyloplast movement (Fig. 2B and Supplementary Movie
166
S2). In the second outermost layer, nuclei relocated from the proximal (upper) to the
167
middle region of each cell about a few hours before the neighboring outermost layer
168
initiated detachment (Fig. 2B, -8 h, red arrowhead). This nuclear migration was followed
169
by the relocation of amyloplasts around the time when the neighboring outermost layer
170
initiated detachment at the proximal LRC region (Fig. 2B, 0 h, dark blue arrowhead). In
171
later stages, the amyloplasts surrounded the centrally-localized nucleus (Fig. 2B, 13 h,
172
dark blue arrowhead). In the outermost cells, nuclei migrated further to localize to the
173
distal pole of the cell (Fig. 2B, 13 h, purple arrowheads).
174
Dynamic change in vacuolar morphology was also visualized using plants
175
expressing a tonoplast marker (VHP1-mGFP) (Segami et al., 2014) (Supplementary Fig.
176
S2 and Supplementary movie S3). Vacuoles in the inner columella cells were smaller and
177
spherical, whereas those in the outer cells were larger and tubular (Supplementary Fig.
178
S2, 5-23 h). Notably, in the outermost layer, vacuoles were dramatically enlarged, and
179
eventually occupied the entire volume of detaching root cap cells (Supplementary Fig.
180
9
S2, 35-47 h). Confocal imaging of plants expressing both tonoplast and nuclear markers
181
(VHP1-mGFP and pRPS5a:H2B-tdTomato) (Adachi et al., 2011; Segami et al., 2014)
182
revealed that both nuclei and amyloplasts were embedded in the meshwork of vacuolar
183
membranes in the outermost cell layer, whereas, in the inner cell layer, amyloplasts were
184
localized in a space devoid of vacuolar membranes (Fig. 2C). Taken together, our time-
185
lapse microscopic imaging revealed a highly organized sequence of organelle
186
rearrangement in the outer root cap cells, as well as its close association with cell position
187
and cell detachment.
188
189
Autophagy is activated in the outermost root cap cells before their detachment
190
Autophagy is an evolutionarily conserved self-digestion system in eukaryotes and
191
operates by transporting cytosolic components and organelles to the vacuole for nutrient
192
recycling and homeostatic control (Mizushima and Komatsu, 2011). The rapid
193
disappearance of amyloplasts and the formation of large vacuoles observed in the
194
outermost root cap cells made us hypothesize that autophagy operates behind their
195
dynamic subcellular rearrangements before the cell detachment. To test this hypothesis,
196
we examined whether autophagosomes, spherical membrane structures characteristics of
197
autophagy, are formed in the root cap cells at the time and space corresponding to the
198
organelle rearrangement.
199
We first observed an autophagosome marker, 35Spro:GFP-ATG8a, which
200
ubiquitously expresses GFP-tagged Arabidopsis ATG8a proteins, one of the nine ATG8
201
proteins encoded in the Arabidopsis genome (Yoshimoto et al., 2004). ATG8 is a
202
10
ubiquitin-like protein, and upon autophagy activation, incorporated into the
203
autophagosome membranes as a conjugate with phosphatidylethanolamine (Liu and
204
Bassham, 2012). Our time-lapse confocal imaging revealed uniform localization of GFP-
205
ATG8a fluorescence in the inner cell layers, suggesting low autophagic activity in these
206
cells (Fig. 3B and Supplementary Movie S4). In contrast, in detaching outermost cells,
207
dot-like signals of GFP-ATG8a became evident and their number and size increased (Fig.
208
3C, -24.0-1.5 h). In later stages, GFP-ATG8a signals largely disappeared in the outermost
209
cells before their detachment (Fig. 3C, 10 h). After the detachment of the outermost cell
210
layer, the inner cells (the new outermost cells) remained showing uniform GFP-ATG8
211
signals (Fig. 3C, 18.5 h). In the later phase of cell detachment, GFP-ATG8a signals
212
exhibited ring-like shapes, a typical image of autophagosomes in confocal microscopy
213
(Fig. 3C, 1.5 h, red arrowhead and a magnified image in the inset).
214
To further confirm whether the GFP-ATG8a-labelled puncta correspond to the
215
typical double membrane-bound autophagosome, we performed correlative light and
216
electron microscopy (CLEM) analysis (Fig. 4) (Wang and Kang, 2020). GFP
217
fluorescence precisely colocalized with spherical structures typical of autophagosomes
218
(Fig. 4C-4F). Together, our observations confirmed that autophagy is activated in the
219
outermost columella cells before their detachment.
220
221
Autophagy promotes organelle rearrangement in the outermost root cap cells
222
To examine whether autophagy plays a role in the maturation of columella cells, we first
223
tested the effect of E-64d, a membrane-permeable protease inhibitor that promotes the
224
11
accumulation of autophagic bodies inside the vacuole (Inoue et al., 2006; Merkulova et
225
al., 2014). In the outermost columella cells of E64d-treated roots, autophagic body-like
226
aggregates accumulated inside the enlarged vacuoles, suggesting the occurrence of active
227
autophagic degradation in these cells (Fig. S3B, compare with S3A).
228
We next carried out the phenotypic characterization of autophagy-deficient
229
mutants. ATG genes encoding autophagy components are known to exist in the genomes
230
of Arabidopsis and other model plant species (Hanaoka et al., 2002; Liu and Bassham,
231
2012). Among them, ATG5 belongs to the core ATG genes and is essential for
232
autophagosome formation as ATG8. In the loss of function atg5-1 mutant (Yoshimoto et
233
al., 2009), GFP-ATG8a signal was uniformly distributed throughout the cytosol both
234
during and after the cell detachment, indicating that autophagosome formation in the
235
detaching columella cells requires functional ATG5 (Fig. S4 and Supplementary movie
236
S5). Furthermore, time-lapse observation revealed a loss of full vacuolation in the
237
detaching outermost cells of atg5-1 (Fig. S5A, Supplementary movie S6). In the
238
detaching outermost cells of wild-type plants, a central vacuole enlarged to occupy the
239
entire cell volume, whereas only a few spherical and small fragmented vacuoles were
240
found in the corresponding cells of atg5-1 (Fig. 5A-5D). Whereas the disappearance of
241
iodine-stained large amyloplasts was not affected in the outer columella cells of atg5-1
242
(Fig. S3C and S3D), plastids in the atg5-1 mutant exhibited abnormal morphologies
243
dominated by tubular structures called stromules (Hanson and Hines, 2018), suggesting
244
a specific role of autophagy in plastid restructuring and/or degradation (Fig. S3E and S3F).
245
We also found that the detaching atg5-1 cells were strongly stained with FDA, a
246
12
compound that emits green fluorescence when hydrolyzed in the cytosol, as compared
247
with the restricted fluorescence in the cortical region of corresponding wild-type cells
248
(Fig. 5E and 5F). Retention of cytosol in detaching columella cells was also observed in
249
FDA-stained roots of additional atg mutants including atg2-1, atg7-2, atg10-1, atg12ab,
250
atg13ab and atg18a (Fig. 5G-5L), as well as in atg5-1 plants expressing GUS-GFP fusion
251
proteins under the outer layer-specific BRN1 promoter (Fig. S5D, compare with S5C).
252
Defects of vacuolization and cytosol digestion in atg5-1 were complemented with an
253
ATG5-GFP transgene, where GFP-tagged GFP5 proteins were expressed under the ATG5
254
promoter (Fig. 5M and 5N). Together, these observations clearly demonstrated a central
255
role of autophagy in cytosol digestion and vacuolization of detaching columella cells.
256
257
Autophagy is required for organized separation of root cap cell layer
258
In the course of time-lapse imaging of atg5-1, we noticed that the autophagy-deficient
259
mutants exhibited a distinct cell detachment behavior as compared with that of wild type.
260
While the outermost root cap cells detach as a cell layer in the wild type (Fig. 6A, white
261
arrowheads, and Supplementary Movie S7) (Kamiya et al., 2016), those of atg5-1
262
detached individually (Fig. 6B, orange arrowheads, and Supplementary Movie S8),
263
indicating that autophagy is required not only for organelle rearrangement but also for the
264
organized separation of root cap cell layers, a behavior typically observed in the root cap
265
of Arabidopsis and related species (Hamamoto et al., 2006; Hawes et al., 2002). The
266
aberrant cell detachment behavior of atg5-1 was complemented by the ATG5-GFP
267
transgene (Fig. 6C, white arrowheads, and Supplementary Movie S9), confirming the
268
13
causal relationship. To clarify whether autophagy activation in the outermost cells is
269
sufficient for organized cell separation, we established atg5-1 plants expressing GFP-
270
tagged ATG5 proteins under the BRN1 and the RCPG promoter, which drive transcription
271
in the outer two cell layers and the outermost root cap layer, respectively (Kamiya et al.,
272
2016). Time-lapse imaging revealed that both of the plant lines restored the organized
273
separation of the outermost root cap cell layer (Fig. 7A and 7B, white arrowheads and
274
Supplementary movie S10 and S11). These observations, in particular, restoration of the
275
layered cell separation by the RCPG promoter-driven ATG-GFP, confirmed that
276
autophagy activation in the detaching cells at the timing of active cell wall degradation is
277
sufficient for the organized separation of the outermost root cap layer.
278
279
280
Discussion
281
282
In this study, we revealed spatiotemporal dynamics of the intracellular reorganization and
283
cell detachment in the Arabidopsis root cap, as well as a role of developmentally regulated
284
autophagy in these processes. In the outermost root cap layer, autophagy is activated in a
285
specific cell layer and at the timing closely associated with the functional transition of
286
columella cells and their detachment. This spatiotemporally regulated activation of
287
autophagy is essential not only for cell clearance and vacuolar enlargement but also for
288
the organized separation of the outermost layer of the root cap.
289
290
14
Motion-tracking time-lapse imaging revealed rapid intracellular rearrangement
291
associated with the functional transition of root cap cells
292
Cells constituting the root cap constantly turn over by balanced production and
293
detachment of cells at the innermost and the outermost cell layers, respectively. During
294
their lifetime, columella cells undergo a functional transition from being gravity-sensing
295
statocytes to secretory cells according to their position (Blancaflor et al., 1998; Maeda et
296
al., 2019; Sack and Kiss, 1989; Vicre et al., 2005). While the previous electron
297
microscopic observations revealed a profound difference in the subcellular structures
298
between the inner statocytes and the outer secretory cells of the Arabidopsis root cap
299
(Maeda et al., 2019; Poulsen et al., 2008; Sack and Kiss, 1989), detailed temporal
300
dynamics of organelles rearrangement in relation to the timing of cell displacement and
301
detachment has not been analyzed.
302
Our time-lapse observation using a motion-tracking microscope system with a
303
horizontal optical axis clearly visualized both morphological and temporal details of
304
organelle rearrangement in this transition (Fig. 8). Cells in the inner two to three layers
305
have unique arrangements of organelles, which is likely optimized for their gravity-
306
sensing function (Blancaflor et al., 1998). In these cells, starch granule-containing
307
amyloplasts and nuclei are localized at the distal (lower) and proximal (upper) end of
308
each cell, respectively, whereas small tubular vacuoles preferentially occupy the proximal
309
(upper) half of each cell (Fig. 2) (Leitz et al., 2009; Sack and Kiss, 1989). This organelle
310
arrangement changed dynamically in the outermost cell layer. The first conspicuous sign
311
of rearrangement is relocation of nuclei from the upper to the central region, which
312
15
happens even before the layer containing these columella cells starts to detach at the
313
proximal LRC region (Fig. 2). Around the time of the detachment of this cell layer,
314
amyloplasts 'float up' to the middle region of the cell (Fig. 2). Later, amyloplasts disappear
315
and vacuoles start to expand to occupy the entire cell volume by the time these cells
316
slough off from the root tip (Fig. 2 and Supplementary Fig. S2). The development of large
317
central vacuoles likely constitutes a central component of functional specialization of
318
these cells for storage (Driouich et al., 2013; Hawes et al., 2016; Vicre et al., 2005). A
319
novel role of central vacuoles for cell death promotion has been also proposed for LRC
320
cells (Fendrych et al., 2014).
321
Here, the central question is what controls the spatiotemporal activation of this
322
dramatic rearrangement of organelles in the root cap. The NAC-type transcription factors
323
BRN1 and BRN2 are expressed specifically in the outer two cell layers of the root cap
324
and required for cell detachment (Bennett et al., 2010; Kamiya et al., 2016), seemingly
325
becoming good candidates for the upstream regulators. However, the outermost root cap
326
cells of brn1 brn2 mutants, though defective in cell detachment, were found to be
327
normally vacuolated and lacking amyloplasts as those of wild type, indicating that at least
328
a part of the organelle rearrangement is regulated independently of BRN1 and BRN2
329
(Bennett et al., 2010; Kamiya et al., 2016). On the other hand, our previous study
330
suggested the existence of unknown positional cues that, together with another NAC-type
331
transcription factor SMB, promote the outer layer-specific expression of BRN1 and BRN2
332
(Kamiya et al., 2016). Future identification of factors transmitting such positional
333
16
information will provide a clue to understanding a mechanism underlying position-
334
dependent organelle rearrangement in the root cap.
335
336
Autophagy is activated in the outermost root cap cells to promote cell clearance and
337
vacuolization
338
Our time-lapse imaging revealed specific activation of autophagy in the outermost root
339
cap layer in concert with the progression of the cell separation (Fig. 3). As expected,
340
mutants defective in the canonical autophagy pathway exhibited compromised cell
341
clearance and vacuolization of detaching root cap cells (Fig. 5). Because detached root
342
cap cells are dispersed into the rhizosphere and act in plant defense through their secretory
343
capacity (Driouich et al., 2013; Hawes et al., 2016), degradation of starch-containing
344
amyloplasts and vacuolar expansion appear to be a reasonable differentiation trajectory
345
in view of energy-recycling and storage.
346
Autophagosomes are double-membrane vesicles that engulf a wide range of
347
intracellular components and transport them to vacuoles for degradation by lytic enzymes.
348
Rapid reduction of GFP-ATG8a signals and accumulation of autophagic body-like
349
structures inside the vacuoles after the application of the proteinase inhibitor E64d
350
(Supplementary Fig. S3) support occurrence of active autophagic flow and vacuolar
351
degradation in the outermost root cap layer. Such active autophagic transport may act to
352
supply membrane components and to facilitate water influx into the vacuoles by
353
increasing osmotic pressure, leading to enhanced vacuolization of the outermost root cap
354
cells.
355
17
While the autophagy-deficient atg5-1 mutant was capable of eliminating
356
Lugol-stained amyloplasts from mature columella cells as the wild type, morphology of
357
plastids in the detaching root cap cells was abnormal in atg5-1, having tubular structures
358
typical of stromules (Supplementary Fig. S3). Storomules arise from chloroplasts under
359
starvation or senescence conditions. In such stress conditions, chloroplast contents are
360
degraded via piecemeal-type organelle autophagy, in which stromules or chloroplast
361
protrusions are believed to be engulfed by an autophagosome (Ishida et al., 2008),
362
whereas damaged chloroplasts can be engulfed as a whole by an isolated membrane and
363
transported into vacuoles (Izumi et al., 2013). Stromule formation in the autophagy-
364
deficient atg5-1 mutant suggests that amyloplast degradation in the outermost root cap
365
cells proceeds in two steps; first by autophagy-independent degradation of starch granules
366
and stromule formation, followed by the piecemeal chloroplast autophagy. It should be
367
noted, however, that autophagy-dependent amyloplast degradation also occurs as a part
368
of root hydrotropic response, where some starch-containing amyloplasts are engulfed
369
directly by the autophagosome-like structures (Nakayama et al., 2012). Together, these
370
observations suggest that multiple amyloplast degradation pathways exist in the
371
Arabidopsis root cap with different contributions of autophagy.
372
While the present study clearly demonstrated the role of autophagy in the
373
organelle rearrangement in the root cap, spatiotemporal regulation of autophagy
374
activation is yet to be investigated. The root cap autophagy seems to operate via canonical
375
macro-autophagy pathway mediated by the components encoded by the ATG genes (Fig.
376
5) (Liu and Bassham, 2012) (Fig. 5). Autophagy is induced by various stress conditions,
377
18
such as nutrient starvation, as well as abiotic and biotic stresses, where SNF-related
378
kinase 1 (SnRK1) and target of rapamycin (TOR) protein kinase complexes function as
379
key regulators (Liu and Bassham, 2012; Mizushima and Komatsu, 2011). In contrast, the
380
root cap autophagy can occur in plants growing on a sterile nutrient-rich medium in our
381
experiments, suggesting that root cap autophagy is activated independently of nutrient
382
starvation and biotic stress. Instead, activation of the root cap autophagy appears to be
383
closely associated with the process of cell detachment, which in turn is known to be
384
regulated by intrinsic developmental programs (Dubreuil et al., 2018; Shi et al., 2018).
385
Again, BRN1 and BRN2 are unlikely to regulate the root cap autophagy, because cell
386
clearance and vacuolization normally occur in the outermost root cap cells of brn1 brn2
387
mutants.
388
389
Autophagy is required for the organized separation of the Arabidopsis root cap cells
390
Autophagy promotes organelle rearrangement associated with the differentiation of
391
secretory cells that subsequently slough off to disperse into the rhizosphere. Based on this,
392
we expected that the loss of autophagy would inhibit or delay cell detachment in the root
393
cap. Somewhat unexpectedly, however, autophagy-deficient atg5-1 mutants showed a
394
phenotype suggestive of enhanced cell detachment (Fig. 6). In Arabidopsis and related
395
species, the outermost root cap cells separate as a cell layer, rather than as isolated cells
396
(Driouich et al., 2010; Driouich et al., 2007; Kamiya et al., 2016). Although the
397
physiological significance of this detachment behavior has not been demonstrated so far,
398
it has been hypothetically linked with a capacity of secreting mucilage, a mixture of
399
19
polysaccharides implicated in plant defense, aluminum-chelating, and lubrication
400
(Driouich et al., 2010; Maeda et al., 2019).
401
Previous genetic studies suggested a key role of cell wall pectins in the control
402
of root cap cell detachment; when pectin-mediated cell-cell adhesion was compromised
403
by mutations in genes encoding putative pectin-synthesizing enzymes or overexpression
404
of RCPG, a root cap-specific putative pectin-hydrolyzing enzyme, root cap cells slough
405
off as isolated cells (Driouich et al., 2010; Kamiya et al., 2016). Moreover, the
406
morphology of detaching root cap cell layers was altered in the loss-of-function rcpg
407
mutant, likely due to a failure of separating cell-cell adhesion along the lateral cell edge
408
(Kamiya et al., 2016). The similarity between the altered cell detachment behaviors
409
between atg5-1 and pectin-deficient plants suggests a role of autophagy in the control of
410
cell wall integrity during the root cap cell detachment. Both transport and modification
411
of cell wall pectins require Golgi and Golgi-derived vesicles (Driouich et al., 2012; Wang
412
et al., 2017). In outer root cap cells, small vesicles accumulate for their secretory functions
413
(Driouich et al., 2013; Maeda et al., 2019; Wang et al., 2017), and a mutation disrupting
414
this secretory pathway results in the failure of root cap cell detachment (Poulsen et al.,
415
2008). If autophagy is required for timely attenuation of such vesicular transport during
416
the cell detachment program, lack of autophagy should lead to prolonged secretion of cell
417
wall modifying enzymes such as RCPG, resulting in enhanced loosening of cell-cell
418
adhesion. Indeed, we could recognize broader gaps at the apoplastic junctions at the distal
419
cell-cell adhesion points in atg5-1 than those in the wild type (Supplementary movie S7
420
and S8). Future studies comparing secretory dynamics of cell wall-modifying enzymes in
421
20
various genetic backgrounds using our live-imaging system will elucidate the molecular
422
mechanism controlling the cell detachment behaviors in the root cap and the role of
423
autophagy.
424
In summary, our study revealed the role of spatiotemporally regulated
425
autophagy in cell clearance and vacuolization in root cap differentiation as well as in cell
426
detachment. While autophagy has been known to promote tracheary element
427
differentiation in Arabidopsis and anther maturation in rice, roles of autophagy in these
428
instances are linked to PCD (Escamez et al., 2016; Kurusu and Kuchitsu, 2017).
429
Considering that autophagy is required for functional transition and detachment of living
430
columella cells, our study revealed a previously undescribed role of developmentally
431
regulated autophagy in plant development.
432
433
21
Materials and Methods
434
435
Plant materials and growth conditions
436
Arabidopsis thaliana L. Heynh (Arabidopsis) accession Col-0 was used as the wild type.
437
The Arabidopsis T-DNA insertional lines, atg5-1 (SAIL_129_B07), atg7-2 (GK-
438
655B06), atg2-1 (SALK_076727), atg10-1 (SALK_084434), atg12a (SAIL_1287_A08),
439
atg12b (SALK_003192), atg13a (GABI_761_A11), atg13b (GK-510F06) and atg18a
440
(GK_651D08) have been described previously (Doelling et al., 2002; Hanaoka et al.,
441
2002; Izumi et al., 2013; Thompson et al., 2005; Yoshimoto et al., 2004; Yoshimoto et
442
al., 2009). 35Spro:CT-GFP, RPS5apro:H2B-tdTomato and VHP1-mGFP has been
443
described previously (Adachi et al., 2011; Köhler et al., 1997; Segami et al., 2014). Seeds
444
were grown vertically on Arabidopsis nutrient solution supplemented with 1 % (w/v)
445
sucrose and 1 % (w/v) agar under the 16h light/8h dark condition at 23 ºC.
446
447
Generation of transgenic plants
448
For ATG5pro:ATG5:GFP, a 4.5-kb genomic fragment harboring the ATG5
449
coding region and the 5’-flanking region was amplified by PCR and cloned into
450
pAN19/GFP-NOSt
vector,
which
contained
GFP-coding
sequence
and
the
451
Agrobacterium (Rhizobium) nopaline synthase terminator (NOS). The resulting ATG5-
452
GFP fragment was then transferred to pBIN4 to give ATG5pro:ATG5:GFP/pBIN41.
453
22
Layer-specific rescue constructs of ATG5-GFP were constructed by amplifying
454
the ATG5-GFP fragment from ATG5pro:ATG5:GFP/pBIN41, and inserting them to
455
pDONR221 by the GatewayTM technology. The ATG5-GFP fragment was then
456
transferred to pGWB501:BRN1pro and pGWB501:RCPGpro, which respectively
457
contained the BRN1 and RCPG promoter flanking the Gateway cassette in pGWB501
458
(Nakagawa et al., 2007). The cytosolic marker GUS-GFP was similarly constructed by
459
inserting a GUS-GFP fragment into pENTR D-TOPO, and then by transferring the insert
460
to pGWB501:BRN1pro to give BRN1pro:GUS-GFP.
461
For DR5v2:H2B:tdTomato, a DR5v2 promoter fragment was amplified by PCR
462
from the DRv2n3GFP construct (Liao et al., 2015), and inserted into pGWB501 by the
463
In-Fusion technique to give pGWB501:DR5v2. The H2B-tdTomato fragment in pENTR
464
was transferred to the pGWB501:DR5v2. Integrity of the cloned genes was verified by
465
DNA sequencing. Transformation of Arabidopsis plants was performed by the floral dip
466
method using Rhizobium (formerly Agrobacterium) tumefaciens, strain C58MP90.
467
468
Microscopy
469
Time-lapse imaging of the root cap was performed using two microscopic systems
470
developed in the corresponding authors' laboratory, which can automatically track the tip
471
of vertically growing roots. Technical details will be published elsewhere. Briefly, an
472
inverted microscope (ECLIPSE Ti-E and ECLIPSE Ti2-E, Nikon, Tokyo, Japan) was
473
tilted by 90 degrees to vertically orient the sample stage. The motorized stage was
474
controlled by the Nikon NIS-elements software with the “keep object in view” plugin to
475
23
automatically track the tip of growing roots. Three-day-old seedlings were transferred to
476
a chamber slide (Lab-Tek chambered coverglass, Thermofisher, Waltham, MA) and
477
covered with a block of agar medium.
478
Confocal laser scanning microscopy was carried out with a Nikon C2 confocal
479
microscope. Roots were stained with 10 µg/ml of propidium iodide (PI). Fluorescein
480
diacetate (FDA) staining was performed by soaking the roots in a solution containing 2
481
μg/ml of FDA.
482
Iodine staining was performed as described previously (Segami et al., 2018).
483
Root fixed in 4% (w/v) paraformaldehyde in PBS for 30 min under a vacuum at room
484
temperature. The fixed sample was washed twice for 1 min each in PBS and cleared with
485
ClearSee (Kurihara et al., 2015). The samples were transferred to 10% (w/v) xylitol and
486
25% (w/v) urea to remove sodium deoxycholate, and then stained in a solution containing
487
2 mM iodine (Wako), 10 % (w/v) xylitol, and 25 % (w/v) urea.
488
Correlative light and electron microscopy (CLEM) analysis was performed as
489
described previously (Wang and Kang, 2020; Wang et al., 2019). GFP-ATG8a seedlings
490
were grown vertically under 16 h light-8 h dark cycle at 22 °C for seven days. Root tips
491
samples expressing GFP were cryofixed with an EM ICE high-pressure freezer (Leica
492
Microsystems, Austria) and embedded in Lowicryl HM20 resin at -45ºC. TEM sections
493
of 150nm thickness were collected on copper or gold slot grids coated with formvar and
494
examined for GFP after staining the cell wall with Calcofluor White. The grids were post-
495
stained and GFP-positive cells were imaged under an H-7650 TEM (Hitachi High-Tech,
496
24
Japan) operated at 80kV. For electron tomography, tilt series were collected with a TF-
497
20 intermediate voltage TEM (Thermo Fisher Scientific, USA). Tomogram calculation
498
and three-dimensional model preparation were carried out with the 3dmod software
499
package (bio3d.colorado.edu).
500
501
Acknowledgments
502
We thank Masanori Izumi (RIKEN, Japan), Kohki Yoshimoto (Meiji University, Japan),
503
Masayoshi Maeshima (Nagoya University, Japan), Shoji Segami (NIBB, Japan), and
504
Maureen R. Hanson (Cornell University, USA) for providing plant materials, Dolf
505
Weijers (Wageningen University, Netherlands) for providing the DR5v2 construct, and
506
Masako Kanda for technical assistance.
507
508
Competing interests
509
The authors declare no competing interests.
510
511
Funding
512
This work was supported by MEXT/JSPS KAKENHI grants 20H05330 to T.G. and
513
19H05671, 19H05670 and 19H03248 to K.N., and by the Hong Kong Research Grant
514
Council (GRF14121019, 14113921, AoE/M-05/12, C4002-17G) to B.-H. K..
515
516
25
References
517
Adachi, S., Minamisawa, K., Okushima, Y., Inagaki, S., Yoshiyama, K., Kondou, Y.,
518
Kaminuma, E., Kawashima, M., Toyoda, T., Matsui, M., et al. (2011).
519
Programmed induction of endoreduplication by DNA double-strand breaks in
520
Arabidopsis. Proc. Natl. Acad. Sci. USA 108, 10004-10009.
521
Arnaud, C., Bonnot, C., Desnos, T. and Nussaume, L. (2010). The root cap at the
522
forefront. C. R. Biol. 333, 335-343.
523
Baetz, U. and Martinoia, E. (2014). Root exudates: the hidden part of plant defense.
524
Trends Plant Sci. 19, 90-98.
525
Bennett, T., van den Toorn, A., Sanchez-Perez, G. F., Campilho, A., Willemsen, V.,
526
Snel, B. and Scheres, B. (2010). SOMBRERO, BEARSKIN1, and BEARSKIN2
527
regulate root cap maturation in Arabidopsis. Plant Cell 22, 640-654.
528
Blancaflor, E. B., Fasano, J. M. and Gilroy, S. (1998). Mapping the Functional Roles
529
of Cap Cells in the Response of Arabidopsis Primary Roots to Gravity. Plant
530
Physiol. 116, 213-222.
531
Cannesan, M. A., Durand, C., Burel, C., Gangneux, C., Lerouge, P., Ishii, T., Laval,
532
K., Follet-Gueye, M. L., Driouich, A. and Vicré-Gibouin, M. (2012). Effect of
533
Arabinogalactan Proteins from the Root Caps of Pea and Brassica napus on
534
Aphanomyces euteiches Zoospore Chemotaxis and Germination. Plant Physiol.
535
159, 1658-1670.
536
del Campillo, E., Abdel-Aziz, A., Crawford, D. and Patterson, S. E. (2004). Root cap
537
specific expression of an endo-beta-1,4-D-glucanase (cellulase): a new marker to
538
study root development in Arabidopsis. Plant Mol. Biol. 56, 309-323.
539
Doelling, J. H., Walker, J. M., Friedman, E. M., Thompson, A. R. and Vierstra, R.
540
D. (2002). The APG8/12-activating Enzyme APG7 Is Required for Proper
541
Nutrient Recycling and Senescence in Arabidopsis thaliana. J. Biol. Chemi. 277,
542
33105-33114.
543
Dolan, L., Janmaat, K., Willemsen, V., Linstead, P., Poethig, S., Roberts, K. and
544
Scheres, B. (1993). Cellular organisation of the Arabidopsis thaliana root.
545
Development 119, 71-84.
546
Driouich, A., Durand, C., Cannesan, M.-A., Percoco, G. and Vicré-Gibouin, M.
547
(2010). Border cells versus border-like cells: are they alike? J. Exp. Bot. 61, 3827-
548
26
3831.
549
Driouich, A., Durand, C. and Vicre-Gibouin, M. (2007). Formation and separation of
550
root border cells. Trends Plant Sci. 12, 14-19.
551
Driouich, A., Follet-Gueye, M.-L., Bernard, S., Kousar, S., Chevalier, L., Vicré-
552
Gibouin, M. and Lerouxel, O. (2012). Golgi-mediated synthesis and secretion
553
of matrix polysaccharides of the primary cell wall of higher plants. Front. Plant
554
Sci. 3, 79-79.
555
Driouich, A., Follet-Gueye, M. L., Vicre-Gibouin, M. and Hawes, M. (2013). Root
556
border cells and secretions as critical elements in plant host defense. Curr. Opin.
557
Plant Biol. 16, 489-495.
558
Dubreuil, C., Jin, X., Grönlund, A. and Fischer, U. (2018). A Local Auxin Gradient
559
Regulates Root Cap Self-Renewal and Size Homeostasis. Curr. Biol. 28, 2581-
560
2587.e2583.
561
Escamez, S., Andre, D., Zhang, B., Bollhoner, B., Pesquet, E. and Tuominen, H.
562
(2016). METACASPASE9 modulates autophagy to confine cell death to the
563
target cells during Arabidopsis vascular xylem differentiation. Biol. Open 5, 122-
564
129.
565
Fendrych, M., Van Hautegem, T., Van Durme, M., Olvera-Carrillo, Y., Huysmans,
566
M., Karimi, M., Lippens, S., Guerin, C. J., Krebs, M., Schumacher, K., et al.
567
(2014). Programmed cell death controlled by ANAC033/SOMBRERO
568
determines root cap organ size in Arabidopsis. Curr. Biol. 24, 931-940.
569
Gilroy, S. and Swanson, S. J. (2014). Gravitropic Signaling in Plants. In eLS.
570
Hamamoto, L., Hawes, M. C. and Rost, T. L. (2006). The production and release of
571
living root cap border cells is a function of root apical meristem type in
572
dicotyledonous angiosperm plants. Ann. Bot. 97, 917-923.
573
Hanaoka, H., Noda, T., Shirano, Y., Kato, T., Hayashi, H., Shibata, D., Tabata, S.
574
and Ohsumi, Y. (2002). Leaf Senescence and Starvation-Induced Chlorosis Are
575
Accelerated by the Disruption of an Arabidopsis Autophagy Gene. Plant Physiol.
576
129, 1181-1193.
577
Hanson, M. R. and Hines, K. M. (2018). Stromules: Probing Formation and Function.
578
Plant Physiol. 176, 128-137.
579
Hawes, M., Allen, C., Turgeon, B. G., Curlango-Rivera, G., Minh Tran, T., Huskey,
580
D. A. and Xiong, Z. (2016). Root Border Cells and Their Role in Plant Defense.
581
27
Annu. Rev. Phytopathol. 54, 143-161.
582
Hawes, M. C., Bengough, G., Cassab, G. and Ponce, G. (2002). Root Caps and
583
Rhizosphere. J. Plant Growth Regul. 21, 352-367.
584
Hawes, M. C. and Lin, H. J. (1990). Correlation of Pectolytic Enzyme Activity with the
585
Programmed Release of Cells from Root Caps of Pea (Pisum sativum). Plant
586
Physiol. 94, 1855-1859.
587
Inoue, Y., Suzuki, T., Hattori, M., Yoshimoto, K., Ohsumi, Y. and Moriyasu, Y.
588
(2006). AtATG genes, homologs of yeast autophagy genes, are involved in
589
constitutive autophagy in Arabidopsis root tip cells. Plant Cell Physiol. 47, 1641-
590
1652.
591
Ishida, H., Yoshimoto, K., Izumi, M., Reisen, D., Yano, Y., Makino, A., Ohsumi, Y.,
592
Hanson, M. R. and Mae, T. (2008). Mobilization of rubisco and stroma-localized
593
fluorescent proteins of chloroplasts to the vacuole by an ATG gene-dependent
594
autophagic process. Plant Physiol. 148, 142-155.
595
Izumi, M., Hidema, J., Makino, A. and Ishida, H. (2013). Autophagy Contributes to
596
Nighttime Energy Availability for Growth in Arabidopsis. Plant Physiol. 161,
597
1682-1693.
598
Kamiya, M., Higashio, S. Y., Isomoto, A., Kim, J. M., Seki, M., Miyashima, S. and
599
Nakajima, K. (2016). Control of root cap maturation and cell detachment by
600
BEARSKIN transcription factors in Arabidopsis. Development 143, 4063-4072.
601
Köhler, R. H., Cao, J., Zipfel, W. R., Webb, W. W. and Hanson, M. R. (1997).
602
Exchange of Protein Molecules Through Connections Between Higher Plant
603
Plastids. Science 276, 2039-2042.
604
Kumpf, R. P. and Nowack, M. K. (2015). The root cap: a short story of life and death.
605
J. Exp. Bot. 66, 5651-5662.
606
Kurihara, D., Mizuta, Y., Sato, Y. and Higashiyama, T. (2015). ClearSee: a rapid
607
optical clearing reagent for whole-plant fluorescence imaging. Development 142,
608
4168-4179.
609
Kurusu, T. and Kuchitsu, K. (2017). Autophagy, programmed cell death and reactive
610
oxygen species in sexual reproduction in plants. J. Plant Res. 130, 491-499.
611
Leitz, G., Kang, B. H., Schoenwaelder, M. E. and Staehelin, L. A. (2009). Statolith
612
sedimentation kinetics and force transduction to the cortical endoplasmic
613
reticulum in gravity-sensing Arabidopsis columella cells. Plant Cell 21, 843-860.
614
28
Liao, C. Y., Smet, W., Brunoud, G., Yoshida, S., Vernoux, T. and Weijers, D. (2015).
615
Reporters for sensitive and quantitative measurement of auxin response. Nat.
616
Methods 12, 207-210.
617
Liu, Y. and Bassham, D. C. (2012). Autophagy: pathways for self-eating in plant cells.
618
Annu. Rev. Plant Biol. 63, 215-237.
619
Maeda, K., Kunieda, T., Tamura, K., Hatano, K., Hara-Nishimura, I. and Shimada,
620
T. (2019). Identification of Periplasmic Root-Cap Mucilage in Developing
621
Columella Cells of Arabidopsis thaliana. Plant Cell Physiol. 60, 1296-1303.
622
Merkulova, E. A., Guiboileau, A., Naya, L., Masclaux-Daubresse, C. and Yoshimoto,
623
K. (2014). Assessment and optimization of autophagy monitoring methods in
624
Arabidopsis roots indicate direct fusion of autophagosomes with vacuoles. Plant
625
Cell Physiol. 55, 715-726.
626
Mizushima, N. and Komatsu, M. (2011). Autophagy: Renovation of Cells and Tissues.
627
Cell 147, 728-741.
628
Mizushima, N. and Levine, B. (2010). Autophagy in mammalian development and
629
differentiation. Nat. Cell Biol. 12, 823-830.
630
Nakagawa, T., Suzuki, T., Murata, S., Nakamura, S., Hino, T., Maeo, K., Tabata, R.,
631
Kawai, T., Tanaka, K., Niwa, Y., et al. (2007). Improved Gateway binary
632
vectors: high-performance vectors for creation of fusion constructs in transgenic
633
analysis of plants. Biosci. Biotechnol. Biochem. 71, 2095-2100.
634
Nakayama, M., Kaneko, Y., Miyazawa, Y., Fujii, N., Higashitani, N., Wada, S.,
635
Ishida, H., Yoshimoto, K., Shirasu, K., Yamada, K., et al. (2012). A possible
636
involvement of autophagy in amyloplast degradation in columella cells during
637
hydrotropic response of Arabidopsis roots. Planta 236, 999-1012.
638
Poulsen, L. R., López-Marqués, R. L., McDowell, S. C., Okkeri, J., Licht, D., Schulz,
639
A., Pomorski, T., Harper, J. F. and Palmgren, M. G. (2008). The Arabidopsis
640
P4-ATPase ALA3 Localizes to the Golgi and Requires a β-Subunit to Function in
641
Lipid Translocation and Secretory Vesicle Formation. Plant Cell 20, 658-676.
642
Sack, F. D. and Kiss, J. Z. (1989). Rootcap structure in wild type and in a starchless
643
mutant of Arabidopsis. Am. J. Bot. 76, 454-464.
644
Segami, S., Makino, S., Miyake, A., Asaoka, M. and Maeshima, M. (2014). Dynamics
645
of vacuoles and H+-pyrophosphatase visualized by monomeric green fluorescent
646
protein in Arabidopsis: artifactual bulbs and native intravacuolar spherical
647
29
structures. Plant Cell 26, 3416-3434.
648
Segami, S., Tomoyama, T., Sakamoto, S., Gunji, S., Fukuda, M., Kinoshita, S.,
649
Mitsuda, N., Ferjani, A. and Maeshima, M. (2018). Vacuolar H(+)-
650
Pyrophosphatase and Cytosolic Soluble Pyrophosphatases Cooperatively
651
Regulate Pyrophosphate Levels in Arabidopsis thaliana. Plant Cell 30, 1040-1061.
652
Shi, C.-L., von Wangenheim, D., Herrmann, U., Wildhagen, M., Kulik, I., Kopf, A.,
653
Ishida, T., Olsson, V., Anker, M. K., Albert, M., et al. (2018). The dynamics
654
of root cap sloughing in Arabidopsis is regulated by peptide signalling. Nat.
655
Plants 4, 596-604.
656
Strohm, A. K., Baldwin, K. L. and Masson, P. H. (2012). Molecular mechanisms of
657
root gravity sensing and signal transduction. Wiley Interdiscip. Rev.: Dev. Biol. 1,
658
276-285.
659
Thompson, A. R., Doelling, J. H., Suttangkakul, A. and Vierstra, R. D. (2005).
660
Autophagic Nutrient Recycling in Arabidopsis Directed by the ATG8 and ATG12
661
Conjugation Pathways. Plant Physiol. 138, 2097-2110.
662
Vicre, M., Santaella, C., Blanchet, S., Gateau, A. and Driouich, A. (2005). Root
663
border-like cells of Arabidopsis. Microscopical characterization and role in the
664
interaction with rhizobacteria. Plant Physiol. 138, 998-1008.
665
von Wangenheim, D., Hauschild, R., Fendrych, M., Barone, V., Benková, E. and
666
Friml, J. (2017). Live tracking of moving samples in confocal microscopy for
667
vertically grown roots. eLife 6, e26792.
668
Wang, P., Chen, X., Goldbeck, C., Chung, E. and Kang, B.-H. (2017). A distinct class
669
of vesicles derived from the trans-Golgi mediates secretion of xylogalacturonan
670
in the root border cell. Plant J. 92, 596-610.
671
Wang, P. and Kang, B. H. (2020). Correlative Light and Electron Microscopy Imaging
672
of the Plant trans-Golgi Network. Methods Mol. Biol. 2177, 59-67.
673
Wang, P., Liang, Z. and Kang, B. H. (2019). Electron tomography of plant organelles
674
and the outlook for correlative microscopic approaches. New Phytol. 223, 1756-
675
1761.
676
Willemsen, V., Bauch, M., Bennett, T., Campilho, A., Wolkenfelt, H., Xu, J.,
677
Haseloff, J. and Scheres, B. (2008). The NAC domain transcription factors FEZ
678
and SOMBRERO control the orientation of cell division plane in Arabidopsis root
679
stem cells. Dev. Cell 15, 913-922.
680
30
Xuan, W., Band, L. R., Kumpf, R. P., Van Damme, D., Parizot, B., De Rop, G.,
681
Opdenacker, D., Moller, B. K., Skorzinski, N., Njo, M. F., et al. (2016). Cyclic
682
programmed cell death stimulates hormone signaling and root development in
683
Arabidopsis. Science 351, 384-387.
684
Yoshimoto, K., Hanaoka, H., Sato, S., Kato, T., Tabata, S., Noda, T. and Ohsumi, Y.
685
(2004). Processing of ATG8s, Ubiquitin-Like Proteins, and Their Deconjugation
686
by ATG4s Are Essential for Plant Autophagy. Plant Cell 16, 2967-2983.
687
Yoshimoto, K., Jikumaru, Y., Kamiya, Y., Kusano, M., Consonni, C., Panstruga, R.,
688
Ohsumi, Y. and Shirasu, K. (2009). Autophagy negatively regulates cell death
689
by controlling NPR1-dependent salicylic acid signaling during senescence and the
690
innate immune response in Arabidopsis. Plant Cell 21, 2914-2927.
691
692
31
Figures legends
693
694
Fig. 1. A diagram illustrating structure and cell detachment process of Arabidopsis
695
root cap.
696
Landmark events constituting the cell separation sequence are marked by arrowheads.
697
Definition of the proximodistal polarity used in this study is shown on the left.
698
699
Fig. 2. Organelle rearrangement takes place in the outer root cap layers
700
(A) Time-lapse images visualizing the sequences of root cap cell detachment and
701
relocation of amyloplasts. Representative images before (left panel), at the beginning
702
(central panel), and around the end (right panel) of cell layer detachment are shown. Light
703
blue and dark blue arrowheads indicate sedimenting and floating amyloplasts,
704
respectively. Green arrowhead points to a highly vacuolated cell. Corresponding video is
705
available as Supplementary movie S1.
706
(B) Time-lapse images showing intracellular relocation of nuclei (red fluorescence of
707
DR5v2:H2B-tdTomato) and amyloplasts (gray particles in the bright field). Orange and
708
red arrowheads point to the nuclei localized in the proximal (upper) and the middle
709
regions of the cell, respectively. Light blue and dark blue arrowheads point to the
710
amyloplasts in the distal (bottom) and the middle regions of the cell, respectively. Purple
711
arrowheads point to the nuclei localized at the distal pole of the cells. Corresponding
712
video is available as Supplementary movie S2.
713
(C) Confocal images visualizing differential localization of organelles between the inner
714
32
and the outermost cell layers. Orange and red arrowheads point to red-fluorescent nuclei
715
in the proximal (upper) and the middle regions in the cell, respectively. Light blue and
716
dark blue arrowheads point to the amyloplasts in the distal (bottom) and the middle
717
regions in the cell, respectively. Green color indicates vacuolar membranes.
718
Time tables shown in (A) and (B) represent durations of the cell detachment process (gray
719
box). Timing of image capturing is indicated at the upper right corner of each image
720
where the origin (0 h) is set at the time when the outermost layer started detachment in
721
the proximal LRC region. Cell outlines are delineated by white dotted lines. Scale bar, 20
722
µm.
723
724
Fig. 3. Autophagosomes are formed specifically in the outermost root cap layer
725
Representative confocal time-lapse images of the 35Spro:GFP-ATG8a root. Bright-field
726
(A) and GFP-ATG8a fluorescence (B, C) images are shown. Images in (C) are magnified
727
images of the boxed regions in (B). White arrowheads in (C) indicate autophagosomes
728
marked by GFP-ATG8a. They showed the typical donut-shaped autophagosome images
729
in the later phase of detachment (red arrowhead at 1.5h, inset: enlarged view). Timing of
730
image capturing is indicated at the upper right corner of each image where the origin (0
731
h) is set at the time when the outermost layer started detachment in the proximal LRC
732
region. Scale bar, 50 µm (A, B), 20 µm (C), 2 µm (C, inset). A corresponding video is
733
available as Supplementary movie S4.
734
735
Fig. 4. CLEM imaging revealed localization of GFP-ATG8a in autophagosomes
736
33
(A, B) GFP fluorescence (A) and TEM (B) images of a section from a GFP-ATG8a root
737
cap.
738
(C-E) Magnification of the region boxed in (A) and (B). GFP-ATG8a (C), TEM (D), and
739
their merged image (E) are shown. Red arrowhead in (E) indicates an autophagosome
740
with GFP-ATG8a fluorescence.
741
(F) A 3D electron tomographic model built for an amyloplast (blue), two mitochondria
742
(brown,) and an autophagic compartment (magenta) overlaid with the TEM image.
743
Scale bar, 10 µm (A, B); 500 nm (C-F).
744
745
Fig. 5. Vacuolization and cytosol digestion were inhibited in detaching columella
746
cells in atg mutants
747
(A-D) Vacuolar morphologies in wild-type (A, B) and atg5-1 (C, D) columella cells. (A,
748
C) VHP1-mGFP fluorescence (green). (B, D) Merged images with PI-stained cell walls
749
(red).
750
(E-L) Retention of cytosol in the detaching root cap cells of various atg mutants (F-L) as
751
compared with wild type (E). Cytosol and cell walls were stained with FDA (green) and
752
PI (red), respectively.
753
(M, N) Vacuolization and cytosol digestion defects of detaching atg5-1 root cap cells
754
were complemented by the ATG5-GFP transgene (white arrowheads). Note the uniform
755
ATG5:GFP expression by the ATG5 promoter.
756
Scale bar, 10 µm (A-D); 50 µm (E-N).
757
758
34
Fig. 6. Autophagy activation is required for organized separation of the outermost
759
root cap cell layer
760
(A-C) Time-lapse images of root cap detachment processes in wild-type (A), atg5-1 (B),
761
and ATG5pro:ATG5:GFP atg5-1 (C) plants at the time points indicated at the top. Note
762
that the outermost root cap cells detach as a layer (white arrowheads) in wild type (A)
763
and ATG5:GFP atg5-1 (C), whereas they detach individually in atg5-1 (B, orange
764
arrowheads). Scale bar, 50 µm. Corresponding videos are available as Supplementary
765
movie S7-S9.
766
767
Fig. 7. Autophagy activation at the timing of cell wall degradation is sufficient for
768
organized cell separation
769
(A-D) Time-lapse images of root cap detachment processes in BRN1pro:ATG5-GFP
770
atg5-1 (A, B) and RCPGpro:ATG5:GFP atg5-1 (C, D) at the time points indicated at the
771
top right corner of each panel. Note that the outermost root cap cells detach as a cell layer
772
in both genotypes (white arrowheads), as compared with individual detachment in atg5-
773
1 (Fig. 6B). Bright-field (A, C) and GFP fluorescence (B, D) images were shown. Scale
774
bar, 50 µm. Corresponding videos are available as Supplementary movies S10 and S11.
775
776
Fig. 8. Schematic illustration of the sequence of organelle rearrangement and
777
autophagy activation during maturation and detachment of columella cells.
778
779
Fig. S1. Arabidopsis root cap cells detach at fixed intervals
780
35
(A-D) Time-lapse images showing periodic detachment of Arabidopsis root cap cells.
781
Detachment of the outermost root cap layer initiates at the proximal LRC region and
782
progressively extends toward the central columella region (B, black arrowheads).
783
Detached root cap cells adhere together to keep a cell layer morphology (C, red
784
arrowhead). Detachment of the next cell layer initiates in the same manner as the previous
785
one (D). Elapsed time after the start of culture is indicated in each panel. Scale bar, 100
786
µm.
787
(E) A time table showing periodic detachment of root cap cell layers in five (#1-5) root
788
samples each experiencing three rounds of root cap detachment. Gray, blue, and orange
789
boxes indicate the duration from the start (initial detachment at the proximal LRC region)
790
and the end (complete detachment at the columella region) of the first, second, and third
791
cell layer, respectively. The x-axis indicates elapsed time (h) from the start of culture.
792
Red lines indicate average time points of the start of detachment.
793
(F) Intervals between the start of detachment between the first and second cell layers
794
(gray bar), and between the second and third cell layer (black bar). Mean and SE are
795
shown (n = 5).
796
797
Fig. S2. Morphological transition of vacuoles during the detachment of root cap cells
798
(A, B) Time-lapse images showing vacuolar morphology by the tonoplast-localized
799
VHP1-mGFP fluorescence (A) and bright-field images (B). In the outermost cells,
800
vacuoles are initially small and fragmented (up to 17 h), and gradually expand to form
801
large central vacuoles before the cell detachment (41 h). Elapsed time after the start of
802
36
observation is indicated in each panel. A corresponding video is available as
803
Supplementary movie S3.
804
(C-E) The entire cell volume was occupied by a large central vacuole in detaching root
805
cap cells. Images of VHP1-mGFP fluorescence (C) and its overlay with a DIC image (D)
806
were shown. (F) is a Z-stack projection encompassing 50-µm depth. Note that cells at the
807
center of the detached cell layer possess large central vacuoles as visualized by VHP1-
808
mGFP (white arrowheads), whereas those at the periphery do not show fluorescence
809
(orange arrowheads) likely due to the loss of cell viability.
810
Scale bar, 20 µm.
811
812
Fig. S3. Accumulation of autophagic body-like structures in the E64d-treated wild-
813
type root cap cells and abnormal plastid morphology in atg5-1
814
(A, B) Accumulation of autophagic body-like structures inside the vacuoles of the wild-
815
type outermost root cap cells after E-64d treatment (B, orange arrowheads), as compared
816
with the translucence vacuolar images of a non-treated control (A, white arrowheads). 5-
817
day-old seedlings grown on the medium with or without 10 µM E-64d were observed.
818
Scale bar, 20 µm.
819
(C, D) Amyloplasts in the outermost root cap cells lost starch granules in both wild type
820
and atg5-1. Black arrowheads indicate the detaching outermost cell layers. Scale bar, 50
821
µm.
822
(E, F) Amyloplasts exhibit abnormal morphologies in the outermost root cap cells of
823
atg5-1 (F) as compared with those in the wild type (E). Plastids are visualized by the CT-
824
37
GFP fluorescence marker line. Note that small spherical plastids accumulate in the wild-
825
type cells (white arrowheads), whereas those with tubular morphologies dominate in
826
atg5-1 cells (orange arrowheads). Scale bar, 20 µm.
827
828
Fig. S4. Autophagosomes do not form in the detaching root cap cells of atg5-1
829
Time-lapse images of the 35Spro:GFP-ATG8a atg5-1 root tip. Bright-field (A) and GFP-
830
ATG8a fluorescence images (B, C) are shown. Images in (C) are magnified views of
831
boxed regions in (B) of respective time points. Note that the GFP-ATG8a signals were
832
uniformly distributed throughout the cytosol. Occasionally observed punctate signals did
833
not form a donut-shape typical of an autophagosome (D, E). Elapsed time after the start
834
of observation is indicated at the top. Scale bar, 50 µm (A, B); 20 µm (C); 10 µm (D, E).
835
A corresponding video is available as Supplementary movie S5.
836
837
Fig. S5. Vacuolization and cytosol digestion do not occur in detaching atg5-1 cells
838
(A, B) Time-lapse images showing vacuolar morphology by the tonoplast-localized
839
VHP1-mGFP fluorescence (A), and corresponding bright-field images (B) in atg5-1. In
840
the outermost cells, vacuoles are initially small and fragmented and gradually expand as
841
those in wild type, but fail to expand fully (43 h). Elapsed time after the start of
842
observation is indicated at the upper right corner of each panel. Corresponding video is
843
available as Supplementary movie S6.
844
(C, D) Cytosolic GUS-GFP proteins expressed under the outer layer-specific BRN1
845
promoter revealed cytosol digestion in the detaching root cap cells of wild type, as
846
38
compared with its retention in atg5-1 (white arrowheads).
847
Scale bar, 20 µm (A, B); 50 µm (C, D).
848
849
Supplementary Movie S1. Time-lapse movie showing root cap cell detachment and
850
organelle rearrangement in wild-type root cap cells
851
Scale bar, 20 µm.
852
853
Supplementary Movie S2. Time-lapse movie showing intracellular relocation of
854
nuclei (red, DR5v2:H2B-tdTomato) and amyloplasts (gray particles in the bright
855
field) in the root cap cells
856
Scale bar, 20 µm.
857
858
Supplementary Movie S3. Time-lapse movie showing morphological transition of
859
vacuoles during cell detachment
860
Scale bar, 20 µm.
861
862
Supplementary Movie S4. Time-lapse movie showing autophagosome formation in
863
the outermost root cap cells visualized by 35Spro:GFP-ATG8a
864
Scale bar, 20 µm.
865
866
Supplementary Movie S5. Time-lapse movie showing the absence of autophagosome
867
formation in 35Spro:GFP-ATG8a in atg5-1.
868
39
Scale bar, 20 µm.
869
870
Supplementary Movie S6. Time-lapse movie showing morphological transition of
871
vacuoles during cell detachment in atg5-1.
872
Scale bar, 20 µm.
873
874
Supplementary Movie S7. Time-lapse movie showing root cap cell detachment in the
875
wild type
876
Scale bar, 50 µm.
877
878
Supplementary Movie S8. Time-lapse movie showing root cap cell detachment in
879
atg5-1
880
Scale bar, 50 µm.
881
882
Supplementary Movie S9. Time-lapse movie showing root cap cell detachment in
883
atg5-1 complemented with ATG5pro:ATG-GFP
884
Scale bar, 50 µm.
885
886
Supplementary Movie S10. Time-lapse movie showing root cap cell detachment in
887
atg5-1 complemented with BRN1pro:ATG-GFP
888
Scale bar, 50 µm.
889
890
40
Supplementary Movie S11. Time-lapse movie showing root cap cell detachment in
891
atg5-1 complemented with RCPG1pro:ATG5-GFP
892
Scale bar, 50 µm.
893
894
41
Separation of living cells
by cell-wall degradation
Cleavage of LRC layer
Division of initial cells
Proximal
Distal
Columella
Lateral root
cap (LRC)
Programmed cell death
(PCD) of proximal LRC
cells
Fig. 1. A diagram illustrating structure and cell detachment process
of Arabidopsis root cap.
Landmark events constituting the cell separation sequence are marked
by arrowheads. Definition of the proximodistal polarity used in this
study is shown on the left.
42
DR5v2:H2B-tdTomato (nucleus)
B
pRPS5a:H2B:tdTomato
pRPS5a:H2
(nucleus)
Merge
VHP1
P -
P1 mGFP
VHPP1-
GFP
mG
m
(vacuole)
C
0
5
10
-5
-10
-15
-20
-5
10
20
-10
0
5
15
25
30
-4 h
0.5 h
18 h
end
(h)
(h)
A
start
-12 h
-8 h
0 h
13 h
5
0
start
Fig. 2. Organelle rearrangement takes place in the outer root cap layers
43
Fig. 2. Organelle rearrangement takes place in the outer root cap layers
(A) Time-lapse images visualizing the sequences of root cap cell detachment and relocation
of amyloplasts. Representative images before (left panel), at the beginning (central panel),
and around the end (right panel) of cell layer detachment are shown. Light blue and dark
blue arrowheads indicate sedimenting and floating amyloplasts, respectively. Green
arrowhead points to a highly vacuolated cell. Corresponding video is available as
Supplementary movie S1.
(B) Time-lapse images showing intracellular relocation of nuclei (red fluorescence of
DR5v2:H2B-tdTomato) and amyloplasts (gray particles in the bright field). Orange and red
arrowheads point to the nuclei localized in the proximal (upper) and the middle regions of
the cell, respectively. Light blue and dark blue arrowheads point to the amyloplasts in the
distal (bottom) and the middle regions of the cell, respectively. Purple arrowheads point to
the nuclei localized at the distal pole of the cells. Corresponding video is available as
Supplementary movie S2.
(C) Confocal images visualizing differential localization of organelles between the inner and
the outermost cell layers. Orange and red arrowheads point to red-fluorescent nuclei in the
proximal (upper) and the middle regions in the cell, respectively. Light blue and dark blue
arrowheads point to the amyloplasts in the distal (bottom) and the middle regions in the cell,
respectively. Green color indicates vacuolar membranes.
Time tables shown in (A) and (B) represent durations of the cell detachment process (gray
box). Timing of image capturing is indicated at the upper right corner of each image where
the origin (0 h) is set at the time when the outermost layer started detachment in the proximal
LRC region. Cell outlines are delineated by white dotted lines. Scale bar, 20 µm.
44
Bright field
GFP-ATG8
A
B
C
-24.0 h
-15.5 h
-7.0 h
1.5 h
10.0 h
18.5 h
Fig. 3. Autophagosomes are formed specifically in the outermost root cap layer
Representative confocal time-lapse images of the 35Spro:GFP-ATG8a root. Bright-field
(A) and GFP-ATG8a fluorescence (B, C) images are shown. Images in (C) are magnified
images of the boxed regions in (B). White arrowheads in (C) indicate autophagosomes
marked by GFP-ATG8a. They showed the typical donut-shaped autophagosome images in
the later phase of detachment (red arrowhead at 1.5h, inset: enlarged view). Timing of
image capturing is indicated at the upper right corner of each image where the origin (0 h)
is set at the time when the outermost layer started detachment in the proximal LRC region.
Scale bar, 50 µm (A, B), 20 µm (C), 2 µm (C, inset). A corresponding video is available as
Supplementary movie S4.
45
A
B
C
D
E
F
GFP-ATG8a
GFP-ATG8a
TEM
TEM
GFP-ATG8a + TEM
Fig. 4. CLEM imaging revealed localization of GFP-ATG8a in
autophagosomes
(A, B) GFP fluorescence (A) and TEM (B) images of a section from a
GFP-ATG8a root cap.
(C-E) Magnification of the region boxed in (A) and (B). GFP-ATG8a
(C), TEM (D), and their merged image (E) are shown. Red arrowhead in
(E) indicates an autophagosome with GFP-ATG8a fluorescence.
(F) A 3D electron tomographic model built for an amyloplast (blue), two
mitochondria (brown,) and an autophagic compartment (magenta)
overlaid with the TEM image.
Scale bar, 10 µm (A, B); 500 nm (C-F).
46
VHP1-mGFP
VHP1-mGFP + PI
WT
atg5-1
FDA (green) / PI (red)
A
B
C
D
E
F
G
H
I
J
K
L
WT
atg5-1
atg2-1
atg7-2
atg10-1
atg12ab
atg13ab
atg18a
ATG5
G -
5-GFP
ATGG5-
FP
GF
G
Cell wall
M
N
ATG5pro:ATG5:GFP
(atg5-1)
Fig. 5. Vacuolization and cytosol digestion were inhibited in detaching columella
cells in atg mutants
(A-D) Vacuolar morphologies in wild-type (A, B) and atg5-1 (C, D) columella cells. (A,
C) VHP1-mGFP fluorescence (green). (B, D) Merged images with PI-stained cell walls
(red).
(E-L) Retention of cytosol in the detaching root cap cells of various atg mutants (F-L) as
compared with wild type (E). Cytosol and cell walls were stained with FDA (green) and
PI (red), respectively.
(M, N) Vacuolization and cytosol digestion defects of detaching atg5-1 root cap cells
were complemented by the ATG5-GFP transgene (white arrowheads). Note the uniform
ATG5:GFP expression by the ATG5 promoter.
Scale bar, 10 µm (A-D); 50 µm (E-N).
47
WT
atg5-1
atg5-1 with
ATG5pro:ATG5:GFP
A
B
C
Fig. 6. Autophagy activation is required for organized separation of the outermost root
cap cell layer
(A-C) Time-lapse images of root cap detachment processes in wild-type (A), atg5-1 (B), and
ATG5pro:ATG5:GFP atg5-1 (C) plants at the time points indicated at the top. Note that the
outermost root cap cells detach as a layer (white arrowheads) in wild type (A) and
ATG5:GFP atg5-1 (C), whereas they detach individually in atg5-1 (B, orange arrowheads).
Scale bar, 50 µm. Corresponding videos are available as Supplementary movie S7-S9.
48
RCPGpro:ATG5-GFP (atg5-1)
BRN1pro:ATG5-GFP (atg5-1)
0.0 h
7.0 h
19.5 h
20.5 h
30.0 h
20.5 h
34.0 h
43.5 h
48.0 h
50.0 h
A
B
C
D
Fig. 7. Autophagy activation at the timing of cell wall degradation is sufficient for
organized cell separation
(A-D) Time-lapse images of root cap detachment processes in BRN1pro:ATG5-GFP atg5-1
(A, B) and RCPGpro:ATG5:GFP atg5-1 (C, D) at the time points indicated at the top right
corner of each panel. Note that the outermost root cap cells detach as a cell layer in both
genotypes (white arrowheads), as compared with individual detachment in atg5-1 (Fig. 6B).
Bright-field (A, C) and GFP fluorescence (B, D) images were shown. Scale bar, 50 µm.
Corresponding videos are available as Supplementary movies S10 and S11.
49
autophagosomes
vacuolization
cytosol
digestion
nuclei
translocation
cell wall
degradation
detachment
second
outermost
layer
outermost
layer
: Nucleus
: Amyloplast with starch granules (statolith)
: Shrinking amyloplast
: Vacuole
autophagy
amyloplast
floating-up
Fig.
8.
Schematic
illustration
of
the
sequence
of
organelle
rearrangement and autophagy activation during maturation and
detachment of columella cells.
50
1st to 2nd
2nd to 3rd
0
10
20
30
40
Interval (h)
37.3 h
(±2.3)
39.3 h
(±4.4)
Start of observation
End of observation
1st detachment
2nd detachment
3rd detachment
0
50
100
150
200
250
Time of culture (h)
#5
#4
#3
#2
#1
79.2 h
(±4.8)
116.5 h
(±3.5)
155.8 h
(±6.7)
3rd detachment
A
B
C
D
E
F
2nd detachment
Fig. S1. Arabidopsis root cap cells detach at fixed intervals
(A-D) Time-lapse images showing periodic detachment of Arabidopsis root cap cells.
Detachment of the outermost root cap layer initiates at the proximal LRC region and
progressively extends toward the central columella region (B, black arrowheads).
Detached root cap cells adhere together to keep a cell layer morphology (C, red
arrowhead). Detachment of the next cell layer initiates in the same manner as the
previous one (D). Elapsed time after the start of culture is indicated in each panel. Scale
bar, 100 µm.
(E) A time table showing periodic detachment of root cap cell layers in five (#1-5) root
samples each experiencing three rounds of root cap detachment. Gray, blue, and orange
boxes indicate the duration from the start (initial detachment at the proximal LRC
region) and the end (complete detachment at the columella region) of the first, second,
and third cell layer, respectively. The x-axis indicates elapsed time (h) from the start of
culture. Red lines indicate average time points of the start of detachment.
(F) Intervals between the start of detachment between the first and second cell layers
(gray bar), and between the second and third cell layer (black bar). Mean and SE are
shown (n = 5).
51
VHP1-mGFP
VHP1-mGFP + DIC
VHP1-mGFP
(Z-projection)
C
D
E
VHP1-mGFP
Bright field
5 h
11 h
17 h
23 h
29 h
35 h
41 h
47 h
A
B
5 h
11 h
17 h
23 h
29 h
35 h
41 h
47 h
Fig. S2. Morphological transition of vacuoles during the detachment of root cap
cells
(A, B) Time-lapse images showing vacuolar morphology by the tonoplast-localized
VHP1-mGFP fluorescence (A) and bright-field images (B). In the outermost cells,
vacuoles are initially small and fragmented (up to 17 h), and gradually expand to form
large central vacuoles before the cell detachment (41 h). Elapsed time after the start of
observation is indicated in each panel. A corresponding video is available as
Supplementary movie S3.
(C-E) The entire cell volume was occupied by a large central vacuole in detaching
root cap cells. Images of VHP1-mGFP fluorescence (C) and its overlay with a DIC
image (D) were shown. (F) is a Z-stack projection encompassing 50-µm depth. Note
that cells at the center of the detached cell layer possess large central vacuoles as
visualized by VHP1-mGFP (white arrowheads), whereas those at the periphery do not
show fluorescence (orange arrowheads) likely due to the loss of cell viability.
Scale bar, 20 µm.
52
Control
E-64d (proteinase inhibitor)
E
WT
35Spro:CT-GFP (plastid)
Z-projection
Starch granule (iodine staining)
A
B
E
F
F
WT
atg5-1
C
D
Fig. S3. Accumulation of autophagic body-like structures in the E64d-treated
wild-type root cap cells and abnormal plastid morphology in atg5-1
(A, B) Accumulation of autophagic body-like structures inside the vacuoles of the
wild-type outermost root cap cells after E-64d treatment (B, orange arrowheads), as
compared with the translucence vacuolar images of a non-treated control (A, white
arrowheads). 5-day-old seedlings grown on the medium with or without 10 µM E-64d
were observed. Scale bar, 20 µm.
(C, D) Amyloplasts in the outermost root cap cells lost starch granules in both wild
type and atg5-1. Black arrowheads indicate the detaching outermost cell layers. Scale
bar, 50 µm.
(E, F) Amyloplasts exhibit abnormal morphologies in the outermost root cap cells of
atg5-1 (F) as compared with those in the wild type (E). Plastids are visualized by the
CT-GFP fluorescence marker line. Note that small spherical plastids accumulate in the
wild-type cells (white arrowheads), whereas those with tubular morphologies
dominate in atg5-1 cells (orange arrowheads). Scale bar, 20 µm.
53
0 h
9 h
18 h
27 h
36 h
45 h
Bright field
GFP-ATG8
A
B
C
D
E
2 h
19.5 h
Fig. S4. Autophagosomes do not form in the detaching root cap cells of atg5-1
Time-lapse images of the 35Spro:GFP-ATG8a atg5-1 root tip. Bright-field (A) and
GFP-ATG8a fluorescence images (B, C) are shown. Images in (C) are magnified
views of boxed regions in (B) of respective time points. Note that the GFP-ATG8a
signals were uniformly distributed throughout the cytosol. Occasionally observed
punctate signals did not form a donut-shape typical of an autophagosome (D, E).
Elapsed time after the start of observation is indicated at the top. Scale bar, 50 µm
(A, B); 20 µm (C); 10 µm (D, E). A corresponding video is available as
Supplementary movie S5.
54
15 h
19 h
23 h
27 h
31 h
35 h
39 h
43 h
15 h
19 h
23 h
27 h
31 h
35 h
39 h
43 h
Bright field
VHP1-mGFP atg5-1
A
B
C
D
BRN1pro:GUS-GFP
WT
atg5-1
Fig. S5. Vacuolization and cytosol digestion do not occur in detaching atg5-1
cells
(A, B) Time-lapse images showing vacuolar morphology by the tonoplast-
localized VHP1-mGFP fluorescence (A), and corresponding bright-field images
(B) in atg5-1. In the outermost cells, vacuoles are initially small and fragmented
and gradually expand as those in wild type, but fail to expand fully (43 h).
Elapsed time after the start of observation is indicated at the upper right corner of
each panel. Corresponding video is available as Supplementary movie S6.
(C, D) Cytosolic GUS-GFP proteins expressed under the outer layer-specific
BRN1 promoter revealed cytosol digestion in the detaching root cap cells of wild
type, as compared with its retention in atg5-1 (white arrowheads).
Scale bar, 20 µm (A, B); 50 µm (C, D).
55
Supplementary Movie S1. Time-lapse movie showing root cap cell
detachment and organelle rearrangement in wild-type root cap cells
Scale bar, 20 µm.
56
Supplementary
Movie
S2.
Time-lapse
movie
showing
intracellular
relocation of nuclei (red, DR5v2:H2B-tdTomato) and amyloplasts (gray
particles in the bright field) in the root cap cells
Scale bar, 20 µm.
57
Supplementary Movie S3. Time-lapse movie showing morphological
transition of vacuoles during cell detachment
Scale bar, 20 µm.
58
Supplementary Movie S4. Time-lapse movie showing autophagosome
formation in the outermost root cap cells visualized by 35Spro:GFP-ATG8a
Scale bar, 20 µm.
59
Supplementary Movie S5. Time-lapse movie showing the absence of
autophagosome formation in 35Spro:GFP-ATG8a in atg5-1.
Scale bar, 20 µm.
60
Supplementary Movie S6. Time-lapse movie showing morphological transition
of vacuoles during cell detachment in atg5-1.
Scale bar, 20 µm.
61
Supplementary Movie S7. Time-lapse movie showing root cap
cell detachment in the wild type
Scale bar, 50 µm.
62
Supplementary Movie S8. Time-lapse movie showing
root cap cell detachment in atg5-1
Scale bar, 50 µm.
63
Supplementary Movie S9. Time-lapse movie showing
root cap cell detachment in atg5-1 complemented with
ATG5pro:ATG-GFP
Scale bar, 50 µm.
64
Supplementary Movie S10. Time-lapse movie showing root cap cell
detachment in atg5-1 complemented with BRN1pro:ATG-GFP
Scale bar, 50 µm.
65
Supplementary Movie S11. Time-lapse movie showing root cap cell detachment
in atg5-1 complemented with RCPG1pro:ATG5-GFP
Scale bar, 50 µm.
66
| 2022 | Autophagy promotes organelle clearance and organized cell separation of living root cap cells in | 10.1101/2022.02.16.480624 | [
"Goh Tatsuaki",
"Sakamoto Kaoru",
"Wang Pengfei",
"Kozono Saki",
"Ueno Koki",
"Miyashima Shunsuke",
"Toyokura Koichi",
"Fukaki Hidehiro",
"Kang Byung-Ho",
"Nakajima Keiji"
] | creative-commons |
1
Profiles of secoiridoids and alkaloids in tissue of susceptible and resistant green ash progeny
reveal patterns of induced responses to emerald ash borer in Fraxinus pennsylvanica
Robert K. Stanley1, David W. Carey2, Mary E. Mason2, Therese M. Poland3, Jennifer L. Koch2, A.
Daniel Jones4, Jeanne Romero-Severson1*
1Department of Biological Sciences, University of Notre Dame, Notre Dame, IN 46556, USA
e-mail: rstanle3@nd.edu
e-mail: jromeros@nd.edu
2Northern Research Station, Forest Service, U.S. Department of Agriculture, Delaware, OH
43015, USA
e-mail: david.carey@usda.gov
e-mail: mary.mason@usda.gov
e-mail: jennifer.koch@usda.gov
3Northern Research Station, Forest Service, U.S. Department of Agriculture, Lansing, MI 48910,
USA
e-mail: therese.poland@usda.gov
4Department of Biochemistry and Molecular Biology, Michigan State University, East Lansing,
MI 49503, USA
2
e-mail: jonesar4@msu.edu
*Corresponding author
Classification: Plant Biology, Chemistry
Keywords: Emerald Ash Borer, Fraxinus pennsylvanica, Invasive Species, Plant Defenses,
Untargeted Metabolomics,
Preprint Server: Biorxiv
This PDF file includes
Main Text
Figures 1 to 5
Table 1
3
Abstract
The emerald ash borer (Agrilus planipennis, EAB) invasion in North America threatens most
North American Fraxinus species, including green ash (F. pennsylvanica), the mostly widely
distributed species (1, 2). A small number of green ash (“lingering ash”, 0.1-1%) survive years of
heavy EAB attack (3) and kill more EAB larvae when challenged in greenhouse studies than
susceptible controls (4). We combined untargeted metabolomics with intensive phenotyping of
segregating F1 progeny from susceptible or lingering ash parents to detect chemotypes
associated with defensive responses to EAB. We examined three contrasting groups: low larval
kill (0-25% of larvae killed), high larval kill (55-95% of larvae killed) and uninfested. Contrasting
the chemotypes of these groups revealed evidence of an induced response to EAB. Infested
trees deployed significantly higher levels of select secoiridoids than uninfested trees. Within the
infested group, the low larval kill (LLK) individuals deployed significantly higher levels of select
secoiridoids than the high larval kill (HLK) individuals. The HLK individuals deployed significantly
higher concentrations of three metabolites annotated as aromatic alkaloids compared to the
LLK and uninfested individuals. We propose a two-part model for the North American Fraxinus
response to EAB wherein every individual has the capacity to detect and respond to EAB, but
only certain trees mount an effective defense, killing enough EAB larvae to prevent or minimize
lethal damage to the vascular system. Integration of intensive phenotyping of structured
populations with metabolomics reveals the multi-faceted nature of the defenses deployed in
naïve host populations against invasive species.
Significance
4
Long-lived forest trees employ evolutionarily conserved templates to synthesize an array of
defensive metabolites. The regulation of these metabolites, honed against native pests and
pathogens, may be ineffective against novel species, as illustrated by the high mortality (>99%)
in green ash infested by the invasive emerald ash borer (EAB). However, high standing genetic
variation may produce a few individuals capable of an effective defense, as seen in the rare
surviving green ash. In an investigation of this plant-insect interaction, we annotated
metabolites associated with generalized but ineffective responses to EAB, and others
associated with successful defensive responses. Untargeted metabolomics combined with
intensive phenotyping of structured populations provides a framework for understanding
resistance to invasive species in naïve host populations.
INTRODUCTION
Invasive pests and pathogens, now widely dispersed through globalization, threaten
nearly two thirds of North American forests (1). Exacerbated by climate change, these
increasingly severe infections and infestations destabilize forested ecosystems and inflict
billions of dollars in direct costs to individuals and local communities (5-14). Two of these
pathogens and pests, chestnut blight (Cryphonectria parasitica) and emerald ash borer (Agrilus
planipennis, EAB) have had a profound impact on public awareness: chestnut blight because
this disease caused the ecological extinction of the iconic American chestnut (Castanea
dentata) and EAB because of the widespread, rapid and continuing loss of ash trees from
streets, parks, and forests (14). The severe impact of the loss of chestnut, pales in comparison
5
to the economic and ecological damage already inflicted by EAB, the most destructive and
economically devastating invasive insect pest of forest trees in North American history (15).
EAB, a beetle native to Asia, was discovered in Michigan, United States and in Ontario,
Canada in 2002 (16). EAB attacks ash (Fraxinus) species; larvae hatch from eggs laid in bark
furrows and burrow into the living tissue directly beneath the bark (2). The larvae feed on the
vascular cambium, cork cambium, phloem, and xylem inflicting severe vascular damage that
ultimately kills the tree. Larvae feed during the summer and the fall and may take one or two
years to complete development. The larvae pass through four developmental instars and then
chew a pupal chamber in the outer sapwood or inner bark in which they overwinter as mature
larvae folded over in a J-shape. During the spring, they enter the pupal stage and transform into
adults that emerge in late spring and early summer through characteristic D-shaped exit holes
(17, 18).
EAB infestation has resulted in the rapid loss of hundreds of millions of green ash, not
only in forests and rural areas but also in cities, where green ash was once one of the most
widely planted street and park trees in the United States (19, 20). Green ash, the most widely
distributed Fraxinus species in North America, is a dioecious, diploid, and deciduous tree
species native to the eastern and central United States and eastern Canada (16, 18, 20).
Mortality in green ash from EAB infestation can approach 100% within six years of the local
detection of EAB (20). EAB invasion threatens not only green ash, but survival of the majority of
native North American Fraxinus species including white (F. americana), pumpkin (F. profunda),
Carolina (F. caroliniana) and black ash (F. nigra) (20, 21). Fraxinus are ecologically important in a
wide range of forested ecosystems and are also extensively utilized for soil conservation, rural
6
water management, riparian zone stabilization, flood control, and urban green spaces in North
America (19).
Long term forest plot monitoring initiated in 2005, two years (3, 20, 22) after the initial
detection of EAB in North America, revealed a small number of green ash (0.1-1%) that survive
for years after all other surrounding green ash have died (3). These “lingering ash” (L) have
been and continue to be propagated as potential sources of genetic resistance for a breeding
program (23, 24). Eleven years of replicated egg bioassay tests, conducted by placing controlled
densities of EAB eggs on test trees and monitoring larval development and survival, revealed
reproducible larval kill capabilities with phenotypic distributions among trees that suggest
quantitative inheritance (24, 25).
Durable genetic resistance in the host, the most effective control measure for any pest
or pathogen (26, 27), was not initially considered a strategic goal for saving North American
Fraxinus species from EAB. The assumption was that a species cannot have any resistance to a
pest with which it has not coevolved (28, 29). However, many studies have shown that in many
cases native species do marshal heritable defensive responses to non-native invaders (30, 31).
Successful breeding programs have produced American beech (Fagus grandifolia) resistant to
beech bark disease (Neonectria spp transmitted by Cryptococcus fagisuga) (32), eastern white
pine (Pinus strobus) resistant to white pine blister rust (Cronartium ribicola) (33) and Port
Orford cedar (Chamaecyparis lawsoniana) (22) resistant to the root rot pathogen Phytophthora
lateralis. The success of these and other programs demonstrates that heritable resistance exists
in wild populations and can be used to develop resistant populations for species restoration
through breeding (30). Once a genetic component is confirmed, a detailed study of the
7
mechanisms of the response can contribute to a body of knowledge on the omics of heritable
defensive responses.
Previous investigations on the role of secondary metabolites as defenses against EAB
have focused on comparing small numbers of cultivars from susceptible Fraxinus spp. to the
naturally resistant F. mandshurica cultivar ‘mancana’(34). Application of methyl jasmonate in
infested susceptible F. americana individuals induced production of verbascoside and
suppressed EAB larval development (35).These studies collectively proposed a positive
association between lignan glycosides and host resistance, particularly pinoresinol and
verbascoside, as well as suggesting a role for secoiridoid glycosides (36).
Secoiridoids are also implicated in the response of European Ash (F. excelsior) to ash
dieback disease (ADB) caused by Hymenoscyphus fraxineus. Ash dieback ultimately infects the
woody stem tissue, killing the tree (37, 38). High concentrations of specific secoiridoids were
identified with tolerant genotypes in one study, and with susceptible genotypes in another.
Both groups of investigators proposed that the different levels of secoiridoids are the result of
differential transcriptional regulation (39, 40). Investigations of ash dieback phenotypes, in
these and other studies suggest that susceptibility to ADB in F. excelsior, is a quantitative trait
(41).
Other recent investigations of resistance to wood-boring insects have shown that some
trees utilize secondary metabolite-based constitutive and induced defensive responses against
specific insect pests (42-44). The concentration and profiles of certain plant secondary
metabolites strongly predict resistance in maritime pine (Pinus pinaster) to the pine weevil
(Hylobius abietis), after accounting for genetic relatedness among the host trees (42). Other
8
investigations have shown that the response consists of altered rates of synthesis for existing
metabolites, rather than the synthesis of unique compounds (42, 45).
Here we combine untargeted metabolomics and intensive phenotyping on structured
populations using an experimental design that accounts for the confounding effect of genetics
and environment to detect chemotypes associated with defensive responses to EAB. We
hypothesized that the full sibling progeny of Susceptible x Susceptible (SxS), and Lingering x
Lingering (LxL) parents would produce a wide range of larval kill phenotypes and that the family
means of the progeny from LxL parents would be significantly higher than the family means of
the progeny of SxS parents. If both these hypotheses are correct, and the defense is associated
with secondary metabolites, we expect a contrast in chemotypes between the high larval kill
(HLK, tree defenses killed 55-95% of larvae) and low larval kill (LLK, tree defenses killed 0-25%
of larvae) phenotypes. If infestation induces a response, we expect that the chemotypes of
infested individuals will be distinct from uninfested individuals within families. Our data showed
that some secondary metabolites including select secoiridoids occurred at higher
concentrations in infested individuals regardless of larval kill phenotype, while a smaller
number of compounds, annotated as aromatic alkaloids were found in higher concentrations in
high percent larval kill individuals. Our work will spur future investigations for the molecular
basis of durable genetic resistance to EAB in green ash and provide a framework for discovering
resistance to invasive species in naïve host populations.
RESULTS
9
Analysis of EAB-resistance in full-sibling families of reveals that resistance to EAB in green ash
is a multigenic quantitative trait
Seedlings (2-3 years old) from two green ash F1 families produced through crosses
between lingering parents (LxL) and one family produced by a cross between susceptible
parents (SxS) were infested with EAB to confirm the genetic basis of the larval kill phenotype
(Figure 1a).. One-way ANOVA and Tukey-Kramer multiple comparison tests revealed that the
mean percent larval kill of (LxL) families Pe-Y and Pe-Z were significantly different from the
mean percent larval kill of the (SxS) family Pe-C (p < 0.01), but there was no significant
difference among the L x L families’ means (Figure. 1b). The shape and range of the larval kill
distributions strongly suggests that the phenotype is a quantitative trait and provides support
for the hypothesis of complex inheritance (Figure. 1b).
Each family produced a range of larval kill phenotypes, (Pe-C: 0-44%, Pe-Y: 8-95%, Pe-Z:
0-75%). Based on the distribution of phenotypes across families (Figure. 1b), we classified
individual trees with larval kill values of 25% or lower as LLK and those with larval kill values
greater than 55% as HLK. The value of 55% is higher than the highest larval kill value for the
collection of lingering ash parents described in a previous report, and the value of 25% is higher
than the parents of family Pe-C, and most of the progeny (88%) in the susceptible family Pe-C
(Figure 1) (24). As a comparison, the resistant Asian ash F. mandshurica typically kills 80-90% of
EAB larvae when tested with the egg bioassay (24). The lingering families in this study included
some progeny that performed similarly to resistant Manchurian ash individuals.
Generation of untargeted metabolomic profiles.
10
We produced untargeted metabolomic profiles from acetonitrile:isopropanol:water
extractions using ultra-high performance liquid chromatography/ high resolution mass
spectrometry (UHPLC/MS). The levels of metabolites were normalized to a constant internal
standard and replicated, with a constant mass of tissue extracted. An analysis of the relative
standard deviation (RSD) of pooled controls had a median of 29.8% for all features considered
in downstream analyses (Figure S1)
Metabolite based OPLS-DA models correctly identify progeny classes.
We conducted pairwise comparisons of the metabolite profiles of HLK, LLK, and
uninfested (UNI) individuals within families to determine if metabolites were associated with
infestation status or the larval kill phenotype. We assessed 194 metabolite features (Figure 2)
with pairwise one-way analysis of variance (ANOVA) tests for the following contrasts: Family C:
UNI vs LLK; Family Y: UNI vs LLK, UNI vs HLK, LLK vs HLK; Family Z: UNI vs LLK, UNI vs HLK, LLK vs
HLK. Between 9 and 49 features were significant (p < 0.05) in each comparison (Figure 2, Table
S2).
We then took these features and used orthogonal partial least squares-discriminant
analyses (OPLS-DAs) to examine their ability to accurately identify the correct progeny
classification (Figure 2). To prevent overfitting of our model we performed a three-fold cross
validation on our data and report the average prediction accuracies as the performance of our
model. Overall, our model performed quite well, with over 70 % of individuals correctly
assigned to progeny larval-kill phenotype class across all models, with the majority of other
individuals being unclassified, not incorrectly assigned (Figure 3a, Table S2). Our confidence in
these models was supported by principal component analyses (PCAs) yielding similar
11
separations(Figures 3b, 3c), suggesting that OPLS-DAs are producing statistically meaningful
group separations (46). This workflow can serve as a template for assessing the relationship of
chemotypes and complex phenotypes in a non-model system.
Chemotypes across families distinguish a general defense response from a successful defense
response.
We focused our attention on the 32 features that had a significant p-value in more than
one family’s comparisons or were suggested as important for EAB defense in previous
investigations. This latter category included verbascoside (35) and salidroside (47). We
generated electrospray ionization tandem mass spectra (ESI-MS/MS) for the features detected
in our analysis and annotated them based on comparisons with MS/MS databases including the
Massbank of North America, along with published literature and purchased standards. The
annotation confidence is labeled according to the recommendations of the Metabolomics
Standards Initiative (MSI) (48). Our annotations revealed compounds from a wide variety of
chemical families, including the first record of specific alkaloids present in green ash tissue
(Table 1, Figure 4, Figure 5).
We annotated 12 secoiridoids with similar structures to secoiridoids previously
hypothesized to be indicative of a resistance mechanism (36). We found that these secoiridoids
were elevated in both low and high larval kill phenotypes compared to uninfested controls
(Figure S2). However, five secoiridoids had significantly higher concentrations in low larval kill
phenotypes compared to high larval kill phenotypes with no significant difference in
concentration in the other seven secoiridoids (Figure 5b-d).
12
One secoiridoid (m/z 569.23) was elevated in all infested comparisons (LLK v UNI, HLK v
UNI) across all families, and may have some value as an indicator of infestation across many ash
genotypes. Another two secoiridoids, nueznehide (m/z 704.2781) and GL5 (m/z 928.3429),
found in higher concentrations in trees that are highly susceptible to ash dieback in previous
investigations (39, 49), were significantly higher in low larval kill individuals compared to
uninfested individuals across families. Additionally, we found that concentrations of
verbascoside, a phenylethanoid glycoside also proposed as a component of the resistance
response (36), were highest in low larval kill individuals. Overall, our data suggests that these
specific secoiridoids and verbascoside may be indicative of a general wound response, but do
not appear to be responsible for the high larval kill phenotype. The only compounds that were
higher in high larval kill individuals compared to low larval kill individuals were three
compounds annotated as aromatic alkaloids and the phenylethanoid glycoside salidroside
(Table 1, Figure 5, Figure S3). These alkaloids are the first reported in green ash and suggest a
novel role of alkaloid in defense against herbivory in forest trees.
DISCUSSION
We investigated the ability of select green ash individuals to respond to EAB using
structured populations, a reproducible phenotyping method, and an untargeted metabolomics
approach. We found that all green ash seedlings analyzed displayed metabolic changes in
response to infestation, but in most individuals, this response was ineffective to kill many EAB
larvae. OPLS-DA and multivariate analyses showed that high and low performing individuals had
chemotypes distinct from each other and from uninfested individuals. These chemotypes are
distinguishable based on the relative concentrations of select metabolites (Table 2, Figure 5,
13
Table S1), not their presence or absence, suggesting genetic regulation of multiple synthesis
pathways may be responsible for the high larval kill phenotype. We provide an initial
annotation of metabolites for further study, including secoiridoids that may prove to be reliable
indicators of infestation across all genetic backgrounds, and three aromatic alkaloids that may
be part of an effective defensive response.
Defensive responses based on multigenic mechanisms confer durable genetic
resistance, the most effective control measure for any pest or pathogen. In our study, full
sibling F1 progeny of lingering ash parents performed better on average than the F1 progeny of
susceptible parents and produced progeny with phenotypes ranging from 0 % larvae killed to
95% larvae killed. This is the result expected when a phenotype is the result of complex genetic
mechanisms involving multiple loci (50). A multigenic mechanism for the lingering ash
phenotype is also consistent with two recent candidate gene studies, one on the pan-genome
of EAB resistant Fraxinus species, and the other which utilized the 2021 release of the green ash
genome (37, 51). Additional studies will be necessary to fully elucidate the genetic architecture
of these defensive responses. Although we did not examine other components of the lingering
ash phenotype, including adult feeding preferences or attractiveness of egg-laying sites to
female EAB, our controlled greenhouse experiments did allow us to examine the defensive
mechanisms deployed in the woody tissues, where the primary host insect interaction occurs.
The chemotypes of high larval kill, low larval kill, and uninfested individuals from the
same parents could be distinguished based on relative concentrations of groups of metabolites.
Comparisons of the same contrast across multiple families reveals that secoiridoids are
associated with a generalized infestation response that does not predict effective defensive
14
responses. This association is consistent with previous studies that suggested that infested
trees, or trees artificially stressed with methyl jasmonate produced higher amounts of these
metabolites (52). High concentrations of specific secoiridoids in F. excelsior are proposed to be
indicative of tolerance (40) or susceptibility to ash dieback (37), and were predicted to provide
a future robust reservoir of anti-feeding deterrents to EAB (49). Our data suggests that these
specific secoiridoids function best as indicators of a generalized stress response and not
necessarily of resistance to EAB. The study design allowed us to disentangle the generalized
stress response from an effective defense response as indicated by percent larval kill. Our
results suggest that part of the effective defensive response may consist of four metabolites,
annotated as three aromatic alkaloids and salidroside, that were significantly elevated in high
larval kill individuals compared to low larval kill or uninfested individuals. Overall, our study has
distinguished, for the first time, between an effective defensive response and a generalized
defensive response to EAB.
Based on our results, we propose a two-part model for the North American Fraxinus
response to EAB wherein every individual has the biochemical capacity to synthesize chemical
defenses as a response to EAB, but only certain trees deploy an effective induced defense
response that kills enough EAB larvae to prevent or minimize lethal damage to the vascular
system. This model is consistent with forest observations and controlled studies that show most
individuals in North American ash species can kill a few larvae, but cannot withstand a heavy
infestation (24, 25). The high concentrations of secoiridoid glycosides in infested individuals,
especially those with the most live larvae, suggests that even susceptible ash trees detect that
they have been wounded by EAB larvae, and attempt to respond but are unable to do so in a
15
manner that results in effectively killing the larvae. A previous study demonstrated that
application of methyl jasmonate induced a defensive response and suppressed EAB larval
development or killed larvae in susceptible Fraxinus individuals (35), supporting the hypothesis
that even susceptible trees have the necessary synthetic machinery, but lack the ability to
conduct a tailored reconfiguration of their metabolism, as outlined by Schuman and Baldwin
(53), to kill the EAB larvae.
This study provides a list of metabolites that could be targeted in future work focusing
on the response of green ash to EAB. Key questions for future experiments include determining
if the compounds identified extend to additional lingering ash families and gaining a better
understanding of the timing and spatial distribution of effective defense responses. Additional
phenotypic, genomic, proteomic, transcriptomic, and metabolomic analyses will benefit from
the recent release of the green ash genome (51). This future work on the interaction of green
ash and EAB will contribute to our understanding of how forest trees recognize and defend
themselves against stem-boring insects.
In summary, our data supports the hypothesis that the high larval kill phenotype is a
multi-genic and heritable trait. We have also shown that green ash responds to EAB infestation
with increased concentrations of secoiridoids, regardless of the larval kill phenotype. While
infestation with EAB induces a response in all green ash tested, the induced response is
ineffective in most cases. In the individuals that mount a successful response, we found higher
concentrations of three aromatic alkaloids and salidroside, a result that merits further
investigation. Similar metabolites were seen across all phenotypes, but the concentrations
varied, suggesting that the high larval kill phenotype is based on complex regulatory
16
mechanisms. Elucidation of the genetic mechanisms driving defensive responses to EAB in
green ash will be an essential part of a multidisciplinary effort for saving North American
Fraxinus species and guide future investigations of resistance in native species to invasive
threats.
Materials & Methods
Study System and Phenotyping
Green ash were selected in the forest based on two criteria: 1) a healthy canopy at least
two years after the mortality rate of the stand exceeded 95 percent, and 2) a minimum
diameter at breast height (DBH, 1.37 m from the ground) of 26 cm, indicating they were over
the minimum size preferred by EAB when the infestation was at peak levels (24). These
‘lingering ash’ trees show evidence of less severe emerald ash borer infestation compared to
susceptible phenotypes in the forest, often accompanied by evidence of vigorous wound
healing, and maintain a healthy crown for years after local conspecifics have died(3, 54). Over
the last 14 years, individuals meeting these criteria have been clonally propagated through
grafting and subjected to greenhouse bioassays that provided evidence of the ability of some
selected lingering ash trees to mount defensive responses against EAB (24). Although there is
evidence of multiple types of defenses, this work is focused on EAB egg bioassays (described
below) to assess host defenses that result in larval mortality. Clonal replicates of lingering green
ash genotypes, some used as parents in this study, consistently kill more early instar larvae (35
to 50 percent) than the susceptible green ash controls (0 to 10 percent) (24).
Plant Material
17
Plant material was comprised of 97 two-year-old potted F. pennsylvanica seedlings
reared in an outdoor growing area, then transferred into an environmentally controlled
greenhouse in the spring of the treatment year to allow acclimatization prior to the start of the
EAB treatment. The individuals tested were generated by controlled cross-pollinations of
lingering or susceptible green ash to produce full sibling families of known parentage.
Individuals belonged to one of three families: Pe-C (21 individuals, susceptible parentage Pe-97
x “Summit”), Pe-Y (42 individuals, lingering parentage, Pe-53 x Pe-56), or Pe-Z (35 individuals,
lingering parentage Pe-53 x Pe-59). Both susceptible parents, (Pe-97) and the cultivar
“Summit”, had susceptible phenotypes in egg bioassays and did not persist on the landscape
after the arrival of EAB. “Summit”, in particular, has been proven susceptible in our egg
bioassay (16 replications), in common garden studies (55), and by its rapid demise under
natural EAB infestation in city streets and parks (16).
Emerald Ash Borer Resistance Bioassays
EAB eggs were raised and prepared as described in Koch et al 2015 (24). Twelve eggs
were applied to each tree at a density of 400 eggs per square meter, as previously described
(24). Eight weeks after eggs were applied, larval galleries were carefully dissected, starting at
the entry hole, and followed to determine the outcome of each larva that successfully hatched
and entered the tree. Larvae were designated as alive, tree-killed (killed by a host defense
response), or dead by other means such as parasitism, cannibalism, or fungal infection. The
proportion of tree-killed larvae was calculated based on the total number of larvae that
hatched and entered the tree. One-way ANOVA and Tukey-Kramer multiple comparison tests
were used to analyze the performance of families Pe-Y, Pe-Z and Pe-C.
18
Metabolite analyses of F. pennsylvanica woody tissue by UHPLC-MS.
Trees were destructively sampled to collect tissue for metabolite analyses eight weeks
after eggs were placed, during phenotyping. The entire stem, 2.5 cm above the highest EAB
larval galleries, was collected and stored immediately on dry ice, before being transferred to -
80°C storage. This ensured the collection of the vascular cambium, the cork cambium, the
phloem, and the ray parenchyma. All samples were ground under liquid nitrogen in a Spex
Sample Prep freezer mill and stored at -80°C prior to extraction
For each sample, 1 g of frozen powdered plant tissues was extracted in 10 ml of
acetonitrile/isopropanol/water (3:3:2) containing 1.00 mM telmisartan (internal standard) and
0.01% formic acid and incubated in the dark at 4°C for 24 h. samples were then centrifuged at
4°C and 10,000g for 10 minutes, supernatants were transferred to fresh tubes, and 50:1 diluted
aliquots were prepared by adding deionized water. An additional aliquot of undiluted extracted
sample has been archived at -80 ˚C.
UHPLC/MS analyses were performed using a Shimadzu LC-20AD ternary pump coupled
to a SIL-5000 autosampler, column oven, and Waters Xevo G2-XS QTof mass spectrometer
equipped with an electrospray ionization source. The operation parameters for the positive-ion
mode analyses are as previously detailed(56). A 10- µL volume of each diluted extract was
analyzed using a 20-minute gradient method on an Ascentis Express C18UHPLC column
(2.1x100mm, 2.7µm) with mobile phases consisting of 10 mM ammonium formate in water,
adjusted to pH 2.8 with formic acid (solvent A) and acetonitrile (solvent B). The 20-min method
gradient was as follows: 1% B at 0.00 to 1.00 min, then step to 5% B at 1.01 min, linear gradient
to 25% B at 8.00 min, then a linear gradient to 75% B at 12.50 min, another linear gradient to
19
98% B at 15.00 min, and a hold at 98% B until 18.00 min, a step to 1% B at 18.01 min, and a hold
at 1% B until 20.00 min.
Analyte samples were injected in a randomized order while process blank and quality
control samples were injected at regular intervals. All calculated peak areas were normalized to
the peak area for the internal standard telmisartan utilizing Progenesis QI v2.4software
(Nonlinear Dynamics Ltd., Newcastle, UK). Standards of oleuropein, apigenin and salidroside
were purchased from Sigma Aldrich, prepared in the extraction solvent, and run at 5 µg/mL.
Untargeted Metabolomics Data Processing
For untargeted metabolomic analysis, data were initially processed using Progenesis
software. Leucine enkephalin lockmass correction (m/z 556.2766) was applied during run
importation and all runs were aligned to retention times of a bulk pool run automatically
selected by the software from a selection of QC samples. Peak picking and deconvolution was
conducted as previously described (57). After deconvolution, 1,278 compound ions remained.
To remove features from the dataset introduced by solvents, glassware, or instrumentation and
to remove lipids, several filters were applied to the 1,278 compound ions remaining after
deconvolution. Concentrations of each feature were normalized to the internal telmisartan
standard (m/z = 515.2448). Compounds with the highest mean abundance in process blank
samples, maximum abundance less than 0.1% of the most abundant compound in the dataset,
or retention times greater than 16 minutes were excluded from the dataset. This reduced the
total number of metabolic features to 323. Further analysis and statistical comparisons of
compound signals extracted by Progenesis QI software was executed using EZinfo v3.0.2
software (Umetrics, Umeå, Sweden).
20
One way analysis of variance (ANOVA) tests were used to assess significance between
each pairwise comparison for individual metabolic features in the seven following contrasts:
Family C: UNI vs LLK ; Family Y: UNI vs LLK, UNI vs HLK, LLK vs HLK; Family Z: UNI vs LLK, UNI vs
HLK, LLK vs HLK. Features that were significant (p < 0.05) were included in pairwise orthogonal
partial least squares discriminant analysis (OPLS-DA) and principal component analysis (PCA)
analyses (Table S1). OPLS-DAs and PCAs were run using pareto scaling. To prevent overfitting of
our model we performed a threefold cross validation on our data and took the averages as the
performance of our model. For all metabolic features extracted with Progenesis QI and used in
downstream analyses with EZinfo, spectra were processed using MassLynx v4.2 software
(Waters Corporation, Milford, MA,USA) as previously detailed (57) (Table S2).
Of the metabolites considered, thirty-two had significance in more than one family, or
had a previously proposed purpose and were annotated. Annotation of the electrospray
ionization tandem mass spectrometry (ESI-MS/MS) data relied on comparisons with MS/MS
databases such as the Massbank of North America as well as previous studies and purchased
standards. The confidence levels in the metabolite annotation were following
recommendations of the Metabolomics Standards Initiative (48). The quantities present in
individual tissue extracts were too small for complete structure elucidation.
Acknowledgements
Acknowledgments: The authors thank Warren Chatwin and Christina Murray for their helpful
comments on the manuscript. The authors thank Aletta Doran, Julia Wolf, Gavin Nupp, Miranda
McKibben, and Jarod Sanchez for their work propagating and maintaining the study trees and
21
their assistance conducting the EAB resistance bioassays. The authors also thank Patrick
Cunniff, Brandon Chou, Kingsley Owusu Otoo and Julie Huston for assistance in collecting and
organizing tissue samples and managing logistics. J.R-S acknowledges support from USDA-USFS
APHIS grants 18-IA-11242316-105 and 20-JV-11242303-050. J.R-S also acknowledges support
from the Tree Fund Foundation, Tree Fund grant 18-JD-01. R.K.S. acknowledges support from
NIH training grant T32GM075762. JK acknowledges support from USDA APHIS 18-IA-11242316-
105, Michigan Invasive Species Grant Program grant IS18-119, the Commonwealth of
Pennsylvania Department of Conservation and Natural Resources Bureau of Forestry 18-CO-
11242316-014, and the U.S Forest Service Special Technology Development Program grant NA-
2017-01. A.D.J. acknowledges support from Michigan AgBioResearch through the USDA
National Institute of Food and Agriculture, Hatch project number MICL02474, and USDA-USFS
grant 20-JV-11242303-050.
Competing Interest Statement
All the authors declare that they have no competing interests.
Data Availability
The data that support the findings of this study are available upon request from the
corresponding author. The raw data will also be submitted to MetaboLights or similar
repository.
References
1.
K. M. Potter, M. E. Escanferla, R. M. Jetton, G. Man, Important Insect and Disease
Threats to United States Tree Species and Geographic Patterns of Their Potential
Impacts. Forests 10, 304 (2019).
22
2.
D. G. McCullough, Challenges, tactics and integrated management of emerald ash borer
in North America. Forestry: An International Journal of Forest Research
10.1093/forestry/cpz049, cpz049 (2019).
3.
K. S. Knight et al. (2012) Dynamics of surviving ash (Fraxinus spp.) populations in areas
long infested by emerald ash borer (Agrilus planipennis). in Proceedings of the fourth
international workshop on the genetics of host-parasite interactions in forestry: Disease
and insect resistance in forest trees (Pacific Southwest Research Station, Forest Service,
U.S. Department of Agriculture, Albany, CA), pp 143-152.
4.
J. L. Koch, D. W. Carey, M. E. Mason, T. M. Poland, K. S. Knight, Intraspecific
variation in Fraxinus pennsylvanica responses to emerald ash borer (Agrilus planipennis).
New Forests 46, 995-1011 (2015).
5.
C. F. Miniat et al., "Impacts of Invasive Species on Forest and Grassland Ecosystem
Processes in the United States" in Invasive Species in Forests and Rangelands of the
United States, T. M. Poland et al., Eds. (Springer International Publishing, Cham, 2021),
pp. 41-55.
6.
W. R. L. Anderegg et al., Climate-driven risks to the climate mitigation potential of
forests. Science 368, eaaz7005 (2020).
7.
J. A. Hicke et al., Effects of biotic disturbances on forest carbon cycling in the United
States and Canada. Glob Change Biol 18, 7-34 (2012).
8.
K. Hoover, A. A. Riddle, Forest carbon primer. Congressional Research Service:
Washington, DC, USA (2020).
9.
K. F. Kovacs et al., Cost of potential emerald ash borer damage in U.S. communities,
2009–2019. Ecological Economics 69, 569-578 (2010).
10.
J. E. Aukema et al., Economic impacts of non-native forest insects in the continental
United States. PLoS one 6, e24587 (2011).
11.
C. J. A. Bradshaw et al., Massive yet grossly underestimated global costs of invasive
insects. Nature Communications 7, 12986 (2016).
12.
G. M. Lovett et al., Nonnative forest insects and pathogens in the United States: Impacts
and policy options. Ecol Appl 26, 1437-1455 (2016).
13.
F. Krist Jr et al., National Insect and Disease Forest Risk Assessment: 2013-2027. US
Department of Agriculture. FHTET-14-01 (2014).
14.
K. Potter, M. Escanferla, R. Jetton, G. Man, Important Insect and Disease Threats to
United States Tree Species and Geographic Patterns of Their Potential Impacts. Forests
10, 304 (2019).
15.
K. M. Potter, M. E. Escanferla, R. M. Jetton, G. Man, B. S. Crane, Prioritizing the
conservation needs of United States tree species: Evaluating vulnerability to forest insect
and disease threats. Global Ecology and Conservation 18, e00622 (2019).
16.
T. M. Poland, D. G. McCullough, Emerald ash borer: invasion of the urban forest and the
threat to North America’s ash resource. Journal of Forestry 104, 118-124 (2006).
17.
D. A. Herms, D. G. McCullough, Emerald ash borer invasion of North America: history,
biology, ecology, impacts, and management. Annual review of entomology 59, 13-30
(2014).
18.
G. Popkin, Rising from the ashes Science 370, 756-759 (2020).
19.
B. B. Hanberry, Rise of Fraxinus in the United States between 1968 and 20131. The
Journal of the Torrey Botanical Society 141, 242-249 (2014).
23
20.
K. S. Knight, J. P. Brown, R. P. Long, Factors affecting the survival of ash (Fraxinus
spp.) trees infested by emerald ash borer (Agrilus planipennis). Biological Invasions 15,
371-383 (2013).
21.
Anonymous (IUCN 2021. The IUCN Red List of Threatened Species. Version 2021-1.
https://www.iucnredlist.org. Downloaded on 6/27/2021.
22.
R. A. Sniezko et al., Proceedings of the fourth international workshop on the genetics of
host-parasite interactions in forestry: Disease and insect resistance in forest trees. Gen.
Tech. Rep. PSW-GTR-240. Albany, CA: Pacific Southwest Research Station, Forest
Service, US Department of Agriculture. 372 p 240 (2012).
23.
J. L. Koch et al. (2012) Breeding strategies for the development of emerald ash borer-
resistant North American ash. in In: Sniezko, Richard A.; Yanchuk, Alvin D.; Kliejunas,
John T.; Palmieri, Katharine M.; Alexander, Janice M.; Frankel, Susan J., tech. coords.
Proceedings of the fourth international workshop on the genetics of host-parasite
interactions in forestry: Disease and insect resistance in forest trees. Gen. Tech. Rep.
PSW-GTR-240. Albany, CA: Pacific Southwest Research Station, Forest Service, US
Department of Agriculture. pp. 235-239., pp 235-239.
24.
J. Koch, D. Carey, M. Mason, T. Poland, K. Knight, Intraspecific variation in Fraxinus
pennsylvanica responses to emerald ash borer (Agrilus planipennis). New Forests 46,
995-1011 (2015).
25.
J. Romero-Severson, J. L. Koch (2017) Saving green ash. in Proceedings of Workshop on
Gene Conservation of Tree Species-Banking on the Future, May 2016, pp 102-110.
26.
R. A. Sniezko, J. Koch, Breeding trees resistant to insects and diseases: putting theory
into application. Biological Invasions 19, 3377-3400 (2017).
27.
M.-L. Desprez-Loustau et al., An evolutionary ecology perspective to address forest
pathology challenges of today and tomorrow. Annals of Forest Science 73, 45-67 (2016).
28.
K. J. K. Gandhi, D. A. Herms, Direct and indirect effects of alien insect herbivores on
ecological processes and interactions in forests of eastern North America. Biological
Invasions 12, 389-405 (2010).
29.
A. M. Mech et al., Evolutionary history predicts high‐impact invasions by herbivorous
insects. Ecology and evolution 9, 12216-12230 (2019).
30.
C. C. Pike, J. Koch, C. D. Nelson, Breeding for Resistance to Tree Pests: Successes,
Challenges, and a Guide to the Future. Journal of Forestry 119, 96-105 (2020).
31.
R. A. Sniezko, J.-J. Liu, Prospects for developing durable resistance in populations of
forest trees. New Forests, 1-17 (2021).
32.
J. L. Koch, R. L. Heyd, Battling beech bark disease: establishment of beech seed orchards
in Michigan. Newsletter of the Michigan Entomological Society. 58 (1 & 2): 11-14. 58,
11-14 (2013).
33.
C. C. Pike et al., Improving the resistance of eastern white pine to white pine blister rust
disease. Forest ecology and management 423, 114-119 (2018).
34.
J. G. Whitehill et al., Interspecific comparison of constitutive ash phloem phenolic
chemistry reveals compounds unique to Manchurian ash, a species resistant to emerald
ash borer. Journal of Chemical Ecology 38, 499-511 (2012).
35.
J. G. A. Whitehill, C. Rigsby, D. Cipollini, D. A. Herms, P. Bonello, Decreased
emergence of emerald ash borer from ash treated with methyl jasmonate is associated
with induction of general defense traits and the toxic phenolic compound verbascoside.
Oecologia 176, 1047-1059 (2014).
24
36.
C. Villari, D. A. Herms, J. G. Whitehill, D. Cipollini, P. Bonello, Progress and gaps in
understanding mechanisms of ash tree resistance to emerald ash borer, a model for wood‐
boring insects that kill angiosperms. New Phytologist 209, 63-79 (2016).
37.
L. J. Kelly et al., Convergent molecular evolution among ash species resistant to the
emerald ash borer. Nature ecology & evolution 4, 1116-1128 (2020).
38.
R. Enderle, J. Stenlid, R. Vasaitis, An overview of ash (Fraxinus spp.) and the ash
dieback disease in Europe. CAB Rev 14, 1-12 (2019).
39.
E. S. A. Sollars et al., Genome sequence and genetic diversity of European ash trees.
Nature 541, 212-216 (2017).
40.
M. Nemesio-Gorriz et al., Canditate metabolites for ash dieback tolerance in Fraxinus
excelsior. Journal of Experimental Botany 71, 6074-6083 (2020).
41.
L. McKinney et al., The ash dieback crisis: genetic variation in resistance can prove a
long‐term solution. Plant Pathology 63, 485-499 (2014).
42.
X. López-Goldar et al., Inducibility of plant secondary metabolites in the stem predicts
genetic variation in resistance against a key insect herbivore in maritime pine. Frontiers
in Plant Science 9, 1651 (2018).
43.
X. López‐Goldar et al., Genetic variation in the constitutive defensive metabolome and
its inducibility are geographically structured and largely determined by demographic
processes in maritime pine. Journal of Ecology 107, 2464-2477 (2019).
44.
M. Volf, J. Hrcek, R. Julkunen‐Tiitto, V. Novotny, To each its own: differential response
of specialist and generalist herbivores to plant defence in willows. Journal of Animal
Ecology 84, 1123-1132 (2015).
45.
K. F. Raffa, E. B. Smalley, Interaction of pre-attack and induced monoterpene
concentrations in host conifer defense against bark beetle-fungal complexes. Oecologia
102, 285-295 (1995).
46.
B. Worley, R. Powers, PCA as a practical indicator of OPLS-DA model reliability.
Current Metabolomics 4, 97-103 (2016).
47.
S. Chakraborty et al., Effects of water availability on emerald ash borer larval
performance and phloem phenolics of Manchurian and black ash. Plant, Cell &
Environment 37, 1009-1021 (2014).
48.
L. W. Sumner et al., Proposed minimum reporting standards for chemical analysis.
Metabolomics 3, 211-221 (2007).
49.
J. D. Sidda et al., Diversity of secoiridoid glycosides in leaves of UK and Danish ash
provide new insight for ash dieback management. Scientific reports 10, 1-12 (2020).
50.
D. Falconer, T. Mackay, Introduction to quantitative genetics 4th edition. Harlow, UK:
Longmans (1996).
51.
M. Huff et al., A high‐quality reference genome for Fraxinus pennsylvanica for ash
species restoration and research. Molecular ecology resources (2021).
52.
J. G. Whitehill, C. Rigsby, D. Cipollini, D. A. Herms, P. Bonello, Decreased emergence
of emerald ash borer from ash treated with methyl jasmonate is associated with induction
of general defense traits and the toxic phenolic compound verbascoside. Oecologia 176,
1047-1059 (2014).
53.
M. C. Schuman, I. T. Baldwin, The layers of plant responses to insect herbivores. Annual
review of entomology 61, 373-394 (2016).
25
54.
K. C. Steiner, L. E. Graboski, K. S. Knight, J. L. Koch, M. E. Mason, Genetic, spatial,
and temporal aspects of decline and mortality in a Fraxinus provenance test following
invasion by the emerald ash borer. Biological Invasions 21, 3439-3450 (2019).
55.
E. J. Rebek, D. A. Herms, D. R. Smitley, Interspecific Variation in Resistance to Emerald
Ash Borer (Coleoptera: Buprestidae) Among North American and Asian Ash (Fraxinus
spp.). Environmental Entomology 37, 242-246 (2008).
56.
D. B. Lybrand, T. M. Anthony, A. D. Jones, R. L. Last, An integrated analytical approach
reveals trichome acylsugar metabolite diversity in the wild tomato Solanum pennellii.
Metabolites 10, 401 (2020).
57.
R. Sadre et al., Metabolite diversity in alkaloid biosynthesis: a multilane (diastereomer)
highway for camptothecin synthesis in Camptotheca acuminata. The Plant Cell 28, 1926-
1944 (2016).
Figure Legends
Figure 1: Egg Bioassay Protocol & Phenotypic Distributions. (a) EAB greenhouse bioassay
protocol. Two year old green ash individuals are artifcially dissected at eight weeks to ascertain
larval fate. (b) Percentage of EAB larvae kill by infested individuals in families Pe-C, Pe-Y and
Pe-Z when sampled 8 weeks post infestation. Family C contained 17 F1 infested progeny from
two susceptible parents. Families Pe-Y and Pe-Z both contained 30 F1 full sib progeny of two
lingering ash parents. Family means of Pe-Y and Pe-Z were significantly different from the
family mean of susceptible family C (p<0.0001)
Figure 2: Data Processing Schematic. Flowthrough of the untargeted metabolomics workflow
beginning with data generation, and highlighting the number of features at each stage of the
analysis.
26
Figure 3: Classification summary. (a) Orthogonal partial least squares projection to latent square
discriminate analysis (OPLS-DA) model performance, averaged across triplicate prediction
models. The graph indicates that percentage that each model classified correctly, incorrection, or
was unable to classify. (b) Principal component analysis plot comparing high and low larval kill
(LK) in family Pe-Y, utilizing 43 features. (c) OPLS-DA model utilizing all test samples in a
comparison of high vs low larval kill using 43 features.
Figure 4: Metabolite Annotations. MS/MS spectra in positive ion mode support annotations of
metabolite structures: (a) product ions of m/z 642.24 ([M+NH4]+) for verbascoside, (product ions
of m/z 271.06 ([M+H]+) for apigenin, (c) product ions of m/z 584.21 ([M+NH4]+) for Excelside
A.
Figure 5: Chemical Families of Annotated Compounds. (a) proportions of the chemical families
in the annotated metabolites. (b) Pairwise comparisons for specific compounds. ‘Number’ is
metabolite number (table 1). LLK v UNI, HLK v UNI, HLK v LLK indicates pairwise comparisons
between larval kill phenotypes or uninfested individuals. Pe-C, Pe-Y, Pe-Z refer to full sibling
families (figure 1). Box with letter indicates the phenotypic category that had significantly
higher concentration of the indicated metabolite (p < 0.05, L in red LLK, H in gray HLK, U in blue
UNI). Annotated metabolites 1-6 are alkaloids, 9-17 are secoiridoid glycosides, 18-20 are
secoiridoids, 24 is salidroside and 25 is verbascoside.
����������
����������
����������
�
��
��
��
���
���������������������
�
�
�
EAB eggs applied to two year old gra�ed trees
EAB egg on filter
taped to stem
Healthy
larva
Host-killed
larva
Data collected 8
weeks later
Egg hatched Y/N
Larva dead/alive
Larval instar: 1-4
Larval weight
EAB Larvae hatches, chews
through filter into tree
* p<0.0001
a
b
Alkaloid
Ammonium
Flavone
Lignan
Lignan
glycoside
Phenolic
glycoside
Secoiridoid
Secoiridoid
glycoside
Sugar
Unknown
1278 Features
32 Features
Collected MS/MS
and annotated features
323 Features
194 Features
Removed
- Contaminants
- Lipids
- Low Abundance
OPLS-DA generation with
three fold cross validation,
averages reported
Selected all features
that were signifcant in
more than 1 family or
had a previous proposed role
32 Features
correct
unkown
incorrect
Pe-C
UNIvLLK
Pe-Y
UNIvLLK
Pe-Y
UNIvHLK
Pe-Y
HLKvLLK
Pe-Z
UNIvLLK
Pe-Z
UNIvHLK
Pe-Z
HLKvLLK
percent assigned
Calculated RSD and tested for
signifcant relationships in
pairwise comparisons
Removed
- Adducts
- Multiple Fragments
Pe-C
UNI v LLK
34 features
Pe-Y
UNI v LLK
49 features
Pe-Y
UNI v HLK
44 features
Pe-Y
HLK v LLK
43 features
Pe-Z
UNI v LLK
25 features
Pe-Z
UNI v HLK
35 features
Pe-Z
HLK v LLK
9 features
Average Model Performance
0 50 100
79 %
83 %
83 %
73 %
87 %
71 %
76 %
HLK
LLK
LLK
HLK
�
�
�
Unknown
Correct
Incorrect
PC2:13%
PC1:56%
-80
60
30
0
-40
-140
0
-70
140
70
-70
70
35
0
-35
-120
0
-60
120
60
R2Y: 70%
Q2: 59%
Pe-C
UNI v LLK
Pe-Y
UNI v LLK
Pe-Z
UNI v LLK
Pe-Y
UNI v HLK
Pe-Y
HLK v LLK
Pe-Z
UNI v HLK
Pe-Z
HLK v LLK
percent assigned
0
20
40
60
80
100
Average Model Performance
�
��
���
�����������������
������������������
������
������
���
��
�
���
��
�
�
�
�
��������
����������������
������������������
�������
100
300
200
400
500
600
[M+H]+
[M+NH4]+
642.24
325.09
163.04
O
OH
HO
O
OH
����������
HO
HO
O
O
O
OH
O
O
OH
OH
OH
O
HO
OH
H
HO
����������������
�������������������
[M+NH4]+
������
������
������
O
O
O
O
O
O
O
HO
HO
OH
OH
HO
OH
OH
O
O
100
300
200
100
300
200
400
500
Alkaloid
Ammonium
Flavone
Lignan
Lignan glycoside
Phenolic
glycoside
Secoiridoid
Secoiridoid
glycoside
Sugar
Unknown
b
a
| 2022 | Profiles of secoiridoids and alkaloids in tissue of susceptible and resistant green ash progeny reveal patterns of induced responses to emerald ash borer in | 10.1101/2022.05.18.492370 | [
"Stanley Robert K.",
"Carey David W.",
"Mason Mary E.",
"Poland Therese M.",
"Koch Jennifer L.",
"Jones A. Daniel",
"Romero-Severson Jeanne"
] | null |
1
Dose-dependent dissociation of pro-cognitive effects of donepezil on
attention and cognitive flexibility in rhesus monkeys
Seyed A. Hassani1, Sofia Lendor2, Adam Neumann1, Kanchan Sinha Roy2, Kianoush Banaie
Boroujeni1, Kari L. Hoffman1, Janusz Pawliszyn2*, Thilo Womelsdorf1,3*
1Department of Psychology, Vanderbilt University, Nashville, TN 37240.
2Department of Chemistry, University of Waterloo, 200 University Avenue West, Waterloo, Ontario N2L
3G1, Canada
3Department of Biomedical Engineering, Vanderbilt University, Nashville, TN 37240.
*Corresponding Authors:
thilo.womelsdorf@vanderbilt.edu; Vanderbilt University, Psychology Department, 301 Wilson
Hall, 111 21st Avenue South, 37240-1103 Nashville TN; Tel. 5743270486
janusz@uwaterloo.ca; University of Waterloo, Department of Chemistry, 278A Chemistry 2
building, 200 University Avenue West, N2L 3G1 Waterloo ON; Tel. 519-888-4641
Running Title: Dose dependent effects on attention and cognitive flexibility
Keywords: Acetylcholine; Prefrontal Cortex; Striatum, Neurochemistry; Solid Phase
Microextraction; Stability-Flexibility Trade off
2
ABSTRACT
BACKGROUND: Donepezil exerts pro-cognitive effects by non-selectively enhancing
acetylcholine (ACh) across multiple brain systems. The brain systems that mediate pro-cognitive
effects of attentional control and cognitive flexibility are the prefrontal cortex and the anterior
striatum which have different pharmacokinetic sensitivities to ACh modulation. We speculated
that these area-specific ACh profiles lead to distinct optimal dose-ranges for donepezil to enhance
the cognitive domains of attention and flexible learning.
METHODS: To test for dose-specific effects of donepezil on different cognitive domains we
devised a multi-task paradigm for nonhuman primates (NHPs) that assessed attention and
cognitive flexibility. NHPs received either vehicle or variable doses of donepezil prior to task
performance. We measured donepezil intracerebral and how strong it prevented the breakdown of
ACh within prefrontal cortex and anterior striatum using solid-phase-microextraction
neurochemistry.
RESULTS: The highest administered donepezil dose improved attention and made subjects more
robust against distractor interference, but it did not improve flexible learning. In contrast, only a
lower dose range of donepezil improved flexible learning and reduced perseveration, but without
distractor-dependent attentional improvement. Neurochemical measurements confirmed a dose-
dependent increase of extracellular donepezil and decreases in choline within the prefrontal cortex
and the striatum.
CONCLUSIONS: The donepezil dose for maximally improving attention functions differed from
the dose range that enhanced cognitive flexibility despite the availability of the drug in the major
brain systems supporting these cognitive functions. Thus, the non-selective acetylcholine esterase
inhibitor donepezil inherently trades improvement in the attention domain for improvement in the
cognitive flexibility domain at a given dose range.
INTRODUCTION
The acetylcholinesterase (AChE) inhibitor donepezil (Aricept) is one of few FDA approved
cognitive enhancers that aims to address a wide range of cognitive deficits in subjects with mild
cognitive impairment or dementia (1–3). Basic research suggests that the cognitive domains that
can be enhanced with AChE inhibitors range from selective attention, working memory, response
inhibition, learning, and long-term memory (4–6). Consistent with these reports, clinical studies
assessing donepezil at one or two doses across larger cohorts of subjects with varying stages of
Alzheimer’s disease have found improvements of compound scores of cognitive testing batteries
(4,7–10). It is, however, not clear whether the standard doses of donepezil used in clinical studies
improve multiple cognitive domains directly, or whether at a particular effective dose its major
route of action is to enhance arousal, which then provides an indirect, overall cognitive advantage
for attention, working memory, learning and memory processes (6,11). Assessing whether
donepezil’s major route of action is the arousal domain, or whether it affects multiple specific
cognitive domains simultaneously at a given dose is important for evaluating its therapeutic
3
efficiency and to identify cognitive domains that should be targeted in drug discovery efforts for
improved future cognitive enhancers.
One potential limitation of donepezil and other AChE inhibitors is that they increase acetylcholine
(ACh) concentrations non-selectively across multiple brain systems. Such a non-selective ACh
increase has shortcomings when brain systems are differently sensitive to ACh action so that the
same donepezil dose that is optimally affecting one brain system might over- or under-stimulate
another brain system. In primates, muscarinic ACh subreceptors relevant for attention and memory
functions (12–15), have enhanced densities in prefrontal cortex (PFC) (16), suggesting that PFC
may be more sensitive to modulation by AChE inhibitors than posterior brain areas. Moreover, a
comparison of transcription factor (CREB) activation of the PFC and the striatum to muscarinic
modulation by Xanomeline has reported a 10-fold higher receptor sensitivity of the striatum (17),
consistent with other studies reporting significantly higher muscarinic binding potential and higher
AChE activity in the striatum than in other cortical regions (18). It is unclear how these differences
affect ACh modulation of attention functions that depend on the PFC (19) and on flexible learning
functions that are dependent on the striatum (20,21). One consequence of the brain area specific
sensitivity to ACh levels could be that a Best Dose for enhancing cognitive functions supported by
the striatum might not sufficiently stimulate the PFC, and that a Best Dose for enhancing PFC
functions might overstimulate the striatum.
To test for these possible implications of brain region-specific ACh action, we devised a drug
testing paradigm for monkeys that assessed the effects of three different doses of donepezil across
different domains of arousal, attention, and cognitive flexibility in a single testing session. We
evaluated the attention domain with a visual search task that varied the number and perceptual
similarity of distracting objects and quantified the domain of cognitive flexibility with a learning
task asking monkeys to flexibly adapt to new feature-reward rules and avoid perseverative
responding. This assessment paradigm goes beyond existing nonhuman primate studies of
donepezil that so far have found enhanced short-term memory using delayed match-to-sample
tasks (4,6,10,15,22–29), enhanced arousal and non-selective speed of processing (15,27), or no
consistent effect (18) (Table S1). With our multi-domain task design we found that donepezil
improves attentional control of interference from distractors at doses that caused an overall slower
responding (i.e. reduced speed of processing) and peripheral side effects. In contrast, a lower dose
of donepezil caused no clear attentional effect but improved cognitive flexibility. These findings
document domain-specific dose-response effects of donepezil for attention and cognitive
flexibility.
METHODS AND MATERIALS
Nonhuman Primate Testing Protocol
Three adult male rhesus macaques (Macaca mulatta; ~8-15 kg, 6-9 years old) were used for this
experiment. They were separately given access to a cage-mounted Kiosk Station that provided a
touchscreen interface inside the animal’s housing unit to perform cognitive tasks (Figure 1A)
(30). Monkeys were cognitively assessed at the same time of day for ~20ml/kg fluid reward. The
behavioral tasks, reward delivery, and the registering of behavioral responses were controlled by
4
the Unified Suite for Experiments (USE) (31). The task protocols, matlab analysis procedures and
the open-sourced USE software are available at
http://accl.psy.vanderbilt.edu/resources/analysis-tools/unifiedsuiteforexperiments/.
Drugs and Procedures
Donepezil-hydrochloride was purchased from Sigma-Aldrich (catalog number D6821; St. Louis,
MO, USA). We tested three doses of donepezil: 0.06, 0.1 and 0.3 mg/kg to operate within the
dosing range of previous studies reporting pro-cognitive effects (Table S1). At this IM range,
plasma concentrations of donepezil have been shown to be roughly the same when dosing with
~10x the concentration via PO (15). All drug doses were administered in a double-blind manner.
Animals received saline as vehicle control, or a dose of donepezil IM injection 30 minutes prior to
starting task performance taking into account its expected 1h half-life (32). Drug side effects were
assessed 15 min following drug administration and after completion of the behavioral performance
with a modified Irwin Scale (33–36) for rating autonomic nervous system functioning (salivation,
etc.) and somato-motor system functioning (posture, unrest, etc.). Monkeys’ behavioral status was
video-monitored throughout task performance (Figure 1A).
Behavioral Paradigms
Monkeys performed a visual search (VS) task to measure attentional performance metrics and a
feature-reward learning (FL) task to measure cognitive flexibility metrics in each experimental
session. Each performance day was made up of an initial set of 100 trials of VS, a set of 21 learning
blocks with 35-60 trials each of the FL task, and a final set of 100 trials of the VS task (Figure
1Aii). Details of both tasks are provided in the Supplement. In brief, the VS task required monkeys
to find and touch a target object among 3, 6, 9, or 12 distracting objects in order to receive fluid
reward (Figure 1B). The target was the object that was shown in 10 initial trials without distractors.
Targets and distractors were multidimensional, 3D rendered Quaddle objects (31) that shared few
or many features of different features dimensions (colors, shapes, arms, body patterns), which
rendered search easier when there was no or few similarities among features of targets and
distractors, or more difficult if the target-distractor (T-D) similarity was high (Figure 2A). The
FL-task required monkeys to learn through trial-and-error which object feature is rewarded in a
given block of ~35-60 trials (Figure 1C). In each trial of the block three objects were shown that
varied either in features of one feature dimension (e.g. having different colors or different body
shapes), or that varied in features of two feature dimensions (e.g. having different colors and
different body shapes). Choosing the object with the correct feature was rewarded with a
probability of 0.8. Blocks where only 1 feature dimension varied (1D blocks) were easier as there
was lower attentional load than in blocks with 2 varying feature dimensions (2D blocks).
Neurochemical Confirmation of Drug Effect
To evaluate the levels of donepezil in brain structures that are necessary for successful attention
and learning performance, we measured choline and donepezil concentrations in the prefrontal
cortex and the anterior striatum (caudate nucleus) 15 min after administering a low and high dose
of donepezil (0.06 and 0.3 mg/kg, IM) in a separate experiment. Measures of donepezil were made
at the time when we observed dose-limiting side effects at the 0.3 mg/kg dose and the two tested
5
doses were accompanied by pro-cognitive effects in our task (see results). We used microprobes
that sampled the local neurochemical milieu with the principles of solid phase micro-extraction
(SPME) (for details see Supplement) (37). SPME probes sampled the level of donepezil and the
ACh metabolite (choline) via diffusion at a consistent rate until an equilibrium was reached with
the extracellular concentrations. The neurochemical concentrations were quantified with liquid
chromatography and mass spectrometry as described in detail in (37). The detailed procedures
used here are described in (38).
Statistical Analysis
Data were analyzed with standard nonparametric and parametric tests as described in the
Supplement.
Results
Each monkey was assessed in 38 sessions total including 17 vehicle days and 7 days with each
dose (0.06, 0.1 and 0.3 mg/kg). Drug side effects were noted following IM injections of the 0.3
mg/kg dose in the first 30 min post injection as changes in posture, sedation, vasoconstriction and
paleness of skin, but no adverse effects persisted beyond 30 min. (Table S2). First, we confirmed
that monkeys performed the visual search (VS) task at high 84.4% (± 0.54) accuracy (monkeys Ig:
85.2% ±0.81; Wo: 88.3% ±0.94; Si: 79.8% ±0.97) and showed the expected set-size effect evident
in decreased accuracy and slower reaction times with increasing numbers of distractors (Figure
1D, Figure S1 and S2, Supplemental). When targets were more similar to distractors (high T-D
similarity) VS performance decreased from 92.9% (±0.4) to 85.5% (±0.3) and 81.6% (±1.0) for
low, medium and high T-D similarity, respectively (H(2) = 169.48, p < .001) (Figure 2B). In the
feature learning (FL) task, the monkeys reached learning criterion faster in the easier 1D (low
distractor load) condition (avg. trials to ≥80% criterion: 12.5 ± 0.2 SE), than in the 2D (high
distractor load) condition (avg. trials to ≥80% criterion: 15.6±0.2) (Figure 3A, Supplemental).
Dose-dependent improvement of visual search accuracy and slowing of choice reaction times
Donepezil significantly improved accuracy of the visual search task (F(1,1722) = 18.95, p <
.001)(Figure 1D), but on average slowed search reaction times (F(1,1722) = 4.83, p =
.028)(Figure S1B). The slower choice reaction times were evident already to the single target
object in the 10 target familiarization trials (Figure S1A). These main behavioral drug effects were
evident prominently in the first visual search block (Figure 1D, Figure S1A). We therefore
focused our further analysis on the first search block.
The improved accuracy of visual search was dose-dependent. The 0.1 mg/kg dose enhanced
performance by 2.5% ±1.0, 4.4% ±1.3, 6.1% ±1.4 and 6.3% ±1.6 (mean ±SD) for 3/6/9/12
distractor trials, respectively (X2(1, N1 = 16700, N2 = 2100) = 35.5, p < .001). The 0.3 mg/kg dose
enhanced performance by 2.7% ±1.0, 6.3% ±1.2, 8.5% ±1.3 and 11.0% ±1.4 (mean ±SD) for
3/6/9/12 distractor trials respectively (X2(1, N1 = 16700, N2 = 1900) = 75.9, p < .001) (Figure 1E).
Thus, we found larger improvement the more distractors interfered with the target search. We
confirmed this by fitting a regression line across performance at different number of distractors,
6
which revealed overall significantly shallower slopes with donepezil (slopes: -0.013 ±0.001, -
0.009 ±0.002, -0.015 ±0.003 and -0.005 ±0.002 for vehicle, 0.06, 0.1, and 0.3 mg/kg of donepezil
respectively (H(3) = 11.46, p = .01)). Pairwise comparison showed that the 0.3 mg/kg drug dose
and the vehicle condition showed significantly different slopes (Tukey’s, p = .013)(Figure 1F).
In contrast to improving visual search accuracy, donepezil slowed down reaction times across all
distractor conditions at the 0.3mg/kg dose relative to vehicle by on average 100 ms ±40, 238 ms
±79, 208 ms ±99, 264 ms ±102 (mean ± SD) for 3/6/9/12 distractors respectively (p = .023,
Bonferroni correction)(Figure S1C). The slope of the regression over different number of
distractors did not differ between 0.3 mg/kg dose and vehicle which shows the reaction time effect
is a non-selective effect that is independent of distractors (regression slope on RTs: 0.061 ±0.002,
0.065 ±0.007, 0.067 ±0.007 and 0.076 ±0.009 (H(3) = 3.37, n.s.) for vehicle, 0.06, 0.1, 0.3 mg/kg
of donepezil respectively (Figure S1D).
Across sessions visual search accuracy was correlated with reaction times only for the vehicle
(Pearson, r: -0.30, p < .001) and 0.1 mg/kg donepezil dose condition (Pearson, r: -0.46, p = .034),
but not for the 0.06 and 0.3 mg/kg dose conditions in which monkeys showed improved accuracy,
which suggests the accuracy improvement is independent from a slowing of reaction speed (Figure
S2A-B).
We next tested whether improved control of interference from increasing number of distractor
objects was likewise evident when increasing the similarity of distractor and target features
(Figure 2A). First, we confirmed that higher target-distractor similarity overall reduced
performance (F(2,672) = 16.17, p < .001, Supplemental). Donepezil significantly counteracted
this similarity effect and improved performance at the 0.06 and 0.3 mg/kg doses (F(3,672) = 7.75,
p < .001, Tukey’s, p = .034 and p < .001 respectively). This finding shows that the beneficial effect
of donepezil significantly increased when there was higher demand to control perceptual
interference from distracting objects (Figure 2B). This was also evident as a statistical trend of a
shallower regression slope at 0.06 and 0.3 mg/kg doses of donepezil, which indicates less
interference from distracting features when they were similar to the target (Figure 2C) (H(3) =
2.79, n.s.; slope changes relative to vehicle for 0.06, 0.1 and 0.3 mg/kg doses were: +0.0357
±0.0236, -0.0289 ±0.0334, -0.0656 ±0.0197). The improved search performance with donepezil
for visual search with higher target-distractor similarity and with a higher number of distractors
was evident in significant main effects, but there was no interaction, suggesting they improved
performance independently of each other (F(2, 16688) = 55.24, p < .001; F(3,16688) = 50.25, p <
.001; F(6,16688) = 1.16, n.s. respectively)(Figure 2D). This independence was also suggested by
the absence of a correlation of the target-distractor similarity effect and the number-of-distractor
effect (Pearson, n.s.) (Figure S3).
Dose-dependent improvement of flexible learning performance
Donepezil also improved feature learning performance, but only at the 0.06 mg/kg dose (Figure
3B) and most pronounced for the first third of the behavioral session (F(3,602) = 3.3, p = .020;
Figure 3C). We therefore focused further analysis on the first third of the learning blocks, which
revealed that the learning improvement at the 0.06 mg/kg dose was significant for the low
distractor load condition (significant interaction effect of drug condition and distractor load
7
(Condition x Distractor Load F(3, 1052) = 3.59, p = .013); and for vehicle, 0.06, 0.1 and 0.3 mg/kg
donepezil doses the trials to criterion were 11.3 ±0.4, 7.7 ±0.9, 12.3 ±1.3 and 11.0 ±1.2 trials long
with the 0.06 mg/kg dose and vehicle being significantly different (p = .020, Bonferroni
correction)(Figure 3D). There was no change in learning speed with other doses at low or high
distractor load.
Beyond learning speed, we found overall slower choice reaction times at the 0.3 mg/kg donepezil
dose (Figure 3E) (main effect of drug condition: (F(3,1052) = 12.29, p < .001). While reaction
times were overall slower at high distractor load (F(1,1052) = 7.18, p = .008) there was no
interaction with drug dose (F(3,1052) = 0.26, n.s.). After visually inspecting the results we
separately tested the 0.3 mg/kg dose of donepezil and found it led to significantly slower choice
reaction time than vehicle (Tukey’s, p < .001)(Figure 3E). The changes in choice reaction times
did not correlate with changes in learning performance (number of trials to criterion) at any drug
condition, indicating they were independently modulated (Pearson, all n.s.)(Figure S2D).
We predicted that the faster learning at the 0.06 mg/kg donepezil dose could be due to a more
efficient exploration of objects during learning, which would be reflected in reduced perseverative
choices of unrewarded objects. Overall, perseverative errors (defined as consecutive unrewarded
choices to objects with the same feature dimension) made up 20% of all errors. As expected, we
found significantly shorter sequences of perseveration of choosing objects within distractor feature
dimensions at the 0.06 mg/kg dose of donepezil (Figure 3F). For 0.06, 0.1 and 0.3 mg/kg doses
the average length of perseverations in the distractor dimension was: 2.1 ±0.1, 1.8 ±0, 1.9 ±0.1 and
1.9 ±0.1 trials with the difference between vehicle and the 0.06 dose being significant (p = .021).
Perseverative choices in the target feature dimension were not different between conditions (for
0.06, 0.1 and 0.3 mg/kg donepezil doses the avg. perseveration length in the target dimension was:
1.7 ±0, 1.7 ±0, 1.6 ±0, and 1.7 ±0 trials (n.s.).
Dissociation of attention and learning improvements, but slowing is correlated
The effects of donepezil on feature learning and visual search might be related, but we found that
learning speed and search accuracy was not correlated at those doses at which the drug improved
learning and search (0.06 mg/kg dose) or improved only visual search (0.3 mg/kg dose) (Pearson,
all n.s.). A significant correlation was found only for the 0.1 mg/kg dose (Pearson, r: -0.54; p =
.012) (Figure 4A). Learning at low or high distractor load and the set size (slope) effects in the
visual search task was uncorrelated (Pearson, all n.s.). However, at the 0.3 mg/kg donepezil dose
we found that the target-distractor similarity effect (i.e. the search slope change) in the visual
search task was positively correlated with the difference of the learning speed at high versus low
distractor load in the learning task (Pearson, r: 0.60; p = .008). This effect signifies that better
attentional search of a target among similar distractors is associated with poorer flexible learning
of new targets when there are multiple object features to search through (high distractor load).
In contrast to accuracy, choice reaction times in the learning task and visual search were
significantly correlated for the 0.1 mg/kg donepezil dose (Pearson, r: 0.52; p = .016), the 0.3 mg/kg
dose (Pearson, r: 0.66; p = .002), and the vehicle control condition (Pearson, r: 0.60; p <
.001)(Figure 4B).
8
Determination of extracellular donepezil and choline levels in the prefrontal cortex and
anterior striatum
Visual search and flexible learning are realized by partly independent brain systems, including the
PFC and anterior striatum (39). To determine whether extracellular levels of donepezil were
increased to a similar magnitude in the PFC and anterior striatum, we measured its concentration
after administering doses of either 0.06 and 0.3 mg/kg donepezil IM in the PFC, assumed to be
necessary for efficient interference control during visual search (19), and in the head of the caudate
nucleus which is necessary for flexible learning of object values (20,21). We used a recently
developed microprobe that samples chemicals in neural tissue based on the principles of solid-
phase microextraction (SPME) (37,38). We found that donepezil was available in both brain areas
and its extracellular concentration more than doubled after injecting 0.3 mg/kg than 0.06 mg/kg in
both areas (F(1,16) = 9.69, p = .007), with no significant difference between PFC and caudate
(F(1,16) = 1.44, n.s.)(Figure 5A). Donepezil should cause a depletion of the ACh metabolite
choline (40). Using HPLC/MS analysis of the SPME samples we found in the PFC that 0.06 and
0.3 mg/kg donepezil reduced choline concentrations by 74.2% ±14.9 (p = .005) and 85.7% ± 26.9
(p = .007) of their baseline concentrations, and in the caudate, it reduced choline by 68.4% ±13.8
(p = .022) and 81.0% ±12.9 (p = .009) of respective baseline concentrations (Figure 5B). The
11.5% and 12.6% stronger reduction choline at the 0.3 versus 0.06 mg/kg dose in PFC or caudate
was not significant (n.s.).
To obtain an independent physiological marker of dose-dependent effects we quantified during
actual task performance how donepezil changed the heart rate (HR) before versus after drug
administration (Supplemental). HR showed a transient peak ~20 min after donepezil injection
relative to baseline, which was significant for the 0.3 mg/kg dose (pre-injection 102.3 ±7.1 to post-
injection 121.6 ±2.6; p = .021), but not for the 0.06 mg/kg dose (pre-injection: 90.3 ±4.2 to post-
injection: 94.8 ±5.4; n.s.). The 0.3 mg/kg dose caused a significantly higher HR peak than the 0.06
mg/kg dose (p = .006) (Figure 5C).
Discussion
Here, we dissociated donepezil’s improvement of attentional control of interference during visual
search performance from improvements of cognitive flexibility during feature reward learning. At
the highest dose tested donepezil reduced interference during visual search particularly when there
were many distractors and high similarity of distractors to the target, while concomitantly slowing
down overall reaction times and inducing temporary peripheral side effects. In contrast, at the
lowest dose donepezil did not affect target detection times during visual search, but improved
adapting to new feature-rules and reduced perseverative responding. These findings document a
dose-dependent dissociation of the best dose of donepezil for improving attention and for
improving cognitive flexibility.
Different donepezil dose-ranges for improving interference control and flexible learning
Using a behavioral assessment paradigm with two tasks allowed us to discern differences of the
donepezil dose that maximally improved interference control (in the visual search task) versus the
9
dose that maximally improved flexible learning (in the reward learning task). In both tasks,
donepezil modulated performance early within the session (first of two VS blocks and first third
of FL blocks) consistent with its short half-life and rapid time to peak concentration with IM
delivery (15,32) and therefore our results focused on this time window. At the 0.06 mg/kg dose
donepezil facilitated flexible learning of a new feature reward rule and reduced the length of
perseverative errors (Figure 3C,F). These behavioral effects can be interpreted as improvements
of cognitive flexibility of the monkeys in adjusting to changing task demands. At the same 0.06
mg/kg dose visual search response times were unaffected (Figure S1) and visual search accuracy
was overall improved but independent of the number of distractors, i.e. independent of the degree
of interference (Figure 1E,F). In contrast, at the higher donepezil doses flexible learning behavior
was indistinguishable from the no-drug vehicle control condition showing that improving
flexibility required donepezil at a lower dose.
This conclusion is opposite to the drug effects on visual search performance, which was maximally
improved at the 0.3 mg/kg dose. At this dose, subjects not only showed improved resistance to
interference when there were more distracting objects (Figure 1E,F), but also improved resistance
to distracting objects that were visually similar to the searched-for target (Figure 2B-D). These
findings document that donepezil enhances the robustness to distraction (41,42), which critically
extends insights from existing primate studies with donepezil that mostly used simpler tasks to
infer pro-cognitive effects on working memory or arousal (see Table S1). The process of
attentional control of interference goes beyond a short-term memory effect measured with delayed
match to sample tasks. In the visual search paradigm we used, short-term memory of the target
object is already necessary for performing the easier trials with 3 or 6 distractors, while an attention
specific effect can be inferred when there is greater improvement in performance with increased
attentional demands in trials with 9 or 12 distractors. Thus, our study provides strong evidence that
donepezil can cause specific attentional improvement at relatively higher doses. This finding
supports a prominent neuro-genetic model of cholinergic modulation of attention (43) that has
received recently functional support in studies reporting enhanced distractor suppression in
nonhuman primates with nicotine receptor specific ACh modulation (44–46), and improved
suppression of perceptually distracting flankers in human subjects tested with a single dose (47).
Non-selective slowing of response times and dose-limiting side effects
We found that 0.3 mg/kg donepezil overall slowed response times of the monkeys during visual
search independent of distractor number or target-distractor similarity (Figure S1A,C), and during
feature-reward learning independent of distractor load (Figure 3E). The slowing of reaction times
was independent of overall accuracy levels (Figure 4A), which shows it did not reflect trading off
speed for accuracy. The observed slowing occurred at a dose that improved attention and was
unexpected, because prior studies using the delayed match to sample task did not report changes
in reaction times in monkeys (23,25), or reported normalized reaction times in studies using
donepezil to recover from scopolamine induced deficits (15,26) (Table S1). Our findings therefore
indicate that 0.3 mg/kg of donepezil already induced cholinergic side effects while still improving
cognitive processes. This interpretation is supported by our observation of arousal deficits at the
0.3 mg/kg dose that became apparent in vasoconstriction, changes in posture, visible sedation and
paleness (Table S2). These side effects were strongest within 30 min. after administration of the
drug. Although these side effects did not prevent animals from starting and completing the tasks,
10
they limited the dose range we could test. Such dose-limiting side effects are a well known
limitation of donepezil and other AChE inhibitors where therapeutically effective doses cause in a
subset of patients gastrointestinal issues such as nausea, diarrhea, and arousal deficits (10,48–50).
Our finding adds to this literature that arousal deficits are occurring at a dose range that causes
apparent improvements in attentional control of interference while lower doses that were void of
side effects failed to improve attention. These observations might have clinical implications as
they predict that lower doses of donepezil might not cause improved attention, but primarily
improve cognitive flexibility.
Our finding of dose-limiting side effects and reductions in arousal or speed-of-processing
emphasizes the importance of developing drugs that avoid nonselective overstimulation of intrinsic
cholinergic neurotransmission. Strong candidate compounds include positive allosteric modulators
(PAMs) for nicotinic subreceptors (51) and for M1 and M4 muscarinic receptor (13,14,52,53). The
subtype-specific muscarinic PAMs do not target the orthosteric binding site of acetylcholine (ACh)
that is highly conserved across all five mAChR subtypes, but rather, they act at more
topographically distinct allosteric sites. In addition, PAMs have no intrinsic activity at their
respective muscarinic receptor subtype, but act to boost normal cholinergic signaling thereby
conserving the spatial and temporal endogenous ACh signaling and avoiding overstimulation of
peripheral ACh receptors and subsequent adverse side effects (54–56). Our study thus provides an
important benchmark for the development of new drugs that aim to enhance multiple cognitive
domains while minimizing side effects.
Quantifying extracellular levels of donepezil and choline in prefrontal cortex and striatum
We confirmed the presence of extracellular donepezil in the PFC and the anterior striatum at the
doses tested (Figure 5A) and that it prevented ACh metabolism as evident in 68-86% reduced
choline levels (Figure 5B). To our knowledge this is the first quantification of donepezil’s action
on the breakdown of ACh in two major brain regions in the primate. The observed reduction of
choline is higher than reductions of AChE activity (of ~25- 70%) reported with positron emission
tomography or in brain homogenate (57,58). Previous studies suggest that evaluating blood plasma
levels or cerebrospinal concentrations may not predict how effectively drugs acting on AChE
influence behavioral outcomes (59). One likely reason is that intracerebral concentrations can be
multifold higher than extracerebral concentration levels (57,60) and do not reflect the actual
bioactive concentration available in target neural circuits. By confirming that donepezil prevented
ACh breakdown in the PFC and striatum, we thus established a direct link of behavioral outcomes
and local drug action in two brain structures whose activity causally contributes to attention and
learning. The lateral PFC is causally necessary for attentional control of distractor interference
(61) with ACh depletion of the PFC impairing attention, but not learning (62). In contrast, flexible
learning and overcoming perseverative response tendencies closely depend on the anterior striatum
(63). Both structures closely interact during attention demanding learning processes (64), but can
be dissociated neurochemically (39). Our findings showed that donepezil has a similar effect on
ACh breakdown in both areas, which suggests that differences in behavioral outcomes at a given
donepezil dose are likely due to differences in the sensitivity of these areas to ACh action. Indeed,
prior studies suggest that the striatum has a particularly high muscarinic binding potential (18) and
respond (tenfold) stronger to muscarinic ACh receptor activation compared with the PFC (17). We
11
speculate that these brain area specific neuromodulatory profiles underly the observed dose
specific improvements of cognitive flexibility and attentional control of interference.
The neurochemical measurements of donepezil in PFC and striatum were achieved with a recently
developed microprobe that samples neurochemicals through principles of solid phase
microextraction (SPME) (37,38,65–67), and so far was used for testing the consequence of drugs
only in rodents (66,68,69). We believe that leveraging this technique in primate drug studies will
be important for clarifying whether systemically administered drugs reach the desired target brain
systems in which they are supposed to exert their pro-cognitive effects.
In our study, confirming donepezil’s action in PFC and striatum critically constrains the
interpretation of the behavioral results, suggesting that different behavioral outcome profiles are
not due an uneven drug availability. Rather, the different ‘Best Doses’ for visual search and flexible
learning performance will be best explained by brain area specific pharmacokinetic profiles of
receptor densities, drug clearance profiles, or auto-receptor mechanisms that intrinsically
downregulate local drug actions (70–72).
In summary, our results provide rare quantitative evidence that a prominent ACh enhancing drug
exerts domain specific cognitive improvements of attentional control and cognitive flexibility at a
distinct dose range. A major implication of this finding is that for understanding the strength and
limitations of pro-cognitive drug compounds it will be essential to test their dose-response efficacy
at multiple cognitive domains.
Financial Disclosures
The authors declare no competing financial interests.
Acknowledgements
We thank Dr. Carrie K Jones and Jason Russel for helpful feedback throughout the study and about
the manuscript. Research reported in this publication was supported by the National Institute of
Mental Health of the National Institutes of Health under grants MH123687 (T.W.) The content is
solely the responsibility of the authors and does not necessarily represent the official views of the
National Institutes of Health.
All authors report no biomedical financial interests or potential conflicts of interest.
Appendix A. Supplementary Information
Figures
12
Figure 1. Task design, meta-structure and visual search performance as a function of distractor number. (A) i. Picture of one
of the subjects working in the custom-built kiosk, interacting with the touchscreen and receiving fluid reward. ii. The meta
structure of the Multi-task. Each experimental session consists of 3 super-blocks of VS, FL and VS respectively. Each VS
block is preceded by 10 familiarization trials which is identical to a VS trial but without any distractors. Each VS block contains
trials with 3/6/9/12 distractors randomly selected and counterbalanced over the block. In contrast, each FL block will contain
0/1 irrelevant feature dimensions in addition to the relevant feature dimension (the dimension with the rewarded feature value)
counterbalanced over the session. (B) i. From the grand pool of quaddles which includes four feature dimensions and a variable
number of feature values (9 shapes, 9 patterns, 8 colors, and 11 arms), three feature values from three feature dimensions are
chosen. This 3x3 pool is then counterbalanced for dimension presentation and feature reward association and is utilized for 2
weeks of data collection where all presented quaddles are selected from this 3x3 pool. ii. Example trials. Two example VS
trials (top) within the same block with 3 distractors (left) and 9 distractors (right). Each VS block will contain one of 5
backgrounds, with the VS blocks in the same day never having the same background. All distractors and target objects in VS
blocks are three dimensional objects and distractors may be duplicated in each trial. Quaddles are spatially randomly presented
at the intersections of a 5x4 virtual grid pattern on screen. The red box highlights the rewarded target object, which is invariable
within the VS block, in these examples. Two example FL trials (bottom) within the same block containing 2D quaddles (1
distracting dimension plus the relevant dimension). The rewarded feature value in this block is the checkered pattern
independent of what color feature value it is paired with. Quaddles may be presented in 8 possible locations in a circle each
being 17 degrees of visual radius away from the center of the screen. The red box signifies the rewarded target object, which
is a variable combination of the rewarded feature value (the checkered pattern in this example) with a random feature value of
the distractor dimension (color in this example). (C) The trial structure for both the FL (top) and VS (bottom) blocks of the
task are very similar. A trial is initiated by a 0.3-0.5s touch and hold of a blue square (3° visual radius wide) after which the
blue square disappears for 0.3-0.5s before task objects, which are 2.5° visual radius wide, are presented on screen. Once the
task objects are on screen, the subject is given 5s to visually explore and select an object via a 0.2s touch and hold. A failure
to make a choice within the allotted 5s results in an aborted trial and did not count towards the trial count. Brief auditory
feedback and visual feedback (a halo around the selected object) are provided upon object selection, with any earned fluid
reward being provided 0.2s following object selection and lasting 0.5s along with the visual feedback. Non-rewarded trials had
a different auditory tone and a light blue halo around the selected, non-rewarded object. Rewarded objects had a higher pitch
auditory tone, a light yellow halo around the selected rewarded object and had an accompanying fluid (water) reward. (D)
Average VS performance by distractor number for vehicle and all donepezil doses combined, both separated by the first vs
second VS block. VS performance was significantly different for block number (F(1,1722) = 22.19, p < .001) as well as
condition (F(1,1722) = 19.0, p < .001). The inlet shows individual monkey average VS performance linear fits. (E) Average
VS performance by distractor number between vehicle and 0.06, 0.1, and 0.3 mg/kg donepezil doses for the first VS block (p
< .001). Both the 0.06 and 0.3 mg/kg doses were significantly different from vehicle (p < .001). Error bars here reflect standard
deviation in this panel. (F) The set size effect of VS performance by distractor number for each condition. The 0.3 mg/kg dose
set size effect was significantly shallower from the vehicle set size effect (H(3) = 11,46, p = .010; Tukey’s, p = .013).
13
Figure 2. Visual search task performance and change in difficulty through increasing distractor numbers
and target-distractor similarity. (A) A visual description of the target-distractor similarity measure in
the VS task. For an example target in the red square, 3 example distractors are presented with 0, 1 and
2 features in common respectively from left to right. The cartoon plot below shows the impact of the
average target-distractor similarity of an individual trial on performance. (B) Similar to Figure 1D, but
here we plot average VS performance by T-D similarity. There was a significant effect of T-D similarity
on performance (F(2,627) = 16.17, p < .001) as well as condition with both the 0.06 and 0.3 mg/kg
donepezil doses being significantly different from vehicle (F(3,267) = 7.75, p < .001; Tukey’s p = .034
and p < .001 respectively). (C) The change in the slope of VS performance with 0.06, 0.1 and 0.3 mg/kg
donepezil relative to vehicle. The change in slope by distractor number is plotted on the left y axis (same
data as Figure 1F) (H(3) = 11.46, p = .010) while the change in slope by T-D similarity is plotted on
the right y-axis (H(3) = 2.8, n.s.). (D) A visualization of the combined effect of distractor number and
T-D similarity on performance. From left to right, each cluster of lines represents increasing distractor
numbers while data within each line represents low, medium and high T-D similarity from left to right
respectivel. Both distractor number (F(3,16688) = 50.25, p < .001) and T-D similarity (F(2,16688) =
55.24, p < .001) impact VS performance with no significant interaction (F(6,16688) = 1.16, n.s.).
14
Figure 3. Feature learning task learning curves and performance. (A) Average learning curves of each
monkey and all monkeys combined for both low and high distractor load conditions. In all instances,
monkeys learned faster and with higher plateau performance in low distractor load blocks relative to
high distractor load blocks. (B) All monkey average learning curves for vehicle, 0.06, 0.1 and 0.3 mg/kg
donepezil doses for both low and high distractor load conditions. (C) Temporal progression of learning
speed (LP) for vehicle, 0.06, 0.1 and 0.3 mg/kg donepezil doses for the low distractor load condition
only. At the 0.06 dose, donepezil allows for faster learning in the low attentional load blocks (F(3,602)
= 3.3, p = .020). Similar to the VS task, donepezil’s enhancement is only visible early on and relatively
close to its i.m. administration. (D) Average learning speed of vehicle and donepezil doses for low and
high distractor load blocks across sessions reveals an interaction between drug condition and distractor
load (F(3,1052) = 3.59, p = .013). (E) The same as D but for choice RTs instead of learning speed. The
0.3 mg/kg donepezil dose slows choice reaction times in both low and high distractor load blocks
(Condition F(3,1052) = 12.3, p < .001; Tukey’s, p < .001). (F) Change in the length of perseverative
errors from vehicle, where feature values in the distracting dimension were the target of the
perseverations. Error bars reflect SEMean for inter-monkey variability. Donepezil at the 0.06 mg/kg
dose significantly reduces perseveration length in the distracting dimension (p = .021); other donepezil
doses trends towards this as well.
15
Figure 4. The relationship between the visual search task and the feature learning task. (A).
Correlation coefficients between FL learning speed (LP) and VS performance for vehicle, 0.06,
0.1 and 0.3 mg/kg donepezil doses. Only the 0.1 mg/kg donepezil dose had a significant
correlation between FL and VS task performance (Pearson, r: -0.54; p = 0.012). No doses
showed a significant change in correlation from vehicle. (B) Same as figure A but for FL choice
RTs and VS search RTs. Although vehicle, 0.1 and 0.3 mg/kg donepezil doses had a significant
correlation between choice and search RTs, we found no significant change in correlation
relative to vehicle.
16
Figure 5. In-vivo extracellular measurements of choline, donepezil as well as donepezil’s effect
on heart rate. (A) Quantified concentration of extracellular unbound donepezil with 0.06 and
0.3 mg/kg donepezil administration in the PFC and CD. We are able to reliably detect higher
donepezil concentrations with 0.3 mg/kg dosing relative to 0.06 mg/kg dosing (Condition
F(1,16) = 9.69, p = .007) with SPME. We also see a trend towards higher detectable donepezil
in the caudate relative to the PFC at the 0.3 mg/kg dose tested, however, we found neither
significant group or interaction effects. (B) We used choline concentrations as a metric for
donepezil bio-activity as it de-activates AChE and prevents acetylcholine’s degradation into
choline. We extracted average session-wise change in choline from baseline with 0.06 and 0.3
mg/kg donepezil doses within the PFC and CD.
Although we find significant decreases in choline by up to >80% of baseline concentrations,
we found no significant effect of dosing in either the PFC or CD. (C) The heart rate of our
fourth monkey was monitored during the neurochemical experiments. This revealed a sharp
and transient increase in HR post administration of donepezil at 0.3 mg/kg dose
(Supplemental) which lead to a higher average bpm. We found that we can significantly
distinguish 0.06 and 0.3 donepezil administration via HR (p = .006).
17
References
1. Mori S (2002): Responses to donepezil in Alzheimer’s disease and Parkinson’s disease.
Annals of the New York Academy of Sciences 977: 493–500.
2. Sugimoto H (2001): Donepezil hydrochloride: A treatment drug for alzheimer s disease.
Chemical Records 1: 63–73.
3. Sharma P, Tripathi MK, Shrivastava SK (2020): Cholinesterase as a Target for Drug
Development in Alzheimer’s Disease. Clinical Pharmacokinetics, vol. 2089. Retrieved
from http://link.springer.com/10.1007/978-1-0716-0163-1
4. Yoo JH, Valdovinos MG, Williams DC (2007): Relevance of donepezil in enhancing learning
and memory in special populations: A review of the literature. Journal of Autism and
Developmental Disorders 37: 1883–1901.
5. Cools R, Arnsten AFT (2021): Neuromodulation of prefrontal cortex cognitive function in
primates : the powerful roles of monoamines and acetylcholine. Neuropsychopharmacology
1–20.
6. Floresco SB, Jentsch JD (2011): Pharmacological enhancement of memory and executive
functioning in laboratory animals. Neuropsychopharmacology 36: 227–250.
7. Rogers SL, Doody RS, Mohs RC, Friedhoff LT (1998): Donepezil Improves Cognition and
Global Function in Alzheimer Disease. Clinical Chemistry 158: 1021–1031.
8. Pratt RD, Perdomo CA (2002): Donepezil-treated patients with probable vascular dementia
demonstrate cognitive benefits. Annals of the New York Academy of Sciences, vol. 977 977:
513–522.
9. Foldi NS, White REC, Schaefer LA (2005): Detecting effects of donepezil on visual selective
attention using signal detection parameters in Alzheimer’s disease. International Journal of
Geriatric Psychiatry 20: 485–488.
10. Li DD, Zhang YH, Zhang W, Zhao P (2019): Meta-analysis of randomized controlled trials
on the efficacy and safety of donepezil, galantamine, rivastigmine, and memantine for the
treatment of Alzheimer’s disease. Frontiers in Neuroscience 13.
https://doi.org/10.3389/fnins.2019.00472
11. Everitt BJ, Robbins TW (1997): Central cholinergic systems and cognition. Annual Review of
Psychology 48: 649–684.
12. Langmead CJ, Watson J, Reavill C (2008): Muscarinic acetylcholine receptors as CNS drug
targets. Pharmacology and Therapeutics 117: 232–243.
13. Dean B, Scarr E (2020): Muscarinic M1 and M4 receptors: Hypothesis driven drug
development for schizophrenia. Psychiatry Research 288: 112989.
14. Foster DJ, Bryant ZK, Conn PJ (2021): Targeting muscarinic receptors to treat
schizophrenia. Behavioural Brain Research 405: 113201.
15. Lange HS, Cannon CE, Drott JT, Kuduk SD, Uslaner JM (2015): The M1 muscarinic
positive allosteric modulator PQCA improves performance on translatable tests of memory
and attention in rhesus monkeys. Journal of Pharmacology and Experimental Therapeutics
355: 442–450.
16. Froudist-Walsh S, Xu T, Niu M, Rapan L, Margulies DS, Zilles K, et al. (2021): Gradients of
receptor expression in the macaque cortex. bioRxiv 2021.02.22.432173.
17. Thorn CA, Moon J, Bourbonais CA, Harms J, Edgerton JR, Stark E, et al. (2019): Striatal,
Hippocampal, and Cortical Networks Are Differentially Responsive to the M4- and M1-
18
Muscarinic Acetylcholine Receptor Mediated Effects of Xanomeline. ACS Chemical
Neuroscience 10: 1753–1764.
18. Tsukada H, Nishiyama S, Fukumoto D, Ohba H, Sato K, Kakiuchi T (2004): Effects of Acute
Acetylcholinesterase Inhibition on the Cerebral Cholinergic Neuronal System and Cognitive
Function: Functional Imaging of the Conscious Monkey Brain Using Animal PET in
Combination with Microdialysis. Synapse 52: 1–10.
19. Cohen JY, Pouget P, Woodman GF, Subraveti CR, Schall JD, Rossi AF (2007): Difficulty of
visual search modulates neuronal interactions and response variability in the frontal eye
field. Journal of Neurophysiology 98: 2580–2587.
20. Williams ZM, Eskandar EN (2006): Selective enhancement of associative learning by
microstimulation of the anterior caudate. Nature Neuroscience 9: 562–568.
21. Amita H, Hikosaka O (2019): Indirect pathway from caudate tail mediates rejection of bad
objects in periphery. Science Advances 5: 1–8.
22. Vardigan JD, Cannon CE, Puri V, Dancho M, Koser A, Wittmann M, et al. (2015): Improved
cognition without adverse effects: Novel M1 muscarinic potentiator compares favorably to
donepezil and xanomeline in rhesus monkey. Psychopharmacology 232: 1859–1866.
23. Callahan PM, Hutchings EJ, Kille NJ, Chapman JM, Terry A v. (2013): Positive allosteric
modulator of alpha 7 nicotinic-acetylcholine receptors, PNU-120596 augments the effects
of donepezil on learning and memory in aged rodents and non-human primates.
Neuropharmacology 67: 201–212.
24. Buccafusco JJ, Terry A v. (2004): Donepezil-induced improvement in delayed matching
accuracy by young and old rhesus monkeys. Journal of Molecular Neuroscience 24: 85–91.
25. Buccafusco JJ, Jackson WJ, Stone JD, Terry A v. (2003): Sex dimorphisms in the cognitive-
enhancing action of the Alzheimer’s drug donepezil in aged Rhesus monkeys.
Neuropharmacology 44: 381–389.
26. Buccafusco JJ, Terry A v., Webster SJ, Martin D, Hohnadel EJ, Bouchard KA, Warner SE
(2008): The scopolamine-reversal paradigm in rats and monkeys: The importance of
computer-assisted operant-conditioning memory tasks for screening drug candidates.
Psychopharmacology 199: 481–494.
27. Rupniak NMJ, Tye SJ, Field MJ (1997): Enhanced performance of spatial and visual
recognition memory tasks by the selective acetylcholinesterase inhibitor E2020 in rhesus
monkeys. 406–410.
28. Millan MJ, Agid Y, Brüne M, Bullmore ET, Carter CS, Clayton NS, et al. (2012): Cognitive
dysfunction in psychiatric disorders: characteristics, causes and the quest for improved
therapy. Nature Reviews Drug Discovery 11: 141–168.
29. Thiele A, Bellgrove MA (2018): Neuromodulation of Attention. Neuron 97: 769–785.
30. Womelsdorf T, Thomas C, Neumann A, Watson M, Boroujeni KB, Hassani SA, et al.
(2021): A Kiosk Station for the Assessment of Multiple Cognitive Domains and Enrichment
of Monkeys. bioRxiv 2021.03.06.434198.
31. Watson MR, Voloh B, Thomas C, Hasan A, Womelsdorf T (2019): USE: An integrative suite
for temporally-precise psychophysical experiments in virtual environments for human,
nonhuman, and artificially intelligent agents. Journal of Neuroscience Methods 326.
https://doi.org/10.1016/j.jneumeth.2019.108374
32. Shiraishi T, Kikuchi T, Fukushi K, Shinotoh H, Nagatsuka SI, Tanaka N, et al. (2005):
Estimation of plasma IC50 of donepezil hydrochloride for brain acetylcholinesterase
19
inhibition in monkey using N-[11C] methylpiperidin-4-yl acetate ([11C]MP4A) and PET.
Neuropsychopharmacology 30: 2154–2161.
33. Irwin S (1968): Comprehensive observational assessment: Ia. A systematic, quantitative
procedure for assessing the behavioral and physiologic state of the mouse.
Psychopharmacologia 13: 222–257.
34. Patel S, Freedman S, Chapman KL, Emms F, Fletcher AE, Knowles M, et al. (1997):
Biological profile of L-745,870, a selective antagonist with high affinity for the dopamine
D4 receptor. Journal of Pharmacology and Experimental Therapeutics 283: 636–647.
35. Andersen MB, Fink-Jensen A, Peacock L, Gerlach J, Bymaster F, Lundbæk JA, Werge T
(2003): The muscarinic M1/M4receptor agonist xanomeline exhibits antipsychotic-like
activity in cebus apella monkeys. Neuropsychopharmacology 28: 1168–1175.
36. Gould RW, Russell JK, Nedelcovych MT, Bubser M, Blobaum AL, Bridges TM, et al.
(2020): Modulation of arousal and sleep/wake architecture by M1 PAM VU0453595 across
young and aged rodents and nonhuman primates. Neuropsychopharmacology 45: 2219–
2228.
37. Lendor S, Hassani S-A, Boyaci E, Singh V, Womelsdorf T, Pawliszyn J (2019): Solid Phase
Microextraction-Based Miniaturized Probe and Protocol for Extraction of Neurotransmitters
from Brains in Vivo. Analytical Chemistry 91: 4896–4905.
38. Hassani SA, Lendor S, Boyaci E, Pawliszyn J, Womelsdorf T (2019): Multineuromodulator
measurements across fronto-striatal network areas of the behaving macaque using solid-
phase microextraction. Journal of Neurophysiology 122: 1649–1660.
39. Cools R (2019): Chemistry of the Adaptive Mind: Lessons from Dopamine. Neuron 104:
113–131.
40. Parikh V, Pomerleau F, Huettl P, Gerhardt GA, Sarter M, Bruno JP (2004): Rapid assessment
of in vivo cholinergic transmission by amperometric detection of changes in extracellular
choline levels. European Journal of Neuroscience 20: 1545–1554.
41. Noonan MAP, Crittenden BM, Jensen O, Stokes MG (2018): Selective inhibition of
distracting input. Behavioural Brain Research 355: 36–47.
42. Womelsdorf T, Everling S (2015): Long-Range Attention Networks: Circuit Motifs
Underlying Endogenously Controlled Stimulus Selection. Trends in Neurosciences 38: 682–
700.
43. Sarter M, Lustig C, Blakely RD, Koshy Cherian A (2016): Cholinergic genetics of visual
attention: Human and mouse choline transporter capacity variants influence distractibility.
Journal of Physiology Paris 110: 10–18.
44. Azimi M, Oemisch M, Womelsdorf T (2019): Dissociation of nicotinic α7and α4/β2 sub-
receptor agonists for enhancing learning and attentional filtering in nonhuman primates.
Psychopharmacology 237: 997–1010.
45. Hassani SA, Oemisch M, Balcarras M, Westendorff S, Ardid S, van der Meer MA, et al.
(2017): A computational psychiatry approach identifies how alpha-2A noradrenergic
agonist Guanfacine affects feature-based reinforcement learning in the macaque. Scientific
Reports 7: 40606.
46. Hahn B (2015): Nicotinic Receptors and Attention.
47. Gratton C, Yousef S, Aarts E, Wallace DL, D’Esposito M, Silver MA (2017): Cholinergic,
but not dopaminergic or noradrenergic, enhancement sharpens visual spatial perception in
humans. Journal of Neuroscience 37: 4405–4415.
20
48. Jones RW, Soininen H, Hager K, Aarsland D, Passmore P, Murthy A, et al. (2004): A
multinational, randomised, 12-week study comparing the effects of donepezil and
galantamine in patients with mild to moderate Alzheimer’s disease. International Journal of
Geriatric Psychiatry 19: 58–67.
49. Wilkinson DG, Passmore AP, Bullock R, Hopker SW, Smith R, Potocnik FC v, et al. (2002):
A multinational, randomised, 12-week, comparative study of donepezil and rivastigmine in
patients with mild to moderate Alzheimer’s disease. International journal of clinical
practice 56: 441–6.
50. Courtney C, Farrell D, Gray R, Hills R, Lynch L, Sellwood E, et al. (2004): Long-term
donepezil treatment in 565 patients with Alzheimer’s disease (AD2000): randomised
double-blind trial. Lancet 363: 2105–2115.
51. Terry A v., Callahan PM (2019): Nicotinic acetylcholine receptor ligands, cognitive function,
and preclinical approaches to drug discovery. Nicotine and Tobacco Research 21: 383–394.
52. Jones CK, Byun N, Bubser M (2012): Muscarinic and nicotinic acetylcholine receptor
agonists and allosteric modulators for the treatment of schizophrenia.
Neuropsychopharmacology 37: 16–42.
53. Conn PJ, Jones CK, Lindsley CW (2009): Subtype-selective allosteric modulators of
muscarinic receptors for the treatment of CNS disorders. Trends in Pharmacological
Sciences 30: 148–155.
54. Foster DJ, Conn PJ (2017): Allosteric Modulation of GPCRs: New Insights and Potential
Utility for Treatment of Schizophrenia and Other CNS Disorders. Neuron 94: 431–446.
55. Rook JM, Bertron JL, Cho HP, Garcia-Barrantes PM, Moran SP, Maksymetz JT, et al.
(2018): A Novel M 1 PAM VU0486846 Exerts Efficacy in Cognition Models without
Displaying Agonist Activity or Cholinergic Toxicity. ACS Chemical Neuroscience 9: 2274–
2285.
56. Gould RW, Grannan MD, Gunter BW, Ball J, Bubser M, Bridges TM, et al. (2018):
Cognitive enhancement and antipsychotic-like activity following repeated dosing with the
selective M4 PAM VU0467154. Neuropharmacology 128: 492–502.
57. Geerts H, Guillaumat PO, Grantham C, Bode W, Anciaux K, Sachak S (2005): Brain levels
and acetylcholinesterase inhibition with galantamine and donepezil in rats, mice, and
rabbits. Brain Research 1033: 186–193.
58. Kaasinen V, Någren K, Järvenpää T, Roivainen A, Yu M, Oikonen V, et al. (2002): Regional
effects of donepezil and rivastigmine on cortical acetylcholinesterase activity in
Alzheimer’s disease. Journal of Clinical Psychopharmacology 22: 615–620.
59. Imbimbo B pietro (2001): Pharmacodynamic-tolerability relationships of cholinesterase
inhibitors for Alzheimer’s disease. CNS Drugs 15: 375–390.
60. Karasova JZ, Hrabinova M, Krejciova M, Jun D, Kuca K (2020): Donepezil and
rivastigmine: Pharmacokinetic profile and brain-targeting after intramuscular administration
in rats. Iranian Journal of Pharmaceutical Research 19: 95–102.
61. Rossi AF, Bichot NP, Desimone R, Ungerleider LG (2007): Top-down attentional deficits in
Macaques with lesions of lateral prefrontal cortex. Journal of Neuroscience 27: 11306–
11314.
62. Voytko M lou, Olton DS, Richardson RT, Gorman LK, Tobin JR, Price DL (1994): Basal
forebrain lesions in monkeys disrupt attention but not learning and memory. Journal of
Neuroscience 14: 167–186.
21
63. Hikosaka O, Kim HF, Amita H, Yasuda M, Isoda M, Tachibana Y, Yoshida A (2019): Direct
and indirect pathways for choosing objects and actions. European Journal of Neuroscience
49: 637–645.
64. Voloh B, Oemisch M, Womelsdorf T (2020): Phase of firing coding of learning variables
across the fronto-striatal network during feature-based learning. Nature Communications
11: 1–16.
65. Pawliszyn J (2000): Theory of Solid-Phase Microextraction. Journal of Chromatographic
Science 38: 270–278.
66. Cudjoe E, Bojko B, de Lannoy I, Saldivia V, Pawliszyn J (2013): Solid-Phase
Microextraction: A Complementary In Vivo Sampling Method to Microdialysis.
Angewandte Chemie International Edition 52: 12124–12126.
67. Reyes-Garcés N, Diwan M, Boyacl E, Gómez-Ríos GA, Bojko B, Nobrega JN, et al. (2019):
In vivo brain sampling using a microextraction probe reveals metabolic changes in rodents
after deep brain stimulation. Analytical Chemistry 91: 9875–9884.
68. Cudjoe E, Pawliszyn J (2014): Optimization of solid phase microextraction coatings for
liquid chromatography mass spectrometry determination of neurotransmitters. Journal of
Chromatography A 1341: 1–7.
69. Cuthbert BN, Insel TR (2013): Toward the future of psychiatric diagnosis: The seven pillars
of RDoC. BMC Medicine 11. https://doi.org/10.1186/1741-7015-11-126
70. Coppola JJ, Ward NJ, Jadi MP, Disney AA (2016): Modulatory compartments in cortex and
local regulation of cholinergic tone. Journal of Physiology Paris 110: 3–9.
71. Venkatesan S, Jeoung H-S, Chen T, Power SK, Liu Y, Lambe EK (2020): Endogenous
Acetylcholine and Its Modulation of Cortical Microcircuits to Enhance Cognition.
Behavioral Pharmacology of the Cholinergic System. pp 47–69.
72. de Boer P, Westerink BHC, Horn AS (1990): The effect of acetylcholinesterase inhibition on
the release of acetylcholine from the striatum in vivo: Interaction with autoreceptor
responses. Neuroscience Letters 116: 357–360.
SUPPLEMENTAL INFORMATION
Dose-dependent dissociation of pro-cognitive effects of donepezil on
attention and cognitive flexibility in rhesus monkeys
Seyed A. Hassani, Jason Russell, Sofia Lendor, Adam Neumann, Kanchan Sinha Roy, Kianoush
Banaie Boroujeni, Janusz Pawliszyn, Kari L. Hoffman, Carrie K. Jones, Thilo Womelsdorf
Content:
• Supplemental Methods
• Supplemental Results
• Supplemental Tables S1-S2
• Supplemental Figures S1-S4
Hassani et al.
Supplement
2
SUPPLEMENTAL METHODS
Ethics Statement
All animal related experimental procedures were in accordance with the National Institutes of
Health Guide for the Care and Use of Laboratory Animals, the Society for Neuroscience
Guidelines and Policies, and approved by Vanderbilt University Institutional Animal Care and Use
Committee.
Drug Procedures
For the double blinded drug administration, one experimenter prepared drug doses while another
handled injections and observations for potential side effects using a modified Irwin-rating scale.
Ratings were assigned on a scale of 0, 1, or 2 per monkey reflecting no change, a slight change or
a significant change respectively. Donepezil volumes were separated into vials for storage, and
were sonicated and vortexed with sterile saline immediately before injection. Depending on the
weight of the animals, the appropriate volume (0.1-0.7 ml) of donepezil was then drawn for the
planned injection dose; all daily injections were thus prepared together.
Visual Stimuli
The behavioral tasks used 3-dimensionally rendered visual objects, so called quaddles, which
varied in four visual feature dimensions (shape, color, pattern, and arms of a 3D rendered object)
described in detail elsewhere (1). Each visual feature dimension can be parametrically changed
which we then used to generate a number of variants, feature values, of each of the mentioned
visual features (e.g. up-, and downward bended arms with blunt pointy or flared shape). From here
on out, we will refer to the used visual feature spaces as ‘feature dimensions’ and any specific
variant of each visual feature as ‘feature values’. During training, all monkeys were exposed to a
so-called ‘neutral’ quaddle object composed of a spherical shape, grey color, uniform pattern, and
straight arms, which were features values that were never rewarded and served as a null feature
value for each dimension. Therefore, to practically achieve objects with only color and pattern
feature dimensions, and therefore without shape and arms, we kept the shape and arm dimension
constant at the neutral quaddle’s value for shape and arms while having color and pattern feature
values that were different from the neutral quaddle’s color and pattern.
Behavioral Tasks
Monkeys performed a visual search (VS) task and a feature-reward learning (FL) task in each
experimental session(1). For each experimental session and for the VS task, we selected randomly
3 feature dimensions from the pool of 4 possible dimensions (shape, color, pattern, arms) and we
we chose randomly 3 feature values per dimension (e.g. the 3 shape feature values pyramidal,
oblong and cubical) (Figure 1Bi).
Hassani et al.
Supplement
3
Visual search with different target-distractor similarity. The visual search (VS) task quantified
how much visual distractors slow down the detection of a target object and how the distraction
varied with the feature similarity of targets and distractors. The task required finding a cued object
amongst distractors on the screen by touching it for a minimum of 0.2s. At the beginning of each
VS block, the monkey learned which object is the target object in 10 familiarization trials that
presented the target object without any distractors. Touching the object triggered fluid reward. The
target was always an object that varied in three feature dimensions from the feature values of the
neutral object, i.e. a so called 3-D target. This is proceeded by 100 trials, each with a random
counterbalanced distribution of 3/6/9/12 distractors. Distractors were also 3-D objects with feature
values selected at random and thus could share 0/1/2 features with the rewarded object and could
be identical to other distractors within the same trial. If the distractors were dis-similar from the
target, independent of the number of distractors, trials may have a pop out effect with the target
being easily distinguishable while if distractors were similar to the target, trials may resemble a
conjunction search more closely (Figure 2A). Objects are presented at random within the
intersections of a 4x5 grid (example trials in Figure 1Bii).
Individual VS trials are initiated via a 0.3-0.5s long touch to a centrally presented blue square that
is 3° radius wide with a sidelength of 3.5 cm (baseline). This was then followed a 0.3-0.5s period
where the blue square disappears and there are no objects on screen except for the background
image. The task objects are then presented allowing the animal to freely explore for a maximum
of 5s (search + selection). During this 5s window, the animal could at any point touch and hold for
0.2s an object in order to select it. The selected object would then prompt both visual and auditory
feedback 0.2s after the selection lasting 0.5s. The color of the visual feedback and the pitch of the
auditory feedback correspond to the valence of the selected object’s value either signaling a correct
or incorrect choice. Correct choices were followed by fluid reward (water) (Figure 1C).
The VS task at the beginning and end of the experimental session utilized targets and distractors
that were composed of features from the same 3x3 feature space. Targets were never identical
between these two blocks but may appear as distractors in the other VS block. Similarly, all
distractors were created at random from the same 3x3 feature space as well and therefore would
be similar between the two blocks. The background image of the two VS blocks always differed
and acted as a cue for the VS rule set but are different in order to prevent the association of the
rewarded target object with a particular background image. Thus, the first and second VS search
block varied in the target object pulled from the same 3x3 feature space, the background image, as
well as the timing of their occurrence being at the start or the end of the daily session (Figure
1Aii).
Feature-reward learning at different distractor load. The feature-reward learning (FL) task
quantifies how fast and accurate subjects adjust to changing reward rules, indexing cognitive
flexibility. The task required monkey to learn by trial-and-error which object feature is associated
with reward. The same feature remained rewarded within blocks of 35-60 trials. Monkeys had to
choose one amongst three objects (1 target and 2 distractors) where a single feature value in a
single feature dimension is linked to reward with a p = 0.85 reward probability. Distractors contain
the same dimensions as the target but have different, non-repeated feature values. All objects are
presented in 1 of 8 possible locations randomly, all with 17 degrees eccentricity from the central
touch location (example trials in Figure 1Bii). With one experimental session we ran 21 FL blocks.
Hassani et al.
Supplement
4
The feature-reward association must be learned through trial and error and may switch after 35,
40, or 45 trials from the start of the block if the learning criterion is reached (80% over 10 trials)
or in 60 trials otherwise (uniform max FL block trial number). Block changes are un-cued but can
be inferred if there is a change in the object feature dimensions presented and the newly rewarded
feature value may be in the same dimension or a different dimension relative to the previously
rewarded feature value; the two types of shifts are semi-randomly determined to occur in similar
frequencies. The temporal structure and sequence of epochs in the FL task is the same as the VS
task.
The structure of the trials within the FL task was very similar to that of the VS task. Trials are
initiated in a similar manner via a 0.3-0.5s touch on a central blue square. This is followed by a
0.3-0.5s period where the blue square is not present and task objects have not yet been made visible
yet. Three task objects are then presented for up to 5 sec during which at any point the subject is
allowed to make a 0.2s touch and hold on an object to select it. Following a 0.2s delay after the
selection of an object, auditory and visual feedback as well as potentially fluid reward are presented
for 0.5 s. The pitch of the auditory feedback and the color of the visual feedback vary depending
on the presence of reward and not on making a high reward probability choice (Figure 1C).
Neurochemical Quantification of Drug Effect
To confirm the bioactivity of donepezil in the brain we measured the neurochemistry in the
prefrontal cortex and the head of the caudate nucleus after IM administering a low (0.06 mg/kg)
and high (0.3 mg/kg) dose of donepezil. We used microprobes that sampled the local
neurochemical milieu with the principles of solid phase micro-extraction (SPME) probes (2)
previously shown to provide comparable and complimentary outcomes to micro-dialysis (3,4).
These probes sample the drug and metabolites of the neurotransmitters (e.g choline) via diffusion
until an equilibrium is reached with the extracellular concentrations. The detailed procedures used
here are described in (5). In brief, for each brain area a microdrive was prepared holding a cannula
and SPME probe inside it, as well as a microdrive with an electrode to record activity prior to
SPME sampling. The electrode was driven to the target location in prefrontal cortex / striatum.
The target location was confirmed by measuring spiking activity of neurons from the electrode.
The cannula shielded SPME was then lowered to just above the target area and the SPME probe
was then exposed to gray matter of the target area for 20 minutes before being retracted into their
respective cannula and drive back out of the brain. Samples were then stored in a -80°C freezer,
stored for less than 2 weeks and shipped overnight in dry ice to Waterloo, Ontario (Canada) where
they were desorbed and underwent liquid chromatography separation and mass spectrometry
quantification.
The SPME probes were desorbed into 50 µL of acetonitrile/methanol/water 40:30:30 solution
containing 0.1% formic acid and internal standard citalopram-D6 at 20 ng/mL for 1 h with agitation
at 1500 rpm. The LC-MS/MS analysis was carried out using an Ultimate 3000RS HPLC system
coupled to a TSQ Quantiva triple quadrupole mass spectrometer (Thermo Fisher Scientific, San
Jose, CA, USA). Data acquisition and processing were performed using Xcalibur 4.0 and Trace
Finder 3.3 software (Thermo Fisher Scientific, San Jose, CA, USA). The chromatographic
separation employed Hypersil Gold C18 column, 50 x 2.1 mm, 1.9 μm particle size (Thermo
Scientific, Ashville, NC, United States) held at 35°C. The aqueous mobile phase (A) consisted of
Hassani et al.
Supplement
5
water/acetonitrile/methanol 90:5:5 with 0.1 % formic acid, while the organic mobile phase (B)
consisted of acetonitrile/water 90:10 with 0.1% formic acid. The following chromatographic
gradient at a flow rate of 400 µL/min was applied (%B): 0-0.5 min 0 %; 0.5-3 min linear gradient
to 100 %; 3-3.65 min held at 100 %; 3.65-3.7 min linear gradient to 0 %; re-equilibration at 0 %
until 4.5 min. The injection volume was 5 μL. The MS/MS analysis was performed in positive
ionization mode under selected reaction monitoring (SRM) conditions; for the analyte donepezil
the quantifier transition monitored was m/z 380.3 -> 243.2 and the qualifier transition was 380.3 -
> 91.1, while one transition at m/z 331.1 -> 109.1 was monitored for internal standard citalopram-
D6. The capillary voltage was set at 3.5 kV, with the remaining electrospray source conditions set
to the following values: vaporizer temperature 358 °C, ion transfer tube temperature 342 °C, sheath
gas pressure 45, auxiliary gas pressure 13, and sweep gas pressure 1 (arbitrary units). The
instrumental stability throughout the sequence was monitored by analysis of an instrumental QC
sample consisting of the target analyte and internal standards spiked into a neat desorption solvent
at 20 ng/mL.
The concentration of donepezil in brain tissue was determined using a modified external surrogate
matrix-matched calibration approach developed in previous work (5–7). The surrogate matrix
consisted of agarose gel (1% agarose in PBS solution, w/v) mixed with lamb brain homogenate in
the ratio 1:1 (v/w). Prior to combining the agarose gel with the brain homogenate, the latter was
spiked with donepezil in the concentration range 5-750 ng/g. Extractions were carried out in static
mode from 1g of the matrix with extraction time matching the in vivo experiments. The probes
were subsequently rinsed with water and desorbed into 50 µL of the desorption solvent containing
internal standard citalopram-D6 at 20 ng/mL. The analytical response in the form of relative peak
area ratios (analyte to internal standard) was converted to amounts extracted by employing an
instrumental calibration curve consisting of donepezil in neat desorption solvent in the range 0.1-
100 ng/mL. The resulting matrix-matched calibration curve was expressed as amounts extracted
in the function of concentration in tissue. A weighted linear regression equation was fitted to the
analytical response in the function of concentration. Limits of quantitation (LOQ) were determined
as the lowest concentration of analyte producing a signal to noise ratio ≥ 5, with a relative standard
deviation (RSD) of 4 replicate measurements below 20%, and accuracy within 20%. Accuracy was
calculated as the relative percent error of concentrations of analytes in the calibrator samples
determined experimentally with the use of calibration curves versus theoretical (spiked)
concentrations (8).
A single, fourth, Macaca mulatta (male, 8 years old) with an implanted recording chamber above
the left hemisphere was chaired, head-fixed and performed the VS task (data not included in
analysis) to emulate performance by the other 3 subjects. Details about the surgical implantation
of the recording chamber and headpost are reported in (5). Performance of the VS, virtually
identical to the VS task reported above, was done with eye saccades using a Tobii Spectrum eye
tracker instead of touch screen. Subject underwent 6 instances of both 0.06 and 0.3 mg/kg
donepezil dosing in the primate chair at the same time of day as the other 3 NHPs received
donepezil. Injections were done after the animal had been performing one VS block for roughly
20min, followed by a 15min period of quiet wakefulness after which they proceeded to do a second
VS block. SPME sampling events took place once at the beginning of each VS block with probes
being exposed to tissue for 20min in both instances.
Hassani et al.
Supplement
6
During each SPME sampling event heart rate (HR) was monitored using a pulse oximeter
(PalmSAT 2500, Nonin Inc, MN), with the sensor clipped at the ear lobe of the subject and a
sampling rate of 0.25 Hz. HR data was collected 20 min before task start both before and after
donepezil injection. The data was smoothed with an centered 8 sample window (40 sec) with 1
sample shifts (4 sec) and normalized to the average HR 5 min before task start.
Literature Survey
In order to place our results within the broader published work, we identified 9 papers involving
donepezil and non-human primates (9–17). These papers have relevant details such as the task(s)
performed, donepezil dosage and administration method and more extracted, where appropriate
(Table S1). Notably, 6 of the identified papers provided donepezil in conjunction with other
pharmaceutical agents such as Scopolamine. The papers were found by conducting an online
search of the NIH (PubMed) database, as well as google scholar. The keyword search terms of
‘Donepezil’, ‘Aricept’ and ‘E2020’ were used with the terms ‘NHP’, ‘monkey’, ‘primate’,
‘cognition’, and ‘brain’ or some combination of them. We did not consider 8 studies that utilized
donepezil in primates as they lacked a cognitive component. They did however provide insight in
dosing ranges for different dosing routes, dose-limiting side effects and donepezil’s kinetics (18–
25).
Data Analysis
All behavioral analysis was completed using MATLAB (Mathworks Inc., MA).
Analysis of Visual Search. The set size effect of the VS performance was either defined as
proportion of correct trials by the distractor number or by the average t-d similarity of trials. The
set-size effect was estimated by a linear regression which is specified as either utilizing distractor
number or t-d similarity. The average t-d similarity of a trial was calculated by averaging the
number of shared feature values (0/1/2) of all distractors in said trial to the target. Reaction times,
referred to as choice RTs for the VS task, were defined as the time from the initiation of a trial by
pressing and holding the central blue square to the initiation of touch to the selected object leading
to feedback. Reaction time data only takes into consideration rewarded trials. Descriptive statistics
are provided as means with ±SEMean unless specified otherwise. Similarly, error bars in figures
are either mean ±SEMean or median ±SEMedian unless specified otherwise. After pooling data
from all three subjects, the measure of interest is averaged across appropriate trials or blocks to
get a per session value.
Analysis of Feature-Reward Learning. FL blocks were either labeled as ‘low distractor load’ if no
distracting feature dimension was present, or as ‘high distractor load’ if a single distracting feature
dimension was presented alongside the feature dimension to which the rewarded feature value
belonged to. We calculated learning curves by averaging smoothed trial-wise performance aligned
to block reversals. We defined learning speed by calculating at which trial, since block start, the
subjects started performing at ≥80% over 10 trials, the maximally rewarded object. This trial was
termed the ‘learning point’ (LP). For analysis, blocks were excluded if the monkey took a break
Hassani et al.
Supplement
7
of at least several minutes. Furthermore, blocks were excluded where the LP was calculated to be
trial 1 (reflecting ≥80% performance in the first 10 trials since reversal) as well as if the LP
occurred beyond the 40th trial. Reaction times, referred to as choice RTs for the FL task, were
temporally defined the same as for the VS task and also only include rewarded trials.
Perseverative errors were defined as two or more consecutive choices of low probability rewarded
objects with at least 1 shared feature value. Analysis of perseverative errors for feature values in
the same feature dimension as the target feature are separated from those where the perseverated
feature value was in the distracting dimension. For perseverative errors to occur in the distractor
dimension, the block is required to contain a distracting dimension to begin with and is therefore
necessarily a high distractor load block. Perseverative errors in the target feature dimension could
occur in both low and high distractor load blocks.
Statistical Analysis of Drug Effects
Comparisons between vehicle and donepezil doses (0.06, 0.1 and 0.3 mg/kg) were done for all
doses combined followed by post-hoc pair-wise statistics with multiple comparisons corrections
unless specified otherwise. Probability level of less than 5% (p < 0.05) was considered statistically
significant.
SUPPLEMENTAL RESULTS
Overall Visual Search Performance
We performed and report here the results of various analysis to evaluate the overall performance
of the animals on the tasks, or to test specific performance metrics that provide a more
comprehensive overview of how the drug conditions did or did not affect task performance.
For the visual search task, 10 familiarization trials with no distractors were presented prior to each
of the two visual search blocks. The reaction times to detect these single objects on the screen will
be referred to as speed of processing (SoP). They were completed in 628 ms ±133 (Ig: 616 ms
±8.5; Wo: 693 ms ±6.6; Si: 588 ms ±5.3) with the first block having faster SoP at 611 ms ±7.7
relative to the second block with 646 ms ±4.8 (p < .001)(Figure S1A).
On average, monkeys performed the VS task with 84.4% (± 0.54) accuracy (Monkey Ig: 85.2%
±0.81; Wo: 88.3% ±0.94; Si: 79.8% ±0.97) and with 1158 ms ±9.7 search times (Ig: 1281 ms
±18.3; Wo: 1171 ms ±15.9; Si: 1020 ms ±13.3). Increasing numbers of distractors slowed search
RTs, with 3/6/9/12 distractors having 1019 ms, 1216 ms, 1409 ms, and 1552 ms search times
respectively (Distractor Number F(3,1240) = 241.32, p < .001) as well as decreasing accuracy,
with 3/6/9/12 distractors having 91.7% ±0.6, 87.1% ±0.6, 82.9% ±0.8, and 80.0% ± 0.9 accuracy
respectively (all pair-wise comparisons were significant using Tukey's HSD multiple comparisons
test among proportions at an alpha of 0.05, except for 0.1 mg/kg donepezil dose and vehicle).
Search RTs did not vary significantly with regards to VS block number (Block Number F(1,1240)
= 3.18, n.s.) (Figure S1B) while trial outcomes did vary significantly with VS block number (X2(1,
N1 = 14500, N2 = 16700) = 40.0, p < .001) (Figure 1D). This difference in performance between
Hassani et al.
Supplement
8
the two VS blocks may be due to fatigue, as reflected by the significantly reduced SoP, or
otherwise satiation.
Both the change in search time and performance by distractor number were fit by a linear
regression, revealing that each additional distractor increased search duration on average by 60 ±
1.6 ms (Ig: 72 ± 2.7 ms/distractor; Wo: 57 ± 2.8 ms/distractor; Si: 49 ± 2.1 ms/distractor) as well
as decreasing performance by 1.3% ± 0.1 per additional distractor (Ig: 1.2% ± 0.1; Wo: 0.9% ±
0.1; Si: 1.8% ± 0.1) (inlets in Figures 1D and S1B show individual monkey fits for vehicle). The
set size effect on search RT was on average larger in the first than the second VS block (first VS
block: 63 ms/distractor; second VS block: 56 ms/distractor; p = .0254; Ig: p = .0604; Wo: p =
.0401; Si: p = .7199). The set size effect on performance was on average the same in the first and
the second VS block (first VS block: -1.3% change in performance per distractor; second VS block:
-1.3% change in performance per distractor; p = n.s.; Ig: p = n.s.; Wo: p = n.s.; Si: p = n.s.).
We analyzed how the similarity of distractors with the target influenced search RT and
performance. Distracting stimuli could have 0, 1 or 2 shared feature values with the target and the
thus some trials could provide a greater challenge for the monkeys given the average target-
distractor similarity (t-d similarity)(Figure 2A). We found that search RT increased with average
t-d similarity from 1227 ms ±9 to 1410 ms ±7 and 1334 ms ±17 for low, medium and high t-d
similarity respectively (Similarity F(2,14467) = 107.1, p < .001)(Figure S4A). VS performance
decreased with t-d similarity from 92.9% ±0.4 to 85.5% ±0.3 and 81.6% ±1.0 for low, medium
and high t-d similarity respectively (Similarity F(2,672) = 16.17, p < .001) (Figure 2B). Both
distractor number, and t-d similarity impact VS performance significantly (Distractor number
F(3,16688) = 50.25, p < .001; T-D similarity main effect p < 0.001)(Figure 4B-C) but no
significant interaction was found between the two variables (T-D Similarity x Distractor Number
F(6,16688) = 1.16, n.s.)(Figure 4D) with VS RT showing a similar relationship with significant
main effects (distractor number main effect p < 0.001; T-D similarity F(2,16688) = 55.24, p <
.001) but not interaction (F(6,16688) = 1.16, n.s.)(Figure 2D). Individual sessions also showed no
strong correlation between the set size effect of performance by distractor number relative to the
set size effect of performance by t-d similarity (Pearson, n.s.)(Figure S3).
Overall Feature Reward Learning Performance
For the feature reward learning (FL) task, monkeys reached learning criterion on average in 63
±1% of the 21 daily learning blocks (Ig: 71 ±1%; Wo: 61 ±2%; Si: 56 ±1%) once exclusion criteria
were applied (see methods). Learning criterion was reached more often in the low distractor load
than high distractor load blocks with proportion of learned blocks being 70% and 56% respectively
(Ig: 80 vs 62% of blocks; Wo: 66 vs 56%; Si: 63 vs 49%). Average learning curves for low and
high distractor load blocks of each individual monkey, as well as the average across monkeys is
provided in Figure 3A. Monkeys reached the learning criterion on average within 12.5±0.2 and
15.6±0.2 trials in the low and high distractor load condition (Ig: 9.9±0.2 and 14.9±0.3; Wo:
13.5±0.4 and 17.0±0.4; Si: 14.9±0.4 and 15.0±0.4). The average choice reaction time of a correct
FL trial was 986 ±2 ms with faster reaction times in the low than high distractor load blocks (p <
.001; 965 ±3 ms and 1013 ±3 ms respectively).
Visual search performance with donepezil
Hassani et al.
Supplement
9
The speed of processing (SoP; reaction time to a single object during familiarization trials) showed
a significant main effect of block number (Block Number F(1,424) = 6.29, p < .001), as well as a
significant main effect of drug condition (Condition F(3,424) = 15.16, p < .001)(Figure 2B). Pair-
wise statistics comparing the first block SoPs of the control condition and 0.06, 0.1 and 0.3 mg/kg
doses (Tukey’s, n.s, n.s, and p < .001 respectively) suggests that the main effect of condition is
driven by the 0.3 mg/kg dose SoP.
Feature learning performance with donepezil
For the low distractor load condition the proportion of learned blocks were on average 72.3% ±1.4,
75.2% ±2.4, 78.0% ±3.9 and 75.9% ±3.2 in the vehicle, and 0.06 / 0.1 / 0.3 mg/kg) days, which
was not significant (n.s.). Similarly, for the high distractor load condition the proportion of learned
blocks was 60.0% ±1.5, 74% ±4.0, 62.3% ±3.8 and 65.0% ±4.2 in the vehicle, and 0.06 / 0.1 / 0.3
mg/kg) days (n.s.). Comparisons between the proportion of blocks learned in low and high
distractor load conditions revealed a significant difference for vehicle and drug conditions with
more blocks learned in the low distractor load condition than in the high distractor load condition
(X2 values contrast 1D vs 2D learning blocks for vehicle, 0.06, 0.1, and 0.3 mg/kg conditions: X2(1,
N1 = 1757, N2 = 1750) = 58.6, p < .001; X2(1, N1 = 222, N2 = 219) = 8.2, p = .004; X2(1, N1 = 219,
N2 = 222) = 12.6, p < .001; X2(1, N1 = 209, N2 = 190) = 5.6, p = .018 for vehicle, 0.06, 0.1 and 0.3
mg/kg doses respectively). Monkey Ig had a higher overall proportion of blocks learned than both
Monkey Wo and Si with vehicle (p < .001), however, there were no statistically significant
differences between monkeys between low and high distractor loads in vehicle or any drug
conditions (n.s.).
In addition to learning speed we also analyzed in detail the choice reaction times across drug
conditions. Relative to the low distractor load condition, in the high distractor load condition, FL
choice RTs slowed from 993 ±11 to 1060 ±14, from 964 ± 31 to 1048 ±33, from 988 ±27 to 1015
±36, and from 1126 ±29 to 1167 ±31 ms for the vehicle, 0.06, 0.1 and 0.3 mg/kg donepezil doses
respectively (Figure 3E).
There was also significant inter-subject variability in choice RTs with monkey Si having
significant faster choice RTs in the FL task (Subject F(2,1052) = 183.53, p < .001)(Figure S2C)
as well as a significant monkey-drug interaction (F(6,1052) = 3.5, p = .002). Alongside the general
slowing with the 0.3 mg/kg donepezil dose (see main text), we found in a pair-wise analysis a
significant slowing of search RTs with the 0.3 mg/kg donepezil dose for monkey Si (Tukey’s, p <
.001), and a significant faster search RTs with the 0.1 mg/kg donepezil dose for monkey Wo
(Tukey’s, p = .003).
The main text reports the length of consecutive, perseverative errors. Perseverative errors may
occur in the same dimension as the target feature value (12% of all errors), possible in low and
high attentional load blocks, or they may occur in the distracting dimension (26% of all errors)
only possible in high attentional load blocks. The proportions of perseverative errors within the
target dimension were 12% ±1, 11% ±2, 12% ±2 and 11% ±2 for vehicle, 0.06, 0.1 and 0.3 mg/kg
donepezil doses (n.s.), while within the distracting dimension they were 26% ±0, 23% ±6, 22%
±2, and 24% ±2 for vehicle, 0.06, 0.1 and 0.3 mg/kg donepezil doses (n.s.).
Hassani et al.
Supplement
10
We next analyzed whether donepezil modified how flexible subjects learned a new target feature
depending on whether the target feature was from a novel feature dimension and whether the target
was a previous distractor. First, we asked whether donepezil modulates learning differently
depending on whether a newly rewarded (target) feature values belonged to the same feature
dimension as the target feature in the previous block, or to a new target feature dimension. This
analysis quantifies whether learning a new feature set was easier or more difficult than re-assigning
a reward association within the previously relevant feature set. In our task a shift to a new target
feature of a new dimension should be easier because it occurred by presenting new objects that
were not shown in the previous block. We thus compared learning speed for blocks where the
rewarded feature dimension was not presented in the previous block and blocks where the
rewarded feature value was from the same dimension as the previously rewarded feature. We found
that donepezil did not alter learning for block transitions to ‘new target feature dimensions’ versus
‘another feature of the same dimension’ (Condition F(3,2708) = 0.55, n.s.; Block Switch F(1,2708)
= 2.7, n.s.; Condition x Block Switch F(3,2708) = 1.15, n.s.).
Secondly, we quantified whether donepezil modulated how subjects learned a new target feature
value when that target feature was a distractor in the previous block. Difficulties in attending a
previous distractor is sometimes referred to as latent inhibition. There were only few learning
blocks available in which the target feature dimension was a distracting feature dimension in the
previous block which we contrasted to blocks where the rewarded feature dimension was not
presented in the previous block. We found that donepezil did not alter learning speed for blocks
where the ‘target was a previous distractor’, versus when the ‘target was a new feature’ (Condition
F(3,1450) = 0.31, n.s.; Block Switch F(1,1450) = 0.02, n.s.; Condition x Block Switch F(3,1450)
= 0.2, n.s.).
Hassani et al.
Supplement
11
SUPPLEMENTAL TABLES
Table S1. A summary of the literature testing donepezil’s cognitive effects in nonhuman
primates.
Table S1. A summary of the literature testing donepezil’s cognitive effects in nonhuman primates.
Relevant Task(s)
Subject Details
Dosage &
Administration
Cognitive Domain
Relevant Results
Reference
Object retrieval detour
(ORD) cognition test
Macaca mulatta
(male and female)
0.3, 0.56, 1, 1.8, 3, 5
mg/kg PO*
Reasoning & problem
solving (exec function)
Significant interaction of trial type (easy
vs difficult) and treatment. Main effect of
treatment on the difficult condition
Vardigan et al., 2015
(9)
Paired-associates
learning (PAL);
Continuous-performance
task (CPT)
Macaca mulatta (18
males)
0.3-3 mg/kg PO
(PAL task); 0.1-0.25
mg/kg IM (CPT
task)*
PAL: Working memory;
CPT: Attention/vigilance
(exec functioning)
Attenuated scopolamine-induced
impairments in PAL (at 1.0 and 3.0
mg/kg PO) and CPT (at 0.25 mg/kg IM)
Lange et al., 2015 (10)
Delayed matching-to-
sample (DMTS) task
Macaca mulatta (4
aged male, 3 aged
female)
0.003-0.2 mg/kg PO*
Working memory
Enhanced DMTS accuracy in ‘long
delay’ condition at 0.01, 0.025, 0.05, 0.1
and 0.2 mg/kg doses. No changes in ITI
or choice latency
Callahan et al., 2013
(11)
Self-ordered spatial
search
Macaca fascicularis
(6 females; ~15 years
old)
3 mg/kg PO*
Working memory,
Attention/vigilance
Attenuated the scopolamine-induced
impairments in the self-ordered spatial
search task
Uslaner et al., 2013 (12)
Delayed matching-to-
sample (DMTS) task
Macaca mulatta (17
male and 16 female;
average ~18 years
old)
10, 25, 50, 100 ug/kg
IM
50-400 ug/kg PO*
Working memory
Accuracy in long delay condition with
IM administration was improved
Buccafusco et al., 2008
(13)
Delayed matching-to-
sample (DMTS) task
Macaca mulatta (8
male & 9 female; 9-
29 years old)
10, 25, 50, 100 ug/kg
IM
Working memory
Accuracy increased in medium and long
delayed trials (with 25 ug/kg dose being
the most efficacious)
Buccafusco & Terry
2004 (14)
Oculomotor delay
response task (ODR);
Visually guided saccade
task (VGS)
Macaca mulatta
(male; 5 ~5 years old
and 5 ~20 years old)
50, 250 ug/kg IV
ODR: Working memory;
VGS: attention/vigilance
ODR performance was improved in aged
monkeys (not young monkeys). No
changes reported in VGS (assay of motor
performance, not cognition)
Tsukada et al., 2004
(15)
Delayed matching-to-
sample (DMTS) task
Macaca mulatta
(male & female; >20
years old)
0.01, 0.025, 0.05, 0.1
mg/kg IM
Working memory
Accuracy improved independent of drug
dose but dependent on delays
(improvement occurred in medium and
long delays)
Buccafusco et al., 2003
(16)
Spatial delayed response
task (SDRT); Visual
recognition task (VRT)
Macaca mulatta (9
males)
0.01-1.75 mg/kg IM
(SDRT task); 0.003-
0.06 mg/kg IM (VRT
task)*
SDRT: Working
memory; VRT:
attention/vigilance
SDRT: Donepezil rescued effects of
scopolamine in difficult trials
VRT: Performance was enhanced with
donepezil pre-treatment (by ~10%)
Rupniak et al., 1997
(17)
*: Other drugs co-administered at some/all doses.
Hassani et al.
Supplement
12
Table S2. A summary of observed dose-limiting side effects. The effects of donepezil (0.06, 0.1
and 0.3 mg/kg IM) on autonomic and somatomotor system function were evaluated. The mean
score of 3 monkeys was classified as follows: - no effect; + 0-0.15; ++ 0.16-0.3; +++ 0.31-0.45.
Donepezil
0.06 mg/kg
0.1 mg/kg
0.3 mg/kg
Observation
Pre-task
Post-task
Pre-task
Post-task
Pre-task
Post-task
Autonomic Nervous System
Salivation
-
-
-
-
-
-
Lacrimation
-
-
-
-
-
-
Urination
-
-
-
-
-
-
Defecation (amount)
-
-
-
-
+
-
Defecation (consistency)
-
-
-
-
+
-
Emesis
-
-
-
-
-
-
Miosis
-
-
-
-
-
-
Mydriasis
-
-
-
-
-
-
Ptosis
-
-
-
-
+
-
Exophtalmos
-
-
-
-
-
-
Piloerection
-
-
-
-
-
-
Respiratory Rate
-
-
-
-
-
-
Yawn
-
-
-
-
+
-
Vasodilation
-
-
-
-
-
-
Vasoconstriction
-
-
-
-
+++
-
Irritability
-
-
-
-
-
-
Body Temp.
-
-
-
-
-
-
Somatomotor Systems
Physical Appearance
-
-
-
-
+++
-
Tremor
-
-
-
-
-
-
Leg Weakness
-
-
-
-
-
-
Catalepsy
-
-
-
-
-
-
Visuo-Motor Coordination
-
-
-
-
-
-
Posture
-
-
-
-
+++
-
Unrest
-
-
-
-
-
-
Stereotypies
-
-
-
-
-
-
Arousal
-
-
-
-
-
-
Sedation
-
-
-
-
+++
-
Oral Dyskinesia
-
-
-
-
++
-
Bradykinesia
-
-
-
-
+
-
Dystonia
-
-
-
-
-
-
Table S2. A summary of observed dose-limiting side effects. The effect of Donepezil (0.06, 0.1 and 0.3 mg/kg IM) on autonomic and somatomotor system
function were evaluated. The mean score of 3 monkeys was classified as follows: - no effect; + 0-0.15; ++ 0.16-0.3; +++ 0.31-0.45
Hassani et al.
Supplement
13
SUPPLEMENTAL FIGURES
Figure S1. Search reaction time in the visual search task and its
relationship with distractor number. A. The average speed of processing
(SoP), for each condition separated by block. The SoP is significantly
changed between the first and second VS block (Block Number F(1,424)
= 6.29, p < .001) as well as between conditions (Condition F(3,424) =
15.16, p < .001; ANOVA). The average SoP in the first VS block is
significantly slowed with a 0.3 mg/kg dose of donepezil (Tukey’s, p <
0.001). B. Average search RT per distractor number for vehicle and all
donepezil doses combined, both separated by the first vs second VS block.
The number of distractors slowed search RTs (Distractor F(3,1722) =
333.1, p < .001) while the VS block number did not significantly impact
search RTs (Block F(1,1722) = 0.64, n.s.). Donepezil administration,
averaged over all doses, had a significant effect on search RT (Condition
F(1,1722) = 4.83, p = .028), in particular in the first VS block. The inlet
shows individual monkey average search RT linear fits. C. The difference
in search RT by distractor number between donepezil 0.06, 0.1, 0.3 mg/kg
doses and vehicle for the first VS block (F(3,896) = 15.15, p < 0.001) with
the 0.3 mg/kg dose having significantly slower search RT than vehicle (p
= .023). Error bars are standard deviations. D. The set size effect of search
RT by distractor number for each condition. No significant difference was
observed between drug conditions and vehicle.
Hassani et al.
Supplement
14
Figure S2. The relationship between performance and reaction times in
both the visual search task and feature learning task. A. The session-wise
correlation between VS performance and search RT by individual monkey.
No difference between monkeys was found. B. Same as A but for all
monkeys combined and separated by condition. Only vehicle and 0.1
mg/kg doses had a significant correlation, however no significant change
in correlation relative to vehicle was found. C. Similar to A but looking at
the correlation between FL performance (learning speed) and choice RT.
Monkey Si was found to have significantly faster choice RTs (Subject
F(2,1052) = 183.53, p < .001). D. The same as C but for all monkeys
combined and separated by condition. No conditions exhibited significant
correlations.
Hassani et al.
Supplement
15
Figure
S3.
The
relationship
between the set-size effect of
visual search performance as a
function of distractor number
versus target-distractor similarity.
Session-wise
linear
fits
to
performance by distractor number
(x-axis)
and
target-distractor
similarity (y-axis). There was no
significant correlation at any
condition.
Hassani et al.
Supplement
16
Figure S4. Search reaction times within the visual search task as
a function of target-distractor similarity and distractor number. A.
Search reaction time plots as a function of t-d similarity instead of
distractor number (as was the case in Figure S1) for vehicle and
all donepezil doses. There was a significant main effect of
condition with the 0.3 mg/kg donepezil dose being significantly
different from vehicle (F(3,267) = 7.75, p < .001; Tukey’s, p <
.001). B. Visualization of the combined effect of distractor
number and t-d similarity on search RT. From left to right, each
cluster of lines represents increasing distractor numbers while
data within each line represents low, medium and high t-d
similarity from left to right respectively. Both distractor number
(F(3, 14458) = 294.93, p < .001) and t-d similarity (F(2,14458) =
16.87, p < .001) impact VS performance with no significant
interaction (F(6, 14458) = 1.19, n.s.).
Hassani et al.
Supplement
17
Supplemental References
1. Watson MR, Voloh B, Thomas C, Hasan A, Womelsdorf T (2019): USE: An integrative suite
for temporally-precise psychophysical experiments in virtual environments for human,
nonhuman, and artificially intelligent agents. Journal of Neuroscience Methods 326.
https://doi.org/10.1016/j.jneumeth.2019.108374
2. Pawliszyn J (2000): Theory of Solid-Phase Microextraction. Journal of Chromatographic
Science 38: 270–278.
3. Cudjoe E, Bojko B, de Lannoy I, Saldivia V, Pawliszyn J (2013): Solid-Phase
Microextraction: A Complementary In Vivo Sampling Method to Microdialysis.
Angewandte Chemie International Edition 52: 12124–12126.
4. Cudjoe E, Pawliszyn J (2014): Optimization of solid phase microextraction coatings for liquid
chromatography mass spectrometry determination of neurotransmitters. Journal of
Chromatography A 1341: 1–7.
5. Hassani SA, Lendor S, Boyaci E, Pawliszyn J, Womelsdorf T (2019): Multineuromodulator
measurements across fronto-striatal network areas of the behaving macaque using solid-
phase microextraction. Journal of Neurophysiology 122: 1649–1660.
6. Lendor S, Gómez-Ríos GA, Boyacl E, vander Heide H, Pawliszyn J (2019): Space-resolved
tissue analysis by solid-phase microextraction coupled to high-resolution mass spectrometry
via desorption electrospray ionization. Analytical Chemistry 91: 10141–10148.
7. Lendor S, Hassani S-A, Boyaci E, Singh V, Womelsdorf T, Pawliszyn J (2019): Solid Phase
Microextraction-Based Miniaturized Probe and Protocol for Extraction of Neurotransmitters
from Brains in Vivo. Analytical Chemistry 91: 4896–4905.
8. Food and Drug Administration (FDA) (2001): Guidance for Industry Bioanalytical Method
Validation. Veterinary Medicine 1–25.
9. Vardigan JD, Cannon CE, Puri V, Dancho M, Koser A, Wittmann M, et al. (2015): Improved
cognition without adverse effects: Novel M1 muscarinic potentiator compares favorably to
donepezil and xanomeline in rhesus monkey. Psychopharmacology 232: 1859–1866.
10. Lange HS, Cannon CE, Drott JT, Kuduk SD, Uslaner JM (2015): The M1 muscarinic
positive allosteric modulator PQCA improves performance on translatable tests of memory
and attention in rhesus monkeys. Journal of Pharmacology and Experimental Therapeutics
355: 442–450.
11. Callahan PM, Hutchings EJ, Kille NJ, Chapman JM, Terry A v. (2013): Positive allosteric
modulator of alpha 7 nicotinic-acetylcholine receptors, PNU-120596 augments the effects
of donepezil on learning and memory in aged rodents and non-human primates.
Neuropharmacology 67: 201–212.
12. Uslaner JM, Eddins D, Puri V, Cannon CE, Sutcliffe J, Chew CS, et al. (2013): The
muscarinic M1 receptor positive allosteric modulator PQCA improves cognitive measures
in rat, cynomolgus macaque, and rhesus macaque. Psychopharmacology 225: 21–30.
13. Buccafusco JJ, Terry A v., Webster SJ, Martin D, Hohnadel EJ, Bouchard KA, Warner SE
(2008): The scopolamine-reversal paradigm in rats and monkeys: The importance of
computer-assisted operant-conditioning memory tasks for screening drug candidates.
Psychopharmacology 199: 481–494.
14. Buccafusco JJ, Terry A v. (2004): Donepezil-induced improvement in delayed matching
accuracy by young and old rhesus monkeys. Journal of Molecular Neuroscience 24: 85–91.
Hassani et al.
Supplement
18
15. Tsukada H, Nishiyama S, Fukumoto D, Ohba H, Sato K, Kakiuchi T (2004): Effects of Acute
Acetylcholinesterase Inhibition on the Cerebral Cholinergic Neuronal System and Cognitive
Function: Functional Imaging of the Conscious Monkey Brain Using Animal PET in
Combination with Microdialysis. Synapse 52: 1–10.
16. Buccafusco JJ, Jackson WJ, Stone JD, Terry A v. (2003): Sex dimorphisms in the cognitive-
enhancing action of the Alzheimer’s drug donepezil in aged Rhesus monkeys.
Neuropharmacology 44: 381–389.
17. Rupniak NMJ, Tye SJ, Field MJ (1997): Enhanced performance of spatial and visual
recognition memory tasks by the selective acetylcholinesterase inhibitor E2020 in rhesus
monkeys. 406–410.
18. Gould RW, Russell JK, Nedelcovych MT, Bubser M, Blobaum AL, Bridges TM, et al.
(2020): Modulation of arousal and sleep/wake architecture by M1 PAM VU0453595 across
young and aged rodents and nonhuman primates. Neuropsychopharmacology 45: 2219–
2228.
19. Kikuchi T, Okamura T, Arai T, Obata T, Fukushi K, Irie T, Shiraishi T (2010): Use of a
novel radiometric method to assess the inhibitory effect of donepezil on
acetylcholinesterase activity in minimally diluted tissue samples. British Journal of
Pharmacology 159: 1732–1742.
20. Asai M, Fujikawa A, Noda A, Miyoshi S, Matsuoka N, Nishimura S (2009): Donepezil- and
scopolamine-induced rCMRglu changes assessed by PET in conscious rhesus monkeys.
Annals of Nuclear Medicine 23: 877–882.
21. Shiraishi T, Kikuchi T, Fukushi K, Shinotoh H, Nagatsuka SI, Tanaka N, et al. (2005):
Estimation of plasma IC50 of donepezil hydrochloride for brain acetylcholinesterase
inhibition in monkey using N-[11C] methylpiperidin-4-yl acetate ([11C]MP4A) and PET.
Neuropsychopharmacology 30: 2154–2161.
22. Nishiyama S, Tsukada H, Sato K, Kakiuchi T, Ohba H, Harada N, Takahashi K (2001):
Evaluation of PET ligands (+) N-[11C]ethyl-3-piperidyl benzilate and (+) N-[11C]propyl-3-
piperidyl benzilate for muscarinic cholinergic receptors: A PET study with microdialysis in
comparison with (+)N-[11C]methyl-3-piperidyl benzilate in the conscious mo. Synapse 40:
159–169.
23. Tsukada H, Nishiyama S, Ohba H, Sato K, Harada N, Kakiuchi T (2001): Cholinergic
neuronal modulations affect striatal dopamine transporter activity: PET studies in the
conscious monkey brain. Synapse 42: 193–195.
24. Tsukada H, Sato K, Kakiuchi T, Nishiyama S (2000): Age-related impairment of coupling
mechanism between neuronal activation and functional cerebral blood flow response was
restored by cholinesterase inhibition: PET study with microdialysis in the awake monkey
brain. Brain Research 857: 158–164.
25. Tsukada H, Kakiuchi T, Ando I, Ouchi Y (1997): Functional activation of cerebral blood
flow abolished by scopolamine is reversed by cognitive enhancers associated with
cholinesterase inhibition: A positron emission tomography study in unanesthetized
monkeys. Journal of Pharmacology and Experimental Therapeutics 281: 1408–1414.
| 2021 | Dose-dependent dissociation of pro-cognitive effects of donepezil on attention and cognitive flexibility in rhesus monkeys | 10.1101/2021.08.09.455743 | [
"Hassani Seyed A.",
"Lendor Sofia",
"Neumann Adam",
"Roy Kanchan Sinha",
"Boroujeni Kianoush Banaie",
"Hoffman Kari L.",
"Pawliszyn Janusz",
"Womelsdorf Thilo"
] | creative-commons |
P a g e 1 | 60
1
The Nerve Growth Factor IB-like Receptor Nurr1 (NR4A2) Recruits CoREST
2
Transcription Repressor Complexes to Silence HIV Following Proviral Reactivation
3
in Microglial Cells
4
5
Fengchun Ye*, David Alvarez-Carbonell, Kien Nguyen, Saba Valadkhan, Konstantin
6
Leskov, Yoelvis Garcia-Mesa, Sheetal Sreeram, and Jonathan Karn1
7
8
1Department of Molecular Biology and Microbiology. Case Western Reserve University,
9
Cleveland, Ohio, United States of America
10
11
12
*Corresponding author
13
E-mail: fxy63@case.edu
14
15
16
Key word: HIV, silencing, microglia, Nurr1, CoREST
17
18
Running title: Nurr1 mediates HIV latency in microglial cells
19
P a g e 2 | 60
20
ABSTRACT
21
Human immune deficiency virus (HIV) infection of microglial cells in the brain leads
22
to chronic neuroinflammation, which is antecedent to the development of HIV-associated
23
neurocognitive disorders (HAND) in the majority of patients. Productively HIV infected
24
microglia release multiple neurotoxins including proinflammatory cytokines and HIV
25
proteins such as envelope glycoprotein (gp120) and transactivator of transcription (Tat).
26
However, powerful counteracting silencing mechanisms in microglial cells result in the
27
rapid shutdown of HIV expression to limit neuronal damage. Here we investigated
28
whether the Nerve Growth Factor IB-like nuclear receptor Nurr1 (NR4A2), which is a
29
repressor of inflammation in the brain, acts to directly restrict HIV expression. HIV
30
silencing was substantially enhanced by Nurr1 agonists in both immortalized human
31
microglial cells (hµglia) and induced pluripotent stem cells (iPSC)-derived human
32
microglial cells (iMG). Overexpression of Nurr1 led to viral suppression, whereas by
33
contrast, knock down (KD) of endogenous Nurr1 blocked HIV silencing. Chromatin
34
immunoprecipitation (ChIP) assays showed that Nurr1 mediates recruitment of the
35
CoREST/HDAC1/G9a/EZH2 transcription repressor complex to HIV promoter resulting in
36
epigenetic silencing of active HIV. Transcriptomic studies demonstrated that in addition
37
to repressing HIV transcription, Nurr1 also downregulated numerous cellular genes
38
involved in inflammation, cell cycle, and metabolism, thus promoting HIV latency and
39
microglial homoeostasis. Thus, Nurr1 plays a pivotal role in modulating the cycles of
40
proviral reactivation by cytokines and potentiating the proviral transcriptional shutdown.
41
These data highlight the therapeutic potential of Nurr1 agonists for inducing HIV silencing
P a g e 3 | 60
42
and microglial homeostasis and amelioration of the neuroinflammation associated with
43
HAND.
44
AUTHOR SUMMARY
45
HIV enters the brain almost immediately after infection where it infects perivascular
46
macrophages, microglia and, to a less extent, astrocytes. In previous work using an
47
immortalized human microglial cell model, we observed that integrated HIV constantly
48
underwent cycles of reactivation and subsequent silencing. In the present study, we found
49
that the Nurr1 nuclear receptor is a key mediator of HIV silencing. The functional
50
activation of Nurr1 by specific agonists, or the over expression of Nurr1, resulted in rapid
51
silencing of activated HIV in microglial cells. Global gene expression analysis confirmed
52
that Nurr1 not only repressed HIV expression but also regulated numerous genes
53
involved in microglial homeostasis and inflammation. Thus, Nurr1 is pivotal for HIV
54
silencing and repression of inflammation in the brain and is a promising therapeutic target
55
for treatment of HAND.
56
P a g e 4 | 60
57
INTRODUCTION
58
Human immune deficiency virus (HIV) invades the brain soon after primary
59
infection [1]. The virus infects astrocytes, perivascular macrophages, and microglial cells,
60
but not neurons [2, 3]. However, because microglial cells are much longer-lived than
61
astrocytes and perivascular macrophages and can support productive HIV replication,
62
they are mostly likely to be the main cellular reservoir of HIV in the brain [4, 5]. In later
63
stages of HIV infection, many infected patients develop HIV-associated neurocognitive
64
disorders (HAND) [6]. Although combination antiretroviral therapy (cART) dramatically
65
lowers the levels of HIV RNA in the brain [7, 8], it does not reduce the incidence of HAND
66
[9, 10]. Initial studies indicated, paradoxically, that HAND did not correlate with the
67
number of HIV-infected cells or viral antigens in the central nervous system (CNS) [11,
68
12], but instead correlates strongly with systemic inflammation and CNS inflammation
69
[13]. However, the early studies neglected both the side effects of anti-HIV drugs on
70
neuronal damage, which could mask the benefits of reduced HIV expression by cART
71
and the impact of HIV latency on the development of HAND.
72
Over the past decade, the intimate relationship between neuroinflammation,
73
neurodegeneration and abnormal activation of microglial cells has been implicated in a
74
wide range of diverse neurological diseases [14-19]. There are compelling reasons to
75
believe that the physiology of microglia also plays a critical role in the development of
76
HAND. Infected macrophages/microglia in the CNS serve as long-lived cellular reservoirs
77
of HIV-1, even in well-suppressed patients receiving ART [20]. Microglia constitute the
78
first barrier of the innate immune response in the brain and become activated and
79
polarized to maintain the integrity of the CNS [21, 22]. In the normal CNS environment,
P a g e 5 | 60
80
healthy neurons provide signals to microglia via secreted and membrane bound factors
81
such as CX3CL1 and neurotransmitters that induce HIV-silencing. By contrast, damaged
82
neurons not only cause activation of the microglia but also induce HIV reactivation [23].
83
Activated microglia secrete exaggerated amounts of neurotoxins such as tumor
84
necrosis factor-alpha (TNF-), nitric oxide, interleukin-6 (IL-6), interleukin-1 beta (IL-1β),
85
reactive free radicals, and matrix metallopeptidases (MMPs) [24, 25]. The production of
86
these cytotoxic factors is augmented by HIV infection [23, 26]. Mounting evidence
87
indicates that HIV proteins such as transactivator of transcription (Tat), negative
88
regulatory factor (Nef), envelope glycoprotein gp120, and viral RNA are not only directly
89
neurotoxic, but also contribute to inflammation in the brain by activating microglial cells
90
[27-35]. On the other hand, some of the inflammatory cytokines such as TNF- strongly
91
induce HIV expression in microglial cells through autocrine signaling, creating cycles of
92
HIV reactivation and chronic inflammation in the brain [36]. It is therefore important to
93
determine the factors responsible for inducing HIV reactivation and inflammation and
94
explore cellular mechanisms that antagonize these factors in order to develop treatment
95
for HAND.
96
A major constraint for studying HIV infection and replication in the brain is the
97
difficulty of obtaining native microglial cells from brain biopsies. We therefore developed
98
a microglial cell model by immortalizing human primary microglial cells with the simian
99
virus large T-antigen (SV40) and the human telomerase reverse transcriptase (hTERT)
100
[37]. The immortalized microglia retain the typical structure and morphology of primary
101
microglial cells, express microglial cell markers, and display microglial cell activities such
102
as migration and phagocytosis [37].
P a g e 6 | 60
103
A unique feature of HIV infection of microglial cells is that the virus is able to quickly
104
establish latency [23, 36-38]. In microglial cells, transcription initiation is primarily
105
regulated by NF-B. In resting microglia (M0 stage), NF-B is sequestered in the
106
cytoplasm [23, 36-38]. However, unlike memory T-cells, P-TEFb is not disrupted,
107
although it is inhibited by CTIP2 [39, 40]. The provirus is also silenced epigenetically
108
through the CoREST and polycomb repressive complex 2 (PRC2) histone
109
methyltransferase machinery [4, 41-44]. Activation of microglia by pro-inflammatory
110
signals, such as TNF-, reversed these molecular restrictions and leads to the
111
reactivation of dormant proviruses and neuropathology.
112
In contrast to T cells where integrated HIV eventually establishes permanent
113
latency until it is activated by cellular signaling events, HIV in microglial cells undergoes
114
cycles of spontaneous reactivation and subsequent silencing [36]. For example, using a
115
co-culture of (iPSC)-derived human microglial cells (iMG) that were infected with HIV and
116
neurons, we demonstrated that HIV expression in iMG was repressed when co-cultured
117
with healthy neurons but induced when co-cultured with damaged neurons [23]. The
118
dynamics of spontaneous reactivation of latent HIV and subsequent silencing of active
119
HIV constantly typically generates two populations in culture: the GFP- population with
120
transient latent HIV, and the GFP+ population undergoing active HIV transcription.
121
Spontaneous HIV reactivation in microglial cells could be attenuated by activation of the
122
glucocorticoid receptor (GR) with its ligand dexamethasone (DEXA) [45], which blocked
123
recruitment of NF-B and AP-1 for HIV transactivation [36, 45]. However, since we
124
consistently observed spontaneous reactivation of latent HIV and subsequent silencing
125
of the active HIV in the absence of dexamethasone, and in co-cultures with neurons, we
P a g e 7 | 60
126
reasoned that there exist other cellular factors that promote HIV silencing in microglial
127
cells.
128
In the present study, we examined whether three members of the Nerve Growth
129
Factor IB-like nuclear receptor family, which includes nuclear receptor 77 (Nur77,
130
NR4A1), nuclear receptor related 1 (Nurr1, NR4A2), and neuron-derived receptor 1
131
(Nor1, NR4A3), contribute to HIV silencing in microglial cells. These receptors play
132
complementary roles in neurons and microglia to limit inflammatory responses. In
133
neurons, these receptors act as positive transcriptional regulators that control expression
134
of dopamine transporter and tyrosine hydroxylase for differentiation of dopamine neuron,
135
as well as other key genes involved in neuronal survival and brain development [46-49].
136
By contrast, these nuclear receptors can also act as negative transcriptional regulators in
137
microglia cells and suppress expression of inflammatory cytokines such as TNF- and
138
IL-1β [50]. Because of these combined mechanisms, Nerve Growth Factor IB-like nuclear
139
receptors play a critical role in protection of the brain during neurodegenerative diseases
140
such as Parkinson’s disease and Alzheimer’s disease [51-56].
141
Here we report that Nurr1 plays a pivotal role in silencing active HIV in microglial
142
cells by recruiting the CoREST/HDAC1/G9a/EZH2 transcription repressor complex to HIV
143
promoter. Our data also demonstrate that Nurr1 promotes microglial homoeostasis and
144
suppression of inflammation in the brain.
145
P a g e 8 | 60
146
RESULTS
147
Nurr1 agonists strongly induce HIV silencing in microglial cells
148
To study the role of nuclear receptors in the control of HIV expression in the
149
microglia, we used our immortalized human microglial (hµglia) cells [37], which were
150
infected with a recombinant HIV-1 reporter that carried an EGFP marker for “real-time”
151
monitoring of HIV latency and reactivation (Fig 1A). One representative clone, HC69 [37,
152
45], was used for all experiments described in this study. Under normal culture conditions,
153
most cells were GFP-negative (GFP-) (Fig 1B & C). Exposure of HC69 cells to TNF-
154
(400 pg/ml) for 24 hours (hr), induced GFP expression (GFP+) in over 90% of the cells,
155
demonstrating that majority of the integrated HIV provirus was in a latent state under
156
normal culture conditions. To examine whether the reactivated HIV could revert to
157
latency, we conducted a chase experiment by culturing the activated HC69 cells for 96 hr
158
in fresh medium following TNF- stimulation for 24 hr and washing with PBS. Notably,
159
the numbers of GFP+ cells decreased from 93.1% to 61.4% at the end of the chase
160
experiment, suggesting the existence of an intrinsic cellular mechanism that silences the
161
activated HIV. This substantial decrease of GFP+ expression was unlikely to be caused
162
by GR-mediated HIV silencing [45], because the cells were cultured in the absence of GR
163
ligand glucocorticoid or dexamethasone.
164
To understand the regulatory mechanisms of HIV expression in microglial cells,
165
we had previously undertaken a global screening for HIV silencing cellular factors [57, 58]
166
by using a HIV-infected rat microglial cell model (CHME cells) [37, 59]. The latently
167
infected CHME/HIV cells were superinfected with lentiviral vectors carrying a synthetic
168
shRNA library from Cellecta Inc. (Mountain View, CA) containing a total of 82,500
P a g e 9 | 60
169
shRNAs targeting 15,439 mRNA sequences [60-62].. Cells carrying reactivated
170
proviruses were then purified by sorting and the shRNA sequences were identified by
171
next-generation sequencing and classified by Ingenuity Pathway Analysis (QIAGEN).
172
This powerful new technology, which we have also applied to the identification of latency
173
factors in T-cells and TB-infected myeloid cells [58, 63], has revealed a wide range of
174
factors and pathways critical for maintaining proviral latency in microglial cells. Analysis
175
of the top 25 % “hits” led to our unexpected discovery that members of the nuclear
176
receptors (NRs) families including Thyroid Hormone Receptor-like family members
177
PPARα, PPARβ, PPARγ, and RARβ ranked in the top 5%, the Retinoid X Receptor-like
178
family members RXRα and RXRβ together with the glucocorticoid receptor (GR, NR3C)
179
ranked in the top 15%, and the Nerve Growth Factor IB-like family members NR4A1
180
(Nur77), NR4A2 (Nurr1) and NR4A3 (Nor1) ranked in the top 25%.
181
Agonists of the NR4A nuclear receptor family (Nur77 (NR4A1), Nurr1 (NR4A2),
182
and Nor1 (NR4A3)) have been shown to ameliorate neuron degeneration in animal
183
models [53, 64-67]. To confirm a role for the nuclear receptors in HIV silencing, we first
184
treated spontaneously activated HC69 cells with the Nurr1 agonist 6‐mercaptopurine (6-
185
MP) [68, 69]. As shown in Fig 2A, the frequency of GFP+ cells decreased in a 6-MP dose-
186
dependent manner. Data from Western blot analysis showed that HC69 cells
187
constitutively expressed Nurr1, as well as a very low level of Nor1 (Fig 2B), but Nur77
188
expression in these cells was below the detection limit. Treatment with 6-MP slightly
189
increased expression of Nurr1. Expression of HIV, as measured by the levels of Nef
190
protein, was strongly inhibited in a dose-dependent manner. Notably, as a control for the
191
role of Nurr1 in cellular gene expression, 6-MP also substantially reduced expression of
P a g e 10 | 60
192
MMP2, which is a well-known repression target of Nurr1 and a neurotoxin involved in the
193
development of HAND [70, 71].
194
In addition, the Retinoid X Receptor-like family members also play a critical role in
195
silencing inflammation in the brain [72, 73]. We also screened various agonists of the
196
nuclear receptors for their effect on HIV expression in HC69 microglia (Fig 2C). We
197
induced maximum HIV expression in HC69 cells with high dose (400 pg/ml) TNF- for 24
198
hr, followed by a chase experiment during which the induced cells were cultured in the
199
absence or presence of various agonists, alone or in combination. Consistent with our
200
previous gene manipulation data, the RXRα/β/γ agonist bexarotene (BEX) [74-77]
201
silenced HIV expression on its own, although it was less potent than 6-MP. Interestingly,
202
combinations of 6-MP with DEXA and BEX displayed additive HIV silencing effects,
203
suggesting that they each had distinct mechanisms of action.
204
Nurr1 overexpression enhances HIV silencing
205
To further examine how the nuclear receptors contribute to HIV silencing, we
206
constructed lentiviral vectors expressing N-terminal 3X-FLAG-tagged Nur77, Nurr1, and
207
Nor1 respectively under the control of a CMV promoter. Infection of HC69 cells with the
208
different lentiviruses generated cell lines that stably expressed FLAG-tagged Nur77,
209
Nurr1, Nor1, and empty vector, respectively, as confirmed by RNA-Seq studies (Fig 3A)
210
western blots (Fig 3B).
211
To examine how overexpression of each of these nuclear receptors modulated HIV
212
proviral activation and silencing, we stimulated all four cell lines with high dose (400 pg/ml)
213
TNF- for 24 hr to induce HIV transcription through activation of NF-B [38], followed by
214
a 48 hr chase experiment in which TNF- was removed by washing the cells with PBS
P a g e 11 | 60
215
followed by the addition of media lacking TNF- (Fig 3C). As shown by western blot in
216
Fig 3D, TNF- strongly induced the expression of HIV Nef protein, which we used as a
217
marker of HIV reactivation, in all cell lines at 24 hr. Notably, Nef expression decreased in
218
all four cell lines 48 hr after TNF- withdrawal. However, the reduction in Nef expression
219
was much more pronounced in HC69 cells that express 3X-FLAG-Nurr1, suggesting that
220
overexpression of Nurr1 enhanced silencing of active HIV in HC69 cells.
221
We rigorously confirmed these findings using the RNA-Seq data (Fig 3E) to
222
measure the fluctuations in both HIV and Nurr1 expression. In the Nurr1 overexpressing
223
cells, even in unstimulated conditions, the level of HIV proviral expression was strongly
224
reduced. Following stimulation with either a low dose (20 pg/ml) or high dose (400 pg/ml)
225
TNF-α, both vector-infected and Nurr1 overexpressing cells showed an increase in
226
proviral expression. While the level of HIV expression was similar between control cells
227
(vector-infected) and Nurr1-overexpressing cells after high dose TNF- stimulation, Nurr1
228
overexpressing cells had much lower proviral expression level after low dose TNF-
229
stimulation (Fig 3E). The level of HIV mRNA after withdrawal of high dose TNF- was
230
three times lower in Nurr1 overexpressing cells than in vector-infected cells (Fig 3E),
231
strongly suggesting that overexpression of Nurr1 enhanced silencing of active HIV in
232
HC69 cells.
233
Nurr1 knockdown blocks HIV silencing
234
As a complementary approach we performed shRNA-mediated knock down (KD)
235
of endogenous Nurr1 in HC69 cells. Cell lines that stably expressed Nurr1-specific or
236
control shRNA were verified for effective Nurr1 KD by RNA-Seq analyses (Fig 4A)
237
Following the protocol described in Fig 4B, control and KD cells were activated with a
P a g e 12 | 60
238
high dose (400 pg/ml) TNF- for 24 hr, followed by a 72 hr chase. Western blot analyses
239
confirmed the Nurr1 knock down efficiency (Fig 4C). The blots also showed that HIV Nef
240
protein, which is a measure of HIV transcription, was strongly induced at 24 hr post TNF-
241
stimulation in both the control and the Nurr1 KD cells. However, after the chase, Nef levels
242
decreased significantly in the control cells but remained high in Nurr1 KD cells (Fig 4C).
243
Similar results were obtained using flow cytometry (Fig 4D). Compared to cells
244
expressing control shRNA with 10.5% GFP+ cells, the Nurr1 KD cells displayed 58.8%
245
GFP+ cells even before TNF- stimulation, which most likely resulted from failure of
246
silencing spontaneously reactivated HIV in these cells due to Nurr1 depletion (Fig 4D).
247
As expected, after exposure to high dose TNF- for 24 hr, both the control and Nurr1 KD
248
cell lines expressed equally high levels of GFP expression (Fig 4D), displaying 86.3%
249
and 91.2% GFP+ cells respectively. However, 72 hrs after TNF- withdrawal, GFP
250
expression decreased significantly in cells expressing the control shRNA (47.2% GFP+)
251
but remained high (74.6% GFP+) in the Nurr1 KD cells (Fig 4D). Finally, the overall mRNA
252
level of the HIV measured by RNA-Seq was about 1.7 times higher in Nurr1 KD at the
253
end of the chase experiment (Fig 4E).
254
Thus, both the overexpression and the reciprocal KD experiments confirmed an
255
essential role of Nurr1 in the silencing HIV in microglial cells.
256
Nurr1 drives activated microglial cells towards homeostasis
257
Our RNA-Seq data also provided important insights into the cellular pathways that
258
were impacted by Nurr-1 over- and under-expression. We focused our attention on the
259
changes in cellular transcriptome during the chase step following TNF- induction since,
260
as described above, this is the stage where Nurr1 has the greatest impact on HIV gene
P a g e 13 | 60
261
expression. As shown by the differential gene expression curves in Fig 5A, a small subset
262
of genes are selectively up and down regulated during the chase. A larger number of
263
genes were differentially expressed in Nurr1 overexpressing cells compared to control
264
cells (Fig 5A & S1 Fig.). Pathways that showed the most statistically significant changes
265
in response to Nurr1 overexpression included the downregulation of key pathways with
266
critical roles in cellular proliferation and metabolism including: MYC, E2F and MTORC
267
signaling and G2M checkpoint (Fig 5B). By contrast, KD of Nurr1 by shRNA did not
268
selectively activate any major signaling pathways.
269
It is important to note that Nurr1 overexpression did not significantly interfere with
270
the TNF- signaling pathway during any step of these experiments (Fig 5B), suggesting
271
that the cellular proliferation pathways we have identified are directly regulated by Nurr1.
272
To further address this issue and determine whether Nurr1 simply accelerated the
273
reversal of the normal microglial response to TNF- stimulation during the chase, or if it
274
regulated a distinct set of genes and pathways, we performed a gene trajectory analysis
275
(Fig 6A, S2 Fig).
276
For the trajectory analysis we included RNA-Seq data from cells that were treated
277
with the low dose of TNF- (20 pg/ml), to simulate a sub-optimal activation signal. A
278
pseudo-trajectory was defined containing three steps: Step 1 defines the changes in gene
279
expression following stimulation with low dose TNF- compared to untreated cells. Step
280
2 defines additional changes after stimulation with high dose TNF- compared to cells
281
treated with low dose TNF-. Step 3 defines the gene expression changes following the
282
chase step compared to cells treated with high dose TNF- (S2 Fig). For each of these
283
steps we calculated whether the expressed protein-coding genes were either upregulated
P a g e 14 | 60
284
(designated as “u”), downregulated (designated as “d”) or did not show differential
285
expression in a statistically significant manner (designated as “n”). Genes that showed
286
similar patterns of changes during each step were placed in the same category and
287
named according to their pattern of change during these treatment steps. For example,
288
those that did not show a change after low dose TNF- stimulation (thus marked as n for
289
Step 1), but were downregulated after high dose TNF- treatment compared to cells
290
treated with low dose TNF- (marked as d for Step 2), and showed upregulation during
291
the chase study compared to cells treated with high dose TNF- (marked as u for step
292
3), were therefore designated as ndu.
293
Most genes did not show any change in their expression following the above
294
treatments (designated as the “nnn” group) in both control (vector) and Nurr1-
295
overexpressing cells (Fig 6A, S2 Fig), and as expected, control cells had higher numbers
296
of nnn group genes than Nurr1 overexpressing cells.
297
Among those genes that showed an expression change in Nurr1 overexpressing
298
cells, the majority belonged to genes that were not differentially expressed after either a
299
low or high dose TNF- treatment and exclusively changed their expression profiles
300
during the chase step (i.e., nnu and nnd trajectories, Fig 6A, S2 Fig). We also noted that
301
the number of genes in these two trajectories were markedly higher in Nurr1
302
overexpressing cells compared to control cells (i.e., over 800 and 1400 genes for nnu and
303
nnd trajectories, respectively) while the number of genes in other trajectories with the
304
exception of nnn differed by less than 100 genes (Fig 6A, S2 Fig).
305
We next confirmed that the genes showing the nnu and nnd trajectories in Nurr1
306
overexpressing cells were derived from a subset of the nnn trajectory genes in the control
P a g e 15 | 60
307
group and were therefore exclusively altered during the chase step in Nurr1
308
overexpressing cells. To further characterize the Nurr1-specific changes in expression
309
patterns, we used the list of genes in each of the trajectories identified in Nurr1
310
overexpressing cells and defined their trajectory in control cells (S3 Fig). This analysis
311
showed that over 1400 and ~800 of the genes that fall into the nnd or nnu trajectories in
312
Nurr1 overexpressing cells, respectively, have the nnn trajectory in control cells (S3 Fig).
313
Thus, the main transcriptomic outcome of Nurr1 overexpression compared to control cells
314
is the induction of changes in expression of a group of genes exclusively during the chase
315
step. Importantly, this group of genes are not differentially expressed in the control cells
316
during either of the three steps of these studies, nor during the TNF- stimulation steps
317
in Nurr1 overexpressing cells and therefore, the action of Nurr1 during the chase step
318
does not correspond to a reversal of the TNF--induced changes.
319
In order to define the functional impact of this Nurr1-specific set of genes, we
320
performed pathway analysis on the subset of genes that had either nnu or nnd trajectories
321
in Nurr1 overexpressing cells, and a nnn trajectory in control cells (Fig 6B). Strikingly,
322
these Nurr1-induced changes in gene expression during the chase step once again
323
highlighted the downregulation of several key proliferative pathways, including: MYC, E2F
324
and MTORC signaling, G2M checkpoint regulation, metabolic pathways such as oxidative
325
phosphorylation, and inflammatory pathways such as IFN- and IFN-γ response
326
pathways (Fig 6B).
327
Heat maps of the differentially expressed genes further emphasized that the vast
328
majority of genes in each pathway were downregulated in Nurr1 overexpressing cells.
329
For example, among 69 and 60 represented MYC and E2F target genes, 66 and 54 were
P a g e 16 | 60
330
downregulated in Nurr1 overexpressing cells, respectively (S4 Fig). Finally, another
331
compelling way of visualizing these results is to apply the pattern of expression of the
332
Nurr1-specific genes to the KEGG cell cycle pathway (S5 Fig). The strong downregulation
333
by Nurr1 at multiple steps in the cell cycle control pathway is immediately obvious.
334
Finally, we note as another measure of the specificity of the Nurr1 pathway, that
335
the most enriched transcription factor binding motifs in proximity of the promoters of
336
differentially expressed genes following TNF- stimulation all display promoter motifs
337
consistent with TNF- activation (S6 Fig).
338
Thus, the main impact of Nurr1 on the overall cellular response to inflammatory
339
cytokines, in this case TNF-, was to accelerate the cellular return to homeostasis by
340
shutting down pathways involved in inflammation and microglial activation. While these
341
anti-inflammatory, pro-homeostasis effects could indirectly lead to HIV proviral
342
transcriptional shutdown, the enhanced downregulation of HIV expression in Nurr1
343
overexpressing cells, even under basal untreated conditions (Fig 3D), suggests that in
344
addition to its pro-homeostasis effects, Nurr1 may also directly regulate the expression of
345
the HIV provirus.
346
Nurr1 promotes the recruitment of the CoREST/HDAC1/G9a/EZH2 repressor
347
complex to the HIV promoter
348
Previous studies demonstrated that Nurr1 interacted with the corepressor 1 of
349
REST (CoREST) repressor complex [50, 78]. The CoREST complex is comprised of
350
multiple components including CoREST, repressor element-1 silencing transcription
351
factor (REST), HDAC1/2, euchromatic histone lysine N-methyltransferase 2 (EHMT2),
352
also known as G9a, lysine (K)-specific demethylase 1A (KDM1A), and enhancer of zeste
P a g e 17 | 60
353
2 polycomb repressive complex 2 subunit (EZH2) [79, 80]. In microglial cells and
354
astrocytes, after stimulation with lipopolysaccharide (LPS), Nurr1 promoted recruitment
355
of this complex to the promoters of inflammatory genes such as IL-1β leading to
356
epigenetic silencing. We postulated that the Nurr1/CoREST repression pathway might
357
therefore also lead to direct regulation of HIV silencing as illustrated in Fig 7A.
358
To test this hypothesis, we first conducted co-immunoprecipitation (Co-IP) assays
359
to confirm the association of Nurr1 with the CoREST repressor complex in HC69 cells
360
(S7 Fig.). HC69-3X-FLAG-vector and HC69-3X-FLAG-Nurr1 cells were treated with and
361
without a high dose of TNF- for 4 hr (400 pg/ml) or 24 hr. After 24 hr TNF- treatment
362
the cells were chased in the absence of TNF- for a further 24 hr. Total protein lysates
363
from the differently treated cells were immunoprecipitated using a mouse monoclonal
364
anti-FLAG antibody conjugated to magnetic beads. The anti-FLAG beads pulled down
365
not only FLAG-tagged Nurr1 but also CoREST, HDAC1, G9a, and EZH2 from the HC69-
366
3X-FLAG-Nurr1 cell lysates, demonstrating that in the microglial cells Nurr1 bound
367
directly to the CoREST repressor complex. Notably, the amount of CoREST associated
368
with Nurr1 increased after the cells were stimulated with TNF-. In contrast, the amounts
369
of G9a and EZH2 proteins associated with Nurr1 decreased at 4 hr post-TNF-
370
stimulation but rebounded at 24 hr post-TNF- stimulation. Together, these results
371
suggested that the Nurr1/CoREST/HDAC1/G9a/EZH2 complex were most likely
372
dissociated from each other during early time points of TNF- stimulation but were
373
reassembled at later time points.
374
We next conducted ChIP-Seq experiments to demonstrate recruitment of the
375
CoREST repressor complex to the activated HIV promoter in microglial cells. As shown
P a g e 18 | 60
376
in Fig 7B, CoREST, HDAC1, G9a, and EZH2 were all detected on the HIV provirus and
377
were enriched near the promoter region following TNF- activation. However, the
378
recruitment kinetics of each component was distinct, with CoREST being recruited to HIV
379
promoter during early time points of TNF- exposure and HDAC1, G9a, and EZH2 being
380
recruited at late time points. These results are consistent with the Co-IP results shown in
381
S7 Fig. Specifically, the levels of CoREST at the HIV promoter peaked at 4 hr post-TNF-
382
stimulation and decreased at 24 hr post-treatment, whereas the levels of G9a, EZH2, and
383
HDAC1 at the HIV promoter decreased at 4 hr post-TNF- stimulation when compared
384
to un-treated cells. However, these epigenetic silencers returned to HIV promoter in a
385
much more robust manner at 24 hr post-stimulation.
386
To provide direct evidence that Nurr1 mediates the recruitment of the
387
CoREST/HDAC1/G9a/EZH2 complex to HIV promoter, we treated HC69 cells expressing
388
control shRNA and Nurr1 shRNA with high dose TNF-, followed by a 24 hr chase. We
389
then conducted additional ChIP experiments and measured the ChIP products by
390
quantitative PCR (qPCR). As shown in Fig 7C, CoREST was strongly recruited to HIV
391
promoter at 4 hr post TNF- stimulation in HC69 cells expressing control shRNA,
392
however, its recruitment was substantially inhibited in Nurr1 KD cells. Similarly, G9a level
393
in HIV promoter peaked at 24 hr post TNF- stimulation in HC69 cells expressing control
394
shRNA but its recruitment was also reduced in Nurr1 KD cells. Taken together, these
395
results clearly demonstrated a pivotal role for Nurr1 in mediating recruitment of the
396
CoREST/HDAC1/G9a/EZH2 repressor complex to the promoter of active HIV for
397
epigenetic silencing consistent with the model shown in Fig 7A.
398
The CoREST/HDAC1/G9a/EZH2 repressor complex silences HIV in microglial cells
P a g e 19 | 60
399
To further investigate how the CoREST/HDAC1/G9a/EZH2 complex contributes to
400
HIV silencing, we treated HC69 cells with high dose (400 pg/ml) TNF- for 24 hr followed
401
by a chase in the absence or presence of epigenetic inhibitors that target the CoREST
402
complex, specifically: HDAC inhibitor suberoylanilide hydroxamic acid (SAHA), G9a
403
inhibitor UNC0638, and EZH2 inhibitor GSK343. The numbers of GFP+ cells dropped
404
from 88.4% to 67.03% at 48 hr after TNF- withdrawal when cells were cultured in the
405
absence of the inhibitors (Fig 8A). However, in the presence of SAHA, UNC0638, or
406
GSK343, the numbers of GFP+ cells remained higher (i.e., 77.8%, 85.5%, and 84.7%
407
respectively), indicating that functional inhibition of these epigenetic silencers prevented
408
active HIV from reverting to latency.
409
To confirm the role of these epigenetic silencers, we generated HC69 cell lines
410
stably expressing CoREST-specific shRNA or CRISPR/Cas9/guide RNA (gRNA) for G9a
411
or EZH2. We confirmed successful KD or knock out (KO) of these proteins in these cell
412
lines by Western blot analysis (Fig 8B & C). The genetically modified cells were activated
413
with a high dose of TNF- (400 pg/ml) for 24 hr, followed by culturing the cells in the
414
absence of TNF- for 48 hr and measurement of GFP expression. CoREST KD
415
substantially increased GFP expression (80.1% GFP+ vs. 25.8% in control cell) even
416
without TNF- stimulation (Fig 8D). Stimulation with high-dose TNF- for 24 hr resulted
417
in 94.1% and 84.9% GFP+ cells in CoREST KD and control cells respectively. However,
418
after TNF- withdrawal and subsequent culture for 48 h, the numbers of GFP+ cells
419
decreased significantly in cells expressing control shRNA (67.7%) but remained high in
420
CoREST KD cells (91.3%), confirming that CoREST was crucial for the silencing of active
P a g e 20 | 60
421
HIV in microglial cells. Similar results were also seen with the G9a and EZH2 KO cell
422
lines (Fig 8E).
423
Therefore, both the ChIP experiments and gene knockout results demonstrate a
424
pivotal role for the CoREST/HDAC1/G9a/EZH2 transcription repressor complex in
425
silencing active HIV in microglial cells.
426
Nurr1 regulates HIV in iPSC-derived microglial cells
427
Finally, to confirm that Nurr1 is also critical for the silencing of HIV in primary
428
microglial cells, we infected iPSC-derived human microglial cells (iMG) with the same HIV
429
reporter virus described earlier (Fig 1A). About 50% of the iMG became GFP+ two days
430
after HIV infection (Fig 9A). We then treated the infected iMG with 6-MP and another
431
Nurr1 agonist, amodiaquine (AQ) [56, 67], for four days. Both 6-MP and AQ decreased
432
the number of GFP+ cells in a dose-dependent manner (Fig 9B & C) and lowered the
433
levels of HIV un-spliced transcripts (Fig 9D). Both agonists also dose-dependently
434
reduced MMP2 mRNA in iMG (Fig 9E). Collectively, results from both hµglia and iMG
435
strongly suggested an important role for Nurr1 in HIV silencing in microglial cells.
436
P a g e 21 | 60
437
DISCUSSION
438
Epigenetic control of HIV latency in microglial cells
439
Microglial cells are one of the major cellular reservoirs of HIV in the central nervous
440
system (CNS) [4, 5]. These long-lived cells contribute to increased neuroinflammation
441
and oxidative stress [4, 81], and development of HAND by secreting a variety of
442
neurotoxins as well as harmful HIV proteins such as gP120, Tat, Rev, etc. [82, 83].
443
Eradication or complete silencing of HIV-infected microglial cells is therefore crucial not
444
only for an HIV cure, but also to prevent the development of HAND, which affects the
445
majority of HIV infected individuals.
446
Previous studies involving HIV-1 infection of transformed cell lines suggested that
447
epigenetic regulation plays a major role in the establishment and persistence of HIV
448
latency in astrocytes and microglial cells [84, 85]. The cellular COUP transcription factor
449
(COUP-TF) interacting protein (CTIP2) forms a large transcriptional repressor complex
450
with epigenetic silences including the histone deacetylases HDAC1/2, the histone
451
methyltransferases SUV39H1 and SET1, the lysine(K)-specific demethylase KDM1, and
452
heterochromatin protein1 (HP1) [86, 87]. Recruitment of this complex to HIV-1 promoter
453
leads to proviral genome silencing due to reduced histone acetylation and increased
454
levels of histone 3 tri-methylations at lysine 9 (H3K9me3) [86-89]. At the same time,
455
CTIP2 forms another complex with CDK9, Cyclin T1, HEXIM1, 7SK snRNA, and high
456
mobility group AT-hook 1 (HMGA1), which is also recruited to HIV-1 promoter [87, 89]. In
457
the absence of HIV-1 Tat, this complex with inactive pTEFb further supports HIV-1 latency
458
by preventing elongation of RNA polymerase II for active transcription [90]. Nevertheless,
P a g e 22 | 60
459
it remains unknown if these mechanisms also apply to HIV-infected primary microglial
460
cells as transformed cells often behave quite differently.
461
In a previous study [37], we demonstrated that autocrine inflammatory cytokines
462
such as TNF-α were major drivers for spontaneous HIV reactivation in microglial cells,
463
and activation of GR with its specific ligands such as dexamethasone antagonized the
464
effects of cytokines on HIV reactivation [45]. However, we observed that the reactivated
465
HIV was subsequently silenced in microglial cells in the absence of dexamethasone,
466
suggesting the existence of additional HIV silencing mechanisms.
467
Silencing of HIV by Nurr1 and CoREST
468
In the present study, we identified the nuclear receptor Nurr1 as a key HIV silencing
469
factor. Overexpression of Nurr1 had little effect on preventing reactivation of latent HIV
470
but strongly enhanced silencing of active HIV after TNF- stimulation and subsequent
471
withdrawal. Inversely, KD of endogenous Nurr1 in HC69 cells inhibited silencing of active
472
HIV after TNF- withdrawal. Thus, results from both overexpression and KD experiments
473
unequivocally demonstrated a pivotal role of Nurr1 in silencing active HIV.
474
Mechanistically,
we
demonstrated
that
Nurr1
interacted
with
the
475
CoREST/HDAC1/G9a/EZH2 repressor complex as reported previously for cellular early
476
response genes [50]. Nurr1 promoted the recruitment of CoREST complexes to the HIV
477
promoter following TNF- stimulation and subsequent withdrawal.
478
These epigenetic silencers likely silence active HIV by promoting histone de-
479
acetylation and repressive di- or tri-methylations. Consistent with this hypothesis,
480
functional inhibition with specific inhibitors or expressional KD or KO of each component
481
of the repressor complex including CoREST, G9a, and EZH2 strongly inhibited HIV
P a g e 23 | 60
482
silencing. Data from RNA-Seq analysis indicated that Nurr1 might also utilize this “post-
483
TNF- stimulation” epigenetic silencing mechanism to repress many host genes.
484
Regulation of HIV latency in microglial cells by nuclear receptors
485
Nuclear receptors are special transcription factors that turn on or turn off
486
expression of target genes upon specific ligand binding [91]. Accumulating evidence
487
suggest that nuclear receptors play an important role in regulating HIV expression. For
488
example, estrogen receptor (ER) and GR have been found to promote HIV latency in T
489
cells and microglial cells respectively [45, 63]. In this study, by using both immortalized
490
and iPSC-derived human microglial cells, we provided comprehensive data to
491
demonstrate that Nurr1 promoted HIV latency by silencing active HIV. In contrast, we did
492
not see significant effects of Nur77 and Nor1 overexpression on HIV when HC69 cells
493
were stimulated with TNF-. However, we have not examined possible effects of these
494
nuclear receptors on HIV when microglial cells are activated through other signaling
495
pathways such as the “Toll-like” receptor signaling pathway [38]. In addition, it is well
496
known that the different Nerve Growth Factor IB-like nuclear receptors interact with each
497
other or with other nuclear receptors such as GR and the retinoid X receptors (RXR) [92,
498
93]. Therefore, in future experiments, we plan to investigate how Nur77, Nurr1, and Nor1
499
impact HIV expression in microglial cells in response to different stimuli and whether they
500
exert any synergistic effect on HIV expression between themselves or with other
501
interacting nuclear receptors.
502
Role of Nurr1 in maintaining brain homeostasis
503
The roles of Nerve Growth Factor IB-like nuclear receptors in brain development
504
and homeostasis are well established. Both Nor1 and Nurr1 are essential for
P a g e 24 | 60
505
differentiation and survival of dopaminergic neurons [46-49]. Nurr1 deficiency in
506
embryonic ventral midbrain cells results in their failure to migrate and innervation of their
507
striatal target areas [94, 95]. Nurr1 deficiency or reduced expression due to mutations in
508
adults is a major contributing factor in the pathogenesis of Parkinson’s disease [96]. Nurr1
509
is also expressed in non-neuronal cells including monocytes, macrophages, microglia,
510
and astrocytes. Its expression is reduced in the peripheral blood lymphocytes (PBL) of
511
patients with Parkinson’s disease compared with healthy controls [97]. Lower levels of
512
Nurr1 in the brain and blood represents increased risks of Parkinson’s disease and other
513
neurodegenerative diseases in adults [97].
514
Nurr1 protects dopaminergic neurons from inflammation-induced neurotoxicity
515
through the inhibition of pro-inflammatory mediator expression in microglia and astrocytes
516
by recruiting CoREST corepressor complexes to NF-B target genes [50, 98]. A reduction
517
of Nurr1 expression in neurons does not affect their death but enhances expression of
518
inflammatory mediators, and the survival rate of neurons decreases in response to
519
inflammatory stimuli in the Nurr1 deficiency condition [50]. Multiple studies reported that
520
activation of Nurr1 reduces inflammation, protects neurons, and decreases Parkinson’s
521
disease related symptoms [53, 65, 67, 99].
522
Although the pathogenesis of Parkinson’s disease and other types of
523
neurodegeneration remains obscure, increasing evidence suggests that inflammatory
524
responses are responsible for the progression of most neurodegenerative diseases [100].
525
These responses include accumulation of inflammatory mediators, such as inflammatory
526
cytokines and proteases in the substantia nigra and the striatum, as well as activation of
527
the microglia [101], which are also common features of HAND [4, 102].
P a g e 25 | 60
528
Anti-inflammatory role for Nurr1 in HIV-infected microglial cells
529
Little is known on how HIV infection impacts expression or functionality of Nurr1
530
and other nuclear receptors in the brain. Microglial activation is triggered by a series of
531
neurochemical mediators such as IFN-, inducible nitric oxide synthase (iNOS), IL-1β,
532
and TNF- [103-106]. HIV infection of the brain likely further increases the levels of these
533
mediators. Interestingly, data from our RNA-Seq experiments reveal that Nurr1
534
overexpression pushed the activated microglial cells towards homeostasis following TNF-
535
stimulation and subsequent withdrawal by repressing NF-B signaling pathway and
536
genes involved in cellular activity and IFN- and INF- responses. Thus, in addition to
537
silencing HIV, Nurr1 apparently plays a crucial role in suppression of microglia activation.
538
This finding is consistent with a recent report that glycolysis downregulation is a hallmark
539
of HIV‐1 latency in microglial cells [107].
540
Further studies are warranted to determine the expression levels of these nuclear
541
receptors in HIV patients and investigate whether their deficiency or malfunction
542
contributes to development of HAND. Interestingly, multiple Nurr1 agonists exhibit strong
543
therapeutic effects and potentials for Parkinson’s disease in pre-clinical animal study and
544
human trials [108]. In this study, we tested the Nurr1 agonists 6-MP and AQ. Both agents
545
strongly inhibited expression of HIV and the neurotoxin MMP2 in HC69 cells and iMGs.
546
In future studies, it would be of great interest to test additional agonists, particularly those
547
new generations of Nurr1 agonists currently on pre-clinical and human trials, for their anti-
548
HIV activity and eventual application in the clinic for treatment of HAND.
549
P a g e 26 | 60
550
MATERIALS & METHODS
551
Chemicals and reagents
552
TNF- (Invitrogen, Cat. #PHC3015) was used to induce HIV-1 reactivation in
553
microglial cells. Nor1 and Nurr1 agonists 6-mercaptopurine (6-MP) (Millipore-Sigma,
554
Cat#38171) and amodiaquine (AQ) (Millipore-Sigma, Cat#SMB00947) were used to
555
activate the Nerve Growth Factor IB-like nuclear receptors. GSK343 (Sigma Aldrich, Cat#
556
SML0766), UNC0638 (Sigma Aldrich, Cat#U4885), and suberoylanilide hydroxamic acid
557
(SAHA, Millipore-Sigma, Cat#SML0061) were used to examine the effects of EZH2, H9a,
558
and HDAC1/2 on HIV silencing respectively.
559
Numerous antibodies were used for Western blot analysis, co-immunoprecipitation
560
(Co-IP), and chromatin immunoprecipitation (ChIP) assays, including a mouse
561
monoclonal anti-FLAG M2 antibody (Sigma, Cat# F1804), a rabbit polyclonal anti-Nurr1
562
antibody (Sata Cruz Biotechnology, Cat# sc-991), a mouse monoclonal anti-Nurr1
563
antibody (Santa Cruz Biotechnology, Cat# sc-81345), a mouse monoclonal anti-Nor1
564
antibody (Perseus Proteomics, Cat# PP-H7833-00), a rabbit polyclonal anti-Nur77
565
antibody (Cell Signaling, Cat# 3960S), a rabbit monoclonal anti-MMP2 antibody (Cell
566
Signaling, Cat#40994), a rabbit polyclonal anti-CoREST antibody (EMD Millipore, Cat#
567
07-579), a rabbit polyclonal anti-HDAC1 antibody (Santa Cruz Biotechnology, Cat# sc-
568
7872), a rabbit polyclonal anti-G9a antibody (Cell Signaling, cat#3306S), a rabbit
569
polyclonal anti-EZH2 antibody (Cell Signaling, Cat#5246S), a rabbit polyclonal anti-
570
acetylated histone 3 (Ac-H3) antibody (Cell Signaling, Cat#9677S), a rabbit polyclonal
571
anti-H3K27me3 antibody (EDM Millipore, Cat#07-449), a rabbit monoclonal anti-
572
H3K27me2 antibody (Cell Signaling, Cat#9728S), a mouse monoclonal anti-HIV Nef
P a g e 27 | 60
573
antibody (Abcam, Cat#ab42355), and a mouse monoclonal anti-RNA polymerase II
574
antibody (Abcam, Cat#ab817).
575
Cells and flow cytometry analysis of HIV/GFP expression
576
HIV-1 infected immortalized human microglial (hµglia) HC69 cells were cultured
577
and maintained as described as previously [37]. Induced pluripotent stem cells (iPSC)-
578
derived human microglial cells (iMG) (Tempo Bioscience, Cat#SKU 1001.1) were plated,
579
allowed to differentiate and maintained in culture on plates pre-coated with Matrigel matrix
580
(Corning, Cat#356254) according to the manufacturer’s instructions. The iMG were
581
infected with EFGP HIV-1 reporter virus at 1 to 1 (cell-to-virus moiety), which was
582
produced, purified, and titrated as described previously [37]. Two days after infection, the
583
iMG were treated with and without the Nurr1 agonists 6-MP and AQ for four days. Infected
584
with the same EGFP-reporter HIV-1 virus (Fig 1 A), HIV expression in hµglia and iMG
585
cells was measured and quantified with percentage (%) of GFP+ cells by flow cytometry
586
as described previously [45].
587
Lenti-viral construction and production, and generation of stable cell lines
588
Three lentiviral constructs, pLV[Bxp]-Bsd-CMV>3xFLAG-Nur77, pLV[Bxp]-Bsd-
589
CMV>3xFLAG-Nurr1, and pLV[Bxp]-Bsd-CMV>3xFLAG-Nor1 were generated by
590
inserting the full-length open reading frame (ORF) of human NR4A1 (Nur77), NR4A2
591
(Nurr1), and NR4A3 (Nor1) cDNA fragment into the empty vector pLV[Bxp]-Bsd-
592
CMV>3xFLA immediately downstream of the Kozak sequence (VectorBuilder, vector ID:
593
VB180227-1135bmn, VB180227-1134jht, and VB180227-1136rwc). The inserted cDNA
594
was also “in frame” fused with the coding sequence of the N-terminal 3X-FLAG peptide
595
tag, allowing to generate N-terminal 3xFLAG-tagged proteins. Two lentiviral constructs
P a g e 28 | 60
596
expressing human Nurr1-specific shRNA (5’GGTTCGCACAGACAGTTTAAA3’ and
597
5’ATACGTGTGTTTAGCAAATAA3’), one lentiviral construct expressing human
598
CoREST-specific shRNA (5’CCCAATAATGGCCAGAATAAA3’), and two lentiviral
599
constructs expressing control shRNAs (5’CCTAAGGTTAAGTCGCCCTCG3’ and 5’
600
CAACAAGATGAAGAGCACCA3’) were purchased from VectorBuilder. All lentiviral
601
constructs carried an ampicillin resistance gene for selection in bacteria (E. coli) and a
602
blasticidin resistance gene for selection of stable expression in mammalian cells.
603
Infectious viral particles with each of these lentiviral constructs were produced by co-
604
transfecting 293T cells with packaging plasmid psPAX2 (Addgene, Cat#12260) and Env
605
Vector pCMV-VSVg (Addgene, Cat#138479). HC69 cells stably expressing 3X-FLAG-
606
Nur77, 3X-FLAG-Nurr1, 3X-FLAG-Nor1, empty vector, gene-specific shRNA and control
607
shRNA were generated by infection of the cells with purified lentiviral particles for two
608
days, followed by culturing the cells in the presence of blasticidin at 10 g/ml.
609
To investigate the effects of G9a and EZH2 on HIV silencing, we conducted
610
CRISPR/Cas9 mediated “knocking out” (KO) of these genes in HC69 cells, using a dual
611
CRISPR/Cas9 gRNA lentiviral vector. Two different guide RNAs targeting EZH2
612
(TGAGCTCATTGCGCGGGACT
and
GATCTGGAGGATCACCGAGA)
or
G9a
613
(TTCCCCATGCCCTCGCATCC and GTGGCAGCCCCACGGCTGAA) were cloned into
614
lentiCRISPR v2-Blast plasmid following the protocol described previously [109].
615
LentiCRISPR v2-Blast was a gift from Mohan Babu (Addgene plasmid # 83480). VSV-G
616
pseudotyped viruses expressing CRISPR/Cas9 gRNAs were produced in HEK 293T cells
617
by transfection of lentiCRISPR v2 plasmids together with psPAX2 and pCMV-VSV-G.
618
HC69 cells infected with the EZH2 or G9a KO lentiviruses were cultured in the presence
P a g e 29 | 60
619
of blasticidin (10 g/ml). Successful KO of these genes in HC69 cells were verified by
620
Western blot analysis of EZH2 and G9a proteins in the resulting cell lines.
621
Reverse transcription and quantitative polymerase chain reaction (RT-qPCR)
622
Total RNAs from HC69 or HIV-infected IMG cells with different treatments were
623
isolated by using the RNeasy Plus Mini kit from Qiagen (Cat#74134). The purified total
624
RNAs were converted to first-strand cDNAs by using a reverse transcription kit (Bio-Rad,
625
Cat#1708891). The relative levels of HIV-1 un-spliced transcript and human MMP-2
626
mRNA were measured by qRT-PCR using the primers 5’AGGGACCTGAAAGCGAAAG3’
627
(HIV-1 un-spliced-forward) and 5’AATGATACGGCGACGACCNNNNNNNNNN3’ (HIV-1
628
un-spliced-reverse), and 5’ATAACCTGGATGCCGTCGT-3′ (MMP2 forward) and
629
AGGCACCCTTGAAGAAGTAGC-3′ (MMP2 reverse), respectively. The mRNA level of
630
the housekeeping gene β-actin in each sample was used as reference for normalization,
631
which
was
measured
by
qRT-PCR
using
the
primers
5’-
632
TCCTCTCCCAAGTCCACACAGG-3′ (forward) and 5’-GGGCACGAAGGCTCATCATTC-
633
3′ (reverse). Each qRT-PCR was conducted in triplicates.
634
ChIP and ChIP-seq analyses
635
Standard procedures were followed for all ChIP assays. Briefly, cells were fixed
636
with 1% Formaldehyde for 10 minutes (min) at room temperature, followed by incubation
637
in PBS containing 125 mM glysine for 10 min at room temperature. After two washes with
638
ice-cold PBS, cells were re-suspended and allowed to swell in CE buffer [10 mM Hepes,
639
pH7.9, 60 mM KCl, 1 mM EDTA, 0.5% NP-40, 1 mM DTT] on ice for 10 min. After
640
centrifugation at 2,000 g for 10 min at 4C, nuclei were re-suspended in SDS lysis buffer
641
[50 mM Tris-HCl, 1 mM EDTA, 0.5% SDS] and incubated on ice for 10 min. Sheared
P a g e 30 | 60
642
chromatins were prepared by sonicating the nuclei lysate to generate DNA fragments in
643
the range of 250 to 500 bps. ChIP assays with specific antibodies were carried out in
644
ChIP dilution buffer [16.7 mM Tris-HCl, pH 8.1, 167 mM NaCl, 1.2 mM EDTA, 1.1% Triton
645
X-100, and 0.01% SDS] containing 5 g antibody and 50 ul protein-A/protein-G magnetic
646
beads per reaction at 4C for overnight with rotation, followed by consecutive washes with
647
low salt wash buffer [20mM Tris-HCl, pH8.1, 150 mM NaCl, 1 mM EDTA, 1% Triton X-
648
100, 0.1% SDS], high salt wash buffer [20mM Tris-HCl, pH8.1, 500 mM NaCl, 1 mM
649
EDTA, 1% Triton X-100, 0.1% SDS], and RIPA buffer [20 mM Tris-HCl, pH7.5, 150 mM
650
NaCl, 5 mM EDTA, 0.5% Triton X-100, 0.5% sodium deoxycholate, and 0.1% SDS]. The
651
washed beads were then re-suspended in elution buffer [50 mM Tris-HCl, pH 6.5, 20 mM
652
NaCl, 100 mM NaHCO3, 1 mM EDTA, 1% SDS, 100 g/ml proteinase K] and incubated
653
at 50C for 2h. Supernatants from the beads were collected and used for ChIP DNA
654
purification using Qiagen’s PCR purification kit (Cat#28104). Quantification of input and
655
ChIP DNA corresponding to HIV-1 promoter region was conducted by qPCR using
656
specific primers as reported previously [110].
657
For ChIP-seq analyses, the DNA products from each ChIP assay were first end
658
repaired with end repair enzyme mix (New England Biolabs, Inc., Cat#M6630), then
659
ligated to NEBNext adaptor included in the NEBNext® Ultra™ II DNA Library Prep Kit for
660
Illumina® (Cat#E7645L) according to the manufacturer’s instruction, followed by PCR
661
amplification with a specific pair of bar-coded primers. Next, to enrich HIV-1 specific
662
sequences in the library, DNA samples from all ChIP assays were pooled, denatured at
663
98C for 10 min, and then subjected to hybridization with 50 times excessive amount of
664
biotin-labelled and pre-denatured HIV-1 genomic DNA in hybridization buffer containing
P a g e 31 | 60
665
5XSSC and salmon sperm DNA (100 g/ml) at 65C for 1 h. Fragments hybridizing to
666
biotin-labelled HIV-1 DNA were pulled down by incubating the hybridization reaction with
667
streptavidin-conjugated magnetic beads (ThermoFisher Scientific, Cat#88816) at room
668
temperature for 30 min, followed by three times washes with ion wash buffer and elution
669
in water. The enriched ChIP library DNA was PCR amplified with Ion A and Ion P1
670
primers, and PCR fragments in the range from 300 to 500 bps were purified from agarose
671
gel after electrophoresis and loaded for Ion Torrent sequencing.
672
We aligned the sequence reads to NL4.3-Cd8a-EGFP-Nef+ HIV-1 genome. Raw
673
fastq sequencing data were imported to the public server at usegalaxy.org for analysis
674
[111]. We used FASTX-Toolkit for deconvolution of reads. Read mapping was performed
675
by Bowtie2 tool with default settings using the NL4.3-Cd8a-EGFP-Nef+ HIV-1 as a
676
reference genome [112, 113]. DeepTool2 was used to make graphs for distribution of
677
mapped reads along HIV-1 genome [114].
678
RNA-Seq and data analysis
679
Approximately 2 million GFP-negative cells from each of the cell lines HC69-3X-
680
FLAG-vector, HC69-3X-FLAG-Nor1, HC69-3X-FLAG-Nurr1, HC69-control shRNA, and
681
HC69-Nurr1 shRNA were collected from sorting. The isolated cells were expanded in
682
DMEM culture media with low glucose (1g/L) and 1% FBS for 48 hr in the presence of
683
dexamethasone (1g/ml) to maintain HIV latency as reported previously [45]. The cells
684
were next cultured in fresh medium without dexamethasone, un-treated, or treated with
685
low dose (20 pg/ml) and high dose (400 pg/ml) TNF- for 24 h. One portion of the cells
686
treated with high dose TNF- were washed twice with PBS, followed by culturing in fresh
687
medium in the absence of TNF- and dexamethasone for 48 h. Total RNAs from each
P a g e 32 | 60
688
cell line with different treatments were isolated by using the RNeasy Plus Mini kit from
689
Qiagen (Cat#74134). The isolated RNAs were treated with RNase-free DNAse I at 37 C
690
for 30 min to remove genomic DNA, followed by a second-round purification using the
691
same RNA purification kit. For reproducibility concerns, the RNA-Seq analysis consisted
692
of RNA samples from two independent experiments performed several months apart.
693
Total cellular RNA was subjected to 150 base long, paired end RNA-Seq on an
694
NovaSeq 6000 instrument. RNA-Seq reads were quality controlled using Fastqc and
695
trimmed for any leftover adaptor-derived sequences, and sequences with Phred score
696
less than 30 with Trim Galore, which is a wrapper based on Cutadapt and FastQC. Any
697
reads shorter than 40 nucleotides after the trimming was not used in alignment. The pre-
698
processed reads were aligned to the human genome (hg38/GRCh38) with the Gencode
699
release 28 as the reference annotations using STAR version 2.7.2b [115], followed by
700
gene-level quantitation using htseq-count [116]. In parallel, the pre-processed reads were
701
pseudoaligned using Kallisto version 0.43.1 [117], with 100 rounds of bootstrapping to the
702
Gencode release 28 of the human transcriptome to which the sequence of the transfected
703
HIV genome and the deduced HIV spliced transcripts were added. The resulting
704
quantitations were normalized using Sleuth. The two pipelines yielded concordant results.
705
Pairwise differential expression tests were performed using generalized linear models as
706
implemented in edgeR (QL) [118], and false discovery rate (FDR) values were calculated
707
for each differential expression value.
708
Protein-coding genes that were expressed at a minimum abundance of 5
709
transcripts per million (TPM) were used for pathway analysis with fold change values as
710
the ranking parameter while controlling false discovery rate at 0.05. Gene Set Enrichment
P a g e 33 | 60
711
Analysis (GSEA) package was used to identify the enriched pathway and promoter
712
elements using mSigDB and KEGG databases. Pathways that showed an FDR q-value
713
<= 0.25 were considered significantly enriched, per the GSEA package guidelines. The
714
number of genes contributing to the enrichment score was calculated using the leading
715
edge output of GSEA (tag multiplied by size).
716
Identification of marker genes for each study group
717
After filtration of the raw reads to remove low quality reads and mapping the clean
718
reads to the human reference genome using STAR software, differential analysis was
719
performed by edgeR package. For RNA-Seq data analysis, the bulk RNA-Seq data in a
720
form of digital gene expression (DGE) matrix was analyzed using the Seurat package for
721
R, v. 3.1.5 [119]. Variable genes were identified using the FindVariableFeatures function.
722
Top fifteen markers for each cluster were identified using a Wilcoxon Rank Sum test, and
723
a heat map was generated using the DoHeatmap function.
724
P a g e 34 | 60
725
SUPPORTING INFORMATION
726
S1 Fig. Nurr1 overexpression (OE) or knock-down (KD) substantially alters host
727
transcriptome. A, heatmaps representing top 15 gene markers for each treatment group.
728
Statistically-significant (p < 0.001) differentially expressed genes were determined using
729
the Wilcoxon rank-sum test reflecting the impacts of Nurr1 OE by comparing the control
730
cells HC69-3X-FLAG-vector (VT) with Nurr1 overexpressing cells HC69-3X-FLAG-Nurr1
731
(Nurr1 OE), as well as the impacts of KD by comparing the HC69-control shRNA1 and
732
control shRNA2 (Ctl shRNA1/2) cells with HC69-Nurr1 shRNA1 and shRNA2 (Nurr1
733
shRNA1/2) cells, respectively. Various cell lines were cultured in the absence (untreated)
734
or presence of high dose (400 pg/ml) TNF- for 24 hr. In addition, cells were given 48 hr
735
chase after stimulation with high dose (400 pg/ml) TNF- for 24 hr and subsequent
736
withdrawal. B, heatmaps showing top 15 gene transcript markers in samples from panel
737
A rearranged according to their status of treatment with TNF-.. The most enriched gene
738
transcripts as the result of Nurr1 overexpression or KD are listed in columns to the left.
739
The color-coded expression pattern of each gene transcript is shown in a heatmap to the
740
right.
741
S2 Fig. Nurr1 overexpression mainly impacts the recovery step following TNF-
742
stimulation. Trajectories of genes after stimulation with low dose (20 pg/ml) and high
743
dose (400 pg/ml) TNF- for 24 hr and following a 48 hr recovery (chase) after high dose
744
TNF- stimulation for 24 hr and subsequent withdrawal in the Nurr1 overexpression cell
745
line HC69-3X-FLAG-Nurr1 (Nurr1 OE) were shown. Trajectories of the same genes in the
746
control cell line HC69-3X-FLAG-vector (Ctl VT) were also shown, with a semi-transparent
747
line connecting identical genes between the control and Nurr1-overexpressing sides of
P a g e 35 | 60
748
each graph. Each line represented a gene, and the Y axis values indicated the log2
749
expression levels. The number of genes showing each trajectory in Nurr1-overexpressing
750
cells was shown on top. Genes that showed no change, were up regulated, and down
751
regulated in statistically significant manner (FDR<0.05, fold change>2) were indicated
752
with the letters n, u, and d respectively. Grouping of the different trajectories was based
753
on gene responses during stimulation with low dose (Step 1) and high dose (Step 2) TNF-
754
and the recovery time after TNF- stimulation and subsequent withdrawal (Step 3). For
755
instance, the group of genes marked “ndu” represented genes that were not significantly
756
changed in response to stimulation with low dose TNF- but were down regulated with
757
high dose TNF- stimulation and then up regulated during the recovery (chase) period.
758
S3 Fig. HC69-3X-FLAG-Nurr1 and HC69-3X-FLAG-vector cells strongly differ in the
759
recovery step following TNF- stimulation. Genes that showed a different trajectory
760
after TNF- stimulation for 24 hr and following a 48 hr recovery period between HC69-
761
3X-FLAG-vector (control) and HC69-3X-FLAG-Nurr1 cells were identified and groups
762
containing over 100 genes were graphed. Each line represented a gene and a semi-
763
transparent line connected identical genes between control and Nurr1-overexpressing
764
sides of each graph. The Y axis indicated the expression level of each gene throughout
765
the trajectory. Grouping of genes with no statistically significant changes in expression
766
(n), up regulated (d), or down regulated (d) in the three segments was as described in S2
767
Fig.
768
S4 Fig. Nurr1 overexpression (OE) substantially alters host transcriptome. Genes
769
involved in top differentially negatively enriched pathways in Fig 6A are shown in
770
heatmaps. The values shown in the heatmap correspond to the level of differential
P a g e 36 | 60
771
expression between Nurr1 overexpressing cells (marked as “Nurr1”) versus vector-
772
infected control cells (marked as “Vector”) during the chase step. The identities of the
773
plotted pathways and genes involved in the pathways are shown on the top and to the
774
right, respectively.
775
S5 Fig. Nurr1-specific gene expression during the chase step leads to strong
776
downregulated of genes involved in cell cycle. Genes that exclusively change in
777
expression during the chase step only in Nurr1 cells (see S3 Fig) were superimposed on
778
the KEGG cell cycle graph. The color bar on the top right indicates the level of differential
779
expression for each gene in Nurr1 cells during the chase step.
780
S6 Fig. TNF- stimulation leads to strong induction of NF-B-responsive genes
781
along with targets of multiple inflammatory cytokines. The most enriched
782
transcription factor binding motifs in proximity of the promoters of differentially expressed
783
genes are shown. The size of the circles indicates the level of enrichment, while the color
784
intensity reflects the statistical significance as shown by FDR. Positively- and negatively-
785
enriched motifs are shown after each treatment (shown at the bottom) in the left and right
786
panel, respectively. The identity of each motif, as annotated in the C3 lists of the MSIGDB
787
database, is shown to the left.
788
S7 Fig. Nurr1 associates with CoREST, HDAC1, G91, and EZH2 to form a
789
transcription repression complex in microglial cells (HC69). HC69-3X-FLAG-vector
790
and HC69-3X-FLAG-Nurr1 cells were cultured in the absence (untreated) or presence of
791
high dose (400 pg/ml) TNF- for 4 hr and 24 hr respectively. A portion of these cells were
792
also used in a chase experiment by culturing the cells for an additional 24 hr (chase) after
793
stimulation with high dose TNF- for 24 hr and subsequent washing with PBS (TNF-
P a g e 37 | 60
794
24h+24h). Total protein lysates from the differently treated cells were isolated and used
795
for co-immunoprecipitation (Co-IP) with a mouse anti-FLAG monoclonal antibody. The
796
original protein lysates (Input) and the Co-IP products were analyzed by Western blot
797
analysis with antibodies to FLAG, CoREST, HDAC1, G9a, EZH2, and β-tubulin
798
respectively.
799
P a g e 38 | 60
800
ACKNOWLEDGEMENT
801
This study was supported by NIH grants R01 DA043159 and R01 DA049481 to
802
J.K. and R21-AI127252 and two Development Awards from CFAR P30-AI36219 to S.V.
803
We thank Meenakhi Shukla for technical assistance for production of HIV-1 reporter virus.
804
This work made use of the High Performance Computing Resource for Advanced
805
Research Computing and flow cytometry and virology cores of the Center for AIDS
806
Research (CFAR) at Case Western Reserve University.
807
808
AUTHOR CONTRIBUTIONS
809
J.K., F.Y. and D.A. conceived of and oversaw the study. F.Y. performed all the wet
810
bench experiments in the manuscript except as noted and along with J.K., wrote the
811
manuscript. D.A. performed the culture of iPSC derived microglial cells and participated
812
in data analysis and manuscript preparation. K.N. constructed the gene knock out
813
lentiviruses and performed the ChIP-seq data analysis. S.V. processed and analyzed the
814
RNA-Seq dataset and performed the trajectory studies and pathway analysis, participated
815
in manuscript preparation and submitted the RNA-Seq studies performed in this project
816
to SRA (accession number to be provided). K.L. performed the marker gene discovery
817
for the RNA-Seq data. Y.G. performed microglial cell culture and participated in data
818
analysis. S.S. performed the culture of iPSC derived microglial cell and participated in
819
data analysis. All authors read the final manuscript and commented on it.
820
P a g e 39 | 60
821
FIGURE LEGENDS
822
Figure 1. Spontaneous silencing of active HIV in microglial cells. A, genome
823
organization of a d2EGFP reporter HIV-1 that was cloned in the lentiviral vector pHR’. A
824
fragment of HIV-1pNL4-3, containing Tat, Rev, Env, Vpu and Nef with the green
825
fluorescence reporter gene d2EGFP, was cloned into the lentiviral vector pHR’. The
826
resulted plasmid was used to produce the VSV-G HIV particles as described previously
827
[120]. Immortalized human microglial cells (hµglia) were infected with the lenti-HIV viral
828
particles, generating multiple clones with an integrated pro-virus genome. HC69 was a
829
representative of these clones. B, Schematic diagram of experimental scheme to study
830
the role of nuclear receptors in microglial reactivation and reversion to latency. C,
831
Representative phase contrast, GFP, and overlapped images of HC69 cells that were
832
cultured in the absence (untreated, left panel) and in the presence of TNF- (400 pg/ml)
833
for 24 hr (TNF- 24 h, middle panel) respectively, or used in a chase experiment by
834
continuously culturing HC69 cells in the absence of TNF- for 96 hr after stimulating the
835
cells with TNF- (400 pg/ml) for 24 hr and washing with PBS (TNF- 24 h+96 h, right
836
panel). The average percentages of GFP+ cells indicated for each panel were measured
837
by flow cytometry from triplicate wells.
838
Figure 2. Activation of Nurr1 enhances HIV silencing in immortalized human
839
microglial cells (hµglia). A, Impact of Nurr1 agonist 6-MP on HIV expression. Left:
840
Representative flow cytometry histograms. Right: Quantitative results from three
841
independent experiments. For this experiment, we used a batch of HC69 cells with high
842
numbers of GFP+ cells resulting from spontaneous HIV reactivation following multiple
843
passages of culture in the absence of dexamethasone. These cells were cultured in the
P a g e 40 | 60
844
presence of different doses of 6-MP for three days. The percentages of GFP+ cells from
845
the differently treated cells were measured by flow cytometry. The p-values of pair-
846
sample, Student’s t-tests comparing un-treated cells and cells treated with different
847
doses of 6-MP were calculated from three independent experiments. B, Western blot
848
detection of Nurr1, Nor1, HIV-1 Nef protein, and Nurr1 target gene MMP2 in HC69 cells
849
described in A. The level of β-tubulin was used as a loading control. C, the nuclear
850
receptor agonists dexamethasone (DEXA, 1 µM), Bexarotene (BEX, 1 µM) and 6-MP (1
851
µM) have additive effects on HIV silencing in HC69 cells. HC69 cells were first treated
852
with high dose (400 pg/ml) TNF- for 24 hr, followed by a 72 hr chase experiment during
853
which the cells were washed with PBS and cultured in fresh media in the presence of
854
placebo (DMSO) or the various NR agonists, alone or in combination. Expression of Nef
855
and β-tubulin in the differently treated cells was analyzed by Western blot analysis as
856
described in C.
857
Figure 3. Overexpression of Nurr1 in HC69 cells enhances HIV silencing. A,
858
RNA-Seq confirmation of overexpression (OE) of Nurr1 in HC69 cells. Sequence read
859
histograms for the Nurr1 locus is shown for control (vector) and Nurr1 overexpression.
860
Annotated genes for the shown locus are indicated on the top, and the position of the
861
locus on chromosome 2 is shown both at the top and the bottom. A read scale for each
862
row is shown on the right, with the values for the overexpression studies drawn on a log2
863
scale. B, Verification of Nur77, Nurr1, and Nor1 overexpression by Western blot analysis
864
in HC69 cell lines stably expressing 3X-FLAG-tagged Nur77, Nurr1, and Nor1
865
respectively. HC69 cells stably carrying the 3X-FLAG-empty vector were used as a
866
reference for comparison. The level of β-tubulin was used as a loading control. Notably,
P a g e 41 | 60
867
the levels of endogenous Nur77 and Nor1 in HC69 cells were very limited. In contrast,
868
Nurr1 was constitutively expressed in HC69 cells. C, Schematic depicting the TNF-
869
stimulation and chase studies. The four cell lines described in B were either untreated or
870
treated with high dose (400 pg/ml) TNF- for 24 h. To examine HIV silencing, one set of
871
TNF- induced cells were used in a chase experiment by continuous culture of the cells
872
in the absence of TNF- for an additional 48 h. The time points at which TNF- is added
873
or removed are shown by arrows on the top. D, Expression of HIV Nef protein in the
874
different cell lines before and after TNF- stimulation and at the end of the chase
875
experiment was measured by Western blot analysis. The level of -tubulin was used as
876
a loading control. E, Expression level of HIV mRNA (black bar graph) and Nurr1 (red
877
rectangles and lines) in transcripts per million cellular transcripts are shown for each of
878
the treatment steps shown in panel C in both vector-infected cells (on the left) and Nurr1
879
overexpressing cells (on the right half of the graph). For the 24 hr TNF- stimulation step,
880
both a low dose (20 pg/ml) and a high dose (400 pg/ml) are used. The values shown are
881
the average of three replicate RNA-Seq samples with two standard deviations as error
882
bars. The expression values for HIV and Nurr1 are shown on Y axes to the left and right,
883
respectively.
884
Figure 4. Nurr1 knock down (KD) in HC69 cells enhances HIV expression and
885
block proviral silencing during the chase step. A, RNA-Seq confirmation of Nurr1 KD
886
in HC69 cells. Read histograms for the Nurr1 locus is shown for non-targeting shRNA-
887
infected cells, and cells infected with Nurr1 specific shRNA lentiviral constructs.
888
Annotated genes for the shown locus are indicated on the top, and the position of the
889
locus on chromosome 2 is shown both at the top and the bottom. A read scale for each
P a g e 42 | 60
890
row is shown on the right, with the values for the knock down studies drawn on a linear
891
scale. B, Schematic depicting the TNF- stimulation and chase studies. The two shRNA
892
lentiviral transduced cell lines described in A were either untreated or treated with high
893
dose (400 pg/ml) TNF- for 24 hr. One set of TNF- induced cells were used in a chase
894
experiment in the absence of TNF- for an additional 48 hr. The time points at which TNF-
895
is added or removed are shown by arrows on the top. C, Western blot studies measuring
896
the expression of endogenous Nurr1, Nef, and β-tubulin in cells infected with either a non-
897
targeting control shRNA or Nurr1-specific shRNA lentiviral vectors. The expression
898
patterns from the TNF- (400 pg/ml) stimulation and the chase step are shown. D, KD of
899
endogenous Nurr1 strongly inhibits HIV silencing. The percentages of GFP+ cells in the
900
two cell lines, before treatment, at 24 hr post-TNF- (400 pg/ml) stimulation, and at 72 hr
901
after TNF- withdrawal (chase) were analyzed by flow cytometry and calculated from
902
three independent experiments. The difference in GFP expression between the two cell
903
lines at 72 hr chase was statistically significant, with a p = 0.0078. E, Expression level of
904
Nurr1 (red rectangles and lines) and the HIV provirus (black bar graph) in transcripts per
905
million cellular transcripts are shown for each of the treatment steps in both non-targeting
906
shRNA infected cells (on the left) and Nurr1-specific shRNA-infected cells (on the right
907
half of the graph). The values shown reflect the average of three replicate RNA-Seq
908
samples from two distinct shRNA constructs per control and Nurr1 knock down groups,
909
with two standard deviations as error bars. The expression values for HIV and Nurr1 are
910
shown on Y axes to the left and right, respectively.
911
Figure 5. Nurr1 overexpression leads to the inhibition of critical cellular
912
proliferation pathways. A, Patterns of differential gene expression during the chase step
P a g e 43 | 60
913
in vector-infected (top) and Nurr1 overexpressing (Nurr1 OE) cells. Dotted lines indicate
914
the two-fold cut off level. B, Pathway analyses of Nurr1 overexpression at baseline, during
915
TNF- stimulation, and following the recovery period after TNF- stimulation. The
916
identities of specific highly enriched pathways are shown on the Y axis, and the
917
comparisons are shown at the bottom. The color and size of circles correspond to
918
statistical significance, as shown by FDR, and normalized enrichment values,
919
respectively. Positive and negatively enriched pathways are shown in the left and right
920
plot, respectively.
921
Figure 6. Nurr1 overexpression accelerates homeostasis of activated
922
microglial cells by shutting down pathways involved in the maintenance of cellular
923
activation and inflammation. A, Identification of genes selectively altered as a result of
924
Nurr1 overexpression (Nurr1 OE), compared to the control empty vector (Ctl VT) cells, by
925
trajectory analysis. Genes that are unaltered (n), downregulated (d) or upregulated (u)
926
were identified during the activation and the chase steps and were clustered into families
927
with similar profiles. The total number of genes in each category is indicated for both the
928
control and Nurr1-overexpressing cells. Note that the major differences in the gene
929
expression profiles are seen in genes that are either upregulated or downregulated during
930
the chase (highlighted by asterisks). To enable the visualization of the trajectories with
931
low, medium and high membership, the X axis for each group is shown separately. B,
932
Pathway analysis using the Hallmark gene lists of the MSigDB database was performed
933
on non-TNF--responsive genes that are exclusively altered in expression during the
934
chase step in Nurr1 overexpressing cells, corresponding to genes which follow nnu and
935
nnd trajectories in Nurr1 cells and an nnn trajectory in control cells (see S3 Fig). The
P a g e 44 | 60
936
identity of each pathway is shown to the left, and the direction of enrichment (+ or -) is
937
shown at the bottom. The color and size of circles corresponded to statistical significance,
938
as shown by FDR, and normalized enrichment values, respectively.
939
Figure 7. Nurr1 promotes recruitment of the CoREST repressor complex to
940
HIV promoter. A, Schematic illustration of Nurr1-mediated epigenetic silencing of active
941
HIV in microglial cells by recruiting the CoREST/HDAC1/G9a/EZH2 repression complex
942
to HIV promoter. B, ChIP-seq signals (numbers of sequence reads on Y axis) along the
943
reporter HIV-1 pro-viral genome (Figure 1A) on the X axis, resulting from ChIP-seq
944
analysis with antibodies to EZH2, G9a, HDAC1, CoREST, and control IgG, respectively,
945
and sheared chromatins prepared from HC69 cells that were un-treated, induced with
946
TNF- (400 pg/ml) for 4 hr and 24 hr respectively, or used in a chase experiment by
947
continuously culturing HC69 cells in the absence of TNF- for 24 hr after stimulating the
948
cells with TNF- (400 pg/ml) for 24 hr and washing with PBS. Construction of ChIP-seq
949
DNA libraries with the ChIP products, enrichment for HIV-1 specific sequences, and data
950
analysis following Ion Torrent sequencing were described in Materials & Methods.
951
Positions of ChIP sequence reads along the viral genome were marked. C & D, levels of
952
CoREST (C) and G9a (D) in HIV 5’LTR (+30 to +134) in HC69-control shRNA (Control)
953
and HC69-Nurr1 shRNA (Nurr1 KD) cell lines that were treated as described in B. The
954
levels of CoREST and G9a in HIV 5’LTR were measured by qPCR and calculated as
955
percentages of the amounts of ChIP products over input DNA from triplicate qPCR.
956
Figure 8. The CoREST repressor complex plays a pivotal role in silencing
957
active HIV in microglial cells. A, Inhibition of HDAC1, G9a, and EZH2 blocked silencing
958
of activated HIV in HC69 cells. HC69 cells were stimulated with high dose (400 pg/ml)
P a g e 45 | 60
959
TNF- for 24 hr. After washing with PBS, the cells were cultured in the presence of DMSO
960
(placebo, Control), HDAC inhibitor SAHA (2 M), G9a inhibitor UNC0638 (2.5 M), and
961
EZH2 inhibitor GSK343 (2.5 M), respectively, for 48 hr. The levels of GFP expression
962
for each treatment were measured by flow cytometry and calculated from three
963
independent experiments, with p values between the control and treatment with each
964
inhibitor indicated. B, Verification of CoREST KD by Western blot detection of CoREST
965
protein expression in HC69 cell lines stably expressing control shRNA or CoREST-
966
specific shRNA. C, Verification of EZH2 and G9a KO by Western blot detection of G9a
967
and EZH2 protein expression in HC69 cells stably expressing CRISPR/Cas9 and G9a or
968
EZH2 specific gRNA, which were compared to the control HC69 cells stably expressing
969
CRISPR/Cas9 without gRNA. -tubulin was used as a loading control for all Western blot
970
analysis. D, CoREST KD prevents HIV silencing. The HC69-control shRNA and HC69-
971
CoREST-shRNA cells were untreated, induced with high dose (400 pg/ml) TNF- for 24
972
hr, or used in a chase experiment by continuous culturing the cells for 48 hr after TNF-
973
stimulation for 24 hr and washes with PBS. GFP expression levels of all cells were
974
measured by flow cytometry and the mean values were calculated from three
975
independent experiments. Significant differences were observed between the HC69-
976
control shRNA and HC69-CoREST shRNA cell lines. E, G9a and EZH2 KO prevents HIV
977
silencing. Evaluation of the HC69 cell lines expressing G9a or EZH2 specific gRNA or
978
empty vector by flow cytometry following the same protocol as in panel D. There was a
979
significant difference between HC69-vector and HC69 EZH2 or G9a KO cell lines at 48
980
hr after TNF- withdrawal, with p < 0.01.
P a g e 46 | 60
981
Figure 9. Nurr1 Mediates HIV silencing in iPSC-derived microglial cells (iMG).
982
A, Representative phase contrast, GFP, and overlapped images of iMG that were un-
983
infected or infected with the reporter HIV-1 shown in Fig 1A, at 48 hr post-infection (hpi).
984
HIV-infected iMG were treated with different doses of Nurr1 agonist 6-MP or AQ for four
985
days, followed by flow cytometry analysis of GFP expression. B, The average levels of
986
GFP expression in iMG treated with various doses of 6-MP. C, The average levels of GFP
987
expression in iMG treated with various doses of AQ were calculated from three replicates.
988
D, The levels of HIV RNA (un-spliced) in the cells described in panels A and B, were
989
measured by RT-qPCR. E, The mRNA level of Nurr1 target gene MMP2 in the same cells
990
was measured by qRT-PCR. The average levels of HIV transcript and MMP2 mRNA in
991
each sample were calculated from triplicates of qRT-qPCR. Differences in HIV and MMP2
992
mRNA levels between un-treated cells and cells treated with different doses of 6-MP or
993
AQ were statistically significant (** p-values <0.001). HIV transcripts were only detected
994
in infected iMG cells (panel D). MMP2 mRNA was significantly elevated in HIV infected
995
iMG (panel E).
996
P a g e 47 | 60
997
REFERENCES
998
1.
Resnick L, Berger JR, Shapshak P, Tourtellotte WW. Early penetration of the
999
blood-brain-barrier by HIV. Neurology. 1988;38(1):9-14. Epub 1988/01/01. doi:
1000
10.1212/wnl.38.1.9. PubMed PMID: 3422110.
1001
2.
Achim CL, Schrier RD, Wiley CA. Immunopathogenesis of HIV encephalitis. Brain
1002
pathology
(Zurich,
Switzerland).
1991;1(3):177-84.
Epub
1991/04/01.
doi:
1003
10.1111/j.1750-3639.1991.tb00657.x. PubMed PMID: 1669706.
1004
3.
Castellano P, Prevedel L, Eugenin EA. HIV-infected macrophages and microglia
1005
that survive acute infection become viral reservoirs by a mechanism involving Bim.
1006
Scientific reports. 2017;7(1):12866. Epub 2017/10/11. doi: 10.1038/s41598-017-12758-
1007
w. PubMed PMID: 28993666; PubMed Central PMCID: PMCPMC5634422.
1008
4.
Wallet C, De Rovere M, Van Assche J, Daouad F, De Wit S, Gautier V, et al.
1009
Microglial Cells: The Main HIV-1 Reservoir in the Brain. Front Cell Infect Microbiol.
1010
2019;9:362. Epub 2019/11/12. doi: 10.3389/fcimb.2019.00362. PubMed PMID:
1011
31709195; PubMed Central PMCID: PMCPMC6821723.
1012
5.
Dunfee R, Thomas ER, Gorry PR, Wang J, Ancuta P, Gabuzda D. Mechanisms of
1013
HIV-1 neurotropism. Current HIV research. 2006;4(3):267-78. Epub 2006/07/18. doi:
1014
10.2174/157016206777709500. PubMed PMID: 16842080.
1015
6.
Eggers C, Arendt G, Hahn K, Husstedt IW, Maschke M, Neuen-Jacob E, et al. HIV-
1016
1-associated neurocognitive disorder: epidemiology, pathogenesis, diagnosis, and
1017
treatment. J Neurol. 2017;264(8):1715-27. Epub 2017/06/02. doi: 10.1007/s00415-017-
1018
8503-2. PubMed PMID: 28567537; PubMed Central PMCID: PMCPMC5533849.
1019
7.
Clements JE, Gama L, Graham DR, Mankowski JL, Zink MC. A simian
1020
immunodeficiency virus macaque model of highly active antiretroviral treatment: viral
1021
latency in the periphery and the central nervous system. Curr Opin HIV AIDS.
1022
2011;6(1):37-42. Epub 2011/01/19. doi: 10.1097/COH.0b013e3283412413. PubMed
1023
PMID: 21242892; PubMed Central PMCID: PMCPMC3215305.
1024
8.
Desplats P, Dumaop W, Smith D, Adame A, Everall I, Letendre S, et al. Molecular
1025
and pathologic insights from latent HIV-1 infection in the human brain. Neurology.
1026
2013;80(15):1415-23.
Epub
2013/03/15.
doi:
10.1212/WNL.0b013e31828c2e9e.
1027
PubMed PMID: 23486877; PubMed Central PMCID: PMC3662272.
1028
9.
Brew BJ, Barnes SL. The impact of HIV central nervous system persistence on
1029
pathogenesis. AIDS (London, England). 2019;33 Suppl 2:S113-S21. Epub 2019/12/04.
1030
doi: 10.1097/QAD.0000000000002251. PubMed PMID: 31790377.
1031
10.
Brew BJ, Chan P. Update on HIV dementia and HIV-associated neurocognitive
1032
disorders. Curr Neurol Neurosci Rep. 2014;14(8):468. Epub 2014/06/19. doi:
1033
10.1007/s11910-014-0468-2. PubMed PMID: 24938216.
P a g e 48 | 60
1034
11.
Glass JD, Fedor H, Wesselingh SL, McArthur JC. Immunocytochemical
1035
quantitation of human immunodeficiency virus in the brain: correlations with dementia.
1036
Annals of neurology. 1995;38(5):755-62. Epub 1995/11/01. doi: 10.1002/ana.410380510.
1037
PubMed PMID: 7486867.
1038
12.
Masliah E, Heaton RK, Marcotte TD, Ellis RJ, Wiley CA, Mallory M, et al. Dendritic
1039
injury is a pathological substrate for human immunodeficiency virus-related cognitive
1040
disorders. HNRC Group. The HIV Neurobehavioral Research Center. Annals of
1041
neurology. 1997;42(6):963-72. Epub 1997/12/24. doi: 10.1002/ana.410420618. PubMed
1042
PMID: 9403489.
1043
13.
Gannon P, Khan MZ, Kolson DL. Current understanding of HIV-associated
1044
neurocognitive disorders pathogenesis. Curr Opin Neurol. 2011;24(3):275-83. Epub
1045
2011/04/07. doi: 10.1097/WCO.0b013e32834695fb. PubMed PMID: 21467932; PubMed
1046
Central PMCID: PMCPMC3683661.
1047
14.
Jurga AM, Paleczna M, Kuter KZ. Overview of General and Discriminating Markers
1048
of Differential Microglia Phenotypes. Frontiers in cellular neuroscience. 2020;14:198.
1049
Epub 2020/08/28. doi: 10.3389/fncel.2020.00198. PubMed PMID: 32848611; PubMed
1050
Central PMCID: PMCPMC7424058.
1051
15.
Zhou T, Huang Z, Sun X, Zhu X, Zhou L, Li M, et al. Microglia Polarization with
1052
M1/M2 Phenotype Changes in rd1 Mouse Model of Retinal Degeneration. Front
1053
Neuroanat. 2017;11:77. Epub 2017/09/21. doi: 10.3389/fnana.2017.00077. PubMed
1054
PMID: 28928639; PubMed Central PMCID: PMCPMC5591873.
1055
16.
Alsema AM, Jiang Q, Kracht L, Gerrits E, Dubbelaar ML, Miedema A, et al. Profiling
1056
Microglia From Alzheimer's Disease Donors and Non-demented Elderly in Acute Human
1057
Postmortem Cortical Tissue. Front Mol Neurosci. 2020;13:134. Epub 2020/11/17. doi:
1058
10.3389/fnmol.2020.00134. PubMed PMID: 33192286; PubMed Central PMCID:
1059
PMCPMC7655794.
1060
17.
Mesquida-Veny F, Del Rio JA, Hervera A. Macrophagic and microglial complexity
1061
after neuronal injury. Progress in neurobiology. 2021;200:101970. Epub 2020/12/29. doi:
1062
10.1016/j.pneurobio.2020.101970. PubMed PMID: 33358752.
1063
18.
Miedema A, Wijering MHC, Eggen BJL, Kooistra SM. High-Resolution
1064
Transcriptomic and Proteomic Profiling of Heterogeneity of Brain-Derived Microglia in
1065
Multiple Sclerosis. Front Mol Neurosci. 2020;13:583811. Epub 2020/11/17. doi:
1066
10.3389/fnmol.2020.583811. PubMed PMID: 33192299; PubMed Central PMCID:
1067
PMCPMC7654237.
1068
19.
Sheeler C, Rosa JG, Ferro A, McAdams B, Borgenheimer E, Cvetanovic M. Glia
1069
in Neurodegeneration: The Housekeeper, the Defender and the Perpetrator. Int J Mol Sci.
1070
2020;21(23). Epub 2020/12/06. doi: 10.3390/ijms21239188. PubMed PMID: 33276471;
1071
PubMed Central PMCID: PMCPMC7730416.
P a g e 49 | 60
1072
20.
Avedissian SN, Dyavar SR, Fox HS, Fletcher CV. Pharmacologic approaches to
1073
HIV-associated neurocognitive disorders. Curr Opin Pharmacol. 2020;54:102-8. Epub
1074
2020/10/14. doi: 10.1016/j.coph.2020.09.003. PubMed PMID: 33049585; PubMed
1075
Central PMCID: PMCPMC7770011.
1076
21.
Neumann H, Kotter MR, Franklin RJ. Debris clearance by microglia: an essential
1077
link between degeneration and regeneration. Brain : a journal of neurology. 2009;132(Pt
1078
2):288-95. Epub 2008/06/24. doi: 10.1093/brain/awn109. PubMed PMID: 18567623;
1079
PubMed Central PMCID: PMCPMC2640215.
1080
22.
Nimmerjahn A, Kirchhoff F, Helmchen F. Resting microglial cells are highly
1081
dynamic surveillants of brain parenchyma in vivo. Science. 2005;308(5726):1314-8. Epub
1082
2005/04/16. doi: 10.1126/science.1110647. PubMed PMID: 15831717.
1083
23.
Alvarez-Carbonell D, Ye F, Ramanath N, Garcia-Mesa Y, Knapp PE, Hauser KF,
1084
et al. Cross-talk between microglia and neurons regulates HIV latency. PLoS pathogens.
1085
2019;15(12):e1008249. Epub 2019/12/31. doi: 10.1371/journal.ppat.1008249. PubMed
1086
PMID: 31887215; PubMed Central PMCID: PMCPMC6953890.
1087
24.
Smith LK, Kuhn TB, Chen J, Bamburg JR. HIV Associated Neurodegenerative
1088
Disorders: A New Perspective on the Role of Lipid Rafts in Gp120-Mediated
1089
Neurotoxicity. Current HIV research. 2018;16(4):258-69. Epub 2018/10/04. doi:
1090
10.2174/1570162X16666181003144740. PubMed PMID: 30280668; PubMed Central
1091
PMCID: PMCPMC6398609.
1092
25.
Belichenko PV, Miklossy J, Celio MR. HIV-I induced destruction of neocortical
1093
extracellular matrix components in AIDS victims. Neurobiology of disease. 1997;4(3-
1094
4):301-10. Epub 1997/01/01. doi: 10.1006/nbdi.1997.0143. PubMed PMID: 9361307.
1095
26.
Bagetta G, Corasaniti MT, Berliocchi L, Navarra M, Finazzi-Agro A, Nistico G. HIV-
1096
1 gp120 produces DNA fragmentation in the cerebral cortex of rat. Biochemical and
1097
biophysical research communications. 1995;211(1):130-6. Epub 1995/06/06. doi:
1098
10.1006/bbrc.1995.1787. PubMed PMID: 7779077.
1099
27.
Marino J, Maubert ME, Mele AR, Spector C, Wigdahl B, Nonnemacher MR.
1100
Functional impact of HIV-1 Tat on cells of the CNS and its role in HAND. Cell Mol Life
1101
Sci. 2020;77(24):5079-99. Epub 2020/06/25. doi: 10.1007/s00018-020-03561-4. PubMed
1102
PMID: 32577796; PubMed Central PMCID: PMCPMC7674201.
1103
28.
Minagar A, Shapshak P, Fujimura R, Ownby R, Heyes M, Eisdorfer C. The role of
1104
macrophage/microglia and astrocytes in the pathogenesis of three neurologic disorders:
1105
HIV-associated dementia, Alzheimer disease, and multiple sclerosis. J Neurol Sci.
1106
2002;202(1-2):13-23. Epub 2002/09/11. doi: 10.1016/s0022-510x(02)00207-1. PubMed
1107
PMID: 12220687.
1108
29.
Galicia O, Sanchez-Alavez M, Mendez Diaz M, Navarro L, Prospero-Garcia O.
1109
[HIV glycoprotein 120: possible etiological agent of AIDS-associated dementia]. Rev
1110
Invest Clin. 2002;54(5):437-52. Epub 2003/02/18. PubMed PMID: 12587419.
P a g e 50 | 60
1111
30.
Kaushik DK, Gupta M, Basu A. Microglial response to viral challenges: every silver
1112
lining comes with a cloud. Front Biosci (Landmark Ed). 2011;16:2187-205. Epub
1113
2011/05/31. doi: 10.2741/3847. PubMed PMID: 21622170.
1114
31.
Periyasamy P, Thangaraj A, Bendi VS, Buch S. HIV-1 Tat-mediated microglial
1115
inflammation involves a novel miRNA-34a-NLRC5-NFkappaB signaling axis. Brain Behav
1116
Immun. 2019;80:227-37. Epub 2019/03/16. doi: 10.1016/j.bbi.2019.03.011. PubMed
1117
PMID: 30872089; PubMed Central PMCID: PMCPMC6660398.
1118
32.
Catani MV, Corasaniti MT, Navarra M, Nistico G, Finazzi-Agro A, Melino G. gp120
1119
induces cell death in human neuroblastoma cells through the CXCR4 and CCR5
1120
chemokine receptors. Journal of neurochemistry. 2000;74(6):2373-9. Epub 2000/05/23.
1121
doi: 10.1046/j.1471-4159.2000.0742373.x. PubMed PMID: 10820198.
1122
33.
Zhu Q, Song X, Zhou J, Wang Y, Xia J, Qian W, et al. Target of HIV-1 Envelope
1123
Glycoprotein gp120-Induced Hippocampal Neuron Damage: Role of Voltage-Gated K(+)
1124
Channel
Kv2.1.
Viral
Immunol.
2015;28(9):495-503.
Epub
2015/09/24.
doi:
1125
10.1089/vim.2015.0020.
PubMed
PMID:
26393286;
PubMed
Central
PMCID:
1126
PMCPMC4642831.
1127
34.
Rom I, Deshmane SL, Mukerjee R, Khalili K, Amini S, Sawaya BE. HIV-1 Vpr
1128
deregulates calcium secretion in neural cells. Brain research. 2009;1275:81-6. Epub
1129
2009/03/31. doi: 10.1016/j.brainres.2009.03.024. PubMed PMID: 19328187; PubMed
1130
Central PMCID: PMCPMC2692350.
1131
35.
Chompre G, Cruz E, Maldonado L, Rivera-Amill V, Porter JT, Noel RJ, Jr.
1132
Astrocytic expression of HIV-1 Nef impairs spatial and recognition memory. Neurobiology
1133
of disease. 2013;49:128-36. Epub 2012/08/29. doi: 10.1016/j.nbd.2012.08.007. PubMed
1134
PMID: 22926191; PubMed Central PMCID: PMCPMC3530662.
1135
36.
Alvarez-Carbonell D, Ye F, Ramanath N, Dobrowolski C, Karn J. The
1136
Glucocorticoid Receptor Is a Critical Regulator of HIV Latency in Human Microglial Cells.
1137
J Neuroimmune Pharmacol. 2018. Epub 2018/07/11. doi: 10.1007/s11481-018-9798-1.
1138
PubMed PMID: 29987742.
1139
37.
Garcia-Mesa Y, Jay TR, Checkley MA, Luttge B, Dobrowolski C, Valadkhan S, et
1140
al. Immortalization of primary microglia: a new platform to study HIV regulation in the
1141
central nervous system. J Neurovirol. 2017;23(1):47-66. Epub 2016/11/23. doi:
1142
10.1007/s13365-016-0499-3. PubMed PMID: 27873219; PubMed Central PMCID:
1143
PMCPMC5329090.
1144
38.
Alvarez-Carbonell D, Garcia-Mesa Y, Milne S, Das B, Dobrowolski C, Rojas R, et
1145
al. Toll-like receptor 3 activation selectively reverses HIV latency in microglial cells.
1146
Retrovirology. 2017;14(1):9. Epub 2017/02/09. doi: 10.1186/s12977-017-0335-8.
1147
PubMed PMID: 28166799; PubMed Central PMCID: PMCPMC5294768.
1148
39.
Eilebrecht S, Le Douce V, Riclet R, Targat B, Hallay H, Van Driessche B, et al.
1149
HMGA1 recruits CTIP2-repressed P-TEFb to the HIV-1 and cellular target promoters.
P a g e 51 | 60
1150
Nucleic Acids Res. 2014. Epub 2014/03/14. doi: 10.1093/nar/gku168. PubMed PMID:
1151
24623795.
1152
40.
Cherrier T, Le Douce V, Eilebrecht S, Riclet R, Marban C, Dequiedt F, et al. CTIP2
1153
is a negative regulator of P-TEFb. Proc Natl Acad Sci U S A. 2013;110(31):12655-60.
1154
Epub 2013/07/16. doi: 10.1073/pnas.1220136110. PubMed PMID: 23852730; PubMed
1155
Central PMCID: PMC3732990.
1156
41.
Van Lint C, Bouchat S, Marcello A. HIV-1 transcription and latency: an update.
1157
Retrovirology. 2013;10:67. Epub 2013/06/28. doi: 10.1186/1742-4690-10-67. PubMed
1158
PMID: 23803414; PubMed Central PMCID: PMC3699421.
1159
42.
Marban C, Suzanne S, Dequiedt F, de Walque S, Redel L, Van Lint C, et al.
1160
Recruitment of chromatin-modifying enzymes by CTIP2 promotes HIV-1 transcriptional
1161
silencing. EMBO J. 2007;26(2):412-23. PubMed PMID: 17245431.
1162
43.
Quivy V, Van Lint C. Diversity of acetylation targets and roles in transcriptional
1163
regulation: the human immunodeficiency virus type 1 promoter as a model system.
1164
Biochem Pharmacol. 2002;64:925-34.
1165
44.
Van Lint C, Emiliani S, Ott M, Verdin E. Transcriptional activation and chromatin
1166
remodeling of the HIV-1 promoter in response to histone acetylation. EMBO J.
1167
1996;15(5):1112-20.
1168
45.
Alvarez-Carbonell D, Ye F, Ramanath N, Dobrowolski C, Karn J. The
1169
Glucocorticoid Receptor Is a Critical Regulator of HIV Latency in Human Microglial Cells.
1170
Journal of neuroimmune pharmacology : the official journal of the Society on
1171
NeuroImmune
Pharmacology.
2019;14(1):94-109.
Epub
2018/07/11.
doi:
1172
10.1007/s11481-018-9798-1. PubMed PMID: 29987742; PubMed Central PMCID:
1173
PMCPMC6394485.
1174
46.
Zetterström RH, Solomin L, Mitsiadis T, Olson L, Perlmann T. Retinoid X receptor
1175
heterodimerization and developmental expression distinguish the orphan nuclear
1176
receptors NGFI-B, Nurr1, and Nor1. Mol Endocrinol. 1996;10(12):1656-66. Epub
1177
1996/12/01. doi: 10.1210/mend.10.12.8961274. PubMed PMID: 8961274.
1178
47.
Saucedo-Cardenas O, Quintana-Hau JD, Le WD, Smidt MP, Cox JJ, De Mayo F,
1179
et al. Nurr1 is essential for the induction of the dopaminergic phenotype and the survival
1180
of ventral mesencephalic late dopaminergic precursor neurons. Proceedings of the
1181
National Academy of Sciences of the United States of America. 1998;95(7):4013-8. Epub
1182
1998/05/09. doi: 10.1073/pnas.95.7.4013. PubMed PMID: 9520484; PubMed Central
1183
PMCID: PMCPMC19954.
1184
48.
Sonntag KC, Simantov R, Kim KS, Isacson O. Temporally induced Nurr1 can
1185
induce a non-neuronal dopaminergic cell type in embryonic stem cell differentiation. The
1186
European journal of neuroscience. 2004;19(5):1141-52. Epub 2004/03/16. doi:
1187
10.1111/j.1460-9568.2004.03204.x. PubMed PMID: 15016073; PubMed Central PMCID:
1188
PMCPMC2614072.
P a g e 52 | 60
1189
49.
Eells JB, Wilcots J, Sisk S, Guo-Ross SX. NR4A gene expression is dynamically
1190
regulated in the ventral tegmental area dopamine neurons and is related to expression of
1191
dopamine neurotransmission genes. J Mol Neurosci. 2012;46(3):545-53. Epub
1192
2011/09/21. doi: 10.1007/s12031-011-9642-z. PubMed PMID: 21932041; PubMed
1193
Central PMCID: PMCPMC3280384.
1194
50.
Saijo K, Winner B, Carson CT, Collier JG, Boyer L, Rosenfeld MG, et al. A
1195
Nurr1/CoREST pathway in microglia and astrocytes protects dopaminergic neurons from
1196
inflammation-induced
death.
Cell.
2009;137(1):47-59.
Epub
2009/04/07.
doi:
1197
10.1016/j.cell.2009.01.038. PubMed PMID: 19345186; PubMed Central PMCID:
1198
PMC2754279.
1199
51.
Jeon SG, Yoo A, Chun DW, Hong SB, Chung H, Kim JI, et al. The Critical Role of
1200
Nurr1 as a Mediator and Therapeutic Target in Alzheimer's Disease-related
1201
Pathogenesis.
Aging
Dis.
2020;11(3):705-24.
Epub
2020/06/04.
doi:
1202
10.14336/AD.2019.0718. PubMed PMID: 32489714; PubMed Central PMCID:
1203
PMCPMC7220289.
1204
52.
Kandil EA, Sayed RH, Ahmed LA, Abd El Fattah MA, El-Sayeh BM. Modulatory
1205
Role of Nurr1 Activation and Thrombin Inhibition in the Neuroprotective Effects of
1206
Dabigatran Etexilate in Rotenone-Induced Parkinson's Disease in Rats. Molecular
1207
neurobiology. 2018;55(5):4078-89. Epub 2017/06/07. doi: 10.1007/s12035-017-0636-x.
1208
PubMed PMID: 28585189.
1209
53.
Smith GA, Rocha EM, Rooney T, Barneoud P, McLean JR, Beagan J, et al. A
1210
Nurr1 agonist causes neuroprotection in a Parkinson's disease lesion model primed with
1211
the toll-like receptor 3 dsRNA inflammatory stimulant poly(I:C). PloS one.
1212
2015;10(3):e0121072. Epub 2015/03/31. doi: 10.1371/journal.pone.0121072. PubMed
1213
PMID: 25815475; PubMed Central PMCID: PMCPMC4376720 This does not alter the
1214
authors' adherence to PLOS ONE policies on sharing data and materials.
1215
54.
Liu W, Gao Y, Chang N. Nurr1 overexpression exerts neuroprotective and anti-
1216
inflammatory roles via down-regulating CCL2 expression in both in vivo and in vitro
1217
Parkinson's disease models. Biochemical and biophysical research communications.
1218
2017;482(4):1312-9. Epub 2016/12/13. doi: 10.1016/j.bbrc.2016.12.034. PubMed PMID:
1219
27940361.
1220
55.
Moon M, Jung ES, Jeon SG, Cha MY, Jang Y, Kim W, et al. Nurr1 (NR4A2)
1221
regulates Alzheimer's disease-related pathogenesis and cognitive function in the 5XFAD
1222
mouse
model.
Aging
Cell.
2019;18(1):e12866.
Epub
2018/12/06.
doi:
1223
10.1111/acel.12866.
PubMed
PMID:
30515963;
PubMed
Central
PMCID:
1224
PMCPMC6351845.
1225
56.
Li X, Tjalkens RB, Shrestha R, Safe S. Structure-dependent activation of gene
1226
expression by bis-indole and quinoline-derived activators of nuclear receptor 4A2. Chem
1227
Biol Drug Des. 2019;94(4):1711-20. Epub 2019/05/19. doi: 10.1111/cbdd.13564. PubMed
1228
PMID: 31102570; PubMed Central PMCID: PMCPMC6791730.
P a g e 53 | 60
1229
57.
Das B, Dobrowolski C, Luttge B, Valadkhan S, Chomont N, Johnston R, et al.
1230
Estrogen Receptor α (ESR1) is a key regulator of HIV-1 latency that imparts gender-
1231
specific restrictions on the latent reservoir
1232
Proc Natl Acad Sci U S A. 2018;In press.
1233
58.
Li Q, Karim AF, Ding X, Das B, Dobrowolski C, Gibson RM, et al. Novel high
1234
throughput pooled shRNA screening identifies NQO1 as a potential drug target for host
1235
directed therapy for tuberculosis. Sci Rep. 2016;6:27566. Epub 2016/06/15. doi:
1236
10.1038/srep27566.
PubMed
PMID:
27297123;
PubMed
Central
PMCID:
1237
PMCPMC4906352.
1238
59.
Wires ES, Alvarez D, Dobrowolski C, Wang Y, Morales M, Karn J, et al.
1239
Methamphetamine activates nuclear factor kappa-light-chain-enhancer of activated B
1240
cells (NF-kappaB) and induces human immunodeficiency virus (HIV) transcription in
1241
human microglial cells. J Neurovirol. 2012;18(5):400-10. Epub 2012/05/24. doi:
1242
10.1007/s13365-012-0103-4. PubMed PMID: 22618514; PubMed Central PMCID:
1243
PMC3469781.
1244
60.
Shtutman M, Maliyekkel A, Shao Y, Carmack CS, Baig M, Warholic N, et al.
1245
Function-based gene identification using enzymatically generated normalized shRNA
1246
library and massive parallel sequencing. Proc Natl Acad Sci U S A. 2010;107(16):7377-
1247
82. Epub 2010/04/07. doi: 1003055107 [pii]
1248
10.1073/pnas.1003055107. PubMed PMID: 20368428; PubMed Central PMCID:
1249
PMC2867740.
1250
61.
Sims D, Mendes-Pereira AM, Frankum J, Burgess D, Cerone MA, Lombardelli C,
1251
et al. High-throughput RNA interference screening using pooled shRNA libraries and next
1252
generation sequencing. Genome Biol. 2011;12(10):R104. Epub 2011/10/25. doi:
1253
10.1186/gb-2011-12-10-r104. PubMed PMID: 22018332; PubMed Central PMCID:
1254
PMC3333774.
1255
62.
Llewellyn GN, Alvarez-Carbonell D, Chateau M, Karn J, Cannon PM. HIV-1
1256
infection of microglial cells in a reconstituted humanized mouse model and identification
1257
of compounds that selectively reverse HIV latency. Journal of neurovirology. 2017. Epub
1258
2017/12/20. doi: 10.1007/s13365-017-0604-2. PubMed PMID: 29256041.
1259
63.
Das B, Dobrowolski C, Luttge B, Valadkhan S, Chomont N, Johnston R, et al.
1260
Estrogen receptor-1 is a key regulator of HIV-1 latency that imparts gender-specific
1261
restrictions on the latent reservoir. Proc Natl Acad Sci U S A. 2018;115(33):E7795-e804.
1262
Epub 2018/08/01. doi: 10.1073/pnas.1803468115. PubMed PMID: 30061382; PubMed
1263
Central PMCID: PMCPmc6099847.
1264
64.
Wang J, Bi W, Zhao W, Varghese M, Koch RJ, Walker RH, et al. Selective brain
1265
penetrable Nurr1 transactivator for treating Parkinson's disease. Oncotarget.
1266
2016;7(7):7469-79. Epub 2016/02/11. doi: 10.18632/oncotarget.7191. PubMed PMID:
1267
26862735; PubMed Central PMCID: PMCPMC4884932.
P a g e 54 | 60
1268
65.
Dong J, Li S, Mo JL, Cai HB, Le WD. Nurr1-Based Therapies for Parkinson's
1269
Disease.
CNS
Neurosci
Ther.
2016;22(5):351-9.
Epub
2016/03/26.
doi:
1270
10.1111/cns.12536.
PubMed
PMID:
27012974;
PubMed
Central
PMCID:
1271
PMCPMC4833611.
1272
66.
Chen XX, Qian Y, Wang XP, Tang ZW, Xu JT, Lin H, et al. Nurr1 promotes
1273
neurogenesis of dopaminergic neuron and represses inflammatory factors in the transwell
1274
coculture system of neural stem cells and microglia. CNS neuroscience & therapeutics.
1275
2018;24(9):790-800. Epub 2018/02/17. doi: 10.1111/cns.12825. PubMed PMID:
1276
29450981; PubMed Central PMCID: PMCPMC6489950.
1277
67.
Kim CH, Han BS, Moon J, Kim DJ, Shin J, Rajan S, et al. Nuclear receptor Nurr1
1278
agonists enhance its dual functions and improve behavioral deficits in an animal model
1279
of Parkinson's disease. Proceedings of the National Academy of Sciences of the United
1280
States
of
America.
2015;112(28):8756-61.
Epub
2015/07/01.
doi:
1281
10.1073/pnas.1509742112. PubMed PMID: 26124091; PubMed Central PMCID:
1282
PMCPMC4507186.
1283
68.
Wansa KD, Harris JM, Yan G, Ordentlich P, Muscat GE. The AF-1 domain of the
1284
orphan nuclear receptor NOR-1 mediates trans-activation, coactivator recruitment, and
1285
activation by the purine anti-metabolite 6-mercaptopurine. The Journal of biological
1286
chemistry. 2003;278(27):24776-90. Epub 2003/04/24. doi: 10.1074/jbc.M300088200.
1287
PubMed PMID: 12709428.
1288
69.
Ordentlich P, Yan Y, Zhou S, Heyman RA. Identification of the antineoplastic agent
1289
6-mercaptopurine as an activator of the orphan nuclear hormone receptor Nurr1. The
1290
Journal of biological chemistry. 2003;278(27):24791-9. Epub 2003/04/24. doi:
1291
10.1074/jbc.M302167200. PubMed PMID: 12709433.
1292
70.
Rodriguez-Calvo R, Ferran B, Alonso J, Marti-Pamies I, Aguilo S, Calvayrac O, et
1293
al. NR4A receptors up-regulate the antiproteinase alpha-2 macroglobulin (A2M) and
1294
modulate MMP-2 and MMP-9 in vascular smooth muscle cells. Thromb Haemost.
1295
2015;113(6):1323-34. Epub 2015/03/27. doi: 10.1160/TH14-07-0645. PubMed PMID:
1296
25809189.
1297
71.
Bozzelli PL, Caccavano A, Avdoshina V, Mocchetti I, Wu JY, Conant K. Increased
1298
matrix metalloproteinase levels and perineuronal net proteolysis in the HIV-infected brain;
1299
relevance to altered neuronal population dynamics. Experimental neurology.
1300
2020;323:113077. Epub 2019/11/05. doi: 10.1016/j.expneurol.2019.113077. PubMed
1301
PMID: 31678140.
1302
72.
van Neerven S, Kampmann E, Mey J. RAR/RXR and PPAR/RXR signaling in
1303
neurological and psychiatric diseases. Progress in neurobiology. 2008;85(4):433-51.
1304
Epub 2008/06/17. doi: 10.1016/j.pneurobio.2008.04.006. PubMed PMID: 18554773.
1305
73.
Zuo Y, Huang L, Enkhjargal B, Xu W, Umut O, Travis ZD, et al. Activation of
1306
retinoid
X
receptor
by
bexarotene
attenuates
neuroinflammation
via
1307
PPARγ/SIRT6/FoxO3a
pathway
after
subarachnoid
hemorrhage
in
rats.
J
P a g e 55 | 60
1308
Neuroinflammation. 2019;16(1):47. Epub 2019/02/23. doi: 10.1186/s12974-019-1432-5.
1309
PubMed PMID: 30791908; PubMed Central PMCID: PMCPMC6385420.
1310
74.
Boehm-Cagan A, Michaelson DM. Reversal of apoE4-driven brain pathology and
1311
behavioral deficits by bexarotene. The Journal of neuroscience : the official journal of the
1312
Society
for
Neuroscience.
2014;34(21):7293-301.
Epub
2014/05/23.
doi:
1313
10.1523/JNEUROSCI.5198-13.2014. PubMed PMID: 24849361; PubMed Central
1314
PMCID: PMCPMC6608187.
1315
75.
Mariani MM, Malm T, Lamb R, Jay TR, Neilson L, Casali B, et al. Neuronally-
1316
directed effects of RXR activation in a mouse model of Alzheimer's disease. Scientific
1317
reports. 2017;7:42270. Epub 2017/02/17. doi: 10.1038/srep42270. PubMed PMID:
1318
28205585; PubMed Central PMCID: PMCPMC5311933.
1319
76.
Ghosal K, Haag M, Verghese PB, West T, Veenstra T, Braunstein JB, et al. A
1320
randomized controlled study to evaluate the effect of bexarotene on amyloid-beta and
1321
apolipoprotein E metabolism in healthy subjects. Alzheimer's & dementia (New York, N
1322
Y). 2016;2(2):110-20. Epub 2016/06/17. doi: 10.1016/j.trci.2016.06.001. PubMed PMID:
1323
29067298; PubMed Central PMCID: PMCPMC5644280.
1324
77.
Casali BT, Reed-Geaghan EG, Landreth GE. Nuclear receptor agonist-driven
1325
modification of inflammation and amyloid pathology enhances and sustains cognitive
1326
improvements in a mouse model of Alzheimer's disease. J Neuroinflammation.
1327
2018;15(1):43. Epub 2018/02/17. doi: 10.1186/s12974-018-1091-y. PubMed PMID:
1328
29448961; PubMed Central PMCID: PMCPMC5815248.
1329
78.
De Miranda BR, Popichak KA, Hammond SL, Jorgensen BA, Phillips AT, Safe S,
1330
et al. The Nurr1 Activator 1,1-Bis(3'-Indolyl)-1-(p-Chlorophenyl)Methane Blocks
1331
Inflammatory Gene Expression in BV-2 Microglial Cells by Inhibiting Nuclear Factor
1332
kappaB.
Mol
Pharmacol.
2015;87(6):1021-34.
Epub
2015/04/11.
doi:
1333
10.1124/mol.114.095398. PubMed PMID: 25858541; PubMed Central PMCID:
1334
PMCPMC4429718.
1335
79.
Tsai MC, Manor O, Wan Y, Mosammaparast N, Wang JK, Lan F, et al. Long
1336
noncoding RNA as modular scaffold of histone modification complexes. Science.
1337
2010;329(5992):689-93. Epub 2010/07/10. doi: 10.1126/science.1192002. PubMed
1338
PMID: 20616235; PubMed Central PMCID: PMC2967777.
1339
80.
Gu H, Liang Y, Mandel G, Roizman B. Components of the REST/CoREST/histone
1340
deacetylase repressor complex are disrupted, modified, and translocated in HSV-1-
1341
infected cells. Proceedings of the National Academy of Sciences of the United States of
1342
America. 2005;102(21):7571-6. Epub 2005/05/18. doi: 10.1073/pnas.0502658102.
1343
PubMed PMID: 15897453; PubMed Central PMCID: PMCPMC1140450.
1344
81.
Vallat-Decouvelaere AV, Gray F, Chretien F, Le Pavec G, Dormont D, Gras G.
1345
[Neurotoxicity and neuroprotection, two aspects of microglial activation in human
1346
immunodeficiency virus (HIV) infection]. Ann Pathol. 2004;24(1):31-44. Epub 2004/06/12.
1347
doi: 10.1016/s0242-6498(04)93895-3. PubMed PMID: 15192535.
P a g e 56 | 60
1348
82.
Thangaraj A, Chivero ET, Tripathi A, Singh S, Niu F, Guo ML, et al. HIV TAT-
1349
mediated microglial senescence: Role of SIRT3-dependent mitochondrial oxidative
1350
stress.
Redox
Biol.
2021;40:101843.
Epub
2021/01/02.
doi:
1351
10.1016/j.redox.2020.101843. PubMed PMID: 33385630; PubMed Central PMCID:
1352
PMCPMC7779826.
1353
83.
Silverstein PS, Shah A, Weemhoff J, Kumar S, Singh DP, Kumar A. HIV-1 gp120
1354
and drugs of abuse: interactions in the central nervous system. Current HIV research.
1355
2012;10(5):369-83. Epub 2012/05/18. doi: 10.2174/157016212802138724. PubMed
1356
PMID: 22591361; PubMed Central PMCID: PMCPMC4086306.
1357
84.
Narasipura SD, Kim S, Al-Harthi L. Epigenetic regulation of HIV-1 latency in
1358
astrocytes.
Journal
of
virology.
2014;88(5):3031-8.
Epub
2013/12/20.
doi:
1359
10.1128/JVI.03333-13.
PubMed
PMID:
24352441;
PubMed
Central
PMCID:
1360
PMCPMC3958059.
1361
85.
Shirazi J, Shah S, Sagar D, Nonnemacher MR, Wigdahl B, Khan ZK, et al.
1362
Epigenetics, drugs of abuse, and the retroviral promoter. Journal of neuroimmune
1363
pharmacology : the official journal of the Society on NeuroImmune Pharmacology.
1364
2013;8(5):1181-96. Epub 2013/11/13. doi: 10.1007/s11481-013-9508-y. PubMed PMID:
1365
24218017; PubMed Central PMCID: PMCPMC3878082.
1366
86.
Rohr O, Aunis D, Schaeffer E. COUP-TF and Sp1 interact and cooperate in the
1367
transcriptional activation of the human immunodeficiency virus type 1 long terminal repeat
1368
in human microglial cells. The Journal of biological chemistry. 1997;272(49):31149-55.
1369
Epub 1998/01/10. doi: 10.1074/jbc.272.49.31149. PubMed PMID: 9388268.
1370
87.
Marban C, Redel L, Suzanne S, Van Lint C, Lecestre D, Chasserot-Golaz S, et al.
1371
COUP-TF interacting protein 2 represses the initial phase of HIV-1 gene transcription in
1372
human microglial cells. Nucleic Acids Res. 2005;33(7):2318-31. Epub 2005/04/26. doi:
1373
10.1093/nar/gki529.
PubMed
PMID:
15849318;
PubMed
Central
PMCID:
1374
PMCPMC1084325.
1375
88.
Marban C, Suzanne S, Dequiedt F, de Walque S, Redel L, Van Lint C, et al.
1376
Recruitment of chromatin-modifying enzymes by CTIP2 promotes HIV-1 transcriptional
1377
silencing.
The
EMBO
journal.
2007;26(2):412-23.
Epub
2007/01/25.
doi:
1378
10.1038/sj.emboj.7601516. PubMed PMID: 17245431; PubMed Central PMCID:
1379
PMCPMC1783449.
1380
89.
Le Douce V, Colin L, Redel L, Cherrier T, Herbein G, Aunis D, et al. LSD1
1381
cooperates with CTIP2 to promote HIV-1 transcriptional silencing. Nucleic Acids Res.
1382
2012;40(5):1904-15. Epub 2011/11/10. doi: 10.1093/nar/gkr857. PubMed PMID:
1383
22067449; PubMed Central PMCID: PMCPMC3300010.
1384
90.
Cherrier T, Le Douce V, Eilebrecht S, Riclet R, Marban C, Dequiedt F, et al. CTIP2
1385
is a negative regulator of P-TEFb. Proceedings of the National Academy of Sciences of
1386
the United States of America. 2013;110(31):12655-60. Epub 2013/07/16. doi:
P a g e 57 | 60
1387
10.1073/pnas.1220136110. PubMed PMID: 23852730; PubMed Central PMCID:
1388
PMCPMC3732990.
1389
91.
De Bosscher K, Desmet SJ, Clarisse D, Estebanez-Perpina E, Brunsveld L.
1390
Nuclear receptor crosstalk - defining the mechanisms for therapeutic innovation. Nat Rev
1391
Endocrinol. 2020;16(7):363-77. Epub 2020/04/19. doi: 10.1038/s41574-020-0349-5.
1392
PubMed PMID: 32303708.
1393
92.
Kurakula K, Koenis DS, van Tiel CM, de Vries CJ. NR4A nuclear receptors are
1394
orphans but not lonesome. Biochimica et biophysica acta. 2014;1843(11):2543-55. Epub
1395
2014/07/01. doi: 10.1016/j.bbamcr.2014.06.010. PubMed PMID: 24975497.
1396
93.
Wallen-Mackenzie A, Mata de Urquiza A, Petersson S, Rodriguez FJ, Friling S,
1397
Wagner J, et al. Nurr1-RXR heterodimers mediate RXR ligand-induced signaling in
1398
neuronal
cells.
Genes
Dev.
2003;17(24):3036-47.
Epub
2003/12/19.
doi:
1399
10.1101/gad.276003.
PubMed
PMID:
14681209;
PubMed
Central
PMCID:
1400
PMCPMC305256.
1401
94.
Kadkhodaei B, Ito T, Joodmardi E, Mattsson B, Rouillard C, Carta M, et al. Nurr1
1402
is required for maintenance of maturing and adult midbrain dopamine neurons. The
1403
Journal of neuroscience : the official journal of the Society for Neuroscience.
1404
2009;29(50):15923-32. Epub 2009/12/18. doi: 10.1523/JNEUROSCI.3910-09.2009.
1405
PubMed PMID: 20016108; PubMed Central PMCID: PMCPMC6666174.
1406
95.
Tornqvist N, Hermanson E, Perlmann T, Stromberg I. Generation of tyrosine
1407
hydroxylase-immunoreactive neurons in ventral mesencephalic tissue of Nurr1 deficient
1408
mice. Brain Res Dev Brain Res. 2002;133(1):37-47. Epub 2002/02/19. doi:
1409
10.1016/s0165-3806(01)00317-0. PubMed PMID: 11850062.
1410
96.
Decressac M, Volakakis N, Bjorklund A, Perlmann T. NURR1 in Parkinson
1411
disease--from pathogenesis to therapeutic potential. Nature reviews Neurology.
1412
2013;9(11):629-36. Epub 2013/10/16. doi: 10.1038/nrneurol.2013.209. PubMed PMID:
1413
24126627.
1414
97.
Liu H, Wei L, Tao Q, Deng H, Ming M, Xu P, et al. Decreased NURR1 and PITX3
1415
gene expression in Chinese patients with Parkinson's disease. Eur J Neurol.
1416
2012;19(6):870-5. Epub 2012/02/09. doi: 10.1111/j.1468-1331.2011.03644.x. PubMed
1417
PMID: 22309633.
1418
98.
Bensinger SJ, Tontonoz P. A Nurr1 pathway for neuroprotection. Cell.
1419
2009;137(1):26-8. Epub 2009/04/07. doi: 10.1016/j.cell.2009.03.024. PubMed PMID:
1420
19345183.
1421
99.
Spathis AD, Asvos X, Ziavra D, Karampelas T, Topouzis S, Cournia Z, et al.
1422
Nurr1:RXRalpha heterodimer activation as monotherapy for Parkinson's disease.
1423
Proceedings of the National Academy of Sciences of the United States of America.
1424
2017;114(15):3999-4004. Epub 2017/03/30. doi: 10.1073/pnas.1616874114. PubMed
1425
PMID: 28348207; PubMed Central PMCID: PMCPMC5393203.
P a g e 58 | 60
1426
100.
Edison P. Neuroinflammation, microglial activation, and glucose metabolism in
1427
neurodegenerative diseases. Int Rev Neurobiol. 2020;154:325-44. Epub 2020/08/03. doi:
1428
10.1016/bs.irn.2020.03.017. PubMed PMID: 32739010.
1429
101.
Purisai MG, McCormack AL, Cumine S, Li J, Isla MZ, Di Monte DA. Microglial
1430
activation as a priming event leading to paraquat-induced dopaminergic cell
1431
degeneration. Neurobiology of disease. 2007;25(2):392-400. Epub 2006/12/15. doi:
1432
10.1016/j.nbd.2006.10.008. PubMed PMID: 17166727; PubMed Central PMCID:
1433
PMCPMC2001246.
1434
102.
Spudich SS, Ances BM. Neurologic complications of HIV infection. Topics in
1435
antiviral medicine. 2012;20(2):41-7. Epub 2012/06/20. PubMed PMID: 22710906;
1436
PubMed Central PMCID: PMCPMC6148861.
1437
103.
Zhang J, He H, Qiao Y, Zhou T, He H, Yi S, et al. Priming of microglia with IFN-
1438
gamma impairs adult hippocampal neurogenesis and leads to depression-like behaviors
1439
and cognitive defects. Glia. 2020. Epub 2020/07/12. doi: 10.1002/glia.23878. PubMed
1440
PMID: 32652855.
1441
104.
Meda L, Cassatella MA, Szendrei GI, Otvos L, Jr., Baron P, Villalba M, et al.
1442
Activation of microglial cells by beta-amyloid protein and interferon-gamma. Nature.
1443
1995;374(6523):647-50. Epub 1995/04/13. doi: 10.1038/374647a0. PubMed PMID:
1444
7715705.
1445
105.
Chao CC, Hu S, Molitor TW, Shaskan EG, Peterson PK. Activated microglia
1446
mediate neuronal cell injury via a nitric oxide mechanism. Journal of immunology
1447
(Baltimore, Md : 1950). 1992;149(8):2736-41. Epub 1992/10/15. PubMed PMID:
1448
1383325.
1449
106.
Hwang IK, Choi JH, Nam SM, Park OK, Yoo DY, Kim W, et al. Activation of
1450
microglia and induction of pro-inflammatory cytokines in the hippocampus of type 2
1451
diabetic
rats.
Neurol
Res.
2014;36(9):824-32.
Epub
2014/02/28.
doi:
1452
10.1179/1743132814Y.0000000330. PubMed PMID: 24571083.
1453
107.
Shytaj IL, Procopio FA, Tarek M, Carlon-Andres I, Tang HY, Goldman AR, et al.
1454
Glycolysis downregulation is a hallmark of HIV-1 latency and sensitizes infected cells to
1455
oxidative stress. EMBO Mol Med. 2021;13(8):e13901. Epub 2021/07/22. doi:
1456
10.15252/emmm.202013901. PubMed PMID: 34289240; PubMed Central PMCID:
1457
PMCPMC8350904.
1458
108.
Jakaria M, Haque ME, Cho DY, Azam S, Kim IS, Choi DK. Molecular Insights into
1459
NR4A2(Nurr1):
an
Emerging
Target
for
Neuroprotective
Therapy
Against
1460
Neuroinflammation and Neuronal Cell Death. Molecular neurobiology. 2019;56(8):5799-
1461
814. Epub 2019/01/27. doi: 10.1007/s12035-019-1487-4. PubMed PMID: 30684217.
1462
109.
Miguel-Escalada I, Bonas-Guarch S, Cebola I, Ponsa-Cobas J, Mendieta-Esteban
1463
J, Atla G, et al. Human pancreatic islet three-dimensional chromatin architecture provides
1464
insights into the genetics of type 2 diabetes. Nat Genet. 2019;51(7):1137-48. Epub
P a g e 59 | 60
1465
2019/06/30. doi: 10.1038/s41588-019-0457-0. PubMed PMID: 31253982; PubMed
1466
Central PMCID: PMCPMC6640048.
1467
110.
Friedman J, Cho WK, Chu CK, Keedy KS, Archin NM, Margolis DM, et al.
1468
Epigenetic silencing of HIV-1 by the histone H3 lysine 27 methyltransferase enhancer of
1469
Zeste 2. J Virol. 2011;85(17):9078-89. Epub 2011/07/01. doi: 10.1128/jvi.00836-11.
1470
PubMed PMID: 21715480; PubMed Central PMCID: PMCPmc3165831.
1471
111.
Afgan E, Baker D, van den Beek M, Blankenberg D, Bouvier D, Cech M, et al. The
1472
Galaxy platform for accessible, reproducible and collaborative biomedical analyses: 2016
1473
update.
Nucleic
Acids
Res.
2016;44(W1):W3-W10.
Epub
2016/05/04.
doi:
1474
10.1093/nar/gkw343.
PubMed
PMID:
27137889;
PubMed
Central
PMCID:
1475
PMCPMC4987906.
1476
112.
Langmead B, Trapnell C, Pop M, Salzberg SL. Ultrafast and memory-efficient
1477
alignment of short DNA sequences to the human genome. Genome Biol. 2009;10(3):R25.
1478
Epub 2009/03/06. doi: 10.1186/gb-2009-10-3-r25. PubMed PMID: 19261174; PubMed
1479
Central PMCID: PMCPMC2690996.
1480
113.
Langmead B, Salzberg SL. Fast gapped-read alignment with Bowtie 2. Nat
1481
Methods. 2012;9(4):357-9. Epub 2012/03/06. doi: 10.1038/nmeth.1923. PubMed PMID:
1482
22388286; PubMed Central PMCID: PMCPMC3322381.
1483
114.
Ramirez F, Ryan DP, Gruning B, Bhardwaj V, Kilpert F, Richter AS, et al.
1484
deepTools2: a next generation web server for deep-sequencing data analysis. Nucleic
1485
Acids Res. 2016;44(W1):W160-5. Epub 2016/04/16. doi: 10.1093/nar/gkw257. PubMed
1486
PMID: 27079975; PubMed Central PMCID: PMCPMC4987876.
1487
115.
Dobin A, Gingeras TR. Mapping RNA-seq Reads with STAR. Curr Protoc
1488
Bioinformatics.
2015;51:11
4
1-
4
9.
Epub
2015/09/04.
doi:
1489
10.1002/0471250953.bi1114s51. PubMed PMID: 26334920; PubMed Central PMCID:
1490
PMCPMC4631051.
1491
116.
Anders S, Pyl PT, Huber W. HTSeq--a Python framework to work with high-
1492
throughput sequencing data. Bioinformatics. 2015;31(2):166-9. Epub 2014/09/28. doi:
1493
10.1093/bioinformatics/btu638. PubMed PMID: 25260700; PubMed Central PMCID:
1494
PMCPMC4287950.
1495
117.
Bray NL, Pimentel H, Melsted P, Pachter L. Near-optimal probabilistic RNA-seq
1496
quantification.
Nat
Biotechnol.
2016;34(5):525-7.
Epub
2016/04/05.
doi:
1497
10.1038/nbt.3519. PubMed PMID: 27043002.
1498
118.
Robinson MD, McCarthy DJ, Smyth GK. edgeR: a Bioconductor package for
1499
differential expression analysis of digital gene expression data. Bioinformatics.
1500
2010;26(1):139-40. Epub 2009/11/17. doi: 10.1093/bioinformatics/btp616. PubMed
1501
PMID: 19910308; PubMed Central PMCID: PMCPMC2796818.
P a g e 60 | 60
1502
119.
Satija R, Farrell JA, Gennert D, Schier AF, Regev A. Spatial reconstruction of
1503
single-cell gene expression data. Nat Biotechnol. 2015;33(5):495-502. Epub 2015/04/14.
1504
doi:
10.1038/nbt.3192.
PubMed
PMID:
25867923;
PubMed
Central
PMCID:
1505
PMCPMC4430369.
1506
120.
Kim YK, Bourgeois CF, Pearson R, Tyagi M, West MJ, Wong J, et al. Recruitment
1507
of TFIIH to the HIV LTR is a rate-limiting step in the emergence of HIV from latency. The
1508
EMBO journal. 2006;25(15):3596-604. Epub 2006/07/29. doi: 10.1038/sj.emboj.7601248.
1509
PubMed PMID: 16874302; PubMed Central PMCID: PMCPMC1538560.
1510
1511
Agag Brev Env Tat i _
= |
LTR Tat il f Vpu Rev EL]
Experimental work flow
HIV provirus NF-«B
\
TNF-o
400 pg / ml ——_
—> —>
GFP+ cells 8.0%
1.0 uM
80 OuM 4404
| 120/
5 - 100
2 4 ~~
5 “4 53.65% re 20.30%
20+ 40 ©
| 20: 3
04 0 a
2
1603 2.5 uM 5.0uM
140: 150 -
120. ra
8100: 00+ oO
£ 80° 15.48% 12.41%
z i 50
40:
20: 0 10 25 5.0
0! ’ :
040 10? 10°10¢108 010 102 10°10°105 6-MP Concentration (1M)
GFP GFP
B Inhibition of HIV by 6-MP C Inhibition of HIV expression by NR agonists
S-MP (uM) 0 05 1.0 25 5.0 TNF-o. (24 hr + 72 hr Chase)
Nuri — —— s
uy
~
Nor ‘ia! Smee! Sse Se? —| 8 9
: s § Sixx
Nur77 -_ Paes 42a oq
Pees a tcdd
—— &£eéo*esF FSS
Net aa SEF eH EH EET S
| _-— «eee - — Nef
TINE (24 TNF £2 TH Ullase)
xy G+ X3g
+ dW-9
*38 + e.g
*x3q * dig
Vx3q + X3g
AN-9
x: 3g
Vx3q
(
Oswq) e081,
“Y pe Dany)
Pae9 “ur
~ Nef
Tubt
Inhibition of HIV expression by NR agonists
Homo sapiens GRCh38_v93 chr2:156313726-156349561 (35.8 kbp)
KKH 3' HH HE
AC074099.1 NR4A2 (NURR1) FLAG
Empty vector (9° 8 5 3
y a adit tho ke Maa a 4 10 B8a Nur77
Nurr1 overexp Ho° ERS Nurr1
val Hits ub . = lh , [10 Zoe
' ; t Nor1
156, 325, 000 156,335,000 156, 345, 000
C Chase Experiment Design Yubume
+ TNF-a Chase E HIV mRNA Expre
}24hr } 48 hr 6000 GaNt
Empty vector =
p. 5000
NR overexpression =
& 4000
D HIV Nef Expression 3 3000
Untreated TNF-o Chase 2 2000
wi
She She Ske > 1000
gSE5 R555 8555 §&
>Z22Z22 52225222 0
Untreated +
Nef Serr - a TNF-« 20 pg/ml +
TNF-a. 400 pg/ml +
TD) or ST TK Chase +
= HIV mRNA Expression
Ctl VT Nurr1 OE
= 600037 SS
= a
& 5000 60
= ¢
§ 4000 452
% 3000 °
g #L30 6
5 2000 oO
> 1000 SE
s
= 0) 0 =
Untreated + +
TNF-a 20 pg/ml + +
TNF-c 400 pg/ml + +
TNF-a + Chase + +
HIV Nef Expression
Chase
TNF-o
Untreated
LU) i” a ‘Si
Beg Seer e — ESS ee
S AOA
13726-156349561 (35.8 kbp) we ss s
° ruc ie
ee
100 oe
aa FOE 33 nur?
100 Ex®
wit , HOSES Note ~ i —
a : Nor1 a
35,000 156,345,000
Tubulin dca
E HIV mRNA Expression
Ctl VT Nurr1 OE
3 hr _ 6000577 SS
= =
& 5000 60E
= ¢
c 4
5 000 a g
3000 hss &
Chase & 2000 x
— Ww
i > 1000 SE
fc EL =I 3
222 0 On =
— Untreated + +
——— TNF-« 20pgiml = + +
TNF-c. 400 pg/ml + +
TNF-a + Chase + +
. | kK Zee
T T T T T T T —
156,325,000 156,335,000 156,345,000
B Chase Experiment Design C_ HIV Nef Expression
> »
<P & SUA
SF SK SS
shRNA Oo YO SS
+ TNF-a Chase
ur
}24hr ¥ oe
48 hr a
Nurrt shRNA —
Tun
TNF-a Chase
Control shRNA
D Flow Cytometry E HIV mRNA Expression
<100 = Ctl shRNA Nurr1 shRNA —
= segs? 0.0078 = 6000 url Sone
5 ~ 2 E
3B 60 5000 42
n ¢
® 5 4000 =
2 40 @ 58
ui % 3000 o
a 20 5 2000] * 79
O 9g a 1 wi
sec a Se >
smRNA52 52 §2 = 0 02
oO oO oO Untreated + +
== TNF-a + +
Untreated TNF-a Chase Chase + +
c HIV MRNA Expression
Ctl shRNA Nurr1 shRN
6000
5000
4000
3000
2000
1000
0
Untreated + +
TNF-a + +
Chase + +
yo wo BR OP
BDiecewrn’d Eewmraeaaianm {TRas\
HIV Expression (TPM)
oO
Flow Cytometry E
)
0.0078
o
p=
> up
Oo °O
Oo oO
=
(9
2
©
So
t+
Oo
N
°o
%) UoIssesdx"Z dj
Untreated TNF-a Chase
Chase Experiment Design
& N & N
shRNA & wc
Nurr1 ————
Nef
Homo sapiens GRCh38_v93 chr2:156313726-156349561 (35.8 kbp)
KKH
3' HHH HR HS
AC074099.1 NR4A2 (NURR1)
20 Bes
Control shRNA - OO
wud ds A A 10 833
20 Eo ©
Nurr1 shRNA . \ tio 5 52
156,325,000 156,335,000 156,345,000
Chase Experiment Design C_ HIV Nef Expression
& \ & \
FSF
shRNA WYO SS
+ TNF-a Chase —
N —— —
Y2anr ¥ 48 hr w oe
Nef
Touin——
TNF-a Chase
Control shRNA
Nurr1 shRNA
Flow Cytometry E HIV mRNA Expression
<100 p= 0.0078 Ctl shRNA Nurr1 shRNA —
x 8 —_ =6000=--—~C—“‘CS ————-5 =
c 4g ~“ —_—_- hue] (“a
vitterential EAPression (lOgz)
empty vector
2000 4000 6000
Nurr1 OE
2000 4000 6000
Protein-coding genes
SONA FS FOE se Bs
FEO Sl OF Che Ce
(@seyd) WNYUS HID SA WNMUS LUNN
— (4NL) VNYUS 130 SA WNUUS |NN
— (payeeuqum) WNYUS HID SA YNMUS LUNN
+ © @e0@¢ e@ ee (aS2YD) 10}08/ SA LUNN
oe @ (4N1) 10790, SA LN
1 |_| | oo (payeauqup) soe, SA LUNN
+o+0+-+-@ . + -@ t (@seyd) WNYUS LID SA WNUUS LUNN
o-9-9-9-4—-# TT (4NL) VNUUS HID SA WNMUS NN
e+ @0e ¢ 4 (payeeuju) YNMUS LID SA WNXUS LUNN
ee. ¢ @ @0e0e e200 © e (aseyd) J0}O9/ SA LUNN
or TT? eT (4N1)40}98/\ SA |.uNN
7-7 TTT (payeasjup)) JO}O@A SA LUNN
DOO VPLYMO>MYNMSCHAECHy*ZoC#VrKS
Sec SSBces Sar sss-s2gseg-oe
SEESeE C 2zeePersaacsoe =
Bn®GESOSSES w& Se SS SS Oo ESSEa ©
®@¢in-Fox Hoo S PSV 2H. 259553 E
Te=StSsolo s FA SCcNscar $s S
ee" ogi< Fa SEZ8EO =
NO
= F Og C535 5 uw
10000, ,
a
So
So
oO
6000
4000
2000
Genes per trajectory
nnn
ctl VT
HG Nurr1 OE
Genes per trajectory
ndu nnu ndn nud nun nnd
ddu dnn dnu dun uud
Gene Trajectories
Apical Junction
Estrogen Early
Myogenesis
Mitotic Spindle
KRAS Down
Estrogen late
Hedgehog
Coagulation
Bile Acid Metabolism
UV Response Down
EM Cell Transition
IFN-o.
Unfolded Protein
Oxidative Phos.
DNA Repair
Allograph Rejection
IFN-y
MYC Targets V2
G2M Checkpoint
MTORC1
Protein Secretion
E2F Targets
MYC Targets V1
©2000 COOOO
SOOCCHHCOCHCOCO?
+ =
nnu + nnd
aa
Microglia Cell Microglia Cell
flammatory
Signals CoREST
Autocrine Signals Complexes
(TNF-a, IL-6, IL1-f)
ChIP-Seq / HC69 Cc ChIP / COREST
70
NF-c. 24 hr + Chase -~ Mm 4. 60
TNF-c. 24 hr Y 2 50
TNF-o 4 hr W 6 49
3
Untreated 8 30
NF-c. 24 hr + Chase = 20
TNF-o 24 hr © 10
TNF-« 4h 6 0
“ r
: e Sy &
Untreated g ~ eS
ys <*
”
NF-o 24 hr + Chase D am
TNF-a. 24 hr g 60 ChIP / G9a
TNF-o 4 hr 2 50
Untreated
S 40
NF-c 24 hr + Chase aetna .
a 3
5 B30
TNF-c. 24 hr er eee | (ee
a x 20
TNF-o. 4 hr 8
Untreated 10
9 9 s 4 4 @
Untreated _ — & ~ oe &
Position (nt)L_ 1 pt ys x
0 1000 2000 3000 4000 5000 6000 o
LTR] v (RRE W@EGEP BRI
Hem —®Control —®—Nurr1 KD
SYNUB B69 - j
SVNU5 ZHZ4 - {
JOOS - } I
¢ § &
=
yO 8
wNuUsIsayoo- |}
yNaus jonuog- J }
B 8
x $s
oO Fe
{ of
O, o
% | 8
< Oo
3 -|8 % :
2|/S yD |=
—-|2
¥ Bsr +4, 718
Slv|ce bo 3
vi A 2
Q Yon =
oO .
a
Ny ~
mo OG 98 So Oo Oo
Oo ODO © vt N
(%) uoisseudxe G49
Effect of EZH2 and G9a KO
E
Effect of COREST KD
)
oS
o
=
° Oo So o
© © vt N
(%) uoIsseldx9 q49
oO
®69
cHZA
Joe
69
cHZ3A
Joe
&69
@HZA
JO}OA
gRNAs
LS3uY°9D
joujuog
LS3Y°D
joujuog
LS3Y°D
joujuoDg
<x
Zz
a
=
7)
“ 95
©
Ay Y.
OZ,
Y
%
Effect of EZH2 and G9a KO
Ww
°
oO
=
olUDODlmUmUCOCOCOSS
on OO TFT N
(%) uolsseidxe q49
oO
e869
@HZ3
JO}OE/,
e869
cHZ3A
40}O8/,
e699
cHZA
JOA,
gRNAs
a
Qos
$p < 0.00001
GFP Expression (%
nyo
PotSOOSG Sd
-MP Concentration (
HIV RT-PCR
DINo HIV
Uninfected HIV-GFP
MRNA Level (Told)
OfpPr NW HA
_ EINo HIV
6-MP AQ
Bo 005 01.0
Concentration (4M)
ol TETEEEDIIOEE DY Orr” w TMIIOETE BY PANS
#100) SSS 5 £100) ses
50) 8838 ge} 8885
260;.°560 9 a)
Oo, je v Ov v ® mT oo 8
a 40 & a ae a 40 a Vv 4 v
tS 20 wm ae = Hal eam es
a a
uw O nu O
oO 0 0.5 1.0 2.5 5.0 oO 0 0.5 1.0 2.5 5.0
6-MP Concentration (uM) AQ Concentration (uM)
HIV RT-PCR E MMP2 RT-PCR
312 CINo HIV S5 _ SINo HIV
2 z z 2
[0.9 £4 t, :
g 23 } ii
® 0.6. ®
aay a 2
£03 <
= 33 + Z1
x m4 a
€ 0.0. £0
6-MP AQ 6-MP AQ
a0 05 1.0 Bo 005 01.0
Concentration (1M) Concentration (4M)
| 2021 | The Nerve Growth Factor IB-like Receptor Nurr1 (NR4A2) Recruits CoREST Transcription Repressor Complexes to Silence HIV Following Proviral Reactivation in Microglial Cells | 10.1101/2021.11.16.468784 | [
"Ye Fengchun",
"Alvarez-Carbonell David",
"Nguyen Kien",
"Valadkhan Saba",
"Leskov Konstantin",
"Garcia-Mesa Yoelvis",
"Sreeram Sheetal",
"Karn Jonathan"
] | creative-commons |
Universal features shaping organelle gene retention
Konstantinos Giannakis1,∗, Samuel J. Arrowsmith2,∗, Luke Richards3, Sara Gasparini4, Joanna M. Chustecki5,
Ellen C. Røyrvik6, Iain G. Johnston1,7,†
1 Department of Mathematics, University of Bergen, Norway; 2 G´en´etique mol´eculaire, g´enomique, microbiologie,
Universit´e de Strasbourg, France; 3 Department of Life Sciences, University of Warwick, UK; 4 Birkeland Centre for Space
Science, University of Bergen, Norway; 5 School of Biosciences, University of Birmingham, UK; 6 Department of Clinical
Sciences, University of Bergen, Norway; 7 Computational Biology Unit, University of Bergen, Norway
∗ these authors contributed equally to this work; † correspondence to iain.johnston@uib.no
Abstract
Mitochondria and plastids power complex life, and
retain their own organelle DNA (oDNA) genomes,
with highly reduced gene contents compared to their
endosymbiont ancestors.
Why some protein-coding
genes are retained in oDNA and some lost remains a
debated question. Here we harness over 15k oDNA
sequences and over 300 whole genome sequences
with tools from structural biology, bioinformatics, ma-
chine learning, and Bayesian model selection to re-
veal the properties of genes, and associated under-
lying mechanisms, that shape oDNA evolution. Strik-
ing symmetry exists between the two organelle types:
gene retention patterns in both are predicted by the
hydrophobicity of a protein product and its energetic
centrality within its protein complex, with additional
influences of nucleic acid and amino acid biochem-
istry.
Remarkably, retention principles from one or-
ganelle type successfully and quantitatively predict re-
tention in the other, supporting this universality; these
principles also distinguish gene profiles in indepen-
dent endosymbiotic relationships.
The identification
of these features shaping organelle gene retention
both provides quantitative support for several existing
evolutionary hypotheses, and suggests new biochemi-
cal and biophysical mechanisms influencing organelle
genome evolution.
Introduction
Mitochondria and plastids (the broader class of or-
ganelles of which chloroplasts are one type) are
bioenergetic organelles derived from the ancient en-
dosymbiotic acquisition of bacterial precursors [1].
The subsequent co-evolution of mitochondria and
plastids with their host cells has shaped complex life
[2, 3, 4]. Across eukaryotes, the genomes of the origi-
nal endosymbionts (estimated to have contained thou-
sands of genes [5]), have been dramatically reduced
through evolutionary time [6, 7, 1]. Genes have either
been lost completely or transferred to the ‘host’ cell
nucleus, so that modern-day organelle DNA (oDNA)
contains few genes, with profound implications for the
balance of control between the nucleus and endosym-
biont, and the inheritance and maintenance of vital ge-
netic information [8].
Selective pressures favouring organelle gene trans-
fer are largely agreed upon [7]. Nuclear encoding al-
lows recombination to avoid Muller’s ratchet (the irre-
versible buildup of damaging mutations) [9, 6], pro-
tection from chemical mutagens [10, 11] and repli-
cation errors [12, 13], and enhanced fixing of useful
mutations [7, 6]. However, these observations raise
the dual question: why are any genes retained in or-
ganelles at all [14]?
This question has been hotly
debated over decades, with many proposed hypothe-
ses. The preferential retention of genes encoding hy-
drophobic products has been suggested, due to the
challenge of correctly targetting and importing such
products to the correct organelle [15, 16, 17].
The
retention of genes playing central roles in controlling
redox activity has also been proposed, to facilitate lo-
cal control of activity [18]. Other hypotheses, including
roles for nucleic acid biochemistry [19], gene expres-
sion levels [20], energetic costs of encoding [21], toxi-
city [22], and others have been proposed, but quanti-
tative testing of these ideas remains limited [19, 23].
Applying tools from model selection to large-scale
genomic data offers unprecedented and powerful op-
portunities to both generate and impartially test evo-
lutionary and mechanistic hypotheses [24] (aligning
with an influential recent commentary on ideas in biol-
ogy [25]). Here, following previous work on mtDNA
evolution [19], we adopt this philosophy to explore
the mechanisms shaping gene loss across organelles.
First, mindful of the dangers of proposing parallels
between different organelles [26], we nonetheless
hypothesised that the same genetic features would
shape retention propensity of genes in mitochondrial
and plastid DNA. Such features would predispose a
gene to be more or less readily retained in oDNA
overall, while the total extent of oDNA retention in a
given species is shaped in parallel by functional and
metabolic features [23, 27] and evolutionary dynam-
ics (characterised statistically in elegant recent work
[28]). We further expect that these genetic features
1
would reflect the above evolutionary tension, between
maintaining genetic integrity and retaining the ability
to obtain and control machinery, that applies to both
organelles [29, 7].
With this general hypothesis in
mind, we proceed by taking an impartial, data-driven
approach using large-scale genomic data to investi-
gate which features of genes and their protein prod-
ucts predict oDNA gene retention presence (whether
any eukaryotes retain a given gene in oDNA) and ex-
tent (how commonly an oDNA gene is retained across
eukaryotes).
Results
Quantifying gene-specific oDNA loss pat-
terns across eukaryotes
To quantitatively explore the features predicting oDNA
gene retention, we first define a retention index for a
given oDNA gene, measuring its propensity to be re-
tained in oDNA. To this end, we acquired data on or-
ganelle gene content across eukaryotes, using 10328
whole mtDNA and 5176 whole ptDNA sequences from
NCBI. We curated these data with two different ap-
proaches, resembling supervised and unsupervised
philosophies, to form consistent records of gene pres-
ence/absence by species (see Methods).
The su-
pervised approach (manual assignment of ambiguous
gene records to a chosen gene label) and the un-
supervised approach (all-against-all BLAST compar-
ison of every gene record from the organelle genome
database) agreed tightly (Supplementary Fig.
S1).
Simply counting observations of each gene across
species is prone to large sampling bias, as some
taxa (notably bilaterians and angiosperms) are much
more densely sampled than others. Instead we recon-
structed gene loss events using oDNA sequences of
modern organisms and an estimated taxonomic rela-
tionship between them (see Methods). Motivated by
hypercubic transition path sampling [19, 30], we then
define the retention index of gene X as the number
of other genes already lost when gene X is lost (re-
sults were robust with alternative definition; see be-
low). This retention index, along with the unique pat-
terns of oDNA gene presence/absence and their tax-
onomic distribution, are illustrated in Fig. 1A (phyloge-
netic embedding in Supplementary Fig. S2).
The retention patterns of genes in mtDNA and
ptDNA across eukaryotes show pronounced structure,
arguing against a null hypothesis of random gene loss.
The several-fold expansion of mtDNA in this study
compared to [19] preserves the same structure, with,
for example, several rpl genes and sdh[2-4] commonly
lost and nad[1-6], cox[1-3] and cytb commonly re-
tained. The ptDNA patterns display pronounced clus-
tering, following previous observations [31], with one
cluster corresponding broadly to Viridiplantae (typi-
cally retaining ndh genes) and the other correspond-
ing broadly to brown and red algae, diatoms, and other
clades (typically lacking ndh genes but retaining more
atp, rps, rpl, psa, and psb). Several ribosomal sub-
units and ndhb are among the most retained in ptDNA,
with a second tier involving many ndh, psa, psb, and
atp genes retained in around half our species. Least
retained ptDNA genes include other members of psa,
psb, rps, and rpl.
Cross-organelle symmetry in the predic-
tion of gene retention by hydrophobicity
and GC content
We next compiled a set of quantitative properties of
genes and their protein products, linked to evolution-
ary hypotheses about the mechanisms shaping oDNA
gene retention [19]. These included gene length and
GC content, statistics of encoding and codon usage,
and protein hydrophobicity, molecular weight, energy
requirements for production, average carboxyl and
amino pKa values for amino acid residues, and oth-
ers (Supplementary Fig.
S3).
Our quantitative es-
timates for each feature were averages over a taxo-
nomically diverse sampling of eukaryotic records (see
Methods). We used Bayesian model selection to ask
which of these properties were most likely to be in-
cluded in a linear model predicting the retention index
of each gene. Following Ref. [19], this approach iden-
tifies likely predictors with quantified uncertainty, while
acting without prior favouring of any given hypotheses,
and automatically guarding against overfitting and the
appearance of correlated predictors providing redun-
dant information. In both mtDNA and ptDNA datasets,
models where high hydrophobicity and high GC con-
tent predict high gene retention were strongly favoured
(Fig. 1B). It is well-known that oDNA generally has
lower GC content than nuclear DNA, because of the
asymmetric mutational pressure arising from the hy-
drolytic deamination of cytosine to uracil, reducing GC
content in the high mutation system of oDNA [32].
However, our results show that higher GC content is
relatively favoured between oDNA genes – and so at
least partly independently of the general oDNA/nDNA
difference [19].
We then tested the capacity of models involving
these features to predict the retention index of oDNA
genes.
We split mtDNA and ptDNA gene sets into
50:50 training and test sets, trained linear models in-
volving hydrophobicity and GC content using the train-
ing data, and examined their performance in predic-
tion retention index in the independent test set. Av-
erage Spearman correlations were ρ = 0.64 and
ρ = 0.62 for training mt and pt sets respectively, and
ρ = 0.63 and ρ = 0.60 for test mt and pt sets respec-
tively (Fig. 1C). Correlations were higher still (ρ > 0.7)
when only subunits of core bioenergetic complexes
were considered (Supplementary Table S1). Follow-
ing our hypothesis that the same features predict re-
tention in the two organelle types, we also performed
cross-organelle experiments. That is, we trained a hy-
drophobicity and GC model using mt genes and ex-
amined its ability to predict pt gene retention, and vice
2
Figure 1: Structure and predictors of oDNA gene retention. (A) Each row of coloured/white pixels is a unique gene presence/absence
pattern found in eukaryotic oDNA, where columns are individual oDNA genes. Darker colours correspond to higher values of our assigned
retention index for a given gene. Each pattern may be present in many species: grey bars on the left of each row show the number of
species with that pattern in a number of eukaryotic clades. The pronounced split in ptDNA patterns reflects the evolutionary pathways
represented, for example, by Rhodophyta and Viridiplantae [3]. Sets of genes encoding subunits of notable organelle protein complexes
are labelled with grey bars under the horizontal axis. Full set of taxon abbreviations is in Supplementary Text; notable taxa are [metaz]oa,
[virid]iplantae, [fungi], [apico]mplexa, [jakob]ida, [rhodo]phyta. (B) Posterior probabilities over the set of features in linear models predicting
retention index. Each model structure is given by a set of codes describing its component features. Hydrophobicity (Hyd) or hydrophobicity
index (HydI) and GC content (GC) feature in all model structures with the highest posterior probabilities (for priors see Methods). +/−
give posterior mean signs of associated coefficients in model for retention index. Full feature list: [Hyd]rophobicity, [HydI] hydrophobicity
index, [GC] content, [Len]gth, [pK1] carboxyl pKa, [pk2] amino pKa, [MW] molecular weight, [AG/CW] energies of gene expression (Sup-
plementary Text). (C) Prediction of retention index with linear models involving hydrophobicity and GC content. oDNA gene sets are split
into training and test sets; trained models predict retention indices well in the independent test sets. (D) Cross-organelle prediction. Linear
models trained on mtDNA gene properties predict retention indices of ptDNA genes well, and vice versa.
3
versa. Strikingly, both organelle gene sets predicted
well the other’s retention patterns (ρ = 0.65 for pt pre-
dicting mt; ρ = 0.55 for mt predicting pt; Fig. 1D, Sup-
plementary Table S1). In other words, a simple model
trained only using mitochondrial gene data can predict
the retention profile of plastid genes, and vice versa.
To relax the assumptions involved in this analysis,
including linear modelling, we paralleled this analysis
with a range of other regression approaches from data
science, including penalised regression and random
forests, and using different definitions of retention in-
dex (Supplementary Text; Supplementary Fig.
S4).
We generally observed hydrophobicity and GC con-
tent being selected as features with good predictive
ability and the capacity to predict one oDNA type’s
behaviour from the other, regardless of statistical ap-
proach taken (Supplementary Table S1); pKa values
were also selected as informative features in some
model types (see below).
Hydrophobicity and protein biochemistry
predicts oDNA gene transfer to the nu-
cleus in both organelles
We next asked which properties predict which or-
ganelle protein-coding genes are universally trans-
ferred to the nucleus across all eukaryotes. To this
end, we compiled sets of annotated nDNA and oDNA
genes encoding subunits of bioenergetic protein com-
plexes in organelles using a custom pattern matching
algorithm and 308 eukaryotic whole genome records
from NCBI (see Methods) (Fig. 2A). As expected, GC
content in organelle-encoded genes was systemati-
cally lower than nuclear-encoded genes. Here, this
signal cannot be regarded as a causal mechanism,
because it is likely due at least in part to the aforemen-
tioned differences in asymmetric mutational pressure
between nDNA and oDNA [32, 19].
More interest-
ingly, the hydrophobicity of organelle-encoded genes
was systematically higher across taxa (agreeing with
recent observations in the mitoribosome [33]), and the
carboxyl pKa values of organelle-encoded genes were
also systematically higher; other features also differed
by encoding compartment (Supplementary Fig. S5).
We used Bayesian model selection with a generalised
linear model (GLM) using gene properties to predict
the encoding compartment (except GC and codon use
statistics, due to the possibility of differences therein
arising simply due to asymmetric mutation). We found
that hydrophobicity and carboxyl pKa consistently ap-
peared in all the model structures with highest pos-
terior probability.
Their appearance together in a
Bayesian model selection framework suggests that
they provide independent information on gene encod-
ing, despite a correlation (albeit rather weak) between
the features (Supplementary Fig.
S3).
GLMs us-
ing hydrophobicity and carboxyl pKa, trained using a
subset of genes from a given species, were able to
to predict the encoding compartment of an indepen-
dent test set from that species with high performance
(True Positive/Negative rates: mt TP 0.90 ± 0.17, TN
0.97 ± 0.10, pt TP 0.75 ± 0.20, TN 0.88 ± 0.18, mean
and s.d. across species). We also verified that these
differences existed within the sets of genes encoding
subunits of different organellar complexes (Fig. 2B).
We employed a range of classification approaches to
quantify these observations, again training on a sub-
set of the observations and testing classification per-
formance on an independent set (Supplementary Fig.
S12). Hydrophobicity and pKa values consistently ap-
peared as strong separating terms, with other features
including production energy and gene length playing a
supporting role (Supplementary Fig. S12). Classifica-
tion accuracy was typically > 0.8 for all complexes us-
ing random forest approaches (Supplementary Table
S4).
For a subset of organelle-localised gene products,
solved crystal structures of their protein complexes
allow another property to be quantified: the binding
energy statistics of the protein product in its protein
complex structure.
Previous work qualitatively sug-
gested that genes encoding subunits with high to-
tal binding energy (strong binding interactions with
neighbouring subunits) and playing central roles in
complex assembly pathways were most retained in
mtDNA [19, 34, 14].
We used a generalised lin-
ear mixed model to quantify and extend this analy-
sis to complexes in both organelle types. We found
that total binding energy predicted whether a gene
was organelle-encoded in any eukaryotes, with the re-
lationship holding across mitochondria and plastids,
though with varying magnitudes in different complexes
(Fig. 2C; Supplementary Fig. S7). We verified the
absence of pronounced correlation structure between
binding energy statistics and hydrophobicity (Supple-
mentary Fig. S8), suggesting that the two features in-
dependently contribute to gene retention [19]. Hence,
hydrophobicity, amino acid biochemistry, and ener-
getic centrality (linked to colocalisation for redox regu-
lation [14]) predict whether a gene is ever retained in
oDNA; of those that are, hydrophobicity and GC con-
tent predict the extent of this retention across eukary-
otes.
Independent
endosymbiotic
genomes
show compatible profiles of hydrophobic-
ity and protein biochemistry
Evolutionary history cannot easily be rerun to inde-
pendently examine these principles. However, the di-
versity of eukaryotic life provides some existing oppor-
tunities to test them. In several eukaryotic species,
unicellular endosymbionts that are not directly related
to mitochondria or plastids have co-evolved with their
‘host’ species, in many cases involving gene loss and
in some cases transfer of genes to the host. Class In-
secta are known to have several examples of reduced
bacterial endosymbionts [35]; other notable examples
include the chromatophore, an originally cyanobac-
terial endosymbiont of Paulinella freshwater amoe-
4
Figure 2:
Features predicting encoding compartment.
(A) Mean and s.e.m.
of selected gene properties for organelle genes
encoded in nuclear DNA (grey), mtDNA (red), and ptDNA (blue), in different species (organised by the phylogeny on the left, expanded
set in Supplementary Fig. S5). (B) Hydrophobicity and carboxyl pKa of organelle genes encoded in nuclear DNA (red) and oDNA (blue),
organised by the protein complex that the gene product occupies (expanded set in Supplementary Fig. S6). (C) Bayesian model selection
with a generalised linear model (GLM) framework for features predicting the encoding compartment of a given gene. Posterior probabilities
are averaged across independent classifications for individual organisms. Each model structure is given by a set of codes describing its
component features; model labels as in Fig. 1. (D) Performance (True/False Positive/Negative) of GLMs involving hydrophobicity and
carboxyl pKa on predicting encoding compartment of genes outside the training set. Each set of points corresponds to a model for one
organism. (E) Binding energy and encoding compartment. Traces show mean and 95% credible intervals for Bayesian generalised linear
mixed model (GLMM) (see Methods for priors). The associated p-value is a frequentist interpretation from bootstrapping, against the null
hypothesis of no relationship. Crystal structures are coloured according to the number of species in our dataset that retain the gene for
each subunit.
5
bae [36], the recently discovered Candidatus Azoam-
icus ciliaticola, a denitrifying gammaproteobacterial
endosymbiont within a Plagiopylea ciliate host [37],
and the Nostoc azollae symbiont of the Azolla water
ferns [38].
Not all of these endosymbiotic relationships have
been shown to involve gene transfer to the host cell
nucleus, although there is evidence for this in the
Paulinella system [39]. All cases do, however, involve
reduction of the endosymbiont genome, as some ma-
chinery in the endosymbiont becomes redundant in
the symbiotic relationship. In a subset of lost genes,
this redundancy arises because host-encoded ma-
chinery can fulfil the required function (other genes will
be lost without such host-encoded compensation, as
their entire function becomes redundant).
For this subset, the same broad principles regard-
ing import of protein machinery may then be expected
to hold as in organelles. Such genes are lost as host-
encoded machinery removes the need for their local
encoding. But such host-encoded machinery must be
physically acquired by the endosymbiont, raising sim-
ilar issues of the mistargeting and import difficulty for
hydrophobic gene products as in the organelle case.
In tandem, any biochemical pressures influencing the
ease of gene expression in the endosymbiont com-
partment may also be expected to shape retention
patterns of this subset of genes. We therefore hypoth-
esised that the principles we find to shape gene reten-
tion in mitochondria and plastids would also show a
detectable signal in these independent endosymbiotic
cases (while expecting a lower magnitude hydropho-
bicity signal, due to loss of some genes without the
requirement for nuclear compensation).
To test this hypothesis, we computed genetic statis-
tics for the genomes of endosymbionts and non-
endosymbiotic close relatives (Methods; Supplemen-
tary Table S2). The hydrophobicity profile of the en-
dosymbionts in 9 of 10 cases was significantly higher
than their non-endosymbiotic relative (Supplementary
Text; Fig.
3).
Genes retained in the photosyn-
thetic chromatophore also had lower carboxyl pKa val-
ues than in a free-living relative; for other endosym-
bionts, this relationship was reversed, with endosym-
biont genes having lower carboxyl pKa values. This is
compatible with a possible mechanistic link between
the pH of the compartment and the dynamics of gene
expression therein (see Discussion).
Our analysis approach involves several choices of
parameter and protocol.
To assess the robustness
of our findings, we have varied these choices and
checked the corresponding change in outputs, de-
scribed in Supplementary Text and the following fig-
ures. The key choices, with figures illustrating their
effects, are in gene annotation (supervised or unsu-
pervised; Supplementary Fig. S1), initial selection of
features (where we followed existing hypotheses and
particularly their summary in [19]) and how to sum-
marise their quantitative values (Supplementary Fig.
S9), definition of retention index (Supplementary Ta-
ble S1; Supplementary Fig. S10), choice of priors in
Figure 3: Gene feature profiles in other endosymbionts. Hy-
drophobicity and carboxyl pKa across genes in endosymbionts (red)
and a non-endosymbiotic close relative (blue). p-values are from
Wilcoxon rank-sum tests.
Bayesian model selection (Supplementary Fig. S11),
and choice of regression and classification methods:
we additionally tested LASSO and ridge regression,
and decision trees and random forests for regres-
sion and classification (Supplementary Figs. S10 and
S12).
Discussion
To summarise, we have found that hydrophobicity and
energetic centrality (the latter linked to colocalisation
for redox regulation [14]), with other features of nu-
cleic acid and amino acid biochemistry, predict the
prevalence of gene retention to a strikingly symmetric
extent in mitochondria, chloroplasts, and independent
endosymbionts. It must be underlined that no single
mechanism has sole predictive power over this be-
haviour. As expected in complex biological systems,
a combination of factors is likely at play, a situation
that has perhaps contributed to the ongoing debate
on this topic. Our findings support some previously
proposed mechanisms for how selective pressures on
gene content may be manifest, while not being incom-
patible with others (for example, recent theory on the
energetic costs of encoding and importing genes [21]).
Due to the physical difficulty of importing hydropho-
bic products or their propensity to be mistargeted to
other compartments, hydrophobic gene retention may
be favoured [15, 17] (though these mechanisms are
not free from debate [18]). The binding energy central-
ity of a subunit in its protein complex was suggested
as a proxy for control over complex assembly, and thus
redox processes, aligning with the CoRR (colocalisa-
tion for redox regulation) hypothesis [18].
GC con-
tent and carboxyl pKa have less established mech-
6
anistic hypotheses. The increased chemical stability
of GC bonds [40] has been suggested to support the
integrity of oDNA in the damaging chemical environ-
ment of the organelle. pKa, reflecting the ease of de-
protonation of amino acid subgroups for different pH
environments, influences the dynamics of peptide for-
mation in translation [41], resulting in pronounced and
diverse pH dependence of peptide formation for dif-
ferent amino acids [42].
Speculatively, we thus hy-
pothesise that the synthesis of protein products en-
riched for higher-pKa amino acids may involve lower
kinetic hurdles in the more alkaline pH of mitochon-
dria, plastids, and the chromatophore, favouring the
retention of the corresponding genes. The pH within
other endosymbionts, which perform less or no proton
pumping, is expected to be lower, in which case the
opposite pKa trend observed in Fig. 3 follows this pat-
tern. This harnessing of large-scale sequence data
with tools from model selection and machine learning
has thus generated, and tested, new understanding
of the fundamental evolutionary forces shaping bioen-
ergetic organelles, providing quantitative support for
several existing hypotheses and suggesting new con-
tributory mechanisms to this important process.
Materials and Methods
Source data. We used the mitochondrion and plas-
tid sequences available from NCBI RefSeq [43], and
annotated eukaryotic whole genome data also from
NCBI. The accessions and references for the en-
dosymbiont/relative pairs are given in Supplementary
Table S2. For biochemical and biophysical gene prop-
erties, we used the values from [19], described in the
Supplementary Text, using BioPython [44] to assign
these to given gene sequences. We averaged gene
statistics over representative species from a collec-
tion of diverse taxa, both using model species (Homo
sapiens, Arabidopsis thaliana, Saccharomyces cere-
visiae, Reclinomonas americana, Chondrus crispus,
Plasmodium falciparum) and randomly selected mem-
bers of different taxa (Supplementary Text; Supple-
mentary Fig.
S9).
We used crystal structures and
associated HTML descriptions from the PDB [45] ref-
erences 1oco, 1q90, 2h88, 2wsc, 5iu0, 5mdx, 5mlc,
5o31, 5xte, 6cp3, 6fkf. We used PDBePISA [46] to
estimate subunit binding energies with two different
protocols, both removing ligands and incorporating
them into the overall binding energy value for a sub-
unit (Supplementary Text). We used estimated tax-
onomies from NCBI’s Common Taxonomy Tree tool
[47].
Gene labelling and evolutionary transitions. Gene
annotations are inconsistent across such a diverse
dataset.
For organelle genomes, we used two ap-
proaches. In a supervised approach, where the full
set of unique labels found was manually curated and
assigned a ‘correct’ label based on biological knowl-
edge. In an unsupervised approach, we used BLASTn
to perform an all-against-all comparison of all genes in
our dataset. We scored each comparison as the pro-
portional length of the region of identity compared to
the reference sequence, multiplied by the proportion
of identities across that region. Scores over 0.75 were
interpreted as ‘hits’ (e.g. 75% identity over the full se-
quence, or full identity over 75% of the sequence). If
more than 25% of appearance of gene label X in the
BLAST output involved a ‘hit’ with gene labels Y, we
interpreted X and Y as referring to the same gene.
This process built a set of pairwise identities, which
we then resolved interatively into groups of gene la-
bels assumed to refer to the same gene. We then as-
signed the most prevalent gene label to all members of
that group. In each case, we retained only genes that
were present in more than ten species in our dataset.
For annotated whole genome data, we used pattern
matching for gene annotations based on regular ex-
pression identifiers to identify nuclear-encoded sub-
units of organellar protein complexes (expressions in
Supplementary Text).
Using these curated gene sets, we assigned ‘bar-
codes’ of gene presence/absence (binary strings of
length L, with 0 denoting gene absence and 1 denot-
ing gene presence) to each species in our dataset.
Each of these species is a tip on an estimated taxo-
nomic tree describing their putative evolutionary rela-
tionship. Assuming that gene loss is rare and gene
gain is very rare, we iteratively reconstructed parent
barcodes on this tree by assigning a 0 for gene X if
all descendants lack X, and 1 otherwise.
We then
identified parent-child pairs where the child barcode
had fewer genes than the parent (the opposite is im-
possible by construction). For each such instance, we
record the transition from parent barcode to child bar-
code as a loss event.
Retention indices. Our simple retention index is de-
fined as follows. Identify the set of transitions that in-
volve the loss of gene X. For each transition in this
set, count the genes retained by the parent and the
genes retained by the child, and take their mean. The
retention index is the mean of this quantity over the set
of transitions where X is lost. The rationale is to char-
acterise the number of genes that have already been
lost when X is lost. If a transition event involves only
the loss of X, the parent-child average will report this
number minus 1/2. If a transition involves the loss of
several other genes in parallel with X, the average of
the before and after counts is used. We also used an
alternative retention index without dependence on an
assumed evolutionary relationship, described in Sup-
plementary Text.
Prediction of retention index.
We used Bayesian
model selection with non-local priors to promote sep-
aration of overlapping models [48]; specifically, mo-
ment (MOM) priors parameterised so that a signal-
to-noise ratio of > 0.2 is anticipated, compatible with
previous findings [19]; a beta-binomial(1, 1) prior dis-
tribution on the model space, and a minimally infor-
mative inverse gamma prior for noise. Further prior
information, and the effects of varying them, are given
in Supplementary Text and Supplementary Fig. S11.
7
We implemented the selection process in the R pack-
age mombf. We additionally used linear modelling pe-
nalised using ridge and LASSO protocols, tree-based,
and random forest regression, described in the Sup-
plementary Text and implemented using glmnet, tree,
and randomForest packages.
Classification of subcellular encoding.
We used
Bayesian model averaging for generalised linear mod-
els (GLMs) predicting encoding compartments with
priors giving probability 1/2 for the inclusion of each
parameter, implemented in BMA. We then trained
GLMs involving hydrophobicity and carboxyl pKa on
a training subset of genes for each species. The train-
ing subset was the union of a random sample of half
the nuclear-encoded genes and half the organelle-
encoded genes in each species, with the test set be-
ing the complement of this set. We also used decision
tree and random forest approaches for the classifica-
tion task, described in the Supplementary Text. For
binding energy values, we used both a Bayesian GLM
treating all complexes independently, with t-distributed
priors with zero mean, implemented in arm; and a
Bayesian generalised linear mixed model with flat pri-
ors over coefficients, residuals, and covariance struc-
ture, implemented in blme. These priors were used to
overcome convergence issues given the perfect sep-
aration of datapoints observed for some protein com-
plexes. Complexes were visualised in PyMOL [49].
Code and dependencies.
Code is written in R,
Python, and C, with a wrapper script for bash, and is
freely available at github.com/StochasticBiology/
odna-loss. The list of libraries used and correspond-
ing citations are in the Supplementary Text.
References
[1] William F Martin, Sriram Garg, and Verena Zi-
morski. Endosymbiotic theories for eukaryote ori-
gin. Philosophical Transactions of the Royal Soci-
ety B: Biological Sciences, 370(1678):20140330,
2015.
[2] Nick Lane and William Martin. The energetics of
genome complexity. Nature, 467(7318):929–934,
2010.
[3] Martin
F
Hohmann-Marriott
and
Robert
E
Blankenship.
Evolution of photosynthesis.
An-
nual review of plant biology, 62:515–548, 2011.
[4] Austin Booth and W Ford Doolittle. Eukaryoge-
nesis, how special really?
Proceedings of the
National Academy of Sciences, 112(33):10278–
10285, 2015.
[5] Bastien Boussau, E Olof Karlberg, A Carolin
Frank, Boris-Antoine Legault, and Siv GE Ander-
sson. Computational inference of scenarios for
α-proteobacterial genome evolution.
Proceed-
ings of the National Academy of Sciences of the
United States of America, 101(26):9722–9727,
2004.
[6] Jeffrey L Blanchard and Michael Lynch. Organel-
lar genes: why do they end up in the nucleus?
Trends in genetics, 16(7):315–320, 2000.
[7] Keith L Adams and Jeffrey D Palmer. Evolution of
mitochondrial gene content: gene loss and trans-
fer to the nucleus. Molecular phylogenetics and
evolution, 29(3):380–395, 2003.
[8] Daniel B Sloan, Jessica M Warren, Alissa M
Williams, Zhiqiang Wu, Salah E Abdel-Ghany,
Adam J Chicco, and Justin C Havird. Cytonuclear
integration and co-evolution. Nature Reviews Ge-
netics, 19(10):635–648, 2018.
[9] Cecilia Saccone, Carmela Gissi, Cecilia Lanave,
Alessandra Larizza, Graziano Pesole, and Aure-
lio Reyes. Evolution of the mitochondrial genetic
system: an overview.
Gene, 261(1):153–159,
2000.
[10] John F Allen and John A Raven. Free-radical-
induced mutation vs redox regulation: costs and
benefits of genes in organelles. Journal of Molec-
ular Evolution, 42(5):482–492, 1996.
[11] Alan F Wright, Michael P Murphy, and Dou-
glass M Turnbull. Do organellar genomes func-
tion as long-term redox damage sensors? Trends
in Genetics, 25(6):253–261, 2009.
[12] Leslie S Itsara, Scott R Kennedy, Edward J Fox,
Selina Yu, Joshua J Hewitt, Monica Sanchez-
Contreras, Fernando Cardozo-Pelaez, and Leo J
Pallanck. Oxidative stress is not a major contribu-
tor to somatic mitochondrial dna mutations. PLoS
genetics, 10(2):e1003974, 2014.
[13] Scott R Kennedy, Jesse J Salk, Michael W
Schmitt, and Lawrence A Loeb. Ultra-sensitive
sequencing reveals an age-related increase in
somatic mitochondrial mutations that are incon-
sistent with oxidative damage.
PLoS genetics,
9(9):e1003794, 2013.
[14] John F Allen and William F Martin.
Why have
organelles retained genomes?
Cell systems,
2(2):70–72, 2016.
[15] Gunnar Von Heijne. On the hydrophobic nature of
signal sequences. European journal of biochem-
istry, 116(2):419–422, 1981.
[16] Jean-Luc Popot and C de Vitry.
On the mi-
croassembly of integral membrane proteins. An-
nual review of biophysics and biophysical chem-
istry, 19(1):369–403, 1990.
[17] Patrik
Bj¨orkholm,
Andreas
M
Ernst,
Erik
Hagstr¨om, and Siv GE Andersson.
Why mito-
chondria need a genome revisited. FEBS letters,
591(1):65–75, 2017.
8
[18] John F Allen. Why chloroplasts and mitochondria
retain their own genomes and genetic systems:
colocation for redox regulation of gene expres-
sion.
Proceedings of the National Academy of
Sciences, 112(33):10231–10238, 2015.
[19] Iain G Johnston and Ben P Williams. Evolution-
ary inference across eukaryotes identifies spe-
cific pressures favoring mitochondrial gene reten-
tion. Cell Systems, 2(2):101–111, 2016.
[20] Benoit Nabholz, Hans Ellegren, and Jochen BW
Wolf.
High levels of gene expression explain
the strong evolutionary constraint of mitochon-
drial protein-coding genes. Molecular biology and
evolution, 30(2):272–284, 2012.
[21] Steven Kelly.
The economics of endosymbi-
otic gene transfer and the evolution of organellar
genomes. bioRxiv, 2020.
[22] William Martin and Claus Schnarrenberger. The
evolution of the calvin cycle from prokaryotic to
eukaryotic chromosomes: a case study of func-
tional redundancy in ancient pathways through
endosymbiosis.
Current genetics, 32(1):1–18,
1997.
[23] Kacper Maciszewski and Anna Karnkowska.
Should i stay or should i go?
retention and
loss of components in vestigial endosymbiotic or-
ganelles. Current opinion in genetics & develop-
ment, 58:33–39, 2019.
[24] Paul Kirk, Thomas Thorne, and Michael PH
Stumpf.
Model selection in systems and syn-
thetic biology. Current opinion in biotechnology,
24(4):767–774, 2013.
[25] Paul Nurse. Biology must generate ideas as well
as data. Nature, 597(7876):305, 2021.
[26] David Roy Smith and Patrick J Keeling. Mitochon-
drial and plastid genome architecture: reoccur-
ring themes, but significant differences at the ex-
tremes. Proceedings of the National Academy of
Sciences, 112(33):10177–10184, 2015.
[27] Lucia Hadariov´a, Matej Vesteg, Vladim´ır Hampl,
and Juraj Krajˇcoviˇc.
Reductive evolution of
chloroplasts in non-photosynthetic plants, algae
and protists.
Current genetics, 64(2):365–387,
2018.
[28] Jan Janouˇskovec, Denis V Tikhonenkov, Fa-
bien Burki, Alexis T Howe, Forest L Rohwer,
Alexander P Mylnikov, and Patrick J Keeling. A
new lineage of eukaryotes illuminates early mi-
tochondrial genome reduction. Current Biology,
27(23):3717–3724, 2017.
[29] Iain G Johnston.
Tension and resolution:
dynamic,
evolving
populations
of
organelle
genomes within plant cells.
Molecular plant,
12(6):764–783, 2019.
[30] Sam F Greenbury,
Mauricio Barahona,
and
Iain G Johnston.
Hypertraps: Inferring proba-
bilistic patterns of trait acquisition in evolution-
ary and disease progression pathways. Cell sys-
tems, 10(1):39–51, 2020.
[31] Tapan Kumar Mohanta, Awdhesh Kumar Mishra,
Adil
Khan,
Abeer
Hashem,
Elsayed
Fathi
Abd Allah, and Ahmed Al-Harrasi. Gene loss and
evolution of the plastome. Genes, 11(10):1133,
2020.
[32] Aurelio Reyes, Carmela Gissi, Graziano Pesole,
and Cecilia Saccone.
Asymmetrical directional
mutation pressure in the mitochondrial genome
of mammals. Molecular Biology and Evolution,
15(8):957–966, 1998.
[33] Lea Bertgen, Timo M¨uhlhaus, and Johannes M
Herrmann. Clingy genes: Why were genes for ri-
bosomal proteins retained in many mitochondrial
genomes? Biochimica et Biophysica Acta (BBA)-
Bioenergetics, page 148275, 2020.
[34] Uwe-G Maier, Stefan Zauner, Christian Woehle,
Kathrin Bolte, Franziska Hempel, John F Allen,
and William F Martin. Massively convergent evo-
lution for ribosomal protein gene content in plas-
tid and mitochondrial genomes. Genome biology
and evolution, 5(12):2318–2329, 2013.
[35] Filip Husnik and Patrick J Keeling. The fate of ob-
ligate endosymbionts: reduction, integration, or
extinction. Current opinion in genetics & devel-
opment, 58:1–8, 2019.
[36] Arwa Gabr, Arthur R Grossman, and Debashish
Bhattacharya.
Paulinella, a model for under-
standing plastid primary endosymbiosis. Journal
of phycology, 56(4):837–843, 2020.
[37] Jon S Graf, Sina Schorn, Katharina Kitzinger, So-
eren Ahmerkamp, Christian Woehle, Bruno Huet-
tel, Carsten J Schubert, Marcel MM Kuypers, and
Jana Milucka.
Anaerobic endosymbiont gener-
ates energy for ciliate host by denitrification. Na-
ture, 591(7850):445–450, 2021.
[38] Liang
Ran,
John
Larsson,
Theoden
Vigil-
Stenman, Johan AA Nylander, Karolina Inin-
bergs, Wei-Wen Zheng, Alla Lapidus, Stephen
Lowry, Robert Haselkorn, and Birgitta Bergman.
Genome
erosion
in
a
nitrogen-fixing
verti-
cally
transmitted
endosymbiotic
multicellular
cyanobacterium. PLoS One, 5(7):e11486, 2010.
[39] Eva CM Nowack, Heiko Vogel, Marco Groth,
Arthur R Grossman, Michael Melkonian, and
Gernot Gl¨ockner.
Endosymbiotic gene trans-
fer and transcriptional regulation of transferred
genes in paulinella chromatophora. Molecular bi-
ology and evolution, 28(1):407–422, 2011.
9
[40] David C Samuels.
Life span is related to the
free energy of mitochondrial dna. Mechanisms
of ageing and development, 126(10):1123–1129,
2005.
[41] R Edward Watts and Anthony C Forster. Chem-
ical models of peptide formation in translation.
Biochemistry, 49(10):2177–2185, 2010.
[42] Magnus Johansson, Ka-Weng Ieong, Stefan Tro-
bro, Peter Strazewski, Johan ˚Aqvist, Michael Y
Pavlov, and M˚ans Ehrenberg. ph-sensitivity of the
ribosomal peptidyl transfer reaction dependent
on the identity of the a-site aminoacyl-trna. Pro-
ceedings of the National Academy of Sciences,
108(1):79–84, 2011.
[43] Nuala A O’Leary, Mathew W Wright, J Rod-
ney Brister, Stacy Ciufo, Diana Haddad, Rich
McVeigh, Bhanu Rajput, Barbara Robbertse,
Brian Smith-White, Danso Ako-Adjei, et al. Refer-
ence sequence (refseq) database at ncbi: current
status, taxonomic expansion, and functional an-
notation. Nucleic acids research, 44(D1):D733–
D745, 2016.
[44] Peter JA Cock, Tiago Antao, Jeffrey T Chang,
Brad A Chapman, Cymon J Cox, Andrew Dalke,
Iddo Friedberg, Thomas Hamelryck, Frank Kauff,
Bartek Wilczynski, et al.
Biopython:
freely
available python tools for computational molec-
ular biology and bioinformatics. Bioinformatics,
25(11):1422–1423, 2009.
[45] Helen M Berman,
John Westbrook,
Zukang
Feng, Gary Gilliland, Talapady N Bhat, Helge
Weissig, Ilya N Shindyalov, and Philip E Bourne.
The protein data bank. Nucleic acids research,
28(1):235–242, 2000.
[46] Sameer Velankar,
Christoph Best,
B Beuth,
CH Boutselakis, N Cobley, AW Sousa Da Silva,
Dimitris Dimitropoulos, Adel Golovin, Miriam Hir-
shberg, Melford John, et al. Pdbe: protein data
bank in europe.
Nucleic acids research, page
gkp916, 2009.
[47] Scott Federhen. The ncbi taxonomy database.
Nucleic
acids
research,
40(D1):D136–D143,
2012.
[48] Valen E Johnson and David Rossell. On the use
of non-local prior densities in bayesian hypothe-
sis tests. Journal of the Royal Statistical Society:
Series B (Statistical Methodology), 72(2):143–
170, 2010.
[49] Warren L DeLano et al. Pymol: An open-source
molecular graphics tool.
CCP4 Newsletter on
protein crystallography, 40(1):82–92, 2002.
Acknowledgments
LR and JMC are supported by the BBSRC via the
MIBTP Doctoral Training Scheme. This project has re-
ceived funding from the European Research Council
(ERC) under the European Union’s Horizon 2020 re-
search and innovation programme (Grant agreement
No. 805046 (EvoConBiO) to IGJ).
10
Supplementary Text
Materials & Methods
Source data. We used the mitochondrion and plastid sequences available from NCBI RefSeq [1], and an-
notated eukaryotic whole genome data also from NCBI. The accessions and references for the endosym-
biont/relative pairs are given in Supplementary Table S2. For biochemical and biophysical gene properties, we
used the values from [2], described in the Supplementary Text, using BioPython [3] to assign these to given
gene sequences. We averaged gene statistics over representative species from a collection of diverse taxa,
both using model species (Homo sapiens, Arabidopsis thaliana, Saccharomyces cerevisiae, Reclinomonas
americana, Chondrus crispus, Plasmodium falciparum) and randomly selected members of different taxa (Sup-
plementary Text; Supplementary Fig. S9). Codes used in the figures are [Hyd]rophobicity, [HydI] hydropho-
bicity index, [GC] content, [Len]gth, [pK1] carboxyl pKa, [pK2] amino pKa, [MW] molecular weight, [AG/CW]
energies of gene expression. We used crystal structures and associated HTML descriptions from the PDB
[4] references 1oco, 1q90, 2h88, 2wsc, 5iu0, 5mdx, 5mlc, 5o31, 5xte, 6cp3, 6fkf. We used PDBePISA [5] to
estimate subunit binding energies with two different protocols, both removing ligands and incorporating them
into the overall binding energy value for a subunit (Supplementary Text). We used estimated taxonomies from
NCBI’s Common Taxonomy Tree tool [6].
Gene labelling and evolutionary transitions.
Gene annotations are inconsistent across such a diverse
dataset. For organelle genomes, we used two approaches. In a supervised approach, where the full set
of unique labels found was manually curated and assigned a ‘correct’ label based on biological knowledge.
In an unsupervised approach, we used BLASTn to perform an all-against-all comparison of all genes in our
dataset. We scored each comparison as the proportional length of the region of identity compared to the refer-
ence sequence, multiplied by the proportion of identities across that region. Scores over 0.75 were interpreted
as ‘hits’ (e.g. 75% identity over the full sequence, or full identity over 75% of the sequence). If more than 25%
of appearance of gene label X in the BLAST output involved a ‘hit’ with gene labels Y, we interpreted X and Y
as referring to the same gene. This process built a set of pairwise identities, which we then resolved interatively
into groups of gene labels assumed to refer to the same gene. We then assigned the most prevalent gene
label to all members of that group. In each case, we retained only genes that were present in more than ten
species in our dataset. For annotated whole genome data, we used pattern matching for gene annotations
based on regular expression identifiers to identify nuclear-encoded subunits of organellar protein complexes
(expressions in Supplementary Text).
Using these curated gene sets, we assigned ‘barcodes’ of gene presence/absence (binary strings of length
L, with 0 denoting gene absence and 1 denoting gene presence) to each species in our dataset. Each of these
species is a tip on an estimated taxonomic tree describing their putative evolutionary relationship. Assuming
that gene loss is rare and gene gain is very rare, we iteratively reconstructed parent barcodes on this tree by
assigning a 0 for gene X if all descendants lack X, and 1 otherwise. We then identified parent-child pairs where
the child barcode had fewer genes than the parent (the opposite is impossible by construction). For each such
instance, we record the transition from parent barcode to child barcode as a loss event.
Retention indices. Our simple retention index is defined as follows. Identify the set of transitions that involve
the loss of gene X. For each transition in this set, count the genes retained by the parent and the genes
retained by the child, and take their mean. The retention index is the mean of this quantity over the set of
transitions where X is lost. The rationale is to characterise the number of genes that have already been lost
when X is lost. If a transition event involves only the loss of X, the parent-child average will report this number
minus 1/2. If a transition involves the loss of several other genes in parallel with X, the average of the before
and after counts is used. We also used an alternative retention index without dependence on an assumed
evolutionary relationship, described in Supplementary Text.
Prediction of retention index. We used Bayesian model selection with non-local priors to promote separation
of overlapping models [7]; specifically, moment (MOM) priors parameterised so that a signal-to-noise ratio of
> 0.2 is anticipated, compatible with previous findings [2]; a beta-binomial(1, 1) prior distribution on the model
space, and a minimally informative inverse gamma prior for noise. Further prior information, and the effects of
varying them, are given in Supplementary Text and Supplementary Fig. S11. We implemented the selection
process in the R package mombf. We additionally used linear modelling penalised using ridge and LASSO
protocols, tree-based, and random forest regression, described in the Supplementary Text and implemented
using glmnet, tree, and randomForest packages.
Classification of subcellular encoding. We used Bayesian model averaging for generalised linear models
(GLMs) predicting encoding compartments with priors giving probability 1/2 for the inclusion of each parameter,
implemented in BMA. We then trained GLMs involving hydrophobicity and carboxyl pKa on a training subset
of genes for each species. The training subset was the union of a random sample of half the nuclear-encoded
genes and half the organelle-encoded genes in each species, with the test set being the complement of
this set. We also used decision tree and random forest approaches for the classification task, described in
11
Method
MT training
MT test
PT training
PT test
PT predicting MT
MT predicting PT
LM (simple)
0.64
0.63
0.62
0.60
0.65
0.55
LM-pruned (simple)
0.73
0.71
0.72
0.72
0.68
0.50
LM (barcode)
0.71
0.69
0.58
0.56
0.72
0.59
LM-pruned (barcode)
0.71
0.70
0.64
0.64
0.67
0.51
Table S1: Mean linear model regression performance (Spearman’s ρ between predicted and observed indices)
predicting retention index in test sets for different cases. Non-standard genes (msh1/muts, matr, mttb) are
removed from mtDNA sets for these experiments. Labels show simple retention index vs barcode retention
index; ‘pruned’ dataset (retaining only mt genes from families nad, sdh, atp, cox, cytb, rp and pt from psa, psb,
rp, rbc, ndh, atp, pet) vs unpruned. Each LM uses only GC content and hydrophobicity.
the Supplementary Text. For binding energy values, we used both a Bayesian GLM treating all complexes
independently, with t-distributed priors with zero mean, implemented in arm; and a Bayesian generalised linear
mixed model with flat priors over coefficients, residuals, and covariance structure, implemented in blme. These
priors were used to overcome convergence issues given the perfect separation of datapoints observed for
some protein complexes. Complexes were visualised in PyMOL [8].
Code and dependencies. Code is written in R, Python, and C, with a wrapper script for bash, and is freely
available at github.com/StochasticBiology/odna-loss. The list of libraries used and corresponding citations
are in the Supplementary Text.
Taxon abbreviations
Eukaryotic clades in the mitochondrial dataset in Fig.
1 are [apico]mplexa,
[bacill]ariophyta,
[bi-
gyr]a,
[cerco]zoa,
[chatto]nellaceae,
[crypto]phyceae,
[disco]sea,
[eumyc]etozoa,
[eusti]gmatophyceae,
[fungi],
[glauco]cystophyceae,
[hapto]phyta,
[heter]olobosea,
[jakob]ida,
[malaw]imonas,
[metaz]oa,
[oligo]hymenophorea, [oomyc]ota, [phaeo]phyceae, [rhodo]phyta, [virid]iplantae.
Clades in the plastid
dataset are [apico]mplexa, [bacill]ariophyta, [chlora]rachniophyceae, [crypto]phyceae, [dicty]ochophyceae,
[dinop]hyceae, [eugle]nida, [eusti]gmatophyceae, [glauc]ocystophyceae, [hapto]phyta, [mallo]monadaceae,
[pelag]omonadales, [phaeo]phyceae, [rhodo]phyta, [virid]iplantae.
Alternative retention index definitions
In addition to our simple retention index, which relies on an estimated phylogeny linking observations in our
dataset, we considered another assumption-free index. Here, we construct the set of unique oDNA pres-
ence/absence patterns in our dataset (as in Fig. 1A), and simply count the occurrences ci of each gene i in
this dataset. The index is given by log ci/ maxj log cj. This index relies on no evolutionary assumptions, and
thus cannot account for the evolutionary relationship between sampled species. Considering only the set of
unique barcodes goes some way towards accounting for the sampling bias in the dataset (for example, almost
all metazoans have the same presence/absence profile, but this profile will only occur once in the unique set).
The distribution of this index had substantial structure (as visible in Fig. 1A, and clear, particularly for plastids,
in Supplementary Fig. S10), but we do not consider further transformations or more tailored analysis here,
instead focusing on the similarity of results with those from the other index.
Biochemical and biophysical properties of genes and products
Our assignment of biochemical and biophysical properties of genes and their products follows that in
Ref.
[2].
The length* (in number of amino acids of gene product) and GC content (trivially counted)
of genes are taken straightforwardly from a sequence.
Chemical properties of amino acids were taken
from the compilation at http://www.sigmaaldrich.com/life-science/metabolomics/learning-center/
amino-acid-reference-chart.html. The hydrophobicity and hydrophobicity index of a gene product was
computed using this compilation (original data from Ref. [9]). Amine group pKa, carboxyl group pKa, and
molecular weight* values were calculated using this compilation (original data from [10]).
Glucose energy costs* were computed using the Aglucose metric, based on the absolute nutrient cost re-
quired for amino acid biosynthesis, from Ref. [11]. Craig-Weber energy costs*, estimating the number of
high-energy phosphate bonds and reducing hydrogen atoms required from the cellular energy pool to produce
an amino acid, were taken from Ref. [12]. These biochemical properties are summarised in Supplementary
Table S5.
12
Figure S1: Correlation between gene counts across species derived using manual and BLAST labelling ap-
proaches. r = 0.9999 for mitochondrial and r = 0.9849 for plastid data; discrepancies are largely down to a
small number of outliers.
Asterisks denote properties that are not averaged over gene length; it was deemed more appropriate to
average other properties over genome length to gain a representative measure. To check for artefacts from
this interpretation, we performed a (much more computationally demanding) model selection process including
both the normalised and un-normalised values for each property; although coverage of individual models was
unavoidably low in this procedure, the same consistent observation of GC content and hydrophobicity as
important features was observed throughout.
To compute a single value for each statistic of interest, a protocol is required to summarise the many differ-
ent values seen for a given gene across the species in our dataset. For robustness, we considered several
different averaging protocols. First, we averaged gene statistics over a set of model species taken from diverse
eukaryotic groups (Homo sapiens, Arabidopsis thaliana, Saccharomyces cerevisiae, Reclinomonas ameri-
cana, Chondrus crispus, Plasmodium falciparum). Second, we randomly selected a member of each clade
branching from the eukaryotic group (see clade names above) and averaged over the set containing these
random samples. Most statistics were very strongly correlated for these different choices (Fig. S9A). The
exception was GC content, which is well known to evolve differently in different clades. To assess the effect
of this difference, we ran the model selection process in the text with randomly-sampled averaging protocols.
We found that despite differences in GC statistics, the selected models, and the presence of GC within them,
remained robust to averaging choice (Fig. S9B).
Regression for retention index
In addition to the Bayesian linear model approach described in the text, we used a variety of different ap-
proaches for retention index regression.
These included decision linear modelling with ridge and LASSO
penalisation, decision tree regression, and random forest regression. The training, test, and cross-organelle
performance of these approaches is given in Table S3.
Pattern matching for nuclear-encoded organelle genes
We used a combination of positive and negative pattern matching with regular expressions to identify annota-
tions for genes encoding subunits of different organelle complexes. The positive matches required were:
CI
/NADH dehydrogenase|[Uu]biquinone oxidoreductase/
CII
/[Ss]uccinate dehydrogenase|[cC]o[qQ] reductase/
CIII
/[Cc]ytochrome [Bb]|[Cc]ytochrome [Cc] reductase/
CIV
/[Cc]ytochrome [cC] oxidase/
CV
/[Aa][Tt][Pp] synthase|ATPase sub/
MitoRibo
/[Rr]ibosomal.*[Mm]itochondri/
PSI
/[Pp]hotosystem I /
PSII
/[Pp]hotosystem II /
Cytb6f
/[Cc]ytochrome [Bb]6|[Cc]ytochrome f|[Pp]lastocyanin reductase/
Rubisco
/bi.phosphate [Cc]arboxylase/
PlastRibo
/[Rr]ibosomal.*[cC]hloroplast/
13
Figure S2: Taxonomic trees for the mt and pt datasets. Blue diamonds give truncation points; associated taxa
are expanded in the next rightward tree. Truncated taxa are broadly chosen to reflect those with less diversity
in oDNA. Bars illustrate number of retained organelle genes in each species (scale differs in each subtree).
Figure S3: Linear correlations between genetic features and retention index, for mt and pt genes.
14
Endosymbiont
NCBI accession
Free-living relative
NCBI accession
References
Nasuia deltocephalinicola
CP013211.1
Herbaspirillum seropedicae
CP002039.1
[14]
Ca. Sulcia muelleri
CP001981.1
Porphyromonas gingivalis1
AE015924.1
[15]
Ca. Tremblaya phenacola
CP003982.1
Sodalis praecaptivus
CP006569.1
[16]
Rhopalodia gibberula SB
AP018341.1
Cyanothece sp. PCC 8801
CP001287.1
[17]
Ca. Hodgkinia cicadicola
CP008699
Rhizobium etli
CP007641.1
[18]
Ca. Pinguicoccus supinus
CP039370.1
Coraliomargarita akajimensis2
CP001998.1
[19]
Ca. Fokinia solitaria
CP025989.1
Pelagibacter ubique3
CP000084.1
[20]
Paulinella chromatophore
CP000815.1
Synechococcus PCC 7002
CP000951
[21]
Ca. Azoamicus ciliaticola
NZ LR794158.1
Legionella clemsonensis4
NZ CP016397
[22]
Nostoc azollae
CP002059.1
Raphidiopsis brookii
ACYB01000001.1
[23]
Table S2: Independent endosymbionts and close free-living relatives. SB, spherical body.
1 Relative does
invade cells but can survive in oral cavity. 2 Partner is not closest sequence found, but is closest annotated
sequence in putative phylogeny.
3 All closest relatives are intracellular Rickettsiales – relative taken from a
sister group. 4 Most relatives, including Legionella, are largely intracellular.
With the following patterns (split for formatting) required to be absent:
/assembly|alternative|containing|dependent|chaperone|kinase|NADH-cytochrome|coupling|maturase/
/vacuolar|biogenesis|repair|LOW QUALITY PROTEIN|synthetase|activator|reticulum|activase/
/synthesis|lyase|like| non|transporting|lipid|autoinhibited|membrane|type|required/
/QUALITY|precursor|inhibitor|proteasomal|proteasome|E1|various|regulatory|Clp/
/calcium|vesicle|b-245|b5|WRNIP|AAA|Cation|family|remodelling/
The outputs of this approach were manually verified to include genes encoding subunits physically present in
their corresponding complex, while excluding assembly factors, regulatory factors, synthesis factors, unrelated
enzymes, and other false positives.
Classification for compartment
We also considered decision tree and random forest approaches for the organelle/nuclear encoding compart-
ment classification problem; performance is shown in Table S4, with illustrations in Fig. S12.
Binding energy calculations
We used PDBePISA [5] to calculate interaction energies between different protein subunits and ligands in
crystal structures. We summed the interaction energies over all interfaces between a given subunit and its
partners to compute a total energetic centrality statistic for each subunit. Several choices of representation are
possible for these calculations. Ligands can be ignored, so that only interaction energies of interfaces directly
linking protein subunits are considered. Alternatively, bonds to ligands can be included as contributing to a
given subunit’s total binding energy. We primarily considered the mean energy per interface, including ligands,
for each subunit, but also verified that our detected relationship existed for different choices including total
energy over interfaces.
Endosymbionts and relatives
We considered a range of endosymbionts highlighted in a comprehensive recent review [13]. For each we
sought to identify a close free-living relative. In some cases all closest relatives of an endosymbiont themselves
adopted a largely or obligate intracellular lifestyle; in these cases we tried to identify the closest relative that
was at least capable of free-living (Table S2).
Packages and libraries
Our pipeline uses the following R packages: ape [24], arm [25], blme [26], BMA [27], caper [28], cowplot [29],
e1071 [30], geiger [31], GGally [32], ggnewscale [33], ggplot2 [34], ggpubr [35], ggpval [36], ggrepel [37],
ggtree [38], ggtreeExtra [39], glmnet [40], gridExtra [41], hexbin [42], igraph [43], lme4 [44], logistf [45], mombf
[46], nlme [47], phangorn [48], phytools [49], randomForest [50], stringdist [51], stringr [52], and tree [53].
We also use BioPython [3] for parsing sequences and computing gene statistics, PyMOL [8] for visualisation,
and BLAST [54] for sequence comparisons.
15
Method
MT training
MT test
PT training
PT test
PT predicting MT
MT predicting PT
Tree
0.79
0.40
0.82
0.45
0.54
0.33
LM
0.70
0.43
0.71
0.66
0.52
0.25
Tree-reduced
0.73
0.48
0.75
0.45
0.55
0.39
LM-Reduced
0.58
0.52
0.61
0.61
0.54
0.48
Ridge
0.68
0.39
0.66
0.71
0.57
0.41
LASSO
0.63
0.44
0.66
0.71
0.57
0.37
SVR
0.81
0.46
0.77
0.62
0.62
0.34
RF
0.92
0.48
0.95
0.62
0.62
0.45
RF-Reduced
0.88
0.50
0.92
0.51
0.57
0.50
RF-Cross
0.94
N/A
0.96
N/A
0.62
0.56
RF-Cross-Reduced
0.90
N/A
0.92
N/A
0.55
0.59
Table S3: Mean regression performance (Spearman’s ρ between predicted and observed indices) predicting
retention index with different approaches. Non-standard genes (msh1/muts, matr, mttb) are not removed for
these experiments. Tree, decision tree regression; LM, linear model; Ridge, ridge regression; LASSO, LASSO
regression; RF, random forest regression. All genetic features included by default; ‘reduced’ corresponds to
models involving only GC content and hydrophobicity. ‘Cross’ refers to cross-organelle experiments where mt
training is used to predict pt test and vice versa (N/A, not applicable: no test set within training organelle).
Complex
Model type
Training
Test
Balance
Complex
Model type
Training
Test
Balance
nad[0-9]
tree
0.99
0.99
0.10
nad[0-9]
RF
1.00
1.00
0.10
sdh[0-9]
tree
0.97
0.91
0.66
sdh[0-9]
RF
1.00
0.95
0.68
cytb
tree
0.99
0.99
0.18
cytb
RF
1.00
0.99
0.18
cox[0-9]
tree
1.00
0.99
0.09
cox[0-9]
RF
1.00
0.99
0.09
atp[0-9]
tree
0.98
0.96
0.16
atp[0-9]
RF
1.00
0.98
0.16
(MT) rp[sl]
tree
0.88
0.85
0.69
(MT) rp[sl]
RF
1.00
0.92
0.69
psa[a-x]
tree
0.99
0.99
0.03
psa[a-x]
RF
1.00
0.99
0.03
psb[a-z]
tree
1.00
0.99
0.01
psb[a-z]
RF
1.00
1.00
0.01
atp[a-z]
tree
0.98
0.97
0.12
atp[a-z]
RF
1.00
0.99
0.12
pet[a-z]
tree
1.00
0.99
0.01
pet[a-z]
RF
1.00
0.99
0.01
rbc
tree
0.99
0.97
0.07
rbc
RF
1.00
0.98
0.07
(PT) rp[sl]
tree
0.99
0.99
0.02
(PT) rp[sl]
RF
1.00
0.99
0.02
nad[0-9]
tree-reduced
0.99
0.99
0.10
nad[0-9]
RF-reduced
1.00
0.99
0.10
sdh[0-9]
tree-reduced
0.97
0.92
0.66
sdh[0-9]
RF-reduced
1.00
0.93
0.66
cytb
tree-reduced
0.98
0.97
0.18
cytb
RF-reduced
1.00
0.98
0.19
cox[0-9]
tree-reduced
0.98
0.98
0.09
cox[0-9]
RF-reduced
1.00
0.98
0.09
atp[0-9]
tree-reduced
0.92
0.91
0.16
atp[0-9]
RF-reduced
1.00
0.92
0.16
(MT) rp[sl]
tree-reduced
0.79
0.76
0.69
(MT) rp[sl]
RF-reduced
1.00
0.77
0.69
psa[a-x]
tree-reduced
0.98
0.97
0.03
psa[a-x]
RF-reduced
1.00
0.97
0.03
psb[a-z]
tree-reduced
0.99
0.99
0.01
psb[a-z]
RF-reduced
1.00
0.99
0.01
atp[a-z]
tree-reduced
0.91
0.90
0.12
atp[a-z]
RF-reduced
1.00
0.91
0.12
pet[a-z]
tree-reduced
0.99
0.99
0.01
pet[a-z]
RF-reduced
1.00
0.99
0.01
rbc
tree-reduced
0.96
0.93
0.06
rbc
RF-reduced
1.00
0.94
0.07
(PT) rp[sl]
tree-reduced
0.98
0.98
0.02
(PT) rp[sl]
RF-reduced
1.00
0.98
0.02
All PT
tree-cross
0.94
0.80
N/A
All PT
RF-cross
1.00
0.60
N/A
All MT
tree-cross
0.98
0.82
N/A
All MT
RF-cross
1.00
0.79
N/A
All PT
tree-cross-reduced
0.94
0.56
N/A
All PT
RF-cross-reduced
1.00
0.47
N/A
All MT
tree-cross-reduced
0.97
0.81
N/A
All MT
RF-cross-reduced
1.00
0.82
N/A
Table S4: Nuclear-organelle classification performance (proportion of test set assigned to correct compart-
ment), by organelle complex, with different approaches (tree, decision tree; RF, random forest). Complexes
are labelled with regular expressions describing their gene labels. All genetic features included by default; ‘re-
duced’ corresponds to models involving only GC content and hydrophobicity. ‘Cross’ refers to cross-organelle
experiments where mt training is used to predict pt test and vice versa. Balance gives the proportion of genes
encoded in one compartment (may fluctuate slightly due to different subsamples being used in model con-
struction): N/A, not applied to cross-organelle classification.
16
Figure S4: Decision tree and random forest regression for retention index. (top) a set of trees learned to
predict retention for different training-test splits, showing the dominant role of GC content and hydrophobicity
as predictive features. (bottom) variance improvement plots for random forest regression of the same task,
illustrating the importance of each feature in the predictive outcome.
17
Figure S5: Statistics of genes encoded in the nucleus (red), mitochondrion (blue), or plastid (green) compart-
ments. Bars give mean and s.e.m. for each species; phylogeny shows the relationship between species. Spe-
cific model species labelled by initials: Danio rerio, Homo sapiens, Drosophila melanogaster, Caenorhabditis
elegans, Glycine max, Arabidopsis thaliana, Physcomitrella patens, Schizosaccharomyces pombe, Plasmod-
ium falciparum, Dictyostelium discoideum.
Hydro
Hydro I
Mol weight / Da
pKa1
pKa2
Aglucose
CWEnergy
Ala
A
41
3
89.1
2.34
9.69
0.5
12.5
Arg
R
-14
1
174.2
2.17
9.04
1.39
18.5
Asn
N
-28
1
132.12
2.02
8.8
0.79
4
Asp
D
-55
1
133.11
1.88
9.6
0.61
1
Cys
C
49
3
121.16
1.96
10.28
0.75
24.5
Gln
Q
-10
2
146.15
2.17
9.13
0.92
9.5
Glu
E
-31
1
147.13
2.19
9.67
0.86
8.5
Gly
G
0
2
75.07
2.34
9.6
0.31
14.5
His
H
8
2
155.16
1.82
9.17
1.46
33
Ile
I
99
4
131.18
2.36
9.6
1.21
20
Leu
L
97
4
131.18
2.36
9.6
1.21
33
Lys
K
-23
1
146.19
2.18
8.95
1.31
18.5
Met
M
74
4
149.21
2.28
9.21
1.25
18.5
Phe
F
100
4
165.19
1.83
9.13
1.84
63
Pro
P
-46
1
115.13
1.99
10.6
0.99
12.5
Ser
S
-5
2
105.09
2.21
9.15
0.49
15
Stop
X
-
-
-
-
-
-
-
Thr
T
13
2
119.12
2.09
9.1
0.69
6
Trp
W
97
4
204.23
2.83
9.39
2.39
78.5
Tyr
Y
63
3
181.19
2.2
9.11
1.77
56.5
Val
V
76
4
117.15
2.32
9.62
0.96
25
Table S5: Amino acid properties used in model selection. Numerical values of the properties described in
the text. Qauantities are unitless unless specific. See text for sources.
18
Figure S6: Hydrophobicity and carboxyl pKa for nuclear- and organelle-encoded complex subunits.
Figure S7: Comparison of Bayesian generalised linear model (GLM) and generalised linear mixed model
(GLMM) for binding energy-retention relationship. The GLM approach (red) treats each complex independently;
the GLMM (blue) describes complex-specific changes to an overall trend. Frequentist p-values against the null
hypothesis of no relationship are 0.00047 (GLM) and 0.0038 (GLMM).
19
Figure S8: Little correlation between hydrophobicity and energetic centrality across gene products involved in
the complexes studied.
References
[1] Nuala A O’Leary, Mathew W Wright, J Rodney Brister, Stacy Ciufo, Diana Haddad, Rich McVeigh, Bhanu
Rajput, Barbara Robbertse, Brian Smith-White, Danso Ako-Adjei, et al. Reference sequence (refseq)
database at ncbi: current status, taxonomic expansion, and functional annotation. Nucleic acids research,
44(D1):D733–D745, 2016.
[2] Iain G Johnston and Ben P Williams. Evolutionary inference across eukaryotes identifies specific pres-
sures favoring mitochondrial gene retention. Cell Systems, 2(2):101–111, 2016.
[3] Peter JA Cock, Tiago Antao, Jeffrey T Chang, Brad A Chapman, Cymon J Cox, Andrew Dalke, Iddo
Friedberg, Thomas Hamelryck, Frank Kauff, Bartek Wilczynski, et al. Biopython: freely available python
tools for computational molecular biology and bioinformatics. Bioinformatics, 25(11):1422–1423, 2009.
[4] Helen M Berman, John Westbrook, Zukang Feng, Gary Gilliland, Talapady N Bhat, Helge Weissig, Ilya N
Shindyalov, and Philip E Bourne. The protein data bank. Nucleic acids research, 28(1):235–242, 2000.
[5] Sameer Velankar, Christoph Best, B Beuth, CH Boutselakis, N Cobley, AW Sousa Da Silva, Dimitris
Dimitropoulos, Adel Golovin, Miriam Hirshberg, Melford John, et al. Pdbe: protein data bank in europe.
Nucleic acids research, page gkp916, 2009.
[6] Scott Federhen. The ncbi taxonomy database. Nucleic acids research, 40(D1):D136–D143, 2012.
[7] Valen E Johnson and David Rossell. On the use of non-local prior densities in bayesian hypothesis tests.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 72(2):143–170, 2010.
[8] Warren L DeLano et al. Pymol: An open-source molecular graphics tool. CCP4 Newsletter on protein
crystallography, 40(1):82–92, 2002.
[9] Oscar D Monera, Terrance J Sereda, Nian E Zhou, Cyril M Kay, and Robert S Hodges. Relationship
of sidechain hydrophobicity and α-helical propensity on the stability of the single-stranded amphipathic
α-helix. Journal of peptide science, 1(5):319–329, 1995.
[10] DR Lide. Handbook of chemistry and physics. CRC Press, Boca Raton, FL, 1991.
[11] Michael D Barton, Daniela Delneri, Stephen G Oliver, Magnus Rattray, and Casey M Bergman. Evolution-
ary systems biology of amino acid biosynthetic cost in yeast. PloS one, 5(8):e11935, 2010.
[12] CL Craig and RS Weber. Selection costs of amino acid substitutions in cole1 and colia gene clusters
harbored by escherichia coli. Molecular biology and evolution, 15(6):774–776, 1998.
20
Figure S9: Effect of different averaging protocols to summarise gene statistics. (top) Correlations between our
model systems average and (A) average across randomly sampled species from different clades and (B) av-
erage across all species in the dataset (expected to be highly weighted towards bilaterians and angiosperms).
(bottom) Result of model selection and testing process with the random species averaging protocol.
21
Figure S10: Model selection and regression for the barcode-based retention index reflects the outcomes from
the simple retention index (Fig. 1); see also Table S1.
[13] Filip Husnik and Patrick J Keeling. The fate of obligate endosymbionts: reduction, integration, or extinction.
Current opinion in genetics & development, 58:1–8, 2019.
[14] Gordon M Bennett and Nancy A Moran. Small, smaller, smallest: the origins and evolution of ancient dual
symbioses in a phloem-feeding insect. Genome biology and evolution, 5(9):1675–1688, 2013.
[15] John P McCutcheon and Nancy A Moran. Parallel genomic evolution and metabolic interdependence in
an ancient symbiosis. Proceedings of the National Academy of Sciences, 104(49):19392–19397, 2007.
[16] Shinichiro Enomoto, Abhishek Chari, Adam Larsen Clayton, and Colin Dale. Quorum sensing attenuates
virulence in sodalis praecaptivus. Cell host & microbe, 21(5):629–636, 2017.
[17] Takuro Nakayama and Yuji Inagaki. Genomic divergence within non-photosynthetic cyanobacterial en-
dosymbionts in rhopalodiacean diatoms. Scientific reports, 7(1):1–8, 2017.
[18] John P McCutcheon, Bradon R McDonald, and Nancy A Moran. Origin of an alternative genetic code in
the extremely small and gc–rich genome of a bacterial symbiont. PLoS genetics, 5(7):e1000565, 2009.
[19] Valentina Serra, Leandro Gammuto, Venkatamahesh Nitla, Michele Castelli, Olivia Lanzoni, Davide
Sassera, Claudio Bandi, Bhagavatula Venkata Sandeep, Franco Verni, Letizia Modeo, et al. Morphol-
ogy, ultrastructure, genomics, and phylogeny of euplotes vanleeuwenhoeki sp. nov. and its ultra-reduced
endosymbiont candidatus pinguicoccus supinus sp. nov. Scientific reports, 10(1):1–27, 2020.
[20] Anna M Floriano, Michele Castelli, Sascha Krenek, Thomas U Berendonk, Chiara Bazzocchi, Giulio
Petroni, and Davide Sassera. The genome sequence of candidatus fokinia solitaria: insights on reductive
evolution in rickettsiales. Genome biology and evolution, 10(4):1120–1126, 2018.
[21] Duckhyun Lhee, Ji-San Ha, Sunju Kim, Myung Gil Park, Debashish Bhattacharya, and Hwan Su Yoon.
Evolutionary dynamics of the chromatophore genome in three photosynthetic paulinella species. Scientific
reports, 9(1):1–11, 2019.
[22] Jon S Graf, Sina Schorn, Katharina Kitzinger, Soeren Ahmerkamp, Christian Woehle, Bruno Huettel,
Carsten J Schubert, Marcel MM Kuypers, and Jana Milucka. Anaerobic endosymbiont generates energy
for ciliate host by denitrification. Nature, 591(7850):445–450, 2021.
[23] Liang Ran, John Larsson, Theoden Vigil-Stenman, Johan AA Nylander, Karolina Ininbergs, Wei-Wen
Zheng, Alla Lapidus, Stephen Lowry, Robert Haselkorn, and Birgitta Bergman. Genome erosion in a
nitrogen-fixing vertically transmitted endosymbiotic multicellular cyanobacterium. PLoS One, 5(7):e11486,
2010.
22
Figure S11: Bayesian model selection for linear models predicting retention index, with different priors from the
default choices in the main text. Left, mitochondrial; right, plastid data. Top, inverse moment (iMOM) prior with
τ = 0.133 and beta-binomial(1,1) prior over models. Centre, moment (MOM) prior with τ = 0.348 and uniform
prior over models. Bottom, IMOM prior with τ = 0.133 and uniform prior over models. MOM vs iMOM changes
structure of non-local priors; model priors assign different prior weights to overall model structures. Features
appearing in models are: 1 (intercept); 2 (length); 3 (hydrophobicity); 4 (hydrophobicity index); 5 (molecular
weight); 6 (amino pKa); 7 (carboxyl pKa); 8 (glucose assembly energy); 9 (alternate assembly energy); 10 (GC
content).
23
Figure S12: Decision tree and random forest classification for encoding compartment. (top) a set of trees
learned to predict encoding compartment for genes in different protein complexes, showing roles for hydropho-
bicity, pKa, and production energy (CW) as predictive features. (bottom) variance improvement plots for random
forest regression for compartment classification across all genes, illustrating the importance of each feature in
the predictive outcome. Complexes are labelled with regular expressions describing their gene labels.
24
[24] E. Paradis and K. Schliep. ape 5.0: an environment for modern phylogenetics and evolutionary analyses
in R. Bioinformatics, 35:526–528, 2019.
[25] Andrew Gelman and Yu-Sung Su. arm: Data Analysis Using Regression and Multilevel/Hierarchical Mod-
els, 2020. R package version 1.11-2.
[26] Yeojin Chung, Sophia Rabe-Hesketh, Vincent Dorie, Andrew Gelman, and Jingchen Liu. A nondegenerate
penalized likelihood estimator for variance parameters in multilevel models. Psychometrika, 78(4):685–
709, 2013.
[27] Adrian Raftery, Jennifer Hoeting, Chris Volinsky, Ian Painter, and Ka Yee Yeung. BMA: Bayesian Model
Averaging, 2021. R package version 3.18.15.
[28] David Orme, Rob Freckleton, Gavin Thomas, Thomas Petzoldt, Susanne Fritz, Nick Isaac, and Will
Pearse. caper: Comparative Analyses of Phylogenetics and Evolution in R, 2018. R package version
1.0.1.
[29] Claus O. Wilke. cowplot: Streamlined Plot Theme and Plot Annotations for ’ggplot2’, 2020. R package
version 1.1.1.
[30] David Meyer, Evgenia Dimitriadou, Kurt Hornik, Andreas Weingessel, and Friedrich Leisch. e1071: Misc
Functions of the Department of Statistics, Probability Theory Group (Formerly: E1071), TU Wien, 2021.
R package version 1.7-8.
[31] MW Pennell, JM Eastman, GJ Slater, JW Brown, JC Uyeda, RG Fitzjohn, ME Alfaro, and LJ Harmon.
geiger v2.0: an expanded suite of methods for fitting macroevolutionary models to phylogenetic trees.
Bioinformatics, 30:2216–2218, 2014.
[32] Barret Schloerke, Di Cook, Joseph Larmarange, Francois Briatte, Moritz Marbach, Edwin Thoen, Amos
Elberg, and Jason Crowley. GGally: Extension to ’ggplot2’, 2021. R package version 2.1.2.
[33] Elio Campitelli. ggnewscale: Multiple Fill and Colour Scales in ’ggplot2’, 2021. R package version 0.4.5.
[34] Hadley Wickham. ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New York, 2016.
[35] Alboukadel Kassambara. ggpubr: ’ggplot2’ Based Publication Ready Plots, 2020. R package version
0.4.0.
[36] Jun Cheng. ggpval: Annotate Statistical Tests for ’ggplot2’, 2021. R package version 0.2.4.
[37] Kamil Slowikowski. ggrepel: Automatically Position Non-Overlapping Text Labels with ’ggplot2’, 2021. R
package version 0.9.1.
[38] Guangchuang Yu, David Smith, Huachen Zhu, Yi Guan, and Tommy Tsan-Yuk Lam. ggtree: an r package
for visualization and annotation of phylogenetic trees with their covariates and other associated data.
Methods in Ecology and Evolution, 8:28–36, 2017.
[39] Shuangbin Xu, Zehan Dai, Pingfan Guo, Xiaocong Fu, Shanshan Liu, Lang Zhou, Wenli Tang, Tingze
Feng, Meijun Chen, Li Zhan, et al. ggtreeextra: Compact visualization of richly annotated phylogenetic
data. Molecular biology and evolution, 38(9):4039–4042, 2021.
[40] Jerome Friedman, Trevor Hastie, and Robert Tibshirani. Regularization paths for generalized linear mod-
els via coordinate descent. Journal of Statistical Software, 33(1):1–22, 2010.
[41] Baptiste Auguie. gridExtra: Miscellaneous Functions for ”Grid” Graphics, 2017. R package version 2.3.
[42] Dan Carr, ported by Nicholas Lewin-Koh, Martin Maechler, and contains copies of lattice functions writ-
ten by Deepayan Sarkar. hexbin: Hexagonal Binning Routines, 2021. R package version 1.28.2.
[43] Gabor Csardi and Tamas Nepusz. The igraph software package for complex network research. InterJour-
nal, Complex Systems:1695, 2006.
[44] Douglas Bates, Martin M¨achler, Ben Bolker, and Steve Walker. Fitting linear mixed-effects models using
lme4. Journal of Statistical Software, 67(1):1–48, 2015.
[45] Georg Heinze, Meinhard Ploner, and Lena Jiricka. logistf: Firth’s Bias-Reduced Logistic Regression,
2020. R package version 1.24.
25
[46] David Rossell, John D. Cook, Donatello Telesca, P. Roebuck, and Oriol Abril. mombf: Bayesian Model
Selection and Averaging for Non-Local and Local Priors, 2021. R package version 3.0.4.
[47] Jose Pinheiro, Douglas Bates, Saikat DebRoy, Deepayan Sarkar, and R Core Team. nlme: Linear and
Nonlinear Mixed Effects Models, 2021. R package version 3.1-152.
[48] K.P. Schliep. phangorn: phylogenetic analysis in r. Bioinformatics, 27(4):592–593, 2011.
[49] Liam J. Revell. phytools: An r package for phylogenetic comparative biology (and other things). Methods
in Ecology and Evolution, 3:217–223, 2012.
[50] Andy Liaw and Matthew Wiener. Classification and regression by randomforest. R News, 2(3):18–22,
2002.
[51] M.P.J. van der Loo. The stringdist package for approximate string matching. The R Journal, 6:111–122,
2014.
[52] Hadley Wickham. stringr: Simple, Consistent Wrappers for Common String Operations, 2019. R package
version 1.4.0.
[53] Brian Ripley. tree: Classification and Regression Trees, 2021. R package version 1.0-41.
[54] Christiam Camacho, George Coulouris, Vahram Avagyan, Ning Ma, Jason Papadopoulos, Kevin Bealer,
and Thomas L Madden. Blast+: architecture and applications. BMC bioinformatics, 10(1):1–9, 2009.
26
| 2021 | Universal features shaping organelle gene retention | 10.1101/2021.10.27.465964 | [
"Giannakis Konstantinos",
"Arrowsmith Samuel J.",
"Richards Luke",
"Gasparini Sara",
"Chustecki Joanna M.",
"Røyrvik Ellen C.",
"Johnston Iain G."
] | creative-commons |
1
The emergence of new lineages of the Monkeypox virus could affect the 2022 outbreak.
Mayla Abrahim1, Alexandro Guterres1, Patrícia Cristina da Costa Neves1 and Ana Paula
Dinis Ano Bom1.
1 Laboratório de Tecnologia Imunológica, Instituto de Tecnologia em Imunobiológicos,
Vice-Diretoria de Desenvolvimento Tecnológico, Bio-Manguinhos, Fundação Oswaldo
Cruz (FIOCRUZ), Rio de Janeiro, RJ, Brazil.
Keywords: Monkeypox; lineages; pathogenesis; mutations; proteins; signature amino acid
Corresponding author: Alexandro Guterres
Laboratório de Tecnologia Imunológica, Bio-Manguinhos, FIOCRUZ.
Pavilhão Rockefeller, Av. Brasil 4365 – Manguinhos
Rio de Janeiro, RJ, 21045-900, Brazil
e-mail: guterres_rj@yahoo.com.br
2
Abstract
Human monkeypox is a contagious zoonotic viral disease caused by Monkeypox
virus and is causing a current outbreak in various regions of the world, being already
considered an epidemic and a global public health problem. From the sequenced
monkeypox genomes of clades B.1, A.1.1 and A.2 available, we performed analyzes of 9
proteins considered important in the pathogenesis of the disease (A9L, A36R, A50L,
B9R, B16L, C3L, C7L, C12L (SPI-1) and H5R) and 4 important proteins for the host's
immune response (A27L, A33R, B5R and L1R). We identified four synonymous
mutations and six amino acid changes, of which four are in conserved domains, such
changes can alter the function of proteins. Furthermore, we did not find the C3L protein
in monkeypox genomes from the 2022 outbreak, an important protein for disease
pathogenicity. Our analyses suggest that lineage/clade A.2 may be suffering the different
effects of various selective pressures than lineage/clade B.1. In conclusion, the mutations
identified in the present study have not yet been associated with genetic alterations,
significant changes in the transmission route, mean age, signs/symptoms at the clinical
presentation, and their evolution could be detected. Therefore, further research in the field
is needed since our findings need to be confirmed by new studies.
3
Introduction
In May 2022, numerous cases of monkeypox started to be identified in several
non-endemic countries. In about a month, more than 3.500 confirmed cases of
monkeypox have been reported in, at least, 50 non-African countries until past week.
(Kraemer et al., 2022; World Health Organization, 2022). These features are totally new
for this disease in humans, since Monkeypox virus was endemic in West and Central
Africa, and only occasionally caused short outbreaks elsewhere in the world, which were
quickly contained or peter out by themselves (Huhn et al., 2005; Reed et al., 2004). In
endemic African countries, published mortality rates vary from 1% to 10%. Despite the
data restriction, the lineage/clade responsible for outbreaks in the Congo Basin appears
to be associated with higher virulence (Likos et al., 2005).
Monkeypox virus is a double-stranded DNA virus with about 200-kb genome,
being a member of the Orthopoxvirus genus from the Poxviridae family. Recently, two
lineages of the Monkeypox virus were identified in the current outbreak in non-endemic
countries (Gigante et al., 2022). The most sequenced lineage/clade, to date, is related with
a 2021 travel-associated case from Nigeria to Maryland in the USA (USA_2021_MD)
that displays high similarity to the predominant 2022 Monkeypox virus outbreak
sequences. The second lineage/clade is related to Monkeypox virus from a 2021 traveler
from Nigeria to Texas in the USA (USA_2021_TX) (Figure 01). In 2005, Likos and
collaborators compared clinical, laboratory and epidemiological features of confirmed
human monkeypox case-patients. They suggested that human disease pathogenicity was
associated with the viral lineage/clade (West African and Congo Basin (Central African)).
A comparison of proteins between Monkeypox virus clades permitted the prediction of
viral proteins that could cause the observed differences in human pathogenicity (Likos et
al., 2005).
4
The re-emergence and dissemination of the Monkeypox virus have resulted in
infections across the globe. Something has changed. Before the 2022 outbreak, cases
outside Africa have previously been limited to a handful that was associated with travel
to Africa or with the importation of infected animals. Moreover, the ongoing cases differ
from previous outbreaks in terms of age (thirties), sex/gender (most cases being males),
and transmission route, being sexual transmission being highly likely. The clinical
presentation is atypical and unusual, being characterized by anogenital lesions and rashes
that relatively spare the face and extremities (Bragazzi et al., 2022).
Methods
Complete genome sequences of deposited Monkeypox virus were retrieved from
the GenBank® (www.ncbi.nlm.nih.gov). Multiple sequence alignment as well as the
comparison of nucleotide sequences were performed with MAFFT version 7.4 employing
the E-INS-I algorithm (https://mafft.cbrc.jp/alignment/software/). The phylogenetic
relations of the complete genome were estimated using Maximum Likelihood Method
implemented in RAxML (https://cme.h-its.org/exelixis/web/software/raxml/) under the
HKY+G+I model of sequence evolution. Statistical support of the clades was measured
by a heuristic search with 1,000 bootstrap replicates in RAxML. The best-fit evolutionary
model was determined using the Bayesian Information Criterion.
Results and Discussion
We analyzed nine proteins (A9L, A36R, A50L, B9R, B16L, C3L, C7L, C12L
(SPI-1 and H5R) identified by Likos and collaborators in the two lineages/clades (B.1
and A.2) identified in the current outbreak and one lineage/clade (A1.1) identified in 2021
(https://nextstrain.org/monkeypox/hmpxv1). Five proteins are involved with either
immune evasion or host range and the remaining four proteins are involved with various
aspects of the viral life cycle in other poxviruses (Black et al., 1998; Legrand et al., 2004;
5
Moon et al., 1999). We observed nucleotide substitutions in six of the nine genes
analyzed. The nucleotide changes result in five missense mutations and four synonymous
mutations. Some of these changes are within protein domains (table 01). Herein, we found
in lineage/clade 2 the development of a signature amino acid sequence in position 442
(aspartic acid > asparagine) in A50L gene. Interestingly, when we evaluated the genomes
of the current outbreak, we did not find the C3L gene. The vaccinia virus complements
control protein is a 35-kDa protein that is encoded by the C3L gene and secreted by cells
infected with the Vaccinia virus. Members of this family can block complement-mediated
induction of the inflammatory response, and engulfment, killing and lysis of bacteria and
viruses (Chen et al., 2005; Isaacs et al., 2003; Kotwal and Moss, 1988).
Orthopoxviruses produces two antigenically distinct infectious virions,
intracellular mature virus (IMV) and extracellular enveloped virus (EEV). Structurally,
EEV consists of an IMV with an additional outer membrane containing proteins that are
absent from IMV. Due to their stability in the environment, IMVs play a predominant
role in host-to-host transmission, whereas EEVs play an important role in dissemination
within the host (Vanderplasschen et al., 1998). Additionally, we analyzed 4 genes used
successfully in vaccine studies and important to the host's immune response in the two
antigenically distinct infectious virions. Hooper and collaborators reported that a gene-
based vaccine comprised of the A27L and L1R proteins associated with IMV and, A33R
and B5R proteins associated with EEV may be a useful candidates to protect against other
orthopoxviruses, including those that cause smallpox and Monkeypox virus (Hooper et
al., 2003). Again, we found lineage/clade A.1 signature amino acid in position 221
(proline > serine) for B5R protein (table 01). However, the remaining genes had no
nucleotide changes. These data demonstrate that such proteins are attractive targets for
future studies in vaccine production since only B5L had an amino acid substitution.
6
Additionally, it has already been described that the combination of the four VACV genes
(A27L + A33R + L1R + B5R) can provide an alternative vaccine for poxvirus, without
the known side effects and serious adverse reactions.
The genomic surveillance has been vital to the early detection of mutations,
monitoring of virus evolution and evaluating the degree of similarities between
circulating. Molecular clock analyses assumed an evolutionary rate of 5 x 10-6 (Firth et
al., 2010). These mutations arise as a natural by-product of viral replication. Our analyses
suggest that lineage/clade A.2 may be suffering the different effects of various selective
pressures than lineage/clade B.1. Some studies analyzing single genes or whole genomes
have suggested a relation between lineage/clade with differences in the human
monkeypox disease pathology. Combined, these observations propose that the effect of
changes among a moderately small number of genes could account for the modifications
in viral clearance and pathogenesis of human infections. (Esposito and Knight, 1985;
Likos et al., 2005; Reed et al., 2004).
Therefore, with the emergence of new lineages/clades the evaluation of novel
Monkeypox variants should include an assessment of the following questions: What effect
do these mutations have on transmissibility and spread, antigenicity, aspects of
pathogenesis, or virulence? Although it is not yet associated with genetic alterations,
significant changes in the transmission route, mean age, signs/symptoms at the clinical
presentation, and their evolution could be detected (Bragazzi et al., 2022; Patrocinio-Jesus
and Peruzzu, 2022).
Regardless of why the mutations were selected, it is reasonable to expect that
many mutations in these genes affect viral fitness. Sometimes a mutation that enhances
one viral property, can reduce another property. Although most cases in current outbreaks
have presented with mild disease symptoms, Monkeypox virus may cause severe disease
7
in certain population groups as immunosuppressed persons, young children and pregnant
women (Di Giulio and Eckburg, 2004). Even if there are few data linking pregnant
women and the effects of human Monkeypox virus infection, there is evidence that viruses
of the Orthopoxvirus genus are associated with an increased risk of maternal and perinatal
morbidity and mortality (Dashraath et al., 2022; Khalil et al., 2022; Mbala et al., 2017).
Understanding how virulence evolves after a virus jumps or adapts to a new host species
is critical to the effective prevention and treatment of viral infections. Finally, it is
possible that an increased understanding of virulence evolution drawn from a relevant
data set (phylogenetics, epidemiology, and experimental studies of virus virulence and
fitness) may contribute to new strategies for human monkeypox control and eradication.
Figure 01. Phylogenetic analysis of human monkeypox virus based on 171 genomes
complete sequences using the Maximum Likelihood Method using RAxML. The
Hasegawa-Kishino-Yano model with gamma-distributed heterogeneity (HKY + G) was
selected as the best-fit evolutionary model. Bootstrap: 1000.
Author Approvals: All authors critically reviewed the manuscript for intellectual content
and approved it in its final version.
Declaration of Competing Interest: The authors report no declarations of interest.
References
Black, E.P., Moussatche, N., Condit, R.C., 1998. Characterization of the Interactions
among Vaccinia Virus Transcription Factors G2R, A18R, and H5R. Virology 245,
313–322. https://doi.org/10.1006/viro.1998.9166
Bragazzi, N.L., Kong, J.D., Mahroum, N., Tsigalou, C., Khamisy‐Farah, R., Converti,
M., Wu, J., 2022. Epidemiological trends and clinical features of the ongoing
monkeypox epidemic: A preliminary pooled data analysis and literature review. J.
8
Med. Virol. https://doi.org/10.1002/jmv.27931
Chen, N., Li, G., Liszewski, M.K., Atkinson, J.P., Jahrling, P.B., Feng, Z., Schriewer, J.,
Buck, C., Wang, C., Lefkowitz, E.J., Esposito, J.J., Harms, T., Damon, I.K., Roper,
R.L., Upton, C., Buller, R.M.L., 2005. Virulence differences between monkeypox
virus isolates from West Africa and the Congo basin. Virology 340, 46–63.
https://doi.org/10.1016/j.virol.2005.05.030
Dashraath, P., Nielsen-Saines, K., Mattar, C., Musso, D., Tambyah, P., Baud, D., 2022.
Guidelines for pregnant individuals with monkeypox virus exposure. Lancet.
https://doi.org/10.1016/S0140-6736(22)01063-7
Di Giulio, D.B., Eckburg, P.B., 2004. Human monkeypox: an emerging zoonosis. Lancet
Infect. Dis. 4, 15–25. https://doi.org/10.1016/S1473-3099(03)00856-9
Esposito, J.J., Knight, J.C., 1985. Orthopoxvirus DNA: A comparison of restriction
profiles and maps. Virology 143, 230–251. https://doi.org/10.1016/0042-
6822(85)90111-4
Firth, C., Kitchen, A., Shapiro, B., Suchard, M.A., Holmes, E.C., Rambaut, A., 2010.
Using Time-Structured Data to Estimate Evolutionary Rates of Double-Stranded
DNA
Viruses.
Mol.
Biol.
Evol.
27,
2038–2051.
https://doi.org/10.1093/molbev/msq088
Gigante, C.M., Korber, B., Seabolt, M.H., Wilkins, K., Davidson, W., Rao, A.K., Zhao,
H., Hughes, C.M., Minhaj, F., Waltenburg, M.A., Theiler, J., Smole, S., Gallagher,
G.R., Blythe, D., Myers, R., Schulte, J., Stringer, J., Lee, P., Mendoza, R.M., Griffin-
Thomas, L.A., Crain, J., Murray, J., Atkinson, A., Gonzalez, A.H., Nash, J., Batra,
D., Damon, I., McQuiston, J., Hutson, C.L., McCollum, A.M., Li, Y., 2022. Multiple
lineages of <em>Monkeypox virus</em> detected in the United States, 2021- 2022.
9
bioRxiv 1–15. https://doi.org/10.1101/2022.06.10.495526
Hooper, J.., Custer, D.., Thompson, E., 2003. Four-gene-combination DNA vaccine
protects mice against a lethal vaccinia virus challenge and elicits appropriate
antibody
responses
in
nonhuman
primates.
Virology
306,
181–195.
https://doi.org/10.1016/S0042-6822(02)00038-7
Huhn, G.D., Bauer, A.M., Yorita, K., Graham, M.B., Sejvar, J., Likos, A., Damon, I.K.,
Reynolds, M.G., Kuehnert, M.J., 2005. Clinical Characteristics of Human
Monkeypox, and Risk Factors for Severe Disease. Clin. Infect. Dis. 41, 1742–1751.
https://doi.org/10.1086/498115
Isaacs, S.N., Argyropoulos, E., Sfyroera, G., Mohammad, S., Lambris, J.D., 2003.
Restoration of Complement-Enhanced Neutralization of Vaccinia Virus Virions by
Novel Monoclonal Antibodies Raised against the Vaccinia Virus Complement
Control Protein. J. Virol. 77, 8256–8262. https://doi.org/10.1128/JVI.77.15.8256-
8262.2003
Khalil, A., Samara, A., O’Brien, P., Morris, E., Draycott, T., Lees, C., Ladhani, S., 2022.
Monkeypox and pregnancy: what do obstetricians need to know? Ultrasound Obstet.
Gynecol. https://doi.org/10.1002/uog.24968
Kotwal, G.J., Moss, B., 1988. Vaccinia virus encodes a secretory polypeptide structurally
related
to
complement
control
proteins.
Nature
335,
176–178.
https://doi.org/10.1038/335176a0
Kraemer, M.U.G., Tegally, H., Pigott, D.M., Dasgupta, A., Sheldon, J., Wilkinson, E.,
Schultheiss, M., Han, A., Oglia, M., Marks, S., Kanner, J., O’Brien, K., Dandamudi,
S., Rader, B., Sewalk, K., Bento, A.I., Scarpino, S. V, de Oliveira, T., Bogoch, I.I.,
Katz, R., Brownstein, J.S., 2022. Tracking the 2022 monkeypox outbreak with
10
epidemiological
data
in
real-time.
Lancet
Infect.
Dis.
22,
941–942.
https://doi.org/10.1016/S1473-3099(22)00359-0
Legrand, F.A., Verardi, P.H., Jones, L.A., Chan, K.S., Peng, Y., Yilma, T.D., 2004.
Induction of Potent Humoral and Cell-Mediated Immune Responses by Attenuated
Vaccinia Virus Vectors with Deleted Serpin Genes. J. Virol. 78, 2770–2779.
https://doi.org/10.1128/JVI.78.6.2770-2779.2004
Likos, A.M., Sammons, S.A., Olson, V.A., Frace, A.M., Li, Y., Olsen-Rasmussen, M.,
Davidson, W., Galloway, R., Khristova, M.L., Reynolds, M.G., Zhao, H., Carroll,
D.S., Curns, A., Formenty, P., Esposito, J.J., Regnery, R.L., Damon, I.K., 2005. A
tale of two clades: monkeypox viruses. J. Gen. Virol. 86, 2661–2672.
https://doi.org/10.1099/vir.0.81215-0
Mbala, P.K., Huggins, J.W., Riu-Rovira, T., Ahuka, S.M., Mulembakani, P., Rimoin,
A.W., Martin, J.W., Muyembe, J.-J.T., 2017. Maternal and Fetal Outcomes Among
Pregnant Women With Human Monkeypox Infection in the Democratic Republic of
Congo. J. Infect. Dis. 216, 824–828. https://doi.org/10.1093/infdis/jix260
Moon, K.B., Turner, P.C., Moyer, R.W., 1999. SPI-1-Dependent Host Range of
Rabbitpox Virus and Complex Formation with Cathepsin G Is Associated with
Serpin Motifs. J. Virol. 73, 8999–9010. https://doi.org/10.1128/JVI.73.11.8999-
9010.1999
Patrocinio-Jesus, R., Peruzzu, F., 2022. Monkeypox Genital Lesions. N. Engl. J. Med.
https://doi.org/10.1056/NEJMicm2206893
Reed, K.D., Melski, J.W., Graham, M.B., Regnery, R.L., Sotir, M.J., Wegner, M. V.,
Kazmierczak, J.J., Stratman, E.J., Li, Y., Fairley, J.A., Swain, G.R., Olson, V.A.,
Sargent, E.K., Kehl, S.C., Frace, M.A., Kline, R., Foldy, S.L., Davis, J.P., Damon,
11
I.K., 2004. The Detection of Monkeypox in Humans in the Western Hemisphere. N.
Engl. J. Med. 350, 342–350. https://doi.org/10.1056/NEJMoa032299
Vanderplasschen, A., Hollinshead, M., Smith, G.L., 1998. Intracellular and extracellular
vaccinia virions enter cells by different mechanisms. J. Gen. Virol. 79, 877–887.
https://doi.org/10.1099/0022-1317-79-4-877
World Health Organization, W., 2022. Monkeypox - United Kingdom of Great Britain
and
Northern
Ireland
[WWW
Document].
URL
https://www.who.int/emergencies/disease-outbreak-news/item/2022-DON383
100
100
4.0E-5
100
oe
AAA
100
TA
BA
Table 1: Nucleotide changes between clades/lineages A.2 and B.1. was used 155 completes genomes 2022 outbreak. Gene homologs in Vaccinia
virus strain Copenhagen (AY315828) are given for each gene as well as notes on proposed functions taken from annotation.
Gene
(VACV-Cop)
Note
NT Change
AA Change
Domains
(Vaccinia Reference)
Accession
Clade
Importance
A36R
IEV transmembrane (Viral
life cycle)
141 C > T
Synonymous
-
ON675438
A.2
Pathogenies
A50R
DNA ligase
413 C > T
Cys138Phe
DNA ligase N
terminus
ON803435
B.1
Pathogenies
1324 G > A
Asp442Asn
ATP dependent DNA
ligase C terminal
region
ON674051
A.2
Pathogenies
ON675438
ON676707
B5R
EEV type-I membrane
661 C > T
Pro221Ser
Sushi
ON674051
A.2
Immune
response
ON675438
ON676707
B9R
Expressed late during
infection
303 C > T
Synonymous
-
ON675438
A.2
Pathogenies
C7L
Host-range factor for
vaccinia virus life cycle in
mammalian cells
151 G > A
Asp051Asn
-
ON675438
A.2
Pathogenies
C12L (SPI-1)
Serpin (serine protease
inhibitor)
272 C > T
Ser091Leu
-
ON674051
A.2
Pathogenies
363 A > G
Synonymous
-
ON675438
A.2
Pathogenies
588 C > T
Synonymous
-
ON675438
A.2
Pathogenies
H5R
Multifunctional protein*
275 C > T
Ser092Phe
Disordered
ON602722
B.1
Pathogenies
* Protein involved in viral DNA replication, postreplicative gene transcription, and virion morphogenesis.
| 2022 | The emergence of new lineages of the Monkeypox virus could affect the 2022 outbreak | 10.1101/2022.07.07.498743 | [
"Abrahim Mayla",
"Guterres Alexandro",
"da Costa Neves Patrícia Cristina",
"Ano Bom Ana Paula Dinis"
] | creative-commons |
Journal, Vol. XXI, No. 1, 1-5, 2013
Additional note
Whole-brain modeling of the differential influences
of Amyloid-Beta and Tau in Alzheimer‘s Disease
Gustavo Patow1,4, Leon Stefanovski2,3, Petra Ritter2,3, Gustavo Deco4 and Xenia Kobeleva5,6,
for the Alzheimer’s Disease Neuroimaging Initiative∗
Abstract
Alzheimer’s Disease (AD) is a neurodegenerative condition associated with extra- and intra-neuronal accu-
mulation of two misfolded proteins, namely Amyloid-Beta (Aβ) and tau. In this paper, we study the effect of
these proteins on neuronal activity, with the aim of assessing their individual and combined impact on neuronal
processes. The technique uses a whole-brain dynamic model to find the optimal parameters that best describe
the effects of Aβ and tau on the excitation-inhibition balance of the local nodes. Our experimental results
show a clear dominance of the neuronal activity of Aβ over tau in the early disease stages (Mild Cognitive
Impairment), while tau dominates over Aβ in the latest stages (AD). Our findings identify a crucial role for Aβ and
tau in contributing to complex neuronal dynamics and demonstrate the viability of using regional distributions of
neuropathology to define models of large-scale brain function in AD. Our study provides further insight into the
dynamics and complex interplay between these two proteins among themselves and with the regional neural
activity, opening the path for further investigations on biomarkers and candidate therapeutic targets in-silico.
Keywords
Alzheimer’s Disease — Whole-Brain model — Amyloid-Beta — Tau
1ViRVIG, Universitat de Girona, Girona, Spain
2Berlin Institute of Health at Charit´e – Universit¨atsmedizin Berlin, Berlin, Germany
3Department of Neurology with Experimental Neurology, Brain Simulation Section, Charit´e – Universit¨atsmedizin Berlin, corporate member of
Freie Universit¨at Berlin and Humboldt-Universit¨at zu Berlin, Berlin, Germany
4Center for Brain and Cognition, Computational Neuroscience Group, Department of Information and Communication Technologies,
Universitat Pompeu Fabra, Spain
5University Hospital Bonn, Clinic for Neurology, Bonn, Germany
6German Center for Neurodegenerative Diseases (DZNE) Bonn, Bonn, Germany
Corresponding author: Gustavo Patow, ViRVIG, Universitat de Girona, 17003, Spain. gustavo.patow@udg.edu
1
1. Introduction
Alzheimer’s Disease (AD) is a neurodegenerative disease that affects first the medial temporal lobe and the limbic system, and
most areas of the neocortex at later disease stages [1, 2, 3]. The disease can remain asymptomatic for years but ultimately
leads to progressive impairment of memory and other cognitive domains, neuropsychiatric symptoms and, ultimately, to severe
impairment in all body functions. This results in both a huge loss of quality of life of affected people and caregivers and high
costs for the society at large. Minor cognitive deficits with little influence on activities of daily living with, are defined as mild
cognitive impairment (MCI). In the typical disease course, the deficits extend later on to other cognitive domains as, e.g., speech
and spatial orientation. When the cognitive impairment is severe enough to affect the activities of daily living, the disease is
usually referred to as dementia (due to AD) [4].
AD pathogenesis is associated with several interlinked pathomechanistic processes, from genomics to connectomics,
including the Notch-1 pathway, neurotransmitters, polygenetic factors, neuroinflammation, and neuroplasticity [5]. However,
the accumulation of misfolded proteins within the brain is considered as the pathological hallmark of AD: namely extracellular
accumulation of Amyloid-beta (Aβ), forming what are known as senile plaques; and intraneuronal aggregation of the
microtubule protein tau, called neurofibrillary tangles [6]. In general, it is known that Aβ plaques and tau tangles spread
independently through the brain as the disease progresses [7]. Both proteins are currently considered as biomarkers that are
used in the diagnostic classification of AD [6]. Although a plethora of treatment strategies has been examined in the last
decades, the neuronal degeneration itself, as well as the cognitive decline cannot be currently stopped by any treatment, AD is
therefore still considered as incurable. Treatments for removal of Aβ (e.g., with Adacanumab and Lecanemab) are currently
discussed in light of inconclusive effects on halting disease progression [8]. Even more, in spite of the large body of research
devoted to the study of AD, many aspects regarding AD pathophysiology and the role of Aβ and tau in the disease process
are still incompletely understood [9, 10, 11]. While several studies have shown abnormal brain network function at various
stages of AD [6, 12, 13], the relationship between pathology (i.e., Aβ and tau) and associated brain dysfunction has not been
described in great detail [10].
Regarding brain dysfunction, several ex-vivo (human) and animal studies have seen a disruption in excitation/inhibition
(E/I) balance (i.e., the relative contributions of excitatory and inhibitory synaptic inputs corresponding to a neuronal event,
such as a response evoked by sensory stimulation) in the form of hyperexcitability consequence of the disruption of glutamate
reuptake, also disrupting cognition-related cortical activity and contributing to intellectual decline in AD [12, 13]. Change et
al. [14] showed tau affects excitatory and inhibitory neurons differently, and that its ablation decreases the baseline activity
of excitatory neurons, while modulating the intrinsic excitability and axon initial segments of inhibitory neurons, promoting
network inhibition. In this line, Bi and co-workers [15] hypothesized that Aβ produces alterations to the GABAergic system
contributing to impairing GABAergic function and thus producing synaptic hyperexcitation, leading to E/I imbalance and
AD pathogenesis. Petrache et al. [16] found a decrease in canonical synaptic signaling mechanisms first affecting the lateral
entorhinal cortex in combination with synaptic hyperexcitation and severely disrupted E/I inputs onto principal cells, and a
reduction in the somatic inhibitory axon terminals in the lateral entorhinal cortex compared with other cortical regions. Recently,
Lauterborn and coauthors [17] studied the synaptic disturbances in E/I balance in forebrain circuits by assessing anatomical and
electrophysiological synaptic E/I ratios in post-mortem parietal cortex samples of AD patients, revealing significantly elevated
E/I ratios.
While interesting results regarding E/I imbalance were derived ex-vivo (in humans), studies in-vivo regarding E/I imbalance
in AD are lacking, as the activity of excitatory vs. inhibitory neuronal populations cannot be directly measured using
neuroimaging. Most works focusing on whole-brain dynamics studied different measures of brain activation patterns, e.g., from
its connectivity, but were not informative regarding the role of excitatory vs. inhibitory neuronal populations [18, 19, 20, 21, 22].
To disentangle mechanistic contributions of separate neuronal populations, whole-brain dynamic models can contribute to
analyze collective properties of the brain [23, 24, 25], such as the fMRI signal [26, 27, 28, 29]. To understand the complex
interplay between pathophysiological processes and brain activity (i.e., the fMRI signal), models might become even more
informative when incorporating heterogeneity of brain dynamics in brain regions, based on empirical data [30, 31, 32].
Earlier work specifically on AD using whole-brain simulations focused only on linking global and local brain dynamics to
individual differences in cognitive performance scores from different subject conditions [18]. Demirtas¸ [20] et al. studied the
effect of heterogeneity of local synaptic strengths on a large-scale dynamical circuit model of human cortex in healthy subjects,
showing that heterogeneity significantly improved the fitting of fMRI resting-state functional connectivity, and was able to
capture sensory-association organization of multiple fMRI features. Following this approach, recent work by Stefanovski
and co-authors [21] focused on the connection of Aβ with neural function in The Virtual Brain (TVB) platform [33], using a
network of interconnected (through the corresponding structural connectivity matrix) Jansen-Rit models [34], addressing the
phenomenon of hyperexcitability in AD, examining how Aβ burden modulates regional Excitation-Inhibition balance, leading
to local hyperexcitation with high Aβ loads in the model, reproducing what has been previously observed in experimental
studies. The resulting simulated local field potentials improved previous diagnostic classifications between AD and controls [22].
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 3/20
However, all these works study the effect of a single burden, namely Aβ, on the brain neuronal dynamics, while the work
we present here focuses mostly on the interplay of both burdens, i.e., Aβ and tau, assessing their relative impacts on these
dynamics.
The main objective of this paper is to use whole-brain modeling techniques to study the impact of both Aβ and tau on the
dynamics of the regional behaviors in AD. As such, we used our results to discern the impact of each protein in isolation and
in combination, being able to assess their relative weights on contributing to abnormal brain activity. We use the Balanced
Dynamic Mean Field (BEI) model [31], where local neuronal dynamics of each region evolve according to a dynamic mean
field model derived from the behavior of interacting excitatory and inhibitory populations. We will show in this work a clear
dominance of the effects of Aβ over tau in the earlier stages of the disease (Mild Cognitive Impairment, MCI), and a dominance
of protein tau over the ones of Aβ on the function of the brain dynamics in advanced stages (manifest dementia).
2. Methods Overview
Model Creation:
Figure 1a presents an overview of our overall analysis strategy, and the details could be found in the
Methods Section. We make use of MRI and positron emission tomography (PET) from the Alzheimer’s Disease Neuroimaging
Initiative (ADNI). In summary, we use diffusion MRI to generate the structural connectomes of healthy controls (HC), mild
cognitive impairment (MCI) and Alzheimer’s Disease (AD) subjects. We use task-free resting-state functional MRI to fit a
whole-brain model in which the local neuronal dynamics of each brain region evolves according to the dynamic mean field
model by Deco et al. [31], which is then connected to a spontaneous blood-oxygenation-level-dependent (BOLD) dynamics.
We refer to this model as the Balanced Excitation-Inhibition (BEI) model, which can be thought of as a homogeneous reference
against which we evaluate the performance of our heterogeneous AD model. Aβ and tau distributions are derived from AV-45
and AV-1451 PET from ADNI. For the heterogeneous model, we incorporate regional heterogeneous distributions of the main
proteins involved in AD, namely Aβ and tau, as first order multiplicative polynomials for each burden and for each type of
population (excitatory/inhibitory) into the local gain parameter M(E,I). Fitting the model to empirical fMRI data allows us to
evaluate which effect of Aβ and tau to the different populations can mechanistically explain the observed behaviors.
Model Fitting:
For both of our models, homogeneous and heterogeneous, we assume that all diffusion MRI-reconstructed
streamline fibers have the same conductivity and thus the coupling between different brain areas is scaled by a single global
parameter, G. We first tune the G parameter of the BEI model to adjust the strength of effective coupling in the model and
identify the brain’s dynamic working-point by fitting the model to three empirical properties that are estimated from the
empirical fMRI data:
• the Pearson correlation between model and empirical estimates of static (i.e., time-averaged) functional connectivity
estimated across all pairs of brain regions (FC);
• similarity in sliding-window functional connectivity dynamics (swFCD);
• the KS distance between a set of dynamic functional connectivity matrices (also called coherence connectivity matrix [35])
built from the average BOLD time series of each ROI, which were Hilbert-transformed to yield the phase evolution of
the regional signals (phFCD).
We then fit the coefficients for the two burdens, for excitatory and inhibitory populations, with a global optimization algorithm,
within directional bounds obtained from previous clinical observations (see below, in Section 5.7).
Result Evaluation:
We evaluate the quality of the results in two ways. First, we shuffle the input burdens, and compare the
result of performing the simulation with
• the optimized parameters with shuffled burdens.
• the optimized parameters with original (i.e., not shuffled) burdens.
• the homogeneous BEI model.
Second, we examine the relevance of each type of burden by optimizing them in isolation of each other (i.e., zeroing the other
one out), and comparing the results. The full comparisons include both burdens in isolation, both burdens simultaneously, and
with the homogeneous (i.e., BEI) model. See Figure 1b.
3. Results
We used diffusion MRI to generate a the Structural Connectomes of 17 healthy control (HC) subjects, 9 mild cognitive
impairment (MCI) subjects and 10 subjects with Alzheimer’s Disease (AD) from ADNI, which are mostly the same participants
as used by Stefanovski et al. [21] and Triebkorn et al. [22]. See Table 1.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 4/20
Diagnosis
n (female)
Mean age
σ
Min. age
Max. age
Mean MMSE
σMMSE
Min. MMSE
Max. MMSE
AD
10 (5)
72.0
9.6
55.9
86.1
21.3
6.8
9
30
HC
17 (10)
70.8
4.3
63.1
78.0
29.3
0.7
28
30
MCI
9 (3)
68.8
5.8
57.8
76.6
27.4
1.5
25
30
Table 1. Epidemiological information of the population used in this study.
3.1 Fitting the Homogeneous Model
As a first step, we evaluated the capability of the homogeneous BEI model to reproduce empirical properties of resting-state FC
data. To this end, we fitted the global coupling parameter, G, without considering heterogeneity by setting all regional gain
parameters M(E,I) = 1 [36, 31]. Then, we evaluated the ability of the model to reproduce three different properties of empirical
resting-state fMRI recordings: edge-level static FC, swFCD, and phFCD (see Methods for further details.) The results of this
analysis are shown in Figure 2A. To remove differences across subjects related to age, we considered averaged values across
subjects over the healthy control group, and took an equivalent number of simulated trials with the same duration as the human
experiments (see Methods). Following previous research [37, 38, 32] fitting the phFCD better captures the spatiotemporal
structure of the fMRI data, being a stronger constraint on the model. Indeed, where FC fits are consistently high across a broad
range of G values, phFCD yields a clear global optimum at G = 3.1. Thus, we choose to use phFCD for all further analysis.
3.2 Introducing Aβ and tau heterogeneity
Once the global coupling parameter has been found, we can introduce the regional heterogeneity in the distributions of Aβ
and tau, and study how their introduction leads to a better representation of neural dynamics, i.e., improves the fitting of
phFCD. Spatial maps for each form of protein burden used in our modeling are shown in Figures 2G (for Aβ) and 2H (for tau)
for one particular individual. For some individuals, (mainly HC subjects, e.g., as subject 003 S 6067 in the ADNI database,
with ρ = 0.92, p < 0.001) the Aβ and tau distributions are strongly correlated, while for others the two maps show a weaker
correlation (e.g., subject 036 S 4430, with ρ = 0.10, p = 0.04.) This observation indicates that each protein burden introduces
a different form of biological heterogeneity to the benchmark BEI model, and thus should be modeled separately in our
simulations.
We introduce these kinds of heterogeneity by modulating the regional gain functions M(E,I) at the optimal working point
of the homogeneous BEI model found at the previous stage (G = 3.1), through the bias and scaling parameters introduced
above, denoted bE
Aβ and sE
Aβ for Aβ, and bE
τ and sE
τ for tau, all for the excitatory case, and similarly for the inhibitory case
with superscript I. We perform a search in parameter space with constraints introduced from experimental observations,
see Section 5.7, to find the optimal working point for the two protein burdens simultaneously, which results in an 8-degree
of freedom optimization, which is reduced to six degrees due to the constraints. For the optimization we used Bayesian
optimization algorithm using Gaussian Processes, see Section 5.10. We can also perform a simplified search, limited to
the two-variable bI
Aβ and sI
Aβ space, i.e., the inhibitory bias and scaling of the Aβ influence on inhibitory neuron parameters
(Equation 9). In this case, the 2D optimization results show a decreasing the neuronal activity with increasing Aβ concentration,
confirming previous results [21]. On average, for each group of subjects, we got the results shown numerically in Table 2.
Cohort
bE
Aβ
sE
Aβ
bE
τ
sE
τ
bI
Aβ
sI
Aβ
AD
0.2 (0.5)
2.3 (1.2)
-0.4 (0.6)
-2.6 (0.8)
0.2 (0.6)
-2.5 (0.8)
MCI
0.4 (0.7)
1.7 (1.5)
-0.5 (0.5)
-2.8 (0.7)
-0.1 (0.8)
-2.1 (1.2)
HC
0.1 (0.8)
1.7 (0.9)
-0.5 (0.6)
-2.8 (1.0)
0.3 (0.6)
-3.1 (1.0)
Table 2. Resulting averaged parameters from the optimization procedure. In parenthesis, the respective standard deviations.
These results can be seen visually at Figure 3. This figure shows that there is a clear regime in which all three empirical
properties are fitted well by the model, particularly for the values shown above, where a fitting of phFCD of 0.13 is achieved for
the AD subjects, while the reference homogeneous value is equal to 0.5.
3.3 Analysis of burden impact
For the optimal parameter values resulting from model fitting, we simulated each dynamical model 10 times for each subject to
account for the inherent stochastic nature of the models and compute the respective measures of model fit. Figure 4 shows
the distributions of fit statistics across runs for the homogeneous and the heterogeneous model for the different cohorts. In
addition, we show results for a null ensemble of models in which the regional burden values were spatially shuffled to generate
surrogates with the same spatial autocorrelation as the empirical data. Across the benchmark property to which the data were
fitted –—phFCD-––, the models taking into account the regional burden heterogeneity perform better than the homogeneous
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 5/20
model (all pass the Mann-Whitney U rank test on two independent samples with p < .0005). We also find a consistent gradient
of performance across all benchmarks, with the heterogeneous model performing best, and the homogeneous model showing
the poorest performance. For each benchmark metric, the performance of the heterogeneous model was better than all other
models (in all cases p < .06). Also, it must be noted that the differences in fit statistics between models are significant, as
shown in Figure 4. For example, for the AD cohort, the correlation of the median phase FCD between the fitted model and
empirical data showed r < 0.1 for the heterogeneous model, and r ≈ 0.2 for the BEI model. In all subject groups, the difference
between these two models is clear, with p < 0.0005.
Finally, we performed an analysis comparing the impact of each type of burden, in isolation or together, onto the simulation
results. In Figures 2D-2F we can see these results for the different cohorts, for Aβ and tau, Aβ alone, tau alone and finally the
homogeneous BEI model, added for reference. As we can see, with respect to the homogeneous model, the best performance is
systematically obtained by the combined action of both Aβ and tau, giving a value with p < 0.0004 in all cases. However, for
each cohort, each protein shows to play a different role in the development of the disease. For AD subjects, the effect of Aβ on
the optimal combined result is small, with a p < 0.0005, while the influence of tau alone has a p value that does not allow us to
distinguish between its effect and the combined effect of both proteins (p = 0.172), implying a clear dominance of tau over Aβ
in this stage of the disease. Also, with respect to the homogeneous BEI model, tau presents p < 0.005, while Aβ alone shows a
much higher value (p = 0.339), not allowing us to clearly distinguish between these two models. In the case of the MCI cohort,
in Figure 2E, we can observe that the effect of Aβ alone clearly gives the major contribution to the final combined fitting, rather
than tau, with a p < 0.0003 between all cases. Finally, in the HC case in Figure 2D, the effects of the Aβ and tau proteins are
close to the homogeneous BEI model, with Aβ presenting a somewhat higher prevalence than tau. However, it is noticeable
that the differences between this case and the previous one are small, showing that Aβ already plays an important role even in
HC subjects.
4. Discussion
In this paper we studied the influence of the regional variability of two pathological proteins, namely Aβ and tau, on cortical
activity and E/I balance in the context of AD. The incorporation of such heterogeneous patterns of neuropathology into
whole-brain models of neuronal dynamics has been made possible by the availability of in-vivo quantitative PET imaging. We
have shown that the heterogeneous model incorporating both types of neuropathological burdens more faithfully reproduces
empirical properties of dynamic FC than the standard model with fixed and homogeneous parameters. Our findings highlight a
central role of both types of burden on the regional neuronal dynamics in AD, supporting the hypothesis of hyper excitation in
AD, and the crucial role of E/I balance. Regarding their influence on brain activity, our results have shown a dominance of Aβ
influence on neural dynamics in earlier stages of AD (i.e., MCI) and even in healthy controls, while the tau influence plays
a larger role in later stages. These key findings highlight their prominent role in contributing to the abnormal brain activity
patterns in the course of AD [39].
4.1 How does burden heterogeneity shape neuronal dynamics?
We introduced burden heterogeneity into our dynamical model by modifying the regional excitability of local population
activity. We achieved this by modifying each region’s gain response function Mi of inhibitory and excitatory populations, in
accordance with previous works exploring the effect of regional parameters on E/I balance [32], thus focusing on how the
interaction of neuronal populations contributes to neuronal dynamics (i.e., FC or FCD) and their relative impacts over time. Our
approach is different from the work by Stefanovski et al. [21], where the Aβ burden was used to modulate regional E/I balance
by negatively modulating the inhibitory time constant, increasing excitatory activity and producing a higher output of the
pyramidal cell populations, resulting in a local hyperexcitation with high Aβ loads. However, as seen in Methods, our results
confirm their findings with respect to the behavior of the Aβ burden in early stages of the disease, resulting in a net increase of
the excitatory activity with increased Aβ burden. There are other approaches available to introduce heterogeneity, such as
an adjustment of the inter-node connectivity to fit empirical and simulated FCs [40]; or variations of within- and inter-area
connectivity [41]. However, based on the empirical evidence that the interplay of both burdens, Aβ and tau, severely disrupt
normal neuronal function, we decided to model their direct effect on the E/I balance.
In this paper we have chosen to incorporate heterogeneity into the model by modulating population gain response functions
H(E,I), since local variations in the E/I balance will affect the net excitability of the population, which in turn is captured by
the gain function parameter, Mi. We thus assume that changes in regional gain are the common final pathways of different
neuropathology-related pathomechanisms which might have an influence on specific neuronal populations or interaction
between populations.
In particular, we introduced regional variations of Mi as the product of linear terms consisting of a constant (bias), and
a scaling factor. This introduced eight degrees of freedom, which we could narrow down to six due to constraints based
on previous literature [11], which helped to considerably reduce the search space. In sum, our model was created based
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 6/20
the assumptions that Aβ leads to GABAergic interneuron dysfunction and impaired glutamate reuptake, while tau leads to
educed synaptic neurotransmitter release in excitatory cells. This amount of degrees is substantially less than used in other
models [41, 40], making a fast parameter optimization feasible. For the optimization we used Bayesian optimization using
Gaussian Processes (see Methods), because of the many minima that could trap traditional optimization methods.
4.2 Evaluating Aβ and tau impact
A large body of scientific literature focused on linking global and local brain dynamics to individual differences in cognitive
performance scores [18], showing that patients with AD and MCI show less variation in neuronal connectivity during resting-
state [42], and even presented benchmarks for predictive models for resting-state fMRI, revealing biomarkers of individual
psychological or clinical traits [19]. However, the pattern of neuronal connectivity alterations has been incompletely understood.
More recent work focus on the effect of Aβ on hyperexcitability, addressing the fact that this protein modulates regional E/I
balance, resulting in local hyperexcitation with high loads [21]. To our knowledge, no prior study has evaluated both types of
neuropathological burden, Aβ and tau, simultaneously.
As explained in Methods, we compared the impact of each type of burden, in isolation or interacting, onto neural dynamics.
We found that the model fitting optimum is systematically obtained by the interaction of both burdens. Also, we have found
that for each condition (i.e., HC, MCI or AD), each protein has a different impact on the disease. In the case of AD, Aβ has a
small impact on the combined result, while tau alone had almost all of the impact, showing its dominance over Aβ. Also, in
comparison to the homogeneous BEI model, in we observed that tau is clearly distinguishable, but Aβ is not. Taken together,
these results imply that we cannot distinguish between the effect on brain activity of both proteins together vs. the effect of tau
alone, while the effect of Aβ is clearly distinguishable from the combined effect. As a consequence, this allows us to conclude
that the impact of tau in this stage (AD) of the disease is clearly dominant over Aβ. In MCI, the influence of Aβ alone is clearly
dominant over tau, see Figure 2E. Finally, when studying the effect of both proteins in the HC case, we can observe that the
effect of the Aβ and tau proteins is close to the homogeneous BEI model, with Aβ presenting a relatively higher influence than
tau. The influence of Aβ both in MCI patients as well as in HC shows that Aβ leads to a measurable change in brain dynamics,
independent of existing cognitive impairment, in elderly subjects. Despite our findings from model fitting, we acknowledge
that we only observe the current influence of Aβ vs. tau in different disease stages in a cross-sectional cohort. Longitudinal
examinations might also replicate the abundant evidence in the literature [11] that both proteins interplay a toxic feedback loop
which is the ultimate responsible (perhaps among other factors) of the development of the disease.
Our analysis shows that edge-level measures of static FC offer loose constraints for model optimization, showing comparably
high fit statistics across a broad range of values of the global coupling parameter. In contrast, fitting to dynamical functional
connectivity shows a clear optimum, mirroring similar results reported previously [43, 32]. We can conclude that fitting models
to both static and dynamic properties is thus important for identifying an appropriate working point for each model.
Across all these properties, we observe that the model that incorporates the heterogeneous burden loads provides a better
match to the data than the homogeneous BEI model, which does not incorporate a fitting of the gain response function of
inhibitory and excitatory populations to the data. This shows that constraining regional heterogeneity by the protein burdens
yields a more faithful replication of empirical phFCD. The superiority of our model using heterogeneous, empirically estimated
parameters, suggests that regional heterogeneity plays a significant role in shaping the effects of Alzheimer’s disease on
spontaneous BOLD-dynamics. However, as we already mentioned, it must be noted that the differences in fit statistics between
models are significant. These results suggest that these empirical fit statistics have good capacity to tease apart dynamical
differences between models, which gives the opportunity to disentangle the influence of different pathomechanisms in vivo.
We observe that, in all cases, the bias parameters for the different burdens (Figure 3) are approximately 0 in all cases,
thus indicating that the influence of the bias parameters with respect to the homogeneous model can be ignored, reducing
computational complexity. The respective scaling parameters take non negligible values, showing a linear relationship between
Aβ and tau on neural dynamics. In our model, in earlier stages of the disease (i.e., MCI) Aβ has a higher scaling parameter
than tau, suggesting a higher contribution to the E/I imbalance. In later stages, we observe the opposite, which might indicate
that tau burden is more closely related to neuronal dysfunction in these stages, which replicates our results regarding the model
fitting using different types of heterogeneous models. We acknowledge that on a pathophysiological level there is a strong
interplay between Aβ and tau and further (causal) research is needed to clearly discern the role each protein plays in the
generation of neuronal dysfunction.
In summary, in this paper we have presented a whole-brain computational model connecting the main protein burdens,
namely Aβ and tau, with the different stages of AD and in HC. The results we obtained not only reproduce previous research
regarding E/I imbalance in AD, but also shed further light on the relative impact of each type of burden during different disease
stages, opening new avenues to focus research efforts. As a general conclusion, our study shows that whole-brain modeling
enables research on disease mechanisms in-vivo, demonstrating its potential to produce improved diagnostics and help in the
discovery of new treatments.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 7/20
5. Methods
5.1 Participants
Empirical data were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.
edu), which is a longitudinal multi-site study designed to develop biomarkers for Alzheimer’s disease (AD) across all stages.
The inclusion criteria for AD patients was the NINCDS-ADRDA criteria, which contains only clinical features [4], and a
MMSE score below 24. For both HC and MCI, the inclusion criteria were a MMSE (Mini Mental State Examination) score
between 24-30, as well as age between 55-90 years. Also, for MCI, participants had to have a subjective memory complaint
and abnormal results in another neuropsychological memory test. Imaging and biomarkers were not used for the diagnosis.
5.2 Data Acquisition and Processing
All the data in this study were previously used in Stefanovski et al. [21] work, so we will present here an abridged version of
the processing performed on the original data and refer to the original work for the details. All images used in this study were
taken from ADNI-3, using data from Siemens scanners with a magnetic field strength of 3T.
5.2.1 Structural MRI
For each included participant, we created a brain parcellation for our structural data using FLAIR, following the minimal
preprocessing pipeline [44] of the Human Connectome Project (HCP) using Freesurfer1 [45], FSL [46, 47, 48] and connectome
workbench2. Therefore, we used T1 MPRAGE, FLAIR and fieldmaps for the anatomical parcellation. We then registered the
subject cortical surfaces to the parcellation of Glasser et al. [49] using the multimodal surface matching (MSM) tool [50]. In
this parcellation, there were 379 regions: 180 left and 180 right cortical regions, 9 left and 9 right subcortical regions, and 1
brainstem region.
5.2.2 PET Images
For Aβ, we used the version of AV-45 PET already preprocessed by ADNI, using a standard image with a resolution of 1.5mm
cubic voxels and matrix size of 160×160×96, normalized so that the average voxel intensity was 1 and smoothed out using
a scanner-specific filter function. Then, a brainmask was generated from the structural preprocessing pipeline (HCP) and
used to mask the PET image. On the other hand, to obtain the local burden of Aβ, we computed the relative intensity to the
cerebellum. We received in each voxel a relative Aβ burden which is aggregated according to the parcellation used for our
modeling approach. Subcortical region PET loads were defined as the average SUVR in subcortical gray matter (GM). With the
help of the connectome workbench tool, using the pial and white matter surfaces as ribbon constraints, we mapped the Cortical
GM PET intensities onto individual cortical surfaces. Finally, using the multimodal Glasser parcellation we derived average
regional PET loads.
For tau, we also used ADNI’s preprocessed version of AV-1451 (Flortaucipir) following the same acquisition and processing,
resulting in a single relative tau value for each voxel. Then, these values were also aggregated to the selected parcellation, also
following the already mentioned steps. The final average regional tau loads were obtained in the Glasser parcellation.
5.2.3 DWI
Individual tractographies were computed only for included HC participants, and they were averaged to a standard brain template
(see below). Preprocessing was mainly done with the MRtrix3 software package3.
In particular, the following steps were performed: First, we denoised the DWI data [51], followed by motion and eddy
current correction4. Then, B1 field inhomogeneity correction (ANTS N4), followed by a brainmask estimation from the
DWI images. Next, we normalized the DWI intensity for the group of participants, which was used to generate a WM
response function [52], and created an average response function from all participants. Next, we estimated the fiber orientation
distribution and the average response function [53] using the subject normalized DWI image, to finally generate a five tissue
type image. Finally, we used the iFOD2 algorithm [54] and the SIFT2 algorithm [55] to get the weighted anatomical constrained
tractography [56], to end up merging all information into the Glasser connectome, resulting in a structural connectome (SC).
5.2.4 fMRI
With respect to the processing of the fMRI data, the images were initially preprocessed in FSL FEAT and independent
component analysis–based denoising (FSLFIX) following a basic pipeline [21]. Time courses for noise-labeled components,
along with 24 head motion parameters, were then removed from the voxel-wise fMRI time series using ordinary least squares
regression.
1https://surfer.nmr.mgh.harvard.edu/fswiki/FreeSurferMethodsCitation
2https://www.humanconnectome.org/software/connectome-workbench
3http://www.mrtrix.org
4https://mrtrix.readthedocs.io/en/latest/dwi_preprocessing/dwipreproc.html
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 8/20
The resulting denoised functional data were spatially normalized to the MNI space using Advanced Normalization Tools
(version 2.2.0). Mean time series for each parcellated region were then extracted, and interregional FC matrices were estimated
using Pearson correlations between each pair of regional time series. Dynamic FC matrices were also calculated for the
empirical data, as outlined below.
5.3 Generation of a Standard Brain Template
As previously done [21], we average the SCs of all HC participants, using an arithmetic mean
Cµ = (
n
∑
i=1
Ci)/n = (C1 +C2 +...+Cn)/n
wherein Cµ is the averaged SC matrix, n is the number of HC participants and Ci is the individual SC matrix.
However, as matrices in this context are large (i.e., 379 regions), the average input to any given node can be too large for the
DMF, making fitting and processing in general more difficult. Thus, we discarded the traditional normalization of dividing the
matrix elements by its maximum, and used a slightly different approach, instead. First, we added one and applied the logarithm
to every entry, as lC = log(Cµ +1). Then, we computed the maximum input any node could receive for a unitary unit input
current, maxNodeInput = max j(∑i(lCi, j)), and finally we normalized by 0.7∗lC/maxNodeInput, where 0.7 was chosen to be
a convenient normalization value. Observe that this constant is actually multiplying another constant G in the model which we
fit to empirical data, so its actual value can safely be changed.
In Figure 5 we can find the SC matrix and organization graph, where we can observe that the general characteristics of
physiological SCs such as symmetry, laterality, homology, and subcortical hubs are maintained in the averaged connectome.
The election of the averaged SC allowed to control all factors (e.g., atrophy), which matched our objective of simulating the
activity from both healthy and “pathogenic” modifications by Aβ and tau.
5.4 Balanced Excitation-Inhibition (BEI) model
In this work we used the Dynamic Mean Field (DMF) model proposed by Deco et al. [31], which consists of a network model
to simulate spontaneous brain activity at the whole-brain level. In this model, each node represents a brain area and the links
represent the white matter connections between them. In turn, each node is a reduced representation of large ensembles of
interconnected excitatory and inhibitory integrate-and-fire spiking neurons (80% and 20% neurons, respectively), to a set of
dynamical equations describing the activity of coupled excitatory (E) and inhibitory (I) pools of neurons, based on the original
reduction of Wong and Wang [58]. In the DMF model, excitatory synaptic currents, I(E), are mediated by NMDA receptors,
while inhibitory currents, I(I), are mediated by GABAA receptors. Both neuronal pools are reciprocally connected, and the
inter-area interactions occur at the excitatory level only, scaled by the structural connectivity Ck j (see Section 5.2.1).
To be more specific, the DMF model is expressed by the following system of coupled differential equations:
I(E)
k
= WE Io +w+ JN S(E)
k
+JNG∑
j
Ck jS(E)
j
−JkS(I)
k +Iext
(1)
I(I)
k
= WI Io +JNS(E)
k
−S(I)
k +λJNG∑
j
Ck jS(E)
j
(2)
r(E)
k
= H(E)(I(E)
k
) =
ME
k (aEI(E)
k
−bE)
1−exp(−dEME
k (aEI(E)
k
−bE))
(3)
r(I)
k
= H(I)(I(I)
k ) =
MI
k(aII(I)
k −bI)
1−exp(−dIMI
k(aII(I)
k −bI))
(4)
˙S(E)
k
= −S(E)
k
τE
+(1−S(E)
k
)γH(E)(I(E)
k
)
(5)
˙S(I)
k
= −S(I)
k
τI
+H(I)(I(I)
k )
(6)
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 9/20
Here, the last two equations should add, when integrating, an uncorrelated standard Gaussian noise term with an amplitude of
σ = 0.01nA (using Euler-Maruyama integration). In these equations, λ is a parameter that can be equal to 1 or 0, indicating
whether long-range feedforward inhibition is considered (λ = 1) or not (λ = 0).
As mentioned, the DMF model is derived from the original Wong and Wang model [58] to emulate resting-state conditions,
such that each isolated node displays the typical noisy spontaneous activity with low firing rate (H(E) ∼ 3Hz) observed in
electrophysiology experiments, reusing most of the parameter values defined there. We also implemented the Feedback
Inhibition Control (FIC) mechanism described by Deco et al. [31], where the inhibition weight, Jn, was adjusted separately for
each node n such that the firing rate of the excitatory pools H(E) remains clamped at 3Hz even when receiving excitatory input
from connected areas. Deco et al. [31] demonstrated that this mechanism leads to a better prediction of the resting-state FC and
to a more realistic evoked activity. We refer to this model as the balanced excitation-inhibition (BEI) model. Although the local
adjustments in this model introduce some degree of regional heterogeneity, the firing rates are constrained to be uniform across
regions so we consider this BEI model as a homogeneous benchmark against which we evaluate more sophisticated models that
allow Aβ and tau to affect intrinsic dynamical properties across regions.
Following the Glasser parcellation [44], we considered N = 379 brain areas in our whole-brain network model. Each area n
receives excitatory input from all structurally connected areas into its excitatory pool, weighted by the connectivity matrix,
obtained from dMRI (see Section 5.2.3). Furthermore, all inter-area E-to-E connections are equally scaled by a global coupling
factor G. This global scaling factor is the only control parameter that is adjusted to move the system to its optimal working
point, where the simulated activity maximally fits the empirical resting-state activity of healthy control participants. Simulations
were run for a range of G between 0 and 5.5 with an increment of 0.05 and with a time step of 1 ms. For each G, we ran 200
simulations of 435 s each, in order to emulate the empirical resting-state scans from 17 participants. The optimum value found,
for the phFCD observable, is for G = 3.1. See Figure 2A.
5.5 Simulated BOLD signal
Once we have obtained the simulated mean field activity, we need to transform it into a BOLD signal we used the generalized
hemodynamic model of Stephan et al. [59]. We compute the BOLD signal in the k-th brain area from the firing rate of the
excitatory pools H(E), such that an increase in the firing rate causes an increase in a vasodilatory signal, sk, that is subject
to auto-regulatory feedback. Blood inflow fk responds in proportion to this signal inducing changes in blood volume vk and
deoxyhemoglobin content qk. The equations relating these biophysical variables are:
dsk
dt = 0.5r(E)
k
+3−ksk −γ( fk −1)
d fk
dt = sk
τ dvk
dt = fk −vα−1
k
τ dqk
dt = fk
1−(1−ρ)f −1
k
ρ
−qk
vα−1
k
vk
(7)
with finally
Bk = v0
�
k1(1−qk)+k2(1− qk
vk
)+k3(1−vk)
�
being the final measured BOLD signal.
We actually used the updated version described later on [59], which consists on introducing the change of variables ˆz = lnz,
which induces the following change for z = fk, vk and qk, with its corresponding state equation dz
dt = F(z), as:
dˆz
dt = d ln(z)
dz
dz
dt = F(z)
z
which results in z(t) = exp(ˆz(t)) always being positive, which guarantees a proper support for these non-negative states, and
thus numerical stability when evaluating the state equations during evaluation.
5.6 Aβ-Tau model:
In our heterogeneous model, Aβ and Tau are introduced, at the formulae for the neuronal response functions, H(E,I) (excita-
tory/inhibitory), into the gain factor M(E,I)
k
for the k-th area as
ME
k = (1+bE
Aβ +sE
AβAβk)(1+bE
τ +sE
τ tauk)
(8)
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 10/20
MI
k = (1+bI
Aβ +sI
AβAβk)(1+bI
τ +sI
τtauk)
(9)
where b(E,I)
(Aβ,τ) are the excitatory/inhibitory Aβ and tau bias parameters, while s(E,I)
(Aβ,τ) are the respective scaling factors.
These are the 8 (from which actually only 6 are needed as tau only affects excitatory neurons [60], see next section) parameters
that we will optimize for each subject individually.
5.7 Constraints
Based on previous neuroscientific experiments [11], we included constraints on the direction of effect of Aβ and tau (i.e.,
inhibitory vs. excitatory influence). We introduced the following constraints:
• Aβ produces inhibitory GABAergic interneuron dysfunction [61, 62], thus we can infer that sI
Aβ < 0.
• Aβ produces impaired glutamate reuptake [61, 62], so we can introduce the bound sE
Aβ > 0.
• Tau appears to target excitatory neurons [60], so we can safely consider that bI
τ = sI
τ = 0.
• Tau binds to synaptogyrin-3, reducing excitatory synaptic neurotransmitter release [63], thus sE
τ < 0.
Although the interplay between Aβ and tau is not completely known [11], but there is evidence that Aβ promotes tau
by cross-seeding [64, 65], thus the cross term factors (i.e., the ones resulting from the multiplication of Aβ and tau scaling
parameters) play a crucial role to elucidate the final impact of the combined burden.
5.8 Observables
edge-centric FC
The static edge-level FC is defined as the N ×N matrix of BOLD signal correlations between brain areas
computed over the entire recording period (see Figure 5). We computed the empirical FC for each human participant and for
each simulated trial, as well as for the group-averages SC matrix of the healthy cohort. All empirical and simulated FC matrices
were compared by computing the Pearson correlation between their upper triangular elements (given that the FC matrices are
symmetric).
swFCD
The most common and straightforward approach to investigate the temporal evolution of FC is the sliding-window
FC dynamics (swFCD) [66, 67, 68, 69, 70, 43]. This is achieved by calculating the correlation matrix, FC(t), restricted to a
given time-window (t −x : t +x), and successively shifting this window in time resulting in a time-varying FCNxNxT matrix
(where N is the number of brain areas and T the number of time windows considered). Here, we computed the FC over a sliding
window of 30 TRs (corresponding approximately to 1.5 minutes) with incremental shifts of 3 TRs. This FCD matrix is defined
so that each entry, (FCD(tx,ty)) corresponds to the correlation between the FC centered at times tx and the FC centered at ty. In
order to compare quantitatively the spatio-temporal dynamical characteristics between empirical data and model simulations,
we generate the distributions of the upper triangular elements of the FCD matrices over all participants as well as of the FCD
matrices obtained from all simulated trials for a given parameter setting. The similarity between the empirical and model FCD
distributions is then compared using the KS distance, DKS, allowing for a meaningful evaluation of model performance in
predicting the changes observed in dynamic resting-state FC. However, the fundamental nature of the swFCD technique implies
the choice of a fixed window length, which limits the analysis to the frequency range below the window period, so the ideal
window length to use remains under debate [71, 72, 73, 74, 75].
phFCD
In an attempt to overcome the limitations of the sliding-window analysis, a few methods were proposed to estimate
the FC(t) at the instantaneous level. For instance, phase Functional Connectivity Dynamics (phFCD) consists in computing
the phase coherence between time series at each recording frame [76, 77, 78, 35]. In brief, the instantaneous BOLD phase
of area n at time t, θn(t), is estimated using the Hilbert transform. Given the phase, the angle between two BOLD signals is
given by their absolute phase difference: Θnp = |θn(t)−θp(t)|. Then, the phFCD(t) between a pair of brain areas n and p is
calculated as:
phFCDnp(t) = cos(Θnp(t)),n, p ∈ N = 1,.. .,N
with N the number of brain regions considered in the parcellation used.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 11/20
5.9 2D Aβ Optimization
We can use our model to verify the results by Stefanovski et al. [21] by limiting our analysis to the parameters of Aβ at the
inhibitory level (i.e., the inhibitory bias bI
Aβ and scaling sI
Aβ parameters only, defined in Equation 9). This way, we can replicate,
up to a certain degree, the results from that paper, being limited by the fact that we use a different model, based on the BEI
model instead of the Jansen-Rit model [34]; a different expression for the burden, i.e., a linear approximation instead of a
sigmoid; different units, etc. See Figure 3. By analyzing the obtained data at the optimal fit, the same behavior of decreasing
the neuronal activity of inhibitory neurons with the scaling parameter sI
Aβ, corresponding to an increase in Aβ concentration,
can be observed, as shown in Figure 6.
5.10 Full Optimization
To efficiently optimize the 6-dimensional function described before for the three bias and scaling values, we used a Bayesian
optimization algorithm using Gaussian Processes (GP), which approximates the function using a multivariate Gaussian. In
particular, our implementation uses the gp optimize method from the scikit-optimize Python library, which uses a GP kernel
between the parameters to obtain the covariance of the function values. With this information, the algorithm chooses the next
parameter to evaluate by selecting the acquisition function over the Gaussian prior.
Data and code availability statement
All code for implementing computational models and reproducing our results is available at https://github.com/
dagush/WholeBrain
CRediT authorship contribution statement
Gustavo Patow: Conceptualization, Formal analysis, Software, Writing – original draft, Writing – review & editing. Leon
Stefanovski: Data Curation, Writing – review & editing. Petra Ritter: Data Curation, Writing – review & editing. Gustavo
Deco: Conceptualization, Writing – review & editing. Xenia Kobeleva: Conceptualization, Writing – review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Funding
This research was partially funded by grant PID2021-122136OB-C22 from the Ministerio de Ciencia e Innovaci´on, Spain
of GP. This work was supported by an add-on fellowship of the Joachim Herz Foundation of XK. PR had the support of the
following grants: H2020 Research and Innovation Action Grant Human Brain Project SGA2 785907 (PR), H2020 Research and
Innovation Action Grant Human Brain Project SGA3 945539 (PR), H2020 Research and Innovation Action Grant Interactive
Computing E-Infrastructure for the Human Brain Project ICEI 800858 (PR), H2020 Research and Innovation Action Grant
EOSC VirtualBrainCloud 826421 (PR), H2020 Research and Innovation Action Grant AISN 101057655 (PR), H2020 Research
Infrastructures Grant EBRAINS-PREP 101079717 (PR), H2020 European Innovation Council PHRASE 101058240 (PR),
H2020 Research Infrastructures Grant EBRAIN-Health 101058516 (PR), H2020 European Research Council Grant ERC
BrainModes 683049 (PR), JPND ERA PerMed PatternCog 2522FSB904 (PR), Berlin Institute of Health & Foundation Charit´e
(PR), Johanna Quandt Excellence Initiative (PR), German Research Foundation SFB 1436 (project ID 425899996) (PR),
German Research Foundation SFB 1315 (project ID 327654276) (PR), German Research Foundation SFB 936 (project ID
178316478) (PR), German Research Foundation SFB-TRR 295 (project ID 424778381) (PR), German Research Foundation
SPP Computational Connectomics RI 2073/6-1, RI 2073/10-2, RI 2073/9-1 (PR).
Acknowledgements
Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National
Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012).
ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering,
and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery
Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai
Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 12/20
Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.;
Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale
Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal
Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research
is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for
the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research
and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern
California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.
References
[1] H. Braak and E. Braak. Neuropathological stageing of alzheimer-related changes. Acta Neuropathologica, 82(4):239–259,
sep 1991.
[2] G.L. Wenk. Neuropathologic changes in alzheimer’s disease: potential targets for treatment. J Clin Psychiatry, 67(3):3–7,
2006.
[3] Colin L. Masters, Randall Bateman, Kaj Blennow, Christopher C. Rowe, Reisa A. Sperling, and Jeffrey L. Cummings.
Alzheimer's disease. Nature Reviews Disease Primers, 1(1), October 2015.
[4] G. McKhann, D. Drachman, M. Folstein, R. Katzman, D. Price, and E. M. Stadlan. Clinical diagnosis of alzheimer’s
disease: Report of the NINCDS-ADRDA work group under the auspices of department of health and human services task
force on alzheimer’s disease. Neurology, 34(7):939–939, July 1984.
[5] Leon Stefanovski, Jil Mona Meier, Roopa Kalsank Pai, Paul Triebkorn, Tristram Lett, Leon Martin, Konstantin B¨ulau,
Martin Hofmann-Apitius, Ana Solodkin, Anthony Randal McIntosh, and Petra Ritter. Bridging scales in alzheimer's
disease: Biological framework for brain simulation with the virtual brain. Frontiers in Neuroinformatics, 15, April 2021.
[6] Clifford R. Jack, David A. Bennett, Kaj Blennow, Maria C. Carrillo, Billy Dunn, Samantha Budd Haeberlein, David M.
Holtzman, William Jagust, Frank Jessen, Jason Karlawish, Enchi Liu, Jose Luis Molinuevo, Thomas Montine, Creighton
Phelps, Katherine P. Rankin, Christopher C. Rowe, Philip Scheltens, Eric Siemers, Heather M. Snyder, Reisa Sperling,
Cerise Elliott, Eliezer Masliah, Laurie Ryan, and Nina Silverberg. NIA-AA research framework: Toward a biological
definition of alzheimer’s disease. Alzheimer’s & Dementia, 14(4):535–562, April 2018.
[7] Johannes Weickenmeier, Ellen Kuhl, and Alain Goriely. Multiphysics of prionlike diseases: Progression and atrophy. Phys.
Rev. Lett., 121:158101, Oct 2018.
[8] G. Caleb Alexander and Jason Karlawish. The problem of aducanumab for the treatment of alzheimer disease. Annals of
Internal Medicine, 174(9):1303–1304, sep 2021.
[9] Lars M. Ittner and J¨urgen G¨otz. Amyloid-β and tau — a toxic pas de deux in alzheimer's disease. Nature Reviews
Neuroscience, 12(2):67–72, December 2010.
[10] Margaret Lock. The Alzheimer conundrum : entanglements of dementia and aging. Princeton University Press, Princeton,
2013.
[11] Marc Aurel Busche and Bradley T. Hyman.
Synergy between amyloid-β and tau in alzheimer’s disease.
Nature
Neuroscience, 23(10):1183–1193, August 2020.
[12] Eva Vico Varela, Guillaume Etter, and Sylvain Williams. Excitatory-inhibitory imbalance in alzheimer's disease and
therapeutic significance. Neurobiology of Disease, 127:605–615, July 2019.
[13] Fernando Maest´u, Willem de Haan, Marc Aurel Busche, and Javier DeFelipe. Neuronal excitation/inhibition imbalance:
core element of a translational perspective on alzheimer pathophysiology. Ageing Research Reviews, 69:101372, August
2021.
[14] Che-Wei Chang, Mark D. Evans, Xinxing Yu, Gui-Qiu Yu, and Lennart Mucke. Tau reduction affects excitatory and
inhibitory neurons differently, reduces excitation/inhibition ratios, and counteracts network hypersynchrony. Cell Reports,
37(3):109855, October 2021.
[15] Danlei Bi, Lang Wen, Zujun Wu, and Yong Shen. GABAergic dysfunction in excitatory and inhibitory (e/i) imbalance
drives the pathogenesis of alzheimer's disease. Alzheimer’s & Dementia, 16(9):1312–1329, June 2020.
[16] Alexandra L Petrache, Aarib Rajulawalla, Anqi Shi, Andrea Wetzel, Takashi Saito, Takaomi C Saido, Kirsten Harvey, and
Afia B Ali. Aberrant excitatory–inhibitory synaptic mechanisms in entorhinal cortex microcircuits during the pathogenesis
of alzheimer’s disease. Cerebral Cortex, 29(4):1834–1850, February 2019.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 13/20
[17] Julie C. Lauterborn, Pietro Scaduto, Conor D. Cox, Anton Schulmann, Gary Lynch, Christine M. Gall, C. Dirk Keene,
and Agenor Limon. Increased excitatory to inhibitory synaptic ratio in parietal cortex samples from individuals with
alzheimer’s disease. Nature Communications, 12(1), May 2021.
[18] J. Zimmermann, A. Perry, M. Breakspear, M. Schirner, P. Sachdev, W. Wen, N.A. Kochan, M. Mapstone, P. Ritter, A.R.
McIntosh, and A. Solodkin. Differentiation of alzheimer’s disease based on local and global parameters in personalized
virtual brain models. NeuroImage: Clinical, 19:240 – 251, 2018.
[19] Kamalaker Dadi, Mehdi Rahim, Alexandre Abraham, Darya Chyzhyk, Michael Milham, Bertrand Thirion, and Ga¨el
Varoquaux. Benchmarking functional connectome-based predictive models for resting-state fmri. NeuroImage, 192:115 –
134, 2019.
[20] Murat Demirtas¸, Joshua B. Burt, Markus Helmer, Jie Lisa Ji, Brendan D. Adkinson, Matthew F. Glasser, David C. Van
Essen, Stamatios N. Sotiropoulos, Alan Anticevic, and John D. Murray. Hierarchical heterogeneity across human cortex
shapes large-scale neural dynamics. Neuron, 101(6):1181–1194.e13, March 2019.
[21] Leon Stefanovski, Paul Triebkorn, Andreas Spiegler, Margarita-Arimatea Diaz-Cortes, Ana Solodkin, Viktor Jirsa,
Anthony Randal McIntosh, Petra Ritter, and for the Alzheimer’s Disease Neuroimaging Initiative . Linking molecular
pathways and large-scale computational modeling to assess candidate disease mechanisms and pharmacodynamics in
alzheimer’s disease. Frontiers in Computational Neuroscience, 13:54, 2019.
[22] Paul Triebkorn, Leon Stefanovski, Kiret Dhindsa, Margarita-Arimatea Diaz-Cortes, Patrik Bey, Konstantin B¨ulau, Roopa
Pai, Andreas Spiegler, Ana Solodkin, Viktor Jirsa, Anthony Randal McIntosh, and Petra Ritter and. Brain simulation
augments machine-learning–based classification of dementia. Alzheimer’s & Dementia: Translational Research & Clinical
Interventions, 8(1), January 2022.
[23] Gustavo Deco and Morten L. Kringelbach. Great expectations: Using whole-brain computational connectomics for
understanding neuropsychiatric disorders. Neuron, 84(5):892–905, December 2014.
[24] Matthieu Gilson, Nikos E. Kouvaris, Gustavo Deco, Jean-Franc¸ois Mangin, Cyril Poupon, Sandrine Lefranc, Denis Rivi`ere,
and Gorka Zamora-L´opez. Network analysis of whole-brain fMRI dynamics: A new framework based on dynamic
communicability. NeuroImage, 201:116007, November 2019.
[25] Morten L. Kringelbach and Gustavo Deco. Brain states and transitions: Insights from computational neuroscience. Cell
Reports, 32(10):108128, September 2020.
[26] Anandamohan Ghosh, Y. Rho, A. R. McIntosh, R. K¨otter, and V. K. Jirsa. Noise during rest enables the exploration of the
brain's dynamic repertoire. PLoS Computational Biology, 4(10):e1000196, October 2008.
[27] G. Deco, V. Jirsa, A. R. McIntosh, O. Sporns, and R. Kotter. Key role of coupling, delay, and noise in resting brain
fluctuations. Proceedings of the National Academy of Sciences, 106(25):10302–10307, June 2009.
[28] C. J. Honey, O. Sporns, L. Cammoun, X. Gigandet, J. P. Thiran, R. Meuli, and P. Hagmann. Predicting human resting-state
functional connectivity from structural connectivity. Proceedings of the National Academy of Sciences, 106(6):2035–2040,
February 2009.
[29] G. Deco and V. K. Jirsa. Ongoing cortical activity at rest: Criticality, multistability, and ghost attractors. Journal of
Neuroscience, 32(10):3366–3375, March 2012.
[30] Ed Bullmore and Olaf Sporns. Complex brain networks: graph theoretical analysis of structural and functional systems.
Nature Reviews Neuroscience, 10(3):186–198, February 2009.
[31] Gustavo Deco, Adri´an Ponce-Alvarez, Patric Hagmann, Gian Luca Romani, Dante Mantini, and Maurizio Corbetta. How
local excitation–inhibition ratio impacts the whole brain dynamics. Journal of Neuroscience, 34(23):7886–7898, 2014.
[32] Gustavo Deco, Morten L. Kringelbach, Aurina Arnatkeviciute, Stuart Oldham, Kristina Sabaroedin, Nigel C. Rogasch,
Kevin M. Aquino, and Alex Fornito. Dynamical consequences of regional heterogeneity in the brain’s transcriptional
landscape. Science Advances, 7(29):eabf4752, July 2021.
[33] Paula Sanz Leon, Stuart Knock, M. Woodman, Lia Domide, Jochen Mersmann, Anthony McIntosh, and Viktor Jirsa. The
virtual brain: a simulator of primate brain network dynamics. Frontiers in Neuroinformatics, 7:10, 2013.
[34] Ben H. Jansen and Vincent G. Rit. Electroencephalogram and visual evoked potential generation in a mathematical model
of coupled cortical columns. Biological Cybernetics, 73(4):357–366, September 1995.
[35] Joana Cabral, Morten L. Kringelbach, and Gustavo Deco. Functional connectivity dynamically evolves on multiple
time-scales over a static structural connectome: Models and mechanisms. NeuroImage, 160:84–96, October 2017.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 14/20
[36] G. Deco, A. Ponce-Alvarez, D. Mantini, G. L. Romani, P. Hagmann, and M. Corbetta. Resting-state functional connectivity
emerges from structurally and dynamically shaped slow linear fluctuations. Journal of Neuroscience, 33(27):11239–11252,
July 2013.
[37] Andrew Zalesky, Alex Fornito, Ian H. Harding, Luca Cocchi, Murat Y¨ucel, Christos Pantelis, and Edward T. Bullmore.
Whole-brain anatomical networks: Does the choice of nodes matter? NeuroImage, 50(3):970–983, April 2010.
[38] Jesper L. R Andersson, Mark S. Graham, Enik˝o Zsoldos, and Stamatios N. Sotiropoulos. Incorporating outlier detection and
replacement into a non-parametric framework for movement and distortion correction of diffusion MR images. NeuroImage,
141:556–572, November 2016.
[39] Kathryn P. Riley, David A. Snowdon, and William R. Markesbery. Alzheimer's neurofibrillary pathology and the spectrum
of cognitive function: Findings from the nun study. Annals of Neurology, 51(5):567–577, apr 2002.
[40] Peng Wang, Ru Kong, Xiaolu Kong, Rapha¨el Li´egeois, Csaba Orban, Gustavo Deco, Martijn P. van den Heuvel, and
B.T. Thomas Yeo. Inversion of a large-scale circuit model reveals a cortical hierarchy in the dynamic resting human brain.
Science Advances, 5(1):eaat7854, January 2019.
[41] Rishidev Chaudhuri, Kenneth Knoblauch, Marie-Alice Gariel, Henry Kennedy, and Xiao-Jing Wang. A large-scale circuit
mechanism for hierarchical dynamical processing in the primate cortex. Neuron, 88(2):419–431, October 2015.
[42] Pablo N´u˜nez, Jes´us Poza, Carlos G´omez, V´ıctor Rodr´ıguez-Gonz´alez, Arjan Hillebrand, Miguel A Tola-Arribas, M´onica
Cano, and Roberto Hornero. Characterizing the fluctuations of dynamic resting-state electrophysiological functional
connectivity: reduced neuronal coupling variability in mild cognitive impairment and dementia due to alzheimer’s disease.
Journal of Neural Engineering, 16(5):056030, September 2019.
[43] Gustavo Deco, Morten L. Kringelbach, Viktor K. Jirsa, and Petra Ritter. The dynamics of resting fluctuations in the brain:
metastability and its dynamical cortical core. Scientific Reports, 7(1), June 2017.
[44] Matthew F. Glasser, Stamatios N. Sotiropoulos, J. Anthony Wilson, Timothy S. Coalson, Bruce Fischl, Jesper L. Andersson,
Junqian Xu, Saad Jbabdi, Matthew Webster, Jonathan R. Polimeni, David C. Van Essen, and Mark Jenkinson. The minimal
preprocessing pipelines for the human connectome project. NeuroImage, 80:105–124, October 2013.
[45] Martin Reuter, Nicholas J. Schmansky, H. Diana Rosas, and Bruce Fischl. Within-subject template estimation for unbiased
longitudinal image analysis. NeuroImage, 61(4):1402–1418, July 2012.
[46] Stephen M. Smith, Mark Jenkinson, Mark W. Woolrich, Christian F. Beckmann, Timothy E.J. Behrens, Heidi Johansen-
Berg, Peter R. Bannister, Marilena De Luca, Ivana Drobnjak, David E. Flitney, Rami K. Niazy, James Saunders, John
Vickers, Yongyue Zhang, Nicola De Stefano, J. Michael Brady, and Paul M. Matthews. Advances in functional and
structural MR image analysis and implementation as FSL. NeuroImage, 23:S208–S219, January 2004.
[47] Mark W. Woolrich, Saad Jbabdi, Brian Patenaude, Michael Chappell, Salima Makni, Timothy Behrens, Christian Beckmann,
Mark Jenkinson, and Stephen M. Smith. Bayesian analysis of neuroimaging data in FSL. NeuroImage, 45(1):S173–S186,
March 2009.
[48] Mark Jenkinson, Christian F. Beckmann, Timothy E.J. Behrens, Mark W. Woolrich, and Stephen M. Smith. FSL.
NeuroImage, 62(2):782–790, August 2012.
[49] Matthew F. Glasser, Timothy S. Coalson, Emma C. Robinson, Carl D. Hacker, John Harwell, Essa Yacoub, Kamil Ugurbil,
Jesper Andersson, Christian F. Beckmann, Mark Jenkinson, Stephen M. Smith, and David C. Van Essen. A multi-modal
parcellation of human cerebral cortex. Nature, 536(7615):171–178, July 2016.
[50] Emma C. Robinson, Saad Jbabdi, Matthew F. Glasser, Jesper Andersson, Gregory C. Burgess, Michael P. Harms, Stephen M.
Smith, David C. Van Essen, and Mark Jenkinson. MSM: A new flexible framework for multimodal surface matching.
NeuroImage, 100:414–426, October 2014.
[51] Jelle Veraart, Dmitry S. Novikov, Daan Christiaens, Benjamin Ades-aron, Jan Sijbers, and Els Fieremans. Denoising of
diffusion MRI using random matrix theory. NeuroImage, 142:394–406, November 2016.
[52] J.-Donald Tournier, Fernando Calamante, and Alan Connelly. Determination of the appropriate value and number of
gradient directions for high-angular-resolution diffusion-weighted imaging. NMR in Biomedicine, 26(12):1775–1786,
August 2013.
[53] J-Donald Tournier, Fernando Calamante, and Alan Connelly. Robust determination of the fibre orientation distribution in
diffusion MRI: Non-negativity constrained super-resolved spherical deconvolution. NeuroImage, 35(4):1459–1472, May
2007.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 15/20
[54] J Donald Tournier, Fernando Calamante, and Alan Connelly. Improved probabilistic streamlines tractography by 2nd
order integration over fibre orientation distributions. In Proceedings of the international society for magnetic resonance in
medicine, volume 18, page 1670, 2010.
[55] Robert E. Smith, Jacques-Donald Tournier, Fernando Calamante, and Alan Connelly. SIFT2: Enabling dense quantitative
assessment of brain white matter connectivity using streamlines tractography. NeuroImage, 119:338–351, October 2015.
[56] Robert E. Smith, Jacques-Donald Tournier, Fernando Calamante, and Alan Connelly. Anatomically-constrained tractogra-
phy: Improved diffusion MRI streamlines tractography through effective use of anatomical information. NeuroImage,
62(3):1924–1938, September 2012.
[57] Thomas M. J. Fruchterman and Edward M. Reingold. Graph drawing by force-directed placement. Software: Practice and
Experience, 21(11):1129–1164, November 1991.
[58] K.-F. Wong and X.-J. Wang. A recurrent network mechanism of time integration in perceptual decisions. Journal of
Neuroscience, 26(4):1314–1328, January 2006.
[59] Klaas Enno Stephan, Lars Kasper, Lee M. Harrison, Jean Daunizeau, Hanneke E.M. den Ouden, Michael Breakspear, and
Karl J. Friston. Nonlinear dynamic causal models for fMRI. NeuroImage, 42(2):649–662, August 2008.
[60] Hongjun Fu, Andrea Possenti, Rosie Freer, Yoshikazu Nakano, Nancy C. Hernandez Villegas, Maoping Tang, Paula V. M.
Cauhy, Benjamin A. Lassus, Shuo Chen, Stephanie L. Fowler, Helen Y. Figueroa, Edward D. Huey, Gail V. W. Johnson,
Michele Vendruscolo, and Karen E. Duff. A tau homeostasis signature is linked with the cellular and regional vulnerability
of excitatory neurons to tau pathology. Nature Neuroscience, 22(1):47–56, December 2018.
[61] Jorge J. Palop and Lennart Mucke. Network abnormalities and interneuron dysfunction in alzheimer disease. Nature
Reviews Neuroscience, 17(12):777–792, November 2016.
[62] Benedikt Zott, Manuel M. Simon, Wei Hong, Felix Unger, Hsing-Jung Chen-Engerer, Matthew P. Frosch, Bert Sakmann,
Dominic M. Walsh, and Arthur Konnerth. A vicious cycle of β amyloid–dependent neuronal hyperactivation. Science,
365(6453):559–565, August 2019.
[63] Joseph McInnes, Keimpe Wierda, An Snellinx, Laura Bounti, Yu-Chun Wang, Ilie-Cosmin Stancu, Nuno Ap´ostolo,
Kris Gevaert, Ilse Dewachter, Tara L. Spires-Jones, Bart De Strooper, Joris De Wit, Lujia Zhou, and Patrik Verstreken.
Synaptogyrin-3 mediates presynaptic dysfunction induced by tau. Neuron, 97(4):823–835.e8, February 2018.
[64] Bruno Vasconcelos, Ilie-Cosmin Stancu, Arjan Buist, Matthew Bird, Peng Wang, Alexandre Vanoosthuyse, Kristof Van
Kolen, An Verheyen, Pascal Kienlen-Campard, Jean-No¨el Octave, Peter Baatsen, Diederik Moechars, and Ilse Dewachter.
Heterotypic seeding of tau fibrillization by pre-aggregated abeta provides potent seeds for prion-like seeding and propagation
of tau-pathology in vivo. Acta Neuropathologica, 131(4):549–569, January 2016.
[65] Sarah L Griner, Paul Seidler, Jeannette Bowler, Kevin A Murray, Tianxiao Peter Yang, Shruti Sahay, Michael R Sawaya,
Duilio Cascio, Jose A Rodriguez, Stephan Philipp, Justyna Sosna, Charles G Glabe, Tamir Gonen, and David S Eisenberg.
Structure-based inhibitors of amyloid beta core suggest a common interface with tau. eLife, 8:e46924, oct 2019.
[66] ¨Unal Sako˘glu, Godfrey D. Pearlson, Kent A. Kiehl, Y. Michelle Wang, Andrew M. Michael, and Vince D. Calhoun.
A method for evaluating dynamic functional network connectivity and task-modulation: application to schizophrenia.
Magnetic Resonance Materials in Physics, Biology and Medicine, 23(5-6):351–366, February 2010.
[67] Daniel A. Handwerker, Vinai Roopchansingh, Javier Gonzalez-Castillo, and Peter A. Bandettini. Periodic changes in fMRI
connectivity. NeuroImage, 63(3):1712–1719, November 2012.
[68] David T. Jones, Prashanthi Vemuri, Matthew C. Murphy, Jeffrey L. Gunter, Matthew L. Senjem, Mary M. Machulda,
Scott A. Przybelski, Brian E. Gregg, Kejal Kantarci, David S. Knopman, Bradley F. Boeve, Ronald C. Petersen, and
Clifford R. Jack. Non-stationarity in the “resting brain’s” modular architecture. PLoS ONE, 7(6):e39731, June 2012.
[69] R. Matthew Hutchison, Thilo Womelsdorf, Joseph S. Gati, Stefan Everling, and Ravi S. Menon. Resting-state networks
show dynamic functional connectivity in awake humans and anesthetized macaques. Human Brain Mapping, 34(9):2154–
2177, March 2012.
[70] Enrique C.A. Hansen, Demian Battaglia, Andreas Spiegler, Gustavo Deco, and Viktor K. Jirsa. Functional connectivity
dynamics: Modeling the switching behavior of the resting state. NeuroImage, 105:525–535, January 2015.
[71] R. Matthew Hutchison, Thilo Womelsdorf, Elena A. Allen, Peter A. Bandettini, Vince D. Calhoun, Maurizio Corbetta,
Stefania Della Penna, Jeff H. Duyn, Gary H. Glover, Javier Gonzalez-Castillo, Daniel A. Handwerker, Shella Keilholz,
Vesa Kiviniemi, David A. Leopold, Francesco de Pasquale, Olaf Sporns, Martin Walter, and Catie Chang. Dynamic
functional connectivity: Promise, issues, and interpretations. NeuroImage, 80:360–378, October 2013.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 16/20
[72] Nora Leonardi and Dimitri Van De Ville. On spurious and real fluctuations of dynamic functional connectivity during rest.
NeuroImage, 104:430–436, January 2015.
[73] Andrew Zalesky and Michael Breakspear. Towards a statistical test for functional connectivity dynamics. NeuroImage,
114:466–470, July 2015.
[74] Timothy O. Laumann, Abraham Z. Snyder, Anish Mitra, Evan M. Gordon, Caterina Gratton, Babatunde Adeyemo,
Adrian W. Gilmore, Steven M. Nelson, Jeff J. Berg, Deanna J. Greene, John E. McCarthy, Enzo Tagliazucchi, Helmut
Laufs, Bradley L. Schlaggar, Nico U. F. Dosenbach, and Steven E. Petersen. On the stability of BOLD fMRI correlations.
Cerebral Cortex, 27(10):4719–4732, 09 2016.
[75] Maria Giulia Preti, Thomas AW Bolton, and Dimitri Van De Ville. The dynamic functional connectome: State-of-the-art
and perspectives. NeuroImage, 160:41–54, October 2017.
[76] Enrico Glerean, Juha Salmi, Juha M. Lahnakoski, Iiro P. J¨a¨askel¨ainen, and Mikko Sams. Functional magnetic resonance
imaging phase synchronization as a measure of dynamic functional connectivity. Brain Connectivity, 2(2):91–101, April
2012.
[77] Adri´an Ponce-Alvarez, Gustavo Deco, Patric Hagmann, Gian Luca Romani, Dante Mantini, and Maurizio Corbetta. Resting-
state temporal synchronization networks emerge from connectivity topology and heterogeneity. PLOS Computational
Biology, 11(2):e1004100, February 2015.
[78] Gustavo Deco and Morten Kringelbach. Metastability and coherence: Extending the communication through coherence
hypothesis using a whole-brain computational perspective. Trends in Neurosciences, 39(6):432, June 2016.
∗Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI)
database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation
of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI
investigators can be found at: https://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_
Acknowledgement_List.pdf
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 17/20
Tractography
Parcellation
Functional Brain
Dynamics (fMRI)
Aβ
Tau
Whole‐Brain
model
Inhibitory
Excitatory
Inhibitory
Excitatory
Burden modulation
(a) Integrating protein burden data
Inhibitory
Excitatory
Inhibitory
Excitatory
Burden modulation
Aβ
Tau
SC
Empirical phFCD
Minimize
DKS
Simulated phFCD
G ×
BOLD signal
(b) Fitting the phFCD in the whole-brain model
Figure 1. Illustrative overview of our processing pipeline. (a) Basic ingredients for the integration of protein burden data from
structural (dMRI, top left), functional (fMRI, top right), and burden (PET, right) using the same parcellation for each
neuroimaging modality (top, middle) for generating a whole-brain computational model (bottom left). Each node of the model
is using a realistic underlying biophysical neuronal model including AMPA, GABA, and NMDA synapses as well as
neurotransmitter gain modulation of these. (b) Fitting the measures in the whole-brain model: First, we simulate the BOLD
timeseries for each brain region in the parcellation, for each subject. These timeseries are defined by its inputs, namely a
previously found global coupling constant G, an individual Structural Connectivity (SC) matrix, and the corresponding
individual Aβ and tau burdens. Subsequently, we compute a time-versus-time matrix of phFCD. This is compared to a
reference empirical phFCD for that same subject using the Kolmogorov-Smirnov distance (KS), DKS, which is minimized to
find the generative parameters of the model. This process is repeated for the other two measures already mentioned, FC and
swFCD.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 18/20
Fitting comparison
G optimization
Aβ+tau
Aβ
tau
BEI
(HC)
Aβ+tau
Aβ
tau
BEI
(MCI)
Aβ+tau
Aβ
tau
BEI
(AD)
FC Dynamics (FCD)
A
B
C
D
E
F
Aβ
tau
G
H
KS phFCD
KS phFCD
KS phFCD
Functional fitting
Regions
Regions
Regions
Figure 2. Optimization and evaluation of the model: First, using only HC subjects, the global coupling parameter is found, and
then the model free parameters are adjusted to minimize the distance between the empirical and simulated fMRI data, taking
into account the regional burden distributions. (A) Minimization of the global coupling parameter G between 0 and 5.5, for
Functional connectivity (FC), sliding-window Functional connectivity Dynamics (swFCD) and phase FCD (phFCD). Given
their strong similarity in the results, phFCD was used for all subsequent computations. (B, C) Shows the normalized (in [0,1])
FCD distributions for the empirical data (top) and the simulated model (bottom). For an exemplary resulting timeseries, please
refer to the bottom-left part of Figure 1b. (D, E, F): Analysis of the impact (smaller values are better) of the different burdens
when optimized in isolation with respect to their impact in the phFCD (KS distance), and with respect to the homogeneous state
as a reference. As can be seen, the results for AD clearly show that tau alone accounts for the vast majority of the weight of the
impact on brain activity, while for MCI patients it is Aβ who dominates. The case for HC patients is not so clear, but we also
see a predominance of Aβ, although in a less conclusive manner. (G) Aβ and (H) Tau burdens of one subject (036 S 4430 in
ADNI’s database). Colors correspond to the normalized burden of each protein.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 19/20
Figure 3. Parameter values found after the optimization stage for HC, MCI and AD subjects. Observe that all b(E,I)
(Aβ,τ), the
excitatory/inhibitory Aβ and tau bias parameters, have negligible values, while the scaling parameters s(E,I)
(Aβ,τ) present non-null
values. Of note, the p-values between the different scaling parameters across the cohorts are different in a moderately
significant way (p < 0.03), remarkably between HC and AD, but usually not between MCI and AD. In these plots, boxes
extend from the lower to upper quartile values of the data, adding an orange line at the median. Also, whiskers are used to show
the range of the data, extending from the box.
Figure 4. Comparison between the homogeneous model, the result obtained and the same parameter values but with shuffled
burdens. As can be seen, the differences in fit statistics between models are significant. In particular, for the AD cohort, the
median phFCD correlation between model and data showed r < 0.1 for the heterogeneous model, and r ≈ 0.2 for the BEI
model. In all subject groups, the difference between these two models is clear, with p < 0.0005.
Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer‘s Disease — 20/20
Figure 5. Visualization of the SC graph, in matrix form (left) and as a graph showing the strongest 5% of connections. Node
positions are computed with Fruchterman and Reingold’s [57] algorithm, which assumes stronger forces between tightly
connected nodes. Besides the high degree of symmetry, we can observe the laterality is kept in the graph structure (also for
subcortical regions). Node size linearly represents the graph theoretical measure of structural degree for each node. As we can
see, the most important hubs are in the subcortical regions.
Figure 6. Excitatory and inhibitory mean firing rates as a function of the Aβ inhibitory scaling sI
Aβ, with all the other
parameters of the model at the (averaged) fitted optimum values. For the purpose of clarity, the horizontal axis for the scaling
has been taken as absolute values, to illustrate the behavior with increasing Aβ loads. The vertical axis shows the firing rates of
both excitatory and inhibitory populations. It can be clearly seen that the net effect of the burden is to increase the overall
region firing rate, measured at the excitatory population. For the sake of clarity, the inhibitory firing rate has been vertically
inverted (negated) to show their decreased effect on the excitatory population, thus confirming previous findings [21]. The
vertical discontinuous line shows the optimum found for sI
Aβ.
| 2022 | Whole-brain modeling of the differential influences of Amyloid-Beta and Tau in Alzheimer’s Disease | 10.1101/2022.10.30.514365 | null | creative-commons |
BRAFV600E Expression in Mouse Neuroglial Progenitors Increase Neuronal Excitability,
Cause Appearance of Balloon-like cells, Neuronal Mislocalization, and Inflammatory
Immune response.
Roman U. Goz1,2, Ari Silas1, Sara Buzel1, Joseph J. LoTurco1
1Departament of Physiology and Neurobiology, University of Connecticut, Storrs, CT
2Departament of Psychology University of Connecticut, Storrs, CT.
Corresponding author: joseph.loturco@uconn.edu
Abstract
BACKGROUND: Frequent de-novo somatic mutations in major components (PI3KCA, AKT3,
TSC1, TSC2, mTOR, BRAF) of molecular pathways crucial for cell differentiation, proliferation, growth
and migration (mTOR, MAPK) has been previously implicated in malformations of cortical development
(MCDs) and low-grade neuroepithelial tumors (LNETs) 1-7 . LNETs are the most frequent tumors found
in patients undergoing resective surgery for refractory epilepsy treatment. BRAFV600E is found in up to
70% of LNETs. Previous studies suggest a causal relationship between those de-novo somatic mutations
in mTOR, MAPK pathways and seizures occurrence, even without presence of malformation or a tumor 2,
3, 8-13. Recently Koh and colleagues 14 showed that BRAFV600E mutation may cause seizures through
activation of RE1-silecing transcription factor (REST). Additionally, they showed a significant
downregulation of synaptic transmission and plasticity pathways and decreased expression of multiple ion
channels subunits including HCN1, KCNQ3, SCN2A and SCN3B. The downregulation of those genes
including GABA receptors subunits and protein expression specific to interneurons subpopulations (SST,
VIP) suggests that a dysregulated inhibitory circuits are responsible for seizures in GGs. The
experimental manipulation - In-Utero electroporation of episomal activating Cre plasmids that they used
to test their hypothesis in mice however activated mutant BRAFV637 only in excitatory neurons. And the
downregulated genes in mice were confirmed by qRT-PCR in the whole tissue samples. The question of
how electrophysiological properties of the affected and surrounding neurons are changed were not
addressed. The changes in ion conductances and neuronal circuits responsible for seizures could be only
inferred from gene expression profiles. Purpose of the current work was to investigate how overactive
human BRAFV600E mutated protein incorporated into the mouse genome through piggyBase
transposition increase neuronal excitability in ex-vivo mouse cortical slices and whether it induces
histopathological features and gene expression profile alteration observed in low-grade neuroepithelial
tumors (LNETs).
METHODS: Using In-Utero Electroporation we have introduced human BRAFV600E protein
into radial glia progenitors in mouse embryonic cortex on the background of piggyBac transposon system
that allows incorporation of the DNA sequence of interest into the genome. Immunohistochemistry was
used for examination of known markers in LNETs. RNA sequencing on Illumina NextSeq 500 was used
to examine alterations in gene expression profiles. Whole-cell current- and voltage-clamp was used to
examine changes in electrophysiological properties. Unsupervised Hierarchical Clustering Analysis was
used to examine grouping of different conditions based on their gene expression profile and
electrophysiological properties. Video electrocorticographic recordings were used to test whether
BRAFV600E transgenic mice have spontaneous seizures.
RESULTS: Under GLAST driving promoter BRAFV600E induced astrogenesis, caused
morphological alterations in transgenic cells akin to balloon-like cells, and delayed neuronal migration.
Under NESTIN driver promoter BRAFV600E increased neurogenesis, induced balloon-like cells and
caused some cells to remain close to the lateral ventricle displaying large soma size compared to neurons
in the upper cortical layers. Some of the balloon-like cells were immunopositive for astroglial marker
glial fibrillary acidic protein (GFAP), and for both upper and lower cortical layers markers (Cux1 and
Ctip2). Gene ontology analysis for BRAFV600E gene expression profile showed that there is a tissue-
wide increased inflammatory immune response, complement pathway activation, microglia recruitment
and astrocytes activation, which supported increased immunoreactivity to microglial marker iba1, and to
GFAP respectively. In current clamp BRAFV600E neurons have increased excitability properties
including more depolarized resting membrane potential, increased input resistance, low capacitance, low
rheobase, low action potential (AP) voltage threshold, and increased AP firing frequency. Additionally,
BRAFV600E neurons have increased SAG and rebound excitation, indicative of increased
hyperpolarization activated depolarizing conductance (IH), which is confirmed in voltage-clamp. The
sustained potassium current sensitive to tetraethylammonium was decreased in BRAFV600E neurons.. In
4 out of 59 cells, we have also observed a post-action potential depolarizing waves, frequencies of which
increased in potassium current recording when Ca2+ was substituted to Co2+ in the extracellular solution
(5/24). We show that using 20 electrophysiological properties BRAFV600E neurons segregate separately
from other conditions. Comparison of electrophysiological properties of those neurons with neurons
bearing somatic mutations in mechanistic target of rapamycin (MTOR) pathway regulatory components,
overactivation of which is been shown in malformations of cortical development (MCDs), showed that
expression of PIK3CAE545K under GLAST+ promoter and TSC1 knockdown (KD) with CRISPR-Cas9
have different effects on neuronal excitability.
Keywords:
BRAF V600E
low-grade neuroepithelial tumor
ictogenesis
inflammation
neuroepithelial progenitors
malformation of cortical development
focal cortical dysplasia
hyperexcitability
potassium current
hyperpolarization activated depolarizing current
Acknowledgements
We want to thank Dr. Bo Reese from the Center for Genome Innovation, Institute for Systems Genomics,
University of Connecticut, Storrs, CT for helping with RNA-sequencing.
Current work was supported by NIH/NICHD grant #
And, NIH Akiko Nishiyama grant #S10OD016435 for acquisition of Leica SP8 microscope.
Author contribution
R.U.G. performed IUE, electrophysiological ex-vivo patch-clamp recordings, analysis, IHC, image
acquisition, and final figures. R.U.G., A.S., S.B. counted the cells with ImageJ-Fiji and analyzed it.
R.U.G. and J.J.L. performed RNA extraction, alignment and DE gene quantification and functional
enrichment analysis. R.U.G. and J.J.L. wrote the paper. J.J.L. wrote the grant, acquired funding and
provided the equipment and biomolecular research tools.
Competing Interests
The authors declare no competing interests.
1. Introduction
Low grade neuroepithelial tumors (LNETs) are the second most frequent structural pathology in
patients referred for resective surgery of intractable focal epilepsy and present in 25-80% of those cases
14-36. The major subtypes of LNETs - predominant in young patients ganglioglioma (GG) that represent
2- 5% of pediatric brain tumors 16, 33, and dysembryoplastic neuroepithelial tumors (DNETs), in about 20-
36% of cases are associated with focal cortical dysplasia (FCD) 19. FCD is considered a common cause of
drug refractory epilepsy and is found in 25-46% of cases in both children and adults 19-22, 24, 26, 27, 37-40. In
current International League Against Epilepsy FCD associated with LNETs is defined as FCD IIIb 41.
However, in some cases other types of FCD has been found in resected cortical tissue from adjacent to
LNETs areas 42, 43. FCD is a part of larger group of malformations of cortical development (MCDs),
which also include tuberous sclerosis complex disorder (TSC), different severity megalencephalies,
lisencephaly, microcephaly 33, 44. Inspite of abundant data on occurrence and structural pathological
components with shared morphometric features in LNETs and MCDs the mechanisms that cause
ictogenesis and subsequently development of epilepsy are not currently well understood. Contribution of
cortical structure disruption to seizures and epilepsy may depend on specific case, the size of the affected
cortical area, susceptibility of the affected cortical area to such disruptions, the population of progenitor
cells involved within the boundaries of affected developmental stages 39, 45-48, genetic etiology, epigenetic
modulation49, 50 and environmental effects. In some cases it may be malformation independent and
originate in adjacent to malformed cortex areas 2, 3, 8-13. Cortical structure disruption may include
dyslamination, presence of heterotopic neurons, dysmorphic cytomegalic neurons 51, 52, interneurons 53,
immature misoriented small neurons 54, and in sever FCDs cells without clear neuronal or glial
differentiation - balloon cells 41, giant cells in TSC 55, or atypical ganglion cells in LNETs 35. On the
molecular level, mechanisms that contribute to seizures development involve genetic alterations. Genetic
alterations in single components, as well as tissue-wide gene profile showed mutations in mechanistic
target of rapamycin pathway (MTOR) in MCDs 44, 56, 57 and in mitogen activated protein kinase pathway
(MAPK, also RAS-RAF-ERK) in LNETs 1, 7, 14, 58-60.
Recent development in DNA/RNA sequencing technologies simplified study of genetic alterations in
MCDs and LNETs on a tissue-wide scale. While some studies concentrated on comparison of tens to
hundreds of selected genes from microdessected heterotopic neurons, atypical ganglion cells and
astrocytes from FCD and GG 61, 62 showing that there was a differential expression in glutamate and
GABA receptors, and selected growth factors between the cells in tumor or malformation affected area
and cells from control tissues, latter studies used either microarrays or RNA sequencing to interrogate
global expression profiles and chromosomal reorganization in LNETs, low grade gliomas and TSC 14, 63-
71. Selective sequencing can still be applied to discover mutations in known malformations associated
genes 4. Those studies concentrated on discovery of additional somatic postzygotic mutations and
chromosome reorganization in LNETs and MCDs, while few of them has also reported on increased
inflammatory response and activation of complimentary cascade in GG 69 and in TSC 63, 68. Interestingly
Stone et al. 64 used RNA expression profile and DNA methylation profile in LNETs (GG, DNETs, and
with uncertain histologic type) to show that most of those segregate into two distinct groups, one group
with astrocytic differentiation and is driven by BRAFV600E mutation and the second group had
oligodendroglial differentiation and driven by FGFR1 mutation.
BRAF V600E mutation is found in up to 70% of LNETs 1, 7, 14, 58-60. Furthermore, recent study showed
presence of this mutation in few FCD associated with LNETs cases 72.
In the recently developed mouse models of MCD and GG genetic alterations in MTOR and MAPK
components in a small population of cortical cells was enough to disrupt cortical structure, cell
morphology and cause seizures. Moreover, administration of MTOR and MAPK specific components
inhibitors was enough to decrease seizures and prevent structural malformation. 2, 3, 9, 14, 73. However, the
intrinsic electrophysiological mechanisms that may lead to seizures at the single cell level in those studies
were not interrogated. This may be due to previous studies on FCD and TSC cases that showed no
significant increase in intrinsic excitability of malformed components, including cytomegalic neurons,
balloon cells and immature misoriented neurons 51, 53, 54, 74-77.
Here we hypothesized that expression of BRAFV600E mutation associated with LNETs alters gene
expression in the affected cortical tissue and increase intrinsic neuronal excitability in BRAFV600E
neurons, altering passive and active electrophysiological properties. To this end we used In-utero
electroporation that allows to introduce genetic manipulation into radial glia progenitor population
affecting a small percentage of cells (5-10%) 78, 79. This manipulation reflects the percentage of mutated
alleles found in MCDs 2, 4, 9, 73, 80, 81 and in GG 14. Gene expression was examined with RNA sequencing
and intrinsic neuronal properties were examined ex-vivo in cortical slices with whole-cell patch clamp.
Gene ontology analysis of the tissue-wide expression profiles showed that there was a significant increase
in immune response, as well as classic complement pathway activation in BRAFV600E cortical tissue.
The decreased biological protein pathways included potassium channels. BRAFV600E expressing
neurons had hyperexcitable intrinsic properties most prominent of each was increased action potential
firing and low current threshold required to fire action potential (rheobase). Other electrophysiological
properties that contribute to hyperexcitability of those neurons include more depolarized resting
membrane potential, increased input resistance, lower capacitance, more hyperpolarized action potential
voltage threshold. In current-clamp experiments significant SAG and rebound excitation in BRAFV600E
neurons were observed, a phenomenon associated with hyperpolarization activated depolarizing current
(IH). This was confirmed in voltage-clamp showing presence of hyperpolarization activated depolarizing
currents (IH) in BRAFV600E neurons only and not in control conditions. Consistent with that SAG and
rebound excitation were blocked by ZD7288. Also, in voltage-clamp experiments, we show that
BRAFV600E expressing neurons had smaller sustained potassium currents sensitive to
tetraethylammonium (TEA) compared to their untransfected neighbors. Finally, using unsupervised
hierarchical clustering analysis on electrophysiological properties we show that most BRAFV600E
neurons segregate closer together and other experimental conditions comprise the second major group.
When comparing those electrophysiological properties with somatic mutations that has been found in
FCD and TSC 3, 73 (expression of PIK3CA E545K, or CRISPR-Cas9 TSC1 KD) we show that those
mutations have different effect on neuronal electrophysiology.
2. Materials and Methods
2.1 Plasmids and sgRNA sequences
pGlast-PBase and pNestin-PBase were made as previously described 82. “PBase was inserted
downstream of the Nestin second-intron enhancer in the plasmid Nestin/hsp68-EGFP provided by Steven
Goldman 83. This 637-bp enhancer of the second intron of rat Nestin gene (GenBank: AF004334.1) was
located between bases 1162 and 1798 and is sufficient to control gene expression in the central nervous
system neuroepithelial progenitor cells 84. For pGLAST-PBase PBase was inserted downstream of the
GLAST promoter obtained from Dr. D.J. Volsky 85. This 1973-bp GLAST promoter was from human
excitatory amino acid transporter 1 (GenBank: AF448436.1). pPBCAG-monomeric red fluorescent
protein (mRFP), and pPBCAG-EGFP are constructed as previously described 86.” Human BRAFV600E -
pBABEbleo-Flag-BRAFV600E was donated by Dr. Christopher Counter and obtained from addgene
(Plasmid #53156) 87; and human PIK3CAE545K – pBabe-puro-HA-PIK3CAE545K was donated by Dr.
Jean Zhao 88 and also obtained from addgene (Plasmid #12525). The BRAFV600E and PIK3CAE545K
inserts were amplified with standard PCR and cloned into pPBCAG-EGFP replacing EGFP sequence
using EcoRI and NotI sites. Hemagglutinin (HA), a 27 nucleotides epitope tag (5’-
AGCGTAATCTGGAACATCGTATGGGTA-3’) was inserted into pPBase-BRAFV600E after
BRAFV600E sequence and before NotI site. pPBase-BRAFwt was generated with quick change II XL
single nucleotide site directed mutagenesis kit from Agilent according to the manufacturer protocol, to
change E, a glutamic amino acid back to V - valine at position 600 restoring the mutated sequence back to
its wild type. The sequence restoration to BRAFwt was confirmed with Sanger sequencing.
Channelrhodopsin plasmid (pcDNA3.1hChR2-EYFP) was a gift from K Diesseroth, Stanford University,
Stanford, CA, and was subcloned into the pCAG plasmid and used before 89. Guide RNA for TSC1 (T4 –
5’-CCATGCTGGATCCTCCACACTG-3’) and TSC2 (T7 – 5’-CCAAATCCCAGGTGTGCAGAAGG-
3’) were chosen based on 73. These sequences were cloned into pX330 vector (Addgene, plasmid #42230)
90 following normal cloning procedure.
2.2 Animals
Pregnant CD1 mice were obtained from Charles River Laboratories (Wilmington, MA, USA) and
maintained at the University of Connecticut vivarium on 12 h light cycle and fed ad libitum. Animal
gestational ages were determined via palpation prior to and confirmed during the surgery based on crown-
ramp length 91. Female and male mice were used for cortical transgene delivery with In-utero
electroporation. All procedures and experimental approaches were approved by the University of
Connecticut IACUC.
2.3 In utero electroporation
In-utero electroporation was performed as previously described 92. Briefly, mice were
anesthetized with a mixture of ketamine/xylazine (100/10 mg/kg i.p.). Metacam analgesic was
administered daily at dosage of 1 mg/kg s.c. for 2 days following surgery. To visualize the plasmid
during electroporation, plasmids were mixed with 2 mg/ml Fast Green (Millipore Sigma, F7252). In all
conditions, pPBCAG-EGFP, pPBCAG-mRFP, pPB-BRAFV600E, pPB-BRAFwt, pPB-PIK3CAE545K,
pGLAST-PBase, pNESTIN-PBase were used at the final concentration of 1.0 µg/µl. Electroporation was
performed at embryonic day 14 or 15. During surgery, the uterine horns were exposed and one lateral
ventricle of each embryo was pressure injected with 1-2 µl plasmid DNA. Injections were made through
the uterine wall and embryonic membranes by inserting pulled glass microelectrodes (Drummond
Scientific) into the lateral ventricle and injecting by pressure pulses delivered with Picospritzer II
(General Valve). Electroporation was accomplished with a BTX 8300 pulse generator (BTX Harvard
Apparatus) and BTX tweezertrodes. A voltage of 35-45 V was used for electroporation.
2.4 Image acquisition, cell counting and measurement
Images were acquired on inverted Leica TSC SP8 confocal microscope with four PMT detectors
and one HyD detector equipped with 405 nm diode laser, argon (458/488/514 nm) laser, 561 nm DPSS
laser and 633 nm HeNe laser. Sets of images for all the experimental and control conditions in each group
(GLAST+, NESTIN+) were acquired on the same day with the same excitation power and gain settings.
Some of the images were acquired with Zeiss Axiozoom.V16 with 405/488/568/647 filters and
Lumencor’s SOLA SE 365 light engine with ~3.5W white light output through 3 mm dia liquid light
guide (LLG) with PlanNeoFluar Z 2.3x with 0.57 n.a. lens. Axiozoom was equipped with sCMOS
pco.edge 4.2 camera with CIS2020A sensor. All the images were further processed in ImageJ-Fiji
package (version 1.51w, NIH, RRID: SRC_003070) 93. For manual cell counting and distance to pia
measurement images were converted to black for EGFP and white background and pia was oriented as a
horizontal plane and cells were counted with cell count plugin in Fiji by A. S., S.B. and R.G. Soma size
was measured with a freehand selection tool and measure under the same brightness/contrast and color
balance settings in all conditions. Balloon-like cells and aggregates were scanned and counted by S.B. and
R.G. with Axiozoom Zeiss .V16. Image processing for publication was done in Fiji and Corel Draw
Graphics Suite X8 (Corel, Ottawa, Canada; RRID: SCR_002865).
2.5 Immunohistochemistry
Animals were deeply anesthetized with isoflurane and perfused transcardially with 4%
paraformaldehyde/PBS (4% PFA). Samples were post fixed overnight in 4% PFA. For
immunofluorescence, brains were sectioned at 50-µ thickness on a vibratome (Leica VT 1000S). Sections
were processed as free-floating and stained with rabbit polyclonal anti-GFAP (1:2000 dilution, DAKO
Z0334, GenBank L19867, RRID:AB_10013482), mouse monoclonal anti-Aldehyde Dehydrogenase 1
family 1, member L1 (ALDH1L1, 1:50 dilution, NeuroMab, cat. #75-140, RRID:AB_10673448, clone
N103/39, accession number P28037), goat polyclonal anti-Iba1 (1:200 dilution, Invitrogen, cat. #
PIPA518039, accession number P55008), nuclear staining with Hoechst 33342, trihydrochloride,
trihydrate – 10 mg/ml in water (1:3000 dilution, Molecular probes by life technologies, cat. # H3570).
After blocking in PBS containing 5% of normal goat serum (Millipore Sigma, NS02L) and 0.5% Triton
X-100 (Millipore Sigma, X100) for 2 h at room temperature, tissue was washed three times in PBS with
2.5% normal goat serum and 0.2% Triton X-100 (washing solution), followed by incubation with primary
antibodies overnight at 40C in the washing solution. On the following day tissue was washed again in
washing solution and incubated with the appropriate secondary antibodies in washing solution (all Alexa
Fluor in 1:1000, Invitrogen) for 2 h at room temperature (Alexa Fluor 568 anti-mouse IgG, Alexa Fluor
647 anti-rabbit IgG, Alexa Fluor 568 anti-goat IgG). After 2 h the tissue was washed again with washing
solution once, stained with Hoechst 33342 and washed again three times. Tissue was mounted on
Fisherbrand Colorfrost Plus Microscope slides (Cat #12-550-19) submerged in ProLong gold antifade
(Life technologies, cat. #36930) and coverslipped with Fisherfinest premium cover glass (cat. #12-548-
5P,5J,B, sizes 24X60-1, 24X40-1,22X22-1 respectively). When prolong gold antifade has cured the
coverslips edges were covered with transparent nail polish.
2.6 RNAseq, gene ontology analysis and Gene Analytics
Mouse brains were extracted at P65 from 4 GLAST+ BRAFV600E animals, 4 GLAST+ BRAFwt
animals, 4 GLAST+ control-FP animals (2 from GLAST+ BRAFV600E litter and 2 from GLAST+
BRAFwt litter) that were electroporated at E14 after deep anesthetization with isoflurane. The fluorescent
EGFP area of somatosensory cortex was dissected and the white matter remains were cut out from those
tissue chunks. The remaining cortical tissue chunks were further dissociated and processed with Ambien
RNA extraction kit according to manufacturer protocol. The range of RNA amount for the samples was
from 300-900 ng per sample measured with nanodrop-1000 spectrophotometer. Quality was assessed by
RNA Integrity Numbers and values ranged from 6.6 to 8.5 for first stranded cDNA library preparation
and analysis with Illumina NextSeq 500 – mid output v2 (150 cycles). Libraries were sequenced at a
depth of 9.6 to 18 million reads per sample. Quality control, library preparation and sequencing were
done at UCONN Center for Genome Innovation, Institute for System Genomics. For the further
processing and analysis of sequencing results new tuxedo protocol was used 94 on UCONN High
Performance Computing cluster. Fasq files containing sequencing fragments were aligned with HISAT2
(RRID:SCR_015530) 95 using index build for mouse genome fasta file downloaded from Ensemble data
base (ftp://ftp.ensembl.org/pub/release-
92/fasta/mus_musculus/dna/Mus_musculus.GRCm38.dna.primary_assembly.fa.gz), produced sequence
alignment maps (sam) files were sorted with samtools (RRID:SCR_002105) outputting binary alignment
maps (bam) files, that were assembled, merged with the mouse reference genome from Ensemble data
base (ftp://ftp.ensembl.org/pub/release-92/gtf/mus_musculus/Mus_musculus.GRCm38.92.gtf.gz) and
quantified with Stringtie 96. Stringtie raw counts for further differential expression analysis were extracted
with Python script (http://ccb.jhu.edu/software/stringtie/dl/prepDE.py) using Python 2.7
(RRID:SCR_002918). Differential gene expression was estimated with DESeq2 1.18.1
(RRID:SCR_015687) 97 in R 3.4.4 (RRID:SCR_001905) 98 using R Studio GUI 99 and FDR corrected p-
values (q values) of 0.05 were considered significant 100. The results of differentially expressed genes
were further analyzed for functional enrichment with DAVID 6.8 101, 102
(https://david.ncifcrf.gov/home.jsp)
Gene Analytics web-service (geneanalytics.genecards.org) 103 was used to analyze differentially
expressed genes. Pseudorandom mouse gene list was generated in molecular biology online apps web tool
(http://molbiotools.com/randomgenesetgenerator.html) using Mersenne Twister 104 pseudorandom
number generator algorithm (from personal communication with Vladimír Čermák, the site developer).
2.7 Slice preparation
The P15-P70 (average P36.05, mode=36 median=35, stdev=9.86) CD1 were deeply anesthetized
with isoflurane and then decapitated. Brains were rapidly removed and immersed in ice-cold oxygenated
(95% oxygen and 5% carbon dioxide) dissection buffer containing (in mM/L): 83 NaCl, 2.5 KCl, 1
NaH2PO4, 26.2 NaHCO3, 22 dextrose, 72 sucrose 0.5 CaCl2, and 3.3 MgCl2, Coronal slices (400 µm)
were cut with a vibratome (VT1200S; Leica, Nussloch, Germany), incubated in dissection buffer for 40
min at 340C, and then stored at room temperature for the reminder of the recording day. Most of the slice
recordings were performed at 340C, besides voltage-clamp recordings of calcium and potassium currents.
Slices were visualized with inversion recovery differential interference microscopy (E600FN; Nikon,
Tokyo, Japan) and a CCD camera (QICAM; QImaging, Surrey, British Columbia, Canada). Individual
neurons were visualized with a 40X Nikon Fluor water immersion (0.8 n.a.) objective.
2.8 Electrophysiology
For all experiments except potassium and calcium currents recordings, extracellular recording
buffer was oxygenated (95% oxygen and 5% carbon dioxide) and contained (in mM/L): 125 NaCl, 25
NaHCO3, 1.25 NaH2PO4, 3 KCl, 25 dextrose, 1 MgCl2, and 1.3 CaCl2, 295-305 mOsm. For potassium
currents recordings the extracellular calcium was substituted with 1.3 CoCl2 · 6H2O (Millipore Sigma,
CAS number 7791-13-1). For calcium currents recordings extracellular recording buffer was oxygenated
(95% oxygen and 5% carbon dioxide) and contained (in mM/L): 130 NaCl, 25 NaHCO3, 1.25 HaH2PO4,
3 KCL, 25 dextrose, 1 CaCl2, 1.3 MgCl2, 40 TEA, 0.001 TTX, 0.01 gabazine, 0.05 D-APV, 0.01 NBQX,
3 4AP, 0.05 ZD7288. Patch pipettes were fabricated from borosilicate glass (N51A; King Precision
Glass, Claremont, California) to a resistance of 2-7 MΩ. The resultant errors were minimized with bridge
balance and capacitance compensation. For current-clamp experiments, voltage-clamp recording of
hyperpolarization activated currents and potassium currents pipettes were filled with an internal solution
containing (in mM/L): 125 K-gluconate, 10 HEPES, 4 Mg2-ATP, 3 Na-GTP, 0.1 EGTA, 10 Na-
phosphocreatine, 0.05% biocytin, adjusted to pH 7.3 with potassium hydroxide and to 275-285 mOsm
with double-distilled water. For voltage-clamp recordings of calcium currents pipettes were filled with an
internal solution containing (in mM/L): 110 CsMeSO4, 10 CsCl, 5 CaCl2, 10 EGTA, 10 HEPES, 4 Mg2-
ATP, 0.3 Na-GTP, 10 Na-phosphocreatine, 0.05% biocytin, 25 TEA. Signals were amplified with
Multiclamp 700B amplifier (Molecular Devices, Sunnyvale, CA), sampled at 20 kHz, digitized (ITC-18;
HEKA instruments, Bellmore, NY) and filtered at 2 kHz with an 8-pole low-pass Bessel filter. Data were
monitored, acquired, and in some cases analyzed with Axograph X software (Berkley, CA). Series
resistance was monitored throughout the experiments by applying a small test voltage step and measuring
the capacitive currents. Series resistance was 5 to approximately 25 MΩ, and only cells with <20%
change in series resistance and holding current were included in the analysis. Reported membrane
potentials and holding potentials were not corrected for liquid junction potential ~ 10 mV unless
otherwise specified.
For estimation of the effect of early opening of potassium channels on AP firing frequency
retigabine was used (10 µM/L; Alomone labs, Jerusalem, Israel, Cat # R-100), the 100 mM stock was
prepared in DMSO and diluted into extracellular recording buffer. In specified experiments D-APV
dissolved in double deionized water (ddw) to 50-100 mM stock (50 µM/L; Abcam, Cambridge, MA, cat.
# ab120003, lot #GR205917) was used to block specifically NMDAR, and NBQX dissolved in DMSO to
100 mM stock (10 µM/L; Abcam, Cambridge, MA, cat #ab120045, lot #GR133243) was used to block
AMPARs; SR95531 dissolved in ddw to 25 mM stock (gabazine, 10 µM/L; Abcam, Cambridge, MA, cat.
#ab120042, lot #GR69200) was used to block GABAAR; TTX-citrate dissolved in ddw to 10 mM (1
µM/L; Abcam, Cambridge, MA, cat. #ab120055, lot #GR246757) was used to block sodium currents;
4AP dissolved in ddw to 200 mM (30 µM/L – 3 mM/L; Millipore Sigma, Burlington, MA, cat. #275875)
was used to block fast activating fast inactivating potassium currents; TEA (25-40 mM/L; Abcam,
Cambridge, MA, cat. #ab120275, lot #GR69136) was used to block sustained potassium currents;
ZD7288 dissolved in ddw (50 µM/L; Cayman Chemicals, Ann Arbor, MI, cat. # 1522820000040, batch
number 0476856-6) was used to block hyperpolarization activated depolarizing currents (IH).
For action potential (AP) firing frequency, input resistance measurement, 1 second current steps
were applied at 10 pA increment from -40 pA to 300 pA. For TSC1/2 KD neurons those steps increased
to more than 300 pA. Input resistance was measured from last 100 ms of 1 s hyperpolarizing and
depolarizing subthreshold current steps and current-voltage relationship was fit with linear regression to
estimate input resistance from both depolarizing and hyperpolarizing current steps. Resting membrane
potential (RMP) was measured in the beginning of current-clamp protocols before application of current
step pulses. SAG ratio was defined as 𝑆𝐴𝐺 = (1 −
𝑉𝑅𝑀𝑃− 𝑉𝑠𝑠
𝑉𝑅𝑀𝑃− 𝑉𝑚𝑖𝑛) ∗ 100%, VRMP – Resting Membrane
Potential, Vss – stable-state voltage in the last 100 ms of 1 second -40 pA pulse, Vmin – minimal initial
voltage deflection in response to 1 second -40 pA pulse. Rebound excitation was measured as an
overshoot above RMP at the end of -40 pA 1 second current step, in some cells resulting in AP firing. AP
voltage threshold was defined as the point at which the first derivative of voltage to time (dV/dt) crossed
50 V/s. Rheobase is the minimal current step required to elicit first AP firing. AP peak was measured
from RMP.
To record potassium currents the neurons were held at -90 mV and the 500 ms voltage steps
proceeded with 10 mV increments from -100 to +20 mV. For sustained potassium currents the amplitude
was measured at the last 100 ms of 500 ms voltage steps. The capacitive currents were canceled with
internal Multiclamp 700B compensation circuit. Cell capacitance and input resistance in those
experiments was monitored and measured before compensation from +5 mV 150 ms voltage steps with
Axograph X built in procedure designed to measure series resistance, membrane capacitance and
membrane resistance. The measurement was done offline after offline leak current subtraction with scaled
voltage steps of opposite polarity to steps that elicit potassium currents.
To record hyperpolarization activated depolarizing currents (IH) two protocols were used. In the
main protocol, used for the analysis the neurons were held at -50 mV and 1.5 s voltage steps proceeded
with 5 mV increment from -120 mV to -35 mV without capacitance and series resistance compensation.
The peak measurement of IH was done from the point at the beginning of the observe current to the stable
state at the last 100 ms of 1.5 s voltage steps. The tail currents were measured after the voltage steps
ceased. All the measurements were done offline after offline leak current subtraction with scaled voltage
steps of opposite polarity to steps that elicit IH. The second protocol was used to increase stability of the
recorded cells. In this protocol neurons were held at -70 mV and 1 s voltage steps proceeded with 5 mV
increment from -100 mV to -45 mV.
To record calcium currents neurons were held at -80 mV and 200 ms voltage steps proceeded
with 5 mV increment from -90 mV to 0 mV. Maximal negative deflection was used for peak current
estimation at each voltage step and to construct activation curve.
Spontaneous Post-Synaptic currents (sPSCs) recording of 1-5 min was done with the chart
procedure in Axograph X and was sampled at 5 kHz. The sPSCs were detected with semiautomatic
sliding template method as previously described 105 and were visually confirmed. The parameters of the
template, including amplitude, 10-90% rise time, and decay time were determined on the bases of an
average of real events as well as previously reported values. The detection threshold is 2.5 times of the
noise SD. The sliding template length was chosen to be 10 ms for all neurons.
2.9 Headmount installation and video ECoG recordings
EEG system, including headmounts, preamplifiers, amplifiers and video-ECoG recording system
was purchased from pinnacle technology Inc. (Lawrence, KS). BRAFV600E and control-FP
electroporated mice of at least 6 months of age were used for those experiments. Surgery was performed
under general, continuous isoflurane/O2 anesthetic inhalation system at 1-1.5 litters/minute, with
intraperitoneal injection of metacam analgesic at 5mg/kg before beginning of the procedure. The mice
were stabilized in a mouse stereotaxic apparatus (Stoelting, Wood Dale, IL). The fur was shaved off from
the mouse head with small trimmers and disinfected with chlorhexidine, 2% (Henry Schein Animal
Health, Dublin, OH). The rostral-caudal incision in the skin was made with 25 mm cutting edge surgical
scissors to allow sufficient space on the mouse skull for the headmount (about 1.5 cm). Before mounting
hydrogen peroxide was applied to the surface and surgical cotton-tipped sterile q-tips were used to
remove periosteum. Four pilot holes were made in the skull through the openings in the headmount with
25-gauge BD needle at the following approximate coordinates relative to bregma: -2 mm; ML: ± 1.5 mm,
lambda: -2 mm. Two small pockets were made in the nuchal muscle for EMG electrodes insertion. After
insertion of EMG electrodes, the headmount (#8201) with platinum and iridium leads was placed on the
surface of the skull covered with cyanoacrylate glue. Four stainless steel ECoG screws, 2-0.10” in front
and 2-0.12” screws in back (#8209 and #8212 respectively) were inserted through the headmount
openings and manually rotated into the pilot holes. Before the screws were fully locked in place, two-part
silver epoxy was applied between the screw heads and the headmount to ensure electrical conductivity.
After securing the headmount with screws, dental acrylic cement was applied with a small brush dipped
in acetone to the area surrounding the headmount and the base of the headmount. The dental cement cured
within 2-5 minutes. One to five skin sutures were applied to close the skin incision. After the surgery
mice were places in a clean cage on a warm hitting pad to recover. ECoG data acquisition started 5-7 days
after the headmount surgery.
Simultaneous three mice (up to four) video ECoG recordings were performed using pinnacle
#8206 data conditioning and acquisition system (DCAS) with 2 electrocorticographic (ECoG) and 1
electromyographic (EMG) channels and dome cameras with infrared light source for night time recording
(#9022). The recordings continued for at least five days. Mice were housed in the circular acrylic cage on
10” x 10” base with 10” diameter/8” height. Water and food was provided ad libitum. In the beginning of
the video EEG data acquisition the headmount was connected to the 3-channel mouse preamplifier
(#8202-SL, the sleep configuration). The preamplifier had AGND ground connection for the animal to
avoid input amplifier overcharge. This X100 preamplifier was connected to the secondary amplifier,
AD/DA and filtering system – DCAS #8206 (X50.78), that was mounted on a swivel plate to allow mice
to move freely. DCAS #8206 together with dome cameras was connected to the desktop PC. Sirenia
acquisition software version 1.7.10 was used for simultaneous ECoG/Video recording. The ECoG data
was sampled at 600 Hz and low-pass filtered with 8th order progressive elliptic analogue hardware
implemented filter at 25 Hz (6 dB/octave). The EMG data was sampled at 600 Hz and low-pass filtered at
100 Hz. Video recording was acquired at 20 f/s in grayscale with 60% image quality to avoid filling up
hard drive capacity too fast in X1 MJPG compression at 640X480 pixels resolution. The recorded data
was analyzed in Sirenia Seizure Pro software version 1.7.10.
2.10 Unsupervised hierarchical clustering analysis
The hierarchical clustering was done with Gene Cluster 3.0
(http://bonsai.hgc.jp/~mdehoon/software/cluster/software.htm), a freeware developed in Michael Eisen
lab in Berkeley, using pairwise average linkage with Euclidean distance calculated to determine the
difference between clusters of neurons by the length of the branch 106. We have used 20 recorded
electrophysiological parameters, for which each observation was standardized by centering to the mean
and dividing by standard deviation. The missing values were imputed by Non-Linear Iterative Partial
Least Squares (NIPALS) algorithm in XLSTAT (https://www.xlstat.com/en/) Microsoft Excel (Microsoft,
Redmond, Washington) addon 107, 108. The heatmap visualization for RNAseq data was done with
ClustVis 109 (https://biit.cs.ut.ee/clustvis/). For electrophysiological data it was done in Genepattern 110
(https://genepattern.broadinstitute.org/, RRID:SCR_003201) with HierarchicalClusteringViewer module
v.11.3.
2.11 Statistical analysis
All data measurements were kept in Excel (Microsoft, Redmond, Washington) and in Origin
(OriginLab, Northampton, MA; RRID: SCR_002815). All the electrophysiological data was analyzed in
SPSS v.24 (RRID:SCR_002865) 111, for large samples one-way analysis of variance (ANOVA) with
Tukey posthoc correction was used, when the samples had non-homogenous variance (significant Levene
test for equality of variance) Welch test with Games-Howell posthoc correction was used. For small
samples from different observations independent samples two-tailed student t-test was used and
depending on Levene test significance the t statistics for equal or unequal variance was reported. For
measurement coming from the same neurons before and after treatment paired samples two-tailed student
t-test was used. Graphical visualization of data was prepared in Origin and exported to Corel Draw
Graphic Suite X8 for further processing. Arithmetical averages and SEMs were reported for all results
unless otherwise specified.
3. Results
3.1 BRAFV600E alters neuronal migration and morphology
To test whether BRAFV600E is sufficient to cause developmental disruptions in cortex we
introduced BRAFV600E transgenes or control-FP transgenes (mRFP) into populations of neocortical
progenitors using the binary piggyBac transposon system 82, 112-114. We directed transgenes into either a
population of GLAST+ neural progenitors that generates both pyramidal neurons and astrocytes 82, 115, or a
NESTIN+ population that generates primarily pyramidal neurons (Figure 1A) 116-118. We compared the
numbers, positions and morphologies of cells generated from these progenitors in three different transgene
conditions (BRAFV600E, BRAFwt, and mRFP). Consistent with the upper layer laminar fates of neurons
with birth dates at E14, when the transgenes were introduced, the neurons generated in all conditions
expressed the upper layer pyramidal neuron marker CUX1, but not the lower layer pyramidal neuron marker
CTIP2 (Figure 1C). Although positive for upper layer markers, there was a significant increase in the
number of neurons in the BRAFV600E transgene conditions that were in deeper layers relative compared
to control conditions (Figure 1C lower panel, Figure 2C). Interestingly, the pattern of altered neuronal
positioning was different for BRAFV600E introduced into GLAST+ progenitors relative to NESTIN+
progenitors, with a subpopulation of neurons generated from NESTIN+ progenitors displaced even deeper
into cortex, into the subventricular white matter. The dyslamination observed in our experiments concur
the results obtained with episomal Cre expression in BRAFV637E transgenic mice 14. However,
quantification of neuronal soma sizes indicated no significant difference in the sizes of neurons generated
from GLAST+ progenitors and control conditions, whereas neurons from NESTIN+ progenitors that were
also displaced into white matter had significantly larger somas with ganglion cell like morphologies (Figure
2A,B). Those results provide wider insight into the neuroglial progenitors affected population compared to
Koh et al. work 14. Together these results indicate that BRAFV600E disrupts the normal laminar positioning
attained during migration, and the different progenitor populations (NESTIN+ and GLAST+) respond
differently to overactive mutant BRAF. To separate the effect of ectopic expression of human BRAF protein
from the effect of BRAFV600E mutation, which has been shown in COS7 cell cultures to have 500-fold
higher basal kinase activity compared to BRAFwt 119, we examined neurons position and morphology in
cortical slices with introduced wild type human BRAF transgene. There was a smaller but significant effect
compared to control-FP transgenic brains on neuronal migration. Additionally, the increase in the number
of neurons in BRAFwt found in lower cortical layers was lower compared to BRAFV600E transgenic
neurons (Figure 1E).
3.2 BRAFV600E causes development of "Balloon-like" cells
In addition to the delayed neuronal migration observed in BRAFV600E transgenic mouse brains
we tested whether there was an effect on cellular morphology. Balloon cells are a distinctive cell type
characteristic of FCDIIb 41, 120, 121 and have also been described in FCDs in the vicinity of LNETs positive
for BRAFV600E 43, 72. Additionally, in LNETs atypical cytomegalic ganglion cells has been shown to have
like balloon cells morphology and protein expression 61. In cortical tubers resected from TSC patients the
giant cells also show similar morphology 122-124. These unusual cells have distinctive morphologies and
label positive for a mixture of molecular markers for neurons, glia, and neural progenitors. While not in
every brain in our data set, we frequently found "balloon-like" cells in BRAFV600E transgene conditions,
and these cells were positive for cux1, ctip2, GFAP, and but were negative for NeuN. There was no
observed immunoreactivity to Vimentin or Nestin in balloon-like cells (data not shown). Additionally, those
cells were negative for caspase-3 (data not shown), an apoptotic marker 125 and did not display DNA
fragmentation and cellular membrane blebbing. However, we cannot exclude possibility that some of the
balloon-like cells may undergo apoptosis or pyroptosis 126. In brains analyzed prior to P30 we found isolated
balloon-like cells; however, by P30 aggregates of balloon-like cells were apparent (Figure 1B,1D). The
observations indicate that BRAFV600 is sufficient to drive the development of balloon-like cells, that these
cells can be generated in the lineages of either NESTIN+ or GLAST+ progenitors and appear in increasing
numbers in the juvenile period, after P30. Those cells were not described in Koh et al.14
3.3 BRAFV600E causes increased astrocytogenesis and glial activation
Overactive signaling through the RAS-RAF-MEK-ERK pathway is known to increase astrocyte
differentiation and proliferation 127-130. Consistent with this we found that BRAFV600E transgenesis in
GLAST+ progenitors resulted in a significant increase in the number of astrocytes relative to neurons. In
contrast, BRAFV600E transgenes introduced into NESTIN+ progenitors did not result in an appreciable
number of astrocytes (Figure 3C). LNETs with BRAFV600E mutation has been shown to display high
immunoreactivity to Glial Fibrillary Acidic Protein (GFAP) 35. Additionally, Koh et al. 14 showed increased
immunoreactivity of glial lineage in GG patients and their mouse model. Consistent with this we found a
significant increase in number of intensely positive GFAP cells in the regions of cortex containing
BRAFV600E expressing neurons compared to control-FP and BRAFwt. We found that transgenesis of
either NESTIN+ progenitors or GLAST+ progenitors with BRAFV600E resulted in comparable increases
in the intensity of GFAP staining. Taken together this would suggest that BRAFV600E somatic mutations
in proneuronal progenitors and in neurons is sufficient to result in elevated GFAP expression and potentially
astrocyte activation (Figure 3A). GFAP immunoreactive cells were also positive for astrocytes marker
Aldehyde Dehydrogenase 1 family member L1 (ALDH1L1) 131 (Figure 3B).
To determine whether the elevated intensity of GFAP staining we observed was due to reactive
gliosis and potential inflammatory responses in the regions of mutation bearing cells, we performed an
RNA-seq experiment to compare the gene expression profiles of patches of cortex containing BRAFV600E,
BRAFwt, or mRFP transgenes. We estimate that approximately 5-10% of cells are transfected cells bearing
transgenes in a cortex 78, 79, and so the majority of any change in transcript is likely driven by changes in
gene expression profiles in untransfected reacting cells. Using an unsupervised hierarchical clustering
analysis of all genes in 12 samples, 4 in each transgene condition, we found that BRAFV600E, control-FP,
and BRAFwt conditions clustered separately, except for one BRAFV600E sample which clustered with
BRAFwt conditions (Figure 4A). Differential expression and gene ontology analysis indicated a significant
increase in the expression of genes in the inflammatory immune response pathway (H2-Aa, CD74, H2-
Ab1, CD48, CD109, Cxcl16, Ccr1), and classic complement pathway components (C3, Serpinf1, C4b,
C1s1, C1ra, Serpina3i, Serpina3b, Serpina3n) (Figure 4B,C,D,E,G; Table S1). The Iba1, a microglia marker
was also increased in GLAST+ BRAF V600E condition compared to control-FP, it was also increased
compared to BRAFwt at p=0.011 level. A marker of microglia activation HLA-DR(CD74) was
significantly increased in GLAST+ BRAFV600E compared to BRAFwt and to control-FP (Figure 4C,D).
Similarly, markers of astrocyte activation, GFAP and Vimentin, were also significantly upregulated in the
RNAseq profiles of the four BRAFV600E samples relative to the other transgene conditions. Overall, the
pattern of gene expression changes in cortical tissue containing a subpopulation of cells expressing
BRAFV600E is consistent with these cells causing a glial activation and neuroinflammatory response
(Figure 4G). It is also consistent with previous studies showing increased inflammatory immune response
and complement pathway activation in ganglioglioma, and, in tissue resected from epilepsy patients with
tuberous sclerosis 14, 68, 69. Interestingly, the decreased expression of potassium channels (Figure 4G) is
consistent with previous study by Aronica et al. 69 and Koh et al. 14. Additional ontology analysis with Gene
Analytics web tool showed significant enrichment of genes associated with Tuberous Sclerosis (Figure 4H).
3.4 BRAFV600E increases excitability of pyramidal neurons
To test the hypothesis that neurons with BRAFV600E mutation have increased excitability we
performed whole-cell patch clamp recording from pyramidal neurons in upper layers 2/3. We recorded from
neurons in all three transgene conditions and in both transfected and in neighboring neurons not positive
for fluorescent markers of transgenesis. In current-clamp recordings we found that BRAFV600E neurons
displayed significantly higher action potential (AP) firing frequencies to 1 second depolarizing current
pulses (Figure 5A upper panel, 5B, p<0.001 for 20-300 pA current steps). This significantly increased firing
rate was true for neurons from both the NESTIN+ and GLAST+ progenitor populations. Neither BRAFwt
nor neighboring untransfected neurons in BRAFV600E conditions showed elevated firing frequencies
above fluorescent protein transfected controls (control-FP).
To compare the effect of developmentally induced chronic overactivation of RAF-RAS-ERK
pathway on neuronal electrophysiology to the effect of chronic overactivation of mTOR pathway we’ve
performed whole-cell patch clamp experiments in cortical slices transgenic for GLAST+ PIK3CA E545K
3 and CRISPR-Cas9 induced mutation in TSC1 gene 73. Phosphatidylinositol-4,5-Bisphosphate 3-Kinase
Catalytic Subunit Alpha and TSC1 are key regulatory upstream components of mTOR. Substitution of
glutamic amino acid to lysine (E545K) in PIK3CA is a “hot spot” mutation and is found in FCD and cause
constitutive activation. Disruption of TSC1/TSC2 complex that inhibits activation of mTOR is associated
with Tuberous Sclerosis and was found in FCD too. GLAST+ progenitors were transfected with PIK3CA
E545K on PiggyBac transposon background using IUE at E14-E15 in the same way as BRAFV600E, and
for CRISPR-Cas9 TSC1 guide-RNA we used T4 from Lim et al. 2017. The AP firing frequency, AP ISF,
rheobase, RMP and Rin was closer to control conditions in previous experiments (Figure 6A upper panel,
6B, C, D, F, G). However, AP voltage threshold was not different from GLAST+ BRAFV600E neurons.
This suggest that alterations in neuronal electrophysiological properties affected differently by pathological
mutations in mTOR pathway key protein components. Given those findings we decided not to pursue
further inquiry into PIK3CA E545K and CRISPR-Cas9 TSC1 KD conditions.
Instantaneous AP frequency (ISF) measured at +300 pA 1 second current step was significantly
higher in BRAFV600E neurons (Figure 5C). In addition, in 4 out of 59 GLAST+ BRAFV600E neurons
and in 1 out of 9 NESTIN+ BRAFV600E neurons we observed an unusual bursting pattern and post-action
potential depolarization waves that were not observed in any of the non-BRAFV600E conditions (Figure
5G left panel). The number of neurons with those events was increased when recording potassium currents
with Co2+ (1 mM) substituting Ca2+ (1 mM) in aCSF solution (5 out of 24). In addition to the AP firing at a
significantly higher frequency BRAFV600E neurons also had lower rheobase (n=49), minimal depolarizing
current step required to elicit first AP, and a lower voltage threshold to fire action potentials (Figure 7B, C,
E). Passive membrane properties were also significantly different in BRAFV600E neurons. The resting
membrane potential (RMP) was more depolarized in BRAFV600E neurons (n=61, -64.91 ± 0.76 mV)
compared to untransfected neighbor neurons (n=23, -73.33 ± 1.55 mV, p<0.001), and input resistances
measured to hyperpolarizing and depolarizing current pulses were significantly increased in BRAFV600E
neurons (n=61) compared to all other non-BRAF V600E conditions (Figure 7F, G, H). The elevated resting
membrane potential did not explain the increased firing rates in BRAFV600E neurons, as untransfected
neighboring neurons (n=5) did not achieve AP firing frequencies similar to BRAFV600E neurons when
depolarized to -60 mV, and similarly the few BRAFV600E neurons with more negative resting membrane
potentials (n=14, average RMP=-70.66 ± 0.58 mV) generated high frequency trains of action potentials
similar to more depolarized BRAFV600E neurons. Also, subthreshold input resistances did not correlate
significantly with action potential frequencies in either BRAFV600E or control neurons. Taken together,
BRAFV600E transgenes significantly alter the electrophysiological properties of pyramidal neurons
making them more excitable.
3.5 BRAFV600E decreases delayed rectifier potassium currents
Since the combined increased AP firing frequency, SAG ratio, rebound excitation, more
depolarized resting membrane potential and higher input resistance were observed only in BRAFV600E
expressing neurons, and not in any non-BRAFV600E neurons to test alterations in what ionic conductances
make BRAFV600E transgenic neurons hyperexcitable we have performed whole-cell voltage clamp in
GLAST+ BRAFV600E neurons and their untransfected neighbors only. Recorded calcium currents did not
show any significant difference (data not shown). Recording of potassium currents showed a decreased
sustained current sensitive to 25 mM TEA, a known potassium channel inhibitor, measured at the last 100
ms of 500 ms depolarizing voltage pulses across examined range of voltage steps in GLAST+ BRAFV600E
neurons (n=18) compared to their untransfected neighbors (n=8, p<0.01; Table 1, 2), preserving kinetic
properties of activation (Figure 8A, B, C, D, F).
3.6 Elevated IH in BRAFV600E neurons
In response to hyperpolarizing current pulses in whole-cell current clamp mode BRAFV600E
neurons in either GLAST+ or NESTIN+ condition displayed an initial deflection, SAG ratio that was absent
in untransfected neighbor neurons, control-FP neurons, and in BRAFwt neurons (Figure 5A lower panel,
5E). SAG ratio was calculated as 𝑆𝐴𝐺 = (1 −
𝑉𝑅𝑀𝑃− 𝑉𝑠𝑠
𝑉𝑅𝑀𝑃− 𝑉𝑚𝑖𝑛) ∗ 100%, VRMP – Resting Membrane Potential,
Vss – stable-state voltage in the last 100 ms of 1 second -40 pA pulse, Vmin – minimal initial voltage
deflection in response to 1 second -40 pA pulse. GLAST+ BRAFV600E neurons (n=57) had average SAG
ratio of 23.41 ± 1.33% that was significantly larger than in their untransfected neighbor neurons of 4.46 ±
1.48% (n=20, p=0.001); and in GLAST+ control-FP neurons 6.82 ± 1.86% (n=15, p<0.001); and in
GLAST+ BRAFwt neurons 6.59 ± 1.94% (n=17, p<0.001) (Figure 5E). In NESTIN+ BRAFV600E average
SAG was 24.48 ± 2.27 % (n=8) and it was significantly increased compared to all non-BRAF V600E
conditions (p<0.001, Figure 5E); in NESTIN+ control-FP the average SAG ratio was 10.03 ± 1.77% (n=5);
and in NESTIN+ untransfected neighbor the average SAG ratio was 6.59 ± 2.41% (n=4).
Average rebound excitation measured as an overshoot above RMP at the end of 1 second -40 pA
current step was also larger in GLAST+ BRAFV600E neurons (n=42) 2.28 ± 0.24 mV compared to their
untransfected neighbor neurons 0.69 ± 0.18 mV (n=22, p<0.01); to GLAST+ control-FP neurons 0.37 ±
0.11 mV (n=14, p<0.01); to GLAST+ BRAFwt neurons 0.34 ± 0.06 mV (n=17, p<0.01); to NESTIN+
untransfected neighbor 0.18 ± 0.27 mV (n=4, p<0.05); to NESTIN+ control-FP 0.47 ± 0.13 mV (n=5,
p<0.05); in NESTIN+ BRAFV600E rebound excitation was increased (n=7) 1.82 ± 0.17 mV compared to
non-BRAFV600E conditions (to GLAST+ BRAFwt - p<0.001; to GLAST+ untransfected neighbor –
p=0.02; to GLAST+ control-FP – p<0.001; to NESTIN+ untransfected neighbor – p<0.001; to NESTIN+
control-FP – p<0.001) (Figure 5F). In 20.34% of GLAST+ BRAFV600E neurons (12/59) and in 11.11%
of NESTIN+ BRAFV600E (1/9) rebound excitation resulted in AP firing (Figure 5G upper right panel).
Increased SAG ratio and rebound excitation has been previously shown in layer 5 cortical,
hippocampal and non-cortical neurons in mice, rats and cats to be associated with hyperpolarization
activated conductances 132-136. To test whether BRAFV600E expressing neurons have increased
hyperpolarization activated conductance we recorded cells in whole-cell voltage clamp configuration and
show that BRAFV600E neurons have Ih that is absent in all other conditions and have half activation
voltage of V1/2 = -82.79 mV and the slope factor k = 11.58-1 mV using recording protocol that hold the cell
at -50 mV and the first voltage step is at -120 mV with 5 mV increase for 1.5 seconds (Figure 9; Table 3).
This current was blocked with application of 50 µM ZD7288, a known Ih inhibitor in perfusion system and
recorded at least 5 minutes later. Ih peak was only significantly increased in GLAST+ BRAFV600E
neurons (n=17) compared to their untransfected neighbors (n=6, p<0.05), however when normalized to cell
capacitance the Ih peak density was significantly increased in GLAST+ BRAFV600E neurons compared
to their untransfected neighbors (p<0.001), and to NESTIN+ control-FP (n=4, p<0.001). In NESTIN+
BRAFV600E neurons (n=4) the Ih peak density was significantly increased compared to GLAST+
untransfected neighbors (p<0.05), and to NESTIN+ control-FP (p<0.05). Consistent with that application
of ZD7288 decreased SAG and rebound excitation in BRAFV600E neurons (data not shown).
Hyperpolarization activated conductance is generated through ion channels with subunit composition of
HCN1-4 137, 138. Koh et al. 14 finding that HCN1 is downregulated in GG patients and in mouse model
suggest that HCN channels subunit composition may change and, possibly, the expression may be
redistributed across different cellular compartments.
4. Discussion
Here we showed that introduction of human BRAFV600E, an LNETs associated mutation that
constitutively activate BRAF in a RAS-independent manner, into radial glia progenitors using different
driver promoters - GLAST and NESTIN, increased astrogenesis in the first case and neurogenesis in the
second case. The results from GLAST experiments consistent with previous studies that showed increased
astrogenesis when constitutive MEK1 a downstream target of BRAF was expressed in hGFAPCre/CAG-
loxpSTOPloxp-Mek1S218, S222E mouse line 128 and also in tamoxifen induced knockdown of NF1, a RAS-
GTPase activating protein in hGFAPCre driven mouse line 129 and in GG patients and BRAFV637E
transgenic mouse line that were electroporated with episomal Cre plasmid 14. Additionally, Gronych et al.
139 showed that introducing truncated BRAFV600E containing kinase domain using retroviral vector into
neonatal Ntv mice under promoter derived from the human NESTIN gene was sufficient to model tumor
induced astrogenesis observed in pilocytic astrocytoma, another LNET entity. However full length
BRAFV600E did not have such an effect suggesting that there is an increased negative regulation of
BRAF activity through possible phosphorylation of inhibiting residues on C-terminus domain in later
progenitors pool available at birth 119, 140-143. It may also reflect possible increased requirement of Hsp90
stabilizing binding in the full length BRAFV600E protein compared to truncated version when introduced
in postnatal animals 144, 145. Those experimental studies mostly targeted progenitor population that may
already have switched to glial fate. Together with our work this suggest that there are, probably, at least
two separate populations of progenitors that may overlap at some developmental stage 116-118, 146, this
notion is also supported by the previous work in which the effects of overactivation of RAS-RAF-ERK
pathway was examined in Neurog2 driven and in Ascl1 driven radial glia progenitors, that were proposed
as a progenitor molecular fate switch. Neurog2 driving the excitatory neuronal differentiation, and RAS-
RAF-ERK pathway activation cause switching off Neurog2 and turning on Ascl1, through direct
phosphorylation by ERK, subsequently driving inhibitory interneuronal differentiation at the low levels of
RAS-RAF-ERK pathway activation, oligodendrogenesis and astrogenesis at the high levels of RAS-RAF-
ERK pathway activation. Which was proposed as a probable explanation for different LNETs
histopathology 127.
4.1 BRAFV600E LNETs, MCD histopathology and inflammation
Increased number of mislocalized neurons in lower cortical layers in both our GLAST+ and
NESTIN+ BRAFV600E transgenic slices, together with balloon-like cells and clusters, increased
astrogenesis in GLAST+ suggest that we partially recaptured histopathology of LNETs. Further,
NESTIN+ BRAFV600E slices also showed increased soma size of the transgenic cells in the
subventricular area compared to neurons in the upper cortical layers. Membrane blebbing, and DNA
fragmentation, signs of apoptosis and pyroptosis were not observed in those balloon-like cells, additional
examination with anti-caspase-3 immunostaining in the selected slices did not show immunoreactivity.
Increased inflammatory immune response and activation of classic complement pathway in
current work is consistent with microarray study in GG resected tissue 69 and recent publication by Koh
and colleagues 14, it was also reported in cortical tubers resected tissue 63, 68. However inflammatory
response in those studies may result from seizure activity. In current work video-ECoG recording did not
show any behavioral manifestations of seizures in the seven recorded animals with BRAFV600E under
GLAST promoter, also presence of electrocorticographic seizures was rarely observed suggesting that
seizures cannot account for activation of inflammatory pathways using current experimental design with
GLAST+ driving promoter. This is supported by RNA sequencing results that did not show increase of
IL-1R1, which mediates biological response to IL-1β and is increased in neurons and subsequently in
astrocytes after seizures 147-149. However, this possibility cannot be completely excluded. Activation of
inflammatory innate immunity has been shown to precipitate seizures in mouse model of kainate-induced
seizures 150, 151 through phosphorylation of NR2B subunits of NMDA channels by Src serine/threonine
kinase. Sequential injection of complement complex components was sufficient to induce seizures in rats’
hippocampus 152. In case our manipulation have similar results, but requires a longer time to induce
seizures or dependent on the presence of “threshold” number of affected neuronal component (GLAST+
vs. NESTIN+), inhibition of inflammatory pathways may be used to reduce seizures 150, 153.
4.2 BRAFV600E and neuronal hyperexcitability
Increased excitability properties in BRAFV600E neurons observed in current work are described
for the first time. Previous studies that concentrated on cortical tissue resected from FCD patients and
TSC patients 51, 54, 74-76 did not observe significant increase in action potential firing properties in
examined malformed components and in mouse model of synapsin-driven TSC1 KO 77. Those studies
suggested that the difference in synaptic circuit excitability account for seizures observed in FCD and
TSC patients. In current work BRAFV600E neurons with depolarized resting membrane potential,
increased input resistance, low capacitance fired three times more action potentials then untransfected
neighbor neurons, or control neurons transfected with fluorophore only. In addition, neurons transgenic
for wild type BRAF displayed similar properties to control conditions. Significant increase in
hyperpolarization activated depolarizing conductance (IH) observed in BRAFV600E neurons may explain
more depolarized resting membrane potential. Indeed, IH inhibition with ZD7288 hyperpolarized
membrane potential in neurons, but it did not alter significantly action potential firing frequency. The
reducing effects of IH on neuronal excitability suggest that increased IH conductance may be
compensatory and counteract the hyperexcitability changes observed in BRAFV600E neurons, thus
working to reduce input resistance 138, 154. IH in dendrites has been previously shown to decrease
amplitudes of propagating EPSPs and to dampen temporal dendritic summation 155-157. In addition to
increased IH conductance BRAFV600E neurons had decreased sustained potassium currents, which may
contribute to action potential adaptation 158-160. Indeed, when retigabine 161 a Kv7.2, Kv7.3 and Kv7.4
activator was acutely applied to BRAFV600E neurons it decreased action potential firing frequency by
30%. This was consistent with previous studies in hippocampal pyramidal neurons 162, 163, that also
showed opposite effect of blocking Kv7 channels with XE-991, which increased action potential firing
frequency. Opening of multiple different voltage sensitive potassium channels may contribute to
sustained, non-inactivating potassium currents, one of this channels is Kv1, global inhibition of which
with α-dendrotoxin in avian nucleus magnocellularis neurons has been shown to depolarize membrane
potential by about 5 mV, increase input resistance two-fold, hyperpolarized action potential voltage
threshold by 8-10 mV and decreased rheobase 164. This suggest that retigabine in case of BRAFV600E
associated epilepsy may be used to reduce seizures.
In a small number of BRAFV600E neurons we have also observed post-action potential
depolarization waves, that were increased in potassium currents recording, when we substituted Ca2+ ions
with Co2+ ions in the extracellular solution. One of the possible explanations to post-action potential
depolarizing waves may be presence of gap junctions 165, or pannexin hemichannels as has been shown in
severe inflammation with epileptic seizures Rasmussen encephalitis 166.
Reflecting on previous studies in FCD and TSC resected cortical tissue 51, 54, 74-77 that showed
increased action potential dependent glutamatergic synaptic events frequencies in TSC compared to FCD
tissue and the opposite effect on GABAergic events frequencies. In addition, they showed large amplitude
GABAergic pacemaker rhythmic events in immature looking pyramidal neurons and decreased
frequencies of action potential dependent mixed glutamatergic and GABAergic events in normal looking
neurons from severe FCD cortical areas. This led us to examine the action potential dependent mixed
glutamatergic and GABAergic events in our preparation as an initial step. Interestingly, in current work
action potential dependent mixed synaptic events frequencies were increased in untransfected neighbor
neurons compared to BRAFV600E neurons, and was comparable to PIK3CAE545K neurons, TSC1 KD
neurons, their untransfected neighbors, and TSC2 untransfected neighbors. sPSCs frequencies in
BRAFV600E neurons were higher compared to control-FP and to BRAFwt neurons, additionally sPSCs
amplitudes, although significantly different did look similar in all conditions, which shows that increased
IH is not sufficient to counteract hyperexcitability changes in synaptic activity.
Depolarized membrane potential due to IH, lower rheobase, increased input resistance,
hyperpolarized action potential voltage threshold, probably due to decreased Kv1 mediated IK+ currents
allows BRAFV600E to fire action potentials in response to smaller depolarizing current input. Increased
action potential dependent synaptic activity, which is largely mediated by miniature Post-Synaptic
Currents (mPSCs) suggest that the neuronal network is more excitable than in control-FP and in BRAFwt
conditions and that increased frequencies may increase probability of action potential firing. Since those
neurons have a low action potential voltage threshold the may fire more action potentials in response to
similar synaptic inputs as in control-FP. This was not tested yet, and further experiments to elucidate it
may need to be performed.
4.3 Caveats in current work
In current work we have tested the effect of acute BRAFV600E inhibition with specific blocker
Vemurafenib (PLX4032, PLX4720) on excitability in BRAFV600E neurons. This FDA approved cancer
medication was developed for unresectable or metastatic melanoma treatment 167, 168. Preincubation of
BRAFV600E transfected cortical slices in 10-50 µM of Vemurafenib for 1-5h did decrease action
potential firing frequency, but this effect was indistinguishable in slices preincubated in comparable
amount of solvent (DMSO). This was also consistent with previous work examining the effect of DMSO
on neuronal excitability in layer 2 of perirhinal cortex 169. Similar results were obtained for Rapamycin
experiments, which was also dissolved in DMSO.
Histopathological examination of immunopositivity to CD34, a hematopoietic stem cell marker
previously shown to label extensively LNETs 35, showed only few immunopositive neuronal cells.
Electrocorticographic recordings did not show any behavioral manifestation of seizures in
GLAST+ BRAFV600E transgenic mice. The increased astrogenesis and decrease neuronal content
suggest that this manipulation may produce a different malformation which is not Ictogenic in mice.
However, electrocorticographic recordings were not performed in NESTIN+ BRAFV600E mice, which
has more neurons compared to GLAST+ BRAFV600E and may have sufficient number of BRAFV600E
transgenic neurons that are as hyperexcitable as GLAST+ BRAFV600E neurons to initiate seizures. This
would be addressed in the next set of experiments in ex-vivo cortical slices with ChR2 added to the
plasmid mix. Stimulation of BRAFV600E transgenic neurons with ChR2 in NESTIN+ condition with
high power blue laser under high extracellular potassium concentration may differentiate BRAFV600E
condition from control-FP based on the amount of stimulation required to initiate ictal activity (threshold)
170. Also electrocorticographic recording of NESTIN+ BRAFV600E freely moving mice to examine
whether seizures are manifested in those animals.
4.4 Short summary
In summary, our study proposes that constitutively active RAS-RAF-ERK pathway resulting
from BRAFV600E mutated protein switch cell fate dependent on the driver promoter, have
histopathological features akin to LNETs, alters gene signature in the affected cortical tissue, increasing
inflammation, and, increase excitability in neurons with BRAFV600E mutation that may contribute to
seizure generation.
References
1. Andre, V. M. et al. Cytomegalic interneurons: a new abnormal cell type in severe pediatric cortical
dysplasia. J. Neuropathol. Exp. Neurol. 66, 491-504 (2007).
2. Anthony, T. E. & Heintz, N. Genetic lineage tracing defines distinct neurogenic and gliogenic stages of
ventral telencephalic radial glial development. Neural Dev. 3, 30-8104-3-30 (2008).
3. Aronica, E. et al. Gene expression profile analysis of epilepsy-associated gangliogliomas.
Neuroscience 151, 272-292 (2008).
4. Baek, S. T. et al. An AKT3-FOXG1-reelin network underlies defective migration in human focal
malformations of cortical development. Nat. Med. 21, 1445-1454 (2015).
5. Balosso, S. et al. A novel non-transcriptional pathway mediates the proconvulsive effects of
interleukin-1beta. Brain 131, 3256-3265 (2008).
6. Barkovich, A. J., Dobyns, W. B. & Guerrini, R. Malformations of cortical development and epilepsy.
Cold Spring Harb Perspect. Med. 5, a022392 (2015).
7. Bast, T., Ramantani, G., Seitz, A. & Rating, D. Focal cortical dysplasia: prevalence, clinical
presentation and epilepsy in children and adults. Acta Neurol. Scand. 113, 72-81 (2006).
8. Becker, A. J. et al. Mutational analysis of TSC1 and TSC2 genes in gangliogliomas. Neuropathol.
Appl. Neurobiol. 27, 105-114 (2001).
9. Ben-Ari Fuchs, S. et al. GeneAnalytics: An Integrative Gene Set Analysis Tool for Next Generation
Sequencing, RNAseq and Microarray Data. OMICS 20, 139-151 (2016).
10. Berg, A. T. et al. Frequency, prognosis and surgical treatment of structural abnormalities seen with
magnetic resonance imaging in childhood epilepsy. Brain 132, 2785-2797 (2009).
11. Blumcke, I. et al. Low-grade epilepsy-associated neuroepithelial tumours - the 2016 WHO
classification. Nat. Rev. Neurol. 12, 732-740 (2016).
12. Blumcke, I., Aronica, E., Urbach, H., Alexopoulos, A. & Gonzalez-Martinez, J. A. A neuropathology-
based approach to epilepsy surgery in brain tumors and proposal for a new terminology use for long-term
epilepsy-associated brain tumors. Acta Neuropathol. 128, 39-54 (2014).
13. Blumcke, I., Beck, H., Lie, A. A. & Wiestler, O. D. Molecular neuropathology of human mesial
temporal lobe epilepsy. Epilepsy Res. 36, 205-223 (1999).
14. Blumcke, I. et al. The clinicopathologic spectrum of focal cortical dysplasias: a consensus
classification proposed by an ad hoc Task Force of the ILAE Diagnostic Methods Commission. Epilepsia
52, 158-174 (2011).
15. Blumcke, I. & Wiestler, O. D. Gangliogliomas: an intriguing tumor entity associated with focal
epilepsies. J. Neuropathol. Exp. Neurol. 61, 575-584 (2002).
16. Boer, K. et al. Gene expression analysis of tuberous sclerosis complex cortical tubers reveals
increased expression of adhesion and inflammatory factors. Brain Pathol. 20, 704-719 (2010).
17. Bollag, G. et al. Clinical efficacy of a RAF inhibitor needs broad target blockade in BRAF-mutant
melanoma. Nature 467, 596-599 (2010).
18. Brady, D. C. et al. Copper is required for oncogenic BRAF signalling and tumorigenesis. Nature 509,
492-496 (2014).
19. Cahoy, J. D. et al. A transcriptome database for astrocytes, neurons, and oligodendrocytes: a new
resource for understanding brain development and function. J. Neurosci. 28, 264-278 (2008).
20. Cary, L. C. et al. Transposon mutagenesis of baculoviruses: analysis of Trichoplusia ni transposon
IFP2 insertions within the FP-locus of nuclear polyhedrosis viruses. Virology 172, 156-169 (1989).
21. Cataltepe, O., Turanli, G., Yalnizoglu, D., Topcu, M. & Akalan, N. Surgical management of temporal
lobe tumor-related epilepsy in children. J. Neurosurg. 102, 280-287 (2005).
22. Cavanagh, J. B. On certain small tumours encountered in the temporal lobe. Brain 81, 389-405
(1958).
23. Cepeda, C. et al. Pediatric cortical dysplasia: correlations between neuroimaging, electrophysiology
and location of cytomegalic neurons and balloon cells and glutamate/GABA synaptic circuits. Dev.
Neurosci. 27, 59-76 (2005).
24. Cepeda, C. et al. Enhanced GABAergic network and receptor function in pediatric cortical dysplasia
Type IIB compared with Tuberous Sclerosis Complex. Neurobiol. Dis. 45, 310-321 (2012).
25. Cepeda, C. et al. Epileptogenesis in pediatric cortical dysplasia: the dysmature cerebral developmental
hypothesis. Epilepsy Behav. 9, 219-235 (2006).
26. Cepeda, C. et al. Immature neurons and GABA networks may contribute to epileptogenesis in
pediatric cortical dysplasia. Epilepsia 48 Suppl 5, 79-85 (2007).
27. Cepeda, C. et al. Comparative study of cellular and synaptic abnormalities in brain tissue samples
from pediatric tuberous sclerosis complex and cortical dysplasia type II. Epilepsia 51 Suppl 3, 160-165
(2010).
28. Cepeda, C. et al. In Rasmussen encephalitis, hemichannels associated with microglial activation are
linked to cortical pyramidal neuron coupling: a possible mechanism for cellular hyperexcitability. CNS
Neurosci. Ther. 21, 152-163 (2015).
29. Cepeda, C. et al. Pacemaker GABA synaptic activity may contribute to network synchronization in
pediatric cortical dysplasia. Neurobiol. Dis. 62, 208-217 (2014).
30. Cepeda, C. et al. Morphological and electrophysiological characterization of abnormal cell types in
pediatric cortical dysplasia. J. Neurosci. Res. 72, 472-486 (2003).
31. Chen, F. & LoTurco, J. A method for stable transgenesis of radial glia lineage in rat neocortex by
piggyBac mediated transposition. J. Neurosci. Methods 207, 172-180 (2012).
32. Chen, F., Maher, B. J. & LoTurco, J. J. piggyBac transposon-mediated cellular transgenesis in
mammalian forebrain by in utero electroporation. Cold Spring Harb Protoc. 2014, 741-749 (2014).
33. Cheng, P., Alberts, I. & Li, X. The role of ERK1/2 in the regulation of proliferation and
differentiation of astrocytes in developing brain. Int. J. Dev. Neurosci. 31, 783-789 (2013).
34. Chung, L. et al. Lovastatin suppresses hyperexcitability and seizure in Angelman syndrome model.
Neurobiol. Dis. 110, 12-19 (2018).
35. Clements, J. D. & Bekkers, J. M. Detection of spontaneous synaptic events with an optimally scaled
template. Biophys. J. 73, 220-229 (1997).
36. Cong, L. et al. Multiplex genome engineering using CRISPR/Cas systems. Science 339, 819-823
(2013).
37. Conti, V. et al. Focal dysplasia of the cerebral cortex and infantile spasms associated with somatic
1q21.1-q44 duplication including the AKT3 gene. Clin. Genet. 88, 241-247 (2015).
38. Crino, P. B. The mTOR signalling cascade: paving new roads to cure neurological disease. Nat. Rev.
Neurol. 12, 379-392 (2016).
39. Crino, P. B. mTOR signaling in epilepsy: insights from malformations of cortical development. Cold
Spring Harb Perspect. Med. 5, 10.1101/cshperspect.a022442 (2015).
40. Crino, P. B. Evolving neurobiology of tuberous sclerosis complex. Acta Neuropathol. 125, 317-332
(2013).
41. Crino, P. B. mTOR: A pathogenic signaling pathway in developmental brain malformations. Trends
Mol. Med. 17, 734-742 (2011).
42. Crino, P. B. Molecular pathogenesis of tuber formation in tuberous sclerosis complex. J. Child
Neurol. 19, 716-725 (2004).
43. Crino, P. B., Duhaime, A. C., Baltuch, G. & White, R. Differential expression of glutamate and
GABA-A receptor subunit mRNA in cortical dysplasia. Neurology 56, 906-913 (2001).
44. Dhillon, A. S., Meikle, S., Yazici, Z., Eulitz, M. & Kolch, W. Regulation of Raf-1 activation and
signalling by dephosphorylation. EMBO J. 21, 64-71 (2002).
45. Ding, S. et al. Efficient transposition of the piggyBac (PB) transposon in mammalian cells and mice.
Cell 122, 473-483 (2005).
46. Dodgshun, A. J. et al. Disseminated glioneuronal tumors occurring in childhood: treatment outcomes
and BRAF alterations including V600E mutation. J. Neurooncol. 128, 293-302 (2016).
47. Eisen, M. B., Spellman, P. T., Brown, P. O. & Botstein, D. Cluster analysis and display of genome-
wide expression patterns. Proc. Natl. Acad. Sci. U. S. A. 95, 14863-14868 (1998).
48. Fassunke, J. et al. Array analysis of epilepsy-associated gangliogliomas reveals expression patterns
related to aberrant development of neuronal precursors. Brain 131, 3034-3050 (2008).
49. Figueres-Onate, M., Garcia-Marques, J., Pedraza, M., De Carlos, J. A. & Lopez-Mascaraque, L.
Spatiotemporal analyses of neural lineages after embryonic and postnatal progenitor targeting combining
different reporters. Front. Neurosci. 9, 87 (2015).
50. Fink, S. L. & Cookson, B. T. Caspase-1-dependent pore formation during pyroptosis leads to osmotic
lysis of infected host macrophages. Cell. Microbiol. 8, 1812-1825 (2006).
51. Frater, J. L., Prayson, R. A., Morris III, H. H. & Bingaman, W. E. Surgical pathologic findings of
extratemporal-based intractable epilepsy: a study of 133 consecutive resections. Arch. Pathol. Lab. Med.
124, 545-549 (2000).
52. Gamkrelidze, G., Giaume, C. & Peusner, K. D. The differential expression of low-threshold sustained
potassium current contributes to the distinct firing patterns in embryonic central vestibular neurons. J.
Neurosci. 18, 1449-1464 (1998).
53. Gotz, M. Glial cells generate neurons--master control within CNS regions: developmental
perspectives on neural stem cells. Neuroscientist 9, 379-397 (2003).
54. Gotz, M., Hartfuss, E. & Malatesta, P. Radial glial cells as neuronal precursors: a new perspective on
the correlation of morphology and lineage restriction in the developing cerebral cortex of mice. Brain
Res. Bull. 57, 777-788 (2002).
55. Grbovic, O. M. et al. V600E B-Raf requires the Hsp90 chaperone for stability and is degraded in
response to Hsp90 inhibitors. Proc. Natl. Acad. Sci. U. S. A. 103, 57-62 (2006).
56. Gronych, J. et al. An activated mutant BRAF kinase domain is sufficient to induce pilocytic
astrocytoma in mice. J. Clin. Invest. 121, 1344-1348 (2011).
57. Gu, N., Vervaeke, K., Hu, H. & Storm, J. F. Kv7/KCNQ/M and HCN/h, but not KCa2/SK channels,
contribute to the somatic medium after-hyperpolarization and excitability control in CA1 hippocampal
pyramidal cells. J. Physiol. 566, 689-715 (2005).
58. Guan, D., Armstrong, W. E. & Foehring, R. C. Electrophysiological properties of genetically
identified subtypes of layer 5 neocortical pyramidal neurons: Ca(2)(+) dependence and differential
modulation by norepinephrine. J. Neurophysiol. 113, 2014-2032 (2015).
59. Guerrini, R. et al. Diagnostic methods and treatment options for focal cortical dysplasia. Epilepsia 56,
1669-1686 (2015).
60. Harvey, A. S., Cross, J. H., Shinnar, S., Mathern, G. W. & ILAE Pediatric Epilepsy Surgery Survey
Taskforce. Defining the spectrum of international practice in pediatric epilepsy surgery patients. Epilepsia
49, 146-155 (2008).
61. Hawkins, V. E. et al. HCN channels contribute to serotonergic modulation of ventral surface
chemosensitive neurons and respiratory activity. J. Neurophysiol. 113, 1195-1205 (2015).
62. He, C., Chen, F., Li, B. & Hu, Z. Neurophysiology of HCN channels: from cellular functions to
multiple regulations. Prog. Neurobiol. 112, 1-23 (2014).
63. He, C., Chen, F., Li, B. & Hu, Z. Neurophysiology of HCN channels: from cellular functions to
multiple regulations. Prog. Neurobiol. 112, 1-23 (2014).
64. He, J. et al. Cellular mechanisms of cobalt-induced hippocampal epileptiform discharges. Epilepsia
50, 99-115 (2009).
65. Hoischen, A. et al. Comprehensive characterization of genomic aberrations in gangliogliomas by
CGH, array-based CGH and interphase FISH. Brain Pathol. 18, 326-337 (2008).
66. Huang da, W., Sherman, B. T. & Lempicki, R. A. Bioinformatics enrichment tools: paths toward the
comprehensive functional analysis of large gene lists. Nucleic Acids Res. 37, 1-13 (2009).
67. Huang da, W., Sherman, B. T. & Lempicki, R. A. Systematic and integrative analysis of large gene
lists using DAVID bioinformatics resources. Nat. Protoc. 4, 44-57 (2009).
68. IBM Corp. IBM SPSS Statistics for Windows, Version 24.0. Armonk, NY: IBM Corp. (2016).
69. Jansen, L. A. et al. PI3K/AKT pathway mutations cause a spectrum of brain malformations from
megalencephaly to focal cortical dysplasia. Brain 138, 1613-1628 (2015).
70. Jung, S. et al. Downregulation of dendritic HCN channel gating in epilepsy is mediated by altered
phosphorylation signaling. J. Neurosci. 30, 6678-6688 (2010).
71. Kakkar, A. et al. Alterations in BRAF gene, and enhanced mTOR and MAPK signaling in
dysembryoplastic neuroepithelial tumors (DNTs). Epilepsy Res. 127, 141-151 (2016).
72. Kim, D., Langmead, B. & Salzberg, S. L. HISAT: a fast spliced aligner with low memory
requirements. Nat. Methods 12, 357-360 (2015).
73. Kim, S. Y., Choi, S. Y., Chao, W. & Volsky, D. J. Transcriptional regulation of human excitatory
amino acid transporter 1 (EAAT1): cloning of the EAAT1 promoter and characterization of its basal and
inducible activity in human astrocytes. J. Neurochem. 87, 1485-1498 (2003).
74. Kobow, K. & Blumcke, I. Epigenetics in epilepsy. Neurosci. Lett. (2017).
75. Koelsche, C. et al. Mutant BRAF V600E protein in ganglioglioma is predominantly expressed by
neuronal tumor cells. Acta Neuropathol. 125, 891-900 (2013).
76. Koh, H. Y. et al. BRAF somatic mutation contributes to intrinsic epileptogenicity in pediatric brain
tumors. Nat. Med. 24, 1662-1668 (2018).
77. Kohling, R., Senner, V., Paulus, W. & Speckmann, E. J. Epileptiform activity preferentially arises
outside tumor invasion zone in glioma xenotransplants. Neurobiol. Dis. 22, 64-75 (2006).
78. Kornev, A. P., Haste, N. M., Taylor, S. S. & Eyck, L. F. Surface comparison of active and inactive
protein kinases identifies a conserved activation mechanism. Proc. Natl. Acad. Sci. U. S. A. 103, 17783-
17788 (2006).
79. Kriegstein, A. R. & Gotz, M. Radial glia diversity: a matter of cell fate. Glia 43, 37-43 (2003).
80. Kuba, H., Yamada, R., Ishiguro, G. & Adachi, R. Redistribution of Kv1 and Kv7 enhances neuronal
excitability during structural axon initial segment plasticity. Nat. Commun. 6, 8815 (2015).
81. Lamarre, S. et al. Optimization of an RNA-Seq Differential Gene Expression Analysis Depending on
Biological Replicate Number and Library Size. Front. Plant. Sci. 9, 108 (2018).
82. Lawrence, J. J. et al. Somatodendritic Kv7/KCNQ/M channels control interspike interval in
hippocampal interneurons. J. Neurosci. 26, 12325-12338 (2006).
83. Leach, J. L., Greiner, H. M., Miles, L. & Mangano, F. T. Imaging spectrum of cortical dysplasia in
children. Semin. Roentgenol. 49, 99-111 (2014).
84. Lee, J. H. et al. De novo somatic mutations in components of the PI3K-AKT3-mTOR pathway cause
hemimegalencephaly. Nat. Genet. 44, 941-945 (2012).
85. Lee, J. T. et al. PLX4032, a potent inhibitor of the B-Raf V600E oncogene, selectively inhibits
V600E-positive melanomas. Pigment Cell. Melanoma Res. 23, 820-827 (2010).
86. Li, S. et al. RAS/ERK signaling controls proneural genetic programs in cortical development and
gliomagenesis. J. Neurosci. 34, 2169-2190 (2014).
87. Li, X. et al. MEK Is a Key Regulator of Gliogenesis in the Developing Brain. Neuron 75, 1035-1050
(2012).
88. Lim, J. S. et al. Somatic Mutations in TSC1 and TSC2 Cause Focal Cortical Dysplasia. Am. J. Hum.
Genet. 100, 454-472 (2017).
89. Lim, J. S. et al. Brain somatic mutations in MTOR cause focal cortical dysplasia type II leading to
intractable epilepsy. Nat. Med. 21, 395-400 (2015).
90. Ljungberg, M. C., Sunnen, C. N., Lugo, J. N., Anderson, A. E. & D'Arcangelo, G. Rapamycin
suppresses seizures and neuronal hypertrophy in a mouse model of cortical dysplasia. Dis. Model. Mech.
2, 389-398 (2009).
91. Lothian, C. & Lendahl, U. An evolutionarily conserved region in the second intron of the human
nestin gene directs gene expression to CNS progenitor cells and to early neural crest cells. Eur. J.
Neurosci. 9, 452-462 (1997).
92. Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-
seq data with DESeq2. Genome Biol. 15, 550-014-0550-8 (2014).
93. Luyken, C. et al. The spectrum of long-term epilepsy-associated tumors: long-term seizure and tumor
outcome and neurosurgical aspects. Epilepsia 44, 822-830 (2003).
94. Magee, J. C. Dendritic integration of excitatory synaptic input. Nature Reviews Neuroscience 1, 181-
190 (2000).
95. Magee, J. C. Dendritic Ih normalizes temporal summation in hippocampal CA1 neurons. Nat.
Neurosci. 2, 848 (1999).
96. Magee, J. C. Dendritic hyperpolarization-activated currents modify the integrative properties of
hippocampal CA1 pyramidal neurons. J. Neurosci. 18, 7613-7624 (1998).
97. Maher, B. J. & LoTurco, J. J. Disrupted-in-schizophrenia (DISC1) functions presynaptically at
glutamatergic synapses. PLoS One 7, e34053 (2012).
98. Marchi, N. et al. Antagonism of peripheral inflammation reduces the severity of status epilepticus.
Neurobiol. Dis. 33, 171-181 (2009).
99. Maroso, M. et al. Toll-like receptor 4 and high-mobility group box-1 are involved in ictogenesis and
can be targeted to reduce seizures. Nat. Med. 16, 413-419 (2010).
100. Martin, K. R. et al. The genomic landscape of tuberous sclerosis complex. Nat. Commun. 8, 15816
(2017).
101. Martinoni, M. et al. BRAF V600E mutation in neocortical posterior temporal epileptogenic
gangliogliomas. J. Clin. Neurosci. 22, 1250-1253 (2015).
102. Marucci, G. et al. Mutant BRAF in low-grade epilepsy-associated tumors and focal cortical
dysplasia. Ann. Clin. Transl. Neurol. 1, 130-134 (2014).
103. Marucci, G., Martinoni, M. & Giulioni, M. Relationship between focal cortical dysplasia and
epilepsy-associated low-grade tumors: an immunohistochemical study. APMIS 121, 22-29 (2013).
104. Matsumoto M., N. T. Mersenne twister: a 623-dimensionally equidistributed uniform pseudo-
random number generator. ACM Transactions on Modeling and Computer Simulation 8, 3-30 (1998).
105. Maynard, L. M. et al. Epilepsy prevalence and severity predictors in MRI-identified focal cortical
dysplasia. Epilepsy Res. 132, 41-49 (2017).
106. McIlwain, D. R., Berger, T. & Mak, T. W. Caspase functions in cell death and disease. Cold Spring
Harb Perspect. Biol. 5, a008656 (2013).
107. Metsalu, T. & Vilo, J. ClustVis: a web tool for visualizing clustering of multivariate data using
Principal Component Analysis and heatmap. Nucleic Acids Res. 43, W566-70 (2015).
108. Mizuguchi, M. & Takashima, S. Neuropathology of tuberous sclerosis. Brain Dev. 23, 508-515
(2001).
109. Mo, Z. L., Adamson, C. L. & Davis, R. L. Dendrotoxin-sensitive K(+) currents contribute to
accommodation in murine spiral ganglion neurons. J. Physiol. 542, 763-778 (2002).
110. Morris, H. H. et al. Ganglioglioma and intractable epilepsy: clinical and neurophysiologic features
and predictors of outcome after surgery. Epilepsia 39, 307-313 (1998).
111. Mu, J., Slevin, J. C., Qu, D., McCormick, S. & Adamson, S. L. In vivo quantification of embryonic
and placental growth during gestation in mice using micro-ultrasound. Reprod. Biol. Endocrinol. 6, 34-
7827-6-34 (2008).
112. Nagarajan, L. et al. Seizure outcomes in children with epilepsy after resective brain surgery. Eur. J.
Paediatr. Neurol. 19, 577-583 (2015).
113. Nakashima, M. et al. Somatic Mutations in the MTOR gene cause focal cortical dysplasia type IIb.
Ann. Neurol. 78, 375-386 (2015).
114. Nguyen, L. H. et al. mTOR inhibition suppresses established epilepsy in a mouse model of cortical
dysplasia. Epilepsia 56, 636-646 (2015).
115. Palmini, A., Paglioli, E. & Silva, V. D. Developmental tumors and adjacent cortical dysplasia: single
or dual pathology? Epilepsia 54 Suppl 9, 18-24 (2013).
116. Pelliccia, V. et al. Early and late epilepsy surgery in focal epilepsies associated with long-term
epilepsy-associated tumors. J. Neurosurg., 1-6 (2017).
117. Pertea, M., Kim, D., Pertea, G. M., Leek, J. T. & Salzberg, S. L. Transcript-level expression analysis
of RNA-seq experiments with HISAT, StringTie and Ballgown. Nat. Protoc. 11, 1650-1667 (2016).
118. Pertea, M. et al. StringTie enables improved reconstruction of a transcriptome from RNA-seq reads.
Nat. Biotechnol. 33, 290-295 (2015).
119. Pinto, L. & Gotz, M. Radial glial cell heterogeneity--the source of diverse progeny in the CNS.
Prog. Neurobiol. 83, 2-23 (2007).
120. Prabowo, A. S. et al. BRAF V600E mutation is associated with mTOR signaling activation in
glioneuronal tumors. Brain Pathol. 24, 52-66 (2014).
121. Prabowo, A. S. et al. Landscape of chromosomal copy number aberrations in gangliogliomas and
dysembryoplastic neuroepithelial tumours. Neuropathol. Appl. Neurobiol. 41, 743-755 (2015).
122. Qaddoumi, I. et al. Genetic alterations in uncommon low-grade neuroepithelial tumors: BRAF,
FGFR1, and MYB mutations occur at high frequency and align with morphology. Acta Neuropathol. 131,
833-845 (2016).
123. Ravizza, T. et al. The IL-1beta system in epilepsy-associated malformations of cortical development.
Neurobiol. Dis. 24, 128-143 (2006).
124. Ravizza, T. et al. Innate and adaptive immunity during epileptogenesis and spontaneous seizures:
evidence from experimental models and human temporal lobe epilepsy. Neurobiol. Dis. 29, 142-160
(2008).
125. Ravizza, T. & Vezzani, A. Status epilepticus induces time-dependent neuronal and astrocytic
expression of interleukin-1 receptor type I in the rat limbic system. Neuroscience 137, 301-308 (2006).
126. R-Core-Team. R: A language and environment for statistical computing. R Foundation for Statistical
Computing, Vienna, Austria. (2013).
127. Reich, M. et al. GenePattern 2.0. Nat. Genet. 38, 500-501 (2006).
128. Rice, H., Suth, S., Cavanaugh, W., Bai, J. & Young-Pearse, T. L. In utero electroporation followed
by primary neuronal culture for studying gene function in subset of cortical neurons. J. Vis. Exp. (44). pii:
2103. doi, 10.3791/2103 (2010).
129. Rich, K. M., Goldring, S. & Gado, M. Computed tomography in chronic seizure disorder caused by
glioma. Arch. Neurol. 42, 26-27 (1985).
130. Riviere, J. B. et al. De novo germline and postzygotic mutations in AKT3, PIK3R2 and PIK3CA
cause a spectrum of related megalencephaly syndromes. Nat. Genet. 44, 934-940 (2012).
131. Roskoski, R.,Jr. RAF protein-serine/threonine kinases: structure and regulation. Biochem. Biophys.
Res. Commun. 399, 313-317 (2010).
132. Rossi, M. A. Focal cortical dysplasia-associated tumors: resecting beyond the lesion. Epilepsy Curr.
14, 264-265 (2014).
133. Roy, A. et al. Mouse models of human PIK3CA-related brain overgrowth have acutely treatable
epilepsy. Elife 4, 10.7554/eLife.12703 (2015).
134. Roy, N. S. et al. Promoter-targeted selection and isolation of neural progenitor cells from the adult
human ventricular zone. J. Neurosci. Res. 59, 321-331 (2000).
135. RStudio-Team. RStudio: Integrated Development for R. version 1.0.143. RStudio, Inc., Boston, MA
(2016).
136. Samadani, U., Judkins, A. R., Akpalu, A., Aronica, E. & Crino, P. B. Differential cellular gene
expression in ganglioglioma. Epilepsia 48, 646-653 (2007).
137. Schick, V. et al. Alterations of phosphatidylinositol 3-kinase pathway components in epilepsy-
associated glioneuronal lesions. Epilepsia 48 Suppl 5, 65-73 (2007).
138. Schindelin, J. et al. Fiji: an open-source platform for biological-image analysis. Nat. Methods 9, 676-
682 (2012).
139. Schindler, G. et al. Analysis of BRAF V600E mutation in 1,320 nervous system tumors reveals high
mutation frequencies in pleomorphic xanthoastrocytoma, ganglioglioma and extra-cerebellar pilocytic
astrocytoma. Acta Neuropathol. 121, 397-405 (2011).
140. Schramm, J., Kral, T., Grunwald, T. & Blumcke, I. Surgical treatment for neocortical temporal lobe
epilepsy: clinical and surgical aspects and seizure outcome. J. Neurosurg. 94, 33-42 (2001).
141. Schridde, U., Strauss, U., Brauer, A. U. & van Luijtelaar, G. Environmental manipulations early in
development alter seizure activity, Ih and HCN1 protein expression later in life. Eur. J. Neurosci. 23,
3346-3358 (2006).
142. Siddiqi, F. et al. Fate mapping by piggyBac transposase reveals that neocortical GLAST+
progenitors generate more astrocytes than Nestin+ progenitors in rat neocortex. Cereb. Cortex 24, 508-
520 (2014).
143. Silver, J. M., Rawlings, C. E.,3rd, Rossitch, E.,Jr, Zeidman, S. M. & Friedman, A. H.
Ganglioglioma: a clinical study with long-term follow-up. Surg. Neurol. 35, 261-266 (1991).
144. Sisodiya, S. M. Malformations of cortical development: burdens and insights from important causes
of human epilepsy. Lancet Neurol. 3, 29-38 (2004).
145. Sisodiya, S. M., Fauser, S., Cross, J. H. & Thom, M. Focal cortical dysplasia type II: biological
features and clinical perspectives. Lancet Neurol. 8, 830-843 (2009).
146. Spencer, D. D., Spencer, S. S., Mattson, R. H. & Williamson, P. D. Intracerebral masses in patients
with intractable partial epilepsy. Neurology 34, 432-436 (1984).
147. Stas, J. I., Bocksteins, E., Jensen, C. S., Schmitt, N. & Snyders, D. J. The anticonvulsant retigabine
suppresses neuronal KV2-mediated currents. Sci. Rep. 6, 35080 (2016).
148. Stone, T. J. et al. Comprehensive molecular characterisation of epilepsy-associated glioneuronal
tumours. Acta Neuropathol. (2017).
149. Tamagnini, F., Scullion, S., Brown, J. T. & Randall, A. D. Low concentrations of the solvent
dimethyl sulphoxide alter intrinsic excitability properties of cortical and hippocampal pyramidal cells.
PLoS One 9, e92557 (2014).
150. Thom, M., Blumcke, I. & Aronica, E. Long-term epilepsy-associated tumors. Brain Pathol. 22, 350-
379 (2012).
151. Timofeev, I., Bazhenov, M., Sejnowski, T. & Steriade, M. Cortical hyperpolarization-activated
depolarizing current takes part in the generation of focal paroxysmal activities. Proc. Natl. Acad. Sci. U.
S. A. 99, 9533-9537 (2002).
152. Tran, N. H., Wu, X. & Frost, J. A. B-Raf and Raf-1 are regulated by distinct autoregulatory
mechanisms. J. Biol. Chem. 280, 16244-16253 (2005).
153. Verba, K. A. et al. Atomic structure of Hsp90-Cdc37-Cdk4 reveals that Hsp90 traps and stabilizes an
unfolded kinase. Science 352, 1542-1547 (2016).
154. Wahl-Schott, C. & Biel, M. HCN channels: structure, cellular regulation and physiological function.
Cell Mol. Life Sci. 66, 470-494 (2009).
155. Wan, P. T. et al. Mechanism of activation of the RAF-ERK signaling pathway by oncogenic
mutations of B-RAF. Cell 116, 855-867 (2004).
156. Wang, Y. et al. Neocortical hyperexcitability in a human case of tuberous sclerosis complex and
mice lacking neuronal expression of TSC1. Ann. Neurol. 61, 139-152 (2007).
157. Wang, Y. et al. ERK inhibition rescues defects in fate specification of Nf1-deficient neural
progenitors and brain abnormalities. Cell 150, 816-830 (2012).
158. Way, S. W. et al. Loss of Tsc2 in radial glia models the brain pathology of tuberous sclerosis
complex in the mouse. Hum. Mol. Genet. 18, 1252-1265 (2009).
159. Wold, H. "Nonlinear Iterative Partial Least Squares (NIPALS) Modeling: Some Current
Developments" in P.R. Krishnaiah [ed.]. Multivariate Analysis II, Proceedings of an International
Symposium on Multivariate Analysis held at Wright State University, Dayton, Ohio, June 19-24, 1972.
New York, Academic Press, 383-407 (1973).
160. Wold, H. Aspects opératoires des modèles économétriques et sociologiques. Développement actuel
de l'estimation 'F.P.' (Point fixe) et de la modélisation NIPALS" (linéarisation par itération de moindres
carrés partiels). Économie Appliquée 2-3-4 (1973).
161. Wolf, H. K. & Wiestler, O. D. Surgical pathology of chronic epileptic seizure disorders. Brain
Pathol. 3, 371-380 (1993).
162. Wong, M. Mammalian target of rapamycin (mTOR) activation in focal cortical dysplasia and related
focal cortical malformations. Exp. Neurol. 244, 22-26 (2013).
163. Wu, S. C. et al. piggyBac is a flexible and highly active transposon as compared to sleeping beauty,
Tol2, and Mos1 in mammalian cells. Proc. Natl. Acad. Sci. U. S. A. 103, 15008-15013 (2006).
164. Wyllie, E. et al. Seizure outcome after epilepsy surgery in children and adolescents. Ann. Neurol. 44,
740-748 (1998).
165. Xiong, Z. Q., Qian, W., Suzuki, K. & McNamara, J. O. Formation of complement membrane attack
complex in mammalian cerebral cortex evokes seizures and neurodegeneration. J. Neurosci. 23, 955-960
(2003).
166. Yasin, S. A. et al. Balloon cells in human cortical dysplasia and tuberous sclerosis: isolation of a
pathological progenitor-like cell. Acta Neuropathol. 120, 85-96 (2010).
167. Zaika, O. et al. Angiotensin II regulates neuronal excitability via phosphatidylinositol 4,5-
bisphosphate-dependent modulation of Kv7 (M-type) K+ channels. J. Physiol. 575, 49-67 (2006).
168. Zanello, M. et al. Clinical, Imaging, Histopathological and Molecular Characterization of Anaplastic
Ganglioglioma. J. Neuropathol. Exp. Neurol. 75, 971-980 (2016).
169. Zhang, J. et al. Whole-genome sequencing identifies genetic alterations in pediatric low-grade
gliomas. Nat. Genet. 45, 602-612 (2013).
170. Zhao, J. J. et al. The oncogenic properties of mutant p110alpha and p110beta phosphatidylinositol 3-
kinases in human mammary epithelial cells. Proc. Natl. Acad. Sci. U. S. A. 102, 18443-18448 (2005).
Figure 1. BRAFV600E expression in neuroglial progenitors at embryonic ages E14-E15 causes neuron
migrational delay and cellular dysmorphogenesis. A. In-Utero Electroporation (IUE) experimental
design. B. Representative image of GLAST+ BRAFV600E electroporated mouse brain slice at post-natal
day 180 showing aggregates of balloon-like cells near the pia and aggregates of a crescent shaped cells in
layers 4 and 5. C. Cux1, an upper cortical layer marker and Ctip2, a lower cortical layer marker shows
that most of the GLAST+ BRAFV600E transfected neurons reach appropriate cortical layers relative to
embryonic age at IUE. Insets show zoomed in neurons positive for Cux1, and negative for Ctip2 in upper
layer 2. Lower panel shows 3 neurons located in the lower cortial layers and positive for cux1. D.
Balloon-like cells and aggregates (panels I and II, Roman numerals – R.n.) of balloon-like cells found in
Glast+ BRAFV600E transfected brains. Panels II,IV,VI,VIII (even R.n.) show anti-HA stain confirmation
of pPB-BRAFV600E presence. Panels III-IX with odd R.n. show balloon-like cell stained positive for
both Cux1 and Ctip2. Panels XI-XVII, odd R.n. show a balloon-like cell stained positive for Glial
Fibrilary Acidic Protein (GFAP) an astrogllial markers. Panels XII-XVII, even R.n. show a ballon-like
cell negative for NeuN, a neuronal marker. Panels XIX-XXI showing coronal section from Nestin+
BRAFV600E with balloon-like cells aggregates in a piriform cortex area (white arrowhead). E.
Frequency of slices with balloon-like cells and balloon-like cells aggregates in Glast+ BRAFV600E
transfected brains according to the post-natal age. Each bar represent one brain. Scale bars for B. - 500
µm; C. - 500 µm and 50 µm zoomed in images, Lower panel – 100 µm for the wide field view, 50 µm for
single neurons, 10 µm for single zoomed in neuron; for D. I,X – 50 µm; II-IX, and XI-XVIII, R.n. – 20
µm; * - p<0.05, ** - p<0.01, *** - p<0.001. Error bars are ±2SEM.
Figure 2. BRAFV600E expression in NESTIN+ neuroglial progenitors. A.Representative image of EGFP
positive cells in subventricular area compared to cells in layer 2/3 of somatosensory cortex that are
quantified in B. Area measured in NESTIN+ BRAFV600E EGFP positive cells in A in the upper panel,
and diameter of those cells in the lower panel showing increased size of subventricular located cells
(n=17) compared to layer 2/3 neurons (n=11, paired sample T=7.883, p<0.001 for area and T=8.224,
p<0.001 for diameter) . C. Gross neuron counts (left panel) and scaled to max neuron counts (right panel)
- distance to pia measurement shows that there is a decrease in EGFP positive neuronal content in Glast+
BRAFV600E transfected mouse cortical slices and that higher number of the BRAFV600E transfected
EGFP positive neurons targeted for upper cortical layers under both Glast+ and Nestin+ do not reach their
deisgnated location compared to Glast+ control-FP transfected brain slices (p<0.001), and to Glast+
BRAFwt transfected brain slices (p<0.001), there was also higher number of Glast+ BRAFwt transfected
neurons that did not reach designated cortical layers compared to Glast+ control-FP (p<0.001); significant
difference in neuronal distance to pia was also present between Glast+ BRAFV600E and Nestin+
BRAFV600E (p=0.001) Due to non-homogenous variance - Levene test (3, 28606) = 1674.918,
p<0.0001, Welch test with Games-Howell post-hoc correction were used (3, 9803.913) = 970.974,
p<0.001. The results were confirmed with cumulative distribution nonparametric Mann-Whitney U test –
Glast+ BRAFV600E to Glast+ control-FP, U= 31122376, p<0.001; Glast+ BRAFV600E to Glast+
BRAFwt, U= 18003030, p<0.001; Glast+ BRAFwt to Glast+ control-FP, U= 26624785, p<0.001;
Nestin+ BRAFV600E to Glast+ control-FP U= 11954144.5, p<0.001; Nestin+ BRAFV600E to Glast+
BRAFwt, U= 7186563, p<0.001; Nestin+ BRAFV600E to Glast+ BRAFV600E, U= 12109290.5,
p<0.001. Scale bars for A – 500 µm * - p<0.05,** - p<0.01, *** - p<0.0001. Error bars are ±2SEM.
Figure 3. BRAFV600E transgene in GLAST+ neuroglial progenitors increases gliogenesis and induce
reactive astrogliosis. A. Whole-slice image showing increased GFAP immunoreativity in somatosensory
cortex transfected with GLAST+ BRAFV600E compared to GLAST+ control-FP and to GLAST+
BRAFwt transfected brains(upper 3 panels). Whole-slice image showing increased GFAP
immunoreactivity in more frontal part of somatosensory cortex transfected with NESTIN+ BRAFV600E
(lower panel). B. GFAP positive cells were also immunopositive to astrocytes marker ALDH1L1(white
arrowheads). C. Neuron to astrocytes percent ratio showing increased astrogliosis (EGFP positive cells)
in BRAFV600E electroporated murine cortical slices. Due to non-homogenous variance (Levene test (3,
243) =4.574 and 4.575 for neurons and astrocytes respectively p=0.004), Welch test F(3, 106.49)=183.71,
p<0.001 with Games-Howell posthoc correction was used for statistical comparison. Astrocytes
percentage was increased in Glast+ BRAFV600E electroporated slices (n=44 slices, 8 brains) compared
to Glast+ control-FP only (n=43 slices, 7 brains p<0.001); and compared to Glast+ BRAFwt
electroporated slices (n=18 slices, 5 brains p<0.001). And neuronal percentage was decreased in Glast+
BRAFV600E electroporated slices compared to Glast+ control-FP (p<0.001); and compared to Glast+
BRAFwt electroporated slices (p<0.001); but not in Glast+ BRAFwt electroporated slices compared to
Glast+ control-FP electroporated slices for both astrocytes and neurons percentage (p=0.996). Astrocytes
percentage was decreased in Nestin+ BRAFV600E (n=20 slices, 3 brains) compared to Glast+ BRAFwt
(p<0.001); to Glast+ control-FP (p<0.001); to Glast+ BRAFV600E (p<0.001). Neuronal percentage was
increased in Nestin+ BRAFV600E compared to Glast+ BRAFwt (p<0.001); to Glast+ control-FP
(p<0.001); and to Glast+ BRAFV600E (p<0.001). Scale bars for A(upper panel) 1 mm;(lower panel) 500
µm. B.- 50 µm; * - p<0.05,** - p<0.01, *** - p<0.0001. Error bars are ±2SEM.
Figure 4. Unsupervised Hierarchical Clustering Analysis of GLAST+ BRAFV600E, GLAST+ control-FP
and GLAST+ BRAFwt tissue-wide expression profile. A. Four clusters of three conditions with four
replicates each 100, GLAST+ BRAFV600E, GLAST+ control-FP, GLAST+ BRAFwt, 742 genes with at
least log2 fold change =1 were used. B. Scatter plot for BRAFV600E to control-FP showing fold change
and p-values, data with p<0.01 is in red and was used for further functional enrichment analysis. Some of
the genes with the highest fold change are shown. Scatter plot for BRAFV600E to BRAFwt showing fold
change and p-values, data with p<0.01 is in red and was used for further functional enrichment analysis. .
Some of the genes with the highest fold change are shown. C. Expression vallues in transcripts per
million (TPM) of GFAP and vimentin in three conditions; GFAP is increased in BRAFV600E (20.78 fold
increase, p=0.000197, FDR=0.014) and BRAFwt (7.99 fold increase, p=0.0166, FDR=0.32) compared to
control-FP (Table S1 and S3); BRAFV600E to BRAFwt (2.67 fold increase p=0.03066, FDR=0.185;
Table S2); vimentin was also increased in BRAFV600E compared to control-FP (11.07 fold increase
p=0.00008, FDR=0.014) and to BRAFwt (3.63 fold increase p=0.004, FDR=0.131); BRAFwt to control-
FP (3.13 fold increase p=0.025, FDR=0.35); HLA-DR(CD74), a microglia marker was significantly
increased in BRAFV600E compared to control-FP (231.76 fold change p=0.004, FDR=0.03); and in
BRAFV600E compared to BRAFwt (154.39 fold change p=0.007, FDR= 0.136); C3 was significantly
increased in BRAFV600E neurons compared to control-FP (681.23 fold change p= 0.0002, FDR= 0.014);
BRAFV600E to BRAFwt (8.23 fold change p=0.015, FDR=0.15); BRAFwt to control-FP (84.93 fold
change p=0.028, FDR=0.356); C4b was significantly increased in BRAFV600E compared to control-FP
(37.75 fold change p=0.00057, FDR=0.016); H2-Aa, MHC II protein, was significantly increased in
BRAFV600E compared to control-FP (258.75 fold change p=0.0025, FDR=0.027); and in BRAFV600E
compared to BRAFwt (128.49 fold change p=0.011, FDR=0.14); H2-Ab1, also MHC II protein was
significantly increased in BRAFV600E compared to control-FP (134.29 fold change p=0.0031,
FDR=0.029); in BRAFV600E compared to BRAFwt (82.69 fold change p0.011, FDR=0.13); while
GAPDH was insignificantly changed, BRAFV600E to control-FP (1.10 fold decrease p=0.025,
FDR=0.175), to BRAFwt (1.03 fold decrease p=0.52, FDR=0.718); BRAFwt to control-FP (1.13 fold
decrease p=0.31, FDR=0.7). D. Represantative microglia in BRAFV600E immunoreactive to Iba1 mixed
with balloon-like cells. E. Venn diagram of all the upregulated genes with at least log2 fold change=1, and
p<0.01. F. Venn diagram of all the downregulated genes with at least log2 fold change=1, and p<0.01. G.
Fold enrichment from david functional annotation analysis of 402 overrepresented genes in BRAFV600E
compared to both control-FP and BRAFwt and 262 genes underrepresented in BRAFV600E compared to
both control-FP and BRAFwt. m – number of specific biological process associated genes out of 402
genes (n), M- total number of genes associated with specific biological process, N-total number of genes.
H. Gene analytics analysis of 402 overrepresented genes in BRAFV600E compared to both control-FP
and BRAFwt (upper panel). Same analysis of 500 random mouse genes. The scores are based on m/M
ratio, on significantly differentially 200 upregulated or 200 downregulated genes in disease tissues from
gene expression omnibus (GEO) database, or literature with at least 2-fold change and p<0.05, and
genetic association to the disease based on several MalaCards data sources (ClinVar, OMIM, Orphanet,
Uniprot, GeneTest), and the GeneCards-inferred relation to the disease with more frequently mentioned
genes having higher scores. For each gene, the maximal score of all the above mentioned possible scores
is used as the final gene score. The disease score is based on the final scores of all the matched genes.
Scale bar 50 µm in D.
Figure 5. BRAFV600E expressing neurons are hyperexcitable. A. Represantative traces of four GLAST+
experimental conditions (upper panel). Response to -40 pA 1 sec current step showing SAG ratio and
rebound excitation measurement (lower panel). B. Input-Output curve shows more than 2 times higher AP
frequency firing in GLAST+ BRAFV600E transfected neurons (n=54 T= range of 4.37-6.73, p<0.001)
and NESTIN+ BRAFV600E transfected neurons (n=8 T= range of 5.53-9.64, p<0.001) compared to all
other conditions, GLAST+ untransfected neighbor (n=8), GLAST+ control-FP (n=11), GLAST+
BRAFwt (n=13), NESTIN+ control-FP (n=3, 240-300 pA), and NESTIN+ untransfected neighbor (n=5,
130-300 pA). For statistical comparison for 10 pA - 150 pA steps, due to significant difference in
variances (Levene’s test (18, 149) = range of 8.816-1.68, p<0.001, p=0.013 for 140 pA and p=0.05 for
150 pA) Welch test with Games-Howell post-hoc correction was used; for 160-300 pA one-way ANOVA
F(18, 150) = 13.42-16.55 with Tukey post-hoc correction was used. Only cells with 7 and more APs at
300 pA 1 sec current step are chosen for the comparison. C. Instantaneous frequency (ISF) of APs in the
train at 300 pA 1 second depolarizing current step was significantly higher in GLAST+ BRAFV600E
transfected neurons compared to all other conditions (p<0.001). AP ISF at +300 pA 1 second step due to
nonhomogeneous variance for AP #1 and AP #2, Levene test (8, 132)=5.08 and 3.97 respectively, Welch
test (8, 22.91 and 23.18)=27.79 and 15.82 respectively with Games-Howell posthoc correction was used,
from AP #3 to AP #20 One-way ANOVA F(7-8 , 82-132 )=2.24-19.6, p<0.001, with Tukey posthoc
correction was used; there was significant difference in ISF between GLAST+ BRAFV600E neurons
(n=54) to NESTIN+ untransfected neighbor (n=5, p<0.001) and NESTIN+ control-FP (p<0.001);
NESTIN+ BRAFV600E (n=8) to their untransfected neighbor (p<0.001) and NESTIN+ control-FP
(p<0.001). D.Input-Output curve for GLAST+ BRAFV600E, GLAST+ BRAFwt held at -60 mV (n=4),
and GLAST+ untransfected neighbor neurons (n=5) held at -60 mV shows still significant difference in
AP firing frequency. Due to non-homogenous variance for 30 and 40 pA steps Levene (2, 53)=5.5 and
4.41, p=0.007 and p=0.017 Welch test with Games-Howell post-hoc correction was used; for steps 50-
300 pA One-way ANOVA with Tukey post-hoc correction was used. For 30-40 pA steps Welch
F(2,10.85) = 10.43, p=0.003 and (2, 10.22)=9.78, p=0.004 respectively. For 70-300 pA steps ANOVA
F(2, 54)=3.82 – 13.84, p=0.028 and p=0.017 for 70 and 80 pA steps, p=0.007-0.001 for 90-120 pA steps;
p<0.001 for 130-300 pA steps. E. One-way ANOVA with Tukey post-hoc correction of SAG ratio values
in different conditions shows larger SAG in most of the recorded GLAST+ BRAFV600E (n=58)
transfected neurons, F(3,105) = 35.98 (GLAST+ BRAFV600E to GLAST+ control-FP (n=15) – p<0.001;
untransfected neighbor neurons (n=18) – p<0.01 , GLAST+ BRAFwt (n=17) – p<0.001). SAG ratio was
significantly larger in GLAST+ BRAFV600E (n=58) compared to NESTIN+ control-FP (n=5 T=2.94,
p=0.005); to NESTIN+ untransfected neighbor (n=4 T=3.3, p=0.002); NESTIN+ BRAFV600E (n=8) to
NESTIN+ control-FP (n=5 T=4.48, p<0.001); NESTIN+ BRAFV600E to their untransfected neighbor
(T=4.87, p<0.001); NESTIN+ BRAFV600E to GLAST+ BRAFwt (T=5.53 p<0.001); NESTIN+
BRAFV600E to GLAST+ untransfected neighbor (T=7.3 p<0.001); NESTIN+ BRAFV600E to GLAST+
control-FP (T=5.8 p<0.001). F. Due to non-homogenous variances (Levene’s test (3, 91) = 7.61, p<0.001)
Welch test with Games-Howell post-hoc correction was used for statistical comparison of rebound
excitation values and shows larger rebound excitation in GLAST+ BRAFV600E (n=42) transfected
neurons compared to all non-BRAFV600E GLAST+ conditions Welch (3, 42.36) = 20.83, p<0.001;
GLAST+ control-FP (n=14, p<0.001), and GLAST+ untransfected neighbor neurons (n=18, p<0.001),
GLAST+ BRAFwt (n=17, p<0.001). Rebound excitation was larger in GLAST+ BRAFV600E neurons
(n=42) compared to NESTIN+ untransfected neighbor (n=4 T=3.3, p=0.010); to NESTIN+ control-FP
(n=5 T=2.94, p=0.013); it was increased in NESTIN+ BRAFV600E (n=8) compared to their
untransfected neighbor (n=4 T=5.514, p<0.001); to NESTIN+ control-FP (n=5 T=6.035,
p<0.001).BRAFV600E neurons with rebound APs were omitted from statistical comparison. G. Left
upper panel - representative whole-cell current-clamp recording traces of GLAST+ BRAFV600E
transfected neuron showing depolarization waves with smaller spikes riding on top of them following
each full size AP (4 out of 58 neurons, 6.9%), this bursting was not observed in control conditions or
BRAFwt condition. Left lower panel – zoomed in depolarization waves. Right panel – representative
trace of rebound AP observed in 12 out of 58 (20.7%) GLAST+ BRAFV600E transfected neurons. * -
p<0.05, ** - p<0.01, *** - p<0.001. Error bars are ±SEM for B, C, and D; ±2SEM for E-F.
Figure 6. Differential effect of three experimental manipulations on whole-cell current-clamp properties.
A. Representative traces from GLAST+ neurons expressing PIK3CA E545K (blue) and BRAFV600E
mutations (red), and CRISPR knockdown of TSC1 gene showing AP firing at +300 pA (upper panel), and
membrane potential response to hyperpolarizing current step of -40 pA. B. Average AP firing frequency
in all three conditions, BRAFV600E (n=54), TSC1 KD (n=11), PIK3CA E545K (n=6). Cells with
maximal values are shown for PIK3CA E545K (half-filled blue circles), and for TSC1 KD (brown
spheres). C. AP instantaneous frequency at +300 pA current step except TSC1 KD, which is shown for
+450 pA current step with maximal value cell shown for +300 pA current step (brown spheres). D.
Rheobase for all three conditions was compared with Welch test (2, 14.38) =72.98 due to
nonhomogeneous variance (Levene test (2, 68)=15.36, p<0.001), together with Games-Howell posthoc
correction BRAFV600E to TSC1 KD (p<0.001), BRAFV600E to PIK3CA E545K (p=0.011, due to small
sample for PIK3CA E545 student T=3.68, p=0.004 was used), . E. AP 50 V/s voltage threshold is not
different compared with Welch (2, 19.59) =2.48, p=0.11) together with Games-Howell BRAFV600E to
TSC1 KD (p=0.57), BRAFV600E to PIK3CA E545K (p=0.12); Levene test (2, 73) =4.93, p=0.01. F.
Resting Membrane Potential (RMP recorded before application of current steps) was compared with One-
way ANOVA F(2,86) =28.72, together with Tukey posthoc corrections BRAFV600E to TSC1 KD
(p<0.001), and BRAFV600E to PIK3CA E545K (p<0.001). G. Input resistance (Rin) from depolarizing
current steps (due to Ih activation in BRAFV600E) was compared with Welch (2, 25.70) =74.48, due to
nonhomogeneous variance (Levene test (2, 61) =3.40, p=0.04), together with Games-Howell posthoc
correction BRAFV600E to TSC1 KD (p<0.001), and BRAFV600E to PIK3CA E545K (p<0.001). * -
p<0.05, ** - p<0.01, *** - p<0.001.
Figure 7. Properites of first action potential at rheobase are altered in GLAST+ and NESTIN+
BRAFV600E expressing neurons A. Representative first APs at rheobase from GLAST+ BRAFV600E,
control-FP, untransfected-neighbor, BRAFwt all the cells had similar RMPs (BRAFV600E - -72.87 mV,
BRAFwt - -72.32 mV, control-FP - -72.48 mV, untransfected neighbor - -72.31 mV). B. First order
derivative over time (dV/dt) of representative APs (phase-space plot) from A. showing hyperpolarised AP
voltage threshold in BRAFV600E transfected neurons. C. One-way ANOVA F(6, 116)=9.72, p<0.001
with Tukey post-hoc correction showed that AP voltage threshold at 50 V/s was more hyperpolarized in
GLAST+ BRAFV600E transfected neurons (n=54) compared to GLAST+ control-FP (n=18, p<0.001),
and untransfected neighbor (n=16, p<0.001); GLAST+ BRAFwt (n=17, p=0.007) neurons; NESTIN+
control-FP (n=5, p<0.001); it was lower but not statistically significant compared to NESTIN+
BRAFV600E untransfected neighbor (n=4 T=1.81, p=0.076). AP voltage threshold at 50 V/s was
significantly more hyperpolarized in NESTIN+ BRAFV600E neurons (n=9) compared to NESTIN+
control-FP neurons (n=5 p<0.001); to NESTIN+ BRAFV600E untransfected neighbor (n=4 T=2.63,
p=0.023); to GLAST+ untransfected neighbor (p=0.001); to GLAST+ control-FP (p<0.001); to GLAST+
BRAFwt (p=0.012). D. One-way ANOVA F(6, 122) =6.29, p < 0.001 with Tukey post-hoc correction
showed that AP peak measured from RMP was larger in GLAST+ untransfected neighbor neurons (n=16)
compared to GLAST+ BRAFV600E (n=58, p=0.001) and to GLAST+ BRAFwt transfected neurons
(n=18, p=0.04); to NESTIN+ BRAFV600E (n=9, p<0.001). GLAST+ control-FP (n=18) to NESTIN+
BRAFV600E (p=0.039). NESTIN+ BRAFV600E to NESTIN+ untransfected neighbor (n=4, p=0.033); to
NESTIN+ control-FP (n=5, p=0.002) . E. Due to non-homogenous variance – Levene test (6, 114) =5.55,
p<0.001 Welch test with Games-Howell post-hoc correction was used to compare rheobase between
experimental conditions, which showed lower current required to fire AP in GLAST+ BRAFV600E
transfected neurons (n=49) Welch test (6, 20.94) =36.71, p<0.001, compared to control-FP (n=19,
p<0.001), and untransfected neighbor neurons (n=17, p<0.001), to GLAST+ BRAFwt (n=18, p<0.001).
Rheobase was significantly lower in GLAST+ BRAFV600E (n=49) compared to NESTIN+ control-FP
(n=5 p<0.001); to NESTIN+ untransfected neighbor (n=4 T=6.46, p<0.001); NESTIN+ BRAFV600E
neurons (n=9) to NESTIN+ control-FP (n=5 T=11.52, p<0.001); to their untransfected neighbor (n=4
T=6.66, p<0.001); to GLAST+ untransfected neighbor (p<0.001); to GLAST+ control-FP (p<0.001); to
GLAST+ BRAFwt (p=0.004). F. GLAST+ BRAFV600E transfected neurons (n=61) had more
depolarized resting membrane potential compared to GLAST+ control-FP neurons (n=20, p<0.001), and
to their untransfected neighbor neurons (n=23, p<0.001); to NESTIN+ untransfected neighbor (n=5,
p=0.001); to NESTIN+ control-FP (n=5, p=0.031). GLAST+ BRAFwt transfected neurons (n=18) had
more depolarized RMP compared to untransfected neighbor neurons (p=0.044); to GLAST+ control-FP
(p=0.050); to NESTIN+ untransfected neighbor (p=0.034); to NESTIN+ control-FP (T=2.25 p=0.036).
NESTIN+ BRAFV600E to GLAST+ control-FP (p=0.033), and to their untransfected neighbor
(p=0.005). One-way ANOVA F(6, 133) = 9.59, p<0.001 with Tukey post-hoc correction test was used for
statistical comparison. In case of small n student t-test was used. RMP measured in current clamp before
beginning of steps protocol. G. Due to nonhomogeneous variance (Levene test (6, 112) =5.58, p<0.001)
Welch test (6, 24.56) =7.04, p<0.001 with Games-Howell post-hoc correction was used for input
resistance comparison. Averaged input resistance (Rin) as a function of membrane potential response
(Vm) to depolarizing current (I) steps was significantly larger in GLAST+ BRAFV600E transfected
neurons (n=39) compared to GLAST+ control-FP (n=21, p<0.001), and their untransfected neighbor
neurons (n=23, p=0.009); to GLAST+ BRAFwt (n=18, p=0.001). GLAST+ BRAFV600E to NESTIN+
untransfected neighbor (n=5, p=0.018); to NESTIN+ control-FP (n=5, p=0.003). Average Rin from
depolarizing current steps was significantly larger in NESTIN+ BRAFV600E (n=8) compared to
GLAST+ control-FP (n=21 T=2.496, p=0.040); to NESTIN+ untransfected neighbor (n=5 T=2.418,
p=0.043); to NESTIN+ control-FP (n=5 T=2.512, p=0.038). H. Linear fit of averaged membrane
potential responses to current step protocol from -40 to +50 pA 1 second pulse with 10 pA increment
shows a different input resistance in between the recorded conditions. * - p<0.05, ** - p<0.01, *** -
p<0.001. Error bars are ±SEM for A. and H, ±2SEM for C-G.
Figure 8. Sustained potassium currents are decreased in GLAST+ BRAFV600E neurons compared
to their untransfected neighbors. A. Representative traces of potassium currents in GLAST+
BRAFV600E neuron recorded in the presence of 3 mM 4AP, 1µM TTX, 10µM NBQX, 50µM D-AP5,
10µM SR, 50µM ZD7288, and 1 mM Co2+ substitution for Ca2+ (5 min in, holding voltage is -80 mV,
holding current -32.41 pA) with whole-cell capacitance compensated, grey bar indicate the region where
the measurement was made in all conditions (upper panel); middle panel is showing the traces of the same
neuron 9 min after application of 25 mM TEA with previous inhibitors cocktail (holding current -
40.25 pA); lower panel is showing subtracted traces before and after 25 mM TEA with voltage step
protocol. B. Representative traces of potassium currents in GLAST+ untransfected neighbor recorded in
the presence of the same inhibitors cocktail as for A (6 min in, holding voltage is -80 mV, holding current
is -48.04 pA, upper panel) with whole-cell capacitance compensated; middle panel is showing traces from
the same neuron 9 min after application of 25 mM TEA with previous inhibitors cocktail (holding current
-71.67 pA); lower panel is showing subtracted traces before and after 25 mM TEA. C. Average sustained
current activation curve showing decreased peaks at all tested voltages in GLAST+ BRAFV600E neurons
compared to their untransfected neighbor before and after application of 25 mM TEA. D. Normalized to
the maximum current and averaged sustained current activation curve have similar kinetics between two
conditions before and after application of 25 mM TEA. E. Maximal sustained current measured at +20
mV voltage step showing lower values in GLAST+ BRAFV600E neurons (n=18) compared to their
untransfected neighbors (n=8, T=2.92, p=0.008), as well as after application of 25 mM TEA (T=2.411,
p=0.028). It was also decreased in the same neurons when comparing before and after 25 mM TEA –
GLAST+ BRAFV600E (paired sample T=2.474, p=0.035); and their untransfected neighbor (paired
sample T=3.827, p=0.006); and comparing untransfected neighbor neurons to GLAST+ BRAFV600E
after application of 25 mM TEA (T=6.919, p<0.001). F. Current density was not different in
untransfected neighbors’ comparison. G. Capacitance was measured from -5 mV steps at the beginning of
each trace recording using built-in procedure in Axograph acquisition software before application of
whole-cell capacitance compensation. There was no statistically significant difference in capacitance
measurements compared to untransfected neighbors condition. J. Input resistance was significantly
increased in GLAST+ BRAFV600E (n=18) compared to their untransfected neighbors (n=8, T=3.293,
p=0.003), it was increased in GLAST+ BRAFV600E neurons (n=7) compared to their untransfected
neighbors (n=4), but not statistically significant (T=2.223, p=0.053); it was significantly increased in
GLAST+ BRAFV600E neurons (n=7) after application of 25 mM TEA compared to their untransfected
neighbors before application of TEA (n=8, T=3.664, p=0.009). * - p<0.05, ** - p<0.01, *** - p<0.001.
Error bars are ±SEM for C. and D, ±2SEM for E-H.
Figure 9. Hyperpolarization activated depolarizing current (Ih) recorded in whole-cell voltage-clamp
configuration is increased in BRAFV600E expressing cortical neurons of layers 2/3. A., B., C., E.
Representative traces of currents in response to hyperpolarizing voltage step protocols shown in the lower
panels. Note that GLAST+ BRAFwt was recorded at the same holding potential as all other conditions
with the first hyperpolarizing voltage steps been -100 mV and not -120mV. D. Ih activation curve from
the voltage steps protocol shown in C. (tail currents, dashed circle) with maximal activation around -120
mV half activation -82.79 mV and k slope factor of 11.58-1 mV, which are averaged and fit with
Boltzmann curve (n=18). F., G., H. Application of 50 µM ZD7288, a known Ih inhibitor in perfusion
system for 5 minutes blocked Ih. I. Ih peak current measured as shown in C., recorded with protocol
shown in A. lower panel. The significant increase was only found in GLAST+ BRAFV600E neurons
(n=17) compared to their untransfected neighbors (n=6, T=2.117, p=0.046); J. right panel – Ih peak
density was increased in GLAST+ BRAFV600E neurons (n=16) compared to their untransfected
neighbor (n=6, T=3.918, p<0.001), to NESTIN+ control-FP (n=4, T=5.546, p<0.001); it was also
significantly increased in NESTIN+ BRAFV600E neurons (n=4) compared to GLAST+ untransfected
neighbors (n=6, T=3.275, p=0.011), and to NESTIN+ control-FP (n=4, T=3.066, p=0.022). Error bars are
±SEM and for H and I; ±2SEM for J.
Figure 10. GLAST+ and NESTIN+ BRAF V600E expressing neurons segregate to separate clusters in
HCA analysis of electrophysiological properties. A. Unsupervised Hierarchical Cluster Analysis was
performed on 20 recorded electrophysiological parameters and showing that most of the BRAFV600E
neurons segregate together by electrophysiological parameters recorded. The parameters are AP width at
50% height from RMP in ms, AP maximal decay slope in V/s, AP decay time from 100% to 50% height
in ms, AP rise time from 10% to 90% height in ms, AP 50 V/s voltage threshold in mV, AP 10 V/s
voltage threshold in mV, rheobase in pA, AHP measured at the end of +300 pA 1 second current step in
mV, AP peak relative to RMP in mV, AP maximal rise slope in V/s, RMP in mV, Rin from
hyperpolarizing pulses in MΩ, Rin from depolarizing pulses in MΩ, average sPSCs amplitude in pA,
average sPSCs instantaneous frequency in Hz, mAHP measured relative to 10 V/s AP voltage threshold
for rheobase APs in mV, SAG ratio in %, AP frequency at +300 pA 1 second current step in Hz, number
of APs at rheobase, rebound excitation measured as an overshoot above RMP (mV). B. Most
contributing electrophysiological parameters to the variability in PCA shown in 3D plots – upper left
panel SAG ratio on the Z axis, AP number at +300 pA 1 second pulse is on the X axis and rebound
excitation is on the Y axis. C. SAG ratio on the Z axis, AHP at the end of +300 pA 1 second pulse on the
X axis and rebound excitation on the Y axis. D. Rheobase on the Z axis, AP maximal rise slope is on the
X axis, and AP 50V/s voltage threshold on the Y axis. E. AP number at +300 pA 1 second pulse on the Z
axis, resting membrane potential (RMP) is on the X axis, Input resistance from depolarizing current
pulses (Rin) is on the Y axis.
Figure 1:
Figure 2:
Figure 3:
Figure 4:
Figure 5:
Figure 6:
Figure 7:
Figure 8:
Figure 9:
Figure 10:
Tables:
Table 1. GOTERM Biological protein production pathways enrichment in GLAST+ BRAFV600E
compared to control-FP and to GLAST+ BRAFwt from 402 upregulated genes at p<0.01
Category GOTERM_BP_DIRECT
Count
%
Fold
Enrichment
Benjamini p-
value corrected
collagen fibril organization
11
2.770781
14.16681
0.000007152
response to hypoxia
19
4.785894
4.970457
0.000062725
immune system process
24
6.04534
3.147433
0.002011456
response to mechanical stimulus
10
2.518892
7.972663
0.002226683
positive regulation of angiogenesis
13
3.274559
5.396373
0.002342004
wound healing
10
2.518892
5.343381
0.038195949
inflammatory response
19
4.785894
2.774209
0.058961929
Table 2. GOTERM Biological protein production pathways enrichment in GLAST+ BRAFV600E
compared to GLAST+ control-FP and to GLAST+ BRAFwt from 262 downregulated genes at p<0.01.
Category GOTERM_BP_DIRECT
Count
%
Fold
Enrichment
Benjamini p-
value corrected
neuron projection development
10
3.802281
5.807425
0.023973
cytoskeleton organization
9
3.422053
6.985663
0.027481
protein phosphorylation
20
7.604563
2.802889
0.031236
potassium ion transport
10
3.802281
6.356159
0.035131
ephrin receptor signaling pathway
6
2.281369
11.2637
0.044375
phosphorylation
20
7.604563
2.638014
0.044999
Table 3. Sustained K+ current average of maximal values and current density.
Condition
IK-Sustained maximal
(pA) last 100 ms of 500
ms +20 mV pulse
IK-Sustained
maximal
density (pA/pF)
Cm (pF)
n
GLAST+ untransfected neighbor
4062.66 ± 210.81
34.94 ± 11.76
178.04 ± 31.74
8
GLAST+ BRAFV600E
3034.19 ± 282.49**
21.30 ± 2.53
157.76 ± 13.29
18
** - t(23.5)=2.92, P=0.008
Table 4. Sustained K+ current average of maximal values and current density – TEA sensitive
Condition
IK-Sustained peak (pA) last
100 ms of 500 ms
+20 mV pulse TEA
sensitive
n
IK-Sustained density TEA
sensitive (pA/pF)
Cm (pF)
n
GLAST+ untransfected
neighbor
2484.56 ± 258.29
8
12.23 ± 3.85
239.88 ± 59.23
4
GLAST+ BRAFV600E
1561.06 ± 266.57*
11
7.28 ± 1.36
187.18 ± 18.45
7
* - T(17)=2.41, P=0.028
Table 5. Ih peak and current density.
Condition
Ih peak -50 -120
mV (pA)
Ih peak density
-50 -120 mV
(pA/pF)
Cm (pF)
n
GLAST+ control-FP
-129.22
-1.02
126.94
1
GLAST+ untransfected
neighbor
-174.17 ± 58.24
-0.92 ± 0.30
193.03 ± 24.99
6
GLAST+ BRAFwt
-66.22 ± 13.03*
-0.36 ± 0.08
253.49 ± 41.27
12
GLAST+ BRAFV600E
- 446.58± 58.24
-3.45 ± 0.37
135.95 ± 14.18
18
NESTIN+ control-FP
-197.17 ± 93.19
-0.85 ± 0.28
214.29 ± 44.06
4
NESTIN+ untransfected
neighbor
-117.75
-1.13
104.00
1
NESTIN+ BRAFV600E
-322.76 ± 54.45
-2.97 ± 0.63
84.34 ± 32.33
4
| 2019 | BRAFV600E Expression in Mouse Neuroglial Progenitors Increase Neuronal Excitability, Cause Appearance of Balloon-like cells, Neuronal Mislocalization, and Inflammatory Immune response | 10.1101/544973 | [
"Goz Roman U.",
"Silas Ari",
"Buzel Sara",
"LoTurco Joseph J."
] | creative-commons |
Magnitude and Correlates of Caesarean Section
in Urban and Rural Areas:
A Multivariate Study in Vietnam
Myriam de Loenzien1*,
ORCID 0000-0001-7121-0185
Clémence Schantz1¶,
Bich Ngoc Luu2¶,
Alexandre Dumont1¶
1 Centre Population et Développement UMR 196, Institut de Recherche pour le
Développement, Université Paris Descartes, INSERM, France
2 Institute for Population and Social Studies, National Economic University, Hanoi, Vietnam
* Corresponding author
e-mail: Myriam.de-Loenzien@ird.fr
¶ These authors contributed equally to this work.
Abstract
Caesarean section can prevent maternal and neonatal mortality and morbidity. However,
it involves risks and high costs which can be a burden, especially in low and middle income
countries. The international healthcare community considers the optimal caesarean rate to be
between 10% and 15%. The aim of this study is to assess its magnitude and correlates among
women of reproductive age in urban and rural areas in Vietnam. We analyzed microdata from
the national Multiple Indicator Cluster Survey (MICS) conducted in 2013-2014 using
representative sample of households at the national level as well as regarding the urban and
the rural areas. A total of 1,378 women who delivered in institutional settings in the two years
preceding the survey were included. Frequency and percentage distributions of the variables
were performed. Bivariate and multivariate logistic regression analysis were undertaken to
identify the factors associated with caesarean section. Odds ratios with 95% confidence
interval were used to ascertain the direction and strength of the associations. The overall CS
rate among the women who delivered in healthcare facilities in Vietnam is particularly high
(29.2%) with regards to WHO standards. After controlling for significant characteristics,
living in urban areas more than doubles the likelihood of undergoing a CS (OR = 2.31; 95%
CI 1.79 to 2.98). Maternal age at delivery over 35 is a major positive correlate of CS. Beyond
this common phenomenon, distinct lines of socioeconomic and demographic cleavage operate
in urban versus rural areas. The differences regarding correlates of CS according to the place
of residence suggest that specific measures should be taken in each setting to allow women to
access childbirth services appropriate to their needs. Further research is needed on this topic.
Keywords
Delivery, caesarean section, childbirth, urban, rural, correlates, Vietnam
1
Introduction
Caesarean section can prevent maternal and neonatal mortality and morbidity. However,
it involves risks and high costs which can be a burden, especially in low and middle income
countries. The international healthcare community considers the optimal caesarean rate to be
between 10% and 15% (1). Urbanization, which is related not only to a population moving
from a rural to an urban area and an increased concentration of people living in urban areas
but also to the whole process of societal adaptation to subsequent changes, has been identified
as a prominent contributing factor to caesarean section (CS) practices in several countries and
areas (2)(3)(4)(5)(6)(7)(8). However, this influence is controversial (9)(10)(11).
Vietnam, which transformed from a low to a middle income country in the last decade,
has witnessed increasing CS rates concomitantly with urbanization. In this country, the
proportion of women undergoing CS increased from 3.4% in 1997 (12) to 27.5% in 2014
(13), which largely exceeded the levels recommended by the World Health Organization
(WHO) (10 to 15%) (14). This percentage is among the highest in the region (14)(15), and
this trend shows no sign of abatement. The increase is occurring in a context of rapid
socioeconomic and demographic changes. During the same period, the proportion of people
living in urban areas rapidly increased from 23.7% in 1999 to 29.6% in 2009 (16).
We propose to measure the influence of living in urban versus rural areas on childbirth
practices and to explore the possible pathways of the influence of the place of residence on
CS in Vietnam. Using microdata from a nationally representative sample, we provide the
sociodemographic profile associated with high CS rates. Subsequently, we present correlates
of CS rates by making a distinction between women who live in rural and urban areas. Our
main argument is that beyond the apparent overall convergence, CS practices diverge not only
in magnitude between rural and urban areas but also regarding their dynamic.
2
Literature review
Relationship between CS and place of residence
For several decades, living in urban areas in low- and middle-income countries in Asia,
Africa and Latin America has been associated with higher CS rates after controlling for
multiple socioeconomic, biomedical and institutional factors (2)(3)(4)(5)(6)(7)(8). However,
this relationship appears to be nonsignificant in various settings (9)(10). Some studies have
even shown a reverse trend. In Hawaii, despite a lower risk of delivery by CS, women who
deliver in rural hospitals have higher rates of primary CS than do women who deliver in urban
hospitals, even after controlling for maternal risk factors (11).
Further analyses taking into account the level of urbanization complement these results.
In Taiwan, CS rates increased with an advancing urbanization level (17). Similarly, a study
using data from 29 countries in Asia, Africa and Latin America showed higher CS rates in
urban areas than in periurban areas (18). Conversely, a study in Cambodia showed that CS
rates were lower for women living in Phnom Penh than for women living in its surrounding
area (19).
Some studies go more in depth by conducting intersectional analyses between the place
of residence and the wealth effect. One analysis that adopted data from demographic and
health surveys performed in low- and middle-income countries in Africa, Asia and Latin
America showed that the CS rates in most countries were higher in urban areas than in rural
richer households, which represented half of the rural population. In turn, most rural richer
households had higher CS rates than did rural poorer households (20).
More refined indicators of childbirth practices have also been used. In more-developed
countries, research indicates a higher level of non-medically indicated labor induction in
urban areas but a more rapid rise in rural areas, such as the trends observed in the United
States (21). In states in Burkina Faso, CS deliveries for nonabsolute medical indications were
3
more frequent among women living in urban areas even after controlling for other factors
(22).
In this study, we consider that the decision to undergo caesarean section results from a
negotiation between the caregiver and the patient, which is determined by proximate
determinants. Among them, patient’s and health caregivers’ perceptions play major roles, as
well as the availability and accessibility of healthcare facilities, equipment, personnel and
technologies (23). These proximate determinants are in turn determined by distal
determinants, such as biomedical factors, but also social, cultural and political characteristics
at individual, interindividual and collective levels. These characteristics include women’s
human, economic and social capital but also cultural beliefs, values and norms regarding
family and gender relations (24), interactions between social groups (9), the media and formal
institutions, welfare state and national policies as well as economic conditions (3)(25). Due to
contrasted modes of socialization and levels of equipment, we expect underlying processes
related to these phenomena to differ between rural and urban areas.
CS and urbanization in Vietnam
In Vietnam, urbanization has rapidly developed and continues to exhibit a rising trend in
the context of economic growth and on-going demographic transition. Urbanization
accelerated in the 1990s (26) following the reforms in the mid-1980s from a centralized
system to a market-oriented economy under state guidance (27). The country has shifted its
policy from the promotion of intermediate-level cities in the 1990s-2000s to more investment
in great metropolitan areas aimed at acting as drivers of the economy (28). The urbanization
rate increased from 19.2% in the 1980s to 29.6% in 2009 and reached 34.0% in 2015
(16)(29). Simultaneously, rural-urban inequalities have decreased (30). This reduction has
mainly been due to migration in a context of improving economic conditions, the
4
development of industrialization, increasing international integration and profound
demographic and technological changes (31).
Overall, the proportion of women who deliver by CS has multiplied by almost 7 within a
17-year period. From a very low level in 1997 (3.4%), this proportion reached the level
proven to be the threshold of the absence of efficiency of CS (10%) in 2002 (9.9%) (12) (32)
(14). This rate reached almost three times this value in 2013-14 (27.5%) (33)(13).
CS rates have increased at a higher pace in rural areas, where they have multiplied by 9
(from 2.3% to 21.0%), than they have in urban areas, where they have multiplied by 3 (from
13.6% to 43.3%). Consequently, the urban-rural ratio of the proportion of women who
underwent CS dropped (from 5.9 in 1997 to 2.1 in 2014). This increase in CS rates occurred
in the context of a marked development of childbirth biomedicalization fostered by recent
investment in district hospitals by the government (34).
Whereas only a minority of pregnancies were followed up by a doctor in 1997 (28.2%),
almost all pregnancies were followed up in 2014 (90.3%). The number of prenatal care visits
has dramatically increased. Very few women attended 7 visits or more in 1997 (2.3%), and a
higher proportion of women utilized this number of visits 17 years later (39.0%). As a result,
neonatal mortality rates dropped (from 20. to 11.4 deaths per 1000 live births), as did
maternal mortality rates (from 100 to 54 maternal deaths per 100,000 live births) (34).
Vietnam ended its demographic transition, with fertility reaching the replacement level since
2005 (35).
Studies in this country suggest that urban areas are linked to a higher level of CSs.
However, these studies referred to a period when CS rates were still low (9) or to specific
geographical areas (36). Therefore, there is a need to update general trends in this country and
to better understand the correlates of such difference between rural and urban areas.
5
Materials and Methods
Data
We used data from the 2013-2014 Multiple Indicator Clusters Survey (MICS). Urban and
rural areas within each region were identified as the main sampling strata. The sample was
selected in two stages: census enumeration areas were selected within each stratum, and
households were selected within each enumeration area (13). This dataset provides
statistically representative samples of women aged 15-49 years at the national and regional
levels and for each type of setting (rural versus urban). Among these datasets, we focus on
women who had a singleton birth at least once during the 2 years prior to the survey. This
population represents 1,453 women. Among these women, we take into account those who
delivered at an institutional setting, accounting for 94.4% of the population. Only the last birth
of each woman was considered.
Outcome measures and covariates
The outcome variable for this study is the woman’s mode of delivery, either through CS
or vaginal delivery.
The main covariate for this study is the place of residence (urban versus rural area). The
influence of this variable is explored by controlling for other socioeconomic and demographic
correlates. We take into account the place of delivery (public versus private health sector).
Based on empirical observations of the distribution, the number of antenatal care visits was
distinguished between 6 visits or fewer versus 7 or more to maximize the rural-urban
difference. We also took into account the birth weight of the newborn as perceived by the
mother (less than 2.8 kg, 2.8 to 3.5 kg, 3.6 kg and over), the maternal age at delivery (15 to 19
years, 20 to 34 years or 35 years and over) and the woman’s past experience of childbirth
(primiparous versus multiparous). Due to the preference for sons associated with CS in
6
Vietnam, we also explored the influence of the sex of the newborn (37). Additional
sociodemographic and cultural characteristics included the level of education of the women
(primary or less, secondary or tertiary), the region (North Central, Mekong River delta, Red
River delta, Northern Midlands, Central Highlands or the Southeast), the quintile of wealth of
the household (poorest, poor, middle, rich or richest), and the ethnicity of the household head
(Kinh ethnic group versus minority ethnic groups).
Analysis
We conducted bivariate analysis and stepwise logistic regression to assess the
characteristics associated with CS practice as opposed to vaginal delivery. Multivariate
logistic regression models allowed for comparisons between models for all women and for
only women living in rural or urban areas. For each of these 3 groups, two models were
tested. The first model (restricted model) included only sociodemographic variables, whereas
the complete model took into account all the available variables, including sociocultural
characteristics. Bivariate analyses used the women’s sample weights. The multivariate results
included all variables that reached the minimum level of significance in the bivariate analysis.
These results are controlled for the cluster effect. For each model and for the chi-square tests,
we draw on two levels of risk (p < 0.05 and p < 0.10). All statistical analyses were performed
with IBM PASW Statistics 18 software at Paris Descartes University.
Compliance with ethical standards
The Vietnam General Statistics Office (GSO) and the United Nations Children’s Fund
(UNICEF) approved the tools of the Vietnam Multiple Indicator Cluster Survey (MICS)
before the survey was conducted, in accordance with the ethical standards laid down in the
1964 Declaration of Helsinki and its later amendments or comparable ethical standards.
Participation was voluntary, and informed consent was obtained from all the individual
participants included in the study. The MICS data are freely available through the UNICEF
7
MICS website, and there is no need to obtain ethical approval before using the data. To access
data from the MICS website, a written request was submitted to UNICEF, and permission was
granted.
Results
Table 1 provides an overview of the social and demographic profiles of the women who
delivered in healthcare facilities and the corresponding rates of CS.
Table 1 Social and demographic profiles of the women who had a singleton birth in the
two years preceding the survey in healthcare facilities and corresponding rates of
caesarean section (CS) (n = 1350 deliveries)
Distribution
CS rates
Urban
Rural
Urban
Rural
%
%
p
%
p
%
p
Place of delivery
Public health sector
92.5
97.2
**
42.4
*
22.9
Private health sector
7.5
2.8
61.3
11.5
Antenatal care visits
6 or fewer
39.8
67.0
**
29.7
**
22.2
7 or more
60.2
33.0
53.2
23.6
Weight of the newborn
Less than 2.8 kg
12.0
16.8
**
40.0
23.6
2.8 to 3.5 kg
70.4
72.9
41.8
21.4
3.6 kg and over
17.6
10.3
54.8
29.9
Perceived size of
Average
76.1
79.8
**
39.9
**
22.7
the newborn
Smaller than average
8.2
9.7
51.5
19.8
Larger than average
15.7
10.5
58.5
24.7
Maternal age at delivery
15-19
5.3
8.9
**
22.7
**
17.1
**
20-34
84.6
83.1
43.0
22.1
35-49
10.1
8.0
61.9
34.2
8
Distribution
CS rates
Urban
Rural
Urban
Rural
%
%
p
%
p
%
p
Parity
Multiparous
57.1
56.8
42.4
20.5
*
Primiparous
42.9
43.2
45.5
25.2
Women’s education
Primary or less
9.9
14.3
**
31.7
**
23.1
Secondary
51.1
66.6
38.7
21.5
Tertiary
39.0
19.0
53.7
26.8
Wealth quintile
Poorest
3.4
20.6
**
21.4
**
20.8
Poor
8.9
25.7
36.1
22.9
Middle
14.7
24.2
31.1
23.9
Rich
25.1
21.9
34.6
23.3
Richest
48.0
7.6
55.3
21.1
Ethnicity
Kinh, Hoa
93.5
85.8
**
45.1
*
23.6
Ethnic minorities
6.5
14.2
25.9
17.3
Region
North Central
20.2
22.0
**
51.2
26.2
**
Mekong R. Delta
12.0
19.6
52.0
21.3
Red River Delta
23.1
25.5
44.8
17.6
Northern Midlands
9.2
14.3
33.3
26.9
Central Highlands
5.5
6.9
26.1
15.4
Southeast
29.9
11.6
42.7
28.4
Number of women
415
935
415
935
**: p ≤ 0.05, *: p ≤ 0.10
Source: 2013-14 MICS
Overall, almost one-third of the women live in an urban area (30.7%). Several correlates
linked to higher levels of CS are more prevalent in the urban areas than in the rural areas.
First, among the women who deliver in institutional settings, almost two-thirds of those living
in urban areas have more than 7 antenatal care visits, whereas this is the case for only one-
9
third of those living in rural areas. Second, several indicators show more favorable
socioeconomic situations for women living in urban areas: a much larger proportion of
women reach a tertiary level of education in urban (39.0%) than in rural areas (19.0%), the
proportion of women in the richest household quintile reaches a much higher level in urban
areas (48.0%) than in rural areas (7.6%), and lower proportions of women belonging to
minority ethnic groups are observed in urban (6.5%) than in rural areas (14.2%). Third, the
maternal age at delivery is lower in rural areas than in urban areas. The proportion of women
who deliver after 35 is higher in urban (10.1%) than in rural areas (8.0%) and conversely,
fewer women deliver between 15 and 19 in urban (5.3%) than in rural areas (8.9%).
The overall CS rate among the women who delivered in healthcare facilities is
particularly high (29.2%) with regards to WHO standards (14). The CS rate is almost twice as
high in urban (42.4%) than in rural areas (22.9%). The results regarding CS rates confirm that
the urban context is particularly favorable to CS. First, in urban areas, CS rates were almost
doubled among women who had at least 7 antenatal care visits compared to those among
women with 6 visits or fewer. Second, in urban areas, women who have a higher level of
education have higher CS rates, those who live in the richest households also have higher CS
rates, and to a lesser extent those who belong to the Kinh ethnic group have higher CS rates
than those belonging to the minority ethnic groups. Third, a higher maternal age at delivery is
associated with higher CS rates in both rural and urban areas.
A higher number of antenatal care visits, higher levels of education, wealth, and
concentration of Kinh ethnic groups and higher maternal age at delivery in urban areas
combined with higher CS rates among people of these groups help to understand part of the
urban–rural gap regarding CS rates. However, more in-depth analysis is needed to document
the relative influence of each of these correlates on the urban–rural difference in CS rates,
which will be achieved using multivariate analysis.
10
The results of the analysis of correlates of CS for the whole population are displayed in
Table 2. We will first examine a model taking into account only demographic and medical
variables. We will subsequently examine a model that also includes the socioeconomic
variables. The results show that after controlling for significant characteristics, living in urban
areas more than doubles the likelihood of undergoing a CS (OR = 2.31; 95% CI 1.79 to 2.98,
see restricted model). The completed model shows that the influence of the place of residence
on CS weakens when ethnicity is taken into account. This is partly due to the higher
concentration of population belonging to minority ethnic groups in rural areas. Following the
same trend, the weakening of the positive influence of having 7 antenatal care visits or more
suggests that the higher level of medicalization in urban areas mostly regards women from the
Kinh ethnic group. In addition to the place of residence, the number of antenatal care visits
and ethnicity, delivering at 35 years or over remains strongly linked to CS, as is also the case
of maternal perception of the newborn weight as above average and delivery for the first time.
To better understand the contrasts between urban and rural dynamics as well as correlates
of CS, we will separately study women from each place of residence. The results are
displayed in Table 3. They show two models: one concerning urban areas, and the other one
concerning rural areas. In both models, maternal age at delivery over 35 is a major positive
correlate of CS. Beyond this common phenomenon, distinct lines of socioeconomic and
demographic cleavage operate in urban versus rural areas.
In urban areas, women are more than twice more likely to undergo CS when they have at
least 7 antenatal care visits. They are also more than twice more likely to have CS when they
deliver in the private health sector. Perception of one’s baby’s weight as above average also
has a strong effect. When socioeconomic characteristics are taken into account, the distinction
between private and public health sector disappears, whereas living in the Central Highlands
11
or in the Southeast region appears linked to lower levels of CS. This suggests that the
influence of private or public health sector is partly explained by contrasts between regions.
Table 2 Multivariate analysis of the factors associated with caesarean delivery (n =
1350)a
Restricted model
Complete model
OR
95% CI
OR
95% CI
Place of residence (reference = rural)
Urban
2.31**
1.79
2.98
1.99**
1.48
2.67
Place of delivery (reference = public sector)
Private sector
1.35
0.82
2.21
1.27
0.76
2.13
Antenatal care visits (reference = 6 or fewer)
7 visits or more
1.38**
1.06
1.79
1.32
0.98
1.77
Weight of newborn (reference = 2.8 to 3.5 kg)
Less than 2.8 kg
1.05
0.72
1.52
1.08
0.74
1.59
3.6 kg and over
1.60**
1.10
2.32
1.55**
1.06
2.26
Maternal age at delivery (reference = 20-34)
15-19
0.55
0.31
0.98
0.66
0.36
1.19
35-49
2.13**
1.42
3.19
2.20**
1.46
3.33
Parity (reference = multiparous)
Primiparous
1.46**
1.10
1.94
1.41**
1.05
1.89
Women’s education (reference = Primary or less)
Secondary
0.94
0.61
1.47
Tertiary
1.23
0.73
2.08
12
Restricted model
Complete model
OR
95% CI
OR
95% CI
Household wealth quintile (reference = Poorest)
Poor
1.01
0.61
1.69
Middle
0.91
0.53
1.55
Rich
0.85
0.48
1.48
Richest
1.32
0.71
2.47
Ethnicity (reference = Kinh/Hoa)
Ethnic minorities
0.61**
0.40
0.93
Region (reference = North Central)
Mekong River Delta
0.82
0.54
1.25
Red River Delta
0.61**
0.40
0.94
Northern Midlands
1.10
0.73
1.64
Central Highlands
0.54**
0.35
0.84
Southeast
0.75
0.49
1.15
a**: p ≤ 0.05, *: p ≤ 0.10
(Source: 2013-14 MICS)
In rural areas, parity has a significant effect in the restricted model. Primiparous women
are twice more likely to undergo caesarean section than multiparous women. This effect does
not remain when sociocultural factors are taken into account. In the complete model, the odds
of undergoing CS are almost halved for women belonging to minority ethnic groups. This
suggests that the higher level of CS among primiparous women may be partly explained by
the fact that they belong to the Kinh ethnic group, where fertility reaches lower levels.
13
Table 3 Odds of undergoing caesarean section (CS) for all women according to their place of residence: urban (n = 415) and rural (n
= 935) areasa
Urban
Rural
Restricted model
Complete model
Restricted model
Complete model
OR
95% CI
OR
95% CI
OR
95% CI
OR
95% CI
Place of delivery (reference = public sector)
Private sector
2.29**
1.19
4.41
1.88
0.92
3.87
0.51
0.15
1.72
0.48
0.14
1.63
Antenatal care visits (reference = 6 or fewer)
7 visits or more
2.40**
1.62
3.54
2.26**
1.44
3.54
0.97
0.66
1.41
0.96
0.63
1.46
Weight of newborn (reference = 2.8 to 3.5 kg)
Less than 2.8 kg
1.02
0.55
1.92
2.04**
1.02
4.10
1.10
0.69
1.74
1.10
0.69
1.75
3.6 kg and over
1.75**
1.08
2.85
2.90**
1.69
4.99
1.61
0.92
2.81
1.57
0.89
2.78
Maternal age at delivery (reference = 20-34)
15-19
0.46
0.17
1.20
0.59
0.21
1.67
0.59
0.30
1.16
0.67
0.33
1.34
35-49
2.44**
1.30
4.61
2.71**
1.40
5.23
2.06**
1.18
3.58
2.02**
1.15
3.55
Parity (reference = multiparous)
Primiparous
1.24
0.84
1.84
1.23
0.82
1.84
1.61**
1.08
2.41
1.50
0.98
2.31
14
Urban
Rural
Restricted model
Complete model
Restricted model
Complete model
OR
95% CI
OR
95% CI
OR
95% CI
OR
95% CI
Women’s education (reference = Primary or less)
Secondary
1.11
0.57
2.18
1.10
0.69
1.75
Tertiary
1.59
0.97
2.62
1.57
0.89
2.78
Household wealth quintile (reference = middle)
Poorest
0.78
0.26
2.39
1.01
0.58
1.76
Poor
1.38
0.58
3.28
1.04
0.57
1.88
Rich
1.11
0.56
2.21
0.93
0.47
1.84
Richest
1.88
0.95
3.74
0.86
0.35
2.13
Ethnicity (reference = Kinh/Hoa)
Ethnic minorities
0.89
0.39
2.03
0.55**
0.33
0.90
Region (reference = North Central)
Mekong River Delta
1.04
0.55
1.99
0.75
0.43
1.32
Red River Delta
0.60
0.31
1.15
0.59
0.33
1.06
Northern Midlands
0.54
0.26
1.11
1.35
0.82
2.22
Central Highlands
0.37**
0.18
0.77
0.67
0.38
1.17
Southeast
0.49**
0.27
0.90
1.08
0.60
1.96
15
a**: p ≤ 0.05, *: p ≤ 0.10
(Source: 2013-14 MICS)
16
Discussion
The findings of this study confirm our primary assumption that the place of residence has
a significant effect on CS practices in Vietnam in 2013-14. This outcome contrasts with the
narrowing rural–urban gap in childbirth medicalization (13)(12). It updates the previous
results showing a nonsignificant influence of urbanization on CS in the early 2000s (9). At
the same time, despite growing levels of urbanization, nearly half of all CSs still occur in
rural areas, as has been the case for the last two decades (33)(32)(12). This trend can be
explained by the combination of doubling urbanization rates since 1997 and a more rapid
increase of CS rates in rural areas than in urban areas.
The main determinant of CS is maternal age at delivery. This has also been reported in
previous research in Vietnam (38)(36) and other countries (7)(8)(17)(10). On average, the
mean age at childbearing is 24.7 years in Vietnam, and this indicator has remained stable
between 24 and 24.8 years over the last decade (35). Our study reveals that a maternal age of
over 35 years at childbirth more than doubles the likelihood of undergoing a CS and that this
effect is stronger in urban areas, where childbearing is experienced slightly later than it is in
rural areas.
To understand the factors leading to a CS, we have to take into account the
circumstances of childbirth as well as the whole process of pregnancy. This need is
underlined by the positive influence of a high number of antenatal care visits, which prevails
in urban areas. This phenomenon, which is linked to high levels of antenatal ultrasound, may
also be the consequence of pregnancy complications. It has been observed in Vietnam mostly
in relation to prenatal sex selection and fear of birth defects (39). This trend has also been
witnessed in eastern China, where it has been proven to be linked to a high level of CS
practice (40).
17
A contrast exists in the factors associated with caesarean delivery in rural and urban
areas. Various underlying social, demographic and economic rationales are involved. The
perception of the weight of the newborn over or below average significantly increases the
likelihood of undergoing a CS in urban areas, whereas it has no significant effect in rural
areas. This result complements previous findings regarding periurban settings in Northern
Vietnam, which showed no effect of the weight of the newborn on the mode of delivery (38).
This greater use of CS in cases of macrosomia or low-weight newborns may be linked to the
availability of services.
In addition, the influence of social networks in urban areas could be stronger than that in
rural areas due to a higher level of instruction, higher level of exposure to the media and
greater involvement of women in formal professional activities. Interestingly, women’s
education level has no significant effect. The media hold power over healthcare facilities
through the diffusion of information on their practices and results. Part of this power is used
through social networks, by which public opinion is shaped. Women’s abilities to argue their
cases and seek legal recourse in case of medical complications may act as a more powerful
form of pressure on health staff in urban areas (41)(3). (25). The higher levels of human and
social capital of women could make it more difficult for health personnel to resist women’s
requests to undergo CS.
The highest rates of CS are observed among the richest household quintiles. This
confirms a widespread trend in many countries (6)(3)(42). It also illustrates the persistence of
inequalities in Vietnam despite some progress (43)(30). However, the household level of
wealth has no effect after controlling for other sociocultural factors. This absence may be
partly explained by social insurance coverage, which, despite lower levels of coverage in
rural areas than in urban areas, covers 70% of the population (44). A positive link between
18
health insurance and CS practice has been observed in neighboring China (45) and may apply
to Vietnam despite problems with low protection levels (46), especially in rural areas (47).
In contrast with previous research in other low- and middle-income countries, our results
show no influence of the private health sector after other sociocultural factors are taken into
account (41)(8)(48). In Vietnam, the development of private healthcare facilities has
undergone a major evolution following the Doi Moi reforms launched in the mid-1980s.
However, the proportion of women delivering in this sector remains low. The role of the
private health sector may be underestimated due to the offering of private services in public
health facilities. A more in-depth investigation distinguishing between private and public
services within the public health sector could provide more insights. Further explorations of
our data show that the proportion of women who deliver in the private health sector varies
widely across regions. As an example, the proportion is close to zero in the Northern
Midlands and Mountain area but reaches 20% in the urban Mekong River delta.
Hence, urban areas appear heterogeneous across regions, and the pattern of this
heterogeneity is unexpected. The lowest levels of CS are reached in the Central Highlands,
which is understandable given the low level of equipment and the population density in this
area. Surprisingly, a low level was reached in the highly urbanized and densely populated
Southeast region, where indicators of medical equipment were much more favorable for CS.
Further investigations show that delivery in the private health system and a high number of
antenatal care visits are prominent factors of CS rates in this region, suggesting a complex
combination of determinants.
This complexity is also illustrated by the fact that heterogeneity in CS rates depends on
the region in urban areas but not in rural areas, where the key factor is ethnicity, which in turn
is not relevant in urban areas. Women from minority ethnic groups are less likely to perform
a CS regardless of the other characteristics taken into account in our study. This gap is
19
widened by a lower level of birth in health facilities as well as a lower level of assistance by
skilled attendants during delivery among women belonging to minority ethnic groups (34). It
argues in favor of a sociocultural dimension of attitudes and opinions towards childbirth,
which may involve interpersonal communication and transmission. Through our stepwise
methodology, ethnicity appears to be a hidden factor.
CS determinants may combine with each other. Trends towards lower fertility in urban
areas are in favor of higher levels of antenatal care attendance and CS use (35). Experience in
other countries shows that in a context of reduced fertility, couples tend to be more willing to
invest in the monitoring of pregnancy and caesarean delivery (17). This phenomenon should
be distinguished from the concept of “precious pregnancies” attached to low-fertility couples,
which has been subjected to criticism (49). A previous study performed in Vietnam showed
that discussions with relatives also play a moderating role in helping women avoid CS (9).
Such discussions may be more frequent in rural areas than in urban areas, where the family
size is smaller (50). One heuristic concept capable of integrating the factors of CS may be
“urban liveability”, which encompasses not only the physical setting but also social
interactions and has been studied in relation to the social determinants of health in northern
countries (51). Another factor worth exploring is the influence of the household registration
system. In China, this system has proven to be more strongly linked to unmet long-term care
needs than the place of residence (52). The question of whether similar effects on prenatal
healthcare apply in Vietnam can be explored because the health sector is spatially divided for
heath infrastructures and health insurance schemes.
The different CS rates between rural and urban areas may also be explained by different
levels of healthcare equipment. In Vietnam, where the health system is pyramidal with a
special status for main cities (53)(54), the two metropolitan areas of Hanoi in the Red River
delta and Ho Chi Minh City in the Southeast region play key roles. The fact that more than
20
half of the urban population lives in either the Southeast region (29.9%) or the Red River
delta (23.1%) reveals the demographic weight of the two main metropoles (Hanoi in the Red
River delta and Ho Chi Minh City in the Southeast region). At the other extreme, almost half
of women living in rural areas reside in the Red River delta (25.5%) or the North Central
region (22.0%).
The two main metropolitan areas in the country benefit from a concentration of highly
equipped healthcare facilities in densely populated zones served by viable transport and road
networks (28). This situation leads to a high number of deliveries within specialized
healthcare services, as exemplified by the National Hospital of Gynaecology and Obstetrics
in Hanoi, where more than 20,000 deliveries take place annually, with a CS rate of 48%1. The
rural–urban divide is further strengthened by competition between health infrastructures
following the “autonomization” policy launched in the 2000s, which spurs hospitals to make
profits from investments (55). In urban areas where health personnel are more heavily subject
to time pressure and overcrowded services, CS enables more predictable staff management
and shortens the delivery duration (56). Public hospitals at the tertiary level are closely
monitored (53)(54). These hospitals where CSs are performed (Dinh et al., 2012) play a
pioneering role in the elaboration and implementation of health policies at the national level
(57).
This study has limitations. First, we do not know the reason why the CS deliveries under
study have been performed. In particular, we cannot identify medically indicated CS
deliveries and those performed upon the patient’s request. Therefore, we can only uncover
general trends. Second, we do not distinguish between several levels or types of urbanization;
1 For more information, see Nguyen, T. H. P. (2016). Nghiên cứu tình hình mổ lấy thai tại
bệnh viện phụ sản trung ương từ tháng 3/2016 đến 5/2016 [Research on the situation of
caesarean section in Central Hospital of Gynecology and Obstetrics from March 2016 to
May 2016] (Internship medical thesis). Ministry of Health, Central Hospital of Gynecology
and Obstetrics, Hanoi.
21
this type of analysis would require a large sample size. Third, the place of residence may not
coincide with the place of delivery. Therefore, we capture the impact of the long-term
influences of the context rather than the impact of possible adaptation through migration.
Fourth, our statistical analysis provides indications of correlations rather than causal links.
However, we are convinced that this study provides useful insights into the influence of
urbanization on CS through highlighting its major determinants and suggesting a way to
approach this complex phenomenon using existing data representative of the national level.
Conclusion
The overall CS rate among the women who delivered in healthcare facilities in Vietnam
is particularly high (29.2%) with regards to WHO standards (14). After controlling for
significant characteristics, living in urban areas more than doubles the likelihood of
undergoing a CS (OR = 2.31; 95% CI 1.79 to 2.98). Maternal age at delivery over 35 is a
major positive correlate of CS. Beyond this common phenomenon, our study has shown
contrasting models regarding the determinants of recourse to high levels of CS rates between
rural and urban areas. This contrast suggests that actions to reduce unnecessary caesarean
deliveries should be adapted to each context. Indeed, our results show the importance of
taking into account not only medical and sociodemographic factors but also sociocultural
determinants when designing programs to improve women’s childbirth conditions. It is the
case of ethnicity, which needs to be addressed. This approach involves policies at many
different levels regarding not only the regulation of the health sector and training of
healthcare providers but also the sensitization of the entire population, with means
appropriate to their conditions of living. Further research must be conducted to design such
programs and to provide guidance on this complex issue.
22
Acknowledgements
The authors acknowledge the Vietnam General Statistics Office (GSO) and Vietnam
UNICEF for providing the underlying data that made this research possible, with special
thanks to Ms. Nguyen Quynh Trang (UNICEF Vietnam) and Mr. Nguyen Dinh Chung
(GSO). They also thank IRD and CEPED for their support.
References
1.
World Health Organization. WHO Statement on Caesarean Section Rates. Geneva:
Human Reproduction Programme; 2015 p. 8.
2.
Stanton CK, Holtz SA. Levels and trends in cesarean birth in the developing world. Stud
Fam Plann. 2006;37(1):41–48.
3.
Cavallaro FL, Cresswell JA, França GV, Victora CG, Barros AJ, Ronsmans C. Trends
in caesarean delivery by country and wealth quintile: cross-sectional surveys in southern
Asia and sub-Saharan Africa. Bull World Health Organ. 2013 Dec 1;91(12):914-922D.
4.
Khanal V, Karkee R, Lee AH, Binns CW. Adverse obstetric symptoms and rural–urban
difference in cesarean delivery in Rupandehi district, Western Nepal: a cohort study.
Reprod Health [Internet]. 2016 Dec [cited 2017 Dec 8];13(1). Available from:
http://www.reproductive-health-journal.com/content/13/1/17
5.
Kaboré C, Ridde V, Kouanda S, Agier I, Queuille L, Dumont A. Determinants of non-
medically indicated cesarean deliveries in Burkina Faso. Int J Gynecol Obstet. 2016
Nov 1;135(S1):S58–63.
6.
Boatin AA, Schlotheuber A, Betran AP, Moller A-B, Barros AJD, Boerma T, et al.
Within country inequalities in caesarean section rates: observational study of 72 low and
middle income countries. BMJ. 2018 Jan 24;k55.
7.
Yisma E, Smithers LG, Lynch JW, Mol BW. Cesarean section in Ethiopia: prevalence
and sociodemographic characteristics. J Matern Fetal Neonatal Med. 2017 Nov 20;1–6.
8.
Al Rifai RH. Trend of caesarean deliveries in Egypt and its associated factors: evidence
from national surveys, 2005–2014. BMC Pregnancy Childbirth [Internet]. 2017 Dec
[cited 2018 Jan 25];17(1). Available from:
https://bmcpregnancychildbirth.biomedcentral.com/articles/10.1186/s12884-017-1591-2
9.
Leone T, Padmadas SS, Matthews Z. Community factors affecting rising caesarean
section rates in developing countries: An analysis of six countries. Soc Sci Med. 2008
Oct 1;67(8):1236–46.
23
10. Abdul-Rahim HF, Abu-Rmeileh N “Mohammad E, Wick L. Cesarean section deliveries
in the occupied Palestinian territory (oPt): An analysis of the 2006 Palestinian Family
Health Survey. Health Policy. 2009 Dec;93(2–3):151–6.
11. Chang AL, Yoshino KD, Sentell T, Miyamura J, Maddock J, Pacheco M. Comparison
of Cesarean Delivery Rates in Urban and Rural Hospitals [1]: Obstet Gynecol. 2015
May;125:1S.
12. National Committee for Population and Family Planning, The Population and Family
Health Project. Demographic and Health Survey 1997 [Internet]. Hanoi; 1999 [cited
2017 Dec 13]. Available from: https://dhsprogram.com/pubs/pdf/FR104/FR104.pdf
13. General Statistics Office, Unicef. Viet Nam Multiple Indicator Cluster Survey 2014,
Final Report [Internet]. 2015 [cited 2017 Dec 13] p. 364 p. + 64 p. appendices.
Available from: https://mics-surveys-
prod.s3.amazonaws.com/MICS5/East%20Asia%20and%20the%20Pacific/Viet%20Nam
/2013-2014/Final/Viet%20Nam%202013-14%20MICS_English.pdf
14. Betrán AP, Ye J, Moller A-B, Zhang J, Gülmezoglu AM, Torloni MR. The Increasing
Trend in Caesarean Section Rates: Global, Regional and National Estimates: 1990-2014.
Zeeb H, editor. PLOS ONE. 2016 Feb 5;11(2):e0148343.
15. Vogel. Use of the Robson classification to assess caesarean section trends in 21
countries: a secondary analysis of two WHO multicountry surveys. 2015 [cited 2017
Dec 14]; Available from: https://ac.els-cdn.com/S2214109X1570094X/1-s2.0-
S2214109X1570094X-main.pdf?_tid=d26274e2-e0fe-11e7-ae46-
00000aab0f27&acdnat=1513277251_405efe5a3165d046000cdd65fa8a8edc
16. Ministry of Planning and Investment, General Statistics Office. Migration and
Urbanization in Vietnam: Patterns, Trends and Differentials [Internet]. Hanoi; 2011
[cited 2018 Jan 2] p. 106 p. + appendices. Available from:
http://vietnam.unfpa.org/sites/default/files/pub-pdf/7_Monograph-Migration-
Urbanization.pdf
17. Chen C-S, Lin H-C, Liu T-C, Lin S-Y, Pfeiffer S. Urbanization and the likelihood of a
cesarean section. Eur J Obstet Gynecol Reprod Biol. 2008 Dec;141(2):104–10.
18. Souza J, Betran A, Dumont A, de Mucio B, Gibbs Pickens C, Deneux-Tharaux C, et al.
A global reference for caesarean section rates (C-Model): a multicountry cross-sectional
study. BJOG Int J Obstet Gynaecol. 2016 Feb;123(3):427–36.
19. Schantz C, Sim KL, Petit V, Rany H, Goyet S. Factors associated with caesarean
sections in Phnom Penh, Cambodia. Reprod Health Matters. 2016 Nov;24(48):111–21.
20. Ronsmans C, Holtz S, Stanton C. Socioeconomic differentials in caesarean rates in
developing countries: a retrospective analysis. The Lancet. 2006;368(9546):1516–1523.
21. Kozhimannil KB, Hung P, Prasad S, Casey M, Moscovice I. Rural-Urban Differences in
Obstetric Care, 2002–2010, and Implications for the Future: Med Care. 2014
Jan;52(1):4–9.
24
22. Meda IB, Millogo T, Baguiya A, Ouédraogo/Nikiema L, Coulibaly A, Kouanda S. Rate
of and factors associated with indications for cesarean deliveries: Results of a national
review in Burkina Faso. Int J Gynecol Obstet. 2016 Nov 1;135(S1):S51–7.
23. Irani M, Deering S. Challenges affecting access to cesarean delivery and strategies to
overcome them in low-income countries. Int J Gynecol Obstet. 2015 Oct;131(1):30–4.
24. Bui TC, Markham CM, Ross MW, Williams ML, Beasley RP, Tran LT, et al.
Dimensions of gender relations and reproductive health inequity perceived by female
undergraduate students in the Mekong Delta of Vietnam: a qualitative exploration. Int J
Equity Health. 2012;11(1):63.
25. Kottwitz A. Mode of birth and social inequalities in health: The effect of maternal
education and access to hospital care on cesarean delivery. Health Place. 2014
May;27:9–21.
26. Forbes D. Urbanization, Migration and Vietnam’s Spatial Structure. SOJOURN J Soc
Issues Southeast Asia. 1996;11(1):24–51.
27. Beresford M. Doi Moi in review: The challenges of building market socialism in
Vietnam. J Contemp Asia. 2008 May;38(2):221–43.
28. World Bank. Vietnam Urbanization Review. Technical Assistance Report. 2011 p. 238.
29. World Bank. World Development Indicators | DataBank [Internet]. 2017 [cited 2018 Jan
2]. Available from: http://databank.worldbank.org/data/reports.aspx?source=world-
development-indicators
30. Benjamin D, Brandt L, McCaig B. Growth with equity: income inequality in Vietnam,
2002–14. J Econ Inequal. 2017 Mar;15(1):25–46.
31. Welle-Strand A, Vlaicu M, Tjeldvoll A. Vietnam - A New Economic Dragon in
Southeast Asia? J Dev Soc. 2013 May 3;29(2):155–87.
32. Committee for Population, Family and Children, ORC Macro. Vietnam Demographic
and Health Survey 2002 [Internet]. Calverton, Maryland, USA; 2003 [cited 2017 Dec
13]. Available from: https://dhsprogram.com/pubs/pdf/FR139/FR139.pdf
33. General Statistics Office, United Nations Children’s Fund, United Nations Population
Fund. Viet Nam Multiple Indicator Cluster Survey 2011, final report [Internet]. Hanoi,
Vietnam; 2011 [cited 2017 Dec 13] p. 218 + 50 p appendices. Available from:
https://mics-surveys-
prod.s3.amazonaws.com/MICS4/East%20Asia%20and%20the%20Pacific/Viet%20Nam
/2010-2011/Final/Viet%20Nam%202011%20MICS_English.pdf
34. UNFPA, Ministry of Health. Final report. Vietnam Midwifery report 2016. Hanoi; 2017
p. 253.
35. General Statistics Office, United Nations Population Fund. Fetility in Viet Nam:
Differentials, Trends, and Determinants. Hanoi: Vietnamese Agency Publishing House;
2016.
25
36. Tran KT, Eriksson B, Nhat An P, Thi Kim Chuc N, Bondjers G, Gottvall K. Technology
Preference in Choices of Delivery Care Utilization from User Perspective –A
Community Study in Vietnam. Am J Public Health Res. 2013 Jan 31;1(1):10–7.
37. Dinh TPH, Börjesson L, Nguyen TN, Johansson A, Målqvist M. Sex of Newborns
Associated With Place and Mode of Delivery: A Population-Based Study in Northern
Vietnam. Gend Med. 2012 Dec 1;9(6):418–23.
38. Duong DM, Nguyen AD, Nguyen CC, Le VT, Hoang SN, Bui HTT. A Secular Trend in
Birth Weight and Delivery Practices in Periurban Vietnam During 2005-2012. Asia Pac
J Public Health. 2017;29(5_suppl):18S–24S.
39. Gammeltoft T. Haunting images: a cultural account of selective reproduction in
Vietnam. Berkeley, California: University of California Press; 2014. 315 p.
40. Huang K, Tao F, Raven J, Liu L, Wu X, Tang S. Utilization of antenatal ultrasound scan
and implications for caesarean section: a cross-sectional study in rural Eastern China.
BMC Health Serv Res. 2012;12(1):93.
41. Neuman M, Alcock G, Azad K, Kuddus A, Osrin D, More NS, et al. Prevalence and
determinants of caesarean section in private and public health facilities in underserved
South Asian communities: cross-sectional analysis of data from Bangladesh, India and
Nepal. BMJ Open. 2014;4(12):e005982.
42. Khan MN, Islam MM, Rahman MM. Inequality in utilization of cesarean delivery in
Bangladesh: a decomposition analysis using nationally representative data. Public
Health. 2018 Apr;157:111–20.
43. Axelson H, Gerdtham U-G, Ekman B, Hoa DTP, Alfvén T. Inequalities in reproductive,
maternal, newborn and child health in Vietnam: a retrospective study of survey data for
1997–2006. BMC Health Serv Res. 2012;12(1):456.
44. OECD. Social Cohesion Policy Review of Viet Nam [Internet]. OECD Publishing; 2014
[cited 2014 Nov 3]. (Development Centre Studies). Available from: http://www.oecd-
ilibrary.org/education/social-cohesion-policy-review-of-viet-nam_9789264196155-en
45. Long Q, Klemetti R, Wang Y, Tao F, Yan H, Hemminki E. High caesarean section rate
in rural China: Is it related to health insurance (New Co-operative Medical Scheme)?
Soc Sci Med. 2012 Aug;75(4):733–737.
46. Nguyen KT, Khuat OTH, Ma S, Pham DC, Khuat GTH, Ruger JP. Impact of Health
Insurance on Health Care Treatment and Cost in Vietnam: A Health Capability
Approach to Financial Protection. Am J Public Health. 2012 Aug;102(8):1450–61.
47. Palmer MG. Inequalities in Universal Health Coverage: Evidence from Vietnam. World
Dev. 2014 Dec;64:384–94.
48. Phadungkiatwattana P, Tongsakul N. Analyzing the impact of private service on the
cesarean section rate in public hospital Thailand. Arch Gynecol Obstet. 2011
Dec;284(6):1375–9.
26
49. Minkoff HL, Berkowitz R. The Myth of the Precious Baby: Obstet Gynecol. 2005
Sep;106(3):607–9.
50. General Statistics Office. The 2014 Vietnam Intercensal Population and Housing
Survey. Population Sex-Age Structure and Related Socio-Economic Issues in Viet
Nam. Hanoi: Vietnam News Agency Publishing House; 2016 p. 104.
51. Badland H, Whitzman C, Lowe M, Davern M, Aye L, Butterworth I, et al. Urban
liveability: Emerging lessons from Australia for exploring the potential for indicators to
measure the social determinants of health. Soc Sci Med. 2014 Jun;111:64–73.
52. Zhu Y, Österle A. Rural-urban disparities in unmet long-term care needs in China: The
role of the hukou status. Soc Sci Med. 2017 Oct;191:30–7.
53. Priwitzer K. The Vietnamese Health Care System in Change. A Policy Network
Analysis of a Southeast Asian Welfare Regime. Singapore: Institute of Southeast Asian
Studies; 2012. 370 p.
54. Thompson CM. Vietnamese traditional medicine: a social history. 2015.
55. London JD. The promises and perils of hospital autonomy. Soc Sci Med. 2013;96:232–
40.
56. Hellerstein S, Feldman S, Duan T. Survey of Obstetric Care and Cesarean Delivery
Rates in Shanghai, China. Birth. 2016;43(3):193–199.
57. Kerkvliet BJ, Marr DG, Institute of Southeast Asian Studies, editors. Beyond Hanoi:
local government in Vietnam. Copenhagen S, Denmark : Singapore: NIAS Press ;
Institute of Southeast Asian Studies; 2004. 359 p. (Vietnam update series).
| 2019 | Magnitude and Correlates of Caesarean Section in Urban and Rural Areas: A Multivariate Study in Vietnam | 10.1101/554964 | [
"Loenzien Myriam de",
"Schantz Clémence",
"Luu Bich Ngoc",
"Dumont Alexandre"
] | creative-commons |
Viral fitness determines the magnitude of transcriptomic and
epigenomic reprogramming of defense responses in plants
Régis L. Corrêa,1,2,3,6,* Alejandro Sanz-Carbonell,1 Zala Kogej,1,4 Sebastian Y. Müller,3 Sara
López-Gomollón,3 Gustavo Gómez,1 David C. Baulcombe,3 Santiago F. Elena1,5,*
1Instituto de Biología Integrativa de Sistemas (I2SysBio), Consejo Superior de Investigaciones
Científicas (CSIC) - Universitat de València, Paterna, 46980 Valencia, Spain
2Department of Genetics, Federal University of Rio de Janeiro (UFRJ), Rio de Janeiro, 21941-
590, Brazil
3Department of Plant Sciences, University of Cambridge, Cambridge, CB2 3EA, United
Kingdom
4Present address: Department of Biotechnology and Systems Biology, National Institute of
Biology, Ljubljana, 1000, Slovenia
5The Santa Fe Institute, Santa Fe, NM 87501, USA
6Lead contact
*Correspondence: regislcorrea@ufrj.br (R.L.C.), santiago.elena@csic.es (S.F.E.)
SUMMARY
Although epigenetic factors may influence the expression of defense genes in plants, their role
in antiviral responses and the impact of viral adaptation and evolution in shaping these
interactions are still poorly explored. We used two isolates of turnip mosaic potyvirus (TuMV)
with varying degrees of adaptation to Arabidopsis thaliana to address these issues. One of
the isolates was experimentally evolved in the plant and presented increased load and
virulence relative to the ancestral isolate. The magnitude of the transcriptomic responses were
larger for the evolved isolate and indicated a role of innate immunity systems triggered by
molecular patterns and effectors in the infection process. Several transposable elements
(TEs) located in different chromatin contexts and epigenetic-related genes were also affected.
Correspondingly, mutant plants having loss or gain of repressive marks were, respectively,
more tolerant and susceptible to TuMV, with a more efficient response against the ancestral
isolate. In wild-type plants both isolates induced similar levels of cytosine methylation
changes, including in and around TEs and stress-related genes. Results collectively
suggested that apart from RNA silencing and basal immunity systems, DNA methylation and
histone modification pathways may also be required for mounting proper antiviral defenses in
Main text
plants and that the effectiveness of this type of regulation strongly depends on the degree of
viral adaptation to the host.
Keywords
Biotic stress, Epigenome, Methylome, Plant-virus interaction, Potyvirus, RNA-directed DNA
methylation, Systems Biology, Transposable elements, Turnip mosaic virus, Virus adaptation,
WGBS.
INTRODUCTION
Biotic stress responses in plants can be triggered by the recognition of pathogens’ conserved
motifs, proteins or RNA molecules. Pathogen-associated molecular patterns (PAMP) may be
recognized by membrane receptors, triggering a general response referred as PAMP-
triggered immunity (PTI) (Boutrot and Zipfel, 2017). A stronger defense is initiated when
pathogen-specific proteins or other elicitors are recognized by resistance (R) proteins
belonging to the NLR (intracellular nucleotide binding site, leucine-rich repeat containing
receptor) family (Cui et al., 2015). The effector-triggered immunity (ETI) is linked to the
induction of hypersensitive response (HR), restricting pathogen spread. Both PTI and ETI are
associated with the production of hormones that may promote systemic resistance, inducing
the production of resistance pathogenesis-related (PR) proteins, among others (Fu and Dong,
2013). Basal immunity systems are linked mainly to non-viral pathogens, but there is
increasing evidence that they may also play a role against viruses (Teixeira et al., 2019).
RNA-based immunity systems are triggered by the recognition and degradation of
double-stranded RNA (dsRNA) molecules. The mechanism is mostly associated in the
defense against viruses (Wu et al., 2019). Viral dsRNAs are degraded by DICER-LIKE (DCL)
proteins into small RNAs (sRNAs) that are loaded into ARGONAUTE (AGO) proteins and used
as a guide to repress similar single-stranded RNAs. By using RNA-DEPENDENT RNA
POLYMERASE (RDR) proteins to generate new dsRNAs from targets, the RNA silencing
response can also be amplified (Borges and Martienssen, 2015). Viral dsRNAs can also feed
into the PTI pathway (Niehl and Heinlein, 2019). Pathogens on the other hand, may evolve
mechanisms to avoid or inactivate various steps of RNA silencing or PTI/ETI defenses, leading
to Red Queen coevolutionary dynamics.
RNA-based defenses against viruses in plants are part of a broader and conserved
system that includes processes that regulate gene expression and control transposable
elements (TEs) by the addition of epigenetic marks to DNA or DNA-associated histone
proteins (Borges and Martienssen, 2015). Most of the DNA methylation marks in eukaryotes
are linked to cytosine, particularly those followed by guanine (CG). Non-CG methylation,
including CHG and CHH (where H is any nucleotide, except G), however, is also observed. In
plants, the symmetrical CG and CHG methylation are maintained by methyltransferases and
the chromatin remodeling factor DECREASE IN DNA METHYLATION 1 (DDM1) during the
replication process (Sigman and Slotkin, 2016). Signals for restoring asymmetrical CHH
modifications, however, are lost and re-established after every cell division by a sRNA-guided
complex. The mechanism known as RNA-directed DNA methylation (RdDM) is orchestrated
by complexes containing two plant-specific RNA polymerases (Pol IV and V), the RNA
silencing-related factors RDR2, DCL3 and AGO4 for sRNA generation and amplification and
epigenetic factors, e.g., the methyltransferase DOMAINS REARRANGED METHYLASE 2
(DRM2) (Zhang et al., 2018).
RdDM mainly targets small and recently acquired TEs or the borders of long TEs in
euchromatic regions (Sigman and Slotkin, 2016). The mechanism, therefore, establishes a
heterochromatin-like environment within the euchromatin. Environmental stresses may pose
a challenge for the maintenance of this chromatin border, as genes and TEs can mutually
influence their expression under certain conditions (Negi et al., 2016). Changes in cytosine
DNA methylation patterns due to stress have also been reported (Zhang et al., 2018). The
impact of those epigenetic changes in gene expression settings are still elusive, especially for
small and heterochromatin-poor genomes like the Arabidopsis thaliana one. The role of
pathogens, and especially RNA viruses in DNA methylation responses also remains poorly
explored. Contrary to passive abiotic stressors, pathogens can interact and manipulate host
signaling pathways and therefore potentially exploit the intensity or types of epigenetic
responses. In particular, fast-evolving RNA viruses may overcome host defenses by (i) quickly
generating extremely diverse mutant swarms that contain escape variants that are not
controlled by immunity (Andino and Domingo, 2015) or (ii) by encoding specific proteins that
actively interact and block host defenses, being the viral suppressors of RNA silencing (VSR)
relevant players in the context of this study (Wu et al., 2010).
We used A. thaliana ecotype Col-0 and Turnip mosaic virus (TuMV; genus Potyvirus,
family Potyviridae; picorna-like superfamily) pathosystem as a model to explore those topics.
TuMV is an economically relevant virus that infects cruciferous plants, including arabidopsis.
Its compact positive single-stranded, polyadenylated RNA genome produces a single
polyprotein that is processed into 10 major multifunctional proteins (Ivanov et al., 2014) plus
an additional protein encoded in an alternative small ORF (Chung et al. 2008). To test whether
viral evolution and adaptation changes the way viruses might interplay with host epigenetic
regulation, two TuMV isolates with different fitness in arabidopsis were used. We show that
epigenetic pathways have relevant roles in virus infectivity and that the responses are
influenced by pathogen's fitness in the host. We also find several virus-induced DNA
methylation changes, but that their impact on transcriptional changes cannot be generalized.
Overall, no major differences in the methylome exist between both viral isolates, however, the
high-fitness TuMV isolate has a much stronger impact at the transcriptomic level.
RESULTS
Experimental evolution of TuMV in arabidopsis
Host-pathogen interactions in plants, as in other systems, are heavily regulated by a
coevolutionary arms race of defense and counter-defense mechanisms. To check the impact
of virus-acquired adaptations on host's transcriptome and methylome responses, a calla lily
isolate of TuMV (Chen et al., 2003), that had been propagated in Nicotiana benthamiana
plants, was experimentally evolved by serial passages in arabidopsis plants (Figure 1A). By
repeatedly challenging the virus population with the novel host, we expected to evolve a TuMV
isolate better adapted to this particular host than the original isolate, which was naïve to the
plant.
When similar amounts of inoculum were used, the onset of early symptoms in the upper
systemic leaves started ~7 days post-inoculation (dpi), irrespective of the isolate used (Figure
1B). However, plants infected with the evolved virus progressed into strong symptoms faster
than the ancestral-infected ones (Figure 1B and 1C). The largest symptom differences
between the two viruses were observed 10 - 12 dpi (Figure 1B); most evolved virus-infected
plants developed clear and strong leaf yellowing and stunning symptoms, while the ancestral
virus-infected ones were still displaying light symptoms or remained symptomless (Figure 1B
and S1A). The observed difference in symptoms was paralleled with viral load. At early
infection stages (2 and 5 dpi), before symptoms appearance, there was no significant
difference in the levels of TuMV accumulation (Figure S1B; 2-samples t-tests P ≥ 0.1620).
However, after 12 dpi, when symptoms were clearly distinct between isolates, the load of the
evolved virus was significantly higher than the ancestral virus in systemically-infected leaves
(Figure S1B; 2 samples t-test P = 0.0013). In agreement, the evolved virus killed the plants
significantly faster than the ancestral virus (Figure S1C; Kaplan-Meier survival analysis: P =
0.0003).
The genomes of the ancestral and evolved isolates were compared by variant calling
through Illumina polyA-purified RNA sequencing (mRNA-seq) reads. Two single-nucleotide
polymorphisms (SNPs) in the evolved isolate, leading to amino acid substitutions L107F and
D110N, were observed (Figure S1D and S1E). Both substitutions affected the genome-linked
viral protein VPg, which is a multifunctional protein involved in viral replication, genome
stabilization, translation, and suppression of RNA silencing-based defenses (Cheng and
Wang, 2017; Ivanov et al., 2014). These amino acids are located at the end of the third
predicted α-helix (Figure S1D), in a region required for the VPg self-interaction and in close
proximity to regions important for its interaction with the VSR protein HC-Pro and the host
translation initiation factor eIF4E in related viruses (Roudet-Tavert et al., 2007; Yambao et al.,
2003). Collectively, the development of symptoms, virus accumulation and molecular data
indicated that the evolution experiment was effective in producing a TuMV isolate that is more
virulent and better adapted to arabidopsis.
Transcriptomic responses to TuMV infection
The magnitude and nature of plant transcriptomic responses to infection depend on the fitness
of the particular potyviral strain being inoculated (Agudelo-Romero et al., 2008; Cervera et al.,
2018; Hillung et al., 2016), including one study comparing two other TuMV strains (Sánchez
et al., 2015). To confirm this observation in our particular pathosystem, arabidopsis plants
were inoculated with equivalent amounts of transcripts from the ancestral or evolved TuMV
isolates or mock-inoculated, and RNAs extracted from systemic leaves before symptom
appearance (5 dpi) and late infection (12 dpi). In addition, a sample was taken 2 dpi (early
infection) for the evolved virus. Transcriptomes of three biological replicates (plants) from each
condition (mock-inoculated, ancestral virus- and evolved virus-infected) were assessed with
stranded mRNA-seq. The vast majority of the reads in the infected plants mapped to the
arabidopsis genome (Figure S2A), allowing the detection of differentially expressed genes
(DEGs) in all time-points.
When compared to mock-inoculated samples, the number of DEGs was larger in the
response against the evolved virus in all time-points analyzed. The number of DEGs for the
evolved virus at 2 and 5 dpi was about three and seven times higher than for the ancestral at
5 dpi, respectively (Figure 2A and Table S1), indicating that the evolved virus elicited stronger
responses at 2 dpi than the ancestral at 5 dpi. As infection progressed, responses between
isolates tended to equalize, although total number of DEGs were still ~1.5 higher for the
adapted virus at 12 dpi (Figure 2A and S2B). A total of 18 genes were regulated due to the
infection in all time-points for both viruses (Figure S2B), including eight known stress-
responsive genes (Table S2).
Responses against the ancestral virus at 5 dpi were characterized by an enrichment in
genes associated with biotic and abiotic stresses and repression of metabolic and biosynthetic
processes (Figure S2C). The core of defense-related genes associated with general stress-
responses, though, were only observed at 12 dpi for plants infected with the ancestral isolate.
Responses to the evolved virus, on the other hand, were much faster. At 2 dpi, typical shut-
down of general metabolism and photosynthesis was already observed (Figure S2C). All major
classes of regulated genes observed only at 12 dpi for the ancestral virus were already
enriched against the evolved one at 5 dpi. At 12 dpi, those classes were enhanced, with the
additional repression of ribosome constituents (Figure S2C). To highlight the difference
between the viral isolates, the direct comparison of the transcriptomes from plants infected
with the ancestral and evolved strains was performed. This analysis evidenced the stronger
perturbation of the overall physiological homeostasis by the evolved isolate, including the
induction of genes related to general and biotic stresses and transcriptional factors (Figure
2B). Suppression of biotic stress genes in evolved-infected samples was also observed, a
change that may possibly be to the advantage of the virus (Figure 2B).
A network analysis of the identified DEGs was performed using the Arabidopsis
Comprehensive Knowledge Network (AtCKN) (Ramšak et al., 2018). Dynamic views of
AtCKN’s cluster 40, enriched for several well characterized stress-responsive genes, are
available in the Supplemental Files S1 (ancestral) and S2 (evolved). The analysis indicated
that the evolutionary conserved WRKY transcriptional factors may play important roles in
response triggering and dynamics. At 2 dpi, WRKY70, a known activator of salicylic acid (SA)-
related defense genes and a repressor of jasmonic acid (JA)-ones (Li et al., 2004), was
induced against the evolved virus (Figure S2D). At 5 dpi, both isolates induced the expression
of WRKY25 and several of its direct targets (Figure S2D, Table S1, Files S1 and S2). This
gene is a known repressor of SA responses (Zheng et al., 2007). Its induction, together with
other WRKY SA-counteractors (WRKY26/33/38/62) evidenced a possible SA-buffering
mechanism at mid and late infection points (Kc et al., 2008; Zheng et al., 2006). Other
transcriptional factor families also seem to have relevant roles during early responses (Figure
S3A and Table S3). A cross-talk with other hormones was also observed in early and late
infection phases, especially genes related to abscisic acid (ABA), ethylene (ET) and JA (Figure
2C, Table S3). Furthermore, several genes associated with both PTI and ETI systems were
regulated against TuMV, though to a larger extent for the evolved virus, with PR1 having the
highest fold change among them at 12 dpi (Figure 2D, Table S3). A pronounced induction of
PTI and PRs were observed in response to the evolved isolate at 12 dpi when compared to
the ancestral one, which were paralleled with a higher increase of SA genes in this time-point
(Figures 2C and 2D). ETI-related genes seem to be more dynamically regulated when the
isolates are compared, with the bulk difference taking place at 5 dpi, despite the high induction
of some of them at 12 dpi (Figure 2D). Expression of representative genes (PR1, WAK1,
HSP70, and COR15a) associated with biotic and abiotic stresses were confirmed by
quantitative PCR (RT-qPCR) to validate observed mRNA-seq results (Figure S3B), confirming
that, on average, expression of these genes was higher in plants infected with the evolved
strain (post hoc pairwise comparisons with sequential Bonferroni correction P = 0.0046).
Viruses are targeted by RNA silencing defenses. Accordingly, the majority of the RNA
silencing-related genes among the DEGs were induced (Figure 3A, Table S3). Most of the
DNA methylation-related DEGs, on the other hand, were repressed against both isolates
(Figure 3A, Table S3). The changed expression of INCREASE IN BONSAI METHYLATION 1
(IBM1) and REPRESSOR OF SILENCING 1 (ROS1), two genes known to act as methyl
sensors (Lei et al., 2015; Rigal et al., 2012), was confirmed by RT-qPCR (Figure S3C).
Interestingly, the average level of expression was significantly lower in plants infected with the
evolved than with the ancestral virus (post hoc pairwise comparisons with sequential
Bonferroni correction P < 0.0001). The overall responses therefore indicated that both
DNA/histone layers of epigenetic regulation might be altered during virus infection.
Since several epigenetic pathways have TEs as targets, the expression of TE families
was checked with TEtranscripts, a tool developed to handle reads mapping to repetitive
sequences (Jin et al., 2015). At 5 dpi, seven TEs belonging to the Gypsy and Copia families,
usually concentrated in centromeric and pericentromeric regions (Underwood et al., 2017),
respectively, were induced against the evolved isolate (Figure 3B, Table S4). One of them,
the Gypsy ATHILA2, is enriched in the centromere core that is transiently regulated by
temperature shifts and viral infections (Diezma-Navas et al., 2019; Tittel-Elmer et al., 2010).
At 12 dpi, however, both induction and repression of TE families was observed (Figure 3B,
Table S4). The induced elements were again mostly from Gypsy and Copia families, including
AtCOPIA93/Evadé, a TE that is induced against bacterial and viral infections (Diezma-Navas
et al., 2019; Zervudacki et al., 2018). The repressed TEs at 12 dpi, on the other hand, included
the Helitron, Harbinger and Mutator (MuDR) families that are usually located close to genes
(Figure 3B, Table S4). The misregulation of several DNA methylation and histone modification
genes and TEs located in different genomic contexts further suggested that epigenetic factors
may play a role during the infection process.
Effects of epigenetic-related genes in arabidopsis response to TuMV infection
So far, we have presented evidence suggesting that epigenetic factors may play a role during
TuMV infection. To directly test this possibility, arabidopsis mutant genotypes having
compromised or enhanced DNA/histone methylation were challenged against the two TuMV
isolates. Disease severity was checked by scoring the number of days each plant took to reach
strong leaf yellowing symptoms. All tested RdDM mutants, involved mainly in the regulation of
small TEs located within euchromatic environments, were more resistant to TuMV than wild-
type plants; though they were significantly more resistant against the ancestral isolate (Figure
4A; post hoc pairwise comparisons with sequential Bonferroni correction P < 0.0001). Among
the challenged RdDM genotypes, ago4 and rdr2 were the most and least resistant ones,
respectively, while poliv, polv and double drm1 drm2 presented intermediate values. Strong
resistance, especially for the ancestral virus, was also observed in ddm1 mutants, lacking a
master regulator of TEs (Figure 4B; post hoc pairwise comparisons with sequential Bonferroni
correction P ≤ 0.0003).
Histone modification mutants, however, had opposite effects depending on the altered
pathway. Compared to wild-type plants, ibm1 mutants were significantly more susceptible to
the evolved isolate (Figure 4C; post hoc pairwise comparisons with sequential Bonferroni
correction P < 0.0001), but not against the ancestral one (Figure 4C; post hoc pairwise
comparisons with sequential Bonferroni correction P = 0.7778). IBM1 is a histone demethylase
that
removes
TE-associated
H3K9
marks
from
genes,
therefore
reinforcing
euchromatin/heterochromatin borders (Saze et al., 2008). On the other hand, inoculation of
both isolates in mutants of the gene JUMONJI 14 (JMJ14), rendered plants more resistant to
the virus (Figure 4C; post hoc pairwise comparisons with sequential Bonferroni correction P <
0.0001). JMJ14 is also a histone demethylase, but removes H3K4 methylation marks, a
modification usually associated with gene activation (Lu et al., 2010; Searle et al., 2010).
Infection results in the mutant backgrounds therefore indicated that infectivity and
development of symptoms severity may be correlated to altered chromatin states. Mutants
defective in heterochromatin formation (RdDM mutants, ddm1 and jmj14) are more tolerant to
TuMV infection, whilst the one with reduced euchromatin (ibm1) was more susceptible. The
experiments also support the transcriptome findings that epigenetic factors may be required
for virus defense mechanisms in plants.
Virus-induced DNA methylation changes
Since several genes related to cytosine DNA methylation influenced TuMV infectivity, we
asked whether this type of epigenetic modification is altered during the infection process in
wild-type plants. Whole-genome bisulfite sequencing methylome libraries were constructed
and Illumina-sequenced (WGBS-seq) for ancestral and evolved virus-infected plants at three
time points: 2, 5 and 12 dpi. DNA material came from the same samples used for the
transcriptome analysis. The observed differentially methylated regions (DMRs, in 100 bp tiles)
were analyzed separately for the three cytosine methylation contexts (CpG, CHG and CHH)
and divided into hypermethylated and hypomethylated, for gain or loss of methylation in
comparison to mock-inoculated control plants, respectively (Table S5). In contrast to the
transcriptome data, the numbers of DMRs induced by TuMV were relatively even between the
ancestral and evolved isolates along the time-course (Figure S4A). The exception was for
CHG DMRs at 12 dpi, that were clearly more pronounced in evolved virus-infected plants, with
ca. twice of them hypermethylated. DMRs in the CpG context were in general more numerous
during TuMV infection than in the other contexts (Figure S4A). CHG and CHH DMRs,
however, had a marked increase at 12 dpi (Figure S4A).
Most of the observed CpG DMRs were mapped within protein-coding genes (Figure 5A
and 5B). However, DMRs in CpG context proximal to transcriptional start sites (TSS) were
also observed and, to a lesser extent, within TEs (Figure 5A and 5B). CHG and CHH DMRs,
as expected, were enriched in TEs, with increased numbers in later infection times (Figure 5A
and 5B). Plants infected with the evolved virus had about 2-fold more CHG DMRs in TEs at
12 dpi, corresponding to the bulk methylation difference in this context between the isolates
(Figure 5A, 5B and S4A). In agreement with transcriptional profiles, DMRs from all three
contexts were found in TE families located throughout the genome, with Gypsy, MuDR and
Copia the most frequent ones (Figure S4B). While CpG DMRs in TEs tented to have similar
amounts of hyper- or hypo-methylation irrespective of the time point, non-CpG DMRs in those
elements had a clear tendency for hypermethylation at 12 dpi (Figure 5A, 5B and S4B).
Methylome profiles identified therefore the existence of DMRs in both TEs and genes during
TuMV infection, suggesting a possible mutual regulation between them.
Impact of virus-derived methylation changes on the transcriptome
Since methylation of promoters is usually associated with changes in gene expression, the
impact of TSS-proximal DMRs in the expression of protein-coding genes was assessed.
DMRs in the CpG context were the most abundant ones in the region comprising 2 kb
upstream and 200 bp downstream of protein-coding genes’ TSS, followed by CHG and CHH
DMRs. If TSS-proximal methylation has a role in gene expression control, a negative
correlation between them would be expected. However, most genes having TSS-proximal
DMRs were not regulated by the infection at any time-point, regardless of the context (Figure
6A). Cases of negative correlation between TSS-proximal methylation and expression,
though, were observed, especially in the CpG context at 12 dpi (Figure 6A, Table S6). The
observed correlations were mainly linked to TSS-proximal hypermethylation and repression of
gene expression, although few cases in the opposite direction were also observed (Table S6).
These genes were classified according to functional categories. Genes related to RNA
metabolism (biosynthesis and processing) and protein metabolism (modification and
translocation) were the most predominant ones (Figure 6B). Genes related to amino acid,
carbohydrate, coenzyme, lipid, nucleotide, and secondary metabolism were also enriched.
Despite being one of the most responsive in the transcriptome, few stress-related genes had
negative correlation with DMRs (Figure 6B, Table S6).
Since CHG and CHH are the major transposon-associated methylation marks and that
variation in their patterns can influence the expression of nearby genes (Sigman and Slotkin,
2016), we also sought for cases where elements with non-CpG DMRs were close to virus-
induced DEGs. At 5 dpi, the vast majority of methyl-regulated TEs were further than 10 kb
from DEGs, indicating that either their regulation did not influence expression of nearby genes
or that they were located outside of gene-rich areas (Figure 6C). A larger number of regulated
TEs close to DEGs was observed at 12 dpi, although elements far from DEGs were still the
predominant type (Figure 6C). In both time points, there was no clear general correlation
between the state of TE regulation (hyper- or hypo-methylated) and expression direction
(induced or repressed) of nearby genes, probably reflecting the dynamic changes in their
control along the infection time course. At 12, about 80 DEGs, including PTI- and ETI-related
genes, were found to be close to regulated TEs in both isolates (Table S7). There were also
isolate-specific cases: about 150 DEGs were close to regulated TEs in the ancestral and other
300 in the evolved TuMV. Abiotic and biotic stress-related genes were also found among the
isolate-specific TE-close DEGs (Table S7).
DISCUSSION
In this study we have used different approaches to evaluate the impact of epigenetic factors
in triggering stress responses against viruses in plants. Infection experiments in epigenetic-
deficient mutants indicated that RdDM factors, including AGO4, RDR2, POLIV, POLV and
DMR1/2, the chromatin remodeler DDM1 and the histone modification proteins IBM1 and
JMJ14 can control responses against TuMV infection (Figure 4). RdDM-, DDM1- and JMJ14-
deficient plants showed resistance against the virus, while ibm1 mutants were more
susceptible. This agrees with experiments performed in inflorescence of other arabidopsis
epigenetic mutants (drm1 drm2, drm1, drm2, cmt3, and ros1) infected with a tobravirus
(Diezma-Navas et al., 2019). Other studies have also associated loss of DNA methylation
factors with increased resistance against non-viral biotrophic pathogens, but susceptibility to
necrotrophic ones (Dowen et al., 2012; Le et al., 2014; López et al., 2011; López Sánchez et
al., 2016; Luna et al., 2012; Yu et al., 2013). Biotrophic pathogens are thought to be targeted
mainly by SA-mediated defenses and several genes related to this pathway, including PR1
are induced in different RdDM or other DNA (de)methylation mutants (Agorio and Vera, 2007;
Diezma-Navas et al., 2019; Dowen et al., 2012; López et al., 2011; López Sánchez et al.,
2016; Yu et al., 2013). Necrotrophic pathogens, on the other hand, are controlled by JA
defense pathways, repressed in those mutants (López et al., 2011). Since our transcriptome
data evidenced that SA signaling might be important for TuMV responses (Figure S2D and
2C), the general induction of SA-mediated defense pathways in the hypo-methylated mutants
may be one of the mechanistic explanations of their resistance to the virus. The observed
RdDM effects, however, may not be universal for plant viruses, as ago4 mutants have been
shown to be more susceptible to a tobravirus at late infection stages (Diezma-Navas et al.,
2019; Ma et al., 2015). Misexpression of defense genes and changes in resistance have also
been observed in histone modification mutants (Zhu et al., 2016). The genes tested here,
IBM1 and JMJ14, have antagonistic roles in expression regulation. In ibm1 mutants,
thousands of genes are known to gain TE-related repressive marks (Miura et al., 2009). The
increased heterochromatin in this genotype therefore may possibly prevent or delay the
expression of defense genes, promoting the observed susceptibility to TuMV. In contrast,
JMJ14 removes H3K4 active marks from TEs and euchromatin-related marks are increased
in the mutant (Lu et al., 2010; Searle et al., 2010; Yang et al., 2010). Moreover, RdDM is
partially deficient in the absence of the protein (Greenberg et al., 2013). Defense genes may
be therefore more primed in jmj14 than in wild-type plants, corroborating the observed
increased resistance to the virus.
As observed in other types of stresses, differences in methylation patterns in and around
genes and transposons due to TuMV were observed in infected wild-type arabidopsis plants.
The downregulation of some RdDM factors due to the stress (Figure 3A), together with other
factors, including competition with nearby transcriptional machinery, host or viral small RNAs
and viral silencing suppressor proteins, may have contributed to the observed methylation
differences. Most of the DMRs were in the CpG context and mapped inside or around the
transcription start site of protein-coding genes (Figure 5A, 5B and S4A). Transcription of genes
having DMRs around their TSS, however, was largely not affected by the virus (Figure 6A).
Absence of a significant correlation between promoter proximal CpG methylation and
expression were also found in arabidopsis and other plants exposed to stress or in natural
populations (Lafon-Placette et al., 2018; Mager and Ludewig, 2018; Narsai et al., 2017;
Seymour and Becker, 2017; Seymour et al., 2014; Sun et al., 2019; Xu et al., 2018). Contrary
to other plants, about only 5% of the arabidopsis genes are thought to be regulated by
promoter methylation (Zhang et al., 2006). Furthermore, it has been shown that methylation
differences in the plant are higher between tissues than in stress conditions (Seymour et al.,
2014). The dilution effect produced by using whole leaves, with different cell types and likely
varying viral loads, probably precluded the identification of general correlation between
promoter methylation and expression. TuMV-induced genes with negative methylation and
expression, though, were observed. They belonged to several functional categories and genes
related to RNA or protein metabolism were the most frequent ones (Figure 6B, Table S6). Few
biotic stress-related genes were found to have inverted correlations, contrary to what was
observed for tobacco plants infected with cucumber mosaic virus (Wang et al., 2018). Among
the identified stress-related genes, SOMATIC EMBRYOGENESIS RECEPTOR KINASE 1
(SERK1) have already been linked to antiviral defense by channeling dsRNAs into PTI
pathways (Niehl et al., 2016). And the HEAT SHOCK PROTEIN 22 has been associated with
plant memory to cycles of heat stress (Stief et al., 2014). Those correlations should be
interpreted carefully, since it is still not clear how much methylation difference is in fact
required to promote significant transcriptional changes.
TEs known to be located in both euchromatic and heterochromatic environments,
including centromere core, also presented differences in CHG and CHH marks, indicating a
widespread deregulation of methylation machinery (Figure S4B). At 2 - 5 dpi, similar amounts
of hyper- and hypo-methylation were observed in transposons, but a more pronounced
hypermethylation of those elements was observed at 12 dpi (Figures 5A, 5B and S4B).
Accordingly, TEs were found to be generally repressed at 12 dpi, especially against the
evolved isolate (Figure 3B). This agrees with the observed repression of several TEs in
arabidopsis plants infected with a DNA geminivirus (Coursey et al., 2018). Results are also in
line with models predicting that early abiotic stress responses may trigger transient
hypomethylation of TEs due to the overexpression of responsive genes, that can be reversed
by their hypermethylation at a later time-point (Secco et al., 2015). The higher stress intensity
promoted by the evolved isolate may have contributed to a faster regulation shift, explaining
the increased numbers of hypermethylated and downregulated TEs at 12 dpi. Case-specific
exceptions to the model were observed (Figure 3B). For example, the RdDM-targeted
transposons AtGP1 and AtCOPIA93 were induced by TuMV infection at 12 dpi (Mirouze et al.,
2009; Yu et al., 2013). A short version of the AtCOPIA93 element, also known as EVADÉ, has
been shown to be required for the expression of the NLR gene RECOGNITION OF
PERONOSPORA PARASITICA 4 (RPP4) (Zervudacki et al., 2018), a gene that was induced
by both TuMV isolates (Table S1). Although the epigenetic regulation of TEs is reported to
regulate expression of nearby genes, there was no clear general correlation between the state
of TuMV-induced TE regulation (hyper- or hypo-methylated) and the type of nearby gene
regulation (induced or repressed). This lack of correlation may reflect the dynamic changes
along the infection time-course (Figure 6C), but can also be due to the small and
heterochromatin-poor arabidopsis genome. In fact, DNA methylation mutants in several plants
are lethal or severely compromise development, while most of them show light or no
phenotype in arabidopsis (Zhang et al., 2018). There was also little correlation between TE
methyl regulation and expression during the infection, in agreement with studies showing that
their induction under heat stress or virus infection is not associated with DNA methylation
changes (Diezma-Navas et al., 2019; Pecinka et al., 2010). Nonetheless, some DEGs that
were close to TEs having non-CpG DMRs were detected, indicating a possible co-regulation
mechanism. Some of them were similarly regulated by both TuMV isolates, including genes
involved with disease resistance, transcriptional factors, RNA silencing and histone variants
involved with stress responses. Important transcriptional and disease regulators were also
found among isolate-specific cases (Table S7). As for promoter methylation differences,
reported correlations should be carefully interpreted, as extra experimental approaches should
be applied to confirm their causal relationships.
Apart from epigenetic factors and known RNA silencing responses, the transcriptome
data also indicated that several other types of defense mechanisms were mounted against
TuMV, including general shut-down of photosynthesis, metabolic rearrangements and
induction of genes related to all known immunity pathways in plants (Figure 2D, S2C and
S2D). The induction of several genes related to basal immunity systems based on molecular
patterns (PTI) and elicitors (ETI) are in line with the increasing evidence suggesting that those
types of innate defenses with conserved animal counterparts have also roles against viruses
in plants (Teixeira et al., 2019). A possible role of SA in triggering defense responses was
corroborated by the induction of some of its well characterized activators or targets at 12 dpi
(Figure S2D, Table S1, Files S1 and S2). However, the high expression of several known SA-
antagonistic genes in various time points, including WRKY25/26/33/38/62 (Kc et al., 2008;
Zheng et al., 2006, 2007), indicates a possible viral counter-defense strategy. The induction
of anti-SA genes was particularly high for the evolved TuMV strain (Figure S3A and Table S3).
Although only two SNPs in the region coding for the viral multifunction protein VPg were
detected, several lines of evidence demonstrated that the isolate had a higher virulence than
the ancestral stock to arabidopsis plants. Integrating the observed methylomes and
transcriptomes with virus-host protein-protein interaction networks for both isolates will be a
valuable way to find the molecular basis of the adaptive process.
Viral fitness is a complex parameter often used by virologists to quantitatively describe
the reproductive ability and evolutionary potential of a virus in a particular host. As cellular
parasites, viruses utilize all sorts of cellular factors, reprogram gene expression patterns into
their own benefit, and block and interfere with cellular defenses. All these processes take
place in the host complex network of intertwined interactions and regulations. Interacting in
suboptimal ways with any of such elements may have profound effects in the progression of
infection and therefore in viral fitness; inefficient interactions may result in attenuated or even
abortive infections. Despite its relevance, how virus evolution shapes and optimizes these
interactions has received scant attention. In previous experimental evolution studies in which
tobacco etch potyvirus was adapted to different ecotypes of arabidopsis, it was shown that the
transcriptome of the infected plants was differentially affected depending on the degree of
adaptation of the virus and identified potential host drivers of virus adaptation (Agudelo-
Romero et al., 2008; Hillung et al., 2016). Here, we expand these previous studies to
incorporate a new level of regulation of gene expression: DNA and histones epigenetic
modifications. Our results suggest that TuMV isolates that differ in their degree of adaptation
to arabidopsis may exert a differential effect on methylation patterns and in the expression of
genes epigenetically regulated.
ACKNOWLEDGMENTS
We thank Francisca de la Iglesia for technical assistance and for performing the TuMV
evolution experiments in Valencia and Pawel Baster and James Barlow for technical support
in Cambridge. We are grateful to Dr. César Llave and Dr. Virginia Ruiz-Ferrer for providing
the ago4, rdr2, drm1, and drm2 seeds and to Dr. R. Keith Slotkin for the ddm1 ones. This work
was supported by Spain Agencia Estatal de Investigación - FEDER grants BFU2015-65037-
P (S.F.E.) and AGL2016-79825-R (G.G.) and by Generalitat Valenciana grant
PROMETEU/2019/012 (S.F.E). R.L.C was supported by a fellowship from the Brazilian
funding agency CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico
Brasil) for the stay in Valencia and from EMBO/EuropaBio for the stay in Cambridge.
Author Contributions
R.L.C., G.G., D.C.B., and S.F.E. conceived and designed the experiments; R.L.C. and Z.K.
performed infection experiments; R.L.C., A.S.C. and S.Y.M. processed and analyzed the
transcriptome and methylome data; R.L.C. and S.F.E. performed statistical analysis; R.L.C.,
A.S.C., S.Y.M., S.L.G., G.G., D.C.B., and S.F.E. analyzed and interpreted the data; G.G.,
D.C.B. and S.F.E. contributed with reagents/materials/analysis tools; R.L.C. and S.F.E. wrote
the paper.
Star Methods
Contact for Reagent and Resource Sharing
Further information and requests for resources and reagents should be directed to and will be
fulfilled by the Lead Contact, Régis L. Corrêa (regislcorrea@ufrj.br).
Experimental Model and Subject Details
Plant genotypes
For the experimental evolution and infection time-course analysis (all mRNA-seq and WGBS-
seq data), wild-type Arabidopsis thaliana L. plants from the Col-0 ecotype were grown on short
day conditions, i.e., with 8 h of light at 25 °C and 16 h in the dark at 20 °C.
Mutant arabidopsis genotypes were maintained and infected on long day conditions, i.e.,
with 16 h of light at 24 °C and 8 h in the dark at 20 °C. The lines nrpD1a-3 (SALK_128428),
nrpE1 (SALK_017795C), ibm1-4 (SALK_035608C), and jmj14 (SALK_135712C) were
obtained from the Nottingham Arabidopsis Stock Centre (NASC). Lines rdr2-1
(SAIL_1277_H08), drm1-2 drm2-2 and ago4 were kindly provided by Dr. César Llave and
ddm1-2 by Dr. Keith Slotkin. Oligonucleotides used for genotyping are listed in Table S8.
Primers for ddm1-2 and nrpD1a-3 were described elsewhere (Herr et al., 2005; Yadegari et
al., 2000). All mutant genotypes were in the Col-0 background.
Virus isolates
The TuMV isolate YC5 (GenBank, AF530055.2) cloned under the 35S promoter and NOS
terminator originally obtained from calla lily (Zantedeschia sp.) was used as source of virus
inoculum (Chen et al., 2003). The virus was maintained in Nicotiana benthamiana Domin
plants before being inoculated into arabidopsis plants.
Method Details
Evolution experiments
Initial inoculum came from N. benthamiana leaf tissues infected with the YC5 TuMV isolate.
Infected leaves were ground in liquid nitrogen and 100 mg of fine powder mixed with a solution
containing 50 mM phosphate buffer (pH 7), 3% polyethylene glycol 6000 and 10%
Carborundum at 100 mg/ml (diluted in the same PEG/phosphate buffer). Two leaves of 20
arabidopsis plants (5 weeks old) were mechanically inoculated with 5 μL of the sap.
Arabidopsis plants having clear TuMV symptoms at 10 dpi were pooled and used as source
of inocula as described before. A total of 10 passages of this kind were performed.
Survival analysis was done with the survfit function from the survival R package to
compute Kaplan-Meier estimates. Time of “death” was scored when plants were having strong
leaf yellowing symptoms (Figure S1A). Plots were generated with the survminer R package.
R version 3.4.4 in RStudio was used for these analyses.
Infection experiments in wild-type and mutant plants
For the time-course experiments (used for all mRNA-seq and WGBS-seq data), batches of
Col-0 plants were mechanically inoculated with two inocula sources: coming from
benthamiana (as above) and the 9th passage-infected arabidopsis tissues (ancestral and
evolved TuMV, respectively). To ensure that even viral loads were used, concentration of viral
transcripts in both inocula were measured by standard curve RT-qPCR. Total RNA from health
and TuMV-infected arabidopsis and benthamiana plants were extracted using Plant Isolation
Mini Kit (Agilent). Standard curves were constructed from eight serial dilutions of in vitro-
transcribed TuMV RNA using the mMESSAGE mMACHINE® SP6 Transcription Kit (Ambion).
Each of the 5-fold dilutions were done by mixing viral transcripts with total RNA extracted from
health tissues of arabidopsis or benthamiana for taking any PCR inhibitors into account. The
20 μL reactions were performed in an ABI StepOnePlus real-time PCR system (Applied
Biosystems), using the GoTaq 1-Step RT-qPCR system (Promega). Cycling conditions were
as follows: one cycle of 42 °C for 5 min and 95 °C for 10 min; 40 cycles of 95 °C for 5 min and
60 °C for 34 s; and one cycle of 60 °C for 1 min, followed by a melting curve from 60 °C to
95 °C, with 0.3 °C increments. Primers TuMV F117_F and TuMV F118_R used to amplify the
viral capsid coding region are described in Table S8. Results for arabidopsis and benthamiana
were analyzed separately, with their corresponding viral serial dilutions. Inoculations were
performed as described above, but using adjusted tissue amounts from each plant source in
order to have even inocula. Non-inoculated leaves of mock, ancestral or evolved-infected
plants were collected at 2, 5 and 12 dpi and kept frozen at −70 °C until nucleic acid extraction.
Plants collected at 2 and 5 dpi were symptomless. To be sure that the inoculation worked,
they were left alive until the end of the time-course after leave sampling. Only frozen tissues
from plants showing clear symptoms at later stages of infection were further analyzed.
Arabidopsis mutant lines were grown on long day conditions, as described above.
Three-week old plants were infected with adjusted amount of TuMV-infected benthamiana or
arabidopsis (9th passage) saps, as described for wild-type plants. Individual plants were
scored daily for typical TuMV symptoms.
Nucleic acid extraction and library preparation
Total DNA and RNA from TuMV-infected and healthy Col-0 plants were co-extracted using
the protocol described in (Oliveira et al., 2015), with two phenol-chloroform extractions before
lithium precipitation. The quality of the RNAs used for preparing mRNA-seq libraries were
checked with the Bioanalyzer nano kit and quantified with the Qubit RNA BR Assay Kit
(ThermoFisher). Libraries were prepared with the True-seq Stranded mRNA prep kit
(Illumina), using 1 μg of total RNA as input. In total, 24 libraries were prepared, containing
three biological replicates for each of the conditions. Each biological replicate was made by
total RNA from individual plant systemic leaves. Libraries were sequenced with the Illumina
High Output Kit v2 (2 × 75 bp) in a NextSeq 500 benchtop machine (Illumina).
DNAs (100 ng) were bisulfite-treated with the EZ DNA Methylation Gold kit (Zymo
Research), before library preparation with the TruSeq DNA Methylation Kit (Illumina). In total,
18 libraries were prepared, containing two biological replicates for each condition. Each
biological replicate was made by a pool of DNAs extracted from systemic leaves of three
plants. Libraries were sequenced with the High Output Kit v2 (1 × 75 bp) in a NextSeq 550
benchtop machine (Illumina).
mRNA-seq analysis
The quality of the mRNA-seq libraries was checked with FastQC v0.11.7 (https://github.com/s-
andrews/FastQC)
and
trimmed
with
TrimGalore
v0.4.4
(https://github.com/FelixKrueger/TrimGalore), using cutadapt v1.3 (Martin, 2011). Twelve
bases from the 5’ end of reads 1 and 2 were removed before mapping with HiSat2 v2.1.0
(Pertea et al., 2016) to the Ensembl release 39 of the A. thaliana TAIR10 genome assembly.
For viral genome SNP calling, trimmed reads were mapped with HiSat2 to the TuMV isolate
YC5 (GenBank, AF530055.2) with a modified minimum score parameter (L, 0 -0.8) to allow
more mismatches. Resulting SAM files were BAM-converted, sorted, indexed and analyzed
with SAMtools v1.9 (Li et al., 2009). SNP calling was performed using bcftools v1.6 by first
using the “mpileup” subroutine (with default parameters apart from -d10000) followed by the
“call” subroutine as well as the “filter” subroutine filtering out low quality calls (<10). Read
counting in features was done with htseq-count, using The Arabidopsis Reference Transcript
Dataset (AtRTD2) (Zhang et al., 2017) as input annotation file. Differential expression analysis
was done with DESeq2 v1.18.1 (Love et al., 2014), considering only genes having a total of
at least 10 reads for each pairwise comparison. Functional characterization of DEGs was done
with plant GOSilm implemented in the Cytoscape plugin Bingo (Maere et al., 2005) and
MapMan (Thimm et al., 2004). For the analysis of differentially expressed transposons, the
TEtranscripts tool was used (Jin et al., 2015). Trimmed reads were mapped with STAR (Dobin
et al., 2013) to the Ensembl release 39 of the A. thaliana TAIR10 genome assembly. The
arabidopsis
transposon
annotation
file
from
TEtranscripts
(http://labshare.cshl.edu/shares/mhammelllab/www-
data/TEToolkit/TE_GTF/TAIR10_TE.gtf.gz) was used as input to the program.
RT-qPCRs
For RT-qPCRs, 1 μg of Turbo DNAse (ThermoFisher)-treated total RNAs were reverse-
transcribed with Superscript IV (ThermoFisher) with random hexamer primers and used for
amplification in a 10 μL reaction with the Luna® Universal qPCR Master Mix (New England
Biolabs). Oligonucleotides used are listed in Table S8. Amplifications were done in a CFX96
machine (Bio-Rad) with the following cycling conditions: one cycle of 95 °C for 1 min; 40 cycles
of 95 °C for 15 s and 60 °C for 30 s; and one cycle of 60 °C for 1 min, followed by a melting
curve from 60 °C to 95 °C. Reaction efficiencies and the fractional cycle number at threshold
were calculated based on raw fluorescence with the Miner tool (Zhao and Fernald, 2005).
Transcripts were quantified by the comparative ΔΔCT method, and previously known
arabidopsis stable genes PROTEIN PHOSPHATASE 2A SUBUNIT A3 (AT1G13320) and
SAND (AT2G28390) were used as endogenous references (Czechowski et al., 2005). Primer
sequences are described in Table S8. Primers for ROS1 amplification were described
elsewhere (Lei et al., 2015).
Relative gene expression data were fitted to generalized linear mixed models (GLMM)
using plant genotypes and viral inocula as orthogonal factors. A Gamma probability distribution
and a logarithm-link function were chosen based on the minimum Bayes information criterion.
For testing differences among specific samples, post hoc pairwise comparisons with
sequential Bonferroni correction tests were used. These analyses were performed with SPSS
version 26 (IBM Corp.).
WGBS-seq analysis
The quality of the WGBS libraries was checked with FastQC v0.11.7 (https://github.com/s-
andrews/FastQC) and trimmed with cutadapt v1.16 (Martin, 2011). The first nine initial and
two last bases from reads were removed, and remaining ends with qscore lower than 30 were
also trimmed. Reads having less than 20 bases after trimming were also discarded. Mapping
was performed with Bismark - Bisulfite Mapper v0.20.0 (Krueger and Andrews, 2011), using
the Ensembl release 39 of the TAIR10 genome assembly. Removal of PCR duplicates, sorting
and indexing of the resulting BAM files was done with SAMtools v1.9 (Li et al., 2009).
Methylation call extraction and differential analysis were performed with the Methylkit R
package v1.4.1 (Akalin et al., 2012). For each pairwise comparisons (mock vs ancestral TuMV
and mock vs evolved TuMV, for each time-point), bases with low (below 10×) and more than
99.9th percentile of coverage in each sample were discarded before mean read normalization.
Only bases covered in all samples from each pairwise comparisons were further analyzed.
Methylation difference was tested with logistic regression and P-values were adjusted to q-
values with the SLIM method. Differentially methylated regions in 100 bp tiles having q < 0.05
and methylation difference larger than 15% were selected. Assignment of each DMR to
features was done with the GenomicFeatures v1.30.3 package (Lawrence et al., 2013).
Annotation files were obtained from AtRTD2 (Zhang et al., 2017) and TEtranscripts tool (Jin
et al., 2015). Bisulfite non-conversion rates were calculated by mapping reads to arabidopsis
cytoplasmic genomes.
Quantification and Statistical Analysis
General statistical analysis
Specific statistical tests used for each experiment were detailed in Figure Legends and
described in the Method Details section of the Star Methods as needed.
Data and Software Availability
The mRNA-seq and WGBS-seq data have been deposited to the SRA database under ID
codes PRJNA545306 and PRJNA545300, respectively.
Figure titles and legends
Figure 1. Experimental evolution of TuMV in arabidopsis
(A) A TuMV stock originally obtained from calla lily and subsequently maintained in N.
benthamiana plants was used as a source of virus for an evolution experiment by serial
passages in batches of arabidopsis wild-type plants. The first and 10th passages were the
ancestral and evolved isolates used in all experiments, respectively. (B) Symptom severity
associated with ancestral and evolved TuMV isolates from 7 to 17 dpi, according to the scale
defined in Figure S1A. Violin plots represent the symptoms severity level of each of the 20
plants infected with the different isolates. Lines represent the median severity value in each
time-point. (C) Number of days each plant (dot) took to reach strong symptoms (symptom
level 3, according to the scale provided in Figure S1A) after TuMV inoculation. Student two-
samples t tests, ***P < 0.001; **P < 0.01; NS., not significant.
Figure 2. Transcriptome responses to TuMV
(A) Number of DEGs obtained by DESeq2 analysis for each TuMV infection condition
(adjusted P < 0.05). In each time-point, three biological replicates infected with either the
ancestral or evolved TuMV isolate were compared to mock-inoculated ones. (B) Gene
ontology analysis (plant GOSlim) for DEGs between the evolved and ancestral TuMV isolates.
For highlighting the differences between the isolates, TuMV ancestral- and evolved-infected
samples were used as control and treatment in the DESeq2 analysis, respectively. Circle size
represents level of enrichment and color heat maps indicate adjusted P values (padjv). (C)
Transcriptional profiles (log2 fold change) of selected phytohormone genes, including abscisic
acid (ABA), ethylene (ET), jasmonic acid (JA) and salicylic acid (SA). In the left and central
panels both virus isolates were compared against mocks. In the right panel, the evolved isolate
was directly compared against the ancestral one. (D) Transcriptional profiles (log2 fold change)
of selected innate immunity genes, including PAMP-triggered immunity (PTI), effector-
triggered immunity (ETI) and pathogenesis-related (PR) genes. As above, samples from
evolved isolate-infected plants were directly compared against the ancestral isolate-infected
plants in the right panel. dpi: days post-inoculation.
Figure 3. Transcriptional profiles of epigenetic-related selected genes and transposons
(A) Transcriptional profiles (log2 fold change) of selected RNA silencing (yellow lines) and DNA
methylation genes (grey lines). (B) Heat map with fold changes of differentially expressed
transposons (adjusted P < 0.05) obtained with the TEtranscripts tool. dpi: days post-
inoculation.
Figure 4. TuMV infection in epigenetic mutants
Number of days each plant (dot) took to reach strong symptoms after TuMV inoculation in
epigenetic mutants, compared to Col-0 wild-type plants. (A) Panel of selected RdDM mutants.
(B) Chromatin remodeler ddm1 mutant. (C) Histone modification mutants. In all panels, the
variable days to strong symptoms was fitted to generalized linear mixed models (GLMM) with
plant (as indicated in the corresponding abscissa axes) and virus genotypes (ancestral and
evolved) as orthogonal factors; a Normal probability distribution and an identity-link function
were assumed. Post hoc pairwise comparisons with sequential Bonferroni correction tests
were performed; ****P < 0.001; ***P < 0.01; **P < 0.05; *P < 0.1; NS., not significant.
Figure 5. Whole-genome bisulfite sequencing (WGBS) of infected wild-type plants
Number of hyper- or hypo-methylated differentially methylated regions (DMRs) in the three
cytosine contexts (CpG, CHG and CHH) proximal to transcriptional start sites (TSS-prox),
inside genes (GbM) or TEs. (A) Ancestral TuMV-infected plants. (B) Evolved TuMV-infected
plants.
Figure 6. Correlation between TuMV-induced methylation and expression
(A) Number of genes having differentially methylated regions (DMRs) proximal to
transcriptional start sites (TSS-prox) that were found to be regulated at the transcriptional
level. (B) Functional characterization based on MapMan bins of genes having negative
correlation between TSS-prox methylation and expression. (C) Percentage of TEs with hyper-
or hypo-methylated non-CpG DMRs that are close (up to 10 kb) or far from DEGs.
Supplemental Information titles and legends
Figure S1. Biological and molecular differences between the ancestral and evolved
TuMV isolates
(A) Categories of observed symptoms from 11 to 13 dpi. (B) Estimation by RT-qPCR of TuMV
accumulation along the infection time-course. Student’s two-samples t-tests, ***P < 0.001; **P
< 0.01; NS., not significant. (C) Kaplan-Meier survival regression analysis of TuMV-infected
wild-type plants. Analysis based on the time each plant took to reach strong symptoms. (D)
Predicted structures of the ancestral (blue) and evolved (red) VPg proteins. Altered regions
were highlighted in grey. (E) Amino acid sequence alignment between the predicted VPg
regions of the ancestral and evolved TuMV isolates. Shared residues are highlighted by red
dots.
Figure S2. Transcriptome responses to TuMV
(A) Number of mapped reads to the plant or virus genomes at 2, 5 and 12 dpi. (B) Upset plot
with numbers of shared DEGs in each condition. (C) Gene ontology analysis (plant GOSlim)
for the identified DEGs. Circle size represents level of enrichment. (D) Visualization of cluster
40 of the AtCKN Arabidopsis network in response to the evolved TuMV at 2, 5 and 12 dpi.
Figure S3. Transcriptional profiles of biotic- and abiotic-related genes
(A) Transcriptional profiles of transcription factor genes. (B) RT-qPCR confirmation of biotic,
abiotic and development genes in TuMV-infected plants at 12 dpi. Relative gene expression
data were fitted to a generalized linear mixed model (GLMM) with plant genotype (PR1, WAK1,
HSP70, and COR15a) and source of inocula (mock, ancestral and evolved viruses)
incorporated as orthogonal factors. (C) RT-qPCR confirmation of IBM1 and ROS1 genes in
TuMV-infected plants at 12 dpi. Relative gene expression data were fitted to a GLMM with
plant genotype (IBM1 and ROS1) and source of inocula (mock, ancestral and evolved viruses)
incorporated as orthogonal factors. (B) and (C) a Gamma probability distribution and a
logarithm-link function were assumed. Post hoc pairwise comparisons with sequential
Bonferroni correction tests were performed; ***P < 0.001; **P < 0.05; *P < 0.1; NS., not
significant.
Figure S4. Whole-genome bisulfite sequencing (WGBS) of infected wild-type plants
Number of hyper- or hypo-methylated DMRs in the three cytosine contexts (CpG, CHG and
CHH) are presented for each condition. (A) Total number of DMRs found in the genome. (C)
Number of DMRs in selected TE families.
Table S1. List of all DEGs identified in the DESeq2 analysis (adjusted P < 0.05) for each
TuMV infection condition.
Table S2. List of DEGs regulated by all TuMV infection conditions. Related to Figure S2B.
Table S3. List of selected DEGs regulated by TuMV infection. Related to Figures 3A, S3A,
S3B and S3D.
Table S4. List of differentially expressed transposons obtained with TEtranscripts
(adjusted P < 0.05) for each TuMV infection condition.
Table S5. Genomic ranges and values for all identified DMRs in each TuMV infection
condition.
Table S6. Genes having negative correlation between transcriptional start site proximal
methylation and expression. Related to Figures 6A and 6B.
Table S7. DEGs close to TEs with non-CpG DMRs at 12 dpi. Related to Figure 6C.
Table S8. List of primers used in this study.
File S1. Dynamic visualization of cluster 40 of the AtCKN Arabidopsis network in
response to the ancestral TuMV isolate at 5 and 12 dpi.
File S2. Dynamic visualization of cluster 40 of the AtCKN Arabidopsis network in
response to the evolved TuMV isolate at 2, 5 and 12 dpi.
REFERENCES
Agorio, A., and Vera, P. (2007). ARGONAUTE4 is required for resistance to Pseudomonas
syringae in Arabidopsis. Plant Cell 19, 3778–3790.
Agudelo-Romero, P., Carbonell, P., and Perez-Amador, M.A. (2008). Virus adaptation by
manipulation of host’s gene expression. PLoS One 3, e2397.
Akalin, A., Kormaksson, M., Li, S., Garrett-Bakelman, F.E., Figueroa, M.E., Melnick, A., and
Mason, C.E. (2012). methylKit: a comprehensive R package for the analysis of genome-wide
DNA methylation profiles. Genome Biol. 13, R87.
Andino, R., and Domingo, E. (2015). Viral quasispecies. Virology 479-480, 46–51.
Borges, F., and Martienssen, R.A. (2015). The expanding world of small RNAs in plants.
Nat. Rev. Mol. Cell Biol. 16, 727–741.
Boutrot, F., and Zipfel, C. (2017). Function, Discovery, and Exploitation of Plant Pattern
Recognition Receptors for Broad-Spectrum Disease Resistance. Annu. Rev. Phytopathol.
55, 257–286.
Cervera, H., Ambrós, S., and Bernet, G.P. (2018). Viral Fitness Correlates with the
Magnitude and Direction of the Perturbation Induced in the Host’s Transcriptome: The
Tobacco Etch Potyvirus—Tobacco Case Study. Mol. Biol. Evol. 35, 1599-1615.
Chen, C.C., Chao, C.H., Chen, C.C., Yeh, S.D., Tsai, H.T., and Chang, C.A. (2003).
Identification of Turnip mosaic virus Isolates Causing Yellow Stripe and Spot on Calla Lily.
Plant Dis. 87, 901–905.
Cheng, X., and Wang, A. (2017). The Potyvirus Silencing Suppressor Protein VPg Mediates
Degradation of SGS3 via Ubiquitination and Autophagy Pathways. J. Virol. 91, e01478-16.
Coursey, T., Regedanz, E., and Bisaro, D.M. (2018). Arabidopsis RNA Polymerase V
Mediates Enhanced Compaction and Silencing of Geminivirus and Transposon Chromatin
during Host Recovery from Infection. J. Virol. 92, e01320-17.
Cui, H., Tsuda, K., and Parker, J.E. (2015). Effector-triggered immunity: from pathogen
perception to robust defense. Annu. Rev. Plant Biol. 66, 487–511.
Czechowski, T., Stitt, M., Altmann, T., Udvardi, M.K., and Scheible, W.-R. (2005). Genome-
wide identification and testing of superior reference genes for transcript normalization in
Arabidopsis. Plant Physiol. 139, 5–17.
Diezma-Navas, L., Pérez-González, A., Artaza, H., Alonso, L., Caro, E., Llave, C., and Ruiz-
Ferrer, V. (2019). Crosstalk between epigenetic silencing and infection by tobacco rattle
virus in Arabidopsis. Mol. Plant Pathol. 20, 1439-1452.
Dobin, A., Davis, C.A., Schlesinger, F., Drenkow, J., Zaleski, C., Jha, S., Batut, P.,
Chaisson, M., and Gingeras, T.R. (2013). STAR: ultrafast universal RNA-seq aligner.
Bioinformatics 29, 15–21.
Dowen, R.H., Pelizzola, M., Schmitz, R.J., Lister, R., Dowen, J.M., Nery, J.R., Dixon, J.E.,
and Ecker, J.R. (2012). Widespread dynamic DNA methylation in response to biotic stress.
Proc. Natl. Acad. Sci. U. S. A. 109, E2183–E2191.
Fu, Z.Q., and Dong, X. (2013). Systemic acquired resistance: turning local infection into
global defense. Annu. Rev. Plant Biol. 64, 839–863.
Greenberg, M.V.C., Deleris, A., Hale, C.J., Liu, A., Feng, S., and Jacobsen, S.E. (2013).
Interplay between active chromatin marks and RNA-directed DNA methylation in Arabidopsis
thaliana. PLoS Genet. 9, e1003946.
Herr, A.J., Jensen, M.B., Dalmay, T., and Baulcombe, D.C. (2005). RNA polymerase IV
directs silencing of endogenous DNA. Science 308, 118–120.
Hillung, J., García-García, F., Dopazo, J., Cuevas, J.M., and Elena, S.F. (2016). The
transcriptomics of an experimentally evolved plant-virus interaction. Sci. Rep. 6, 24901.
Ivanov, K.I., Eskelin, K., Lõhmus, A., and Mäkinen, K. (2014). Molecular and cellular
mechanisms underlying potyvirus infection. J. Gen. Virol. 95, 1415–1429.
Jin, Y., Tam, O.H., Paniagua, E., and Hammell, M. (2015). TEtranscripts: a package for
including transposable elements in differential expression analysis of RNA-seq datasets.
Bioinformatics 31, 3593–3599.
Kc, K., Lai, Z., Fan, B., Chen, Z., and Kim, K.-C. (2008). Arabidopsis WRKY38 and WRKY62
transcription factors interact with histone deacetylase 19 in basal defense. Plant Cell 20, 1–
2.
Krueger, F., and Andrews, S.R. (2011). Bismark: a flexible aligner and methylation caller for
Bisulfite-Seq applications. Bioinformatics 27, 1571–1572.
Lafon-Placette, C., Le Gac, A.-L., Chauveau, D., Segura, V., Delaunay, A., Lesage-
Descauses, M.-C., Hummel, I., Cohen, D., Jesson, B., Le Thiec, D., et al. (2018). Changes
in the epigenome and transcriptome of the poplar shoot apical meristem in response to
water availability affect preferentially hormone pathways. J. Exp. Bot. 69, 537–551.
Lawrence, M., Huber, W., Pagès, H., Aboyoun, P., Carlson, M., Gentleman, R., Morgan,
M.T., and Carey, V.J. (2013). Software for computing and annotating genomic ranges. PLoS
Comput. Biol. 9, e1003118.
Le, T.-N., Schumann, U., Smith, N.A., Tiwari, S., Au, P.C.K., Zhu, Q.-H., Taylor, J.M., Kazan,
K., Llewellyn, D.J., Zhang, R., et al. (2014). DNA demethylases target promoter transposable
elements to positively regulate stress responsive genes in Arabidopsis. Genome Biol. 15,
458.
Lei, M., Zhang, H., Julian, R., Tang, K., Xie, S., and Zhu, J.-K. (2015). Regulatory link
between DNA methylation and active demethylation in Arabidopsis. Proc. Natl. Acad. Sci. U.
S. A. 112, 3553–3557.
Li, H., Handsaker, B., Wysoker, A., Fennell, T., Ruan, J., Homer, N., Marth, G., Abecasis,
G., and Durbin, R. (2009). The Sequence Alignment/Map format and SAMtools.
Bioinformatics 25, 2078–2079.
Li, J., Brader, G., and Palva, E.T. (2004). The WRKY70 transcription factor: a node of
convergence for jasmonate-mediated and salicylate-mediated signals in plant defense. Plant
Cell 16, 319–331.
López, A., Ramírez, V., García-Andrade, J., Flors, V., and Vera, P. (2011). The RNA
silencing enzyme RNA polymerase V is required for plant immunity. PLoS Genet. 7,
e1002434.
López Sánchez, A., Stassen, J.H.M., Furci, L., Smith, L.M., and Ton, J. (2016). The role of
DNA (de)methylation in immune responsiveness of Arabidopsis. Plant J. 88, 361–374.
Love, M.I., Huber, W., and Anders, S. (2014). Moderated estimation of fold change and
dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550.
Lu, F., Cui, X., Zhang, S., Liu, C., and Cao, X. (2010). JMJ14 is an H3K4 demethylase
regulating flowering time in Arabidopsis. Cell Res. 20, 387–390.
Luna, E., Bruce, T.J.A., Roberts, M.R., Flors, V., and Ton, J. (2012). Next-generation
systemic acquired resistance. Plant Physiol. 158, 844–853.
Ma, X., Nicole, M.-C., Meteignier, L.-V., Hong, N., Wang, G., and Moffett, P. (2015). Different
roles for RNA silencing and RNA processing components in virus recovery and virus-
induced gene silencing in plants. J. Exp. Bot. 66, 919–932.
Maere, S., Heymans, K., and Kuiper, M. (2005). BiNGO: a Cytoscape plugin to assess
overrepresentation of gene ontology categories in biological networks. Bioinformatics 21,
3448–3449.
Mager, S., and Ludewig, U. (2018). Massive Loss of DNA Methylation in Nitrogen-, but Not
in Phosphorus-Deficient Zea mays Roots Is Poorly Correlated With Gene Expression
Differences. Front. Plant Sci. 9, 497.
Martin, M. (2011). Cutadapt removes adapter sequences from high-throughput sequencing
reads. EMBnet.journal 17, 10–12.
Mirouze, M., Reinders, J., Bucher, E., Nishimura, T., Schneeberger, K., Ossowski, S., Cao,
J., Weigel, D., Paszkowski, J., and Mathieu, O. (2009). Selective epigenetic control of
retrotransposition in Arabidopsis. Nature 461, 427–430.
Miura, A., Nakamura, M., Inagaki, S., Kobayashi, A., Saze, H., and Kakutani, T. (2009). An
Arabidopsis jmjC domain protein protects transcribed genes from DNA methylation at CHG
sites. EMBO J. 28, 1078–1086.
Narsai, R., Secco, D., Schultz, M.D., and Ecker, J.R. (2017). Dynamic and rapid changes in
the transcriptome and epigenome during germination and in developing rice (Oryza sativa)
coleoptiles under anoxia and re-oxygenation. Plant J. 89, 805-824.
Negi, P., Rai, A.N., and Suprasanna, P. (2016). Moving through the Stressed Genome:
Emerging Regulatory Roles for Transposons in Plant Stress Response. Front. Plant Sci. 7,
1448.
Niehl, A., and Heinlein, M. (2019). Perception of double-stranded RNA in plant antiviral
immunity. Mol. Plant Pathol. 20, 1203-1210.
Niehl, A., Wyrsch, I., Boller, T., and Heinlein, M. (2016). Double-stranded RNAs induce a
pattern-triggered immune signaling pathway in plants. New Phytol. 211, 1008–1019.
Oliveira, R.R., Viana, A.J.C., Reátegui, A.C.E., and Vincentz, M.G.A. (2015). Short
Communication An efficient method for simultaneous extraction of high-quality RNA and
DNA from various plant tissues. Genet. Mol. Res. 14, 18828–18838.
Pecinka, A., Dinh, H.Q., Baubec, T., Rosa, M., Lettner, N., and Mittelsten Scheid, O. (2010).
Epigenetic regulation of repetitive elements is attenuated by prolonged heat stress in
Arabidopsis. Plant Cell 22, 3118–3129.
Pertea, M., Kim, D., Pertea, G.M., Leek, J.T., and Salzberg, S.L. (2016). Transcript-level
expression analysis of RNA-seq experiments with HISAT, StringTie and Ballgown. Nat.
Protoc. 11, 1650–1667.
Ramšak, Ž., Coll, A., Stare, T., Tzfadia, O., Baebler, Š., Van de Peer, Y., and Gruden, K.
(2018). Network Modeling Unravels Mechanisms of Crosstalk between Ethylene and
Salicylate Signaling in Potato. Plant Physiol. 178, 488–499.
Rigal, M., Kevei, Z., Pélissier, T., and Mathieu, O. (2012). DNA methylation in an intron of
the IBM1 histone demethylase gene stabilizes chromatin modification patterns. EMBO J. 31,
2981–2993.
Roudet-Tavert, G., Michon, T., Walter, J., Delaunay, T., Redondo, E., and Le Gall, O.
(2007). Central domain of a potyvirus VPg is involved in the interaction with the host
translation initiation factor eIF4E and the viral protein HcPro. J. Gen. Virol. 88, 1029–1033.
Sánchez, F., Manrique, P., Mansilla, C., Lunello, P., Wang, X., Rodrigo, G., López-
González, S., Jenner, C., González-Melendi, P., Elena, S.F., et al. (2015). Viral Strain-
Specific Differential Alterations in Arabidopsis Developmental Patterns. Mol. Plant. Microbe.
Interact. 28, 1304–1315.
Saze, H., Shiraishi, A., Miura, A., and Kakutani, T. (2008). Control of genic DNA methylation
by a jmjC domain-containing protein in Arabidopsis thaliana. Science 319, 462–465.
Searle, I.R., Pontes, O., Melnyk, C.W., Smith, L.M., and Baulcombe, D.C. (2010). JMJ14, a
JmjC domain protein, is required for RNA silencing and cell-to-cell movement of an RNA
silencing signal in Arabidopsis. Genes Dev. 24, 986–991.
Secco, D., Wang, C., Shou, H., Schultz, M.D., Chiarenza, S., Nussaume, L., Ecker, J.R.,
Whelan, J., and Lister, R. (2015). Stress induced gene expression drives transient DNA
methylation changes at adjacent repetitive elements. Elife 4, e09343.
Seymour, D.K., and Becker, C. (2017). The causes and consequences of DNA methylome
variation in plants. Curr. Opin. Plant Biol. 36, 56–63.
Seymour, D.K., Koenig, D., Hagmann, J., Becker, C., and Weigel, D. (2014). Evolution of
DNA methylation patterns in the Brassicaceae is driven by differences in genome
organization. PLoS Genet. 10, e1004785.
Sigman, M.J., and Slotkin, R.K. (2016). The First Rule of Plant Transposable Element
Silencing: Location, Location, Location. Plant Cell 28, 304–313.
Stief, A., Altmann, S., Hoffmann, K., Pant, B.D., Scheible, W.-R., and Bäurle, I. (2014).
Arabidopsis miR156 Regulates Tolerance to Recurring Environmental Stress through SPL
Transcription Factors. Plant Cell 26, 1792–1807.
Sun, Y., Fan, M., and He, Y. (2019). DNA Methylation Analysis of the Citrullus lanatus
Response to Cucumber Green Mottle Mosaic Virus Infection by Whole-Genome Bisulfite
Sequencing. Genes 10, 344.
Teixeira, R.M., Ferreira, M.A., Raimundo, G.A.S., Loriato, V.A.P., Reis, P.A.B., and Fontes,
E.P.B. (2019). Virus perception at the cell surface: revisiting the roles of receptor-like
kinases as viral pattern recognition receptors. Mol. Plant Pathol. 20, 1196–1202.
Thimm, O., Bläsing, O., Gibon, Y., Nagel, A., Meyer, S., Krüger, P., Selbig, J., Müller, L.A.,
Rhee, S.Y., and Stitt, M. (2004). MAPMAN: a user-driven tool to display genomics data sets
onto diagrams of metabolic pathways and other biological processes. Plant J. 37, 914–939.
Tittel-Elmer, M., Bucher, E., Broger, L., Mathieu, O., Paszkowski, J., and Vaillant, I. (2010).
Stress-induced activation of heterochromatic transcription. PLoS Genet. 6, e1001175.
Underwood, C.J., Henderson, I.R., and Martienssen, R.A. (2017). Genetic and epigenetic
variation of transposable elements in Arabidopsis. Curr. Opin. Plant Biol. 36, 135–141.
Wang, C., Wang, C., Xu, W., Zou, J., Qiu, Y., Kong, J., Yang, Y., Zhang, B., and Zhu, S.
(2018). Epigenetic Changes in the Regulation of Nicotiana tabacum Response to Cucumber
Mosaic Virus Infection and Symptom Recovery through Single-Base Resolution Methylomes.
Viruses 10, E402.
Wu, Q., Wang, X., and Ding, S.-W. (2010). Viral suppressors of RNA-based viral immunity:
host targets. Cell Host Microbe 8, 12–15.
Wu, X., Valli, A., García, J.A., Zhou, X., and Cheng, X. (2019). The Tug-of-War between
Plants and Viruses: Great Progress and Many Remaining Questions. Viruses 11, E203.
Xu, J., Zhou, S., Gong, X., Song, Y., van Nocker, S., Ma, F., and Guan, Q. (2018). Single-
base methylome analysis reveals dynamic epigenomic differences associated with water
deficit in apple. Plant Biotechnol. J. 16, 672–687.
Yadegari, R., Kinoshita, T., Lotan, O., Cohen, G., Katz, A., Choi, Y., Katz, A., Nakashima, K.,
Harada, J.J., Goldberg, R.B., et al. (2000). Mutations in the FIE and MEA genes that encode
interacting polycomb proteins cause parent-of-origin effects on seed development by distinct
mechanisms. Plant Cell 12, 2367–2382.
Yambao, M.L.M., Masuta, C., Nakahara, K., and Uyeda, I. (2003). The central and C-
terminal domains of VPg of Clover yellow vein virus are important for VPg-HCPro and VPg-
VPg interactions. J. Gen. Virol. 84, 2861–2869.
Yang, W., Jiang, D., Jiang, J., and He, Y. (2010). A plant-specific histone H3 lysine 4
demethylase represses the floral transition in Arabidopsis. Plant J. 62, 663–673.
Yu, A., Lepère, G., Jay, F., Wang, J., Bapaume, L., Wang, Y., Abraham, A.-L., Penterman,
J., Fischer, R.L., Voinnet, O., et al. (2013). Dynamics and biological relevance of DNA
demethylation in Arabidopsis antibacterial defense. Proc. Natl. Acad. Sci. U. S. A. 110,
2389–2394.
Zervudacki, J., Yu, A., Amesefe, D., Wang, J., Drouaud, J., Navarro, L., and Deleris, A.
(2018). Transcriptional control and exploitation of an immune-responsive family of plant
retrotransposons. EMBO J. 37, e98482.
Zhang, H., Lang, Z., and Zhu, J.-K. (2018). Dynamics and function of DNA methylation in
plants. Nat. Rev. Mol. Cell Biol. 19, 489–506.
Zhang, R., Calixto, C.P.G., Marquez, Y., Venhuizen, P., Tzioutziou, N.A., Guo, W.,
Spensley, M., Entizne, J.C., Lewandowska, D., Ten Have, S., et al. (2017). A high quality
Arabidopsis transcriptome for accurate transcript-level analysis of alternative splicing.
Nucleic Acids Res. 45, 5061–5073.
Zhang, X., Yazaki, J., Sundaresan, A., Cokus, S., Chan, S.W., Chen, H., Henderson, I.R.,
Shinn, P., Pellegrini, M., Jacobsen, S.E., et al. (2006). Genome-wide high-resolution
mapping and functional analysis of DNA methylation in arabidopsis. Cell 126, 1189–1201.
Zhao, S., and Fernald, R.D. (2005). Comprehensive algorithm for quantitative real-time
polymerase chain reaction. J. Comput. Biol. 12, 1047–1064.
Zheng, Z., Qamar, S.A., Chen, Z., and Mengiste, T. (2006). Arabidopsis WRKY33
transcription factor is required for resistance to necrotrophic fungal pathogens. Plant J. 48,
592–605.
Zheng, Z., Mosher, S.L., Fan, B., Klessig, D.F., and Chen, Z. (2007). Functional analysis of
Arabidopsis WRKY25 transcription factor in plant defense against Pseudomonas syringae.
BMC Plant Biol. 7, 2.
Zhu, Q.-H., Shan, W.-X., Ayliffe, M.A., and Wang, M.-B. (2016). Epigenetic Mechanisms: An
Emerging Player in Plant-Microbe Interactions. Mol. Plant. Microbe. Interact. 29, 187-96.
Figure 1
Click here to access/download;Figure;Fig1.png
Figure 2
Click here to access/download;Figure;Fig2.png
Figure 3
Click here to access/download;Figure;Fig3.png
Figure 4
Click here to access/download;Figure;Fig4.png
Figure 5
Click here to access/download;Figure;Fig5.png
Figure 6
Click here to access/download;Figure;Fig6.png
| 2019 | Viral fitness determines the magnitude of transcriptomic and epigenomic reprogramming of defense responses in plants | 10.1101/2019.12.26.888768 | [
"Corrêa Régis L.",
"Sanz-Carbonell Alejandro",
"Kogej Zala",
"Müller Sebastian Y.",
"López-Gomollón Sara",
"Gómez Gustavo",
"Baulcombe David C.",
"Elena Santiago F."
] | creative-commons |
Integrated sample inactivation, amplification, and Cas13-based detection of SARS-CoV-2
Jon Arizti-Sanz1,2,*, Catherine A. Freije1,3,*, Alexandra C. Stanton1,3, Chloe K. Boehm1, Brittany A.
Petros1,2,4, Sameed Siddiqui1,5, Bennett M. Shaw1,6, Gordon Adams1, Tinna-Solveig F. Kosoko-
Thoroddsen1, Molly E. Kemball1, Robin Gross7, Loni Wronka8, Katie Caviness8, Lisa E. Hensley7,
Nicholas H. Bergman8, Bronwyn L. MacInnis1,9, Jacob E. Lemieux1,6, Pardis C. Sabeti1,9,10,11,12,+, Cameron
Myhrvold1,10,12,+,§
1Broad Institute of Massachusetts Institute of Technology (MIT) and Harvard, Cambridge, MA, USA.
2Harvard-MIT Program in Health Sciences and Technology, Cambridge, MA, USA. 3Program in Virology,
Harvard Medical School, Boston, MA, USA. 4Harvard-MIT MD-PhD Program, Boston, MA, USA.
5Computational and Systems Biology PhD program, MIT, Cambridge, MA, USA. 6Department of Medicine,
Division of Infectious Diseases, Massachusetts General Hospital, Boston, MA, USA. 7Integrated Research
Facility, Division of Clinical Research, National Institute of Allergy and Infectious Diseases, National
Institutes of Health, Frederick, MD, USA. 8National Biodefense Analysis and Countermeasures Center, Fort
Detrick, MD, USA. 9Harvard T.H. Chan School of Public Health, Boston, MA, USA. 10Department of
Organismic and Evolutionary Biology, Harvard University, Cambridge, MA, USA. 11Howard Hughes Medical
Institute, Chevy Chase, MD, USA. 12Massachusetts Consortium on Pathogen Readiness, Boston, MA, USA.
*These authors contributed equally.
+These authors jointly supervised the work.
§Corresponding author. Email: cmyhrvol@broadinstitute.org (C.M.)
Abstract
The COVID-19 pandemic has highlighted that new diagnostic technologies are essential for controlling
disease transmission. Here, we develop SHINE (SHERLOCK and HUDSON Integration to Navigate
Epidemics), a sensitive and specific integrated diagnostic tool that can detect SARS-CoV-2 RNA from
unextracted samples. We combine the steps of SHERLOCK into a single-step reaction and optimize
HUDSON to accelerate viral inactivation in nasopharyngeal swabs and saliva. SHINE’s results can be
visualized with an in-tube fluorescent readout — reducing contamination risk as amplification reaction
tubes remain sealed — and interpreted by a companion smartphone application. We validate SHINE on
50 nasopharyngeal patient samples, demonstrating 90% sensitivity and 100% specificity compared to
RT-PCR with a sample-to-answer time of 50 minutes. SHINE has the potential to be used outside of
hospitals and clinical laboratories, greatly enhancing diagnostic capabilities.
Introduction
Point-of-care diagnostic testing is essential to prevent and effectively respond to infectious disease
outbreaks. Insufficient nucleic acid diagnostic testing infrastructure (1) and the prevalence of
asymptomatic transmission (2, 3) have accelerated the global spread of severe acute respiratory
syndrome coronavirus 2 (SARS-CoV-2) (4–6), with confirmed case counts surpassing 5 million (7).
Ubiquitous nucleic acid testing — whether in doctor’s offices, pharmacies, or mobile/drive-thru/pop-up
testing sites — would increase diagnostic access and is essential for safely reopening businesses,
schools, and country borders. Easy-to-use, scalable diagnostics with a quick turnaround time and limited
equipment requirements would fulfill this major need and have the potential to alter the trajectory of this
global pandemic.
The paradigm for nucleic acid diagnostic testing is a centralized model where patient samples are sent
to large clinical laboratories for processing and analysis. RT-qPCR, the highly specific and sensitive
current gold-standard for SARS-CoV-2 diagnosis (8), requires laboratory infrastructure for nucleic acid
extraction, thermal cycling, and analysis of assay results. The need for thermocyclers can be eliminated
through the use of isothermal (i.e., single temperature) amplification methods, such as loop-mediated
isothermal amplification (LAMP) and recombinase polymerase amplification (RPA) (9–14). However,
isothermal amplification methods still require technological advances (Qian, Boswell, Chidley, Lu et al.
submitted) to increase sensitivity on unextracted RNA samples and to reduce non-specific amplification
(15, 16), which would enable testing at scale outside of laboratories.
Recently developed CRISPR-based diagnostics have the potential to transform infectious disease
diagnosis. Both CRISPR-Cas13- and Cas12-based assays have been developed for SARS-CoV-2
detection using extracted nucleic acids as input (17–22). One such CRISPR-based diagnostic,
SHERLOCK (Specific High-sensitivity Enzymatic Reporter unLOCKing), involves two separate steps,
starting with extracted nucleic acid: (1) isothermal RPA and (2) T7 transcription and Cas13-mediated
collateral cleavage of a single-stranded RNA reporter (23) (Fig. 1A). Cas13-based detection is highly
programmable and specific, as it relies on complementary base pairing between the target RNA and the
CRISPR RNA (crRNA) sequence (23, 24). However, in their current state, these technologies require
nucleic acid extraction (often using kits that are in short supply) and multiple sample transfer steps,
limiting their widespread use. SHERLOCK can be paired with HUDSON (Heating Unextracted Diagnostic
Samples to Obliterate Nucleases), which eliminates the need for nucleic acid extraction by using heat
and chemical reduction to both destroy RNA-degrading nucleases and lyse viral particles (25). Together,
SHERLOCK and HUDSON can be performed with limited laboratory infrastructure, solely requiring a
heating element. However, the scalability of these methods is currently limited by the need to prepare
multiple reaction mixtures and transfer samples between them.
To address the current limitations of nucleic acid diagnostics, we developed SHINE (SHERLOCK and
HUDSON Integration to Navigate Epidemics) for extraction-free, rapid, and sensitive detection of SARS-
CoV-2 RNA. We established a SARS-CoV-2 assay (18), then combined SHERLOCK’s amplification and
Cas13-based detection steps, decreasing user manipulations and assay time (Fig. 1A). We demonstrated
that SHINE can detect SARS-CoV-2 RNA in HUDSON-treated patient samples with either a paper-based
colorimetric readout, or an in-tube fluorescent readout which can be performed with portable equipment
and with reduced risk of sample contamination.
Results
We first developed a two-step SHERLOCK assay which sensitively detected SARS-CoV-2 RNA at 10
copies per microliter (cp/μL). Using ADAPT, a computational design tool for nucleic acid diagnostics
(Metsky et al. in prep), we identified a region within open reading frame 1a (ORF1a) of SARS-CoV-2 that
comprehensively captures known sequence diversity, with high predicted Cas13 targeting activity and
SARS-CoV-2 specificity (Fig. 1B) (18). Using both colorimetric and fluorescent readouts, we detected 10
cp/μL of synthetic RNA after incubating samples for 1 h or less, but preparing the reactions required 45-
90 minutes of hands-on time depending on the number of samples (Fig. 1C and 1D and fig. S1A). We
tested this assay on HUDSON-treated SARS-CoV-2 viral seedstocks, detecting down to 1.31e5 PFU/ml
via colorimetric readout (Fig. S1B). Finally, in a side-by-side comparison of our two-step SHERLOCK
assay and the CDC RT-qPCR assay, we demonstrated similar limits of detection, reliably identifying 1-
10 cp/μL with stochasticity evident at lower viral titers (Fig. S1C).
Fig. 1. Initial assay development for SHERLOCK-based SARS-CoV-2 detection. (A) Schematic of two- and single-step
SHERLOCK assays using RNA extracted from patient samples with a fluorescent or colorimetric readout. Times, range of
suggested incubation times; pipette, step involving user manipulation; RT-RPA, reverse transcriptase-recombinase polymerase
amplification; C, control line; T, test line. (B) Schematic of the SARS-CoV-2 genome and SHERLOCK assay location. Sequence
conservation across the primer and crRNA binding sites for publicly available SARS-CoV-2 genomes (see Methods for details).
Text denotes nucleotide position with lowest percent conservation across the assay location. ORF, open reading frame; narrow
rectangles, untranslated regions; dashed border, unlikely to be expressed (32). (C) Colorimetric detection of synthetic RNA using
two-step SHERLOCK after 30 min. NTC_r, non-template control introduced in RPA, NTC_d, non-template control introduced in
detection; T, test line; C, control line. (D) Background-subtracted fluorescences of the two-step and original single-step
SHERLOCK protocols using synthetic SARS-CoV-2 RNA after 3 h. The 1 h timepoint from this experiment is shown in Fig. 2E.
NTC, non-template control introduced in RPA. Error bars, s.d. for 2-3 technical replicates.
We sought to develop an integrated, streamlined assay that was significantly less time- and labor-
intensive than two-step SHERLOCK. However, when we combined RT-RPA (step 1), T7 transcription,
and Cas13-based detection (step 2) into a single step (i.e., single-step SHERLOCK), the sensitivity of
the assay decreased dramatically. This decrease was specific for RNA input, and likely due to
incompatibility of enzymatic reactions with the given conditions (limit of detection (LOD) 106 cp/μL; Fig.
1D and fig. S2A). As a result, we evaluated whether additional reaction components and optimized
reaction conditions could increase the sensitivity and speed of the assay. Addition of RNase H, in the
presence of reverse transcriptase, improved the sensitivity of Cas13-based detection of RNA 10-fold
(LOD 105 cp/μL; Fig. 2A and fig. S2B and S2C). RNase H likely enhanced the sensitivity by increasing
the efficiency of RT through degradation of DNA:RNA hybrid intermediates (Qian, Boswell, Chidley, Lu
et al. submitted).
Given that each enzyme involved has optimal activity at distinct reaction conditions, we evaluated the
role of different pHs, monovalent salt, magnesium, and primer concentrations on assay sensitivity.
Optimized buffer, magnesium, and primer conditions resulted in an LOD of 1,000 cp/μL (Fig. 2B and 2C
and fig. S2D and S2E). We then improved the speed of Cas13 cleavage and RT to reduce the sample-
to-answer time. Given the uracil-cleavage preference of Cas13a (24, 26, 27), detection of RNA in the
single-step SHERLOCK assay reached half-maximal fluorescence in ~67% of the time when RNaseAlert
was substituted for a polyU reporter (Fig. 2D, left and fig. S3). In addition, reactions containing
SuperScript IV reverse transcriptase reached half-maximal fluorescence two times faster than RevertAid
(Fig. 2D, right).
Together, these improvements resulted in an optimized single-step SHERLOCK assay that could detect
SARS-CoV-2 RNA with reduced sample-to-answer time and equal sensitivity compared to our two-step
assay. We quantified the LOD of our optimized single-step SHERLOCK assay on synthetic RNA,
detecting as few as 10 cp/μL using a fluorescent readout — 100,000 times more sensitive than the initial
assay — and 100 cp/μL using the lateral-flow-based colorimetric readout (Fig. 2E and 2F and fig. S4).
We then evaluated our assay’s performance on SARS-CoV-2 RNA extracted from patient
nasopharyngeal (NP) swabs. We compared our fluorescent single-step SHERLOCK assay to previously-
performed RT-qPCR using a pilot set of 9 samples. We detected SARS-CoV-2 from 5 of 5 SARS-CoV-
2-positive patient samples tested, demonstrating 100% concordance with RT-qPCR, with no false
positives for 4 SARS-CoV-2-negative extracted samples nor 2 non-template controls (Fig. 2H and 2I and
table S1).
Fig. 2. Optimization of the single-step SHERLOCK reaction. (A) Background-subtracted fluorescence of Cas13-based
detection with synthetic RNA, reverse transcriptase, and RPA primers (but no RPA enzymes) after 3 h. (B) Single-step
SHERLOCK normalized fluorescence using various buffering conditions after 3 h. (C) Background-subtracted fluorescence of
single-step SHERLOCK with synthetic RNA and variable RPA forward and reverse primer concentrations after 3 h. (D) Single-
step SHERLOCK normalized fluorescence over time using two different fluorescent reporters (left) and two different reverse
transcriptases (right). (E) Background-subtracted fluorescences of the original single-step and optimized single-step
SHERLOCK with synthetic RNA after 1 h. Data from the 3 h timepoint from this experiment is shown in Fig. 1D. (F) Colorimetric
detection of synthetic RNA input using optimized single-step SHERLOCK after 3 h. (G) Optimized single-step SHERLOCK
background-subtracted fluorescence using RNA extracted from patient samples after 1 h. (H) Concordance between
SHERLOCK and RT-qPCR for 7 patient samples and 4 controls. For (C and E), see methods for details about normalized
fluorescence calculations. For (B,D,F, and G), NTC, non-template control. For (B,D,E, and F), error bars, s.d. for 2-3 technical
replicates. For (B and D) RNA input at 104 cp/μL.
Finally, we paired HUDSON and SHERLOCK with multiple visual readouts to create SHINE (SHERLOCK
and HUDSON Integration to Navigate Epidemics), a platform whose results are interpretable by a
companion smartphone application (Fig. 3A). In order to reduce total run time, we reduced the incubation
time of HUDSON from 30 min to 10 min for both universal viral transport medium (UTM), used for NP
swab samples, and for saliva, through the addition of RNase inhibitors (25) (Fig. 3B and fig. S5). With
this faster HUDSON protocol, we detected 50 cp/μL of synthetic RNA when spiked into UTM and 100
cp/μL when spiked into saliva, using a colorimetric readout (Fig. S6). However, the lateral flow readout
requires opening of tubes containing amplified products and interpreting the test band by eye, which
introduces risks of sample contamination and user bias, respectively. Thus, we incorporated an in-tube
fluorescent readout with SHINE. Within 1 hour, we detected as few as 10 cp/μL of SARS-CoV-2 synthetic
RNA in HUDSON-treated UTM and 5 cp/μL in HUDSON-treated saliva with the in-tube fluorescent
readout (Fig. 3C and 3D and figs. S7 and S8). To reduce user-bias in interpreting results of this in-tube
readout, we developed a companion smartphone app which uses the built-in smartphone camera to
image the reaction tubes. The application then calculates the distance of the experimental tube’s pixel
intensity distribution from that of a user-selected negative control tube, and returns a binary result
indicating the presence or absence of viral RNA in the sample (Fig. 3A and 3E; see Methods for details).
Thus, SHINE both minimized equipment requirements and user interpretation bias when implemented
with this in-tube readout and the smartphone application.
We used SHINE to test a set of 50 unextracted, NP samples from 30 RT-qPCR-confirmed, COVID-19-
positive patients and 20 COVID-19-negative patients. We used SHINE with the paper-based colorimetric
readout on 6 SARS-CoV-2-positive samples and detected SARS-CoV-2 RNA in all 6 positive samples,
and in none of the negative controls (100% concordance, Fig. 3F). For all 50 samples, we used SHINE
with the in-tube fluorescence readout and companion smartphone application. We detected SARS-CoV-
2 RNA in 27 of 30 COVID-19-positive samples (90% sensitivity) and none of the COVID-19-negative
samples (100% specificity) after a 10-minute HUDSON and a 40-minute single-step SHERLOCK
incubation (Fig. 3G and 3H, fig. S9, and table S1 and S2). Thus, SHINE demonstrated 94% concordance
using the in-tube readout with a total run time of 50 minutes. Notably, the RT-qPCR-positive patient NP
swabs that SHINE failed to detect tended to have higher Ct values than those that SHINE detected as
positive (p = 0.0084 via one-sided Wilcoxon rank sum test; Fig. S10). Moreover, this observation could
be related to sample degradation and differences in sample processing, as SHINE samples went through
additional freeze-thaw cycles and RT-qPCR was performed on extracted and DNase-treated samples.
Fig. 3. SARS-CoV-2 detection from unextracted samples using SHINE. (A) Schematic of SHINE, which is HUDSON paired
with single-step SHERLOCK using an in-tube fluorescent or colorimetric readout. Times, range of suggested incubation times;
C, control line; T, test line. (B) RNaseAlert fluorescence measured after 30 min at room temperature from universal viral transport
medium (UTM), saliva, and phosphate buffered saline (PBS) after heat and chemical treatment. (C) SARS-CoV-2 RNA detection
in HUDSON-treated UTM as measured by single-step SHERLOCK and the in-tube fluorescence readout after 1 h. (D) SARS-
CoV-2 RNA detection in HUDSON-treated saliva as measured by single-step SHERLOCK and the in-tube fluorescence readout
after 1 h. (E) Schematic of the companion smartphone application for quantitatively analyzing in-tube fluorescence and reporting
binary outcomes of SARS-CoV-2 detection. (F) Colorimetric detection of SARS-CoV-2 RNA in unextracted patient NP swabs
using the SHINE after 1 h. (G) SARS-CoV-2 detection from unextracted patient samples using SHINE and smartphone
application quantification of in-tube fluorescence after 40 min. Threshold line determined as average readout value for controls
plus 3 standard deviations. (H) Concordance table between SHINE and RT-qPCR for 50 patient samples.
Discussion
Here, we have described SHINE, a simple method for detecting viral RNA from unextracted patient
samples with minimal equipment requirements. SHINE’s simplicity matches that of the most streamlined
nucleic acid diagnostics. Furthermore, the in-tube fluorescence readout and companion smartphone
application lends themselves to scalable, high-throughput testing and automated interpretation of results.
SHINE’s simplicity and CRISPR-based programmability underscore its potential to address diagnostic
needs during the COVID-19 pandemic, and in outbreaks to come.
Additional advances are still required for diagnostic testing to occur in virtually any location. Ideally, all
steps would be performed at ambient temperature (without heat), in 15 minutes or less, using a
colorimetric readout that does not require tube opening. Existing nucleic acid diagnostics, to our
knowledge, are not capable of meeting all these requirements simultaneously. Sample collection without
UTM (i.e., “dry swabs”) combined with spin-column-free extraction buffers, and incorporation of solution-
based, colorimetric readouts could address these limitations (28–31). Together, these advances could
greatly enhance the accessibility of diagnostic testing and provide an essential tool in the fight against
infectious diseases. By reducing personnel time, equipment, and assay time-to-results without sacrificing
sensitivity or specificity, we have taken steps towards the development of such a tool.
Materials and Methods
Detailed information about reagents, including the commercial vendors and stock concentrations, is
provided in Table S3.
Clinical samples and ethics statement
Clinical samples were acquired from clinical studies evaluated and approved by the Institutional Review
Board/Ethics Review Committee of the Massachusetts General Hospital and Massachusetts Institute of
Technology (MIT). The Office of Research Subject Projection at the Broad Institute of MIT and Harvard
University approved use of samples for the work performed in this study.
Extracted sample preparation and RT-qPCR testing
Nasal swabs were collected and stored in universal viral transport medium (UTM; BD) and stored at -80
°C prior to nucleic acid extraction. Nucleic acid extraction was performed using MagMAX™ mirVana™
Total RNA isolation kit. The starting volume for the extraction was 250 μl and extracted nucleic acid was
eluted into 60 μl of nuclease-free water. Extracted nucleic acid was then immediately Turbo DNase-
treated (Thermo Fisher Scientific), purified twice with RNACleanXP SPRI beads (Beckman Coulter), and
eluted into 15 μl of Ambion Linear Acrylamide (Thermo Fisher Scientific) water (0.8%).
Turbo DNase-treated extracted RNA was then tested for the presence of SARS-CoV-2 RNA using a lab-
developed, probe-based RT-qPCR assay based on the N1 target of the CDC assay. RT-qPCR was
performed on a 1:3 dilution of the extracted RNA using TaqPath™ 1-Step RT-qPCR Master Mix (Thermo
Fisher Scientific) with the following forward and reverse primer sequences, respectively:
GACCCCAAAATCAGCGAAAT, TCTGGTTACTGCCAGTTGAATCTG. The RT-PCR assay was run with
a
double-quenched
FAM
probe
with
the
following
sequence:
5’-FAM-
ACCCCGCATTACGTTTGGTGGACC-BHQ1-3’. RT-qPCR was run on a QuantStudio 6 (Applied
Biosystems) with RT at 48 °C for 30 min and 45 cycles with a denaturing step at 95 °C for 10 s followed
by annealing and elongation steps at 60 °C for 45 s. The data were analyzed using the Standard Curve
(SC) module of the Applied Biosystems Analysis Software.
SARS-CoV-2 assay design and synthetic template information
SARS-CoV-2-specific forward and reverse RPA primers and Cas13-crRNAs were designed as previously
described (18). In short, the designs were algorithmically selected, targeting 100% of 20 published SARS-
CoV-2 genomes, and predicted by a machine learning model to be highly active (Metsky et al. in prep).
Moreover, the crRNA was selected for its high predicted specificity towards detection of SARS-CoV-2,
versus related viruses, including other bat and mammalian coronaviruses and other human respiratory
viruses (https://adapt.sabetilab.org/covid-19/).
Synthetic
DNA
targets
with
appended
upstream
T7
promoter
sequences
(5’-
GAAATTAATACGACTCACTATAGGG-3’) were ordered as double-stranded DNA (dsDNA) gene
fragments from IDT, and were in vitro transcribed to generate synthetic RNA targets. In vitro transcription
was conducted using the HiScribe T7 High Yield RNA Synthesis Kit (New England Biolabs (NEB)) as
previously
described
(23).
In
brief,
a
T7
promoter
ssDNA
primer
(5’-
GAAATTAATACGACTCACTATAGGG-3’) was annealed to the dsDNA template and the template was
transcribed at 37 ºC overnight. Transcribed RNA was then treated with RNase-free DNase I (QIAGEN)
to remove any remaining DNA according to the manufacturer’s instructions. Finally, purification occurred
using RNAClean SPRI XP beads at 2✕ transcript volumes in 37.5% isopropanol.
Sequence information for the synthetic targets, RPA primers, and Cas13-crRNA is listed in Table S4.
Two-step SARS-CoV-2 assay
The two-step SHERLOCK assay was performed as previously described (18, 23, 25). Briefly, the assay
was performed in two steps: (1) isothermal amplification via recombinase polymerase amplification (RPA)
and (2) LwaCas13a-based detection using a single-stranded RNA (ssRNA) fluorescent reporter. For
RPA, the TwistAmp Basic Kit (TwistDx) was used as previously described (i.e., with RPA forward and
reverse primer concentrations of 400 nM and a magnesium acetate concentration of 14 mM) (25) with
the following modifications: RevertAid reverse transcriptase (Thermo Fisher Scientific) and murine RNase
inhibitor (NEB) were added at final concentrations of 4 U/µl each, and synthetic RNAs or viral seedstocks
were added at known input concentrations making up 10% of the total reaction volume. The RPA reaction
was then incubated on the thermocycler for 20 minutes at 41 °C. For the detection step, 1 µl of RPA
product was added to 19 µl detection master mix. The detection master mix consisted of the following
reagents (final concentrations in master mix listed), with magnesium chloride added last: 45 nM
LwaCas13a protein resuspended in 1✕ storage buffer (SB: 50 mM Tris pH 7.5, 600 mM NaCl, 5%
glycerol, and 2 mM dithiothreitol (DTT); such that the resuspended protein is at 473.7 nM), 22.5 nM
crRNA, 125 nM RNaseAlert substrate v2 (Thermo Fisher Scientific), 1✕ cleavage buffer (CB; 400 mM
Tris pH 7.5 and 10 mM DTT), 2 U/µlL murine RNase inhibitor (NEB), 1.5 U/µl NextGen T7 RNA
polymerase (Lucigen), 1 mM of each rNTP (NEB), and 9 mM magnesium chloride. Reporter fluorescence
kinetics were measured at 37 °C on a Biotek Cytation 5 plate reader using a monochromator (excitation:
485 nm, emission: 520 nm) every 5 minutes for up to 3 hours.
Single-step SARS-CoV-2 assay optimization
The starting point for optimization of the single-step SHERLOCK assay was generated by combining the
essential reaction components of both the RPA and the detection steps in the two-step assay, described
above (23, 25). Briefly, a master mix was created with final concentrations of 1✕ original reaction buffer
(20 mM HEPES pH 6.8 with 60 mM NaCl, 5% PEG, and 5 µM DTT), 45 nM LwaCas13a protein
resuspended in 1✕ SB (such that the resuspended protein is at 2.26 µM), 136 nM RNaseAlert substrate
v2, 1 U/µl murine RNase inhibitor, 2 mM of each rNTP, 1 U/µl NextGen T7 RNA polymerase, 4 U/µl
RevertAid reverse transcriptase, 0.32 µM forward and reverse RPA primers, and 22.5 nM crRNA. The
TwistAmp Basic Kit lyophilized reaction components (1 lyophilized pellet per 102 µl final master mix
volume) were resuspended using the master mix. After pellet resuspension, cofactors magnesium
chloride and magnesium acetate were added at final concentrations of 5 mM and 17 mM, respectively,
to complete the reaction.
Master mix and synthetic RNA template were mixed and aliquoted into a 384-well plate in triplicate, with
20 µl per replicate at a ratio of 19:1 master mix:sample. Fluorescence kinetics were measured at 37 °C
on a Biotek Cytation 5 or Biotek Synergy H1 plate reader every 5 minutes for 3 hours, as described
above. We observed no significant difference in performance between the two plate reader models.
Optimization occurred iteratively, with a single reagent modified in each experiment. The reagent
condition (e.g., concentration, vendor, or sequence) that produced the most optimal results — defined as
either a lower limit of detection (LOD) or improved reaction kinetics (i.e., reaction saturates faster) — was
incorporated into our protocol. Thus, the protocol used for every future reagent optimization consisted of
the most optimal reagent conditions for every reagent tested previously.
For all optimization experiments, the modulated reaction component is described in the figures,
associated captions, or associated legends. Across all experiments, the following components of the
master mix were held constant: 45 nM LwaCas13a protein resuspended in 1✕ SB (such that the
resuspended protein is at 2.26 µM), 1 U/µl murine RNase inhibitor, 2 mM of each rNTP, 1 U/µl NextGen
T7 RNA polymerase, and 22.5 nM crRNA, and TwistDx RPA TwistAmp Basic Kit lyophilized reaction
components (1 lyophilized pellet per 102 µl final master mix volume). In all experiments, the master mix
components except for the magnesium cofactor(s) were used to resuspend the lyophilized reaction
components, and the magnesium cofactor(s) were added last. All other experimental conditions, which
differ among the experiments due to real-time optimization, are detailed in Table S5.
Single-step SARS-CoV-2 optimized reaction
The optimized reaction (see Supplementary Protocol for exemplary implementation) consists of a master
mix with final concentrations of 1✕ optimized reaction buffer (20 mM HEPES pH 8.0 with 60 mM KCl and
5% PEG), 45 nM LwaCas13a protein resuspended in 1✕ SB (such that the resuspended protein is at
2.26 µM), 125 nM polyU [i.e., 6 uracils (6U) or 7 uracils (7U) in length, unless otherwise stated] FAM
quenched reporter, 1 U/µl murine RNase inhibitor, 2 mM of each rNTP, 1 U/µl NextGen T7 RNA
polymerase, 2 U/µl Invitrogen SuperScript IV (SSIV) reverse transcriptase (Thermo Fisher Scientific), 0.1
U/µl RNase H (NEB), 120 nM forward and reverse RPA primers, and 22.5 nM crRNA. Once the master
mix is created, it is used to resuspend the TwistAmp Basic Kit lyophilized reaction components (1
lyophilized pellet per 102 µl final master mix volume). Finally, magnesium acetate is the sole magnesium
cofactor, and is added at a final concentration of 14 mM to generate the final master mix.
The sample is added to the complete master mix at a ratio of 1:19 and the fluorescence kinetics are
measured at 37 °C using a Biotek Cytation 5 or Biotek Synergy H1 plate reader as described above.
Visual detection via in-tube fluorescence and via lateral flow strip
Minor modifications were made to the single-step SARS-CoV-2 optimized reaction to visualize the
readout via in-tube fluorescence or lateral flow strip.
For in-tube fluorescence, we generated the single-step master mix as described above, except the 7U
FAM quenched reporter was used at a concentration of 62.5 nM. The sample was added to the complete
master mix at a ratio of 1:19. Samples were incubated at 37 °C and images were collected after 30, 45,
60, 90, 120 and 180 minutes of incubation, with image collection terminating once experimental results
were clear. A dark reader transilluminator (DR196 model, Clare Chemical Research) was used to
illuminate the tubes.
For lateral-flow readout, we generated the single-step master mix as described above, except we used
a biotinylated FAM reporter at a final concentration of 1 µM rather than the quenched polyU FAM
reporters. The sample was added to the complete master mix at a ratio of 1:19. After 1-2 hours of
incubation at 37 °C, the detection reaction was diluted 1:4 in Milenia HybriDetect Assay Buffer, and the
Milenia HybriDetect 1 (TwistDx) lateral flow strip was added. Sample images were collected 5 min
following incubation of the strip.
In-tube fluorescence reader mobile phone application
To enable smartphone-based fluorescence analysis, we designed a companion application. Using the
application, the user captures an image of a set of strip tubes illuminated by a transilluminator. The user
then identifies regions of interest in the captured image by overlaying a set of pre-drawn boxes onto
experimental and control tubes. Image and sample information is then transmitted to a server for analysis.
Within each of the user-selected squares, the server models the bottom of each tube as a trapezoid and
uses a convolutional kernel to determine the location of maximal signal within each tube, using data from
the green channel of the RGB image. The server then identifies the background signal proximal to each
tube and fits a Gaussian distribution around the background signal and around the in-tube signal. The
difference between the mean pixel intensity of the background signal and the mean pixel intensity of the
in-tube signal is then calculated as the background-subtracted fluorescence signal for each tube. To
identify experimentally significant fluorescent signals, a score is computed for each experimental tube;
this score is equal to the distance between the experimental and control background-subtracted
fluorescence divided by the standard deviation of pixel intensities in the control signal. Finally, positive or
negative samples are determined based on whether the score is above (positive, +) or below (negative,
-) 1.5, a threshold identified empirically.
HUDSON protocols
HUDSON nuclease and viral inactivation were performed on viral seedstock as previously described with
minor modifications to the temperatures and incubation times (25). In short, 100 mM TCEP (Thermo
Fisher Scientific) and 1 mM EDTA (Thermo Fisher Scientific) were added to non-extracted viral seedstock
and incubated for 20 minutes at 50 ºC, followed by 10 minutes at 95 ºC. The resulting product was then
used as input into the two-step SHERLOCK assay.
The improved HUDSON nuclease and viral inactivation protocol was performed as previously described,
with minor modifications (25). Briefly, 100 mM TCEP, 1 mM EDTA, and 0.8 U/µl murine RNase inhibitor
were added to clinical samples in universal viral transport medium or human saliva (Lee Biosolutions).
These samples were incubated for 5 minutes at 40 ºC, followed by 5 minutes at 70 ºC (or 5 minutes at
95 ºC, if saliva). The resulting product was used in the single-step detection assay. In cases where
synthetic RNA targets were used, rather than clinical samples (e.g., during reaction optimization), targets
were added after the initial heating step (40 ºC at 5 minutes). This is meant to recapitulate patient
samples, as RNA release occurs after the initial heating step when the temperature is increased and viral
particles lyse.
For optimization of nuclease inactivation using HUDSON, only the initial heating step was performed.
The products were then mixed 1:1 with 400 mM RNaseAlert substrate v2 in nuclease-free water and
incubated at room temperature for 30 minutes before imaging on a transilluminator or measuring reporter
fluorescence on a Biotek Synergy H1 [at room temperature using a monochromator (excitation: 485 nm,
emission: 520 nm) every 5 minutes for up to 30 minutes]. The specific HUDSON protocol parameters
modified are indicated in the figure captions.
SHINE
The SHINE assay consists of the optimized HUDSON protocol (described above) with the resulting
product used as the sample input into our optimized, one-step SHERLOCK protocol (described above).
Data analysis and schematic generation
Conservation of SARS-CoV-2 sequences across our SHERLOCK assay was determined using publicly
available genome sequences via GISAID. Analysis was based on an alignment of 5376 SARS-CoV-2
genomic sequences. Percent conservation was measured at each nucleotide within the RPA primer and
Cas13-crRNA binding sites and represents the percentage of genomes that have the consensus base at
each nucleotide position.
As described above, fluorescence values are reported as background-subtracted, with the fluorescence
value collected before reaction progression (i.e., the latest time at which no change in fluorescence is
observed, usually time 0, 5, or 10 minutes) subtracted from the final fluorescence value (3 hours, unless
otherwise indicated).
Normalized fluorescence values are calculated using data aggregated from multiple experiments with at
least one condition in common. The maximal fluorescence value across all experiments is set to 1, with
fluorescence values from the same experiment set as ratios of the maximal fluorescence value. Common
conditions across experiments are set to the same normalized value, and that value is propagated to
determine the normalized values within an experiment.
The Wilcoxon rank sum test was conducted in MATLAB (MathWorks). Schematics shown in Fig. 1A and
Fig. 3A were created using BioRender.com. All other schematics were generated in Adobe Illustrator
(v24.1.2). Data panels were primarily generated via Prism 8 (GraphPad), except Figure 3E which was
generated using Python (version 3.7.2), seaborn (version 0.10.1) and matplotlib (version 3.2.1) (33, 34).
References
1. S. Kaplan, K. Thomas, Despite Promises, Testing Delays Leave Americans “Flying Blind” (2020),
(available at https://www.nytimes.com/2020/04/06/health/coronavirus-testing-us.html).
2. Y. Bai, L. Yao, T. Wei, F. Tian, D.-Y. Jin, L. Chen, M. Wang, Presumed Asymptomatic Carrier
Transmission of COVID-19. JAMA (2020), doi:10.1001/jama.2020.2565.
3. C. Rothe, M. Schunk, P. Sothmann, G. Bretzel, G. Froeschl, C. Wallrauch, T. Zimmer, V. Thiel, C.
Janke, W. Guggemos, M. Seilmaier, C. Drosten, P. Vollmar, K. Zwirglmaier, S. Zange, R. Wölfel,
M. Hoelscher, Transmission of 2019-nCoV Infection from an Asymptomatic Contact in Germany.
N. Engl. J. Med. 382, 970–971 (2020).
4. C. Wang, P. W. Horby, F. G. Hayden, G. F. Gao, A novel coronavirus outbreak of global health
concern. Lancet. 395, 470–473 (2020).
5. N. Zhu, D. Zhang, W. Wang, X. Li, B. Yang, J. Song, X. Zhao, B. Huang, W. Shi, R. Lu, P. Niu, F.
Zhan, X. Ma, D. Wang, W. Xu, G. Wu, G. F. Gao, W. Tan, China Novel Coronavirus Investigating
and Research Team, A Novel Coronavirus from Patients with Pneumonia in China, 2019. N. Engl.
J. Med. 382, 727–733 (2020).
6. P. Zhou, X.-L. Yang, X.-G. Wang, B. Hu, L. Zhang, W. Zhang, H.-R. Si, Y. Zhu, B. Li, C.-L. Huang,
H.-D. Chen, J. Chen, Y. Luo, H. Guo, R.-D. Jiang, M.-Q. Liu, Y. Chen, X.-R. Shen, X. Wang, X.-S.
Zheng, K. Zhao, Q.-J. Chen, F. Deng, L.-L. Liu, B. Yan, F.-X. Zhan, Y.-Y. Wang, G.-F. Xiao, Z.-L.
Shi, A pneumonia outbreak associated with a new coronavirus of probable bat origin. Nature. 579,
270–273 (2020).
7. World Health Organization (WHO). COVID-19 Situation Report, May 25, 2020, (available at
https://www.who.int/docs/default-source/coronaviruse/situation-reports/20200524-covid-19-sitrep-
125.pdf?sfvrsn=80e7d7f0_2).
8. U.S. Food and Drug Administration, Policy for COVID-19 Tests During the Public Health
Emergency (Revised) (2020), (available at https://www.fda.gov/regulatory-information/search-fda-
guidance-documents/policy-coronavirus-disease-2019-tests-during-public-health-emergency-
revised).
9. T. Notomi, H. Okayama, H. Masubuchi, T. Yonekawa, K. Watanabe, N. Amino, T. Hase, Loop-
mediated isothermal amplification of DNA. Nucleic Acids Res. 28, E63 (2000).
10. G.-S. Park, K. Ku, S.-H. Baek, S.-J. Kim, S. I. Kim, B.-T. Kim, J.-S. Maeng, Development of
Reverse Transcription Loop-Mediated Isothermal Amplification Assays Targeting SARS-CoV-2. J.
Mol. Diagn. (2020), doi:10.1016/j.jmoldx.2020.03.006.
11. Y. H. Baek, J. Um, K. J. C. Antigua, J.-H. Park, Y. Kim, S. Oh, Y.-I. Kim, W.-S. Choi, S. G. Kim, J.
H. Jeong, B. S. Chin, H. D. G. Nicolas, J.-Y. Ahn, K. S. Shin, Y. K. Choi, J.-S. Park, M.-S. Song,
Development of a reverse transcription-loop-mediated isothermal amplification as a rapid early-
detection method for novel SARS-CoV-2. Emerg. Microbes Infect., 1–31 (2020).
12. A. Niemz, T. M. Ferguson, D. S. Boyle, Point-of-care nucleic acid testing for infectious diseases.
Trends in Biotechnology. 29 (2011), pp. 240–250.
13. Abott ID NOWTM COVID-19 - Alere is now, (available at https://www.alere.com/en/home/product-
details/id-now-covid-19.html).
14. O. Piepenburg, C. H. Williams, D. L. Stemple, N. A. Armes, DNA detection using recombination
proteins. PLoS Biol. 4, e204 (2006).
15. H. Zaghloul, M. El-Shahat, Recombinase polymerase amplification as a promising tool in hepatitis
C virus diagnosis. World J. Hepatol. 6, 916–922 (2014).
16. L. Yan, J. Zhou, Y. Zheng, A. S. Gamson, B. T. Roembke, S. Nakayama, H. O. Sintim, Isothermal
amplified detection of DNA and RNA. Molecular BioSystems. 10 (2014), p. 970.
17. Zhang F, Abudayyeh OO, Gootenberg JS, A protocol for detection of COVID-19 using CRISPR
diagnostics, (available at https://www.broadinstitute.org/files/publications/special/COVID-
19%20detection%20(updated).pdf).
18. H. C. Metsky, C. A. Freije, T.-S. F. Kosoko-Thoroddsen, P. C. Sabeti, C. Myhrvold, CRISPR-based
surveillance for COVID-19 using genomically-comprehensive machine learning design. bioRxiv
(2020).
19. J. P. Broughton, X. Deng, G. Yu, C. L. Fasching, V. Servellita, J. Singh, X. Miao, J. A. Streithorst,
A. Granados, A. Sotomayor-Gonzalez, K. Zorn, A. Gopez, E. Hsu, W. Gu, S. Miller, C.-Y. Pan, H.
Guevara, D. A. Wadford, J. S. Chen, C. Y. Chiu, CRISPR-Cas12-based detection of SARS-CoV-2.
Nat. Biotechnol. (2020), doi:10.1038/s41587-020-0513-4.
20. L. Guo, X. Sun, X. Wang, C. Liang, H. Jiang, Q. Gao, M. Dai, B. Qu, S. Fang, Y. Mao, Y. Chen, G.
Feng, Q. Gu, L. Wang, R. R. Wang, Q. Zhou, W. Li, SARS-CoV-2 detection with CRISPR
diagnostics, medRxiv, doi:10.1101/2020.04.10.023358.
21. J. N. Rauch, E. Valois, S. C. Solley, F. Braig, R. S. Lach, N. J. Baxter, K. S. Kosik, C. Arias, D.
Acosta-Alvear, M. Z. Wilson, A Scalable, Easy-to-Deploy, Protocol for Cas13-Based Detection of
SARS-CoV-2 Genetic Material, medRxiv, doi:10.1101/2020.04.20.052159.
22. U.S. Food and Drug Administration, Sherlock CRISPR SARS-CoV-2 Kit, (available at
https://www.fda.gov/media/137747/download).
23. J. S. Gootenberg, O. O. Abudayyeh, J. W. Lee, P. Essletzbichler, A. J. Dy, J. Joung, V. Verdine, N.
Donghia, N. M. Daringer, C. A. Freije, C. Myhrvold, R. P. Bhattacharyya, J. Livny, A. Regev, E. V.
Koonin, D. T. Hung, P. C. Sabeti, J. J. Collins, F. Zhang, Nucleic acid detection with CRISPR-
Cas13a/C2c2. Science. 356, 438–442 (2017).
24. O. O. Abudayyeh, J. S. Gootenberg, S. Konermann, J. Joung, I. M. Slaymaker, D. B. T. Cox, S.
Shmakov, K. S. Makarova, E. Semenova, L. Minakhin, K. Severinov, A. Regev, E. S. Lander, E. V.
Koonin, F. Zhang, C2c2 is a single-component programmable RNA-guided RNA-targeting CRISPR
effector. Science. 353, aaf5573 (2016).
25. C. Myhrvold, C. A. Freije, J. S. Gootenberg, O. O. Abudayyeh, H. C. Metsky, A. F. Durbin, M. J.
Kellner, A. L. Tan, L. M. Paul, L. A. Parham, K. F. Garcia, K. G. Barnes, B. Chak, A. Mondini, M. L.
Nogueira, S. Isern, S. F. Michael, I. Lorenzana, N. L. Yozwiak, B. L. MacInnis, I. Bosch, L. Gehrke,
F. Zhang, P. C. Sabeti, Field-deployable viral diagnostics using CRISPR-Cas13. Science. 360,
444–448 (2018).
26. J. S. Gootenberg, O. O. Abudayyeh, M. J. Kellner, J. Joung, J. J. Collins, F. Zhang, Multiplexed
and portable nucleic acid detection platform with Cas13, Cas12a, and Csm6. Science. 360, 439–
444 (2018).
27. A. East-Seletsky, M. R. O’Connell, D. Burstein, G. J. Knott, J. A. Doudna, RNA Targeting by
Functionally Orthogonal Type VI-A CRISPR-Cas Enzymes. Mol. Cell. 66, 373–383.e3 (2017).
28. S. Srivatsan, P. D. Han, K. van Raay, C. R. Wolf, D. J. McCulloch, A. E. Kim, E. Brandstetter, B.
Martin, J. Gehring, W. Chen, S. Kosuri, E. Q. Konnick, C. M. Lockwood, M. J. Rieder, D. A.
Nickerson, H. Y. Chu, J. Shendure, L. M. Starita, Seattle Flu Study Investigators, Preliminary
support for a “dry swab, extraction free” protocol for SARS-CoV-2 testing via RT-qPCR, bioRxiv,
doi:10.1101/2020.04.22.056283.
29. J. Joung, A. Ladha, M. Saito, M. Segel, R. Bruneau, M.-L. W. Huang, N.-G. Kim, X. Yu, J. Li, B. D.
Walker, A. L. Greninger, K. R. Jerome, J. S. Gootenberg, O. O. Abudayyeh, F. Zhang, Point-of-
care testing for COVID-19 using SHERLOCK diagnostics. medRxiv,
doi:10.1101/2020.05.04.20091231.
30. A. E. Calvert, B. J. Biggerstaff, N. A. Tanner, M. Lauterbach, R. S. Lanciotti, Rapid colorimetric
detection of Zika virus from serum and urine specimens by reverse transcription loop-mediated
isothermal amplification (RT-LAMP). PLoS One. 12, e0185340 (2017).
31. B. A. Rabe, C. Cepko, SARS-CoV-2 Detection Using an Isothermal Amplification Reaction and a
Rapid, Inexpensive Protocol for Sample Inactivation and Purification, medRxiv,
doi:10.1101/2020.04.23.20076877.
32. D. Kim, J.-Y. Lee, J.-S. Yang, J. W. Kim, V. N. Kim, H. Chang, The Architecture of SARS-CoV-2
Transcriptome. Cell. 181, 914–921.e10 (2020).
33. M. Waskom, O. Botvinnik, D. O’Kane, P. Hobson, S. Lukauskas, D. C. Gemperline, T.
Augspurger, Y. Halchenko, J. B. Cole, J. Warmenhoven, J. de Ruiter, C. Pye, S. Hoyer, J.
Vanderplas, S. Villalba, G. Kunter, E. Quintero, P. Bachant, M. Martin, K. Meyer, A. Miles,
Y. Ram, T. Yarkoni, M. L. Williams, C. Evans, C. Fitzgerald, Brian, C. Fonnesbeck, A. Lee,
A. Qalieh, mwaskom/seaborn: v0.8.1 (September 2017) (2017),
doi:10.5281/zenodo.883859.
34. J. D. Hunter, Matplotlib: A 2D Graphics Environment. Comput. Sci. Eng. 9, 90–95 (2007).
Acknowledgements
We would like to thank E. Rosenberg for kindly providing patient samples used in this study; the Harvard
Medical School Systems Biology Department for providing additional laboratory space to perform the
work; those researchers and laboratories who generously made SARS-CoV-2 sequencing data publicly
available to aid in our assay design; members of the Sabeti lab — E. Normandin, K. DeRuff, K. Lagerborg,
M. Bauer, M. Rudy, K. Siddle, A. Lin and A. Gladden-Young — for assisting with patient sample collection
and processing; H. Metsky, for his contributions to the assay design; M. Springer, the Springer lab, and
the Sabeti lab, notably H. Metsky, A. Lin, and N. Welch for their thoughtful discussions and reading of
the manuscript. Funding: Funding was provided by DARPA D18AC00006 and the Open Philanthropy
Project. J.A.-S. is supported by a fellowship from ”la Caixa” Foundation (ID 100010434, code
LCF/BQ/AA18/11680098). B.A.P. is supported by the National Institute of General Medical Sciences
grant T32GM007753. The views, opinions, and/or findings expressed should not be interpreted as
representing the official views or policies of the Department of Defense, US government, National
Institute of General Medical Sciences, or the National Institutes of Health. Competing interests: C.A.F.,
P.C.S., and C.M. are inventors on patent filings related to this work. J.E.L. consults for Sherlock
Biosciences, Inc. P.C.S. is a co-founder of, shareholder in, and advisor to Sherlock Biosciences, Inc, as
well as a Board member of and shareholder in Danaher Corporation.
Items included in Supplementary Materials
Supplementary Text
Figs. S1 to S10
Tables S1 to S4
References (35-38)
Other Supplementary Files
Table S5
Supplementary Protocol
| 2020 | Integrated sample inactivation, amplification, and Cas13-based detection of SARS-CoV-2 | 10.1101/2020.05.28.119131 | [
"Arizti-Sanz Jon",
"Freije Catherine A.",
"Stanton Alexandra C.",
"Boehm Chloe K.",
"Petros Brittany A.",
"Siddiqui Sameed",
"Shaw Bennett M.",
"Adams Gordon",
"Kosoko-Thoroddsen Tinna-Solveig F.",
"Kemball Molly E.",
"Gross Robin",
"Wronka Loni",
"Caviness Katie",
"Hensley Lisa E.",
"Be... | creative-commons |
Spatial organization of nuclear pores in Xenopus laevis oocytes
Linda Ravazzano a, Silvia Bonfanti a,b, Roberto Guerra a, Fabien Montel c, Caterina A. M. La
Portad,e, Stefano Zapperi∗,a,b
Received Xth XXXXXXXXXX 20XX, Accepted Xth XXXXXXXXX 20XX
First published on the web Xth XXXXXXXXXX 200X
DOI: 10.1039/b000000x
Nuclear pores are protein assemblies inserted in the nuclear envelope of eukaryotic cells, acting as main gates for communication
between nucleus and cytoplasm. So far, nuclear pores have been extensively studied to determine their structure and composition,
yet their spatial organization and geometric arrangement on the nuclear surface are still poorly understood. Here, we analyze
super-resolution images of the surface of Xenopus laevis oocyte nuclei during development, and characterize the arrangement
of nuclear pores using tools commonly employed to study the atomic structural and topological features of soft matter. To
interpret the experimental results, we hypothesize an effective interaction among nuclear pores and implemented it in extensive
numerical simulations of octagonal clusters mimicking typical pore shapes. Thanks to our simple model, we find simulated
spatial distributions of nuclear pores that are in excellent agreement with experiments, suggesting that an effective interaction
among nuclear pores exists and could explain their geometrical arrangement. Furthermore, our results show that the statistical
features of the geometric arrangement of nuclear pores do not depend on the type of pore-pore interaction, attractive or repulsive,
but are mainly determined by the octagonal symmetry of each single pore. These results pave the way to further studies needed
to determine the biological nature of pore-pore interactions.
1
Introduction
Genetic information in eukaryotic cells is well protected inside
the cell nucleus that is divided from the outside cytoplasm by a
membrane called nuclear envelope (NE). This segregation has
the advantage of protecting the genome from sources of dam-
age, but on the other hand communications based on exchange
of macro-molecules, such as messenger RNAs (mRNA) or
transcriptomic factors, are of vital importance during all the
cell life cycle, to control protein synthesis and instruct gene
expression1,2. The spatial architecture of the nucleus is cru-
cial for the interaction between the genome and protein com-
ponents of the nuclear complex and has a role in chromatin
reorganization during cellular differentiation.
Nuclear pores, large protein assemblies inserted in the
nuclear envelope,
are responsible for selective nucleo-
cytoplasmic transport, allowing the free diffusion of ions and
small molecules and acting as selective gates for import and
aCenter for Complexity and Biosystems, Department of Physics, University of
Milano, via Celoria 26, 20133 Milano, Italy
b CNR - Consiglio Nazionale delle Ricerche, Istituto di Chimica della Materia
Condensata e di Tecnologie per l’Energia, Via R. Cozzi 53, 20125 Milano,
Italy
c Universit´e de Lyon, ´Ecole Normale Sup´erieure de Lyon, Universit ´e Claude
Bernard, CNRS, Laboratoire de Physique, Lyon 69342, France
d Center for Complexity and Biosystems, Department of Environmental Sci-
ence and Policy, University of Milan, via Celoria 26, 20133 Milano, Italy
e CNR - Consiglio Nazionale delle Ricerche, Istituto di Biofisica, via Celoria
26, 20133 Milano, Italy
export of macromolecules, such as proteins and mRNAs3.
Furthermore, nuclear pores are also involved in the organi-
zation of the genome and they contribute to gene regulation
through physical interactions with chromatin4.
Since the discovery of nuclear pores in the 1950s, their
structure has been the subject of extensive experimental inves-
tigations using Electron Microscopy (EM) and Cryo-Electron
Tomography (cryo-ET)5.
Great effort was spent on deter-
mining the composition of a single pore in terms of different
protein complexes. Nuclear pores appear as modular assem-
blies of discrete constituents arranged with octagonal symme-
try around a central axis6. Further investigations identified
those discrete elements as multiple copies of about 34 protein
subunits (nucleoporins). These peculiar proteins are remark-
ably conserved throughout eukaryotes, showing similar fea-
tures in algae, yeast, vertebrates such as Xenopus Laevis, up
to human.
The characteristic shape of a nuclear pore consists of two
superimposed rings of nucleoporins, with eightfold symme-
try, one on the outer face of the nuclear membrane and one on
the inner face, with eight extended filaments departing from
each ring. In contrast, the center of the pore, which forms
the permeability barrier, is filled with disordered filaments of
phenylanine-glycine (FG) repeats1,5. In this context, the ex-
perimental observations have already been accompanied by in
silico studies involving all atoms and coarse-grained molecu-
lar dynamics simulations in order to to characterize the pecu-
liar structure of nuclear pore complexes (NPCs).7,8,9.
1–10 | 1
Despite great progress in understanding the structure of a
single nuclear pore, little is known on how nuclear pores are
distributed across the surface of the nuclear membrane (in an
average human cell there are approximately 2000–3000 nu-
clear pores) and whether and how they might interact among
each other. Interactions are likely modulated by the nuclear
lamina, a filamentous protein network underlying the nuclear
envelope, but how the interaction occurs is still unclear. Re-
cent observations in Drosophila show that nuclear pores are
arranged in a nonrandom manner with clusters that suggest
the presence of an effective mutual attractive interaction10
Earlier studies, interestingly, revealed that highly prolifer-
ative cells such as embryos or tumors have an high density
of nuclear pores on the nuclear membrane, while terminal dif-
ferentiated cells have fewer, suggesting a link between number
and distribution of pores and cell activity11. This link has been
further explored in another early study focused on the changes
in distribution of nuclear pores during spermatogenesis, fol-
lowing the evolution from spermatocytes to early spermatids.
In particular, a clear change in nuclear pore spatial organiza-
tion, from aggregation with hexagonal packing in pore rich
areas coexisting with large pore-free areas in spermatocytes to
a random distribution of pores in early spermatids, has been
observed12. A further step in our understanding of the role of
nuclear pore organization came from the observation of large
pore-free islands in HeLa S3 human cells. These islands dis-
perse with cell-cycle progression and reveal the importance of
lamin A/C in regulating the pore distribution13.
In a recent paper, Sell´es et al. performed super-resolution
microscopy on Xenopus laevis oocytes observing the varia-
tion of nuclear pore distribution on the nuclear membrane dur-
ing oocyte development14. In the present paper, we analyze
those experimental data14 to investigate the spatial distribu-
tions of nuclear pores across the nuclear membrane during the
development of Xenopus laevis oocytes. To this end, we use
tools typical of soft matter physics, such as the radial distribu-
tion function (RDF), local order parameters and Voronoi tes-
sellation. To model the spatial distribution of nuclear pores
observed experimentally, we introduce an effective interac-
tion among them. We first define a potential with octagonal
symmetry to properly model the shape of nuclear pores and
then perform extensive numerical simulations of interacting
nuclear pores, studying their behavior as the density varies. So
far attempts of modeling NPCs geometrical arrangement on
the nuclear surface and using computer simulations to deepen
our understanding on it were missing. With our work we try
to fill that gap proposing a first attempt to study in silico pe-
culiarities and features of the spatial distribution of nuclear
pores.
2
Materials and Methods
2.1
Experimental images
In this Section we analyze super-resolution experimental im-
ages of nuclear pores of Xenopus laevis oocyte by Sell´es et
al. 14. According to the stage of development of the oocyte15
we identify three groups of images: an early Stage II, an in-
termediate Stage IV, and a later Stage VI. A different number
of samples were taken at each stage, specifically: 6 samples
for Stage II, and 11 samples for both Stage IV and Stage VI.
The images are 2560×2560 pixels (px) wide, with 1 px cor-
responding to 10 nm. Examples of nuclear pore experimental
images are reported in Fig. 1a)-c).
2.1.1
Tracking of Nuclear Pores – We analyze the tra-
jectories of the nuclear pores with Trackpy v0.4.2, a Python
package for particle tracking in 2D, 3D, and higher dimen-
sions16. In particular, we first discriminate the nuclear pores
using the function trackpy.locate, whose working principle is
the following: i) preprocess the image by applying a bandpass
filter (i.e. performing a convolution with a Gaussian to remove
short-wavelength noise, and subtracting out long-wavelength
variations by subtracting a running average, in order to re-
tain intermediate scale features), ii) apply a threshold over
the color channels, and iii) locate all the peaks of brightness,
each referring to the position of a pore16,17. The parameters
used for the tracking are: diameter = 9 px the diameter, and
minmass, the minimum integrated brightness, working as a
threshold value. The latter value is chosen based on the sam-
ples: minmass = 0 (no threshold) is used in the high-density
samples of Stage II, while higher values of this parameter were
necessary to correctly detect pores in the noisier experimental
images of Stage IV and VI. The tracking procedure also allows
us to determine the density of pores defined as the number of
pores per unit area of the nuclear envelope from the experi-
mental samples: 34.9±2.3 NPC/µm2 for Stage II, 25.6±2.3
NPC/µm2 for Stage IV, 20.5 ± 1.7 NPC/µm2 for Stage VI.
The errors here represent the standard deviation computed on
the ensemble of samples for each developmental Stage. The
above density values are slightly lower than those computed
by Sell´es et al. 14. We attribute this discrepancy to the dif-
ferent tracking techniques employed and to an inherent uncer-
tainty related to experimental measurements performed with
super-resolution optical microscopy. In fact, with this tech-
nique, some fluorescent spots may be “fragmented” in the fi-
nal image, due to small microscope movements. Thus it can
happen that a single pore appears divided into several dots,
introducing a certain arbitrariness in the counting of pores.
2 |
1–10
2.2
Numerical Simulations
To model the interaction among nuclear pores, we consider
their peculiar octagonal shape, as observed in early experi-
mental studies6,18 (see e.g. Fig. 1a) and subsequently con-
firmed by structural studies on nucleoporins, and by recent
advances in experimental techniques such as cross-linking
mass spectrometry and cryo–electron tomography5. Coarse-
grained models are of key importance for understanding the
essential behaviors of biological phenomena without resorting
to detailed modeling of the molecular structure19. In the fol-
lowing paragraph, we describe the coarse-grained model pro-
viding the details on the form of the potential for the pore–pore
interaction, and the simulation protocol.
2.2.1
Model Potential for Nuclear Pores – To account
for the composite structure of each nuclear pore and its over-
all octagonal shape, we implement a coarse-grained model of
the pore, which consists of a central particle surrounded by
eight particles located at the vertices of a regular octagon, of
circumradius R. For simplicity, the pore is treated as a rigid
non-deformable object.
Taking experimental data as refer-
ence14,20, in all simulations we set R = 67.5 nm. The overall
interaction potential acting among the particles of two neigh-
boring pores is composed of three terms, each of which con-
sists of a Lennard-Jones (LJ) potential,
V(rij) = 4ε
�� σ
ri j
�12
−
� σ
ri j
�6�
, ri j < rcut
(1)
with rij = |ri −rj| is the distance between particle i and parti-
cle j, and ε, σ, and rcut parameters depend on the interaction
term:
• center-center interaction – the central particle of a pore
interacts repulsively with the central particle of a neigh-
boring pore. For this term we set εcc = 0.01 pg·µm2/µs2,
σcc = 0.12 µm, and the cutoff distance is set rcut
cc =
2
1/6σcc, to make the interaction purely repulsive. This
term is necessary to avoid non-physical configurations,
such as the case of overlapping pores that are otherwise
rarely encountered.
• center-vertex interaction – the central particle of a pore
interacts repulsively with the particle at the vertex of
a neighboring pore. Again, this is introduced to avoid
pore overlap and interpenetration.
LJ parameters for
this term are εcv = 0.01 pg·µm2/µs2, σcv = 0.08 µm,
rcut
cv = 2
1/6σcv.
• vertex-vertex interaction – the particle at the vertex of
a pore interacts with the particle at the vertex of a neigh-
boring pore.
We first considered the full LJ interac-
tion (long-range attractive, short-range repulsive), but the
purely repulsive case was also studied (see Supplemen-
tary Fig. S1). For the LJ interaction we set εvv = 5·10−4
pg·µm2/µs2, σvv = 0.02 µm, and the cutoff distance is
set to rcut
vv = 2.5σvv so to include the attractive part. For
the purely repulsive case we set rcut
vv = 2
1/6σvv. By con-
struction, in our model the overall pore–pore interaction
is mainly driven by the present term, due to the fact that
the central particles are surrounded by vertex particles.
In all cases, the LJ potentials are shifted to zero at the cutoff
distance to avoid any energy discontinuity. Fig. 1d reports the
total interaction energy of an octagonal pore (centered at the
origin) with a corner particle from a second pore, as a function
of the latter’s position. The resultant potential energy surface
(PES) shows a central strongly repulsive region (in blue) and
trapping regions (in red) concentrated near the eight corner
sites.
2.2.2
In silico nuclear pores configurations – The sim-
ulations of nuclear pore assemblies are performed using
LAMMPS22 with a timestep ∆t = 10−5 µs. The starting con-
figurations consist of 1000 randomly placed octagonal pores
confined in a square periodic box of side L = 40 µm. To mimic
the different experimental nuclear pore densities observed dur-
ing oocyte development, the random configurations are alter-
nately subjected to 105 steps of box compression at a constant
temperature T = 2
3
εvv
kB followed by a 2·105 steps of annealing
from high temperature T = εvv
kB down to T = 2
3
εvv
kB , in order to
allow thermally-assisted rearrangements. Such procedure is it-
erated until the required density is obtained, and a final energy
minimization at T = 0 K in 2×106 steps is performed. Follow-
ing the above protocol we obtain configurations with a density
of 20, 26, 36, 46, and 53 NPC/µm2. For each density value we
obtain 10 different realizations starting from different random
initial positions, in order to allow for proper statistical aver-
aging. Examples of nuclear pore configurations obtained with
numerical simulations are shown in Fig. 1e) and f).
2.3
Statistical analysis of nuclear pores structure
In this Section we describe three different quantities used to
provide a statistical comparison of the simulations with the
reference experimental data: the radial distribution function,
the hexatic order parameter and voronoi diagram.
2.3.1
Radial Distribution Function – To gain insight on
the local structure of the nuclear pore complex on the nu-
clear membrane, we analyze the radial distribution function
(RDF)23:
g(r) =
L2
2πrN2
N
∑
i=1
N
∑
j=1
j̸=i
⟨δ(r −rij)⟩
(2)
1–10 | 3
Fig. 1 Comparison of experimental and simulated nuclear pore images. – (a) Energy-filtering transmission electron microscopy
(EFTEM) images of Xenopus oocyte NEs embedded in thick amorphous ice. From those images the eight-fold symmetry of nuclear pores can
be appreciated. The four insets on the bottom left of the figure show higher magnification images. Correlation averages over 100 pores are
plotted in the bottom right inset, clearly revealing the shape of a single pore. Scale bar represents 200 nm. Reprinted from ’Cryo-electron
tomography provides novel insights into nuclear pore architecture: implications for nucleocytoplasmic transport’ Daniel Stoffler et al., Journal
of molecular biology, 2003, 328.1: 119-130.21, Copyright 2003, with permission from Elsevier. (b),(c) Portions of experimental images of
nuclear pores in Xenopus laevis oocyte at different developmental stages, respectively Stage II (b) and Stage VI (c), obtained using
super-resolution microscopy. Scalebars are 1µm. Panels (b) and (c) are adaptations from Sell´es et al. 14. Scalebars are 1 µm. (d) Potential
energy surface obtained from the modeled interaction between a pore and a corner of a neighboring pore: the blue area marks a strongly
repulsive region, while the red areas mark the trapping centers. (e),(f) Example of two configurations of nuclear pores obtained from numerical
simulations for comparison with experimental data. The density is 36 NPC/µm2 for (e) and 20 NPC/µm2 for (f). Scalebars are 1µm.
where N is the number of particles in the system, L is the sys-
tem size as specified above, and ri j is the distance between
particles i and j and the average ⟨ ⟩ is over particles.
The RDF is a key tool for the theory of monoatomic liq-
uids, to characterize amorphous colloidal solids24 and to study
glasses and the glass transition25.
2.3.2
Hexatic Order Parameter – To characterize the ge-
ometrical properties of the structure formed by the nuclear
pores, we compute for each particle the n-fold local orienta-
tional order parameter (hexatic order parameter):
ψn(ri j) =
1
nnn
nnn
∑
j=1
einθ(rij)
(3)
where nnn is the number of nearest neighbors of particle i,
θ(ri j) is the angle formed by the x axis and the vector rij
connecting particles i and j. Experimental nuclear pore com-
plexes of Xenopus laevis at different developmental stages
show ordered regions characterized by triangular and square
lattice14, therefore we focus our analysis on order parame-
ters i) with n = 6 for which |ψ6| = 1 for particles belonging
to a perfect hexagonal structure, and ii) with n = 4 for which
|ψ4| = 1 for particles belonging to a perfect square lattice. The
determination of nearest neighbor particles is performed us-
ing a cutoff distance σcut = 0.15 µm for the simulations and
σcut = 0.20 µm for the experimental images. Those values
have been chosen considering the typical inter-particle dis-
tance in simulated and experimental samples. For all isolated
particles (nnn < 2) we set ψn=0 .
2.3.3
Voronoi Tessellation – We finally analyze the
Voronoi diagram for the nuclear pore configurations, parti-
4 |
1–10
tioning the image into regions of convex polygons around the
center of each pore. This so-called Voronoi cells represent
the area of space containing all points that are closer to one
pore than to any other. For the experimental images, the co-
ordinates of the pore centers were obtained from the tracking
analysis and used as input for the Voronoi tesselation. For the
simulations only the central particle of each octagon is consid-
ered in the Voronoi analysis. By construction, each Voronoi
cell has polygonal shape, with a number of sides that corre-
sponds to the number of neighbors. To compute the Voronoi
tesselation we used the Python library Freud26, that allows to
account for periodic boundary conditions. Using this method
we extract for each pore the number of neighbors (pores are
considered neighbors if they share an edge in the Voronoi dia-
gram) and the size of each associated Voronoi cell.
3
Results
3.1
Global structure of NPC
From the RDF of the experimental samples, we note that at
high density (early stage of development of the oocyte) g(r)
shows a liquid-like shape, with two peaks clearly visible (see
Fig. 2a). In fact in monoatomic liquids, at short range g(r)
shows a pattern of peaks representing the nearest neighbour
distances, and at large r it tends to unity due to total loss
of order23. As the density decreases during oocyte develop-
ment (Fig. 2b,c), the second peak of the experimental g(r)
tends to disappear while the first peak tends to flatten out,
thus converging toward a gas-like phase in which the order is
lost. Previous analysis of the experimental images suggested
a significant presence of square lattice domains of nuclear
pores at low density (Stage II)14. For this reason, we have
looked for a specific peak in the g(r) function. In the case
of a regular square lattice, a second peak should be present at
xsq =
√
2x1 (where x1 represent the spatial coordinate of the
first peak of the g(r) function), however we notice that a peak
in this position is not visible. Nevertheless xsq falls on the
tail of the first peak (which is quite flat), suggesting that some
sparse regions with square lattice could be present inside the
amorphous liquid. Similar considerations have been done for
hexagonal structures at high densities: if a regular hexagonal
packing of pores is present, the g(r) should display a peak at
the position xhex =
√
3x1 and again, this is not the case, with
xhex falling only on the growing part of the second peak for
the high density configurations. Some hexagonal structures
at high density could be present but are not predominant, be-
ing the spatial distribution of NPCs mainly disordered. The
g(r) obtained from simulations with the octagonal attractive
potential (Fig. 2) reproduces the above described liquid-like
profile presenting just a few peaks emerging over an other-
wise flat profile. Furthermore, the positions of the peaks in the
simulated g(r) match nicely with the experiments (especially
for what concerns the first peak), suggesting that our model is
able to capture the relevant features of the effective interaction
among the nuclear pores on the nuclear membrane. This is a
non-trivial result, as the position of the g(r) peaks closely re-
flects the interactions held between the constituents, and are a
signature of each material and its peculiar properties as shown
in previous studies of noble gases or water27–30. Ultimately,
from this analysis, we can rule out the significant presence of
extensive crystalline regions.
3.2
Orientational Order
In Fig. 3, we report the calculated local order parameter
ψ6 and ψ4, as defined in Section 4, for some experimen-
tal samples and for the configurations obtained from simula-
tions. The cutoff to consider a pore as a neighbor was set to
σcut = 0.20 µm for the experiments and σcut = 0.15 µm for
the simulations. These values were chosen considering the
typical pore-pore distances. From Fig. 3a and Fig. 3c it can
be observed that only few pores are associated with square
symmetry (ψ4 ∼ 1). Instead, Fig. 3b and Fig. 3d show much
broader regions associated with triangular lattice structures
(ψ6 ∼ 1). In those samples, in presence of higher pore den-
sities, more nuclear pores belong to a regular structure and
some regions with clear hexagonal order appears. To further
investigate such behavior we computed the distribution of the
local order parameters P(ψ6) and P(ψ4) averaged over all the
samples for each value of the density (see Fig. 4). First, we
can observe that the distributions of the experimental samples
(Fig. 4a,b) are all unimodal, thus confuting the hypothesis of
two different coexisting phases suggested previously14. Sec-
ondly, we notice that for both ψ6 and ψ4 the distributions are
peaked at a value which increases with the density. The largest
mode value is obtained for ψ6 at the highest density (Stage II),
suggesting a preference for hexagonal structures in the dense
limit. The above trends of the distributions with the density
are well reproduced by the simulated NPCs, as reported in
Fig. 4c and Fig. 4d. For a more straightforward comparison
we have reported in Fig. 4e the average value of |ψ6| and |ψ4|
as a function of the pores density for both simulations and ex-
periments. For the former, we observe that both |ψ6| and |ψ4|
values slowly increase with the density, showing the same val-
ues until a density of 36 NPC/µm2 where a bifurcation occurs.
Beyond that density value, |ψ4| seems to saturate while |ψ6|
keeps increasing, thus favoring the hexagonal order at high
density. It is worth to note that, in the explored density range,
the |ψ6| value is far from approaching unity, corresponding
to a crystalline structure, indicating that much larger densities
would be required for such an ordered phase. Finally we note
that in Fig. 4e the points associated to the experimental data
do not fall exactly on the theoretical curves derived from the
1–10 | 5
Fig. 2 Radial Distribution Function of nuclear pores – The g(r) is reported for three different densities of the nuclear pores. Blue curves
are from averages over ten simulations with attractive octagon potential. Dashed red curves are obtained from the experimental images for
Stage II - (a), Stage IV - (b), Stage VI - (c).
simulations. This can be partially explained with the uncer-
tainties connected with the experimental observations of the
nuclear pores, that affect the density evaluation. Despite that,
the experimental points show a trend very close to that of the
simulation, with an initial overlapping of |ψ6| and |ψ4| values,
and a further bifurcation at higher density.
3.3
Properties of Voronoi cells
Examples of Voronoi tesselation performed on high density
NPC from the experimental images and from the simulated
configurations are reported in Fig. 5a and Fig. 5d, respectively.
The comparison in these high density samples shows similar
tessellation patterns in experiments and simulations. A statis-
tical analysis on the number of sides N of the Voronoi cells
at different densities (Fig. 5b and Fig. 5e) clearly shows that
the N = 6 occurrence increases with the density, and viceversa
for the the N = 4 occurrence. Therefore, the Voronoi analy-
sis enforces the idea, already anticipated above by the local
order parameter, that the hexagonal configuration is favored
at high densities at the expense of other kind of local order
arrangements.
Correspondingly, in agreement with experi-
mental observations, the presence of some square structures
at low density is also supported. However, we note at any
density a significant number of cells with N = 5, about half
between those with N = 4 and N = 6. We associate this with
the particular sensitivity of the Voronoi tesselation method to
“defect”, i.e. deviations with respect to the ideal symmetric
cases. To provide a further comparison, we report in Fig. 5c
and Fig. 5f the distribution of the Voronoi cells area. Again,
a good agreement between the experiments and our model is
obtained, with much narrower distributions at higher densi-
ties, shifting toward higher area values and broadening out as
density decreases.
4
Discussion and Conclusions
A first attempt to investigate the spatial distribution of nuclear
pores goes back to the ’70s, when the positions of NPCs on
the surface of rat kidney nuclei was observed and distances
among them measured 31. Already in this study, some regu-
larities were found in the distribution of pore-pore distances
measured in the samples, suggesting a non random spatial dis-
tribution and some peaks corresponding to hexagonal struc-
tures, even if the statistics was too poor to reach further con-
clusions. More recently, Sell´es et al. investigated the angular
distribution between first neighbors of nuclear pores, reveal-
ing no preferential angles for Stage II and IV Xenopus laevis
oocytes (high density) and two distinct peaks at 90◦ and 180◦
for later Stage VI, suggesting the presence of square lattice
regions at low density14. In our study, from the analysis of
the radial distribution function g(r) of the nuclear pores on
the nuclear membrane of Xenopus laevis oocytes, we could
not observe peaks in correspondence of peculiar geometrical
structures, meaning that even though some crystalline regions
are present, they are quite rare and do not statistically influ-
ence the overall NPC spatial distribution.
Interestingly by
analysing the g(r) of the nuclear pores, we were able to iden-
tify an amorphous, liquid-like structure in which, in the early
phase of oocyte development (when NPC density is high),
long-range order is soon lost. On the other hand, as the oocyte
develops, the nuclear pore density decreases and g(r) shows
a behaviour compatible with a more dilute, gas-like system.
From a biological point of view, the early stages of oocyte de-
velopment are associated with intense transcriptional activity,
6 |
1–10
Simulations
(a)
(b)
(c)
(d)
Experiments
Fig. 3 Color maps of local order parameters – Snapshots of pores colored as a function of the local order parameter: (a) a zoomed region
of an experimental sample at Stage VI, for which we estimate a density of 20.5±1.7 NPC/µm2 ; (b) a zoomed region of an experimental
sample at Stage II, for which we estimate a density of 34.9±2.3 NPC/µm2; (c) and (d) a simulation box with pores density 20 NPC/µm2 and
36 NPC/µm2 respectively.
as the oocyte needs to build up a huge reserve of gene prod-
ucts such as mRNAs, tRNAs and proteins in order to correctly
fulfil its future role after fertilisation. Once the necessary ma-
ternal mRNAs have been copied, transcriptional activity in the
later stages of oocyte development becomes lower 32. These
changes in transcriptional activity could be linked to changes
in the spatial distribution of NPCs during oocyte development,
particularly changes in density. It would be extremely interest-
ing to further explore this connection from a biophysical point
of view, e.g. by trying to quantify the flow of matter through
the pores, (as has already been done in some previous kinetic
studies, which showed that a single NPC can allow a mass
flow of nearly 100 MDa/s33), at different stages of the cell’s
life-cycle. The positions of the peaks in the g(r) we computed
for nuclear pores are another key point of our results. Indeed,
it is known from the physics of matter that the positions of the
peaks of the radial distribution function and their relative dis-
tances give actual information about the geometrical arrange-
ment of the particles within a material, and are a signature of
the material itself 34. In particular, the peak positions allow to
indirectly infer the type of interactions among the constituents
of a specific material. Here, we have shown that the eight-fold
potential used to model the NPC in our simulations is able
to nicely reproduce the experimental g(r) peak locations. In
particular, the position of the first peak obtained from the sim-
ulations is in excellent agreement with Sell´es et al. 14 and with
previous observations 12,31. We also checked that assuming a
simple LJ potential acting among the pores (i.e. with spheri-
cal symmetry) with parameters compatible with experimental
pore sizes, the prediction of the g(r) peaks is not in agree-
ment with experimental results (see Supplementary informa-
tion). Even if the hypothesis of a pore-pore potential with
spherical symmetry (perhaps with an effective interaction size
for the pore that does not coincide with its physical size) can
not be excluded, our work suggests that the octagonal shape
of the pore and the associated eight-fold symmetry of its inter-
action potential plays a crucial role in determining the correct
spatial distribution of the pores. These facts are worth to no-
tice, since our simplified model based on the assumption of
an effective eight-fold pore-pore interaction is able to catch a
crucial signature of the spatial distributions of nuclear pores,
the radial distribution function peaks positions, even if the in-
teraction details (e.g. if the pore-pore potential is attractive
or repulsive) are not known (see Supplementary). Hopefully
1–10 | 7
Fig. 4 Distribution of the local order parameters – The distribution of the local order parameters at different pores densities is reported for
(a),(b) experimental samples and (c),(d) for simulated configurations; (e) the average |ψ6| and |ψ4| values as a function of the density. The
dashed lines report the values obtained from the simulations, with shadows highlighting the respective standard deviation. The points
represent the values computed from the experimental samples, with vertical errorbars for the standard deviation, and horizontal errorbars
reporting the error on the density, as described in Section 2.1.1.
this could help to deepen the investigation of the nature of the
pore-pore interplay, allowing to study also in silico an inter-
action that in reality is not fully understood under a biological
point of view. Pore-pore interactions are unlikely to be direct,
but rather mediated by the lamin scaffold through complex in-
teractions that are hard to model explicitly. In this sense, our
assumption that nuclear pore arrangement can be modeled as
assembly of interacting octagons is possibly oversimplified,
since after development pores are stuck within the lamina and
are unlikely to diffuse. We can, however, imagine that as the
nuclear envelope is formed the nuclear pores are arranged in
a way that is dictated by their geometry and which could then
be captured by our model. Since extensive MD simulations
of lamina filaments forming a three-dimensional network be-
neath the nuclear envelope have recently been performed35, it
would be interesting to try to go further in modelling the outer
regions of the cell nucleus, linking the lamina network and the
spatial distribution of the nuclear pores.
Coming back to our analysis, the spatial distribution of
NPCs, investigated through the local order parameters shows
that at high density the pores tend to arrange following the tri-
angular lattice. Even though the g(r) does not show explicit
peaks in correspondence to a triangular lattice, the study of Ψ6
(Fig. 3) and the Voronoi tessellation method (Fig. 5) prove that
at high density islands of six-fold symmetry packed pores ap-
pear. Noticeably, such behavior has been already reported in
previous experimental observations. During apoptosis the dis-
tribution of nuclear pores on the cell nucleus strongly changes,
bringing the NPCs to be highly concentrated in small regions
of the nuclear envelope (on mouse cell nuclei) and leaving the
rest of the surface pore-free. Those clusters of pores showed
a hexagonal packing and were supposed to be correlated with
diffuse chromatine areas36. Occasional areas of very regular
hexagonal packing of nuclear pores have been also observed
to emerge during the development of male germ cells, in ro-
dent spermatocytes12. Those facts open interesting questions
on how the geometrical disposition of the pores in some areas,
or even more simply, their density, are influenced by the un-
8 |
1–10
Experiments
Simulations
(a)
(b)
(c)
(f)
(e)
(d)
Number of facets
Voronoi cells area [μmμmm2]
P(N)
P(V)
Voronoi cells area [μmμmm2]
Number of facets
P(N)
P(V)
Number of facets
Number of facets
Fig. 5 Voronoi tesselation applied to nuclear pores – Examples of Voronoi tesselation are provided for (a) an experimental samples at
Stage II and (d) for a simulation with density 36 NPC/µm2; (b) and (e) the histogram of the number of Voronoi cell facets, for different
densities; (c) and (d) the corresponding distribution of the Voronoi cells area.
derlying nuclear activity and on what are the biological causes
responsible for the effective interaction among NPCs. Consid-
ering the pores under a geometrical and topological point of
view, underlying the importance of their octagonal shape, like
our simple model does, could be extremely interesting also in
the contest of membranes studies. In fact in a recent paper by
Torbati et al.37, the authors studied the mechanical stability
of the lipid bilayer membrane of the nuclear envelope, con-
sidered as two concentric membrane shells fused at numerous
sites with toroid-shaped nuclear pores (here simply modeled
as circular holes). Using mechanistic arguments based on elas-
ticity, they showed that in- and out-of-plane stresses can give
rise to the pore geometry and the geometric topology observed
in cell nuclei, finding simulated interpore distances in good ac-
cord with the ones observed in mammalian cells nuclei. How
octagons can contribute to stabilize the curvature of a spheri-
cal membrane1 and how they tend to be spatially arranged on
such a geometry could be an issue to consider to better clarify
the process of nuclear pores formation.
Acknowledgements
We thank Zoe Budrikis for preliminary developments of the
simulation code.
References
1 K. E. Knockenhauer and T. U. Schwartz, Cell, 2016, 164,
1162–1171.
2 H. Ris, Scanning, 1997, 19, 368–375.
3 S. R. Wente and M. P. Rout, Cold Spring Harbor perspec-
tives in biology, 2010, 2, a000562.
4 A. Buchwalter, J. M. Kaneshiro and M. W. Hetzer, Nature
Reviews Genetics, 2019, 20, 39–50.
5 D. H. Lin and A. Hoelz, Annual review of biochemistry,
2019, 88, 725–783.
6 P. Unwin and R. Milligan, The Journal of cell biology,
1982, 93, 63–75.
7 R. Gamini, W. Han, J. E. Stone and K. Schulten, PLoS
computational biology, 2014, 10, e1003488.
8 L. Miao and K. Schulten, Biophysical journal, 2010, 98,
1658–1667.
1–10 | 9
9 J. R. Perilla, B. C. Goh, C. K. Cassidy, B. Liu, R. C.
Bernardi, T. Rudack, H. Yu, Z. Wu and K. Schulten, Cur-
rent opinion in structural biology, 2015, 31, 64–74.
10 J. Cheng, E. S. Allgeyer, J. H. Richens, E. Dzafic, A. Pa-
landri, B. Lewk´ow, G. Sirinakis and D. St Johnston, Jour-
nal of Cell Science, 2021.
11 G. Maul, L. Deaven, J. Freed, L. M. Campbell and W. Be-
cak, Cytogenetic and Genome Research, 1980, 26, 175–
190.
12 D. W. Fawcett and H. E. Chemes, Tissue and Cell, 1979,
11, 147–162.
13 K. Maeshima, K. Yahata, Y. Sasaki, R. Nakatomi,
T. Tachibana, T. Hashikawa, F. Imamoto and N. Imamoto,
Journal of cell science, 2006, 119, 4442–4451.
14 J. Sell´es, M. Penrad-Mobayed, C. Guillaume, A. Fuger,
L. Auvray, O. Faklaris and F. Montel, Scientific reports,
2017, 7, 1–8.
15 J. N. Dumont, Journal of morphology, 1972, 136, 153–
179.
16 D.
Allan,
C.
van
der
Wel,
N.
Keim,
T.
A.
Caswell, D. Wieker, R. Verweij, C. Reid, Thierry,
L.
Grueter,
K.
Ramos,
apiszcz,
zoeith,
R.
W.
Perry, F. Boulogne, P. Sinha, pfigliozzi, N. Bruot,
L.
Uieda,
J.
Katins,
H.
Mary
and
A.
Ahma-
dia,
soft-matter/trackpy:
Trackpy
v0.4.2,
2019,
https://doi.org/10.5281/zenodo.3492186.
17 J. C. Crocker and D. G. Grier, Journal of colloid and in-
terface science, 1996, 179, 298–310.
18 R. Reichelt, A. Holzenburg, E. Buhle, M. Jarnik, A. Engel
and U. Aebi, Journal of Cell Biology, 1990, 110, 883–894.
19 M. Hoferer, S. Bonfanti, A. Taloni, C. A. La Porta and
S. Zapperi, Physical Review E, 2019, 100, 042410.
20 M. Eibauer, M. Pellanda, Y. Turgay, A. Dubrovsky,
A. Wild and O. Medalia, Nature communications, 2015,
6, 1–9.
21 D. Stoffler, B. Feja, B. Fahrenkrog, J. Walz, D. Typke and
U. Aebi, Journal of molecular biology, 2003, 328, 119–
130.
22 S. Plimpton, Journal of computational physics, 1995, 117,
1–19.
23 J.-P. Hansen and I. R. McDonald, Theory of simple liquids:
with applications to soft matter, Academic Press, 2013.
24 P. K. Jana, M. J. Alava and S. Zapperi, Physical Review E,
2018, 98, 062607.
25 R. Br¨uning, D. A. St-Onge, S. Patterson and W. Kob, Jour-
nal of Physics: Condensed Matter, 2008, 21, 035117.
26 V. Ramasubramani, B. D. Dice, E. S. Harper, M. P.
Spellings, J. A. Anderson and S. C. Glotzer, Computer
Physics Communications, 2020, 254, 107275.
27 J. Yarnell, M. Katz, R. G. Wenzel and S. Koenig, Physical
Review A, 1973, 7, 2130.
28 E. Ermakova, J. Solca, H. Huber and D. Marx, Chemical
physics letters, 1995, 246, 204–208.
29 L. B. Skinner, C. Benmore, J. C. Neuefeind and J. B.
Parise, The Journal of chemical physics, 2014, 141,
214507.
30 G. Camisasca, H. Pathak, K. T. Wikfeldt and L. G. Petters-
son, The Journal of chemical physics, 2019, 151, 044502.
31 G. G. Maul, J. W. Price and M. W. Lieberman, The Journal
of cell biology, 1971, 51, 405–418.
32 A. Schlune, V. Shahin, K. Enss, H. Schillers and H. Ober-
leithner, Journal of cellular biochemistry, 2006, 98, 567–
576.
33 K. Ribbeck and D. G¨orlich, The EMBO journal, 2001, 20,
1320–1330.
34 D. Chandler, Mechanics. Oxford University Press, Oxford,
UK, 1987, 40, year.
35 K. T. Sapra, Z. Qin, A. Dubrovsky-Gaupp, U. Aebi, D. J.
M¨uller, M. J. Buehler and O. Medalia, Nature communi-
cations, 2020, 11, 1–14.
36 E. Falcieri, P. Gobbi, A. Cataldi, L. Zamai, I. Faenza and
M. Vitale, The Histochemical Journal, 1994, 26, 754–763.
37 M. Torbati, T. P. Lele and A. Agrawal, Proceedings of the
National Academy of Sciences, 2016, 113, 11094–11099.
10 |
1–10
| 2021 | Spatial organization of nuclear pores in oocytes | 10.1101/2021.09.01.458492 | [
"Ravazzano Linda",
"Bonfanti Silvia",
"Guerra Roberto",
"Montel Fabien",
"La Porta Caterina A. M.",
"Zapperi Stefano"
] | null |
Model balancing: consistent in-vivo kinetic constants and
metabolic states obtained by convex optimisation
Wolfram Liebermeister1,2
1 Universit´e Paris-Saclay, INRAE, MaIAGE, 78350, Jouy-en-Josas, France
2 Institut f¨ur Biochemie, Charit´e – Universit¨atsmedizin Berlin, Germany
Abstract
Enzyme kinetic constants in vivo are largely unknown, which limits the construction of large metabolic
models. In theory, kinetic constants can be fitted to measured metabolic fluxes, metabolite concentrations,
and enzyme concentrations, but these estimation problems are typically non-convex. This makes them hard
to solve, especially if models are large. Here I assume that the metabolic fluxes are given and show that
consistent kinetic constants, metabolite levels, and enzyme levels can then be found by solving a convex
optimality problem. If logarithmic kinetic constants and metabolite concentrations are used as free variables
and if Gaussian priors are employed, the posterior density is strictly convex. The resulting estimation method,
called model balancing, can employ a wide range of rate laws, accounts for thermodynamic constraints on
parameters, and considers the dependences between flux directions and metabolite concentrations through
thermodynamic forces.
It can be used to complete and adjust available data, to estimate in-vivo kinetic
constants from omics data, or to construct plausible metabolic states with a predefined flux distribution. To
demonstrate model balancing and to assess its practical use, I balance a model of E. coli central metabolism
with artificial or experimental data. The tests show what information about kinetic constants can be extracted
from omics data and reveal practical limits of estimating kinetic constants in vivo.
Keywords: Metabolic model, enzyme kinetic constant, parameter estimation, convex optimality problem, param-
eter balancing, enzyme cost minimisation
1
Introduction
The number of metabolic network reconstructions is constantly growing, and there have been attempts to convert
metabolic networks automatically into kinetic models.
To build models with plausible parameter values and
metabolic states (characterised by enzyme levels, metabolite levels, and fluxes), one needs to reconstruct the
metabolic network, add allosteric regulation arrows, choose enzymatic rate laws, find kinetic constants, and
make sure the model shows plausible metabolic states. These subproblems have been addressed in various ways.
Pathway models have been built from in-vitro enzyme kinetics [1, 2]. To simplify model construction and to replace
unknown rate laws, standardised rate laws have been proposed [3, 4], used for automatic model generation [5],
and evaluated for their practical use [6]. In-vitro kinetic constants, available from the Brenda database [7, 8], are
widely used and unknown kcat values have been estimated by machine learning [9]. Directly inserting measured or
sampled kinetic constants into models can lead to inconsistencies because thermodynamic dependencies between
kinetic constants will be ignored. To address this problem, methods to construct consistent parameter sets have
been devised [10, 11, 12, 13, 14] and applied in modelling [15]. In parallel, there have been attempts to estimate
kinetic constants in vivo from flux, metabolite, and enzyme data [16, 17]. Methods for parameter fitting have
been developed and benchmarked [18, 19], and the question of parameter identifiability has been addressed [20].
1
Large models have been parameterised [21, 22], and pipelines for model parameterization have been developed
[23, 24]. Finally, even if parameters are unknown, methods for parameter sampling and ensemble modelling allow
to find plausible parameter sets [25] and to draw conclusions about possible dynamic behaviour [26].
The key problem here is to obtain realistic, consistent values of kinetic constants. In-vivo values are hard to
measure, and in-vitro values as proxies may be unreliable – or at least, this is hard to assess unless in-vivo values
are known.
So how can we infer model parameters from omics data?
To obtain realistic model parameters
and metabolic states, various types of measurement data must be combined. In an ideal case, in which kinetic
constants and enzyme concentrations are precisely known, metabolite levels and fluxes could be computed by
simulating the model. In another ideal case, in which enzyme levels, metabolite levels and fluxes are precisely
known for a number of states, we can solve for the kinetic constants. In reality, we are in between these two
cases: data of different types are available, but these data are incomplete and too noisy to be used directly in
models. In practice, in model construction we may have different aims, for example (i) finding consistent kinetic
constants in plausible ranges; (ii) adjusting and completing a data set of measured kinetic constants to obtain
consistent model parameters; (iii) estimating in-vivo kinetic constants from omics data (measured enzyme levels,
metabolite levels, and fluxes); (iv) completing and adjusting omics data for consistent metabolic states, which
may involve predictions of physilogical metabolite concentrations [27, 28] or predictions of metabolite and enzyme
concentrations based on resource allocation principles [29, 30, 31].
Taking all this together, our general task is not only to fit kinetic constants, but also to reconstruct consistent
metabolite concentrations, enzyme concentration, and metabolic fluxes, for a given model and based on data for
all these quantities. A notorious problem in model fitting is that data are uncertain, inconsistent and incomplete.
Thus, in our estimation task uncertainties in data and estimated parameters need to be quantified. Luckily, we
can employ some further constraints: our solutions must satisfy some general physical laws, namely thermody-
namic relations (Wegscheider conditions and Haldane relationships) between kinetic constants [3, 4] and relations
between kinetic constants and metabolic variables (e.g. the flux sign constraints, which couple flux directions,
equilibrium constants, and metabolite levels). Moreover, we can use prior distributions (for kinetic constants,
metabolite levels, or enzyme levels); and we can use measured (in vitro) parameter values as additional data (of
course, these data cannot be used as test data anymore). However, one problem remains. The resulting optimality
problems, e.g. in maximum-likelihood estimation, will be non-convex: they may show multiple local optima and
globally optimal solutions may be hard to find, especially with larger networks.
Here I address the general problem of finding consistent kinetic constants and metabolic states, based on het-
erogenous (kinetic, metabolomics, and proteomics) data. Under the strong assumption that the metabolic fluxes
are known (from measurements or previous calculations), I show that parameter fitting in kinetic metabolic models
can be formulated as a convex optimality problem. The estimation method, called model balancing, uses the
following input data: measured or assumed values of kinetic constants (which may be incomplete and uncertain),
and measured or assumed values of metabolite and enzyme levels in a number of metabolic states (which may be
incomplete and uncertain); and it relies on precise metabolic fluxes from one or several metabolic states (which
may be stationary or non-stationary). It determines a set of kinetic parameters and state variables (metabolite and
enzyme levels for all metabolic states in question) that are consistent with the rate laws and all other dependencies
in the model, plausible (i.e. respecting constraints and prior distributions), and resemble the data (showing large
likelihood values). For the estimation, we can either follow a maximum-likelihood approach (leading to a convex
optimality problem) or a Bayesian maximum-posterior approach (where Gaussian priors ensure strict convexity).
The maximum-posterior problem has a single solution which can be found be gradient-descent methods, and
also posterior sampling is facilitated by strict convexity. Model balancing relies on two main assumptions: (i)
all fluxes are predefined and thermodynamically consistent (i.e. infeasible cycle fluxes must be excluded) and (ii)
kinetic constants and metabolite concentrations are treated on logarithmic scale, while enzyme concentrations
are treated on absolute scale. Model balancing builds upon some methods for metabolic model construction:
2
parameter balancing [11, 14], elasticity sampling [32], and enzyme cost minimisation [31], which I review in the
discussion section.
2
Parameter estimation in kinetic models as a convex problem
2.1
Estimating kinetic constants from omics data
To see how information about in-vivo kinetic constants is extracted from omics data, let us review an existing
approach. To estimate kcat values, Davidi et al. [16] compared measured proteomics data to flux data obtained
from flux balance analysis (FBA), without presupposing any knowledge of metabolite concentrations or specific
kinetic laws. The method works as follows. We assume unknown rate laws of the form v = e k(c) (with flux
v and enzyme concentration e), where the catalytic rate k depends on (unknown) metabolite levels (vector c)
and can vary between zero and a value kcat (called turnover rate or catalytic constant). To determine kcat from
data, we consider a cell in different metabolic states and assume that each enzyme reaches its maximal capacity
in at least one of these states. Based on this assumption, a kcat value is estimated by computing the empirical
catalytic rates v(s)/e(s) in all states and taking their maximum value. The method was applied to a large number
of enzymes in E. coli, and the estimated kcat values were found to resemble the measured in-vivo values. Some
of the deviations could be explained by enzyme kinetics and thermodynamics, but this was not quantitatively
modelled. The limitations of this method are clear: since the max function is only sensitive to the highest value,
one high outlier value can completely distort the result. Such outliers may arise if a small protein level, due to
measurement errors, appears even smaller. But aside from this practical problem, what if the basic assumption
is not satisfied? We cannot be sure that an enzymes reaches its maximal capacity in one of the samples, so the
estimated in-vivo kcat value should be seen as a lower bound kcat ≥ maxs v(s)/e(s). But how far is this bound
from the true in-vivo kcat value?
If we manage to explain the (non-maximal) catalytic rates in the different states, can we maybe obtain a better
estimate of the true kcat value, even if this value is reached in none of the samples? To do this, we need to
consider metabolic concentrations and enzyme kinetics, i.e. the functional form of kl(c). A typical form of k(c)
for a uni-uni reaction, the Michaelis-Menten kinetics, is given by v = k+
cat s/Ks−k−
cat p/Kp
1+s/KS+p/Kp
[4] or, in factorised form
[33], by v = k+
cat · ηrev(c) · ηsat(c), where the efficiency terms ηrev (for reversibility, or thermodynamics) and ηsat
(for enzyme saturation and allosteric regulation) are numbers between 0 and 1 depending on metabolite levels.
If the efficiency terms are close to 1, then k approaches its maximal value kcat; but normally k is lower. Given
the rate laws, and given data for fluxes v, metabolite levels c, and enzyme levels e, we might be able to estimate
kcat and KM values, even if the maximal efficiency is not reached in any of the samples.
2.2
Metabolic model and statistical estimation model
Let us start by stating the estimation problem (shown in Figure 1 (a)). We consider a kinetic metabolic model with
thermodynamically consistent modular rate laws [4] and kinetic constants (e.g. equilibrium constants1, catalytic
constants and Michaelis-Menten constants) in a vector q = ln p. The model shows a number of metabolic states,
each characterised by a flux vector v(s), a metabolite concentration vector c(s), and an enzyme concentration
vector e(s). These states can be stationary (with steady-state fluxes) or non-stationary (e.g. snapshots from a
dynamic time course). The model formulae define dependencies among model parameters and state variables.
The kinetic constants in a network are interdependent because of physical laws [3, 4]. Each rate law contains a
1Equilibrium constants are determined by thermodynamics and do not depend on specific enzymes, but for simplicity I will count
them among the kinetic constants.
3
Objective function
(minimal metabolite
plus enzyme cost)
forces θ
Thermodynamic
concentrations e
Enzyme
(several states)
data,
bounds
Prior,
data
Prior,
Data,
bounds
(several states)
concentrations c
Metabolite
Fluxes v
(several states,
predefined)
All kinetic
parameters k
forces θ
Thermodynamic
Free kinetic
concentrations e
Enzyme
(several states)
Prior,
bounds
data,
bounds
Prior,
data
Prior,
Data,
bounds
Data
bounds
pseudo values,
(several states)
concentrations c
Metabolite
Fluxes v
(several states,
predefined)
Enzyme levels
Metabolite levels
Fluxes
Enzyme levels
Metabolite levels
Fluxes
Kinetic constants
catalytic constants, ...)
(Michaelis−Menten constants,
Kinetic metabolic model
Metabolic state 1:
Metabolic state 2:
...
(a) Kinetic model and metabolic states
(e) Model balancing with known kinetics
(c) Model balancing (general problem)
forces θ
Thermodynamic
Free kinetic
Prior,
bounds
data,
bounds
Prior,
Data,
bounds
Data
bounds
pseudo values,
(several states)
concentrations c
Metabolite
All kinetic
parameters k
concentrations e
Enzyme
(several states)
(several states)
concentrations c
Metabolite
Fluxes v
(several states,
predefined)
All kinetic
parameters k
All kinetic
forces θ
Thermodynamic
Enzyme concentrations e
linear in log kind
linear in log k
Metabolite
concentrations c
Fluxes v
convex in log c
and log k
(b) Physical dependencies between model variables
(f) Enzyme cost minimisation
(d) Parameter / state balancing
All kinetic
parameters k
parameters k
Reciprocal
rate laws
Free kinetic
parameters k
parameters k
parameters k
ind
ind
ind
Figure 1: Parameter estimation in kinetic metabolic models. (a) Kinetic model and metabolic states. A model
is parameterised by kinetic constants (e.g. equilibrium constants, catalytic constants, and Michaelis-Menten con-
stants) and gives rise to a number of metabolic states (characterised by enzyme levels, metabolite levels, and
fluxes). These states may be stationary (with steady-state fluxes) or not (e.g. states during dynamic time courses).
(b) Dependencies between kinetic constants and state variables. All kinetic constants are described on logarith-
mic scale, and a subset of kinetic constants determines all other kinetic constants through linear relationships.
If kinetic constants, metabolite levels, and fluxes are known, the enzyme levels can be computed from rate laws
and fluxes: each enzyme level is a convex function of the (logarithmic) kinetic constants and metabolite levels.
(c) Parameter estimation. Kinetic constants and metabolite levels (for a number of metabolic states) are the
free variables of a statistical model. Dependent kinetic constants, thermodynamic driving forces, and enzyme
levels (bottom) are treated as dependent variables, and the fluxes (top right) are predefined. For estimating the
variables, priors and available data may be used. The other subfigures show similar estimation and optimisation
methods, in which (d) only kinetic data are balanced (no metabolic data), (e) only metabolic data are balanced
(kinetic parameters are predefined), or (f) enzyme and metabolite levels are optimised for a low biological cost.
forward and a reverse catalytic constant as well as the Michaelis-Menten constants2, and all these parameters in
a model may depend on each other via Haldane relationships and Wegscheider conditions.
The dependencies between model variables are summarised in appendix A.1. To satisfy all parameter dependencies
2Activation and inhibition constants are independent of all other constants and are therefore independent parameters.
4
in our model, we introduce a set of independent kinetic parameters (independent equilibrium constants, Michaelis-
Menten constants, and velocity constants) from which all remaining constants can be derived (see Figure 1 (b),
top left). The vector p contains all kinetic constants. In each metabolic state s, the rate laws define equalities
v(s)
l
= e(s)
l
kl(p, c(s)) for enzyme levels e(s)
l , metabolite levels c(s)
i , and catalytic rates kl. By inverting this
equation, the enzyme levels can be written as functions
e(s)
l
=
v(s)
l
kl(p, c(s))
(1)
of kinetic constants, metabolite levels, and fluxes (Figure 1 (b), bottom). The signs of thermodynamic forces
given by a vector θ(s) = ln keq − N⊤
all ln c(s) determine the flux directions (where reactions with vanishing fluxes
are always allowed). This law holds for all thermodynamically feasible rate laws.
Given a metabolic model with all its variables and dependencies, we now define estimation problems (see Figure 1
(c)). The most general aim is to estimate kinetic constants, metabolite profiles and enzyme profiles in a number
of metabolic states. Available data may comprise kinetic constants, metabolite and enzyme concentrations, and
possibly thermodynamic forces in a number of metabolic states, and metabolic fluxes in the same metabolic
states. All data may be uncertain and incomplete, except for the fluxes, which must be precisely given. Moreover,
we may use prior distributions and impose upper and lower bounds on the model parameters and on metabolite
and enzyme levels. In the model, all dependencies must be satisfied. To get to a convex optimality problem,
we treat the (logarithmic3) independent kinetic constants and the (logarithmic) dependent kinetic constants
and (logarithmic) metabolite concentrations as free variables, while the (non-logarithmic) enzyme levels and
thermodynamic forces are dependent variables to be computed from kinetic constants, metabolite levels, and
fluxes. The vector of free variables (logarithmic kinetic constants and metabolite concentrations) is constrained
by thermodynamic laws, and the resulting feasible space is a convex polytope. We may consider two variants of
the estimation problem, maximum-likelihood estimation and maximum-posterior estimation [34]. In maximum-
likelihood estimation, we minimise the negative log-likelihood (or “likelihood loss”), a convex function on the
feasible polytope. In maximum-posterior estimation, we consider Gaussian priors, which make the negative log-
posterior density (or “posterior loss”) strictly convex on the feasible polytope. This means: the posterior mode is
unique and can be obtained by convex optimisation. Formulae are summarised in appendix A.1.
2.3
A simplified estimation problem: fitting of metabolite and enzyme levels
Before we get to the full model balancing proablem, let us first assume that the kinetic constants are known and let
us estimate metabolite and enzyme levels for a single steady state4, based on data with error bars for (some or all)
metabolite and enzyme levels. To fit consistent metabolite and enzyme levels to these data, we maximise either
their likelihood or the posterior density. For the log-metabolite vector x, we assume an uncorrelated Gaussian
prior (with mean vector ¯xprior and covariance matrix Cx,prior = Dg(σx,prior)2) and lower and upper bounds
(possibly different for each metabolite). For the enzyme vector e, we assume an uncorrelated Gaussian prior (with
mean vector ¯eprior and covariance matrix Ce,prior = Dg(σe,prior)2). Negative values are not allowed (el ≥ 0).
The possible logarithmic metabolite profiles x form a convex polytope Px in log metabolite space [31]. This
shape of this polytope is defined by physiological upper and lower bounds and by thermodynamic constraints,
depending on flux directions and equilibrium constants. The logarithmic metabolite concentrations xi, our free
variables, determine the enzyme levels el through Eq. (1), and the enzyme levels are convex functions on the
metabolite polytope. As a consequence, the likelihood function is convex. Thus, to define an estimation problems,
3Natural logarithms are used throughout the text.
4Mathematically, this estimation problem resembles Enzyme Cost Minimisation [31]. Both methods are based on kinetic models
with known parameters and predefined fluxes, and both of them optimise metabolite and enzyme levels, but in different ways. In
enzyme cost minimisation, while in the present estimation problem, metabolite and enzyme levels are fitted to measured data.
5
we construct the poltyope, consider prior, likelihood and posterior functions on this polytope, and use them to
estimate metabolite concentrations and corresponding enzyme levels.
Assuming prior distributions for x and e, we define the preprior loss function5
P ′(x, e)
=
(x − ¯xprior)⊤C−1
xprior(x − ¯xprior) + (e − ¯eprior)⊤C−1
eprior(e − ¯eprior),
(2)
the negative logarithmic prior density, where constant terms and the prefactor6
1
2 are ignored. Similarly, using
data for x and e, we define the prelikelihood loss function
L′(x, e)
=
(Px x − ¯xdata)⊤C−1
xdata(Px x − ¯xdata) + (Pe e − ¯edata)⊤C−1
edata(Pe e − ¯edata),
(3)
the negative log-likelihood (again without constant terms and the prefactor).
The vectors ¯xdata and ¯edata
contain mean values and the matrices Cxdata = Dg(σx,data)2 and Cedata = Dg(σe,data)2 contain covariances
for measurement data. The projection matrices Px and Pe map the concentrations of all metabolite and enzyme
levels to those concentrations that appear in the measured data. The function L′ is convex in x and e, and P ′ is
strictly convex. If we add the two functions, we obtain the preposterior loss function R′(x, e) = P ′(x, e)+L′(x, e).
By adding Eqs (2) and (4) and simplifying the quadratic functions (as in [10] and [11]), we obtain the formula
R′(x, e)
=
(x − ¯xpre)⊤C−1
x,pre(x − ¯xpre) + (e − ¯epre)⊤C−1
e,pre(e − ¯epre)
(4)
with covariance matrices and mean vectors
Cx,pre
=
[C−1
x,prior + P⊤
x C−1
x,dataPx]−1
¯xpre
=
Cx,pre [C−1
x,prior ¯xprior + P⊤
x C−1
x,data ¯xdata].
(5)
Analogous formulae hold for ¯epre and C−1
e,pre.
Why is R′ called “preposterior” and not simply “posterior”? The preposterior contains enzyme levels as function
arguments, but the enzyme levels are dependent on metabolite levels and fluxes. By inserting the enzyme demand
function Eq. (1) into Eq. (4), we reobtain the three loss scores, but as functions of x alone:
Prior loss P(x)
=
(x − ¯xprior)⊤C−1
xprior(x − ¯xprior) + (e(x) − ¯eprior)⊤C−1
eprior(e(x) − ¯eprior)
Likelihood loss L(x)
=
(Px x − ¯xdata)⊤C−1
xdata(Px x − ¯xdata) + (Pe e(x) − ¯edata)⊤C−1
edata(Pe e(x) − ¯edata)
Posterior loss R(x)
=
(x − ¯xpre)⊤C−1
xpre(x − ¯xpre) + (e(x) − ¯epre)⊤C−1
epre(e(x) − ¯epre).
(6)
The enzyme demand e(x) is a convex function on the metabolite polytope [31] for a wide range of plausible rate
laws. Therefore, likelihood loss and persterior loss are convex functions, and the posterior mode can be found by
convex optimisation7.
Our estimation method can be extended to problems with several metabolic states, where each condition s has its
own flux distribution, metabolite data, and enzyme data. In fact, in this case we can run the estimation separately
for each state (see also appendix A.1). In an estimation problem with a single metabolic state, non-zero fluxes
can be assumed (because reactions with vanishing flux can be simply omitted). For problems with several states,
vanishing fluxes can be considered (see appendix C.2).
5If desired, prior and likelihood terms for thermodynamic forces may be included.
6In the matlab implementation, in contrast, this prefactor is used.
7Since P ′ and L′ are convex in the vectors x and e, and since e is convex in x, the loss terms P and L are convex in x. If P
is strictly convex in x, the posterior loss P(x) + L(x) + const. is also strictly convex. Since the feasible polytope is convex as well,
computing the posterior mode is a convex optimality problem.
6
2.4
Simultaneous estimation of kinetic constants and metabolic states
We now consider the full model balancing problem, that is, the simultaneous estimation of kinetic constants,
metabolite levels, and enzyme levels. Following [3], we parametrize the model by kinetic constants Keq, KV,
KM, and possibly KA and KI (all on log scale). Some of them may be available as data (for instance, equilibrium
constants Keq can be estimated from thermodynamic calculations) and the true values of all these quantities
need to be estimated. This problem resembles our simplified problem, where the enzyme levels were convex in x.
Now the enzyme levels also depend on kinetic constants, but they are convex in the logarithmic kinetic constants
as well! A description of the algorithm, including the convexity proof, is given in appendix B.1. Here I summarise
some main points.
Since state variables and kinetic constants are estimated together, and since the kinetic
constants are kept constant across metabolic states, the state variable become coupled across metabolic states
and need to be estimated in one go. Instead of a metabolite vector x, we consider a larger vector y, containing
the log-metabolite levels for all metabolic states and the vector of logarithmic kinetic constants. Allowed ranges
and thermodynamic constraints define a feasible polytope for the vector y. The prior, likelihood, and posterior
loss functions contain terms that depend on enzyme levels e(x). If we insert Eq. (1) into these formulae, these
terms are convex in the logarithmic kinetic parameters, and independent of the metabolite levels8. Since el(q, x)
is a convex function of the vector y =
�x
q
�
, all terms of the likelihood loss function are convex in y. The prior
loss function is strictly convex in y if pseudo values for kinetic constants are considered [11] (pseudo values are a
way to define priors by which all model parameters, even dependent ones, have non-flat priors). Details are given
in appendix B.1. Altogether, our estimation problem has the same good properties as the previous, simplified
problem. In practice, the model balancing algorithm can be improved by a number of simplifications and tricks
(appendix C). For example, enzyme levels (and therefore the likelihood function) increase very steeply close to
some polytope boundaries; to avoid numerical problems, regions close to the boundary may be excluded by extra
constraints, and the log(log posterior) may be minimised instead of the log posterior.
3
Example applications
Our test case for model balancing is a model of E. coli central metabolism (Figure 16 in appendix16), including
metabolite, enzyme, and kinetic data, taken from [31]. The model contains no allosteric regulation, but such
regulations could be added and KI and KA values could be estimated. We consider different estimation scenarios,
with artificial data, experimental data from one metabolic state (data from [31]), or experimental data from three
metabolic states (data from [16]). The same algorithm settings (such as priors or bounds) were used in all tests
(with artificial or experimental data). For details on model structure, kinetic and metabolic data, and priors see
appendix D.
3.1
E. coli metabolic model: tests with artificial data
I first generated artificial parameter sets containing kinetic constants and metabolic data (metabolite levels,
enzyme levels, and fluxes). Artificial data were generated by using the same random distributions (means and
widths) that were also used as priors in model balancing. Metabolic state variables were generated from the kinetic
model (parameterised by artificial kinetic constants) by computing steady states with randomly chosen enzyme
levels and external metabolite levels. For details on artificial data, see appendix E. Based on (noise-free or noisy)
artificial data for six simulated metabolic states, model balancing was used to reconstruct the true (noisy-free)
values. In different scenarios (see Figure 17 in appendix E), data were either fitted (metabolite and enzyme levels,
8The preposterior for kinetic constants is given by the posterior obtained from parameter balancing. For more details, see appendix
C.2.
7
and “known” kinetic constants) or predicted based on the other data (“unknown” kinetic constants).
The results of model balancing with artificial data are shown in Figures 3, 4, 5, and 6, where kinetic or state data
were either noise-free or noisy. Figure 3 shows the results for noise-free kinetic and state data. Subfigures show
different simulation and estimation scenarios (rows) and different types of variables (columns). Each subfigure
shows a scatter plot between true and fitted variables (metabolite levels, enzyme levels, and different types of
kinetic constants). Deviations from the diagonal (in y-direction) indicate estimation errors in the kinetic constants.
In the top subfigure row, data for all kinetic constants were given; in the centre row, only data for equilibrium
constants were used, and in the bottom row, no kinetic data were used9. Depending on the scenario, kinetic
constants were then either fitted (red dots) or predicted from data (magenta dots). The quality of the fit or
prediction is assessed by geometric standard deviations10 and linear (Pearson) correlations (for logarithmic values,
except for the case of enzyme levels). For comparison, I also estimated kcat values by maximal apparent kcat
values [16], based on the same artificial data (Figure 7).
The first scenario (top row) shows ideal conditions: we assume noise-free, complete kinetic data and state data.
Not surprisingly, the reconstruction errors are very small, arising from small conflicts between data and priors. The
other rows show the estimation results based on equilibrium constants only (centre row), or using no kinetic data
at all (bottom row). With noise-free data, the reconstructions in these two rows have a similar quality. To assess
the effect of noisy data, I generated artificial metabolic data (metabolite levels, enzyme levels, and fluxes) with
a relative noise level of 20 percent.
With noisy kinetic and/or metabolic data, the estimation results become
worse (Figures 4, 5, and 6), and especially the reconstruction of KM values becomes very poor. Using data on
equilibrium constants improves the results and kcat values can still be partially reconstructed (Figure 6).
Even in
the case without any kinetic data (nor equilibrium constants), model balancing yields better kcat estimates than
the “maximal apparent catalytic rate” method.
The tests with artificial data show that model balancing can adjust noisy data sets, yielding complete, consistent
model parameters and states, and that it can extract information about kcat values from metabolic data.
The
results are better than with the “maximal apparent kcat” method, and known equilibrium constants improve the
results. This is good news, because equilibrium constants are not enzyme-dependent and can be estimated from
molecule structures [35, ?].
KM values are harder to reconstruct: the estimates are in realistic ranges (probably
due to the priors), but they appear to be randomly distributed unless noise-free metabolite and enzyme data are
used.
3.2
E. coli metabolic model with experimental data
As a next test, I balanced the E. coli model with experimental data. As kinetic data, I used in-vitro kinetic
constants collected in [31] for the same model (“original kinetic data”), as well as a completed, balanced version
of this data set (“balanced kinetic data”). For details on model and data, see appendix D.
Figures 8 and 9 show estimation results for a single metabolic state, aerobic growth on glucose; see appendix D.
Since the “true” metabolic data in-vivo kinetic constants, are not known, the reconstructed kinetic constants,
metabolite levels and enzyme levels are compared to the data used for the reconstruction. In a first test, I used a
set of kinetic data obtained by parameter balancing (Figure 8). If all kinetic data are used (top row), a good fit
to these data is achieved. On the contrary, even with noisy kinetic constants slight adjustments suffice to obtain
a consistent kinetic model agreeing with all data available. However, the kinetic constants were fitted and not
predicted (as indicated by red dots). In the centre row, where equilibrium constants were used as the only kinetic
data, there is no such bias. This time, the kinetic constants are actually predicted (as indicated by magenta dots)
9The bottom subfigure rows (estimation without kinetic data) are repeated between Figures 3 and 4, and accordingly between
other figures.
10The geometric standard deviation is defined as exp(σ), where σ is the root mean square of the residuals on (natural) log scale.
8
and show correlations to in-vitro values. However, there may still be some bias because the kinetic constants (to
which I compare the predictions) had been balanced using the same network model and the same priors as used
in model balancing. To avoid this bias, I next ran model balancing with the original in-vitro kinetic data (which
contain much fewer data points for comparison). As shown in Figure 9, the predicted kcat estimates still capture a
trend in the in-vitro data (Pearson correlation 0.64 with usage of Keq data and 0.29 without Keq data).
Again,
in comparison to the method of maximal apparent catalytic constants (see Figure 10) model balancing performs
better.
A single metabolic state contains too little information to estimate the kinetic constants11. Therefore, I repeated
the estimation, no using metabolic data from three different states (growth on glucose, glycerol, and acetate)
and assuming that the kinetic constants do not change between these states (see appendix D). Figures 11, 12,
and 13 show the results. Just like before, a consistent model was obtained by moderate changes in the data. An
estimation using equilibrium constants predicted kcat values more reliably than the “maximal apparent kcat value”
method. Unexpectedly, using three states instead of one did not considerably improve the estimation results.
3.3
Parameter identifiability and choice of priors
To see how much information can be extracted from our data, we need to think about parameter identifiability
and about the choice of priors.
In parameter estimation, parameters or parameter ratios may be non-identifiable, that is, their values cannot be
inferred from the given model and data. In our Bayesian method, Gaussian priors guarantee a uniquely determined
posterior mode, but if parameters are non-identifiable, their values will only reflect the priors (which means that
high values will be underestimated and low values will be overestimated). This problem must arise if there are
fewer data values than variables to be estimated. For example, metabolic data from a single metabolic state
may not suffice to reconstruct the kinetic constants; if more metabolic states are used, the kinetic constant may
become well-defined.
In practice, we are faced with several questions: is the algorithm able to find the posterior
mode? Can we improve the result by using more data (e.g., metabolite levels from more metabolic states)?
If no kinetic data are given, how many metabolic states are needed to identify all kinetic parameters? Which
parameters are hard to reconstruct? And are there kinetic constants that remain non-identifiable, no matter how
much metabolic data we use?
If an enzyme is always saturated with a metabolite, that is, if the metabolite level is always much larger than the
KM value, the KM value is hard to estimate because it has practically no effect on measurable variables. In the
reconstructed parameter set, such KM are likely to carry large errors (i.e. posterior variances). A similar problem
occurs if the KM value of a unimolecular reaction is always much smaller than the metabolite level; in this case, the
enzyme works in its linear range, and only the ratio kcat/KM is identifiable, while the kcat and KM, individually,
are not.
If an enzyme in question is always saturated or always in the linear range, this is less of a problem,
because then parameters that are non-identifiable are also irrelevant for model predictions. However, predictions
for other experiments, in which the enzyme does behave differently, may be poor. Of course, the identifiability
problem is not specific to model balancing; other estimation methods would face the same problem.
In model balancing, like in other estimation methods, priors and measurement error bars must be carefully chosen.
In the tests with artificial data, realistic statistical distributions (for kinetic constants, metabolic variables, and
their measurement errors) were used to generate data, and the same distributions were used as priors when
reconstructing the true values. This is an ideal situation. In real-life applications, if our priors and assumed noise
11We can see this by considering possible parameter variations: if a single state is considered, a change in a KM value can be
compensated by a simultaneous change in kcat values (yielding the same flux at given metabolite and enzyme levels). Therefore,
KM values and kcat values are, individually, non-identifiable. Nevertheless, using priors we may still obtain reasonable estimates of
all parameters.
9
levels are wrong, the reconstruction would be worse than suggested by our tests with artificial data.
To obtain
the realistic distributions of kinetic constant mentioned before, I started from known (or suspected) distributions
(from [31, 14], which relied on [8]), and adjusted them based on data.
By visual inspection during parameter
balancing, I noticed that some priors had to be changed, probably because kinetic constants in central metabolism
are differently distributed than kinetic constants in metabolism in general.
4
Discussion
Various methods and modelling tools have been developed to parameterise kinetic models. They use different
types of knowledge (in-vitro kinetic constants, omics data, and physical parameter constraints) and different ways
estimation approaches (including machine learning, regression models, calculations based on rate laws, and model
fitting). A comparison to model balancing highlights some advantages and limitations of these methods.
1. Parameter estimation or optimisation by random sampling. In theory, parameter fitting and optimisa-
tion can be performed by random screening or by Monte-Carlo methods for optimisation, such as genetic
algorithms or simulated annealing. For example, one may generate a large ensemble of possible parameter
sets, compute for each of them the likelihood or posterior density values, and choose the one that performs
best (see [25] for an example).
Such optimisation methods are generic and easy to implement, but
with large parameter spaces and complicated objective functions the search for optimal solutions becomes
highly inefficient. Moreover, without an analytical grasp of the optimality problem, it is hard to assess how
good the solutions actually are. Proving an objective function to be convex, as done here, makes numerical
problems more transparent. Another question concerns the usage of priors. In sampling kinetic constants,
one may employ realistic priors obtained from parameter balancing, which also account for constraints.
However, putting priors on state variables as well would be difficult.
In model balancing, priors for all
variables are directly integrated into the optimality problem. Thus, compared to simple sampling methods,
convex model balancing has two advantages: first, the estimation problem is formulated in a transparent
way, and second, instead of numerical sampling, possibly with local optima, we directly obtain an optimality
problem for the maximum posterior (and posterior sampling can be done, too).
2. Structural kinetic modelling and elasticity sampling An alternative method for model parameterisation
is Structural Kinetic Modelling (SKM) [26], in which parameters are not fitted but randomly chosen to create
model ensembles. A consistent model state is constructed in two steps: first, a metabolic state is defined by
choosing fluxes and metabolite levels. Then, kinetic constants are chosen at random, but in agreement with
the predefined metabolic state. In practice this is achieved by randomly sampling the saturation values of
enzymes and then reconstructing the corresponding kinetic constants. Elasticity sampling [32], a variant of
this method, considers reversible rate laws and guarantees thermodynamically consistent results. In the first
step, it requires thermodynamically consistent fluxes, metabolite levels, and thermodynamic forces. In the
second step, thermodynamic forces are used to convert saturation values into correct reaction elasticities.
SKM and elasticity sampling can be adapted to account for priors or data of KM values. However, including
data or priors about kcat values and enzyme levels remains difficult, and the method cannot be used to used
to match kinetic constants simultaneously to several metabolic states.
3. Fitting kinetic constants to complete omics data in single reactions If fluxes, metabolite levels, and
enzyme level are known for several steady states, the kinetic constants can be fitted theoretically, reaction
by reaction12 [36, 17]. However, this approach has a number of limitations: for each reaction considered,
complete omics data are required; and if kinetic constants are estimated separately for each reaction,
12In the SIMMER method [17], a Markov chain Monte Carlo approach is used for the optimisation.
The estimation can be
reformulated as a model balancing problem, and be solved by convex optimisation.
10
these constants may violate thermodynamic constraints (unless a safe parameterisation scheme, e.g. with
predefined equilibrium constants, is used).
4. Maximal apparent kcat method A comparison of model balancing to the “maximal apparent kcat” method
showed that model balancing estimates kcat values more reliably, and thus extracts more information from
the available data. Of course, the “maximal apparent kcat” method is not expected to work very well if
only few metabolic states are considered. But this also holds for model balancing! The problem with model
balancing is that the calculations become harder for larger numbers of states, where the “maximal apparent
kcat” method remains the method of choice.
Parameter estimation in kinetic models can easily lead to non-convex optimisation. It may be surprising that a
simple convex estimation method exists. Model balancing relies on two insights: all fluxes must be predefined13,
and logarithmic kinetic constants and metabolite concentration are the right variables for optimisation14. Model
balancing builds on two other methods that share the same features and lead to convex optimality problems:
Parameter Balancing (PB) for the estimation of kinetic constants and Enzyme Cost Minimisation (ECM) the
estimation of optimal metabolic states (see Figure 2).
1. Parameter balancing.
Parameter balancing is an estimation method to obtain consistent kinetic and
thermodynamic constants from kinetic and thermodynamic data. It resembles model balancing, but without
detailed information on rate laws and fluxes. All “multiplicative” constants (such as Michaelis-Menten con-
stants or catalytic constants) are described by logarithmic values. To account for parameter dependencies, all
other kinetic constants are computed from a subset of kinetic constants1516, the free parameters in our linear
regression model. With Gaussian priors and measurement errors (on logarithmic scale), likelihood loss and
posterior loss terms are quadratic and convex. Parameter balancing can also be applied to kinetic and thermo-
dynamic constants (“kinetic parameter balancing”), to metabolite concentrations and thermodynamic forces
in one or more metabolic states (“state balancing”), or to kinetic constants and metabolic states together
(“state/parameter balancing”). Known flux directions can be used as additional data, to define the signs of
thermodynamic forces. Thus, parameter balancing can predict thermodynamically feasible kinetic constants
and metabolite levels and its optimisation takes place on the same set as in model balancing. It provides
reasonable ranges for kinetic constants, but in contrast to model balancing it does not consider rate laws or
quantitative fluxes17, and so it cannot be used to fit kinetic constants to metabolite, enzyme, and flux data.
2. Enzyme cost minimisation. Enzyme cost minimisation [31] predicts optimal enzyme and metabolite levels
in a kinetic model with known parameter values. Unlike parameter balancing, this method uses kinetic rate
laws with given kinetic constants, and it is a biological cost, not a fit to data, that is optimised. ECM assumes
predefined metabolic fluxes and determines metabolite and enzyme levels that realise these desired fluxes at
a minimal cost, where cost functions can be a linear or convex function of the enzyme levels, plus a convex
function of the metabolite levels. The optimisation is carried out in (log-)metabolite space. With given rate
laws, the enzyme levels can be written as functions of metabolite levels and fluxes and the cost function
(scoring enzyme and metabolite levels) is convex on the feasible metabolite polytope.
13Measurement errors in metabolic fluxes will distort our estimation results, but model balancing remains applicable, i.e., the
estimation problem is still convex.
However, fluxes must be thermodynamically consistent, that is, without thermodynamically
infeasible flux cycles.
14Accordingly, kinetic constants and metabolite concentration must be described with log-normal distributions for measurement
errors and priors while enzyme levels must be described on non-logarithmic scale (assuming normal distributions for measurement
errors and priors).
15Mathematically, parameter balancing resembles the component contribution method, which component contribution method [37]
used to determine thermodynamic constants in eQuilibrator [35].
16The equilibrium constants were not parameterised by standard chemical potentials µ◦ (as proposed in [14] for parameter balanc-
ing), but by independent equilibrium constants. This is convenient because we use a smaller set of independent variables and avoid
non-identifiability (while the standard chemical potentials themselves are not in the centre of interest), and the same choice could be
applied in parameter balancing.
17As a practical workaround, balanced kinetic constants can be further adjusted to match quantitative fluxes, but this only works
if a single metabolic state is considered.
11
Kinetic data
Priors,
constraints
Kinetic data,
Metabolic data
Priors,
Priors,
fluxes
and state variables
(metabolite and enzyme levels)
flux directions
Kinetic data
metabolite data
Enzyme
cost
minimisation
Convex
Estimated kinetic constants
Estimated kinetic constants
Estimated kinetic constants
and metabolite concentrations
Optimised state variables
(metabolite and enzyme levels)
model
balancing
parameter
Kinetic
balancing
parameter
balancing
State /
(b) State / parameter balancing
(c) Enzyme cost minimisation
(a) Parameter balancing
Constraints,
Fluxes,
Metabolite
constraints
(known parameters)
Kinetic model
(d) Model balancing
Constraints,
Enzyme and
Figure 2: Model balancing and similar methods for parameter estimation and optimal metabolic states. The
methods differ in their purpose (parameter estimation versus prediction of biologically optimal states), the choice
of free variables (kinetic constants and/or metabolite and enzyme levels), and data used, but they all share
some mathematical features: kinetic constants and metabolite levels are described on logarithmic scale (such
that all dependencies become linear); thermodynamic and physiological constraints are imposed; and fluxes are
predefined. In each of these methods, the search space is a convex polytope and the objective function is convex
(either quadratic or derived from kinetics), leading to convex optimality problem.
Model balancing combines elements from both methods. As in parameter balancing, the free variables are log-
kinetic constants and log-metabolite levels (forming a feasible parameter/concentration polytope), and the prior
and likelihood terms of kinetic and metabolic variables are convex functions. And, as in enzyme cost minimisation,
we assume that the fluxes are given and use the fact that the enzyme levels are convex functions of the (logarithmic)
metabolite levels. This is combined with two additional insights: it uses the fact that enzyme levels are convex
functions in the combined space of kinetic and metabolic variables, and the fact that in this space the prior and
likelihood terms for enzymes are convex functions just like the enzyme levels themselves.
In all three methods, the feasible region is a high-dimensional polytope (for the vector of logarithmic kinetic
constants, metabolite levels, or both). Each dimension refers to one variable, a box is defined by upper and lower
bounds, and linear constraints defined by dependencies are added. The feasible polytope for Model Balancing
is obtained from the polytopes of the other methods by taking their Cartesian product and removing infeasible
regions, in which constraints between kinetic constants and metabolite levels would be violated (shown in Figure
14). Since all variables are estimated at the same time, information about one variable can improve the estimates
of other variables. In parameter balancing, a data value for one kinetic constant may improve the estimates of all
others. Similarly, in model balancing additional metabolite and enzyme data improve the estimation of all kinetic
constants.
Depending on data available, model balancing can be applied in different ways.
1. Infer a missing data types Let us assume that data for two of our data types (kinetic constants, metabolite
levels, and enzyme levels) are available, while the third type of data is missing. There are three cases: we
may estimate in-vivo kinetic constants from fluxes, metabolite levels, and enzyme levels; we may estimate
metabolite levels from fluxes, enzyme levels, and a kinetic model; or we may estimate enzyme levels from
fluxes, metabolite levels, and a kinetic model. If the given data were complete and precise, the third type of
variables could be directly computed. But since we assume that the given data are uncertain and incomplete,
our aim is to infer the missing data while completing and adjusting the others.
2. Obtain complete, consistent metabolic states If all kinetic constants are known, and if metabolite and
enzyme have been measured, we can translate these incomplete and uncertain data into consistent and plausible
metabolic states. As in all the other cases, fluxes must be given and their directions must agree with the
12
assumed equilibrium constants and metabolite bounds.
Even in the worst case, without any enzyme or
metabolite data, we can still guess plausible metabolic states based on fluxes and on the kinetic model and
relying on priors for enzyme or metabolite levels.
3. Ensure thermodynamic constraints and bounds To obtain a consistent model, we may colect data for
kinetic and state variables and translate them into parameters and state variables for our kinetic model. These
values will satisfy the rate laws, agree with physical and physiological constraints, and resemble the data and
prior values. As in all other cases, posterior sampling could be used to decrease and assess uncertainties about
the model parameters.
4. Sampling from the posterior Instead of maximising the posterior density, we may sample from the posterior
to obtain marginal distributions and covariances of kinetic constants and state variables, and parameter sets
can be sampled to obtain a model ensemble. Sampling is facilitated by the fact that the posterior loss function
is convex (and thus, the posterior itself has a single mode).
To simplify this process, the posterior may be
approximated by a multivariate Gaussian distribution, obtained from the posterior mode and the Hessian matrix
in this point.
Model balancing can use various types of knowledge (network structure, data, priors, and constraints), handles
different types of variables (as defined by the dependence scheme used), and makes relatively few assumptions.
For example, many metabolic modelling methods, such as FBA, assume stationary flux distributions. Model
balancing does not make this assumption. Like ECM it applies to non-stationary fluxes, e.g. fluxes appearing
in dynamic time courses. However, the assumed fluxes must be thermodynamically correct.
Here I focused
on maximum-posterior estimation. Of course, the posterior can also be sampled (by Monte-Carlo Markov chain
methods) or be approximated by a multivariate Gaussian, inside the feasible polytope, defined by the posterior
mode and the Hessian matrix in this point.
Model balancing extracts information from heterogeneous data.
Even if almost no data are available, it can be used to obtain plausible models or model ensembles. In the tests
with articificial data, model balancing performed well when precise data were given, and even with imprecise data
it performed better than estimation by maximal catalytic rates.
Usage of equilibrium constants improves the
results, which confirms the importance of known equilibrium constants for constructing reliable kinetic models.
Currently, the main limitation seems to be model size, which impacts memory requirements and calculation time
(results not shown). Thus, for large models, posterior sampling based on the posterior defined here – may be be
the method of choice.
Acknowledgements
I thank Elad Noor for a interesting and enjoyable discussions. Most ideas for this work were developed in the
European Commission 7th Framework project BaSysBio (LSHG-CT-2006-037469).
References
[1] B. Teusink, J. Passarge, C.A. Reijenga, E. Esgalhado, C.C. van der Weijden, M. Schepper, M.C. Walsh,
B.M. Bakker, K. van Dam, H.V. Westerhoff, and J.L. Snoep. Can yeast glycolysis be understood in terms
of in vitro kinetics of the constituent enzymes? Testing biochemistry. European Journal of Biochemistry,
267:5313–5329, 2000.
[2] K. Smallbone et al. A model of yeast glycolysis based on a consistent kinetic characterisation of all its
enzymes. FEBS Letters, 587:2832–2841, 2013.
13
[3] W. Liebermeister and E. Klipp. Bringing metabolic networks to life: convenience rate law and thermodynamic
constraints. Theor. Biol. Med. Mod., 3:41, 2006.
[4] W. Liebermeister, J. Uhlendorf, and E. Klipp. Modular rate laws for enzymatic reactions: thermodynamics,
elasticities, and implementation. Bioinformatics, 26(12):1528–1534, 2010.
[5] F. B¨uchel et al. Path2Models: large-scale generation of computational models from biochemical pathway
maps. BMC Systems Biology, 7:116, 2013.
[6] B. Du, D.C. Zielinski, E.S. Kavvas, A. Dr¨ager, J. Tan, Z. Zhang, K.E. Ruggiero, G.A. Arzumanyan, and B.Ø.
Palsson. Evaluation of rate law approximations in bottom-up kinetic models of metabolism. BMC Systems
Biology, 10(40), 2016.
[7] I. Schomburg, A. Chang, C. Ebeling, M. Gremse, C. Heldt, G. Huhn, and D. Schomburg. BRENDA, the
enzyme database: updates and major new developments. Nucleic Acids Research, 32:D431–433, 2004.
[8] A. Bar-Even, E. Noor, Y. Savir, W. Liebermeister, D. Davidi, D.S. Tawfik, and R. Milo.
The moder-
ately efficient enzyme: evolutionary and physicochemical trends shaping enzyme parameters. Biochemistry,
21:4402–4410, 2011.
[9] D. Heckmann, C.J. Lloyd, N. Mih, Y. Ha, D.C. Zielinski, Z.B. Haiman, A. Amer Desouki, M.J. Lercher, and
B.Ø. Palsson. Machine learning applied to enzyme turnover numbers reveals protein structural correlates
and improves metabolic models. Nature Communications, 9:5252, 2018.
[10] W. Liebermeister and E. Klipp. Bringing metabolic networks to life: integration of kinetic, metabolic, and
proteomic data. Theor. Biol. Med. Mod., 3:42, 2006.
[11] T. Lubitz, M. Schulz, E. Klipp, and W. Liebermeister.
Parameter balancing for kinetic models of cell
metabolism. J. Phys. Chem. B, 114(49):16298–16303, 2010.
[12] P. Saa and L.K. Nielsen. A general framework for thermodynamically consistent parameterization and efficient
sampling of enzymatic reactions. PLoS Computational Biology, 11(4):e1004195, 2015.
[13] J.C. Mason and M.W. Covert. An energetic reformulation of kinetic rate laws enables scalable parameter
estimation for biochemical networks. Journal of Theoretical Biology, 461:145–156, 2019.
[14] T. Lubitz and W. Liebermeister.
Parameter balancing: consistent parameter sets for kinetic metabolic
models. Bioinformatics, 35:3857–3858, 2019.
[15] N.J. Stanford, T. Lubitz, K. Smallbone, E. Klipp, P. Mendes, and W. Liebermeister. Systematic construction
of kinetic models from genome-scale metabolic networks. PLoS ONE, 8(11):e79195, 2013.
[16] D. Davidi, E. Noor, W. Liebermeister, A. Bar-Even, A. Flamholz, K. Tummler, U. Barenholz, M. Goldenfeld,
T. Shlomi, and R. Milo. Global characterization of in vivo enzyme catalytic rates and their correspondence
to in vitro kcat measurements. PNAS, 113(12):3401–3406, 2016.
[17] S.R. Hackett, V.R.T. Zanotelli, W. Xu, J. Goya, J.O. Park, D.H. Perlman, P.A. Gibney, D. Botstein, J.D.
Storey, and J.D. Rabinowitz. Systems-level analysis of mechanisms regulating yeast metabolic flux. Science,
354(6311):aaf2786, 2016.
[18] M. Ashyraliyev, Y. Fomekong-Nanfack, J.A. Kaandorp, and J.G. Blom. Systems biology: parameter estima-
tion for biochemical models. FEBS Journal, 276(4):886–902, 2009.
[19] A.F. Villaverde, F. Froehlich, D. Weindl, J. Hasenauer, and J.R. Banga. Benchmarking optimization methods
for parameter estimation in large kinetic models. Bioinformatics, 35:830–838, 2018.
14
[20] S. Srinivasan, W.R. Cluett, and R. Mahadevan. A scalable method for parameter identification in kinetic
models of metabolism using steady state data. Bioinformatics, page btz445, 2019.
[21] N. Jamshidi and B.Ø. Palsson. Formulating genome-scale kinetic models in the post-genome era. Molecular
Systems Biology, 4:171, 2008.
[22] A. Khodayari and C.D. Maranas. A genome-scale Escherichia coli kinetic metabolic model k-ecoli457 satis-
fying flux data for multiple mutant strains. Nature Communications, 7(13806), 2016.
[23] R.W. Smith, R.P. van Rosmalen, V.A.P. Martins dos Santos, and C. Fleck. DMPy: a Python package for
automated mathematical model construction of large-scale metabolic systems. BMC Systems Biology, 12:72,
2018.
[24] C.J. Foster, S. Gopalakrishnan, M.R. Antoniewicz, and C.D. Maranas. From Escherichia coli mutant 13C
labeling data to a core kinetic model: A kinetic model parameterization pipeline.
PLoS Computational
Biology, 15(9):e1007319, 2019.
[25] D. Christodoulou, H. Link, T. Fuhrer, K. Kochanowski, L. Gerosa, and U. Sauer. Reserve flux capacity in
the pentose phosphate pathway enables Escherichia colis rapid response to oxidative stress. Cell Systems,
6:569578, 2018.
[26] R. Steuer, T. Gross, J. Selbig, and B. Blasius. Structural kinetic modeling of metabolic networks. PNAS,
103(32):11868–11873, 2006.
[27] W. Liebermeister. Predicting physiological concentrations of metabolites from their molecular structure. J.
Comp. Biol., 12(10):1307–1315, 2005.
[28] A. Bar-Even, E. Noor, A. Flamholz, J.M. Buescher, and R. Milo. Hydrophobicity and charge shape cellular
metabolite concentrations. PLoS Computational Biology, 7(10):e1002166, 2011.
[29] N. Tepper, E. Noor, D. Amador-Noguez, H.S. Haraldsd´ottir, R. Milo, J. Rabinowitz, W. Liebermeister, and
T. Shlomi. Steady-state metabolite concentrations reflect a balance between maximizing enzyme efficiency
and minimizing total metabolite load. PLoS ONE, 8(9):e75370, 2013.
[30] A. Flamholz, E. Noor, A. Bar-Even, W. Liebermeister, and R. Milo. Glycolytic strategy as a tradeoff between
energy yield and protein cost. PNAS, 110(24):10039–10044, 2013.
[31] E. Noor, A. Flamholz, A. Bar-Even, D. Davidi, R. Milo, and W. Liebermeister.
The protein cost of
metabolic fluxes: prediction from enzymatic rate laws and cost minimization. PLoS Computational Biology,
12(10):e1005167, 2016.
[32] W. Liebermeister. Elasticity sampling links thermodynamics to metabolic control. Preprint on arXiv.org:
arXiv:1309.0267, 2013.
[33] E. Noor, A. Flamholz, W. Liebermeister, A. Bar-Even, and R. Milo. A note on the kinetics of enzyme action:
a decomposition that highlights thermodynamic effects. FEBS Letters, 587(17):2772–2777, 2013.
[34] A. Gelman, J. B. Carlin, H. S. Stern, and D.B. Rubin. Bayesian Data Analysis. Chapman & Hall, New York,
1997.
[35] A. Flamholz, E. Noor, A. Bar-Even, and R. Milo. eQuilibrator – the biochemical thermodynamics calculator.
Nucleic Acids Research, 40(D1):D770–D775, 2012.
[36] J. Bruck, W. Liebermeister, and E. Klipp. Exploring the effect of variable enzyme concentrations in a kinetic
model of yeast glycolysis. Genome Informatics Series, 20, 2008.
15
[37] E. Noor, H.S. Haraldsdottir, R. Milo, and R.M.T. Fleming. Consistent estimation of Gibbs energy using
component contributions. PLoS Computational Biology, 9:e1003098, 2013.
[38] T. Lubitz, J. Hahn, F.T. Bergmann, E. Noor, E. Klipp, and W. Liebermeister. SBtab: A flexible table format
for data exchange in systems biology. Bioinformatics, 32(16):25592561, 2016.
[39] M. Hucka, A. Finney, H.M. Sauro, H. Bolouri, J.C. Doyle, H. Kitano, A.P. Arkin, B.J. Bornstein, D. Bray,
A. Cornish-Bowden, A.A. Cuellar, S. Dronov, E.D. Gilles, M. Ginkel, V. Gor, I.I. Goryanin, W.J. Hedley, T.J.
Hodgman, J.H. Hofmeyr, P.J. Hunter, N.S. Juty, J.L. Kasberger, A. Kremling, U. Kummer, N. Le Nov`ere,
L.M. Loew, D. Lucio, P. Mendes, E. Minch, E.D. Mjolsness, Y. Nakayama, M.R. Nelson, P.F. Nielsen,
T. Sakurada T J.C. Schaff, B.E. Shapiro, T.S. Shimizu, H.D. Spence, J. Stelling, K. Takahashi, M. Tomita,
J. Wagner, J. Wang, and the SBML Forum. The Systems Biology Markup Language (SBML): A medium
for representation and exchange of biochemical network models. Bioinformatics, 19(4):524–531, 2003.
[40] B.R.B.H. van Rijsewijk, A. Nanchen, S. Nallet, R.J. Kleijn, and U. Sauer. Large-scale 13c-flux analysis
reveals distinct transcriptional control of respiratory and fermentative metabolism in escherichia coli. Mol.
Syst. Biol., 7(477):477, 2011.
[41] A. Schmidt, K. Kochanowski, S. Vedelaar, E. Ahrn´e, B. Volkmer, L. Callipo, K. Knoops, M. Bauer, R. Ae-
bersold, and M. Heinemann. The quantitative and condition-dependent Escherichia coli proteome. Nature
Biotechnology, doi:10.1038/nbt.3418, 2015.
[42] L. Gerosa, B.R.B.H. van Rijsewijk, D. Christodoulou, K. Kochanowski, T.S.B. Schmidt, E. Noor, and
U. Sauer. Pseudo-transition analysis identifies the governing regulation of microbial nutrient adaptations
from steady state data. Cell Systems, 1:270–282, 2015.
16
E. coli model with artificial data (noise-free kinetic data, noise-free metabolic data)
(a) Metabolites
(b) Enzymes
(c) Keq values
(d) k+
cat values
(e) k−
cat values
(f) KM values
With kinetic data
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 1.04 CorrCoeff: 1.00
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 2.95 CorrCoeff: 0.34
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.03 CorrCoeff: 1.00
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 1.05 CorrCoeff: 1.00
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (fit)
GeomDev: 1.05 CorrCoeff: 1.00
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 1.04 CorrCoeff: 1.00
With Keq data only
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 1.03 CorrCoeff: 1.00
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 1.00 CorrCoeff: 1.00
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.00 CorrCoeff: 1.00
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 5.79 CorrCoeff: 0.68
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (fit)
GeomDev: 5.18 CorrCoeff: 0.83
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 3.66 CorrCoeff: 0.67
Without kinetic data
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 1.03 CorrCoeff: 1.00
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 1.00 CorrCoeff: 1.00
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.61 CorrCoeff: 0.99
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 8.92 CorrCoeff: 0.50
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (fit)
GeomDev: 6.00 CorrCoeff: 0.78
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 3.99 CorrCoeff: 0.60
Figure 3: Model balancing results for E. coli central metabolism model with artificial data. The model structure is shown in Figure 16. Each subfigure shows “true” values
(x-axis) versus reconstructed values (y-axis). Similarities are quntified by geometric standard deviations (“GeomDev”) and Pearson correlation coefficients (“CorrCoeff”).
(a) Metabolite levels. (b) Enzyme levels. (c)-(f) Different types of kinetic constants. Rows show different estimation scenarios (see Figure ) Upper row: simple scenario
S1 (noise-free artificial data, data for kinetic constants). Centre row: scenario S1K (noise-free artificial data, kinetic data given only for equilibrium constants). Lower row:
scenario S2 (noise-free artificial data, no data for kinetic constants). Depending on the scenario, kinetic constants are either fitted (red dots) or predicted (magenta dots).
17
E. coli model with artificial data (noisy kinetic data, noise-free metabolic data)
(a) Metabolites
(b) Enzymes
(c) Keq values
(d) k+
cat values
(e) k−
cat values
(f) KM values
With kinetic data
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 1.04 CorrCoeff: 1.00
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 4.20 CorrCoeff: 0.17
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.23 CorrCoeff: 1.00
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 1.34 CorrCoeff: 0.99
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (fit)
GeomDev: 1.37 CorrCoeff: 0.99
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 1.51 CorrCoeff: 0.97
With Keq data only
10 -4 10 -3 10 -2 10 -1 10 0 10 1 10 2
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 1.14 CorrCoeff: 1.00
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 9.05 CorrCoeff: -0.03
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.42 CorrCoeff: 0.99
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 8.63 CorrCoeff: 0.42
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (fit)
GeomDev: 11.23 CorrCoeff: 0.54
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 5.93 CorrCoeff: -0.08
Without kinetic data
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 1.03 CorrCoeff: 1.00
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 1.00 CorrCoeff: 1.00
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.61 CorrCoeff: 0.99
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 8.92 CorrCoeff: 0.50
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (fit)
GeomDev: 6.00 CorrCoeff: 0.78
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 3.99 CorrCoeff: 0.60
Figure 4: Same as Figure 3, with noisy kinetic data
18
E. coli model with artificial data (noise-free kinetic data, noisy metabolic data)
(a) Metabolites
(b) Enzymes
(c) Keq values
(d) k+
cat values
(e) k−
cat values
(f) KM values
With kinetic data
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 1.75 CorrCoeff: 0.98
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 1.89 CorrCoeff: 0.60
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 5.75 CorrCoeff: 0.82
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 4.72 CorrCoeff: 0.89
10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
kcat
-
values [1/s] (fit)
GeomDev: 5.14 CorrCoeff: 0.92
10 -4 10 -3 10 -2 10 -1 10 0 10 1 10 2
KM values [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 4.76 CorrCoeff: 0.77
With Keq data only
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 2.03 CorrCoeff: 0.96
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 5.28 CorrCoeff: 0.44
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.59 CorrCoeff: 0.99
10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
kcat
+
values [1/s] (fit)
GeomDev: 56.50 CorrCoeff: 0.50
10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
kcat
-
values [1/s] (fit)
GeomDev: 52.88 CorrCoeff: 0.40
10 -610 -510 -410 -310 -210 -1 10 0 10 1 10 2
KM values [mM] (true)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 102.34 CorrCoeff: -0.04
Without kinetic data
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 1.29 CorrCoeff: 1.00
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 1.86 CorrCoeff: 0.70
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
Keq values [unitless] (true)
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
Keq values [unitless] (fit)
GeomDev: 476.25 CorrCoeff: 0.36
10 -510 -410 -310 -210 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 227.99 CorrCoeff: 0.23
10 -5
10 -4
10 -310 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
kcat
-
values [1/s] (true)
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
kcat
-
values [1/s] (fit)
GeomDev: 270.34 CorrCoeff: 0.37
10 -510 -410 -310 -210 -1 10 0 10 1 10 2
KM values [mM] (true)
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 79.31 CorrCoeff: 0.03
Figure 5: Results for E. coli central metabolism with noisy artificial data. Top row: estimation scenario S3 (noisy artificial data, data used for kinetic constants). Centre
row: estimation scenario S3K (noisy artificial data, data for equilibrium constants only). Bottom row: estimation scenario S4 (noisy artificial data, no data for kinetic
constants).
19
E. coli model with artificial data (noisy kinetic data, noisy metabolic data)
(a) Metabolites
(b) Enzymes
(c) Keq values
(d) k+
cat values
(e) k−
cat values
(f) KM values
With kinetic data
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 1.76 CorrCoeff: 0.98
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 1.89 CorrCoeff: 0.59
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 5.81 CorrCoeff: 0.82
10 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 5.02 CorrCoeff: 0.87
10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
kcat
-
values [1/s] (fit)
GeomDev: 5.22 CorrCoeff: 0.91
10 -4 10 -3 10 -2 10 -1 10 0 10 1 10 2
KM values [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 4.77 CorrCoeff: 0.77
With Keq data only
10 -5 10 -4 10 -3 10 -2 10 -1 10 0 10 1
Metabolite levels [mM] (true)
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 2.09 CorrCoeff: 0.96
10 -8 10 -7 10 -6 10 -5 10 -4 10 -3 10 -2
Enzyme levels [mM] (true)
10 -8
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 11.16 CorrCoeff: 0.34
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (true)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.74 CorrCoeff: 0.98
10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7
kcat
+
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
kcat
+
values [1/s] (fit)
GeomDev: 100.29 CorrCoeff: 0.55
10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7
kcat
-
values [1/s] (true)
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
kcat
-
values [1/s] (fit)
GeomDev: 100.31 CorrCoeff: 0.46
10 -610 -510 -410 -310 -210 -1 10 0 10 1 10 2
KM values [mM] (true)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 132.75 CorrCoeff: -0.01
Without kinetic data
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
Metabolite levels [mM] (fit)
GeomDev: 1.29 CorrCoeff: 1.00
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (true)
10 -6
10 -5
10 -4
10 -3
10 -2
Enzyme levels [mM] (fit)
GeomDev: 1.86 CorrCoeff: 0.70
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
Keq values [unitless] (true)
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
Keq values [unitless] (fit)
GeomDev: 476.25 CorrCoeff: 0.36
10 -510 -410 -310 -210 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 227.99 CorrCoeff: 0.23
10 -5
10 -4
10 -310 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
kcat
-
values [1/s] (true)
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
kcat
-
values [1/s] (fit)
GeomDev: 270.34 CorrCoeff: 0.37
10 -510 -410 -310 -210 -1 10 0 10 1 10 2
KM values [mM] (true)
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
KM values [mM] (fit)
GeomDev: 79.31 CorrCoeff: 0.03
Figure 6: Same as Figure 5, with noisy kinetic data
20
k+
cat values
k−
cat values
No metabolic noise
With metabolic noise
No metabolic noise
With metabolic noise
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kapp,max
+
values [1/s] (fit)
GeomDev: 1156.03 CorrCoeff: 0.01
10 -210 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
+
values [1/s] (true)
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kapp,max
+
values [1/s] (fit)
GeomDev: 30.12 CorrCoeff: 0.12
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
-
values [1/s] (true)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kapp,max
-
values [1/s] (fit)
GeomDev: 3292.62 CorrCoeff: 0.41
10 -210 -1 10 0 10 1 10 2 10 3 10 4 10 5
kcat
-
values [1/s] (true)
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
kapp,max
-
values [1/s] (fit)
GeomDev: 70.62 CorrCoeff: -0.31
Figure 7: Catalytic constants in E. coli central metabolism (artificial data), estimated by maximal apparent catalytic rates [16]. Note that kcat values can only be estimated
in the direction of fluxes (e.g. k+
cat for reactions with forward fluxe).
21
E. coli model (aerobic growth on glucose), balanced kinetic data
(a) Metabolites
(b) Enzymes
(c) Keq values
(d) k+
cat values
(e) k−
cat values
(f) KM values
With kinetic data
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 1.73 CorrCoeff: 0.96
10 -5
10 -4
10 -3
10 -2
10 -1
Enzyme levels [mM] (data)
10 -5
10 -4
10 -3
10 -2
10 -1
Enzyme levels [mM] (fit)
GeomDev: 2.57 CorrCoeff: 0.61
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 6.18 CorrCoeff: 0.90
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (data)
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 2.13 CorrCoeff: 0.86
10 0
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (data)
10 0
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (fit)
GeomDev: 2.11 CorrCoeff: 0.86
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 1.94 CorrCoeff: 0.81
With Keq data only
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 3.91 CorrCoeff: 0.84
10 -5
10 -4
10 -3
10 -2
10 -1
Enzyme levels [mM] (data)
10 -5
10 -4
10 -3
10 -2
10 -1
Enzyme levels [mM] (fit)
GeomDev: 2.91 CorrCoeff: 0.59
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.51 CorrCoeff: 1.00
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 6.61 CorrCoeff: 0.49
10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7
kcat
-
values [1/s] (data)
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
kcat
-
values [1/s] (fit)
GeomDev: 28.33 CorrCoeff: 0.44
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 3.31 CorrCoeff: 0.23
Without kinetic data
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 1.02 CorrCoeff: 1.00
10 -3
10 -2
10 -1
Enzyme levels [mM] (data)
10 -3
10 -2
10 -1
Enzyme levels [mM] (fit)
GeomDev: 1.02 CorrCoeff: 1.00
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 34.58 CorrCoeff: 0.28
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 5.59 CorrCoeff: 0.22
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (data)
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (fit)
GeomDev: 6.99 CorrCoeff: -0.26
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 3.14 CorrCoeff: 0.30
Figure 8: Results for E. coli central metabolism with experimental data (aerobic growth on glucose). The kinetic data stem from previous parameter balancing based on
in-vitro data. Top: estimation using kinetic data. Centre: estimation using equilibrium constants as the only kinetic data. Centre: estimation using equilibrium constants
as the only kinetic data. Bottom: estimation without usage of kinetic data. The same metabolite, enzyme, and kinetic data were used in [31].
22
E. coli central metabolism model (aerobic growth on glucose), in-vitro kinetic data
(a) Metabolites
(b) Enzymes
(c) Keq values
(d) k+
cat values
(e) k−
cat values
(f) KM values
With kinetic data
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 1.66 CorrCoeff: 0.97
10 -3
10 -2
10 -1
Enzyme levels [mM] (data)
10 -3
10 -2
10 -1
Enzyme levels [mM] (fit)
GeomDev: 1.12 CorrCoeff: 0.99
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 5.52 CorrCoeff: 0.90
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 2.26 CorrCoeff: 0.93
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (data)
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (fit)
GeomDev: 2.41 CorrCoeff: 0.90
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 1.94 CorrCoeff: 0.93
With Keq data only
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 3.94 CorrCoeff: 0.84
10 -5
10 -4
10 -3
10 -2
10 -1
Enzyme levels [mM] (data)
10 -5
10 -4
10 -3
10 -2
10 -1
Enzyme levels [mM] (fit)
GeomDev: 2.92 CorrCoeff: 0.59
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 1.51 CorrCoeff: 1.00
10 1
10 2
10 3
10 4
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
kcat
+
values [1/s] (fit)
GeomDev: 8.97 CorrCoeff: 0.64
10 0
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (data)
10 0
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (fit)
GeomDev: 33.42 CorrCoeff: 0.13
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 7.15 CorrCoeff: 0.12
Without kinetic data
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 1.02 CorrCoeff: 1.00
10 -3
10 -2
10 -1
Enzyme levels [mM] (data)
10 -3
10 -2
10 -1
Enzyme levels [mM] (fit)
GeomDev: 1.02 CorrCoeff: 1.00
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 34.82 CorrCoeff: 0.28
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 8.57 CorrCoeff: 0.29
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (data)
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (fit)
GeomDev: 19.45 CorrCoeff: -0.68
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 6.41 CorrCoeff: 0.39
Figure 9: Results for E. coli central metabolism with experimental data (aerobic growth on glucose). Same as Figure 8, but based on original kinetic in-vitro data instead
of balanced kinetic data.
23
k+
cat values
k−
cat values
Balanced kinetic data
Original kinetic data
Balanced kinetic data
Original kinetic data
10 -1
10 0
10 1 10 2
10 3
10 4
kcat
+
values [1/s] (data)
10 -1
10 0
10 1
10 2
10 3
10 4
kapp,max
+
values [1/s] (fit)
GeomDev: 16.91 CorrCoeff: 0.08
10 1
10 2
10 3
10 4
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
kapp,max
+
values [1/s] (fit)
GeomDev: 9.56 CorrCoeff: 0.11
10 1
10 2
10 3
kcat
-
values [1/s] (data)
10 1
10 2
10 3
kapp,max
-
values [1/s] (fit)
GeomDev: 14.76 CorrCoeff: NaN
10 -1
10 0
kcat
-
values [1/s] (data)
10 -1
10 0
kapp,max
-
values [1/s] (fit)
GeomDev: NaN CorrCoeff: NaN
Figure 10: Catalytic constants in E. coli central metabolism model (aerobic growth on glucose), estimated by maximal apparent catalytic rates [16].
24
E. coli central metabolism model, three conditions (glucose, glycerol, acetate), kinetic data balanced
(a) Metabolites
(b) Enzymes
(c) Keq values
(d) k+
cat values
(e) k−
cat values
(f) KM values
With kinetic data
10 -610 -510 -410 -310 -210 -1 10 0 10 1 10 2
Metabolite levels [mM] (data)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 1.90 CorrCoeff: 0.98
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
Enzyme levels [mM] (data)
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
Enzyme levels [mM] (fit)
GeomDev: 8.26 CorrCoeff: 0.61
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3 10 4
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
Keq values [unitless] (fit)
GeomDev: 23.42 CorrCoeff: 0.62
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (data)
10 0
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 3.33 CorrCoeff: 0.73
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (data)
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (fit)
GeomDev: 2.02 CorrCoeff: 0.86
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 2.11 CorrCoeff: 0.77
With Keq data only
10 -610 -510 -410 -310 -210 -1 10 0 10 1 10 2
Metabolite levels [mM] (data)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 4.38 CorrCoeff: 0.87
10 -6 10 -5 10 -4 10 -3 10 -2 10 -1 10 0
Enzyme levels [mM] (data)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
Enzyme levels [mM] (fit)
GeomDev: 5.89 CorrCoeff: 0.52
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 3.22 CorrCoeff: 0.96
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 8.80 CorrCoeff: 0.33
10 1
10 2
10 3
10 4
10 5
10 6
kcat
-
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
10 6
kcat
-
values [1/s] (fit)
GeomDev: 15.04 CorrCoeff: 0.40
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 3.36 CorrCoeff: 0.27
Without kinetic data
10 -610 -510 -410 -310 -210 -1 10 0 10 1 10 2
Metabolite levels [mM] (data)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 1.05 CorrCoeff: 1.00
10 -710 -610 -510 -410 -310 -210 -1 10 0
Enzyme levels [mM] (data)
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
Enzyme levels [mM] (fit)
GeomDev: 11.40 CorrCoeff: 0.40
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
Keq values [unitless] (fit)
GeomDev: 239.47 CorrCoeff: 0.16
10 1 10 2 10 3 10 4 10 5 10 6 10 7
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
10 6
10 7
kcat
+
values [1/s] (fit)
GeomDev: 27.10 CorrCoeff: 0.04
10 -310 -210 -1 10 0 10 1 10 2 10 3 10 4
kcat
-
values [1/s] (data)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (fit)
GeomDev: 58.62 CorrCoeff: -0.43
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 3.56 CorrCoeff: 0.16
Figure 11: Results for E. coli central metabolism with experimental data (aerobic growth on glucose, glycerol, or acetate). Balanced kinetic data used. Top: estimation
with kinetic data used. Centre: estimation using equilibrium constants as the only kinetic data. Centre: estimation using equilibrium constants as the only kinetic data.
Bottom: estimation without usage of kinetic data.
25
E. coli central metabolism model, three conditions (glucose, glycerol, acetate), in-vitro kinetic data
(a) Metabolites
(b) Enzymes
(c) Keq values
(d) k+
cat values
(e) k−
cat values
(f) KM values
With kinetic data
10 -610 -510 -410 -310 -210 -1 10 0 10 1 10 2
Metabolite levels [mM] (data)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 1.78 CorrCoeff: 0.98
10 -710 -610 -510 -410 -310 -210 -1 10 0
Enzyme levels [mM] (data)
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
Enzyme levels [mM] (fit)
GeomDev: 10.88 CorrCoeff: 0.54
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3 10 4
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
Keq values [unitless] (fit)
GeomDev: 24.51 CorrCoeff: 0.61
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 3.38 CorrCoeff: 0.84
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (data)
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (fit)
GeomDev: 2.33 CorrCoeff: 0.94
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 2.16 CorrCoeff: 0.90
With Keq data only
10 -610 -510 -410 -310 -210 -1 10 0 10 1 10 2
Metabolite levels [mM] (data)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 4.40 CorrCoeff: 0.87
10 -6 10 -5 10 -4 10 -3 10 -2 10 -1 10 0
Enzyme levels [mM] (data)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
Enzyme levels [mM] (fit)
GeomDev: 5.90 CorrCoeff: 0.52
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
Keq values [unitless] (fit)
GeomDev: 3.23 CorrCoeff: 0.96
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
kcat
+
values [1/s] (fit)
GeomDev: 11.81 CorrCoeff: 0.32
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
kcat
-
values [1/s] (fit)
GeomDev: 26.95 CorrCoeff: 0.04
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 7.20 CorrCoeff: 0.20
Without kinetic data
10 -610 -510 -410 -310 -210 -1 10 0 10 1 10 2
Metabolite levels [mM] (data)
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
Metabolite levels [mM] (fit)
GeomDev: 1.05 CorrCoeff: 1.00
10 -710 -610 -510 -410 -310 -210 -1 10 0
Enzyme levels [mM] (data)
10 -7
10 -6
10 -5
10 -4
10 -3
10 -2
10 -1
10 0
Enzyme levels [mM] (fit)
GeomDev: 11.40 CorrCoeff: 0.40
10 -410 -310 -210 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6
Keq values [unitless] (data)
10 -4
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
Keq values [unitless] (fit)
GeomDev: 242.93 CorrCoeff: 0.15
10 1
10 2
10 3
10 4
10 5
10 6
kcat
+
values [1/s] (data)
10 1
10 2
10 3
10 4
10 5
10 6
kcat
+
values [1/s] (fit)
GeomDev: 40.46 CorrCoeff: -0.11
10 -310 -210 -1 10 0 10 1 10 2 10 3 10 4
kcat
-
values [1/s] (data)
10 -3
10 -2
10 -1
10 0
10 1
10 2
10 3
10 4
kcat
-
values [1/s] (fit)
GeomDev: 698.44 CorrCoeff: -0.77
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (data)
10 -3
10 -2
10 -1
10 0
10 1
KM values [mM] (fit)
GeomDev: 6.33 CorrCoeff: 0.31
Figure 12: Results for E. coli central metabolism with experimental data (aerobic growth on glucose, glycerol, or acetate). Original kinetic data used. Top: estimation
with kinetic data used. Centre: estimation using equilibrium constants as the only kinetic data. Centre: estimation using equilibrium constants as the only kinetic data.
Bottom: estimation without usage of kinetic data.
26
k+
cat values
k−
cat values
Balanced kinetic data
Original kinetic data
Balanced kinetic data
Original kinetic data
10 0
10 1
10 2
10 3
10 4
kcat
+
values [1/s] (data)
10 0
10 1
10 2
10 3
10 4
kapp,max
+
values [1/s] (fit)
GeomDev: 74.57 CorrCoeff: 0.22
10 0
10 1
10 2
10 3
10 4
kcat
+
values [1/s] (data)
10 0
10 1
10 2
10 3
10 4
kapp,max
+
values [1/s] (fit)
GeomDev: 231.48 CorrCoeff: 0.26
10 -1
10 0
kcat
-
values [1/s] (data)
10 -1
10 0
kapp,max
-
values [1/s] (fit)
GeomDev: NaN CorrCoeff: NaN
10 -1
10 0
kcat
-
values [1/s] (data)
10 -1
10 0
kapp,max
-
values [1/s] (fit)
GeomDev: NaN CorrCoeff: NaN
Figure 13: Catalytic constants in E. coli central metabolism (glucose, glycerol, actetate), estimated by maximal apparent catalytic rates [16].
27
A
The model balancing problem
A.1
Model variables and constraints
To define a model balancing problem, we need to consider all model parameters and state variables (as “model
variables”) and figure out their dependencies.
We split the model variables into “independent” (or “free”)
variables and “dependent” variables based on the following thoughts. (i) To describe dependencies between kinetic
constants, we treat some of them as free variables (independent log-equilibrium constants, log-Michaelis-Menten
constants, and log-velocity constants), while all others are linearly dependent on them (dependent log-equilibrium
constants, log-catalytic constants). (ii) For each metabolic state, we consider a metabolite log-concentration
vector, an enzyme concentration vector, and a flux vector.
Vectors from different metabolic states (usually
given as columns of a matrix) are concatenated into a large vector. (iii) Since enzyme levels follow from kinetic
constants, metabolite levels, and fluxes, they are treated as dependent variables. (iv) Thermodynamic driving
forces follow from equilibrium constants and metabolite concentrations, and are therefore dependent variables.
The kinetic constants and metabolite levels remain the only free variables. (v) The predefined flux directions
determine the signs of driving forces, implying linear constraints between logarithmic equilibrium constants and
metabolite concentrations. Altogether, we obtain the following variables and dependencies (see Figure 1 (b)).
1. Independent variables Our free variables comprise (i) the independent kinetic constants on logarithmic scale
(independent equilibrium constants ln Kind
eq , Michaelis-Menten constants ln KM, allosteric activation constants
ln KA, allosteric inhibition constants ln KI, and velocity constants ln KV), collected in a vector
qind =
ln kind
eq
ln kV
ln kM
ln kA
ln kI
,
(7)
and (ii) the metabolite log-concentrations from one or more metabolic states s, contained in metabolite vectors
x(s) = ln c(s). We obtain a vector of free variables
y =
x(1)
x(2)
..
qind
.
(8)
With np independent kinetic constants, nm metabolites, and ns metabolic states, the vector has the length
np + nm ns.
2. Dependent variables We consider three types of dependent variables: dependent kinetic constants, enzyme
concentrations, and thermodynamic forces. (i) The dependent kinetic constants on logarithmic scale, (de-
pendent equilibrium constants ln Kdep
eq , forward catalytic constants ln k+
cat), and reverse catalytic constants
ln k−
cat), form in a vector
qdep =
ln kdep
eq
ln k+
cat
ln k−
cat
.
(9)
This vector can be computed from qind by a linear function qdep = Mdep qind. The dependency matrix M
follows from the stroichiometric matrix as described in [11]. Similarly, the vector q of all kinetic constants is
28
given by the linear formula
q =
�qind
qdep
�
=
�
I
Mdep
ind
�
qind = Mall
ind qind.
(10)
(ii) The thermodynamic forces are computed by the linear formula
θ(s) = ln keq − N⊤ x(s)
(11)
or briefly θ = Mθ y with a matrix Mθ obtained from the network structure.
(iii) Based on rate laws and
using Eq. (1), the enzyme concentration vectors e(s) are given by
e(s)
l
=
v(s)
l
kl(q, x(s)).
(12)
3. Feasible region The feasible region for our free variables is defined by two types of constraints. First, lower
and upper bounds on all variables aside from enzyme levels18
qmin ≤ q ≤ qmax,
xmin ≤ x(s) ≤ xmax,
θmin ≤ θ(s) ≤ θmax,
(13)
where s denotes metabolic states. Second, the driving forces must be positive along the fluxes, and the given
flux directions define the signs of all driving forces. For all reactions with non-zero fluxes v(s)
l
̸= 0, this yields
the thermodynamic constraints
v(s)
l
θ(s)
l
> 0,
(14)
which translate into linear constraints for the variable vector y. In reactions with zero flux, driving forces are
unconstrained (unless for some reasons reactions are assumed to be in chemical equilibrium). Together these
constraints can be written as
A y ≤ b,
(15)
with a matrix A and a vector b obtained from reaction stoichiometries and the flux directions.
These
constraints define a convex feasible polytope P. Each polytope point defines a feasible vector y, i.e. a feasible
set of model parameters and metabolic states (i.e. states with positive forward driving forces). Conversely,
any feasible set of kinetic constants and metabolic states (respecting all bounds) corresponds to a point in
the polytope.
4. Priors and likelihood terms The posterior is obtained from prior and likelihood terms. For metabolite levels,
we assume uncorrelated normal priors for the values x(s)
i
(i.e. log-normal priors for concentrations).
The
data values x(s)
i,data, appearing in the likelihood, are assumed to be independent and normally distributed. For
the absolute enzyme levels, we assume normal, independent priors and data values. For logarithmic kinetic
constants, we assume normally distributed data values.
For the independent kinetic constants, we use a
correlated prior, obtained from a prior term for each independent kinetic constant and from pseudo values for
dependent kinetic constants. Formally, pseudovalues are invoked to define a correlated prior, but in practice
they are treated like additional data points (see [11]).
In contrast to similar modelling methods (Parameter Balancing and ECM), model balancing determines q and
x at the same time. The resulting vector y lives in a high-dimensional polytope whose geometric structure is
18Positivity is ensure by the other formulae. With thermodynamically feasible rate laws, the enzyme levels e(s)
l
(q, x(s)) for active
reactions, Eq. (12), are positive and convex on the entire polytope P (see appendix B.1).
29
Polytope of possible solutions
x
q
x,q
State 3
State 1
State 2
Metabolite polytopes
Kinetic constants (types of independent parameters)
ln Keq
ln KV
ln KM
ln KA
x1
x2
x3
Cartesian product
Non−Cartesian product
Non−cartesian product
Figure 14: Search space used in model balancing. The free model variables (metabolite levels and kinetic constants,
all on logarithmic scale) are constrained by physiological ranges and thermodynamic constraints, dependent
on flux directions.
Together, these inequality constraints define a feasible region in the space of logarithmic
variables (bottom).This high-dimensional polytope arises from a “non-Cartesian” product between a metabolite
polytope and a kinetic constant polytope (centre), a Cartesian product from which some parts are removed due to
constraints. The metabolite polytope itself is a Cartesian product of the metabolite polytopes for single metabolic
states; the kinetic constant polytope is a (non-Cartesian) product of polytopes (boxes) for the different types of
kinetic constants (top).
schematically shown in Figure 14. Since each state vector y consists of a vector q and a number of vectors xs,
the polytope resembles a Cartesian product of the polytopes for these single vectors. However, thermodynamic
constraints between kinetic constants and metabolite levels require that some parts of this Cartesian product must
be removed.
To see how the metabolite spaces for several states are combined, let us return to our simplified model balancing
problem from section 2.3. We can solve this problem separately for each of the states, and this is in fact the
easiest thing to do. But we can also fit all metabolic states simultaneously by one big regression model, combining
all metabolite profiles x(s). Each of these profiles must lie in a metabolite polytope P(s)
x , and if the flux directions
in all metabolic states are the same, these polytopes are identical. In contrast, if fluxes change their directions,
the metabolite polytopes P(s)
x
will differ. If we merge all vectors x(s) into a vector x, the feasible polytope for this
vector will be higher-dimensional and will be given by the Cartesian product �
s P(s)
x . As before, we can consider
the prior, likelihood, and posterior (for all metabolic states) as functions on this higher-dimensional polytope,
and the problem remains strictly convex. Since the metabolic states are independent, the prior, likelihood, and
posterior functions can be split into products of priors, likelihoods, and posteriors for the single states, confirming
again that the estimation problems can be separately solved.
30
ln KM
ln KM
ln KM
(a)
(b)
(c)
ln c
ln c
ln c
Enzyme demand is convex
Enzyme demand is constant
Enzyme demand is convex
Proportional variation of ln c and ln K
Variation of ln K
Variation of ln c
M
M
Figure 15: If enzyme demand is convex in in log metabolite levels, it is also convex in the log kinetic constants.
The graphics illustrates this by showing variations of model variables (logarithmic kinetic constants and metabolite
levels) and their effects on enzyme demand (symbolised by contour lines). (a) The enzyme demand (for each
reaction and each metabolic state, at given fluxes) is convex in the (logarithmic) metabolite levels (proof in [31]).
(b) A variation of a KM value will change the enzyme demand, but since KM values always appear in term of
the form c/KM, this change can be compensated by also varying the corresponding metabolite level, and can
therefore also be mimicked by an opposite variation of this metabolite level. (c) It follows that the enzyme demand
is convex in ln KM, and is therefore a convex function in the space of ln c and ln KM.
B
Convexity proof
B.1
The reciprocal catalytic rate is a convex function of log-metabolite levels, KM
values, and kcat values
In our models, we assume reaction rates of the form vl = el kl, with catalytic rates kl depending on metabolite
concentrations ci and kinetic constants (in a vector p, containing all forward and reverse catalytic constants k±
cat,l,
Michaelis-Menten constants KM,li, and possibly activation and inhibition constants KA and KI). In particular,
we assume that enzyme kinetics kl follow modular rate laws (which ar so general that this means hardly any
restriction):
kl =
k+
cat,l
�
j(ci/KM,li)mS
li − k−
cat,l
�
j(ci/KM,li)mP
li
Dl(c, kM)
(16)
with the molecularities mli.
The denominator D depend on the rate law chosen and must be a polynomial with
positive prefactors (or “posinomial”), consisting of terms of the shape ci/KM,li and possibly ci/KI,li or KA,li/ci.
Proposition 1 (Reciprocal rate laws are convex in the logarithmic metabolic concentrations and kinetic constants)
For all rate laws of the form 16, the reciprocal catalytic rate 1/kl is a convex function of the logarithmic metabolite
concentrations ln ci, the logarithmic Michaelis-Menten constants ln KM,li, and the logarithmic catalytic constants
ln k±
cat,l.
Corollary: Since the logarithmic kinetic constants are related by linear dependencies, the reciprocal catalytic rate
1/kl is also a convex function of the metabolite log-concentrations ln ci and the logarithmic independent kinetic
constants considered in model balancing.
Proof (alternative 1) For this proof, we note that 1/k(x) is convex in x if the kinetic constants are fixed and if
x is restricted to the feasible metabolite polytope given these kinetic constants and the predefined flux direction.
This has been shown in [31]. Moreover, we note that in the rate laws considered, concentrations and kinetic
constants always appear in the form of product terms (e.g. k+
cat · c/KM). On log-scale, these terms are sums
(e.g. ln k+
cat + ln c − ln KM). Therefore, if changes of logarithmic concentrations have a certain effect (namel
a “convex” variation of 1/r), then changes of logarithmic kinetic constants should the same type of effects (see
31
Figure 15). To see this in detail, we first show that 1/k is convex in the combined space of x = ln c (relevant
metabolites) and qM = ln KM (relevant Michaelis-Menten values). Since concentrations and Michaelis-Menten
values always appear as ratios, any linear variation of a ln KM value can be mimicked by a variation in x-space:
instead of increasing a Michaelis-Menten value, we can decrease the corresponding metabolite level, with the
same effect on the catalytic rate. Therefore, any linear variation in (x, qM)-space can be mimicked by a linear
variation in (x)-space alone as far as changes in 1/k are concerned. Therefore, convexity of 1/ratelaw in x-space
implies convexity in (x, qM)-space.
Next, we consider variations of the catalytic constants kcat and use the same
trick: we know that 1/k is convex in (x, qM)-space, and describe changes of the catalytic constants as variations
in qcat-space. Again, any linear variation can be mimicked by a linear variation in (x, qM), and so 1/k must
be convex in (x, qM, qcat)-space. So far, we considered only kcat and KM values and neglected the activation
constants KA and inhibition constants KI. In our rate laws these constants appear in similar mathematical terms
as the Michaelis-Menten constants. For example, a rate law with competitive inhibition contains similar terms
with KI values and KM values in its denominator. The terms with KA values, on log scale, carry a minus sign, but
since this term (on log-scale) is linear, the minus sign does not change the convexity. Finally, since (logarithmic)
kinetic constants depend linearly on (logarithmic) independent kinetic constants, the enzyme level is also convex
in the (logarithmic) independent kinetic constants and (logarithmic) metabolite levels.
There is a also a shorter proof. Without loss of generality, we assume that the flux vl is positive. To show that
1/kl is a convex function of
�q
x
�
, we rewrite (see [4], Eq. 26)
1
kl
=
Dl(c, p)
kV
l
��
i(ci/KM,li)mli 2 sinh( hl
2 θl)
(17)
with molecularities mli, the vector p of kinetic constants, and driving force θl = − �
i nil(µ◦
i /RT + ln ci). Since
ex is a convex function, expression (17) will be convex in (ln c, ln p) if its logarithm
ln Dl(c, p) − ln kV
l − 1
2 ln(
�
i
(ci/KM,li)mli) − ln 2 − ln sinh(hl
2 θl)
(18)
is convex in (ln c, ln p). Since the denominator term is a posinomial Dl(c, p) = �
a Aail cαail
i
kβail
il
, ln Dl is
convex [31]. Furthermore, − 1
2 ln(�
i(ci/KM,li)mli) is linear in (ln c, ln p) and therefore convex, and − ln 2 is
constant and therefore convex. Finally, − ln sinh( ·) is convex for any positive arguments, and its argument hlθl
2
is in fact positive (for positive fluxes) and affine in
�q
x
�
.
B.2
Model balancing is a convex problem
Based on the proof in section B.1, we can conclude that model balancing is a convex problem. For a proof, we
need to show that the likelihood loss for enzyme data, and the negative log priors for enzyme levels are convex
functions on the feasible polytope. First, we note that likelihood loss and prior loss are convex functions of the
individual enzyme levels e(s)
l
and that the concatenation of two convex functions yields a convex function. Thus,
it remains to be shown that each enzyme level e(s)
l
is a convex function on the feasible polytope. Second, each
e(s)
l
depends, effectively, only on the kinetic constants of the reaction considered and on the metabolite levels cs
affecting this reaction. Third, given the flux in this state and given the kinetic constants and metabolite levels, the
enzyme level e(s)
l
is proportional to 1/kl in this state (which, as we saw, is convex in ln cs and in the logarithmic
kinetic constants).
32
C
Implementation
A Matlab implementation of Model Balancing, together with example models and data, is available at https:
//github.com/liebermeister/cmb. The file format for models and data (kinetic constants, fluxes, metabolite
levels, protein levels) is SBtab [38] and metabolic networks can be defined in SBML [39] or SBtab files. By
default, the algorithm starts by running model balancing on an average model state (with metabolic state data
given by the geometric mean over the metabolic states).
The resulting kinetic constants are then used as initial
values for the following full calculation with several metabolic states.
C.1
Possible simplifications and variants of model balancing
Model balancing can be adapted in various ways. (i) If a type of data is not used, likelihood terms for this data type
are omitted. Even without any data, priors will keep the results in biochemically plausible ranges. (ii) If certain
parameters (e.g. the equilibrium constants) are precisely known, their values can be predefined (e.g. by treating
them as data with very small standard errors). (iii) Model balancing also applies to models with irreversible rate
laws. In an irreversible rate law, there are fewer kinetic constants (since reverse catalytic constants, equilibrium
constants, and velocity constants do not play a role); the forward kinetic constant is a free parameter, and no
Haldane relationship is considered. Describing (some or all) rate laws as irreversible changes the structure of
the kinetic dependence matrix M. (iv) Different model parameterisations: instead of independent equilibrium
constants, standard chemical potentials may be used as independent parameters [11]. (v) A preposterior for
kinetic parameters may be obtained by previous parameter balancing, and pseudo values for metabolite and
enzyme levels may be obtained by a previous ECM. (vi) To penalise unrealistically high metabolite or enzyme
levels, a regularisation term may be added, for example, proportional to the cost function considered in ECM.
(vii) Omics data may not contain absolute metabolite and enzyme levels, but relative changes between metabolic
states. To account for such data, a variant of the dependence scheme might be considered: for each metabolite,
we split the log-concentrations ln c(s)
i
into a reference value ln ci and a deviation ∆ ln c(s)
i . Uncorrelated priors
for these variables yield a meaningful correlated prior for the metabolite levels, and a similar splitting can be used
for enzyme levels.
C.2
Practical computation details
1. Calculation of the preposterior To compute the preposterior functions (Eq. 5 for metabolite levels, and
similar formulae for enzyme levels and kinetic constants), we need to invert a covariance matrix. This can be
numerically expensive. To compute the preposterior of the independent kinetic constants, we need to solve
Cq,ind,pre
=
[C−1
q,ind,prior + M⊤ C−1
q,dataM]−1
¯qind,pre
=
Cq,ind,pre [C−1
q,ind,prior ¯qind,prior + M⊤ C−1
q,data ¯qdata] .
(19)
The matrix inversion for C−1
q,ind,prior and C−1
q,data (covariance matrices for metabolite and enzyme levels) is easy
because the original covariance matrices are diagonal and the projector matrices P select single vector elements.
However, inverting the term in brackets may be hard. To speed up the calculation, we set A = C−1
q,ind,pre and
obtain the similar formulation
A
=
C−1
q,ind,prior + M⊤ C−1
q,dataM
¯qind,pre
=
A−1 [C−1
q,ind,prior ¯qind,prior + M⊤ C−1
q,data ¯qdata] .
(20)
Now the costly matrix inversion in the first equation is avoided, and the right-hand side in the second equation
33
can be computed without explicitly computing the matrix inverse (e.g. by using the matrix left division operator
\ in matlab). This calculation is faster and works for sparse matrices.
2. Reactions with vanishing flux If reaction flux is non-zero, the flux direction puts a constraint on the driving
force, and the predicted enzyme level is positive. If a reaction is always inactive – that is, in all metabolic states
– the kinetic constants for this reaction are ill-determined, and the reaction can be removed from the model.
But what if a reaction fluxes vanish in some of the metabolic states? The vanishing flux can either be caused
by a vanishing enzyme level, or by a vanishing thermodynamic force. If the reaction is known to be in chemical
equilibrium, we also set the driving force to 0, which leads to an extra equality constraint on metabolite levels.
In this case, the enzyme level can be positive and needs to be estimated (although the economical “principle
of dispensable enzyme” would suggest a vanishing enzyme level in this case). Otherwise, with a zero flux and
a non-zero driving force, the enzyme activity must be zero: for an enzyme without allosteric inhibition, this
means that the enzyme concentration must vanish.
3. Divergence of enzyme levels close to polytope boundaries. Each thermodynamic constraint defines a
boundary of the feasible polytope. Close to this boundary, an enzyme levels goes to infinity and the likelihood
function explodes. This steep increase can cause numerical problems during optimisation. To handle them, we
may apply the logarithm function once more to the (likelihood or posterior) score, and use the resulting function
as our minimisation objective.
This new objective function will still go to infinity at polytope boundaries,
but less steeply. The new objective function may be non-convex, but since it depends monotonically on a
convex function, it will still have a single local minimum. A second way to avoid this problem is to exclude
problematic regions close to the boundary by introducing some extra constraints. In practice. we can make all
thermodynamic constraints a bit tighter, by requiring small, non-zero thermodynamic forces in every reaction
[31].
4. Starting point for optimisation To obtain an initial point for our optimisation, we may first run model bal-
ancing for an average metabolic state. This yields a first guess of the kinetic constants. Alternatively, we can
run model balancing separately for each metabolic state. In each run, we start from the prior mode (or alterna-
tively, from the posterior mode for kinetic constants obtained by Parameter Balancing, and the posterior mode
for each metabolite value). The resulting concentration vectors and the state-averaged (arithmic/geometric)
kinetic constant vector can be used as initial values for the multi-state problem.
5. Running parameter balancing as a separate first step Model balancing can also be run in two steps.
The first step, is a simple parameter balancing problem: we consider only kinetic constants and fit them to
kinetic data. The result is a multivariate Gaussian posterior for all (logarithmic) kinetic parameters [11] that
summarizes all data and prior knowledge about the kinetic constants. In the second step, we use this posterior
as a prior for the kinetic constants, and fit kinetic constants and model states (metabolite and enzyme levels)
to metabolite and enzyme data. Since the kinetic data have already been used to define the prior, they can
be ignored in this part of the estimation. The calculation is equivalent to the method described in this paper.
By processing the kinetic data separately in advance, we can learn more clearly what information is contained
in the kinetic data alone, before combining them with metabolic data. Moreover, a known kinetic “prior” that
includes all information about kinetic data may allow us to further constrain the kinetic constants in order to
reduce the feasible search space.
D
Example model
The E. coli central carbon metabolism model, taken from [31], comprises 40 metabolites and 30 reactions and
contains 107 KM values and 167 kinetic constants in total (KM values as well as forward and reverse kcat values)
34
fructose-1,6P
xylulose-5P
sedoheptulose-7P
ribulose-5P
glucono-lactone-6P
gluconate-6P
glycerone-P
2e-
ATP
ADP
ATP
ADP
Pi
2e-
glucose
PEP
pyruvate
CoA
L-malate
succinate
citrate
isocitrate
cis-aconitate
fumarate
CoA
2e-
2e-
CoA,ATP
ADP, Pi
CoA
2e-
2e
ZWF
PGI
PFK
PTS
GLH
PGD
RPI
RPE
TXT
TAL
TXT
ALD
GAP
PGK
TIM
PGM
PGH
PYK
PDH
glycerate-2P
CSN
ACN
ACN
MDH
FUM
SDH
SCS
ICD
KGD
-
2e
CO2
-
2e-
glycerate-1,3BP
fructose-6P
glucose-6P
glyceraldehyde-3P
ribose-5P
erythrose-4P
glycerate-3P
PEP
pyruvate
oxaloacetate
acetyl-CoA
succinyl-CoA
2-ketoglutarate
ATP
ADP
CO2
CO2
CO2
CO2
Pi
PPC
PTS
PGI
PFKFBA
TPI
GAP
PGK
GPM
ENO
PYK
PDH
CSNACN1
ACN1
ICD
KGD
SCSSDH
FUM MDH
PGL
GND
RPE
RPI
TAL
TKT1TKT1
ZWF
Figure 16: Model of E. coli central carbon metabolism and protein data, both taken from [31].
. The model structure is shown in Figure 16 and described at https://github.com/liebermeister/cmb (in
the file resources/data/data-organisms/escherichia coli/network/ecoli noor 2016.tsv).
To model aerobic growth on glucose, I used a data set from [31], which gathered measured flux data from [40],
proteomics data from [41], and metabolomics data from [42]. To model several metabolic states, I used a data
set from [16], where a larger network model had been considered, proteomics data from different sources were
used, and flux data had been computed by FBA. I linearly the flux data onto the E. coli model to obtain complete
and consistent flux values.
A comparison between the two data sets reveals a discrepancy in scaling: the (FBA-
derived) fluxes from [16] were smaller than the fluxes taken from [31] by an approximate factor of 10,
while
enzyme levels were smaller by an approximate factor of 2.
E
Prior distributions and artificial data
To define priors, pseudo values, and constraints (for kinetic constants, metabolite levels and enzyme levels), I
used the default values from parameter balancing (see www.parameterbalancing.net. However, when running
parameter balancing as a test, I found that the available kcat values were typically much higher than the prior
median value, as expected for enzymes in central metabolism [8]. In line with these data, I changed the prior
for kcat values from a median of 10 s−1 (geometric standard deviation 100) to a median of 200 s−1 (geometric
standard deviation 50). Likewise, I changed the prior width for KM values from a geometric standard deviation
of 10 to a geometric standard deviation of 20 (while keeping the median 0.1 mM unchanged). A table describing
the priors is provided in the github repository, file resources/data/data-prior/cmb prior.tsv. These values, used in
the matlab implementation, can be easily modified.
Artificial kinetic constant data were generated as follows.
Given the network structure, true artificial kinetic
constants were generated by assigning random (log-normal) values to ln Kind
eq , ln KM, and ln KV and computing
35
General scheme
S1: Noise−free data, kinetic constants used as data
S2: Noise−free data, kinetic constants unknown
S3: Noisy data, kinetic constants used as data
S4: Noisy data, kinetic constants unknown
Noisy
articifical
data
add noise
add noise
Kinetic
(data)
(data)
State
"True"
variables
Estimated
(reconstructed)
variables
Kinetic
(true)
simulate
Kinetic
(fit)
State
(true)
State
(fit)
estimate
simulate
Kinetic
(fit)
State
(true)
State
(fit)
Kinetic
(true)
simulate
Kinetic
(fit)
State
(true)
State
(fit)
"True"
variables
Estimated
(reconstructed)
variables
Kinetic
(true)
simulate
Kinetic
(fit)
State
(true)
State
(fit)
"True"
variables
Noisy
articifical
data
Estimated
(reconstructed)
variables
simulate
add noise
Kinetic
(fit)
State
(true)
(data)
State
State
(fit)
"True"
variables
Estimated
(reconstructed)
variables
"True"
variables
Estimated
(reconstructed)
variables
Kinetic
(true)
Kinetic
(true)
estimate
estimate
estimate
Noisy
articifical
data
add noise
add noise
Kinetic
(data)
(data)
State
estimate
Figure 17: Estimation scenarios with artificial data. Left: general procedure. In a given model, kinetic constants
are drawn from random distributions (respecting their interdependencies), and metabolic state data are generated
by combining sampling and simulation runs (top row). From these “true” values, artificial (kinetic and state)
data are generated by adding uncorrelated noise (centre row). Model balancing is used to estimate the kinetic
parameters and metabolic state variables (bottom row), aimed to resemble the true values. Right: I employed
four variants of this procedure (called S1-S4) in which noise is either considered or not (in the latter case, the
noise level is set to zero), and kinetic data are used or not. In another variant, data for equilibrium constants are
used as the only kinetic data.
the other constants. The random values were sampled from the same distributions that are used as priors in
model balancing.
To generate artificial metabolic state data, enzyme levels and external metabolite levels were randomly sampled
from the same distributions that are used as priors in model balancing. Then the model was parameterised with
the “true” artificial kinetic constants and was solved to obtain a steady reference state (steady-state metabolite
concentrations and fluxes). Based on this reference state, a number of metabolic states were constructed by
randomly varying metabolite and enzyme levels (again, following the prior distribution) and computing the (non-
steady) reaction rates19.
The resulting states are seen as the “true values”.
To generate noisy state data,
uncorrelated random noise was added to the “true values”. When generating artificial data, noise was also added
to fluxes but the flux signs were kept unchanged, to ensure thermodynamically feasible flux directions as required
in model balancing.
19Alteratively, one could similate a dynamic time course and take snapshots at different time points.
36
| 2019 | Model balancing: consistent kinetic constants and metabolic states obtained by convex optimisation | 10.1101/2019.12.23.887166 | null | creative-commons |
1
Diving Behavior Reveals Humidity Sensing Ability of Water Deprived
Planarians
Yu Pei1, Renzhi Qian1, Yuan yan1, Yixuan Zhang1, Liyuan Tan1, Xinran Li1, Chenxu Lu1, Yuxuan
Chen1, Yuanwei Chi1, Kun Hao1, Zhen Xu1, Guang Yang1, Zilun Shao1, Yuhao Wang1 and
Kaiyuan Huang1,2,3
1College of Biological Science, China Agricultural University; Beijing, 100193, China.
2Tsinghua Institute of Multidisciplinary Biomedical Research (TIMBR), Tsinghua University;
17 Beijing, 100084, China
3National Institute of Biological Sciences (NIBS); Beijing, 102206, China.
* Kaiyuan Huang
Email: huangkaiyuan@nibs.ac.cn
Author Contributions: Yu Pei, Renzhi Qian and Yuan yan contributed equally to this work.
Kaiyuan Huang, Yu Pei, Renzhi Qian and Yuan yan designed research. Yu Pei, Renzhi Qian,
Yuan yan, Yixuan Zhang, Liyuan Tan, Xinran Li, Chenxu Lu, Yuxuan Chen, Yuanwei Chi and Kun
Hao performed research. Zhen Xu, Guang Yang, Zilun Shao and Yuhao Wang analyzed data.
Kaiyuan Huang and Yu Pei wrote the manuscript.
Competing Interest Statement: The authors declare that there is no competing interest in the
study.
Classification: Biological Sciences
Keywords: planarians, humidity sensing, aquatic animals, decision making, survival seeking.
2
Abstract
1
Humidity sensing ability is crucial to terrestrial animals for fitting the environment. Researchers
2
made great progress in recent study about humidity sensing mechanisms of terrestrial animals.
3
However, it is poorly understood whether humidity sensing exists in aquatic animals. Here, we
4
demonstrate that the aquatic planarians, one of the primitive forerunners of later animals, has the
5
ability of humidity sensing and is capable of using the ability to perceive the direction of water
6
from a drought place to seek survival. The behavior we discovered is described as diving
7
because the worms twist its body to break away from the mucus that make them adhere to the
8
drought place and drop into the water. The behavior is triggered by rapidly increasing humidity.
9
This finding suggests that humidity sensing ability exists in the lower aquatic animals, and the
10
ability might be used to seek for water when aquatic animals are facing desiccation. The finding
11
also suggests that survival-seeking and decision-making behavior have appeared in the primitive
12
planarian worms.
13
14
Main Text
15
16
Introduction
17
18
As a universal medium for biochemical events, water is an indispensable resource for all
19
animals. For terrestrial animals, they are at constant risk for desiccation due to unpredictable
20
climate change. Therefore, humidity sensing has been widely investigated in terrestrial animals
21
for their need for a comfortable environment. Recent studies revealed detailed mechanisms of
22
how terrestrial animals sense humidity(1-3).
23
In contrast to terrestrial animals, aquatic animals have much less possibility to face a situation
24
of desiccation. However, dehydration is usually fatal to aquatic animals for they need water to
25
respire. So, it might also be crucial for some aquatic animals to perceive the direction of water
26
when facing an emergent situation of water depletion. Nevertheless, it is poorly understood
27
whether this ability exists in aquatic animals.
28
Planarian is a kind of aquatic free-living flatworm to have first evolved a centralized brain. As a
29
primitive forerunner of later animals, the planarians can be evolutionarily instructive for the
30
investigation of later animals. For freely living in the natural environment, planarians have evolved
31
various sensory abilities, including sensitivity to light(4), temperature(5), water currents(6),
32
chemical gradients(7), vibration(8), magnetic fields(9) and electric fields(10), but its humidity
33
sensing ability is not yet identified.
34
Unlike most aquatic animals who live freely in the water, planarians usually live and stick under
35
rocks, debris and water plants in streams, ponds, and springs(11). Therefore, they are confronted
36
with frequently falling water levels and might be lifted out of the water. So, it might be important
37
for planarians to perceive the direction of water to seek survival under such emergent situations.
38
Thus, we speculate that planarians have the ability of humidity sensing to carry out such tasks.
39
To prove this hypothesis, we established a behavioral paradigm of planarians called ‘diving’,
40
which will be explained in detail in the result section. And then we demonstrate that the worms
41
can perceive humidity and its increasing speed to judge the direction of the water. Our finding
42
identified the humidity sensing ability of a kind of aquatic animal and explained what this ability is
43
used for, which is yet not discovered in this field. This finding also suggests that survival-seeking
44
and decision-making behavior have appeared in the primitive planarian worms and might shed
45
light on how these abilities evolved. The finding also provides a ‘diving’ behavioral paradigm for
46
future study.
47
48
Results
49
50
1. Rapidly increasing humidity induces diving behavior of planarians
51
3
We established a behavioral paradigm of planarians called ‘diving’ (Fig. 1A, SI movie 1). A
52
planarian worm is put in a petri dish and its surrounding water is wiped out. Then the petri dish is
53
inverted onto a 250 mL beaker containing 200 mL of water. The worm will first struggle to find
54
water and then uplift its head. When the worm decides to dive, it will twist its body to break away
55
from the mucus and drop into the water.
56
We totally tested 20 worms through the diving paradigm and most of the worms started the
57
diving behavior in 60 seconds, showing that they might be able to perceive the water under them.
58
Then we tested 20 worms with a dry 250 mL beaker, all of the worms did not perform the diving
59
behavior and finally stopped moving (Fig. 1B).
60
We speculate that this behavior is related to the increase of humidity. Thus, to control the
61
humidity conditions, all of the diving experiment was carried out in a 38%±1% relative humidity
62
(RH) environment if not otherwise stated. We measured the RH variation in the two processes
63
above. To simulate the situation of a worm in the experiment while measuring the RH variation,
64
we embedded the humidity meter probe in the middle of a foam plastic board and then put the
65
board on the beaker (Fig. 1E). The result reveals a rapid increasing RH in the 250 mL beaker
66
containing 200 mL water, which increases from 38% RH to more than 60% RH in 30 seconds.
67
(Fig. 2A). In contrast, the RH of the dry 250 mL beaker maintained relatively constant at 38%±1%.
68
The time that a worm starts the diving behavior is counted and synchronized with the RH
69
variation. (Fig. 2A, Fig. 2B).
70
The worms dropped at the mean time of 13 seconds and the corresponding RH is 54.1%. We
71
argued that the diving behavior might be induced by high constant humidity rather than rapidly
72
increasing humidity. Therefore, we tested 20 worms with dry 250 mL beaker at an at a constant
73
RH of 65%±5%. Although the worms can struggle to crawl for a long time due to low evaporation
74
rate, none of the worms performed diving behavior. Hence, we conclude that the diving behavior
75
of the worm is induced by rapidly increasing humidity rather than constant high humidity.
76
77
2. Slower increasing humidity hinders planarians’ decision to dive
78
To investigate whether diving behavior can be induced by slower increasing humidity, we
79
tested 20 worms with a 250 mL beaker containing 50 mL and measured its humidity variation.
80
(Fig. 1C, Fig. 2B). The humidity increase rate in this case is about one half of 250 mL beaker
81
containing 200 mL of water. Surprisingly, some of the worms didn’t drop and some of the worms
82
took about minutes to drop. We reasoned that whether to execute the diving behavior involves
83
the worm’s decision. Slower increasing humidity makes some of the worms hesitate to drop,
84
which further proves that only rapidly increasing humidity can solidly induce the diving behavior of
85
planarians.
86
87
3. Rapidly increasing humidity can mislead planarians drop into a dry place
88
To further demonstrate the diving behavior is induced by rapidly increasing humidity, we
89
simulated a situation of rapidly increasing humidity, yet no water is provided if the worm drops.
90
Instead of using a large quantity of water, we sprayed water droplets on the wall of the beaker
91
and put a piece of dry plastic to cover the bottom of the beaker (Fig. 1D). Then we tested 20
92
worms in this beaker and measured the humidity variation, all of the worms started to drop on the
93
dry plastic in 60 seconds (Fig. 2D). This result confirmed the conclusion that rapidly increasing
94
humidity induces the diving behavior of planarians.
95
96
Discussion
97
The investigation of mechanisms of humidity sensing had been focused on terrestrial animals.
98
Including how hygroreceptor works in insects like P. americana(12) and D. melanogaster(13), and
99
the integration of mechano and thermo inputs of C. elegans(2) and humans(3). However, the
100
humidity sensing of aquatic animals was hardly ever considered in previous studies. In the
101
present study, we unveiled the ability of humidity sensing of aquatic planarians by establishing
102
the diving behavioral paradigm, which they use to seek survival under dehydration conditions.
103
4
Our work reveals that in the diving behavioral paradigm, a worm facing dehydration has to
104
make a quick decision whether or not to secede from the attached surface before it cannot move
105
anymore. In this process, the worm continues to raise its head probably to sense the increasing
106
humidity, which would accelerate the rate of evaporation. So, the judgment of the worm must be
107
accurate to deal with such an emergent situation. As our result shows increasing rate of humidity
108
have become a crucial indicator for the worm’s decision.
109
As illustrated above, the diving behavior of planarians can be classified as a decision-making
110
and survival-seeking behavior. Being one of the first kinds of animals to have evolved a
111
centralized brain, planarians’ behavior can provide instructions from the evolutionary perspective
112
for investigating the behaviors of later animals. Our results demonstrate that the decision-making
113
and survival-seeking behavior had already developed in the primitive planarian worms, which
114
might provide a new evolutionary perspective for investigating such behaviors.
115
116
Materials and Methods
117
118
Planarians A laboratory strain of D. japonica, originating from wild collected D. japonica
119
(identified by cytochrome c oxidase subunit 1 gene) from the Cherry-Valley in Beijing Botanical
120
Garden, Haidian district, Beijing, China in 2019. Worms are maintained in Montjuic Water(14) in
121
the dark and fed with chicken liver twice a week. Worms are fed 2 days before experiment. The
122
length of planarians used in the experiment varies from 1.5 cm to 2.5 cm.
123
124
Experimental Setup A 250mL glass beaker and a plastic petri dish is used in the experiment.
125
Kimwipes paper towel is used to wipe water. A UT331+ humidity meter (Uni-Trend Technology
126
(China) Co., Ltd.) is used to measure and record the RH.
127
128
Test Procedure A planarian is transferred to the petri dish containing water from home well by a
129
transfer pipette. Wait until the worm sink and attach to the bottom of the petri dish. Slowly pour
130
the water out while maintaining the worm attached to the bottom. Wipe out the water in the petri
131
dish but do not touch the worm. The petri dish is washed by water and wiped between each
132
worm’s test.
133
134
Humidity Measure Procedure The primer of the humidity meter is embedded into a foam plastic
135
board then cover the beaker and immediately start measuring for 90 second.
136
137
Statistical Analysis All data were analyzed using PRISM (GraphPad Prism 9.0.0(121)).
138
Nonlinear regression (curve fit): polynomial (fourth order) is used to generate fitting results of RH
139
variation.
140
141
Acknowledgments
142
143
We wish to thank Prof. Baoqing Wang, Prof. Zhengxin Ying and Dr. Wei Wu for suggestions and
144
financial support. Figure 1 is created with BioRender.com
145
146
References
147
1.
Enjin A, et al. (2016) Humidity Sensing in Drosophila. Current biology : CB 26(10):1352-
148
1358.
149
2.
Russell J, Vidal-Gadea A, Makay A, Lanam C, & Pierce-Shimomura J (2014) Humidity
150
sensation requires both mechanosensory and thermosensory pathways in
151
Caenorhabditis elegans. Proceedings of the National Academy of Sciences of the United
152
States of America 111.
153
5
3.
Filingeri D, Fournet D, Hodder S, & Havenith G (2014) Why wet feels wet? A
154
neurophysiological model of human cutaneous wetness sensitivity. Journal of
155
Neurophysiology 112:1457.
156
4.
Shettigar N, et al. (2017) Hierarchies in light sensing and dynamic interactions between
157
ocular and extraocular sensory networks in a flatworm. Science Advances 3:e1603025.
158
5.
Inoue T, Hoshino H, Yamashita T, Shimoyama S, & Agata K (2015) Planarian shows
159
decision-making behavior in response to multiple stimuli by integrative brain function.
160
Zoological Letters 1(1):7.
161
6.
Allen GD (1915) Reversibility of the Reactions of Planaria Dorotocephala to a Current of
162
Water. Biological Bulletin 29(2):111-128.
163
7.
Mason P (1975) Chemo-klino-kinesis in planarian food location. Animal behaviour
164
23:460-469.
165
8.
Dessì-Fulgheri F & Messeri P (1973) [Use of 2 different negative reinforcements in light-
166
darkness discrimination of planarians]. Bollettino della Società italiana di biologia
167
sperimentale 49:1141-1145.
168
9.
Brown F & Chow C (1975) Differentiation between Clockwise and Counterclockwise
169
Magnetic Rotation by the Planarian, Dugesia dorotacephala. Physiological Zoology
170
48:168-176.
171
10.
Brown H & Ogden T (1968) The Electrical Response of the Planarian Ocellus. The Journal
172
of general physiology 51:237-253.
173
11.
Vila-Farré M & Rink J (2018) The Ecology of Freshwater Planarians.), Vol 1774, pp 173-
174
205.
175
12.
Tichy H & Kallina W (2010) Insect Hygroreceptor Responses to Continuous Changes in
176
Humidity and Air Pressure. Journal of neurophysiology 103:3274-3286.
177
13.
Liu L, et al. (2007) Drosophila hygrosensation requires the TRP channels water witch and
178
nanchung. Nature 450:294-298.
179
14.
Merryman S, Sánchez Alvarado A, & Jenkin J (2018) Culturing Planarians in the
180
Laboratory.), Vol 1774, pp 241-258.
181
182
183
6
184
Figures
185
186
Figure 1. Illustration of the experiment process. (A-D) The diving experiment. (E) The relative
187
humidity measurement.
188
189
<insert page break here>
190
191
192
7
193
Figure 2. Diving behavior are induced by rapidly increasing humidity. 3 sets of RH data are used
194
for nonlinear curve fitting in (A-C). The lower panel of (A-C) shows that the time a worm starts the
195
diving behavior synchronized with the RH variation, time data is presented as mean ± SEM. (A)
196
Rapidly increasing humidity induces diving behavior of planarians (n=20). (B) Slower increasing
197
humidity hinders planarians’ decision to dive (n=20, 6 worms did not dive, 2 worms used more
198
than 90 seconds). (C) Rapidly increasing humidity can mislead planarians drop into a dry place
199
(n=20). (D) The percentage of worms dived in each group.
200
201
202
203
| 2022 | Diving Behavior Reveals Humidity Sensing Ability of Water Deprived Planarians | 10.1101/2022.10.12.511880 | [
"Pei Yu",
"Qian Renzhi",
"yan Yuan",
"Zhang Yixuan",
"Tan Liyuan",
"Li Xinran",
"Lu Chenxu",
"Chen Yuxuan",
"Chi Yuanwei",
"Hao Kun",
"Xu Zhen",
"Yang Guang",
"Shao Zilun",
"Wang Yuhao",
"Huang Kaiyuan"
] | creative-commons |
Title: Whole-genome fingerprint of the DNA methylome during chemically induced
1
differentiation of the human AML cell line HL-60/S4
2
3
Running title: DNAme of HL-60/S4 differentiation
4
5
Authors: Enoch Boasiako Antwi1,2, Ada Olins3, Vladimir B Teif4, Matthias Bieg1,5,7, Tobias
6
Bauer1,5, Zuguang Gu1,5, Benedikt Brors6, Roland Eils1,5,7,8, Donald Olins3, Naveed Ishaque1,5,7.*
7
8
Affiliations:
9
1 Division of Theoretical Bioinformatics, German Cancer Research Center (DKFZ), Heidelberg,
10
Germany.
11
2 Molecular and Cellular Engineering, Centre for Biological Signalling Studies, Freiburg
12
University, Germany.
13
3 Department of Pharmaceutical Sciences, College of Pharmacy, University of New England,
14
Portland, ME USA.
15
4 School of Biological Sciences, University of Essex, Colchester, UK
16
5 Germany Heidelberg Center for Personalized Oncology (DKFZ-HIPO), German Cancer
17
Research Center (DKFZ), Heidelberg, Germany
18
6 Division of Applied Bioinformatics, German Cancer Research Center (DKFZ), Heidelberg,
19
Germany
20
7 Center for Digital Health, Berlin Institute of Health and Charité - Universitätsmedizin Berlin,
21
Kapelle-Ufer 2, 10117, Berlin, Germany
22
8 Translational Lung Research Center Heidelberg (TLRC), German Center for Lung Research
23
(DZL), University of Heidelberg, Heidelberg, Germany
24
25
Corresponding Author: * Naveed Ishaque, naveed.ishaque@charite.de
26
27
Keywords: DNA Methylation, Promyelocyte, Granulocyte, Macrophage, Differentiation,
28
Epigenetics, Enhancer, Promoter, Multi-omics correlation
29
Summary statement
30
Epigenomics plays a major role in cell identity and differentiation. We present the DNA
31
methylation landscape of leukemic cells during in-vitro differentiation, to add another ‘omics
32
layer to better understand the mechanisms behind differentiation.
33
Abstract
34
Background: Myeloid differentiation gives rise to a plethora of immune cells in the human body.
35
This differentiation leaves strong signatures in the epigenome through each differentiated state
36
of genetically identical cells. The leukemic HL-60/S4 promyelocytic cell can be easily
37
differentiated from its undifferentiated promyelocyte state into neutrophil- and macrophage-like
38
cell states, making it an excellent system for studying myeloid differentiation. In this study, we
39
present the underlying genome and epigenome architecture of HL-60/S4 through its
40
undifferentiated and differentiated cell states.
41
42
Results: We performed whole genome bisulphite sequencing of HL-60/S4 cells and their
43
differentiated counterparts. With the support of karyotyping, we show that HL-60/S4 maintains a
44
stable genome throughout differentiation. Analysis of differential CpG methylation reveals that
45
most methylation changes occur in the macrophage-like state. Differential methylation of
46
promoters was associated with immune related terms. Key immune genes, CEBPA, GFI1,
47
MAFB and GATA1 showed differential expression and methylation. However, we observed
48
strongest enrichment of methylation changes in enhancers and CTCF binding sites, implying
49
that methylation plays a major role in large scale transcriptional reprogramming and chromatin
50
reorganisation during differentiation. Correlation of differential expression and distal methylation
51
with support from chromatin capture experiments allowed us to identify putative proximal and
52
long-range enhancers for a number of immune cell differentiation genes, including CEBPA and
53
CCNF. Integrating expression data, we present a model of HL-60/S4 differentiation in relation to
54
the wider scope of myeloid differentiation.
55
56
Conclusions: For the first time, we elucidate the genome and CpG methylation landscape of
57
HL-60/S4 during differentiation. We identify all differentially methylated regions and positions.
58
We link these to immune function and to important factors in myeloid differentiation. We
59
demonstrate that methylation plays a more significant role in modulating transcription via
60
enhancer reprogramming, rather than by promoter regulation. We identify novel regulatory
61
regions of key components in myeloid differentiation that are regulated by differential
62
methylation. This study contributes another layer of “omics” characterisation of the HL-60/S4
63
cell line, making it an excellent model system for studying rapid in vitro cell differentiation.
64
Introduction
65
Gene expression profiles differ among different cell types and change as stem cells differentiate
66
(Cheng et al., 1996; Le Naour et al., 2001; Natarajan et al., 2012). Genome wide CpG
67
methylation, an epigenetic regulation and modification process, has been shown to exhibit
68
similar dynamic behaviour during differentiation (Brunner et al., 2009; Bock et al., 2012).
69
Usually, these two changes (i.e., gene expression and CpG methylation) have been shown to
70
correlate negatively with each other, depending upon the location of the methylated CpG
71
relative to the gene body (Payer et al., 2008; Chuang, Chen and Chen, 2012; Jones, 2012;
72
Yang et al., 2014). Overall, changes in methylation patterns between cell types and tissues
73
throughout life, work to either activate or shut down specific cellular processes (Smith &
74
Meissner, 2013), making cells exhibit different phenotypic characteristics. Acting as a
75
shutdown mechanism, DNA methylation reinforces gene silencing, when expression is not
76
required in a particular cell type (Lock, et al., 1987).
77
78
Normal myeloid cell differentiation occurs within the bone marrow, where stroma cells secrete
79
cytokines to help activate myeloid-specific gene transcription (De Kleer, et al., 2014). Further
80
differentiation can occur in the peripheral tissues or blood, dependent upon exposure of the
81
myeloid precursors to cytokines and other factors, such as antigens (Geissmann et al., 2010;
82
Álvarez-Errico et al., 2015). The first direct committed step toward myeloid cell development is
83
the differentiation of multipotent progenitors (MPP) cells into common myeloid progenitor cells
84
(CMP) (Kondo, et al., 1997), (Alvarez-Errico, et al., 2015). CMP cells can then differentiate
85
further into the granulocyte-macrophage lineage progenitor (GMP) and megakaryocyte-
86
erythroid progenitor (MEP) (Iwasaki & Akashi, 2007). While CMP cells can differentiate into all
87
myeloid cell types, GMP cells give rise mainly to monocytes/macrophages and neutrophils,
88
together with a minor population of eosinophils, basophils and mast cells (Laiosa, et al., 2006),
89
(Iwasaki & Akashi, 2007), (Alvarez-Errico, et al., 2015).
90
91
The human myeloid leukemic cell line HL-60/S4 is an excellent system to study epigenetic
92
changes during chemically induced in vitro cell differentiation. HL-60/S4 cells are supposedly
93
blocked at the GMP cell state and unable to differentiate any further. The HL-60/S4 cell line is a
94
subline of HL-60 and demonstrates “faster” cell differentiation than the parent HL-60 cells.
95
Undifferentiated HL-60/S4 cells exhibit a myeloblastic or promyelocytic morphology with a
96
rounded nucleus containing 2 to 4 nucleoli, basophilic cytoplasm and azurophilic granules
97
(Birnie, 1988). Retinoic acid (RA) can induce HL-60/S4 differentiation to a granulocyte-like
98
state. 12-O-tetradecanoylphorbol-13-acetate (TPA) can induce differentiation to
99
monocyte/macrophage-like states (Fontana, Colbert and Deisseroth, 1981; Birnie, 1988).
100
101
The extent to which DNA methylation regulates these chemically induced differentiation
102
processes is not known. Likewise, the global genome wide methylation changes associated
103
with these differentiation processes have not been described. This study details the methylation
104
changes (and lack of changes), when HL-60/S4 is differentiated to granulocytes, employing RA,
105
and to macrophage, employing TPA. The information contained within this study is intended as
106
a sequel to previous studies that describe the transcriptomes (Mark Welch et al., 2017),
107
nucleosome positioning (Teif et al., 2017) and epichromatin properties (Olins et al., 2014) of
108
HL-60/S4 cells differentiated under identical conditions. The goal is to integrate these different
109
lines of information into a comprehensive description and mechanistic analysis of the cell
110
differentiation pathways in the human myeloid leukemic HL-60/S4 cell lineage.
111
Results
112
113
Little or no DNA methylation changes are observed upon HL-60/S4 cell differentiation at the
114
megabase scale
115
We performed whole genome bisulphite sequencing (WGBS) of HL-60/S4 in 3 different cell
116
differentiation states: the undifferentiated state (UN), the retinoic acid treated granulocyte state
117
(RA), and the tetradecanoyl phorbol acetate (TPA) treated macrophage state. Comparison of
118
the whole genome coverage profiles for each of the three differentiation states of HL-60/S4
119
revealed that the cell line is hypo-diploid (Mark Welch, Jauch, Langowski, Olins, & Olins, 2017)
120
and is chromosomally stable throughout differentiation (Supplementary Figure S1 A-C). A
121
comparison of HL-60/S4 cells (from 2008 and 2012) by fluorescent in situ hybridization (FISH)
122
karyotyping showed that this cell line is also stable over long time periods (Supplementary
123
Figure S1 D&E). From all the CpGs identified by WGBS on all three cell states, a total of
124
21,974,649 (82.38%) CpGs had >= 10x coverage (Table 1 and Table S1), which spanned the
125
full range of methylation rates, from 0 (completed unmethylated) to 1 (fully methylated). Most of
126
these CpGs are highly and fully methylated (> 0.75 methylation rate), with only small sets of
127
lowly and unmethylated CpGs (< 0.25 methylation rate) and partially methylated CpGs
128
(methylation rate from 0.25 to 0.75) (Figure 1 A and B). Principal component analysis of all
129
CpGs with coverage greater than 10 revealed that the RA treated samples differed only slightly
130
from the untreated sample, while the TPA samples had a much higher methylation variance,
131
compared to the other two samples (Figure 1C). However, little or no methylation differences
132
were observed among the 3 samples, when methylation rates were averaged over 10
133
megabase (Mb) windows (Figure 1D).
134
135
The single CpG methylation landscape of TPA cells differ most, when compared to UN and RA
136
Cells
137
Due to the small changes observed on the megabase scale, we focused on significantly
138
differentially methylated single CpGs positions (DMPs) for further analysis. A total of 41,306
139
unique CpGs were identified to be significantly differentially methylated (Fisher analysis, see
140
Materials and Methods). These DMPs comprise of 12,713, 17,392 and 17,100 CpGs from the
141
comparisons of RA to UN cells, TPA to UN cells and RA to TPA cells, respectively (Figure 2A).
142
A higher proportion of the DMPs identified in the comparison of TPA to UN cells were hyper-
143
methylated; but a similar number of hyper- and hypo-methylated DMPs were observed in the
144
RA to TPA cells comparison. Most of the hyper-methylated DMPs had a methylation rate shift
145
from around 0 to 0.2; hypo-methylated DMPs showed a reverse shift of methylation rate (0.2 to
146
0) (Figure 2C and D).
147
148
Enhancers are most enriched within DMPs
149
The most enriched genomic features in the hyper-methylated DMPs were enhancers,
150
transcription start sites (TSSs) of protein coding genes and CpG islands (CpGIs) for both RA
151
and TPA cells, compared to UN cells (Figure 2B). CTCF was enriched in TPA hyper-methylated
152
DMPs, but not in RA. On the other hand, CpGIs were also the most enriched feature in the
153
hypo-methylated DMPs, when RA was compared to UN cells. Enhancers alone showed a high
154
enrichment in both hyper-methylated and hypo-methylated DMPs, identified when TPA is
155
compared to UN cells (Figure 2B). In contrast to enhancers, simple repeats, epichromatin, and
156
LINE (long interspersed nuclear element) repeats were depleted within hyper- and hypo-
157
methylated regions in both RA and TPA.
158
159
We identified clusters of DMP methylation pattern changes between the 3 cell states of HL-
160
60/S4. We called these cluster “modules”. Module analysis reveals that enhancers are
161
significantly enriched in DMPs that are hypo-methylated in the TPA state, relative to UN and RA
162
(modules M6 and M12). The observed hypo-methylation for TPA treated cells corresponded
163
with lower nucleosome occupancy around the DMPs of M6 and M12 (Supplementary Figure
164
S2). M7 DMPs were similarly hypo-methylated in the TPA, compared to RA and UN cells, but
165
with lower methylation differences (Figure 2E and F). Enrichment of exons, epichromatin and
166
chromatin-interacting domains (Li Teng et al., 2015) were also observed in module M6.
167
168
CpGIs have a very dynamic differential methylation
169
CpGIs are differentially methylated, but mainly in relation to RA treated cells. CpGIs were most
170
enriched in module M1, which has DMPs that are hemi-methylated (approximated 0.5
171
methylation rate) in RA; but these DMPs showed lower methylation in TPA and UN cells.
172
Similar results were seen in module M9, where DMPs were hypo-methylated in RA, compared
173
to TPA and UN cells. Likewise, CpGI enrichment was observed for module M11, where DMPs
174
are hyper-methylated in TPA, compared to RA and UN cells.
175
176
Methylation of transcription start site DMPs correlate weakly with gene expression
177
A total of 110 and 132 genes were found to have their TSS overlapping with DMPs from RA
178
and TPA cells compared to UN cells, respectively. These overlapping DMPs had a methylation
179
rate difference of at least 0.2. RA genes showed a weak and insignificant correlation between
180
the average methylation difference of the DMPs overlapping with the TSS and the -log2 (RNA
181
expression fold change) of genes (Figure 3A). This is confirmed by the comparable number of
182
genes that have positive and negative correlation between TSS DMP methylation and gene
183
expression (Figure 3B).
184
185
However, the scatter plot of TSS overlapping DMP methylation change and -log2 (RNA
186
expression fold change) does show a weak, but significant, negative correlation (Figure 3C) and
187
a higher number of negatively correlating genes, compared to positively correlating ones
188
(Figure 3D).
189
190
Methylation of long distance regulatory regions shows negative correlation with target gene
191
expression
192
Methylation and expression of CEBPE (a major transcription factor involved in myeloid cell
193
differentiation) shows a negative correlation at the 3’ end of the gene; a region identified to be
194
an enhancer in the ROADMAP epigenome project (Figure 4A) (Roadmap Epigenomics
195
Consortium et al., 2015). The downstream region of the CEBPE gene, containing the DMPs
196
whose methylation has strong negative correlation with expression, has been shown through
197
IMPET (integrated methods for predicting enhancer targets) and CHIA-PET (Chromatin
198
Interaction Analysis by Paired-End Tag Sequencing) to interact with the upstream regions that
199
spans part of the gene body and the TSS region (L. Teng et al., 2015). No DMPs were
200
observed overlapping the TSS of the CEBPE gene; hence, no correlation between TSS
201
methylation and expression is available. The RNA expression of CEBPE (as well as CCNF and
202
PGP) in UN, RA, and TPA is shown in Figure 4B.
203
204
Furthermore, the RNA expression of the gene encoding for cyclin F, CCNF, (Figure 4D)
205
correlates weakly with the methylation of DMPs (Figure 4F) that overlap with its gene and TSS.
206
However, CCNF RNA expression has a strong negative correlation with DMPs overlapping the
207
upstream region of the PGP gene, which encodes phophoglycolate phosphatase (Figure 4E
208
correlation*). PGP RNA expression (Figure 4C) does not show a similar correlation. This region
209
has also been identified by ROADMAP as an enhancer.
210
211
Functional annotation of DMPs are mostly immune response related
212
Using DMPs with a methylation fold change greater than or equal to 2, we observed that
213
immune response related cellular functions were the most enriched biological function for all the
214
genes whose TSS overlapped with DMPs, when RA cells were compared to UN cells (Table 2).
215
Similarly, genes with their TSS overlapping DMPs in TPA compared to UN cells, were also
216
mostly related to (or involved with) phosphoproteins, signalling and defence responses,
217
including chemotaxis (Table 3). Similar observations were made when DMPs were merged into
218
DMRs and their functional associations tested in TPA, compared to UN cells (Table 5). For RA
219
compared to UN cells, the functional annotation was general cell function related (Table 4).
220
221
Key myeloid differentiation transcription factors are differentially expressed
222
From analysis of the expression and methylation profiles of important myeloid differentiation
223
regulatory transcriptions factors, it was observed that CEBPA (Supplementary Figure S3 A&B)
224
and GFI1 (Supplementary Figure S3 C&D) may be required to maintain HL-60/S4 in the
225
undifferentiated state (Figure 5). As such, downregulation of CEBPA is necessary for the further
226
differentiation of HL-60/S4 to either the neutrophil-like or macrophage-like state. Meanwhile,
227
SPI1 and CEBPB are upregulated in both differentiated states (Supplementary Figure S3 E&F
228
and K&L).
229
230
Upregulation of CEBPE (Figure 4B) is seen in RA; whereas, it is downregulated in TPA,
231
together with GFI1. In TPA treated cells, MAFB is upregulated, although still at low levels
232
(Supplementary Figure S3 G&H). GATA1 is also down-regulated in RA and upregulated in TPA
233
treated cells (Supplementary Figure S3 I&J).
234
Discussion
235
236
Differential methylation during HL-60/S4 differentiation occurs over small regions
237
Only small differences in DNA methylation were observed during HL-60/S4 cell differentiation
238
at the 10 Mb window scale (Figure 1 E). Despite the lack of large scale methylation changes
239
during the induced differentiation of HL-60/S4 cells, we observed both hyper- and hypo-
240
methylation of a large number of differentially methylated single CpGs (DMPs) with a mean
241
difference in methylation rate of 0.2 (Figure 2 A, C and D). Interestingly, the methylation rates
242
of most of the differentially methylated CpGs ranged from around 0 to 0.4, corresponding to the
243
partially methylated or unmethylated CpGs (Figure 2 C and D). This explains why only very few
244
differentially methylated CpGs could be identified, since CpGs with this methylation rate value
245
range were globally very sparse.
246
247
The DNA methylation landscape of RA cells is closer to undifferentiated HL-60/S4 cells, than to
248
TPA treated cells
249
Despite the generally similar megabase-scale methylation landscape observed in all 3
250
samples, they could be clearly distinguished using principal component analysis (Figure 1 C).
251
Whereas TPA cells were seen to be very different from UN cells based on their whole genome
252
methylation profiles, RA and UN cells were closely positioned on the axes of both principal
253
component 1 and 2. Neutrophil methylation has already been shown to be only slightly, but
254
significantly, different from the promyelocyte precursor cell methylation (Alvarez-Errico, et al.,
255
2015). Thus, the small differences seen between RA (granulocyte-like) cells and the UN
256
(promyelocytic) cell forms are consistent with that previous study.
257
258
Differential methylation is limited to a few CpGs with very low levels of methylation
259
A total of 41,306 CpGs were identified to be differentially methylated; a very small number
260
compared to the genome wide CpG numbers. The numbers of differentially methylated CpGs
261
identified by a comparison of TPA to RA and UN cells were very similar. The lowest numbers of
262
differences of differentially methylated CpGs were seen in a RA comparison to UN cells (Figure
263
2A). Hyper-methylated DMPs were only enriched in protein coding TSS and in CpGI and
264
Enhancers for both RA and TPA, compared to UN. However, CTCF sites were only enriched in
265
hyper-methylated DMPs in the TPA-UN comparison (Figure 2B). Hypo-methylated DMPs were
266
seldom enriched for any particular genomic feature, except for CpGI, which was enriched in the
267
RA-UN comparison, while enhancers showed enrichment within the TPA-UN comparison.
268
269
Changes in gene expression profiles, regulated by enhancers, may play a major role in the
270
differentiation of macrophage-like cells. Enhancers stand out from other genomic features for
271
TPA differentiated cells, which are quite different from UN cells (Figure 1C). The DMP module
272
(M6), which has full methylation of CpGs in UN and RA, but hypo-methylation in TPA, is the
273
same module that shows the highest enhancer enrichment (Figure 2D&E). These observations
274
emphasize the significance of hypo-methylation of enhancers in macrophage-like differentiation,
275
as observed in TPA-treated cells. On the other hand, modules M1 and M4 which showed either
276
hyper- and hypo-methylation, for RA compared to UN, showed little enrichment of any genomic
277
features, except CpGI. This may suggest a fine tuning of expression for already active genes,
278
while hypo-methylation of DMPs in module M6 hints at the activation of expression of genes
279
that might not be expressed in UN or RA cells.
280
281
On a broader view, hypo-methylation of enhancers, epichromatin and chromatin interaction
282
domains in TPA cells suggests a remodeling of the transcriptional regulatory circuits in this
283
state, compared to the RA and UN cell states.
284
285
Interplay of DNA CpG methylation and nucleosome occupancy is genomic context dependant
286
In TPA cells we observed lower nucleosome occupancy and hypo-methylation around the
287
DMPs of module M6 (Figure 2E and Supplementary Figure S2). This module was also enriched
288
for enhancers (Figure 2F). Similar observations were made for modules M7 and M12, albeit
289
with lower levels of methylation change, enhancer enrichment and differential nucleosome
290
occupancy changes. In modules M8 and M11, we observe hyper-methylation in TPA cells but
291
no increase in nucleosome occupancy. These modules had little or no enrichment of
292
enhancers. Similarly, other modules with hypo-methylation for either UN (modules M5 and M10)
293
or RA cells (M8 and M9) did not exhibit reduced nucleosome occupancy, nor were they
294
enriched in enhancers. This suggests that differential nucleosome occupancy that is associated
295
with differential DNA methylation in our differentiation system occurs in the genomic context of
296
enhancers. This is consistent with previous findings of changes of nucleosome occupancy and
297
DNA methylation in regulatory genomics contexts of CTCF binding and promoters (Kelly et al.,
298
2012) during cellular differentiation.
299
300
RA and TPA cells share only a few DMPs
301
We identified 12 clusters of DMP patterns, which we grouped into modules. These modules
302
revealed that most of the identified CpGs were differentially methylated only in TPA cells,
303
compared to the other differentiated states (Figure 2 E). The first 6 modules describe CpGs that
304
were differentially methylated in one cell state, by comparison to one other cell state; while the
305
latter 6 modules are for CpGs that were differentially methylated in one cell state, compared to
306
the other two cell states.
307
308
Since the differentiation of HL-60/S4 into the granulocyte-like or macrophage-like state is a
309
branched process and not linear, the effects of most of the CpGs that are differentially
310
methylated in one direction may not be important to the other differentiation direction; unless, of
311
course, the effect on CpGs is required for the differentiation process. It is conceivable that the
312
effects of differentially methylated CpGs in modules M7-12 may be related to cell differentiation
313
in general, while those in modules M1-6 may be related to specific developments of the
314
different cell states.
315
316
Both positive and negative correlations are observed comparing DNA methylation of TSS
317
regions and levels of gene expression
318
Earlier reports suggested that methylation in the promoter and the first exon inversely
319
correlated with gene expression (Brenet et al., 2011; Jones, 2012). As such, it would be
320
expected that in the HL-60/S4 cell differentiation system, DNA methylation in the TSS region of
321
genes should correlate negatively with gene expression. However, we observed equal
322
numbers of genes that showed either positive or negative correlation between TSS methylation
323
and gene expression was about equal (Figure 3). This observation suggests that there are
324
additional epigenetic modifications required at gene promoters to regulating transcriptional
325
activity (Ford et al., 2017) or that gene expression is determined by the epignenetic state of
326
multiple regulatory elements and not just the promoter (Ong and Corces, 2011).
327
328
Long-range chromatin interactions play an important role in HL-60/S4 differentiation
329
RNA expression of CEBPE exhibits a strong inverse correlation with differential methylation in
330
a downstream region of the CEBPE gene. These regions have been shown to be interacting,
331
employing CHIA-PET in the K562 leukemia cell line (Figure 4A) (Dunham et al., 2012).
332
333
Similarly, a region within the promoter of PGP was identified to contain DMPs which correlated
334
negatively with the RNA expression of CCNF (upstream of PGP) (Figure 4E and F). As these
335
two genes transcribe in opposite directions, they may share the same promoter. However, this
336
region, despite being in PGP, showed negative correlation with only CCNF. Being a Cyclin, it is
337
involved in regulating the progress through the cell cycle, but the exact function in this process
338
of differentiation is not clear. We have also presented evidence that methylation of chromatin
339
interaction partners also plays a crucial role for expression of genes in HL-60/S4 cells (Figure
340
4).
341
342
CEBPA downregulation and differential regulation of CEBPE expression are required of HL-
343
60/S4 differentiation
344
TSS methylation and RNA expression of key myeloid differentiation transcription factors SPI1,
345
CEBPB, CEBPE, CREBBP, CEBPA, DNMTs and HDACs were examined. CEBPA was
346
observed to be hyper-methylated in RA and TPA compared to UN cells (Figure S3). This
347
resulted in significant downregulation of expression of CEBPA in the differentiated states
348
compared to UN cells.
349
350
SPI1 and CEBPA, together with CEBPB are known to be required for the maintenance of CMP
351
and GMP developmental stages of myeloid cells (Alvarez-Errico, et al., 2015). However, it is
352
the counter-interaction between SPI1 and CEBPA transcription factors that decides whether a
353
GMP differentiates or not (Iwasaki & Akashi, 2007), since CEBPA is known to repress
354
macrophage differentiation induced by SPI1. However, down-regulation of CEBPA expression
355
in both RA and TPA suggests that it is significant in maintaining HL-60/S4 in the promyelocytic
356
state. Thus, down-regulating CEBPA is key to macrophage differentiation; whereas, SPI1 is
357
also expressed over 1.5-fold in both RA and TPA compared to UN cells.
358
359
Most of the other transcription factors necessary for the differentiation of macrophage and
360
granulocytes are equally regulated by RA and TPA. An exception is CEBPE, which is
361
upregulated in RA, but downregulated in TPA (Figure 4 A and B). This suggests that it is the
362
downregulation of CEBPE which permits the differentiation of HL-60/S4 into the macrophage-
363
like state.
364
365
Employing these observations, together with the data of Supplementary Figure S3, we have
366
developed a model of the HL-60/S4 differentiation program based upon the transcription
367
factors that may be required (Figure 5). In this model, we propose that down-regulation of
368
CEBPA is necessary for differentiation of HL-60/S4 cells. Whereas, CEBPE is upregulated in
369
the neutrophil-like state, its downregulation and the simultaneous upregulation of MAFB and
370
GATA1 are necessary of macrophage-like differentiation. This supports the idea that CEBPE is
371
necessary for the commitment of HL-60/S4 cells to a neutrophil-like state.
372
373
The upregulation in the expression of GATA1 and MAFB genes supports their role in
374
committing HL-60/S4 cells to a macrophage-like state. We, therefore, postulate that HL-60/S4
375
cells may only differentiate into a macrophage-like state upon down-regulation of CEBPA, in the
376
absence of CEBPE.
377
17
Conclusions
378
379
The HL-60/S4 cell line is an excellent model system for myeloid leukemia and for cell
380
differentiation studies, due to the capability of differentiating the (undifferentiated)
381
promyelocytic cell line into macrophage-like and granulocyte-like states, following TPA and
382
RA treatments, respectively. The 3 different states of this cell line show very high
383
methylation levels for most CpGs, leaving only a few partially methylated or unmethylated
384
CpGs. Genome wide DNA methylation analysis indicates that the methylation level of the
385
granulocyte-like state differs only slightly from the undifferentiated form; whereas, the
386
macrophage-like state is very different from the other two cell states.
387
388
We found 41,306 CpGs (of the ~26.7x106 measured CpGs) showed significant differential
389
methylation upon differentiation of the HL-60/S4 cells, concentrated within a group
390
characterized by very low to partially methylated CpGs. This is substantially fewer than the
391
4.93 million dynamic CpGs involved in B-cell maturation, most of which were found in later
392
stages of differentiation (Kulis et al., 2015). Furthermore, since differentiation into the
393
macrophage-like and granulocyte-like states is a branched set of events, only a few
394
differentially methylated CpGs are shared between the diverged cell states. Hence, most of
395
differentially methylated CpGs are specific to either macrophage-like or granulocyte-like
396
differentiation.
397
398
Similarly, differential methylation was limited to the genomic features that overlapped with
399
CpGs that are not fully methylated. This explains why regulatory genomic features such
400
enhancers, CpG islands and protein-coding gene TSS were enriched, while epichromatin
401
was highly depleted in the differentially methylated regions. This could also imply that once a
402
CpG becomes methylated, it is more likely to remain methylated, which is consistent with
403
observations in previous studies (Senner et al., 2012).
404
405
18
A gene encoding a key transcription factor in the differentiation of myeloid cells (CEBPA)
406
was hyper-methylated in both RA and TPA treated cells. Hyper-methylation of the promoter
407
of this gene, however, negatively correlated with gene expression, implying repression of
408
transcription of CEBPA in both the macrophage-like and granulocyte-like states. CEBPE, on
409
the other hand, was hyper-methylated and expression was down-regulated only in the
410
macrophage-like cell forms. This implies that down-regulation of CEBPE is required for
411
macrophage development. Experiments involving CEBPE “knockout/know-down” are
412
required to examine whether down-regulation of CEBPA in HL-60/S4 cells will promote
413
differentiation into a granulocyte-like state.
414
19
Materials and Methods
415
416
Samples
417
We used the human AML (acute myeloid leukemia) cell line HL-60/S4, available from ATCC
418
(CRL-3306). Differentiation of this cell line was induced with retinoic acid (RA) and 12-O-
419
tetradecanoylphorbol-13-acetate (TPA) to attain the granulocyte-like and macrophage-like
420
states, as previously described (Mark Welch et al 2017). In previous publications (Mark
421
Welch et al 2017, Teif et al 2017), the undifferentiated (UN) HL-60/S4 cells were denoted
422
“0”. In the current study the same undifferentiated cells are denoted “UN”.
423
424
Sequencing and library preparation
425
Whole genome bisulphite sequencing (WGBS) libraries were prepared for untreated (UN),
426
RA, and TPA treated HL-60/S4 cells. Libraries were prepared using the Illumina TruSeq
427
DNA Sample Preparation Kit v2-set A (Illumina Inc., San Diego, CA, USA) according to
428
manufacture guidelines. After the adapters were ligated to the library, they were treated with
429
bisulphite followed by PCR amplification. Sequencing was performed on the Illumina HiSeq
430
2000 using paired end mode with 101 cycles using standard Illumina protocols and the 200
431
cycle TruSeq SBS Kit v3 (Illumina Inc., San Diego, CA, USA).
432
433
Read alignment and methylation calling with BSMAP
434
WGBS sequencing data were analysed using BsMAP (Xi and Li, 2009) and BisSNP
435
packages. In brief, sequencing reads were adaptor-trimmed using CUTADAPT package
436
(Martin, 2011), while read alignments were performed against the human reference genome
437
(hg19 GRCh37 version hs37d5-lambda, 1000 Genomes) using the BsMAP-2.89 package
438
with non-default parameter –v 8 (Xi & Li, 2009). Putative PCR duplicates were filtered using
439
Picard (version 1.61(1094) MarkDuplicates (http://picard.sourceforge.net). Only properly
440
paired or singleton reads with minimum mapping quality score of >=30 and bases with a
441
Phred-scaled quality score of >=10 were considered in methylation calling using the
442
20
BisulfiteGenotyper command. BisulfiteTableRecalibration was called with –maxQ 40.
443
Methylation calling was done with BisSNP package (Liu, et al., 2012) and single-base-pair
444
methylation rates (b-values) were determined by quantifying evidence for methylated
445
(unconverted) and unmethylated (converted) cytosines at all CpG positions. Non-conversion
446
rates were estimated using data from mitochondrion DNA (chrM). Only CpGs with coverage
447
greater than or equal to 10x in all sample replicates were considered in downstream
448
analysis.
449
450
Differentially methylated CpGs calling
451
Fisher exact test with α = 0.05 was applied to all 17,233,911 CpGs individually to extract
452
differentially methylated positions (DMPs).
453
454
Principal component analysis (PCA)
455
Principal component analysis was done on all 17,233,911 CpGs using the princomp
456
command in R.
457
458
Genomic features analysis
459
We extracted genic features (intron, exons, intergenic regions, genes transcription start site
460
(TSS)) together with 4D genomic interaction data from gencode v17 (Harrow et al., 2012),
461
CpG Island, Laminal Associated Domains (LADS) and RepeatMasker definitions from UCSC
462
(Rosenbloom et al., 2013). Using the start and end coordinates of the genes from
463
Genecode17, TSS was defined as the region extending 2kb upstream and 1kb downstream
464
the start of the gene. RepeatMaskers considered in the enrichment analysis are: DNA repeat
465
elements (DNA), Long interspersed nuclear elements (LINE), Low complexity repeats, Long
466
terminal repeats (LTR), Rolling Circle repeats (RC), RNA repeats (RNA, rRNA, scRNA,
467
snRNA, srpRNA and tRNA), Satellite repeats, Simple repeats (micro-satellites) and Short
468
interspersed nuclear elements (SINE). Enhancer were extracted from ENCODE (Dunham et
469
21
al., 2012), FANTOM5 (Andersson et al., 2017) and Vista (Visel et al., 2007). Coordinates of
470
HL-60/S4 Epichromatin are described (Olins et al., 2014).
471
472
Enrichment analysis
473
Genomic feature and chromosome enrichment in the DMPs were estimated using the
474
formula:
475
476
DMP_enrichmentfeature= (overlap_size / data_size) / (feature_size / genome_size)
477
478
Where “data_size” is the size of the data (for either RA or TPA DMPs) been used to
479
calculate the enrichment. Note that the enrichment of the hyper and hypo-methylated DMPs
480
were calculated relative to the “data_size” or the total DMPs or DMRs called for each
481
comparison but not relative to the total of only hyper or hypo-methylated DMPs or DMRs.
482
483
Functional annotation
484
DMR functional annotations was performed with DAVID 6.8 (Huang, Sherman and Lempicki,
485
2009) using the full set of human genes as the background.
486
487
Differential methylation patterns of DMPs analysis
488
DMPs were clustered using the hclust (Murtagh, 1985) with the complete linkage method
489
after the Euclidean distances were calculated using the dist function in R. The hierarchically
490
clustered DMPs were divided into 12 clusters using cutree. The resulting clusters were
491
named as modules, from module M1 to module M12.
492
493
Feature enrichment within modules were estimated using the following formula:
494
495
Module_enrichmentfeature = (mod_feature / feature_size) / (module_size /
496
total_modules)
497
22
498
Where “mod_feature” is the size of a module overlapping with a specific genomic feature
499
and “feature_size” is total size of a genomic feature in all modules. Whereas “module_size”
500
is the total size of a module and “total_modules” is the size of the all modules together.
501
502
Extraction of differentially methylated regions (DMRs)
503
DMR calling was done by first averaging coverage and number of methylation calls in a 3
504
CpGs sliding windows with maximum size of 2kb. Fisher exact testing was done using an
505
alpha value of 0.05 to extract differentially methylated windows. Continuous differentially
506
methylated windows were merged into one and Fisher test with same conditions were
507
applied the second time ensure the regions were significantly differentially methylated.
508
Differentially methylated regions that had 3 CpGs /1kb ratio were extracted before applying
509
the final filter which states that a DMR should consist of at least 3 sliding windows. This step
510
was to eliminate regions that probably had only one truly differentially methylated CpGs. As
511
such, DMRs that were made of less than 3 windows (5 CpGs) were dropped also dropped.
512
513
Differential gene expression
514
Differentially expressed genes data estimated using the RSEM software package (Li &
515
Dewey, 2011) were obtained from our collaborators in The Josephine Bay Paul Center for
516
Comparative Molecular Biology and Evolution (USA) (Mark Welch et al., 2017).
517
518
Correlation between gene expression and TSS methylation of HL-60/S4 genes
519
Methylation and transcriptome data were integrated by first extracting genes with –log2
520
(RNA expression fold change) >= 1.5 and TSS overlapping with at least one DMP as
521
extracted using the Fisher exact test. Using this criterion we identified 114 and 221 genes for
522
RA and TPA respectively, summing up to a total of 280 unique differentially expressed
523
genes.
524
23
Secondly, genes with TSS overlapping with DMPs with methylation rate difference >=0.2
525
were extracted for functional annotation analysis. In this extraction criterion 86 and 112
526
genes were identified for RA and TPA respectively. The correlation between the average
527
methylation change of DMPs overlapping with the TSS of a gene and the –log2 (RNA
528
expression fold change) were estimated for RA and TPA genes in a scatter plot. Similarly,
529
correlation between the average methylation of DMPs overlapping with a gene TSS and the
530
gene’s expression were estimated using values from all samples and the distributions plotted
531
separately for genes whose TSS overlap with RA and TPA DMPs.
532
533
Furthermore, the correlation between the methylation of individual CpGs in the gene body
534
and TSS region and gene expression was estimated for all genes from both extraction
535
criteria together with the gene expression of transcription factors known to be involved in
536
myeloid cell differentiation (Figure 5).
537
538
List of abbreviations
539
HL-60/S4 – human myeloid leukemic cell line HL-60/S4 (ATCC CRL-3306).
540
UN – undifferentiated HL-60/S4
541
TPA – tetradecanoyl phorbol acetate treated HL-60/S4
542
CpG – Cytosine-phosphate-Guanine dinucleotide
543
CpGI – CpG island
544
DMP – differential methylated CpG position
545
DMR – differentially methylated CpG region
546
MPP – multipotent progenitor cells
547
CMP – common myeloid progenitor cells
548
GMP – granulocyte monocyte progenitor cells
549
MEP – megakaryocyte erythrocyte progenitor cell
550
WGBS – whole genome bisulphite sequencing
551
FISH – fluorescent in-situ hybridisation
552
24
M-FISH – multiplex FISH
553
LINE – long interspersed nuclear element
554
TSS – transcription start site
555
CHIA-PET – Chromatin interaction analysis by paired end tag sequencing
556
IM-PET – integrated method for predicted enhancer targets
557
558
Author Contributions
559
DEO, RE conceived the research. RE, DEO, NI supervised the study. ALO, DEO acquired
560
the samples and data. EBA, NI, VT processed the data. EBA, VT, MB, TB, ZG, NI analysed
561
data. All authors interpreted and discussed data. EBA, NI wrote the paper. All authors
562
commented on and critically revised the manuscript.
563
564
Acknowledgements
565
We thank the DKFZ-Heidelberg Center for Personalized Oncology (DKFZ-HIPO) for
566
technical support. We thank Anna Jauch for karyotyping the HL-60/S4 cells using M-FISH.
567
We thank the College of Pharmacy (University of New England) for providing space and
568
facilities to DEO and ALO, enabling the growth and characterization of HL-60/S4 cells.
569
570
Competing interests
571
No competing interests declared
572
573
Funding
574
ALO and DEO were Guest Scientists at the DKFZ Heidelberg, Germany) and recipients of
575
support from the University of New England, College of Pharmacy.
576
577
Data availability
578
Raw sequencing data was deposited at the ENA under accession PRJEB27665.
579
25
Associated processing scripts and differential methylation analysis scripts are available via
580
GitHub:
581
https://github.com/jokergoo/ngspipeline/blob/master/WGBS_pipeline.pl
582
https://github.com/eantwibo/HL60S4_methylation_scripts/
583
26
References
584
Álvarez-Errico, D., Vento-Tormo, R., Sieweke, M. and Ballestar, E. (2015). Epigenetic
585
control of myeloid cell differentiation, identity and function. Nature Reviews Immunology.
586
15(1), 7–17.
587
Andersson, R., Gebhard, C., Miguel-escalada, I., Hoof, I., Bornholdt, J., Boyd, M., Chen, Y.,
588
Zhao, X., Schmidl, C., Suzuki, T., et. al. (2017). An atlas of active enhancers across human
589
cell types and tissues. Nature. 507, 455–461.
590
Birnie, G. D. (1988). The HL60 cell line: a model system for studying human myeloid cell
591
differentiation.. The British journal of cancer. 9, 41.
592
Bock, C., Beerman, I., Lien, W. H., Smith, Z. D., Gu, H., Boyle, P., Gnirke, A., Fuchs, E.,
593
Rossi, D. J. and Meissner, A. (2012). DNA Methylation Dynamics during In Vivo
594
Differentiation of Blood and Skin Stem Cells. Molecular Cell. 47, 633–647
595
Brenet, F., Moh, M., Funk, P., Feierstein, E., Viale, A. J., Socci, N. D. and Scandura, J. M.
596
(2011). DNA methylation of the first exon is tightly linked to transcriptional silencing. PloS
597
one. 6, e14524.
598
Brunner, A. L., Johnson, D. S., Kim, S. W., Valouev, A., Reddy, T. E., Neff, N. F., Anton, E.,
599
Medina, C., Nguyen, L., Chiao, E., et. al. (2009). Distinct DNA methylation patterns
600
characterize differentiated human embryonic stem cells and developing human fetal liver.
601
Genome Res. 19, 1044–1056.
602
Cheng, T., Shen, H., Giokas, D., Gere, J., Tenen, D. G. and Scadden, D. T. (1996).
603
Temporal mapping of gene expression levels during the differentiation of individual primary
604
hematopoietic cells.. Proceedings of the National Academy of Sciences of the United States
605
of America, 93(23), 13158–13163.
606
Chuang, T.-J., Chen, F.-C. and Chen, Y.-Z. (2012). Position-dependent correlations between
607
DNA methylation and the evolutionary rates of mammalian coding exons. Proceedings of the
608
National Academy of Sciences, 109(39), 15841–15846.
609
Dunham, I., Kundaje, A., Aldred, S. F., Collins, P. J., Davis, C. A., Doyle, F., Epstein, C. B.,
610
Frietze, S., Harrow, J., Kaul, R., et. al. (2012). An integrated encyclopedia of DNA elements
611
27
in the human genome. Nature, 489(7414), 57–74.
612
Fontana, J. A., Colbert, D. A. and Deisseroth, A. B. (1981). Identification of a population of
613
bipotent stem cells in the HL60 human promyelocytic leukemia cell line. Proceedings of the
614
National Academy of Sciences. 78 , 3863–3866.
615
Ford, E., Grimmer, M. R., , Sabine Stolzenburg , Ozren Bogdanovic, A. de, Mendoza,
616
Farnham, P. J. and , Pilar Blancafort, R. L. (2017). Frequent lack of repressive capacity of
617
promoter DNA methylation identified through genome-wide epigenomic manipulation.
618
bioRxiv doi:10.1101/170506.
619
Geissmann, F., Manz, M. G., Jung, S., Sieweke, M. H., Merad, M. and Ley, K. (2010).
620
Development of monocytes, macrophages, and dendritic cells. Science. 327 , 656–661.
621
Harrow, J., Frankish, A., Gonzalez, J. M., Tapanari, E., Diekhans, M. and Kokocinski, F.
622
(2012). GENCODE: The Reference Human Genome Annotation for The ENCODE Project.
623
Genome Research, 22, 1760–1774.
624
Huang, D. W., Sherman, B. T. and Lempicki, R. A. (2009). Systematic and integrative
625
analysis of large gene lists using DAVID bioinformatics resources. Nature Protocols, 4(1),
626
44–57.
627
Jones, P. A. (2012). Functions of DNA methylation: islands, start sites, gene bodies and
628
beyond. Nature Reviews Genetics. 13(7), 484–492.
629
Kelly, T. K., Liu, Y., Lay, F. D., Liang, G., Berman, B. P. and Jones, P. A. (2012). Genome-
630
wide mapping of nucleosome positioning and DNA methylation within individual DNA
631
molecules. Genome research, 22(12), 2497–2506.
632
Kulis, M., Merkel, A., Heath, S., Queirós, A. C., Schuyler, R. P., Castellano, G., Duran-ferrer,
633
M., Russiñol, N., Vilarrasa-blasi, R., Ecker, S., et. al. (2015). Whole-genome fingerprint of
634
the DNA methylome during human B cell differentiation. Nature genetics. 47, 746-56.
635
Mark Welch, D. B., Jauch, A., Langowski, J., Olins, A. L. and Olins, D. E. (2017).
636
Transcriptomes reflect the phenotypes of undifferentiated, granulocyte and macrophage
637
forms of HL-60/S4 cells. Nucleus, 8(2), 222–237.
638
Martin, M. (2011). Cutadapt removes adapter sequences from high-throughput sequencing
639
28
reads. EMBnet.journal. https://doi.org/10.14806/ej.17.1.200
640
McLean, C. Y., Bristor, D., Hiller, M., Clarke, S. L., Schaar, B. T., Lowe, C. B., Wenger, A. M.
641
and Bejerano, G. (2010). GREAT improves functional interpretation of cis-regulatory regions.
642
Nature biotechnology. 28(5), 495–501.
643
Murtagh, F. (1985) “Multidimensional Clustering Algorithms”, in COMPSTAT Lectures 4.
644
Wuerzburg: Physica-Verlag.
645
Le Naour, F., Hohenkirk, L., Grolleau, A., Misek, D. E., Lescure, P., Geiger, J. D., Hanash,
646
S. and Beretta, L. (2001). Profiling Changes in Gene Expression during Differentiation and
647
Maturation of Monocyte-derived Dendritic Cells Using Both Oligonucleotide Microarrays and
648
Proteomics. Journal of Biological Chemistry, 276(21), 17920–17931.
649
Natarajan, A., Yardimci, G. G., Sheffield, N. C. and Frazer, K. (2012). Predicting cell-type −
650
specific gene expression from regions of open chromatin the genome. Genome Research.
651
22, 1711–1722.
652
Olins, A. L., Ishaque, N., Chotewutmontri, S., Langowski, J. and Olins, D. E. (2014).
653
Retrotransposon Alu is enriched in the epichromatin of HL-60 cells. Nucleus. 5, 237–246.
654
Ong, C.-T. and Corces, V. G. (2011). Enhancer function: new insights into the regulation of
655
tissue- specific gene expression. Nature Reviews Genetics, 12(4), 283–293.
656
Payer, B., Lee, J. T., Ge, B., Kwan, T., Pastinen, T., Blanchette, M., Payer, B., Lee, J.,
657
Yoder, J., Walsh, C., et. al. (2008). X Chromosome Dosage Compensation: How Mammals
658
Keep the Balance. Annual Review of Genetics, 42(1), 733–772.
659
Roadmap Epigenomics Consortium, Kundaje, A., Meuleman, W., Ernst, J., Bilenky, M., Yen,
660
A., Heravi-Moussavi, A., Kheradpour, P., Zhang, Z., Wang, J., et. al. (2015). Integrative
661
analysis of 111 reference human epigenomes. Nature, 518(7539), 317–329.
662
Rosenbloom, K. R., Sloan, C. A., Malladi, V. S., Dreszer, T. R., Learned, K., Kirkup, V. M.,
663
Wong, M. C., Maddren, M., Fang, R., Heitner, S. G., et. al. (2013). ENCODE Data in the
664
UCSC Genome Browser: Year 5 update. Nucleic Acids Res. 41, D56–63.
665
Senner, C. E., Krueger, F., Oxley, D., Andrews, S. and Hemberger, M. (2012). DNA
666
methylation profiles define stem cell identity and reveal a tight embryonic-extraembryonic
667
29
lineage boundary. Stem Cells, 30(12), 2732–2745.
668
Teif, V. B., Mallm, J., Sharma, T., Welch, D. B. M., Rippe, K., Eils, R., Langowski, J., Olins,
669
A. L., Olins, D. E., Teif, V. B., Mallm, J., Sharma, T., Welch, D. B. M., Rippe, K., Eils, R.,
670
Langowski, J., Olins, A. L. and Nucleosome, D. E. O. (2017). Nucleosome repositioning
671
during differentiation of a human myeloid leukemia cell line. Nucleus. 8, 188–204.
672
Teng, L., He, B., Wang, J. and Tan, K. (2015). 4DGenome: A comprehensive database of
673
chromatin interactions. Bioinformatics, 31(15), 2560–2564.
674
Visel, A., Minovitsky, S., Dubchak, I. and Pennacchio, L. A. (2007). VISTA Enhancer
675
Browser - A database of tissue-specific human enhancers. Nucleic Acids Res. 35, D88–92.
676
Xi, Y. and Li, W. (2009). BSMAP: whole genome bisulfite sequence MAPping program. BMC
677
bioinformatics. 10, 232.
678
Yang, X., Han, H., DeCarvalho, D. D., Lay, F. D., Jones, P. A. and Liang, G. (2014). Gene
679
body methylation can alter gene expression and is a therapeutic target in cancer. Cancer
680
Cell. 26, 577–590.
681
30
Figure Legends
682
Figure 1: Analysis of DNA methylome upon chemical induction of differentiation of HL-60/S4
683
cells. A. Whole genome CpG methylation rate density plot. B. Box plots summarising the
684
distribution of CpG methylation rates per sample replicates for the ~22 million CpGs with
685
coverage greater than or equal to 10x in all samples. The upper and lower limits of the boxes
686
represent the first and third quartiles respectively, and the black horizontal line is the median.
687
The whiskers indicate the variability outside the upper and lower quartiles. C. Principal
688
component analysis of the WGBS data for the three treated samples. D. Circular
689
representation of DNA methylation rates for the different treatments. CpG methylation rates
690
were averaged over 10-Mb windows and are presented as heat map tracks. The heat maps
691
show the DNA methylation change with respect to the sample in the next inner track.
692
693
Figure 2: Differentially methylated CpGs (DMPs) analysis. A. Number of DMPs identified
694
with Fisher exact test for each comparison. RA and TPA are the DMPs identified when RA or
695
TPA was compared to the control (UN), while RA vs TPA is the comparison in which RA was
696
compared to TPA. B. Enrichment of genomic features in the hyper-methylated (left) and
697
hypo-methylated (right) DMPs in RA and TPA, compared UN cells. C. The density plot of the
698
methylation rates of DMPs. Hyper and hypo-methylated DMPs are denoted by (hyper) and
699
(hypo) respectively. On the left and right panels show the distribution of DMPs identified in
700
the RA and TPA compared to UN cells respectively. D. Modules identified from the
701
unsupervised clustering of the DMPs. E. Genomic feature enrichment in the 12 modules
702
identified. F. Enrichment of genomic features in the 12 identified modules.
703
704
Figure 3: Correlation between TSS methylation and gene expression A. The scatter plot of
705
TSS DMPs methylation change and fold change of RNA expression values (-log2
706
transformed) for genes which had their TSS overlapping with DMPs with 0.2 methylation
707
difference between RA and UN cells. B. The distribution of Pearson correlation coefficient
708
values between the average methylation of DMPs overlapping with TSS and expression
709
31
intensity for the genes with differentially methylated TSS in RA relative to UN cells. C. The
710
scatter plot of TSS DMPs methylation change and fold change of expression values (-log2
711
transformed) for genes which had their TSS overlapping with DMPs with 0.2 methylation
712
difference between TPA and UN cells. D. The distribution of Pearson correlation coefficient
713
values between the average methylation of DMPs overlapping with TSS and expression
714
intensity for the genes with differentially methylated TSS in TPA relative to UN cells.
715
716
Figure 4: Gene expression integration with CpG differential methylation shows that CEBPE
717
expression is regulated by methylation of downstream region. A. The promoter region of
718
CEBPE gene shows a strong inverse correlation between expression and methylation of
719
DMPs in its downstream region. Top panel: the interacting regions are regions identified
720
with IM-PET in K562 cells and confirmed by CHIA-PET. Second panel shows genomic
721
coordinated of CEBPE gene on chromosome 14 whereas E029 chromHMM panel shows the
722
genomic features along the CEBPE gene in E029 (primary monocyte cells from peripheral
723
blood). Correlation panel shows the correction of between DMPs along the genomic
724
coordinates and the gene expression of CEBPE for all three states. Methylation panel shows
725
the methylation rate of CpGs along the gene coordinates +/- 2kb. B-D. The expression of
726
CEBPE (B), PGP (C) and (CCNF) for the three different states. E, F. The cyclin-F-box
727
protein coding gene CCNF interacts with a distant upstream region which regulated its
728
expression through methylation. Panel description is similar to figure 4A except the
729
“correlation*” panel in 4E which depicts the correlation between the methylation of CpGs
730
along the PGP gene coordinates and the expression of CCNF. The sequence panel shows
731
the nucleotide sequence of in the differentially methylated region upstream PGP (enhancer
732
region): DMPs are marked in the sequence panel by ** while the blue arrows points to TPA-
733
specific SNP sites within the differentially methylated region. 4F shows similar information
734
depicted in 4A for CCNF gene. The interaction between the CCNF promoter and the region
735
upstream PGP was also identified by ChIA-PET in the K562 cell line.
736
737
32
Figure 5: Chemical differentiation model of HL-60/S4 showing the transcription factors that
738
may play an essential role in determining cell fate. Downregulation or upregulation of gene
739
expression are denoted by “-“ or “+” respectively. Genes with no sign attach implies their
740
levels are maintained at similar levels as in UN (promyelocytic) state.
741
33
Tables and table legends
742
743
UN
RA
TPA
Treatment
None
Retinoic acid
tetradecanoyl phorbol
acetate
State
Undifferentiated
Granulocyte
Macrophage
Measured CpGs
26,681,926
26,681,926
26,647,233
Genome coverage (x)
28.87
29.43
27.56
CpG coverage (x)
21.90
22.60
20.20
ChrM conversion rate
0.999
0.999
0.998
744
Table 1: CpG coverage statistics. A summary of the whole genome bisulphite sequencing
745
(WGBS) data for the undifferentiated HL-60/S4 (UN), and retinoic acid (RA) and
746
tetradecanoyl phorbol acetate (TPA) treated cells.
747
748
34
749
Term
%
Enrichment PValue
Glycoprotein binding
3.85
19.86
0.01
Translation
7.69
4.46
0.01
Defence response
10.26
3.2
0.01
Immunoglobulin-like V-type domain
5.13
8.01
0.01
Signal peptide
26.92
1.6
0.03
Steroid binding
3.85
11.48
0.03
Protein biosynthesis
5.13
5.32
0.04
Lipoprotein
8.97
2.72
0.04
Positive regulation of cell migration
3.85
8.29
0.05
Peroxisome
3.85
8.33
0.05
Enzyme binding
7.69
2.81
0.06
Positive regulation of locomotion
3.85
7.53
0.06
Endocytosis signal motif
2.56
27.58
0.07
Defence response to Gram-positive bacterium
2.56
24.6
0.08
Ankyrin
5.13
3.94
0.08
Phospholipid catabolic process
2.56
22.36
0.08
Cell membrane
17.95
1.59
0.09
SH2 domain binding
2.56
20.41
0.09
Locomotory behavior
5.13
3.59
0.1
Structure of Caps and SMACs
2.56
14.92
0.1
750
Table 2: Immune response related functions are predominant in cellular functions of genes
751
with the most differentially methylated TSS in RA, compared to UN cells. The functional
752
annotation of genes with their TSS overlapping with DMPs, with a methylation rate difference
753
>=0.2 in RA, compared to UN cells, for which gene expression data was available. The p-
754
35
value is the calculated hypergeometric binomial calculated in DAVID.
755
756
36
Term
%
Enrichment PValue
Defense response
11.43
3.47
0
Phosphatase activity
5.71
4.29
0.01
Positive regulation of locomotion
3.81
7.27
0.02
Chemotaxis
5.71
3.9
0.02
Peroxisome
3.81
7.09
0.02
Leukocyte transendothelial migration
3.81
5.75
0.03
Opsonization
1.9
59.33
0.03
Immunoglobulin-like fold
7.62
2.59
0.03
Translation
5.71
3.23
0.04
Immune response
8.57
2.32
0.04
RNA binding
8.57
2.23
0.04
Steroid binding
2.86
8.34
0.05
Phosphoprotein
44.76
1.23
0.06
Intracellular protein transport
5.71
2.86
0.06
p53 signalling pathway
2.86
7.48
0.06
MAPK signalling pathway
4.76
3.17
0.06
Positive regulation of phagocytosis
1.9
29.67
0.06
Regulation of leukocyte activation
3.81
4.29
0.06
Zinc finger region:C3H1
1.9
29.11
0.07
Protein biosynthesis
3.81
4.05
0.07
Signal peptide
22.86
1.4
0.08
Regulation of apoptosis
8.57
1.99
0.08
Macrophage activation
1.9
23.73
0.08
Small GTPase mediated signal transduction
4.76
2.92
0.09
Calcium-binding
1.9
21.03
0.09
37
Antimicrobial
3.81
3.69
0.09
Cell cycle
5.71
2.48
0.09
Cytoskeleton organization
5.71
2.45
0.09
Structure of Caps and SMACs
1.9
14.92
0.1
Palmitate moiety binding
3.81
3.58
0.1
757
Table 3: Immune response related functions are predominant in cellular functions of genes
758
with the most differentially methylated TSS in TPA compared to UN cells. The functional
759
annotation of genes with their TSS overlapping with DMPs with a methylation rate difference
760
>=0.2 in TPA, compared to UN for which gene expression data was available. The p-value is
761
the calculated hypergeometric binomial calculated in DAVID.
762
763
764
765
766
38
Term
n
Enrichment
BinomP
Peroxisome proliferator activated receptor pathway
2
30.84
4.76E-05
Catabolic process
40
1.59
3.70E-04
Phagocytosis
7
3.88
5.84E-04
Organic substance catabolic process
34
1.51
1.65E-03
Cellular catabolic process
31
1.50
2.02E-03
Organelle organization
40
1.38
2.99E-03
Protein folding
6
2.28
3.97E-03
Organophosphate catabolic process
13
2.08
4.29E-03
Regulation of cholesterol transport
3
7.23
4.47E-03
Cell division
12
2.09
4.61E-03
Leukocyte migration involved in immune response
1
38.55
5.22E-03
Quinolinate metabolic process
2
25.70
5.27E-03
Positive regulation of calcium-mediated signalling
3
9.64
5.32E-03
Mitochondrion degradation
2
22.03
5.67E-03
Histone H4-K acetylation
2
11.86
5.88E-03
Nucleotide catabolic process
12
2.10
6.05E-03
Retinoic acid receptor signalling pathway
2
8.57
6.27E-03
Nucleoside phosphate catabolic process
12
2.07
6.37E-03
Purinergic nucleotide receptor signalling pathway
2
8.12
7.28E-03
Neurotransmitter metabolic process
3
8.90
7.58E-03
Heterocycle catabolic process
16
1.60
8.25E-03
Regulation of ER to Golgi vesicle-mediated
transport
2
25.70
9.02E-03
Aromatic compound catabolic process
16
1.59
9.55E-03
Organonitrogen compound metabolic process
31
1.55
9.79E-03
767
39
Table 4: Cellular functions of DMRs which were generated by merging DMPs. The biological
768
process enrichment was performed with DMRs generated from TPA-UN comparison DMPs.
769
The p-value is the calculated binomial calculated by GREAT (McLean et al., 2010).
770
771
40
Term
n
Enrichment P value
Regulation of defense response
53
1.76
1.1E-10
Endocytosis
50
2.14
3.8E-10
Negative regulation of interleukin-8 production
5
12.93
6.5E-08
Regulation of inflammatory response
27
1.98
1.3E-07
Negative regulation of protein modification process
44
1.88
1.1E-06
Negative regulation of transferase activity
27
2.15
1.1E-06
Immune response-activating signal transduction
36
1.95
1.2E-06
Activation of immune response
38
1.75
2.4E-06
Positive regulation of defense response
31
2.05
3.1E-06
Negative regulation of protein kinase activity
25
2.23
3.6E-06
Negative regulation of kinase activity
26
2.19
5.7E-06
Negative regulation of phosphorylation
35
2.23
8.1E-06
Negative regulation of protein phosphorylation
34
2.30
9.0E-06
Platelet activation
29
2.10
1.6E-05
Regulation of peptidase activity
38
1.71
1.9E-05
Toll-like receptor 5 signalling pathway
12
2.86
2.5E-05
Toll-like receptor 10 signalling pathway
12
2.86
2.5E-05
Response to lipoprotein particle stimulus
5
7.76
6.5E-05
Positive regulation of histone H4 acetylation
3
15.51
8.0E-05
Positive regulation of myeloid leukocyte differentiation
9
3.58
1.1E-04
Response to low-density lipoprotein particle stimulus
4
10.34
1.3E-04
Regulation of cysteine-type endopeptidase activity
25
2.03
2.3E-04
Regulation of meiosis
8
4.77
2.5E-04
Positive regulation of behaviour
16
2.73
3.2E-04
Regulation of meiotic cell cycle
8
4.00
5.2E-04
41
Response to estrogen stimulus
20
2.30
7.7E-04
Positive regulation of histone acetylation
6
5.82
1.1E-03
Negative regulation of glycolysis
4
12.41
1.4E-03
772
Table 5: Cellular functions of DMRs which were generated by merging DMPs. The biological
773
process enrichment was performed with DMRs generated from TPA-UN comparison DMPs.
774
The p-value is the calculated binomial calculated by GREAT (McLean et al., 2010).
775
42
Supplementary figures
776
Supplementary figure S1: Genome of HL60/S4 is stable over long time and upon
777
differentiation: Coverage plots of the WGBC data for UN (A), RA (B), and TPA (C) depicting
778
stable genome during differentiation. D and E show 2 examples of M-FISH of
779
undifferentiated HL-60/S4 over a period of 4 years depicting stability of the genome.
780
781
Supplementary figure S2: Average nucleosome occupancy around DMP of the different
782
modules as described in figure 2. Each image shows nucleosome occupancy 2000 bases
783
up- and downstream of DMPs per module. Nucleosome occupancy is shown in black, red
784
and blue for untreated, RA and TPA treated respectively. GC content refers to the
785
percentage of GC at each base relative to the DMP position.
786
787
Supplementary figure S3: Key myeloid differentiation transcription factors are differentially
788
methylated and expressed during expression. DNA methylation landscape and gene
789
expression of transcription factors know to play important role in myeloid differentiation.
790
Gene expression levels for the three differentiation states as shown in the blue bar plots
791
correspond with the methylation and correlation profiles on their left.
792
793
794
43
Supplementary tables
795
UN
RA
TPA
QC-passed reads
1,075,185,936 1,070,133,636 1,096,718,018
Read pairs
453,160,937 453,160,937
433,631,362
Unpaired reads
37,984,963
37,984,963
36,188,487
Unmapped reads (%)
12
10
18
Duplicates (%)
4
4
4
Genome-wide coverage(x)
28.87
29.43
27.56
CpGs identified
26681926
26699651
26647233
CpG coverage
21.9
22.6
20.2
ChrM conversion
0.998761
0.999071
0.998135
796
Supplementary table ST1: Read and alignment statistics of the whole genome bisulphite
797
sequencing data used in this study.
798
799
44
Supplementary tables ST1-ST13: enrichment of GO molecular function terms in modules
800
M1-M12
801
[Submitted as an EXCEL file]
802
803
Supplementary table ST14: enrichment of GO biological process terms in module M6
804
[Submitted as an EXCEL file]
805
806
807
C
D
A
B
Principal component 2 (1.28%)
Principal component 1 (97.54%)
●
●
●
RA
UN
TPA
1.0
0.8
0.6
0.4
0.2
0.0
UN RA TPA
Methylation rate
Samples
RA
UN
0MB
90MB
180MB
1
0MB
90MB
180MB
2
0MB
90MB
180MB
3
90MB
180MB
0MB
4
0MB
90MB
180MB
5
0MB
90MB
6
0MB
90MB
7
0MB
90MB
8
0MB
90MB
9
0MB
90MB
10
0MB
90MB
11
0MB
90MB
12
0MB
90MB
13
B
M
0
90MB
14
0MB
90MB
15
0MB
90MB
16
0MB
17
0MB
18
0MB
19
0MB
20
0MB
21
0MB
22
−0.015
−0.01
−0.005
0
0.005
Methylation change
0.6
0.65
0.7
0.75
0.8
0.85
Methylation rate
TPA
Figure 1
CpG methylation rate
Cells
UN
RA
TPA
1.0
0.0
1.0
0.0
2.0
Desnity
-0.568
-0.578
0.4
0.0
0.8
-0.4
-0.573
-0.2
0.2
0.6
1.0
0
2
4
6
8 10 12
CpG methylation rate
RA(hyper)
RA(hypo)
0.0
0.5
1.0
0
2
4
6
8 10 12
CpG methylation rate
Density
TPA(hyper)
TPA(hypo)
A
B
C
D
E
Figure 2
Enrichment scale
1.01
0.98
0.99
0.97
1.00
1.37
1.06
0.99
0.94
0.92
1.10
1.00
1.17
0.92
0.95
1.15
1.10
1.55
1.03
1.15
0.84
0.75
1.17
0.92
1.16
0.97
0.99
1.15
0.99
0.79
1.08
0.90
0.88
0.83
1.12
0.78
1.24
0.98
1.00
1.07
0.99
0.58
1.00
0.66
0.79
0.67
1.06
0.75
1.66
0.61
1.20
1.35
1.14
0.52
1.29
1.33
1.45
1.34
1.60
1.31
1.13
0.87
0.87
1.00
0.98
3.27
1.56
0.75
0.57
0.90
1.17
1.86
1.04
0.97
1.16
0.91
1.22
0.51
0.76
0.90
0.85
1.00
1.18
0.88
1.14
0.83
1.16
1.26
1.17
1.58
1.13
1.07
0.84
1.03
0.72
1.25
0.98
1.02
1.08
0.92
0.97
0.26
0.91
1.04
1.19
1.10
0.77
0.81
1.02
0.99
0.94
1.04
1.02
1.39
1.03
1.01
0.89
0.97
1.14
1.09
0.87
1.08
1.03
0.91
0.85
1.30
1.21
0.56
0.85
0.74
0.50
0.39
0.98
1.03
1.02
0.95
0.94
0.67
0.97
0.92
1.06
1.14
0.73
0.72
0.96
1.02
1.04
0.97
1.09
0.37
0.80
1.06
1.19
1.03
1.00
0.89
1.10
0.90
1.10
1.19
1.07
0.79
0.97
1.37
1.06
1.20
1.11
0.92
1.07
1.07
1.06
0.62
0.80
0.21
1.25
0.76
1.21
0.40
0.00
0.38
1.37
0.66
0.82
1.20
1.34
0.40
2.40
1.46
2.62
0.76
0.00
2.95
Genes
Exons
TSSpTSS
CpGI
Enhancer
CTCF
Epichromatin
LADs
Interactions
DNA
LINE
LTR
SINE
Satellite
Simple repeats
M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
M11
M12
0
0.5
1
1.5
2
2.5
3
RA
TPA
RA vs TPA
hyper−methylated
hypo−methylated
0
2000
6000
10000
DMPs
Comparisons
0.43
0.68
0.83
1.26
1.43
1.22
0.98
0.39
0.53
0.50
0.38
0.34
0.53
0.49
0.56
0.11
0.56
0.83
0.99
1.38
1.29
1.68
1.26
0.38
0.72
0.58
0.52
0.41
0.71
0.61
0.41
0.17
0.44
0.51
0.66
0.89
1.23
0.82
0.99
0.37
0.50
0.61
0.46
0.41
0.61
0.53
0.73
0.17
0.35
0.48
0.53
0.66
0.79
1.40
0.68
0.31
0.42
0.35
0.36
0.27
0.37
0.34
0.31
0.11
RA
TPA
RA
TPA
Genes
Exons
TSS
pTSS
CpGI
Enhancer
CTCF
Epichromatin
LADs
Interactions
DNA
LINE
LTR
SINE
Satellite
Simple repeats
0
0.5
1
1.5
Enrichment scale
Hyper-methylated
Hypo-methylated
F
Methylation rate
0
0.2
0.4
0.6
0.8
1
UN
RA
TPA
M1 (6,050)
M2 (19,209)
M3 (6,849)
M4 (1,571)
M5 (1,959)
M6 (928)
M7 (1,409)
M8 (774)
M9 (1,289)
M10 (494)
M11 (265)
M12 (509)
B
D
F
ChromHMM chromatin states
TSS
Promoter flanking
Strong transcription
Weak transcription
Genic Enhancers
Active Enhanvers
Weak Enhancers
ZNF genes and repeats
Heterochromatin
Bivalent TSS
Bivalent Enhancers
Repressed polycomb
Weak repressed polycomb
E
CEBPE
PGP
CCNF
Figure 5
CMP
Myeloblast
Promyelocyte
Neutrophil
Macrophage
Monoblast
CEBPA
SPI1
GFI1
CEBPA
CEBPE
SPI1
GFI1
- CEBPA
- CEBPE
+ GFI1
+ MAFB
GATA1
SPI1
+RA
+TPA
-
+
| 2019 | Whole-genome fingerprint of the DNA methylome during chemically induced differentiation of the human AML cell line HL-60/S4 | 10.1101/608695 | [
"Antwi Enoch Boasiako",
"Olins Ada",
"Teif Vladimir B",
"Bieg Matthias",
"Bauer Tobias",
"Gu Zuguang",
"Brors Benedikt",
"Eils Roland",
"Olins Donald",
"Ishaque Naveed"
] | creative-commons |
Brain tissue properties and morphometry assessment after
chronic complete spinal cord injury
Andreas Hug MD1, Adriano Bernini2, Haili Wang1, Antoine Lutti PhD2, Johann M.E. Jende
MD4, Markus Böttinger MD1, Marc-André Weber MD3,5, Norbert Weidner MD1, and Simone
Lang PhD1
1Spinal Cord Injury Center, Heidelberg University Hospital, Heidelberg, Germany
2Laboratory for Research in Neuroimaging (LREN), Department of Clinical Neurosciences
Lausanne University Hospital, University of Lausanne, Lausanne, Switzerland
3Department of Radiology, Heidelberg University Hospital, Heidelberg, Germany
4Department of Neuroradiology, Heidelberg University Hospital, Heidelberg, Germany
5Institute of Diagnostic and Interventional Radiology, University Medical Center Rostock,
Rostock, Germany
Corresponding author:
Andreas Hug MD
Spinal Cord Injury Center
Heidelberg University Hospital
Schlierbacher Landstraße 200a
69118 Heidelberg. Germany
Phone: +49-6221-5626322
Fax: +49-6221-5626345
Email: andreas.hug@med.uni-heidelberg.de
Abstract
There is much controversy about the potential impact of spinal cord injury (SCI) on brain’s
anatomy and function, which is mirrored in the substantial divergence of findings between
animal models and human imaging studies. Given recent advances in quantitative magnetic
resonance imaging (MRI) we sought to tackle the unresolved question about the link between
the presumed injury associated volume differences and underlying brain tissue property
changes in a cohort of chronic complete SCI patients. Using the established computational
anatomy methods of voxel-based morphometry (VBM) and voxel-based quantification (VBQ)
we performed statistical analyses of grey matter volume and parameter maps indicative for
brain’s myelin, iron and free tissue water content in complete SCI patients (n=14) and healthy
individuals (n=14). Our whole-brain analysis showed significant white matter volume loss in
the rostral and dorsal part of the spinal cord consistent with Wallerian degeneration of
proprioceptive axons in the lemniscal tract in SCI subjects, which correlated with spinal cord
atrophy assessed with quantification of the spinal cord cross-sectional area at cervical level.
The latter finding suggests that Wallerian degeneration of the lemniscal tract represents a main
contributor to the observed spinal cord atrophy, which is highly consistent with preclinical
ultrastructural/histological evidence of remote changes in the central nervous system secondary
to SCI. Structural changes in the brain representing remote changes in the course of chronic
SCI could not be confirmed with conventional VBM or VBQ statistical analysis. Whether and
how MRI based brain morphometry and brain tissue property analysis will inform clinical
decision making and clinical trial outcomes in spinal cord medicine remains to be determined.
Introduction
Spinal cord injury (SCI) is a major cause for chronic disability that profoundly affects patients’
autonomy and quality of life. Despite the abundance of empirical evidence on the local effects
of SCI along the spinal cord, our understanding of the concomitant changes in brain anatomy
and function is still limited. Animal models of SCI showed controversial results ranging from
extensive neuronal cell death in cortical areas (Hains et al., 2003) and the rubrospinal tract
(Viscomi and Molinari, 2014) to lack of upper motoneuron degeneration or cell death of
corticospinal neurons (Nielson et al., 2010; Nielson et al., 2011). The lack of in-depth
knowledge about the impact of SCI on brain anatomy in humans highlights the need to provide
in vivo analytic proof of concomitant structural changes that could inform clinical decision
making in respect to treatment and prognosis.
Computational anatomy methods using magnetic resonance imaging (MRI) and
mathematical algorithms to extract relevant brain features allow for statistical analysis of
volume, shape and surface in three dimensional brain space (Ashburner et al., 2003). One of
the well-established methods - voxel-based morphometry (VBM), was used to monitor local
grey matter volume changes following SCI to deliver conflicting results ranging from lack of
SCI related brain anatomy changes (Crawley et al., 2004) to evidence about profound
sensorimotor cortex reorganization (Jurkiewicz et al., 2006; Wrigley et al., 2009a; Henderson
et al., 2011; Freund et al., 2013). More recent reports demonstrate grey matter loss in non-
motor areas including anterior cingulate gyrus, insula, orbitofrontal gyrus, prefrontal cortex and
thalamus (Wrigley et al., 2009b; Grabher et al., 2015; Chen et al., 2017). One of the potential
reasons for the reported controversial findings is the fact that these studies pooled together
patients with incomplete and complete SCI (Crawley et al., 2004; Jurkiewicz et al., 2006; Chen
et al., 2017) not taking into account potential impact of differences in the time span since injury
(Grabher et al., 2015). The non-quantitative character of the used T1-weighted MRI protocols
represents another source for differences between studies. The computer-based estimation of
regional volumes and cortical thickness from T1-weighted data is heavily dependent on the MR
contrast, which is influenced by local histological tissue properties that give potentially rise to
spurious morphological changes (Lorio et al., 2016b).
Recent advances in quantitative MRI (qMRI) circumvent these limitations to provide
quantitative maps indicative for myelin, iron and tissue free water content (Helms et al., 2008;
Draganski et al., 2011; Lutti et al., 2014). Investigations applying qMRI restricted to a set of
regions-of-interest reported progressive volume loss in the internal capsule of SCI patients
paralleled by myelin reduction at 12 months post-injury compared to baseline (Freund et al.,
2013). Using the same technique in the very same cohort, the authors observed also myelin
reduction in thalamus, cerebellum and brainstem in the same period of time (Freund et al., 2013;
Grabher et al., 2015). These combined morphometry and tissue property findings in the early
injury phase contrast with the absence of volume differences when comparing sub-acute
(duration <1 year) and chronic (duration >1 year) patients with complete motor SCI (Chen et
al., 2017).
Here we sought to address previous limitations in the field and investigate the sensitivity
of quantitative brain tissue property MRI mapping to detect brain anatomy changes in a
homogenous cohort of chronic complete SCI subjects We used the established voxel-based
quantification (VBQ) and voxel-based morphometry to study potential differences in myelin,
iron and free tissue water content between SCI subjects and healthy controls. We hypothesized
that chronic complete SCI is associated with brain atrophy in the sensorimotor and non-
sensorimotor system paralleled by specific alterations in myelin, iron and water content.
Materials and Methods
Study participants
All study related procedures were performed after obtaining informed consent according to
protocols approved by the independent local ethics committee. We screened all patients
admitted to the Spinal Cord Injury Center at Heidelberg University Hospital, Germany, for
eligibility participating in the study. The main inclusion criterion was a spinal cord injury grade
(American Spinal Injury Association Scale (AIS) grade A) that dated back at least 3 months
before study participation. The control group was chosen with the intention to minimize age
and gender differences between groups. Clinical scoring and grading were done according to
the International Standards for Neurological Classification of Spinal Cord Injury (ISNCSCI)
(Kirshblum et al., 2011).
MRI acquisition
All MRI data were acquired on a 3 Tesla scanner (Siemens Verio, Siemens Healthineers,
Germany). The imaging protocol consisted of three whole-brain multi-echo 3D fast low angle
shot (FLASH) acquisitions with predominantly magnetization transfer-weighted (MTw: TR/α
= 24.5ms/6°), proton density-weighted (PDw: TR/α = 24.5ms/6°) and T1-weighted (T1w:
24.5ms/21°) contrast (Helms et al., 2009, 2008; Weiskopf et al., 2013). For each contrast we
acquired multiple gradient echoes with minimum at 2.46ms and equidistant 2.46ms echo
spacing. Per echo 176 sagittal partitions with 1mm isotropic voxel size (field of view and matrix
size 256x240) and alternating readout polarity were acquired. The number of echoes was 7/8/8
for the MTw/PDw/T1w acquisitions to keep the TR value identical for all contrasts. We used
parallel imaging along the phase-encoding direction (acceleration factor 2 with GRAPPA
reconstruction) (Griswold et al., 2002) and partial Fourier (factor 6/8) along the partition
direction.
Map calculation
The R2*, MT, PD* and R1 quantitative maps were calculated as previously described
(Draganski et al., 2011). For map calculation we used in-house software running under SPM12
(Wellcome Trust Centre for Neuroimaging, London, UK; www.fil.ion.ucl.ac.uk/spm) and
Matlab 7.11 (Mathworks, Sherborn, MA, USA). R2* maps were estimated from the regression
of the log-signal of the eight PD-weighted echoes. MT and R1 maps were created using the
MTw, PDw and T1w data averaged across all echoes (Helms et al., 2008). All maps were
corrected for local RF transmit field inhomogeneities using the inhomogeneity correction
UNICORT algorithm in the framework of SPM (Weiskopf et al., 2011).
Voxel-based morphometry (VBM) and voxel-based quantification (VBQ)
For automated tissue classification in grey matter (GM), white matter (WM) and cerebrospinal
fluid (CSF) we used the MT maps within SPMs “unified segmentation” approach (Ashburner
and Friston, 2005) with default settings and enhanced tissue probability maps (Lorio et al.,
2016a) that provide optimal delineation of subcortical structures. Aiming at optimal anatomical
precision, we estimated subject specific spatial registration parameters using the diffeomorphic
algorithm based on exponentiated lie algebra - DARTEL (Ashburner, 2007). For VBM analysis,
we scaled the probability maps with the corresponding Jacobian determinants to preserve the
initial total amount of signal intensity. For VBQ analysis, we followed the same strategy by
applying a weighted averaging procedure that ensures the preservation of the initial signal
intensity of the MT, PD* and R2* parameter maps (Draganski et al., 2011). The resulting maps
were spatially smoothed using an isotropic Gaussian convolution kernel of 8 mm full-width-at-
half-maximum.
Spinal cord cross-sectional area (CSA) assessment
For CSA assessment we used the calculated T1-weighted images after contrast adjustment
within snap-ITK (Yushkevich et al., 2006). Delineation of spinal cord was performed by
outlining the spinal cord circumference manually (AH) at the C2/3 level slice by slice in the
axial plane yielding a total of 15 continuous slices. For an approximation of the mean cross-
sectional area of the upper spinal cord the average of these 15 slices was used.
Statistical analysis
We used parametric and nonparametric statistics from the software package JMP® - v12, for
descriptive analysis of clinical data where deemed appropriate.
We created voxel-based parametric regression models with the group factor (SCI x healthy
controls) as main predictor variable, and age, total intracranial volume (TIV - sum of GM, WM,
and CSF tissue classes) and gender as additional variables (Barnes et al., 2010) in an unpaired
two-sample t-test design as implemented in SPM12. For the whole-brain analysis we reduced
the search volume within brain’s GM or WM yielding 10 separate models (2 for VBM – GM-
VBM and WM-VBM, 8 for VBQ - GM-PD*, GM-MT, GM-R1, GM-R2*, WM-PD*, WM-
MT, WM-R1, WM-R2*). Parameter estimates and beta weights were estimated by appropriate
one-sided t-contrast statements with corresponding statistical parametric T-maps. To control
for multiple comparisons in this voxel-based analysis, family-wise error (FWE) correction
methods using Random Field Theory were applied. The peak level height threshold for
statistical significance was set at FWE p<0.05 with no cluster extent threshold.
Results
Population characteristics
MRI scans were obtained during a sampling period of 3.5 years. The main clinical
characteristics of the patient population are summarized in table 1. We recruited 14 patients and
14 control subjects. The mean age in the control and patient group were 46±16 and 55±13 years
(p=0.1147), respectively. The female male ratio was 3:11 in each group (p=1.0). The median
time since SCI was 144 (14-568) months. Lesion severity in all patients was sensorimotor
complete (AIS grade A). The TIV in the healthy control cohort was 1571±171ml (mean ± SD)
and 1479±210ml (mean ± SD) - in SCI subjects (p=0.2169).
VBM and VBQ analysis
In the whole-brain VBM analysis we observed WM volume decreases in the dorsal part of the
rostral cervical spinal cord in SCI subjects compared to healthy controls (mean difference 10±2
µl). There were no other significant grey or white matter brain volume differences (table 2; fig.
1).
The cross-sectional area (CSA) analysis at the cervical level revealed smaller CSA in SCI
subjects compared to controls (60.2±8.1 mm2 versus 74.5±10.4, p<0.001). We found a positive
correlation between the loss of cervical level CSA and WM volume at MNI -2 -48 -68 (r=0.662;
p<0.001). CSA was not associated with any other WM and GM brain volume changes.
The VBQ analysis did not reveal any significant between-group differences.
Discussion
In this study, we identified WM volume loss (approximately 10 µl) in the most rostral and
dorsal part of the spinal cord in chronic sensorimotor complete SCI subjects. This volume loss
correlated positively with spinal cord atrophy at the cervical level. In contrast to published
results despite a reasonable sample size and a highly homogenous population from clinical and
pathophysiological point of view, we were not able to find any morphometry or brain tissue
property differences in SCI patients that reached the accepted levels of statistical significance.
The reduced volume in the dorsal region of the rostral spinal cord in SCI subjects most
likely reflects Wallerian degeneration of large proprioceptive sensory axons, which has been
confirmed histologically (Becerra et al., 1995; Weber et al., 2006). It is unlikely that the volume
change at the rostral spinal cord was generated by a software algorithm induced imaging artifact
due to false classification/registration (Bookstein, 2001). Quality inspection of the normalized
images after application of the DARTEL algorithm did not indicate false classification or
registration. The close positive correlation between spinal cord cross sectional area at the C2/3
level and the WM volume at the uppermost part of the spinal cord also supports the hypothesis
of a true biological effect. Cord atrophy at the cervical and medulla oblongata level is a
consistent finding after SCI, which is associated with more severe disability (Freund et al.,
2010; Freund et al., 2011; Freund et al., 2013). In other neurological diseases such as
Parkinson’s disease volume changes were confirmed in similar regions (Jubault et al., 2009).
WM volume reductions in the brain stem topographically related to the corticospinal tract
and in the left cerebellar peduncle have been previously described in a more heterogeneous
(more incomplete SCI subjects) and less chronic cohort of SCI subjects (Freund et al., 2011;
Freund et al., 2013). However, volume changes in respective neuroanatomical regions were not
reproducible in our more homogeneous chronic SCI cohort. Volume reductions in our cohort
were located more caudal in the most rostrocaudal region of the cervical spinal cord consistent
with the localization of the dorsal funiculus. Whether degenerative changes such as retrograde
axon dieback or neuronal atrophy/cell death occur in corticospinal projections as suggested by
VBM studies (Freund et al., 2011; Freund et al., 2013) is still debated. However, most recent
preclinical studies indicate that respective alterations in the corticospinal tract – at least in
rodents – cannot be observed (Nielson et al., 2010; Nielson et al., 2011). Remote changes in
the brain related to neurogenesis or gliogenesis, which also could have produced volumetric
changes in VBM studies (Killgore et al., 2013), were not observed in animal models of SCI
(Franz et al., 2014). Post-mortem histological data from human SCI subjects are not available
to either confirm or reject such changes.
Previous studies reported inconsistencies in respect to MRI based volumetric changes in
the brain and affected regions in the brain (Crawley et al., 2004; Jurkiewicz et al., 2006;
Wrigley et al., 2009a; Freund et al., 2011; Freund et al., 2013; Hou et al., 2014). In the current
study, we were not able to attribute any VBM or VBQ brain or brainstem differences
unambiguously to chronic sensorimotor complete SCI. The median time since injury was
around 12 years in the analyzed cohort. Only one other study (Wrigley et al., 2009a) analyzed
a comparable SCI group in respect to injury severity and time since injury. However, our data
do not support their finding of extensive reduced GM volumes in motor and non-motor regions
of the brain related to SCI. We identified a correlation of brain volume reductions in respective
regions only related to age (statistical significant covariate associated with smaller GM
volumes; used as nuisance variable in our study), which has been consistently shown in
previous studies not related to SCI (Raz et al., 2007; Thompson and Apostolova, 2007;
Draganski et al., 2011). Lack of statistical power is also unlikely to explain the deviating
findings. 14 SCI subjects were investigated in the present study, whereas 13 (Freund et al.,
2013), 10 (Freund et al., 2011), 15 (Wrigley et al., 2009a), and 17 (Jurkiewicz et al., 2006) SCI
subjects were enrolled in previous MRI studies.
In summary, our VBM and VBQ analyses in a highly homogenous group of chronic SCI
subjects failed to detect main effects of chronic complete SCI on brain’s anatomy. These
findings corroborate the absence of strong preclinical evidence of secondary neurodegenerative
remote changes in the brain, yet confirm histological evidence in respect to remote changes in
the spinal cord. At least with the methodology employed in the current study the application of
MRI technology is not able to detect suitable and clinically meaningful markers in the brain,
which help facilitate clinical decision making or enrich innovative clinical trial designs.
Acknowledgements
This study was supported by grants from the Deutsche Forschungsgemeinschaft (SFB1158) to
Norbert Weidner and rom the Medical Faculty Heidelberg to Andreas Hug. We thank Bogdan
Draganski from the Department of Clinical Neurosciences Lausanne University Hospital,
University of Lausanne, Switzerland, for his expert advice.
References
Ashburner J. A fast diffeomorphic image registration algorithm. Neuroimage 2007; 38(1): 95-
113.
Ashburner J, Csernansky JG, Davatzikos C, Fox NC, Frisoni GB, Thompson PM. Computer-
assisted imaging to assess brain structure in healthy and diseased brains. Lancet Neurol 2003;
2(2): 79-88.
Ashburner J, Friston KJ. Unified segmentation. Neuroimage 2005; 26(3): 839-51.
Barnes J, Ridgway GR, Bartlett J, Henley SM, Lehmann M, Hobbs N, et al. Head size, age and
gender adjustment in MRI studies: a necessary nuisance? Neuroimage 2010; 53(4): 1244-55.
Becerra JL, Puckett WR, Hiester ED, Quencer RM, Marcillo AE, Post MJ, et al. MR-pathologic
comparisons of wallerian degeneration in spinal cord injury. AJNR Am J Neuroradiol 1995;
16(1): 125-33.
Bookstein FL. "Voxel-based morphometry" should not be used with imperfectly registered
images. Neuroimage 2001; 14(6): 1454-62.
Chen Q, Zheng W, Chen X, Wan L, Qin W, Qi Z, et al. Brain Gray Matter Atrophy after Spinal
Cord Injury: A Voxel-Based Morphometry Study. Front Hum Neurosci 2017; 11: 211.
Crawley AP, Jurkiewicz MT, Yim A, Heyn S, Verrier MC, Fehlings MG, et al. Absence of
localized grey matter volume changes in the motor cortex following spinal cord injury. Brain
Res 2004; 1028(1): 19-25.
Draganski B, Ashburner J, Hutton C, Kherif F, Frackowiak RS, Helms G, et al. Regional
specificity of MRI contrast parameter changes in normal ageing revealed by voxel-based
quantification (VBQ). Neuroimage 2011; 55(4): 1423-34.
Franz S, Ciatipis M, Pfeifer K, Kierdorf B, Sandner B, Bogdahn U, et al. Thoracic rat spinal
cord contusion injury induces remote spinal gliogenesis but not neurogenesis or gliogenesis in
the brain. PLoS One 2014; 9(7): e102896.
Freund P, Weiskopf N, Ashburner J, Wolf K, Sutter R, Altmann DR, et al. MRI investigation
of the sensorimotor cortex and the corticospinal tract after acute spinal cord injury: a
prospective longitudinal study. Lancet Neurol 2013; 12(9): 873-81.
Freund P, Weiskopf N, Ward NS, Hutton C, Gall A, Ciccarelli O, et al. Disability, atrophy and
cortical reorganization following spinal cord injury. Brain 2011; 134(Pt 6): 1610-22.
Freund PA, Dalton C, Wheeler-Kingshott CA, Glensman J, Bradbury D, Thompson AJ, et al.
Method for simultaneous voxel-based morphometry of the brain and cervical spinal cord area
measurements using 3D-MDEFT. J Magn Reson Imaging 2010; 32(5): 1242-7.
Grabher P, Callaghan MF, Ashburner J, Weiskopf N, Thompson AJ, Curt A, et al. Tracking
sensory system atrophy and outcome prediction in spinal cord injury. Ann Neurol 2015; 78(5):
751-61.
Hains BC, Black JA, Waxman SG. Primary cortical motor neurons undergo apoptosis after
axotomizing spinal cord injury. J Comp Neurol 2003; 462(3): 328-41.
Helms G, Dathe H, Kallenberg K, Dechent P. High-resolution maps of magnetization transfer
with inherent correction for RF inhomogeneity and T1 relaxation obtained from 3D FLASH
MRI. Magn Reson Med 2008; 60(6): 1396-407.
Henderson LA, Gustin SM, Macey PM, Wrigley PJ, Siddall PJ. Functional reorganization of
the brain in humans following spinal cord injury: evidence for underlying changes in cortical
anatomy. The Journal of neuroscience : the official journal of the Society for Neuroscience
2011; 31(7): 2630-7.
Hou JM, Yan RB, Xiang ZM, Zhang H, Liu J, Wu YT, et al. Brain sensorimotor system atrophy
during the early stage of spinal cord injury in humans. Neuroscience 2014; 266: 208-15.
Jubault T, Brambati SM, Degroot C, Kullmann B, Strafella AP, Lafontaine AL, et al. Regional
brain stem atrophy in idiopathic Parkinson's disease detected by anatomical MRI. PLoS One
2009; 4(12): e8247.
Jurkiewicz MT, Crawley AP, Verrier MC, Fehlings MG, Mikulis DJ. Somatosensory cortical
atrophy after spinal cord injury: a voxel-based morphometry study. Neurology 2006; 66(5):
762-4.
Killgore WD, Olson EA, Weber M. Physical exercise habits correlate with gray matter volume
of the hippocampus in healthy adult humans. Sci Rep 2013; 3: 3457.
Kirshblum SC, Waring W, Biering-Sorensen F, Burns SP, Johansen M, Schmidt-Read M, et al.
Reference for the 2011 revision of the International Standards for Neurological Classification
of Spinal Cord Injury. The journal of spinal cord medicine 2011; 34(6): 547-54.
Lorio S, Fresard S, Adaszewski S, Kherif F, Chowdhury R, Frackowiak RS, et al. New tissue
priors for improved automated classification of subcortical brain structures on MRI.
Neuroimage 2016a; 130: 157-66.
Lorio S, Kherif F, Ruef A, Melie-Garcia L, Frackowiak R, Ashburner J, et al. Neurobiological
origin of spurious brain morphological changes: A quantitative MRI study. Hum Brain Mapp
2016b; 37(5): 1801-15.
Lutti A, Dick F, Sereno MI, Weiskopf N. Using high-resolution quantitative mapping of R1 as
an index of cortical myelination. Neuroimage 2014; 93 Pt 2: 176-88.
Nielson JL, Sears-Kraxberger I, Strong MK, Wong JK, Willenberg R, Steward O. Unexpected
survival of neurons of origin of the pyramidal tract after spinal cord injury. The Journal of
neuroscience : the official journal of the Society for Neuroscience 2010; 30(34): 11516-28.
Nielson JL, Strong MK, Steward O. A reassessment of whether cortical motor neurons die
following spinal cord injury. J Comp Neurol 2011; 519(14): 2852-69.
Raz N, Rodrigue KM, Haacke EM. Brain aging and its modifiers: insights from in vivo
neuromorphometry and susceptibility weighted imaging. Ann N Y Acad Sci 2007; 1097: 84-
93.
Thompson PM, Apostolova LG. Computational anatomical methods as applied to ageing and
dementia. Br J Radiol 2007; 80 Spec No 2: S78-91.
Viscomi MT, Molinari M. Remote neurodegeneration: multiple actors for one play. Mol
Neurobiol 2014; 50(2): 368-89.
Weber T, Vroemen M, Behr V, Neuberger T, Jakob P, Haase A, et al. In Vivo High-Resolution
MR Imaging of Neuropathologic Changes in the Injured Rat Spinal Cord. 2006; 27(3): 598-
604.
Weiskopf N, Lutti A, Helms G, Novak M, Ashburner J, Hutton C. Unified segmentation based
correction of R1 brain maps for RF transmit field inhomogeneities (UNICORT). Neuroimage
2011; 54(3): 2116-24.
Wrigley PJ, Gustin SM, Macey PM, Nash PG, Gandevia SC, Macefield VG, et al. Anatomical
changes in human motor cortex and motor pathways following complete thoracic spinal cord
injury. Cereb Cortex 2009a; 19(1): 224-32.
Wrigley PJ, Press SR, Gustin SM, Macefield VG, Gandevia SC, Cousins MJ, et al. Neuropathic
pain and primary somatosensory cortex reorganization following spinal cord injury. Pain
2009b; 141(1-2): 52-9.
Yushkevich PA, Piven J, Hazlett HC, Smith RG, Ho S, Gee JC, et al. User-guided 3D active
contour segmentation of anatomical structures: significantly improved efficiency and
reliability. Neuroimage 2006; 31(3): 1116-28.
Table 1: Characteristics of SCI subjects. AIS (American Spinal Injury Association Impairment
Scale), NLI (Neurological Level of Injury)
ID
AIS
NLI
Age
(years)
Duration
(months)
P1
A
C7
66
531
P2
A
T7
62
519
P3
A
T10
32
206
P4
A
T5
59
122
P5
A
T3
55
114
P17
A
C5
21
7
P19
A
T5
59
4
P20
A
T8
65
4
P21
A
T6
61
66
P22
A
C6
53
16
P25
A
C3
52
385
P30
A
C6
65
568
P32
A
T9
62
567
P34
A
T5
58
165
Table 2: Voxel based morphometry significant findings (p<0.05 after FWE); WM = white
matter; KE = cluster extent; T = t statistic; Z = z score statistic; coordinates = coordinates in
Montreal Neurological Institute (MNI) space.
tissue
map
contrast
regions
p FWE-
corrected
KE
T
Z
coordinates
x
y
z
WM
SCI<Control
Medulla oblongata
(dorsal column)
0.021
41
5.82
4.52
-2
-48
-68
Figure legends
Figure 1:
For illustration purposes representative section planes with significant voxels in VBM analysis
at the alpha<0.001 uncorrected statistical threshold level are depicted. Color-coded voxels
depict reduced WM volume in SCI subjects compared to healthy control individuals. The color
scale represents T-values (height threshold: T=5.21, p=0.05 FWE)
Figure 1
| 2019 | Brain tissue properties and morphometry assessment after chronic complete spinal cord injury | 10.1101/547620 | [
"Hug Andreas",
"Bernini Adriano",
"Wang Haili",
"Lutti Antoine",
"Jende Johann M.E.",
"Böttinger Markus",
"Weber Marc-André",
"Weidner Norbert",
"Lang Simone"
] | null |
1
Bacterial diversity in deep-sea sediments under influence of asphalt seep at the São Paulo Plateau
1
Luciano Lopes Queirozab, Amanda Gonçalves Bendiaa, Rubens Tadeu Delgado Duartec, Diego Assis
2
das Graçasd, Artur Luiz da Costa da Silvad, Cristina Rossi Nakayamae, Paulo Yukio Sumidaa, Andre O.
3
S. Limaf, Yuriko Naganog, Katsunori Fujikurag, Hiroshi Kitazatog, Vivian Helena Pellizaria
4
a Institute of Oceanography, University of São Paulo: Praça do Oceanográfico, 191 - CEP: 05508-120,
5
São Paulo, Brazil
6
b Microbiology Graduate Program, Department of Microbiology, Institute of Biomedical Science,
7
University of São Paulo, São Paulo, Brazil;
8
c Microbiology, Immunology and Parasitology Department, Federal University of Santa Catarina:
9
CCB-MIP, Campus Trindade - PO Box 476, CEP: 88040-900, Florianópolis, Brazil
10
d Institute of Biological Science, Federal University of Pará: Rua Augusto Correa, 01 - CEP: 66075-
11
110, Belém, Brazil
12
e Institute of Environmental, Chemical and Pharmaceutical Sciences, Federal University of São Paulo:
13
Rua São Nicolau, 210 - CEP: 09913-030, Diadema, Brazil
14
f Centro de Ciências Tecnológicas da Terra e do Mar (CTTMAR), University of Vale do Itajaí: Rua
15
Uruguai, 458, - CEP: 88302-202, Itajaí-SC, Brazil
16
g Japan Agency for Marine-Earth Science and Technology, 2-15 Natsushima, Yokosuka, Kanagawa,
17
237-0061, Japan
18
Corresponding author: Luciano L. Queiroz, lqueiroz@usp.br, ORCID: 0000-0002-5260-0628
19
2
Abstract
20
Here we investigated the diversity of bacterial communities from deep-sea surface sediments under
21
influence of asphalt seeps at the Sao Paulo Plateau using next-generation sequencing (NGS) method.
22
Sampling was performed at North São Paulo Plateau using the human occupied vehicle Shinkai 6500
23
and her support vessel Yokosuka. The microbial diversity was studied at two surficial sediment layers
24
(0-1 and 1-4 cm) of five samples collected in cores in water depths ranging from 2,456-2,728 m.
25
Bacterial communities were studied through sequencing of 16S rRNA gene on the Ion Torrent platform
26
and clustered in operational taxonomic units. We observed high diversity of bacterial sediment
27
communities as previously described by other studies. When we considered community composition,
28
the most abundant classes were Alphaprotebacteria (27.7%), Acidimicrobiia (20%),
29
Gammaproteobacteria (11.3%) and Deltaproteobacteria (6.6%). Most abundant OTUs at family level
30
were from two uncultured bacteria from Actinomarinales (5.95%) and Kiloniellaceae (3.17%). The
31
unexpected high abundance of Alphaproteobacteria and Acidimicrobiia in our deep-sea microbial
32
communities may be related to the presence of asphalt seep at North São Paulo Plateau, since these
33
bacterial classes contain bacteria that possess the capability of metabolizing hydrocarbon compounds.
34
Keywords: asphalt seep, deep-sea sediment, diversity, microorganisms, São Paulo Plateau
35
36
3
Introduction
37
Deep-sea ecosystems represent most of Earth surface. The seabed is composed of several types of
38
habitat as hydrothermal vents, cold seeps, seafloor and subseafloor (Jørgensen and Boetius 2007;
39
Orcutt et al. 2011). Seafloor sediments are particularly interesting due to their geochemical
40
characteristics, sedimentary dynamics and greater habitat stability that are important factors to
41
structuring communities of macro and microorganisms. Research on microbial diversity in superficial
42
sediment habitat has been intensified, in an effort to better understanding how spatial and temporal
43
patterns are determined (Zinger et al. 2011; Nemergut et al. 2013).
44
Spatial distribution of microorganisms in deep-sea habitats has been studied in several locations,
45
from Arctic sediments in the Pacific Ocean (Li et al. 2009), to Siberian continental margin (Bienhold et
46
al. 2012), eastern South Atlantic sediments near Angolan coast (Schauer et al. 2010) and Southwestern
47
Atlantic pockmarks close to the Brazilian coast (Giongo et al. 2015). Although there are few studies in
48
deep-sea habitats from South Atlantic Ocean, the knowledge of how, what and where microorganisms
49
inhabit is incipient compared with similar environments from North Atlantic or other better studied
50
deep-sea basins.
51
The Brazilian coast is known for the presence of large oil fields under seafloor sediments
52
(Coward et al. 1999; Winter et al. 2007). Campos and Santos basins are important oil-producing areas
53
of Brazil, responsible for more than 71% of the country’s oil production (Almada and Bernardino
54
2017). Considering the existence of oil and gas reservoirs in these basins, it was expected that
55
chemosynthetic ecosystems exist and a joint Japanese-Brazilian Iatá-Piúna cruise was conducted to
56
investigate that hypothesis. This cooperative project integrated the Quest for the Limits of Life
57
(QUELLE) 2013 carried out by JAMSTEC (Japan Agency for Marine-Earth Science and Technology).
58
During the cruise that explored the deep seafloor of the North São Paulo Plateau in Espírito Santo
59
4
Basin (2500-3600 m), asphalt seeps were found at a depth of 2,700 m colonised by non-chemosynthetic
60
megafaunal organisms (Fujikura et al. 2017). They also found that, in non-asphalt seeps areas, outcrops
61
of mudstone were covered by black manganese oxide crusts and nodules were also present (Aguiar et
62
al. 2014; Fujikura et al. 2017; Jiang et al. 2018). These two particular conditions of the study area may
63
be important factors determining patterns of bacterial diversity.
64
A previous study carried out by the Iatá-Piúna consortium (Jiang et al. 2018) using PCR-DGGE
65
method found high and widespread dominance of Proteobacteria and Firmicutes at sediment samples,
66
including asphalt seep area. The two predominant species were Erythrobacter citreus strain VSW309
67
detected in hydrothermal vents and Thalossospira xianhensis strain MT02 a hydrocarbon-degrading
68
marine bacterium. They also found that microbial community composition between sediment core
69
depths was different.
70
Here we investigate the diversity of bacterial communities from deep-sea surface sediments
71
under the influence of asphalt seeps at the Sao Paulo Plateau using next generation sequencing (NGS).
72
Bacterial community assembly was accessed using high-throughput 16S rRNA gene sequencing on an
73
Ion Torrent PGM platform and by quantitative amplification (qPCR) with the aim of (1) describing
74
bacterial diversity and (2) estimating bacterial populations present in sediment depth layers.
75
76
Material and Methods
77
Description of sampling sites
78
Sediment samples were collected during 2nd leg of ‘Iatá-piuna cruise’ expedition, a collaborative
79
project between Brazil and Japan inserted in the QUELLE (Quest for Limit of Life) initiative from
80
JAMSTEC (Japan Agency for Marine-Earth Science and Technology). Sediments samples were
81
collected using the HOV ‘Shinkai 6500’ and support vessel ‘Yokosuka’ in Sao Paulo Plateau located off
82
the coast of Espírito Santo and Rio de Janeiro states, composing the Campos and Espírito Santo basins.
83
5
The study area was the North São Paulo Plateau (Figure 1), this region is located between
84
coordinates 20°30’ – 21°30’ S and 39°30’ – 38°30’ W. A total of 5 sediment cores were sampled by
85
push corers (30 cm in length and 10 cm in diameter) operated by the manipulators of Shinkai 6500.
86
Cores were subsampled on board at depth intervals of 0-1 cm, 1-4 cm, 4-7 cm, 7-10 cm and 10-13 cm.
87
Each subsample was placed in sterile sample bags and stored at -20 °C. The top two layers (0-1 and 1-4
88
cm) of sampled sediment cores from North Sao Paulo Plateau were selected (Table 1) due the presence
89
of asphalt seep. Samples N11, N12, and N13 were associated with asphalt seep and N06 and N14 were
90
from background deep-sea, distant between them 2 to 5 km.
91
DNA extraction
92
Total community DNA was extracted from the top layers (0-1 and 1-4 cm) of five sediment cores
93
from North Sao Paulo Plateau (Table 1) using PowerSoil® DNA Isolation Kit (MO BIO Laboratories
94
Inc., Carlsbad, CA, USA) following manufacturer’s instructions with adaptations: sample consisted of
95
0.5 g of homogenised sediment and after mechanic cell lysis, a thermal shock step was added, heating
96
samples at 55 °C for five minutes followed by 1 minute at -20 °C.
97
The integrity of extracted DNA was evaluated by electrophoresis in 1% agarose gel with TAE 1X
98
(Tris 0.04M, glacial acetic acid 1M, EDTA 50 mM; in pH 8), visualised with SYBR®Safe (Invitrogen,
99
Paisley, UK), and Lambda Hind III (Life Technologies, Carlbad, CA., EUA) used as molecular marker.
100
DNA was quantified using NanoDrop ND-1000 spectrophotometer (Thermo Scientific, Waltham, MA,
101
EUA) and Qubit® dsDNA HS (High Sensitivity) Assay (Life Technologies).
102
103
Ion Torrent PGM Sequencing
104
The bacterial 16S rRNA gene V3 and V4 variable regions were amplified with primers 341F (5’-
105
CCTACGGGNGGCWGCAG-3’) and 805R (5’-GACTACHVGGGTATCTAATCC-3’) (Herlemann et
106
al. 2011). PCR mixtures contained 0.5 µM of each primer, 0.7 U of Taq DNA Polymerase (Life
107
6
Technologies, Carlbad, CA., EUA), 1X Buffer, 4 mM of MgCl2, 0.2 mM of each dNTP, 0.3 mg/mL
108
BSA (Bovine Serum Albumin) and 4 ng of DNA template. Cycling conditions consisted of 5 min
109
initial denaturation at 95 °C; 2 cycles of 1 min denaturation at 95 °C, 1 min annealing at 48 °C and 1
110
min extension at 72 °C; 2 cycles of 1 min at 95 °C, 1 min at 50 °C and 1 min at 72 °C; 2 cycles of 1
111
min at 95 °C, 1 min at 52 °C and 1 min at 72 °C; and 22 cycles of 1 min at 95 °C, 1 min at 54 °C and 1
112
min at 72 °C. The first few cycles with increasing annealing temperature is an adaptation to avoid
113
mixed-template PCRs bias in the final products (Ishii and Fukui 2001).
114
Amplicons libraries obtained were purified before emulsion step with Purelink PCR Purification
115
Kit (Life Technologies, Carlbad, CA., EUA) and quantified using Qubit® dsDNA HS (High
116
Sensitivity) Assay (Life Technologies). Emulsion PCR was carried out using Ion OneTouch 2TM
117
Instrument, using the Ion PGMTM Template OT2 Reagents 400 Kit and enriched with OneTouch ES
118
(Life Technologies). The sequencing of libraries was carried out in an Ion PGMTM System, using the
119
Ion PGM Sequencing 400 Kit and deposited in two Ion 318 chip Kit v2 following the manufacturer’s
120
protocol (Life Technologies).
121
122
Quantitative PCR (qPCR)
123
Enumeration of bacterial populations was carried out by qPCR, performed in triplicates using
124
SYBR Green I system detection (Invitrogen). Previous to qPCR, DNA was purified with the
125
OneStepTM PCR Inhibitor Removal Kit (Zymo Research, USA) and diluted 1:5. The bacterial primers
126
used were 27F 5’-AGAGTTTGATCMTGGCTCAG-3’ and 518R 5’- GTATTACCGCGGCTGCTGG-
127
3’ (Muyzer et al. 1993). Each reaction contained 12.5 µL of Platinum® Quantitative PCR SuperMix-
128
UDG (Invitrogen), 0.2 µM of each primer, 0.5 µL of BSA (Bovine Serum Albumin), 5 µL of template
129
DNA and ultra-pure water to complete 25 µL final volume. Amplification conditions to bacterial
130
primers were: initial denaturation at 95 °C for 10 min, followed by 40 cycles of denaturation at 95 °C
131
7
for 1 min, annealing at 56 °C for 30 seconds and elongation at 72 °C for 30 seconds. The specificity of
132
the reaction was verified against the denaturing curve with temperatures ranging from 72 °C to 96 °C.
133
Data were analysed using Applied Biosystems software, and values of cycle threshold (Ct), logarithmic
134
correlation (R2) between number of cycles and DNA quantity in samples and reaction efficiency were
135
calculated. As positive controls of qPCR reactions, serial dilutions of 16S rRNA PCR product of
136
Escherichia coli amplified with the primers 27F-1401R (Lane 1991) were used. Thus, values of Ct
137
obtained in each reaction were utilised to determine the absolute quantity of DNA in samples and result
138
were represented by the 16S rRNA gene copy numbers per gram of sediment.
139
140
Bioinformatics
141
The first filter step was carried out using PGM software to remove low quality and polyclonal
142
sequences. We performed bioinformatics analysis using the Brazilian Microbiome Project (BMP)
143
pipeline (Pylro et al. 2014). BMP pipeline is a combination of VSEARCH (Rognes et al. 2016) and
144
QIIME (version 1.9.0) (Caporaso et al. 2010) software. Using VSEARCH barcodes and primer
145
sequences we removed from fastq file, sequences were filtered by length (fastq_trunclen 200) and
146
quality (fastq_maxee 1.0), sorted by abundance and removed singletons. After that, OTUs were
147
clustered and chimeras were removed. We assigned taxonomy using uclust method in QIIME and
148
SILVA 16S Database (version n132) as reference sequences (Quast et al. 2013). The OTU table file
149
was converted to BIOM and taxonomy metadata was added. Diversity indices of Chao1, Shannon (log
150
base 2) and Simpson were calculated among samples.
151
152
Statistical analyses
153
Alpha-diversity analysis were compared between alphalt and non-asphalt seep areas using a t-test
154
(Sokal and Rohlf 1995). The 50 most abundant OTUs were filtered and a heat map was constructed
155
8
considering taxonomic classification (Class and Order) and abundance using Ward’s hierarchical
156
clustering method (ward.d2) (Murtagh and Legendre 2014). The estimated number of bacterial 16S
157
rRNA gene copies were compared between sediment layers by Wilcoxon-Mann Whitney test (Fay and
158
Proschan 2010). Moreover, we performed beta diversity analyses to compare similarities between
159
samples through Principal Coordinates Analysis (PCoA) and using distance matrix of Bray-Curtis,
160
Jaccard, Unweighted and Weighted Unifrac. The differences were tested using Permutational analysis
161
of variance (PERMANOVA) (Anderson 2001). All analyses were carried out using the statistical
162
software R (R Development Core Team 2014), qiimer, ggplot2 (Wickham 2016), phyloseq (McMurdie
163
and Holmes 2013) and vegan packages (Oksanen et al. 2013).
164
165
Results
166
We obtained 520,863 sequences and 5,229 OTUs clustered at 97% of similarity after quality
167
control and bioinformatics analysis from 10 sediment samples using Ion Torrent PGM. The number of
168
sequences varied among samples, ranging from 1,121 sequences in N12-1 to 120,296 in N13-1.
169
Samples N06-2 (3,444), N11-2 (6,566) and N12-1 (1,121) showed a low number of reads, we rarefied
170
all samples to 25,000 reads and excluded those samples from alpha and beta-diversity analysis.
171
Alpha-diversity analysis showed that the number of observed species (OTU0.03) ranged from 904
172
to 2,282, revealing a wide range of species inhabiting the Sao Paulo Plateau (Table 2). Samples with
173
highest richness indices were N13-1 (2,282), N13-2 (2,184), and N14-1 (2,017). On the other hand,
174
samples with lowest richness indices were N11-1 (904) and N12-4 (1621). Similarly, estimated
175
richness by Chao1 index ranged from 973 to 3072 species. However, in contrast to the higher
176
variability observed in the number of OTUs and estimated richness, the Shannon and Simpson indices
177
were more uniform among samples, ranging from 7.841 to 8.216, and 0.982 to 0.987, respectively
178
9
(Table 2). We did not find significative differences of alpha-diversity between asphalt and non-asphalt
179
seep areas, except to Simpson index (Suppl. Table 1).
180
In general, the community composition from all sediment samples was similar, with most
181
sequences classified within the phyla Proteobacteria (45.7%, mean of all samples), Actinobacteria
182
(20.8%), Chloroflexi (3.74%), Acidobacteria (3%), and Gemmatimonadetes (2.1%) (Figure 2). A
183
similar trend was observed when sequences from the phyla Proteobacteria and Actinobacteria were
184
analysed at class level, with Alphaproteobacteria (27.7%), Acidimicrobiia (20%),
185
Gammaproteobacteria (11.3%) and Deltaproteobacteria (6.6%) composing the sediment community
186
(Figure 3).
187
The first and thirty most abundant OTUs were an uncultured bacterium of the order
188
Actinomarinales (Acidimicrobiia) (5.95% and 2.75%).Second and forty most abundant OTUs were an
189
uncultured bacterium of the order Rhodovibrionales and family Kiloniellaceae(Alphaproteobacteria)
190
(3.17% and 2.59%), followed by an uncultured bacterium of the order AT-s2-59
191
(Gammaproteobacteria) (1.91%). Among OTUs classified at genus level, AqS1 (Gammaproteobacteria:
192
Nitrosococcaceae) (0.69%) was found in all samples.
193
Samples were not clustered by sediment depth or asphalt seep presence/absence in heatmap
194
(Figure 4) and PCoA analyses (Suppl. Figure 1). We did not identify significant correlation between
195
community distance matrix used in the PCoA analysis and samples category (Suppl. Table 2). Heatmap
196
analysis showed the clusterization of two sample groups. Further, OTUs classified at Class and Order
197
were divided in three distinct groups, in which one group was related to Actinomarinales and
198
Rhodovibrionales orders, the second group composed by orders Rhizobiales, Rhodobacterales, AT-s2-
199
59, Steroidobacterales and uncultured Alphaproteobacteria, and a third group formed by less abundant
200
orders.
201
10
The number of 16S rRNA copies per gram of sediment was evaluated by qPCR and ranged from
202
2.36×103 to 1.7×106 copies.g-1. Some samples had low cell numbers such as N11-1 and N11-4 with
203
2.59×104 and 2.36×103 copies.g-1, respectively. The sample with the highest density was N14.1 with
204
1.7×106 copies.g-1 (Suppl. Figure 2 and Suppl. Table 3). No amplification occurred in sample N06-4 In
205
the samples N11, N13 and N14, we observed a decrease in cell number with sediment depth, but this
206
difference was not significant (Suppl. Figure 2).
207
208
Discussion
209
The discovery of asphalt seeps in North São Paulo Plateau was an important milestone in studies
210
of hydrocarbon seep environments and their associated chemosynthetic communities. This asphalt seep
211
is similar to asphalt systems found in Campeche Knolls of southern Gulf of Mexico (MacDonald et al.
212
2004) and Angola Margin (Jones et al. 2014). However, Fujikura et al. (2017) analysed the oil from the
213
asphalt seep and their results indicated that this system was not capable of sustaining chemosynthetic
214
communities. Nevertheless, our study was the first of investigate the diversity of bacterial community
215
using next generation sequencing in asphalt seep and non-asphalt seep sediments in the North São
216
Paulo Plateau.
217
Other studies developed in deep-sea surface sediments found similar values of observed species,
218
Chao1 and Shannon indices (Mahmoudi et al. 2014; Zhang et al. 2015), indicating that these
219
environments harbor highly diverse microbial communities, possibly due to their temporal stability,
220
partitioning of resources and niche diversity, allowing the coexistence of distinct microbial metabolic
221
traits (Lozupone and Knight 2007; Zinger et al. 2011; Bienhold et al. 2016). Differences of alpha
222
diversity values were not observed between samples at the asphalt seep and non-asphalt seep areas.
223
Beta-diversity analysis showed that the microbial communities distribution were not influenced
224
by sediment depth or presence/absence of asphalt seep. Despite this, we observed a prevalence of some
225
11
taxonomic groups accordingly to sediment depth. For example, four of six samples from 0-1 cm layer
226
had as the most abundant OTU an Acidimicrobiia from Actinomarinales order, while in the second
227
layer 1-4 cm, the most abundant OTU in three of five samples was an Alphaproteobacteria from
228
Rhodovibrionales order (Figure 4). Jiang et al. (2018) observed that communities from surface
229
sediments (0-4 cm) were more similar between them than communities from bottom sediments,
230
independently whether samples were asphalt or non-asphalt seeps (16-20 cm). In our study, core
231
sediments were sliced in two surface sediment samples (0-1 and 1-4 cm), which may explain the
232
homogeneity between layers caused by dispersion or even by the mixture of sediments by deep-sea
233
water currents (Meadows and Meadows 1994; Bienhold et al. 2016).
234
Proteobacteria and Actinobacteria comprised the prevalent phyla found in the samples, a pattern
235
commonly observed in marine sediments throughout the world. However, at class level, we found a
236
distinct bacterial community composition, dominated by Alphaproteobacteria, Acidimicrobiia,
237
Gammaproteobacteria and Deltaproteobacteria , in contrast with marine sediments from other regions
238
of the globe, where the predominant taxa in general, from most to least abundant, are
239
Gammaproteobacteria, Deltaproteobacteria, Planctomycetes, Actinobacteria and Acidobacteria
240
(Schauer et al. 2010; Zinger et al. 2011; Jacob et al. 2013).
241
The high abundance of Alphaproteobacteria and Acidimicrobiia in our samples may be explained
242
by the presence of oil from the asphalt seep at São Paulo Plateau (Aguiar et al. 2014; Fujikura et al.
243
2017). Some Alphaproteobacteria taxa are able to degrade hydrocarbon such as the Rhodobacteraceae
244
family (Kostka et al. 2011; Bacosa et al. 2018), which composed 4.5% of sequences in our samples.
245
Bacosa et al. (2018) found an increase in the relative abundance of Rhodobacteraceae in oil treatments
246
and, using a metagenomic approach, they could also reconstruct seven genomes, one of them classified
247
as Rhodobacteraceae and possessing several aromatic degradation genes. We found a high abundance
248
of the Kiloniellaceae in our samples (13%), a family which is represented by the single genera
249
12
Kiloniella and the type species Kiloniella laminariae (Wiese et al. 2009; Imhoff and Wiese 2014).
250
Wiese et al. (2009) showed by phylogenetic analysis that Kiloniella laminariae clustered with an
251
uncharacterized bacterium from hydrothermal plumes and this group forms a large cluster with
252
Terasakiella pusilla and Thalassospira species. Jiang et al. (2018) detected in the same area we have
253
studied the hydrocarbon-degrading bacteria Thalassospira xianhensis using PCR-DGGE method.
254
Alphaproteobacteria contains several species which are highly abundant in superficial pelagic
255
environments and have a broad spatial distribution. The most common example is Pelagibacter ubique,
256
a ubiquitous Alphaproteobacteria present in all oceans that have important functions in biogeochemical
257
cycles (Morris et al. 2002; Sunagawa et al. 2015). However, some studies in deeper pelagic
258
environments also found Alphaproteobacteria composing most of the microbial community
259
(Konstantinidis et al. 2009; Eloe et al. 2011). Therefore, this proximity between deep seawater and
260
sediment surface may allow microbial community interchange, since both environments have similar
261
chemical variables and suitable habitats for these microbial populations (Hamdan et al. 2013; Walsh et
262
al. 2016).
263
Acidimicrobiia was highly abundant in North São Paulo Plateau sediments. The taxon
264
Acidimicrobiia was recently created by updating taxonomic classification of Actinobacteria phylum to
265
include the Acidimicrobidae subclass, assigned as Acidimicrobiia class (Zhi et al. 2009). Most of our
266
sequences assigned to this phylum belonged to the order Actinomarinales, and the two most
267
representative OTUs were classified as uncultured actinobacterium (previously classified as OM1
268
clade). The OM1 clade group was also found in deep-sea waters (Eloe et al. 2011; Quaiser et al. 2011),
269
as an important component of deep-sea sediment core microbiome in several oceans (Bienhold et al.
270
2016).
271
It is generally assumed that microbial cell densities in deep-sea sediment tend to decrease with
272
increasing sediment depth (Orcutt et al. 2011). In our study this tendency was not observed, with
273
13
differences in bacterial densities not being significant, probably by the microbial communities
274
interchanges between deep seawater and sediment surface. Microbial densities in North Sao Paulo
275
Plateau might vary in deeper sediment layers (> 4 cm) not achieved in our study. In addition, the low
276
abundance of 16S rRNA gene copies corroborates with similar deep-sea sediments habitats (Jorgensen
277
et al. 2012), indicating that these habitats in the North Sao Paulo Plateau are oligotrophic, and sustain a
278
low abundant, but diverse microbial community.
279
280
Conclusions
281
Bacterial communities in the North Sao Paulo Plateau are diverse, despite their low abundance,
282
and are dominated by the classes Alphaproteobacteria and Acidimicrobiia. This community structure
283
differs from other communities from similar environments, in which Gammaproteobacteria are usually
284
more abundant. We also found a high number of unclassified sequences mainly related to
285
Actinomarinales order, suggesting that this environment can harbor groups poorly explored to date.
286
The dominance of Alphaproteobacteria potentially involved with hydrocarbon degrading might be
287
likely related to the presence of asphalt seeps, however further studies are needed to answer this
288
question.
289
290
Acknowledgements
291
We would like to thank Japan Agency for Marine-Earth Science and Technology (JAMSTEC),
292
the Oceanographic Institute of the São Paulo University (IOUSP), the Brazilian Geological Survey
293
(CPRM), Petróleo Brasileiro S.A. (Petrobras) and the Embassy of Japan in Brazil for assistance in this
294
study. We would also like to thank Fundação de Amparo a Pesquisa do Estado de São Paulo (FAPESP)
295
for financial support (Project number: 2013/09159-2) and CNPq for scholarship provided to A.O.S.L
296
14
(Process 311010/2015-6); the operating team of the HOV Shinkai 6500 and the crew of the R/V
297
Yokosuka for assistance with the survey; and all team of Laboratório de Ecologia Microbiana
298
(LECOM) for productive discussions about our methods and results, and Kleber do Espirito-Santo
299
Filho for help with maps.
300
Data
301
The nucleotide sequence data reported are available in the NCBI under BioProject PRJNA562874.
302
Authorship
303
The author LQ, RD, CN, PS, AL, YN, KF, HK and VP designed study, LQ, RD,DG,AS and VP
304
performed research, LQ, AB, RD and DG analysed data; LQ, AB, RD and DG contributed new
305
methods or models; and LQ, AB, RD, CN and VP wrote the paper.
306
References
307
Aguiar JE, de Lacerda LD, Miguens FC, Marins RV (2014) The geostatistics of the metal
308
concentrations in sediments from the eastern Brazilian continental shelf in areas of gas and oil
309
production. J South Am Earth Sci 51:91–104. doi: 10.1016/j.jsames.2013.12.005
310
Almada GV de MB, Bernardino AF (2017) Conservation of deep-sea ecosystems within offshore oil
311
fields on the Brazilian margin, SW Atlantic. Biol Conserv 206:92–101. doi:
312
10.1016/j.biocon.2016.12.026
313
Anderson MJ (2001) A new method for non-parametric multivariate analysis of variance. Austral Ecol
314
26:32–46. doi: 10.1111/j.1442-9993.2001.01070.pp.x
315
Bacosa HP, Erdner DL, Rosenheim BE, et al (2018) Hydrocarbon degradation and response of seafloor
316
sediment bacterial community in the northern Gulf of Mexico to light Louisiana sweet crude oil.
317
ISME J 12:2532–2543. doi: 10.1038/s41396-018-0190-1
318
Bienhold C, Boetius A, Ramette A (2012) The energy-diversity relationship of complex bacterial
319
15
communities in Arctic deep-sea sediments. ISME J 6:724–32. doi: 10.1038/ismej.2011.140
320
Bienhold C, Zinger L, Boetius A, Ramette A (2016) Diversity and Biogeography of Bathyal and
321
Abyssal Seafloor Bacteria. PLoS One 11:e0148016. doi: 10.1371/journal.pone.0148016
322
Caporaso JG, Kuczynski J, Stombaugh J, et al (2010) QIIME allows analysis of high-throughput
323
community sequencing data. Nat Methods 7:335–6. doi: 10.1038/nmeth.f.303
324
Coward MP, Purdy EG, Ries A, Smith DG (1999) The distribution of petroleum reserves in basins of
325
the South Atlantic margin. In: Cameron NR, Bate RH, Clure VS (eds) The Oil and Gas Habitats of
326
the South Atlantic, 153rd edn. Geological Society of London, Special Publications, London, pp
327
101–131
328
Eloe EA, Fadrosh DW, Novotny M, et al (2011) Going Deeper: Metagenome of a Hadopelagic
329
Microbial Community. PLoS One 6:e20388. doi: 10.1371/journal.pone.0020388
330
Fay MP, Proschan M a (2010) Wilcoxon-Mann-Whitney or t-test? On assumptions for hypothesis tests
331
and multiple interpretations of decision rules. Stat Surv 4:1–39. doi: 10.1214/09-SS051
332
Fujikura K, Yamanaka T, Sumida PYG, et al (2017) Discovery of asphalt seeps in the deep Southwest
333
Atlantic off Brazil. Deep Res Part II Top Stud Oceanogr 146:35–44. doi:
334
10.1016/j.dsr2.2017.04.002
335
Giongo A, Haag T, Simão TLL, et al (2015) Discovery of a Chemosynthesis-based Community in the
336
Western South Atlantic Ocean. Deep Sea Res Part I Oceanogr Res Pap. doi:
337
10.1016/j.dsr.2015.10.010
338
Hamdan LJ, Coffin RB, Sikaroodi M, et al (2013) Ocean currents shape the microbiome of Arctic
339
marine sediments. ISME J 7:685–96. doi: 10.1038/ismej.2012.143
340
Herlemann DP, Labrenz M, Jürgens K, et al (2011) Transitions in bacterial communities along the
341
2000�km salinity gradient of the Baltic Sea. ISME J 5:1571–9. doi: 10.1038/ismej.2011.41
342
Imhoff JF, Wiese J (2014) The order Kiloniellales. In: The Prokaryotes: Alphaproteobacteria and
343
16
Betaproteobacteria. Springer-Verlag Berlin Heidelberg, pp 301–306
344
Ishii K, Fukui M (2001) Optimization of Annealing Temperature To Reduce Bias Caused by a Primer
345
Mismatch in Multitemplate PCR. Appl Environ Microbiol 67:3753–3755. doi:
346
10.1128/AEM.67.8.3753-3755.2001
347
Jacob M, Soltwedel T, Boetius A, Ramette A (2013) Biogeography of Deep-sea benthic bacteria at
348
regional scale (LTER HAUSGARTEN, Fram Strait, Arctic). PLoS One 8:e72779. doi:
349
10.1371/journal.pone.0072779
350
Jiang K, Zhang J, Sakatoku A, et al (2018) Discovery and biogeochemistry of asphalt seeps in the
351
North São Paulo Plateau, Brazilian Margin. Sci Rep 8:12619. doi: 10.1038/s41598-018-30928-2
352
Jones DOB, Walls A, Clare M, et al (2014) Asphalt mounds and associated biota on the Angolan
353
margin. Deep Sea Res Part I Oceanogr Res Pap 94:124–136. doi: 10.1016/j.dsr.2014.08.010
354
Jørgensen BB, Boetius A (2007) Feast and famine--microbial life in the deep-sea bed. Nat Rev
355
Microbiol 5:770–781. doi: 10.1038/nrmicro1745
356
Jorgensen SL, Hannisdal B, Lanzen A, et al (2012) Correlating microbial community profiles with
357
geochemical data in highly stratified sediments from the Arctic Mid-Ocean Ridge. Proc Natl Acad
358
Sci 109:E2846–E2855. doi: 10.1073/pnas.1207574109
359
Konstantinidis KT, Braff J, Karl DM, DeLong EF (2009) Comparative metagenomic analysis of a
360
microbial community residing at a depth of 4,000 meters at station ALOHA in the North Pacific
361
Subtropical Gyre. Appl Environ Microbiol 75:5345–5355. doi: 10.1128/AEM.00473-09
362
Kostka JE, Prakash O, Overholt WA, et al (2011) Hydrocarbon-degrading bacteria and the bacterial
363
community response in Gulf of Mexico beach sands impacted by the deepwater horizon oil spill.
364
Appl Environ Microbiol 77:7962–7974. doi: 10.1128/AEM.05402-11
365
Lane DJ (1991) 16S/23S rRNA sequencing. In: Stackebrandt E, Goodfellow M (eds) Nucleic acid
366
techniques in bacterial systematics. Wiley, Chichester, pp 115–175
367
17
Li H, Yu Y, Luo W, et al (2009) Bacterial diversity in surface sediments from the Pacific Arctic Ocean.
368
Extremophiles 13:233–246. doi: 10.1007/s00792-009-0225-7
369
Lozupone C a, Knight R (2007) Global patterns in bacterial diversity. Proc Natl Acad Sci U S A
370
104:11436–11440. doi: 10.1073/pnas.0611525104
371
MacDonald IR, Bohrmann G, Escobar E, et al (2004) Asphalt volcanism and chemosynthetic life in the
372
Campeche Knolls, Gulf of Mexico. Science (80- ) 304:999–1002. doi: 10.1126/science.1097154
373
Mahmoudi N, Robeson MS, Castro HF, et al (2014) Microbial community composition and diversity in
374
Caspian Sea sediments. FEMS Microbiol Ecol 91:1–11. doi: 10.1093/femsec/fiu013
375
McMurdie PJ, Holmes S (2013) phyloseq: an R package for reproducible interactive analysis and
376
graphics of microbiome census data. PLoS One 8:e61217. doi: 10.1371/journal.pone.0061217
377
Meadows A, Meadows PS (1994) Bioturbation in deep sea Pacific sediments. J Geol Soc London
378
151:361–375. doi: 10.1144/gsjgs.151.2.0361
379
Morris RM, Rappé MS, Connon SA, et al (2002) SAR11 clade dominates ocean surface
380
bacterioplankton communities. Nature 420:806–10. doi: 10.1038/nature01240
381
Murtagh F, Legendre P (2014) Ward ’ s Hierarchical Agglomerative Clustering Method�: Which
382
Algorithms Implement Ward ’ s Criterion�? J Classif 31:274–295. doi: 10.1007/s00357-
383
Muyzer G, Waal EC de, Uitterlinden AG (1993) Profiling of complex microbial populations by
384
denaturing gradient gel electrophoresis analysis of polymerase chain Profiling of Complex
385
Microbial Populations by Denaturing Gradient Gel Electrophoresis Analysis of Polymerase Chain
386
Reaction-Amplified Genes. Appl Environ Microbiol 59:695–700
387
Nemergut DR, Schmidt SK, Fukami T, et al (2013) Patterns and processes of microbial community
388
assembly. Microbiol Mol Biol Rev 77:342–356. doi: 10.1128/MMBR.00051-12
389
Oksanen J, Blanchet FG, Kindt R, et al (2013) vegan: Community Ecology Package
390
Orcutt BN, Sylvan JB, Knab NJ, Edwards KJ (2011) Microbial ecology of the dark ocean above, at,
391
18
and below the seafloor. Microbiol Mol Biol Rev MMBR 75:361–422. doi:
392
10.1128/MMBR.00039-10
393
Pylro VS, Roesch LFW, Morais DK, et al (2014) Data Analysis for 16S Microbial Profiling from
394
Different Benchtop Sequencing Platforms. J Microbiol Methods 107:30–37. doi:
395
10.1016/j.mimet.2014.08.018
396
Quaiser A, Zivanovic Y, Moreira D, López-García P (2011) Comparative metagenomics of
397
bathypelagic plankton and bottom sediment from the Sea of Marmara. ISME J 5:285–304. doi:
398
10.1038/ismej.2010.113
399
Quast C, Pruesse E, Yilmaz P, et al (2013) The SILVA ribosomal RNA gene database project:
400
Improved data processing and web-based tools. Nucleic Acids Res 41:590–596. doi:
401
10.1093/nar/gks1219
402
R Development Core Team (2014) R: A language and environment for statistical computing.
403
Rognes T, Flouri T, Nichols B, et al (2016) VSEARCH: a versatile open source tool for metagenomics.
404
PeerJ 4:e2584. doi: 10.7717/peerj.2584
405
Schauer R, Bienhold C, Ramette A, Harder J (2010) Bacterial diversity and biogeography in deep-sea
406
surface sediments of the South Atlantic Ocean. ISME J 4:159–170. doi: 10.1038/ismej.2009.106
407
Sokal RR, Rohlf FJ (1995) Biometry, 3rd edn. W. H. Freeman, New York
408
Sunagawa S, Coelho LP, Chaffron S, et al (2015) Ocean plankton. Structure and function of the global
409
ocean microbiome. Science 348:1261359. doi: 10.1126/science.1261359
410
Walsh EA, Kirkpatrick JB, Rutherford SD, et al (2016) Bacterial diversity and community composition
411
from seasurface to subseafloor. ISME J 10:979–989. doi: 10.1038/ismej.2015.175
412
Wickham H (2016) ggplot2. Springer International Publishing, Cham
413
Wiese J, Thiel V, Gärtner A, et al (2009) Kiloniella laminariae gen. nov., sp. nov., an
414
alphaproteobacterium from the marine macroalga Laminaria saccharina. Int J Syst Evol Microbiol
415
19
59:350–356. doi: 10.1099/ijs.0.001651-0
416
Winter WR, Jahnert RJ, França AB (2007) Bacia de Campos. In: Milani EJ, Rangel HD, Bueno GV, et
417
al. (eds) Bacias Sedimentares Brasileiras – Cartas Estratigráficas., 15th edn. Boletim de
418
Geociências da Petrobras, Rio de Janeiro, pp 511–529
419
Zhang J, Sun Q, Zeng Z, et al (2015) Microbial diversity in the deep-sea sediments of Iheya North and
420
Iheya Ridge, Okinawa Trough. Microbiol Res 177:43–52. doi: 10.1016/j.micres.2015.05.006
421
Zhi X-Y, Li W-J, Stackebrandt E (2009) An update of the structure and 16S rRNA gene sequence-
422
based definition of higher ranks of the class Actinobacteria, with the proposal of two new
423
suborders and four new families and emended descriptions of the existing higher taxa. Int J Syst
424
Evol Microbiol 59:589–608. doi: 10.1099/ijs.0.65780-0
425
Zinger L, Amaral-Zettler L a, Fuhrman J a, et al (2011) Global patterns of bacterial beta-diversity in
426
seafloor and seawater ecosystems. PLoS One 6:e24570. doi: 10.1371/journal.pone.0024570
427
428
20
Tables
429
Table 1. Coordinates, depths and alphalt seep presence/absence of sediment samples from North São
430
Paulo Plateau.
431
Dive
No.
Sample
Core
No.
Latitude (S)
Longitude
(W)
Seafloor
depth (m)
Asphalt
seep area
Sediment Layer
1343
06
02
20°41'37.57"
38°38'11.86"
2728
No
N06-1cm
N06-4cm
1345
11
05
20°43'8.05"
38°39'6.23"
2720
Yes
N11-1cm
N11-4cm
1346
12
03
20°43'54.20"
38°39'43.76"
2651
Yes
N12-1cm
N12-4cm
1346
13
04
20°43'54.20"
38°39'43.76"
2651
Yes
N13-1cm
N13-4cm
1347
14
03
20°44'17.20"
38°40'7.61"
2456
No
N14-1cm
N14-4cm
432
21
Table 2. Richness and alpha diversity data found in North region from São Paulo Plateau and obtained
433
by Ion Torrent;
434
Samples
Readsa
Observed
species
(OTU0.03)
Chao1
(sd)
Shannon
Simpson
N06-1cm
94184
1996
2587±59
8.106
0.984
N11-1cm
38311
904
973±16
7.841
0.987
N12-4cm
65601
1702
2076±45
8.120
0.986
N13-1cm
120296
2282
3072±70
8.216
0.985
N13-4cm
88568
2184
2888±66
8.190
0.985
N14-1cm
81735
2017
2661±62
7.914
0.982
N14-4cm
26579
1621
1842±30
8.043
0.983
a Number of reads of all samples were rarified to 25000 reads before alpha-diversity analysis.
435
436
437
22
Figure Captions
438
Figure 1. Map and location of samples in the São Paulo Plateau. (A) Location São Paulo Plateau and
439
(B) distribution of superficial sediments samples in North São Paulo Plateau.
440
Figure 2. Heat map with 50 most abundant bacterial OTUs (classified by Class and Order) among the
441
different depths and asphalt seep presence/absence in North São Paulo Plateau.
442
Figure 3. Relative abundance of the most abundant phyla found in North São Paulo Plateau.
443
Figure 4. Relative abundance of the most abundant classes found in North São Paulo Plateau.
444
Deep Layer
([SUGEEEEENEENNE Asphatt Seep Ow
Gammaproteobacteria - UBA10353 marine group
Dehalococcoidia - S085
Deitaproteobacteria - NB1~)
Deltaproteobacteria - Myxococcales
Dadabacteriia - Dadabacteriaies
Rhodothermia - Rhodothermaies
Acidimicrobiia - Microtrichales
Gemmatimonadetes - Gemmatimonadaies 2o
Gammaproteobacteria - Nitrosococcales
Alphaproteobacteria - Rhodospirilales
Parcubacteria - Candidatus Yanofskybacteria
Thermoleophilia - Gaielales
Alphaproteobacteria - Rhizobiales
Alphaproteobacteria - Rnodobacterales
Alphaproteodacteria - uncultured
Gammaproteobacteria - AT-s2-59
Gammaproteobacteria - Steroidobacterales
‘Acidimicrobiia « Actinomarinales
Alphaproteobacteria - Rhodovibrionales
z
Ss
Relative Abundance
1.004
0.7!
ry
2
3
0.254
0.004
NO6-1 NO6-4 Nii NIL-4 N12-1 N124 Ni3-t Ni3-4 NI44 Nt4-4
Phylum
Actinobac
Bacteroid
Chloroflex
Dadabact
Gemmatin
Proteobac
WNOtatVe ADUNGaNCe
0.75
0.50
0.25
0.00
NO6-1 NO6-4 N11-1 N11-4 N12-1 N12-4 N13-1 N13-4 N14-1 N14-4
Cla:
Acidimicrobiia
Alphaproteoba
Anaerolineae
Dadabacteriia
Dehalococcoid
Deltaproteoba
Gammaproteo
Gemmatimona
JG30-KF-CM
Rhodothermia
Thermoleophil
| 2019 | Bacterial diversity in deep-sea sediments under influence of asphalt seep at the São Paulo Plateau | 10.1101/753616 | [
"Queiroz Luciano Lopes",
"Gonçalves Bendia Amanda",
"Duarte Rubens Tadeu Delgado",
"das Graças Diego Assis",
"Costa da Silva Artur Luiz da",
"Nakayama Cristina Rossi",
"Sumida Paulo Yukio",
"Lima Andre O. S.",
"Nagano Yuriko",
"Fujikura Katsunori",
"Kitazato Hiroshi",
"Pellizari Vivian Helena"... | creative-commons |
Effect of victim relatedness on cannibalistic behaviour of ladybird beetle, Menochilus
1
sexmaculatus Fabricius (Coleoptera: Coccinellidae)
2
Tripti Yadav1, Omkar2, and Geetanjali Mishra3*
3
Author’s details
4
1. Tripti Yadav: Research Scholar, Ladybird Research Laboratory, Department of
5
Zoology,
University
of
Lucknow,
Lucknow-
226007,
India;
Email:
6
triptiyadav3108@gmail.com
7
2. Omkar: Professor, Ladybird Research Laboratory, Department of Zoology, University
8
of Lucknow, Lucknow- 226007, India; Email: omkar.lkouniv@gmail.com
9
3. Geetanjali Mishra*: Professor, Ladybird Research Laboratory, Department of
10
Zoology,
University
of
Lucknow,
Lucknow-
226007,
India;
Email:
11
geetanjalimishra@gmail.com
12
*Corresponding author email: geetanjalimishra@gmail.com
13
14
Effect of victim relatedness on cannibalistic behaviour of ladybird beetle, Menochilus
15
sexmaculatus Fabricius (Coleoptera: Coccinellidae)
16
Tripti Yadav1, Omkar2, and Geetanjali Mishra3*
17
Abstract
18
Cannibalism is taxonomically widespread and has a large impact on the individuals’ fitness
19
and population dynamics. Thus, identifying how the rates of cannibalism are affected by
20
different ecological cues is crucial for predicting species evolution and population dynamics.
21
In current experiment, we investigated how victim relatedness affects the cannibalistic
22
tendencies of different life stages of ladybird, Menochilus sexmaculatus, which is highly
23
cannibalistic. We provided larval instars and newly emerged adults of M. sexmaculatus with a
24
choice of sibling, half-sibling and non-sibling conspecific eggs as victim of cannibalism. First
25
victim cannibalised and latency to cannibalise were observed along with total number of
26
victims cannibalised after 24 hours. First preference of victim did not differ with life stages of
27
the cannibals though the number of victims cannibalized did increase with advancement in
28
stage. Percentage of total eggs cannibalised also varied significantly with life stage and victim
29
relatedness. First and second instars tend to cannibalise more percentage of siblings and non-
30
sibling eggs while third instars cannibalised more percentage of non-sibling eggs; fourth instars
31
and adults on the other hand cannibalised highest percentage of eggs irrespective of their
32
relatedness. Insignificant effect of victim relatedness was observed on latency to cannibalise
33
eggs, though it varied significantly with the cannibal’s life stage. Shortest latency to cannibalise
34
was recorded for first instars and longest for adults and second instars. In conclusion, kin
35
recognition and its avoidance is stage-specific, with fourth instar and newly emerged adults
36
being less discriminatory as compared to early stages owing to increased evolutionary survival
37
pressure.
38
Key words: Kin recognition, cannibalism, conspecific eggs, relatedness
39
Introduction
40
In animal taxa, where parents frequently deposit eggs in clusters in a spatially constrained area,
41
the chances of increased levels of competition between the conspecifics is high (Singh et al.,
42
2019; Zaviezo et al., 2019; Uveges et al., 2021). In such a scenario, discrimination between
43
related individuals becomes essential. The presence of kin recognition and kin discrimination
44
during intensive conspecific interactions has been reported in animals ranging from bacteria to
45
vertebrates (Wall, 2016; Henkel and Setchell, 2018; Kalamara et al., 2018; Mathiron et al.,
46
2019; Anten and Chen, 2021).
47
Organism responsiveness towards the related individuals can have a major impact on its
48
inclusive fitness (West and Gardner, 2013). Thus, in species where individuals can detect the
49
variations in relatedness (kin and non-kin), behavioural variations can be observed. Relatedness
50
is usually assessed via either phenotypic cues signalling a presence of specific shared genes or
51
genotypes, or contextual cues (Penn and Frommen, 2010; Chung et al., 2020;). Sibling
52
cannibalism has been reported across diverse taxa including invertebrates (Chiu et al., 2010;
53
Miranda et al., 2011), arthropods (Johnson et al., 2010; Modanu et al., 2014), fishes ( Liu et
54
al., 2017; Pereira et al., 2017), amphibians (Walls and Blaustein, 1995; Park et al., 2005; Dugas
55
et al., 2016), and birds (Bortolotti et al., 1991; Soler et al., 2022). In contrast, several studies
56
have reported the identification and avoidance of sibling cannibalism across taxa (Dobler and
57
Kolliker, 2010; Schutt, 2017); these organisms avoid cannibalising related but readily
58
cannibalise unrelated young ones.
59
A varied range of behavioural and life-history phenotypes appear to have evolved, especially
60
in parts where intense sibling competition occurs (Pfennig and Collins, 1993; Pfennig, 2021).
61
Certain protozoans (Rosati et al., 1988; Tollrian and Harvell, 1999), rotifers (Gilbert, 2017),
62
nematodes (Lightfoot et al., 2021), insects (Pener and Simpson, 2009), and amphibian larvae
63
(Pfennig and Collins, 1993; Pfennig et al., 1993; Pfennig et al., 1994) exist as one of two
64
structurally and behaviourally distinct morphs, i.e. cannibalistic or non-cannibalistic,
65
depending on the environmental conditions they are raised in (Levis and Ragsdale, 2022;
66
Pfennig, 2021). Since cannibalistic morphs are more likely to injure kin due to possible
67
physical proximity, inclusive fitness theory predicts that they should have more developed kin
68
recognition abilities than non-cannibalistic morphs (Pfennig, 1999, 2021). Several theories
69
such as theory of inclusive fitness (Penn and Frommen, 2010) and selfish gene (Gardner, and
70
Welch, 2011) propose natural selection should favour individuals who can recognize their kin
71
over those who cannot so that copies of individuals who can recognize their kin survive
72
expanding the gene pool encoding this behaviour (Penn and Frommen, 2010; Clune et al.,
73
2011; Mateo, 2015).
74
Kin recognition has been largely studied in eusocial insects with castes of soldiers or guards,
75
e.g. ants, bees and termites (Lize et al., 2013; Vander Meer et al., 2019; Sinotte et al., 2021).
76
Studies in desert isopods (Hemilepistus reaumuri Audouin and Savigny), paper wasps
77
(Ropalidia marginata Lepeletier), and honeybees (Apis mellifera Linnaeus) have reported that
78
they may use phenotypic cues or labels for discrimination between sibling and non-sibling
79
conspecifics. More recently, Sohail et al. (2021) studied the cannibalistic expression of larval
80
instar in green lacewing, Chrysoperla carnea Stephens (Neuroptera: Chrysopidae) towards
81
related and unrelated conspecific eggs and reported that the larvae were more cannibalistic
82
towards unrelated conspecific eggs and the rate of cannibalism increased in presence of
83
conspecifics in the vicinity (Sohail et al., 2021). Also, adult Drosophila melanogaster Meigen
84
is reported to have kin-recognition abilities based on specific cues associated with gut
85
microbiome (Lewis et al., 2014; Lize et al., 2014; Carazo et al., 2015).
86
Coccinellids lay eggs in aggregative clusters in areas with high aphid density, and thus there is
87
a risk of existence of overlapping stages in a given time and space, leading to increased
88
competition between conspecifics over shared resources (Agarwala and Dixon, 1993a; Hodek
89
et al., 2012). The egg laying females can be single or multiply mated and thus the cluster might
90
consist of a mixture of sibling, half-sibling and non-sibling eggs. However, the time frame in
91
which different stages coexists is short. It is highly likely that both larvae, as well as adults,
92
will encounter sibling, half-sibling and non-sibling eggs. In addition, multiple females lay eggs
93
in nearby location making it potentially difficult to identify between related and unrelated
94
conspecifics. In this situation, there are chances that they utilise sensory information to avoid
95
cannibalising their kin. Females of Adalia bipunctata Linnaeus (Agarwala and Dixon, 1993b)
96
and Propylea dissecta Mulsant (Pervez and Khan, 2021) are able to recognise and avoid
97
cannibalising their own eggs. In addition, larvae of Harmonia axyridis Pallas (Joseph et al.,
98
1999), A. bipunctata (Agarwala and Dixon, 1993b), P. dissecta and Coccinella transversalis
99
Fabricius (Pervez et al., 2005) are also reported to have kin recognition abilities through
100
endogenous or chemical cues.
101
Based on literary background, cannibalism of eggs with varying degrees of relatedness (sibling,
102
half-sibling, and non-sibling) by larval and adult stages was tested experimentally to determine
103
whether M. sexmaculatus recognize siblings or not. It was hypothesised that the relatedness of
104
larval and adult stages with the victim will modulate their cannibalistic tendency. Cannibals
105
will recognize and avoid cannibalising sibling eggs in order to maximise their inclusive fitness.
106
Materials and methods
107
Stock culture
108
Adults of Menochilus sexmaculatus (n=60) were collected from the local agricultural fields of
109
Lucknow, India (26°50’N, 80°54’E). The species was selected as an experimental model due
110
to its abundance in local fields, wide prey range, and high reproductive output (Omkar et al.,
111
2005). Adults were fed with ad libitum supply of cowpea aphid, Aphis craccivora Koch
112
(Hemiptera: Aphididae). The aphid colonies were established on Vigna unguiculata L. plants
113
in glasshouse (25 ± 2°C temperature, 65 ± 5% Relative Humidity). Adults were paired and
114
placed in plastic Petri dishes (hereafter, 9.0 × 2.0 cm), which were kept in Biochemical Oxygen
115
Demand incubators (Yorco Super Deluxe, YSI-440, New Delhi, India) at 25 ± 1°C, 65 ± 5%
116
R.H., 14L: 10D. Eggs laid were collected daily, and held in plastic Petri dishes until hatching,
117
which usually occurs within 2-3 days from oviposition. First instars were gently removed using
118
a fine camel-hair paintbrush and assigned individually to clean experimental Petri dishes (size
119
as above) once they began moving on or away from the remnants of their egg clutch.
120
Collection of eggs used in choice treatment
121
For generation of sibling, half sibling and non-siblings, adults were randomly selected from
122
stock culture and paired in different treatments as described below. For the production of
123
siblings, reproductively mature, virgin and unrelated males and females were selected from the
124
stock culture and paired in Petri dishes; one pair per dish. Post mating, males were removed
125
and females were allowed to lay eggs. Eggs collected from females (family I) were divided
126
into two groups, first group was used as experimental replicate and other was used as sibling
127
eggs to be provided as victims in choice treatment.
128
For generation of half siblings, same males (family I) that mated earlier with the females
129
(family I) were again mated with unmated, unrelated females (family II) collected from stock
130
culture. The eggs obtained from these females (family II) were marked as half-sibling eggs and
131
were further used in choice experiment.
132
For non-sibling eggs, unrelated males and females (family III) from different sub populations
133
of stock culture were mated and eggs collected from these females were marked as non-sibling
134
eggs. Fresh eggs were collected daily from respective females and were marked with their
135
family code and relatedness for further use in the experiment as victims.
136
Experimental setup
137
Different life stages were collected from family I, i.e. first, second, third, fourth instars and
138
adult (n=15, each life stage) that were reared individually in Petri dishes on ad libitum supply
139
of A. craccivora. At the start of the experiment, one experimental individual (any immature or
140
adult stage) was placed in the experimental Petri dish containing three equidistantly placed
141
clusters of twenty sibling, twenty half-sibling and twenty non-sibling eggs. The first victim
142
cannibalised, time taken to encounter first victim, time taken for first consumption and first
143
victim cannibalised were recorded for each life stage (first, second, third and fourth instar and
144
adult). The total amount of eggs of differently related eggs, i.e. sibling, half-sibling, and non-
145
sibling, cannibalised by each life stage was also recorded after 24 hours. For recording total
146
amount of eggs cannibalised, the number of eggs provided were stage-specific, i.e. 20 eggs for
147
first, 40 eggs for second, 60 eggs for third, 80 eggs for fourth and 100 eggs for adults of each
148
sibling, half-sibling, and non-sibling eggs. The study was replicated fifteen times for each life
149
stage, i.e. first, second, third and fourth instars and adults.
150
Statistical analysis
151
To analyse the effect of victim relatedness on cannibalistic preferences of different life stages
152
of M. sexmaculatus, victim first cannibalised by each life stage (larval instars and adults) were
153
subjected to Chi-square test using Minitab 20.3 statistical software. Data sets on encounter
154
time, latency to cannibalise and percent egg cannibalised were analysed with Shapiro-Wilk’s
155
and Levene’s tests to test for normal distribution and variance homogeneity, respectively.
156
Further, to analyse the effect of victim encountered on time of first encounter, encounter time
157
was used as response factor and life stage of the cannibal and victim encountered as fixed
158
factors in a Generalised Linear Model (GLM).
159
To analyse the effect on latency to cannibalise, the latency to first victim cannibalised was used
160
as response factor and life stage and victim cannibalised as well as their interaction as fixed
161
factors in a GLM. For percent consumption, the data on percent victim cannibalised were used
162
as response factor, and life stage and relatedness as well as their interaction as fixed factors in
163
a GLM.
164
All the analyses were conducted using the Minitab 20.3 statistical software.
165
Results
166
Chi-square analysis revealed insignificant effect of relatedness on the nature of victim first
167
cannibalised by different life stages (ϰ2=1.92, P>0.05, df=8). In both larval stages as well as
168
adults, victim first cannibalised was random (Figure 1).
169
170
Figure 1. Effect of relatedness on first victim cannibalised by different larval instars and adults
171
of M. sexmaculatus.
172
The time of first encounter was significantly different for different life stages (F=19.74,
173
P<0.05, df=4,74) but was not affected by the relatedness of the victim cannibalised (F=0.36,
174
P>0.05, df=2,74). The longest encounter time was recorded for first instars followed by third
175
instars, fourth instars, adult and second instars (Figure 2).
176
177
Figure 2. Effect of victim relatedness on encounter duration by different larval stages and
178
adults of M. sexmaculatus. Values are mean ± SE. Lowercase and uppercase letters
179
indicate comparison of mean within and between treatments respectively. Similar
180
letters indicate lack of significant difference (P value > 0.05).
181
Similarly, the time of first victim cannibalised was significantly influenced by life stages
182
(F=18.47, P<0.05, df=4,74). However, the consumption duration was not significantly affected
183
by the relatedness of the victim cannibalised (F=1.83, P>0.05, df=2,74). Shortest consumption
184
durations were recorded for first instars followed by third instars, fourth instars, second instars
185
and adults (Figure 3).
186
187
188
189
190
191
192
Figure 3. Effect of victim relatedness on consumption duration by different larval stages and
193
adults of M. sexmaculatus. Values are mean ± SE. Lowercase and uppercase letters
194
indicate comparison of mean within and between treatments respectively. Similar
195
letters indicate lack of significant difference (P value > 0.05).
196
Percent eggs cannibalised by different life stages was significantly affected by both life stage
197
(F=60.46, P<0.05, df=5,149) and relatedness (F=12.49, P<0.05, df=2,149) of the victim. In
198
addition, their interactions were also found to be significant (F=2.74, P<0.05, df=8,149).
199
Comparison of means on life stages revealed highest percent egg cannibalism by fourth instars
200
followed by adults, third instars, second instars, and first instars.
201
In addition, comparison of means on relatedness revealed that the first instars tend to
202
cannibalise more percentage of sibling and non-sibling eggs while second and third instars
203
cannibalised more percentage of non-sibling eggs. Fourth instars and adults, on the other hand,
204
cannibalised highest percentage of eggs irrespective of their relatedness (Figure 4).
205
206
Figure 4. Effect of victim relatedness on percent egg consumption by different larval stages
207
and adults of M. sexmaculatus. Values are mean ± SE. Lowercase and uppercase
208
letters indicate comparison of mean within and between treatments respectively.
209
Similar letters indicate lack of significant difference (P value > 0.05).
210
211
Discussion
212
Current study revealed that the victim relatedness had insignificant effect on cannibalism by
213
different larval instars and adults M. sexmaculatus. First encounter duration and the latency to
214
cannibalise victim were found to be insignificantly affected by victim relatedness, however,
215
both significantly varied with stage of the cannibal. Encounter duration decreased with the
216
advancing stage except for the second instars and cannibalistic latencies followed a reverse
217
trend. The victim first cannibalised by larval stages and adults were random, however, percent
218
total egg consumption increased with advancing stage.
219
Insignificant difference was observed in victim first cannibalised by different larval instars and
220
adults of M. sexmaculatus. However, significant differences in total percent number of eggs
221
(after 24 hours) cannibalised with varying degree of victim relatedness suggests the presence
222
of stage-specific cannibalistic tendencies and kin recognition mechanism in M. sexmaculatus.
223
The percentage of total eggs cannibalised increased with the advancement in stage. First and
224
second instars cannibalised a higher percentage of both sibling and non-sibling eggs while third
225
instars cannibalised a higher percentage of non-sibling eggs. Fourth instars, and the adults, on
226
the other hand, cannibalised the eggs regardless of their relatedness with the victim suggesting
227
that kin recognition changes with stage. For first instars, mobility and the ability to tolerate
228
hunger are relatively low, which might be a reason for high levels of sibling egg cannibalism
229
(Ferran and Dixon, 1993). In addition, earlier studies have reported that the larval instars and
230
adults can assess the surface chemical profile (Agarwala et al, 1998; Omkar et al., 2004). In
231
Hippodamia variegata Goeze (Xie et al., 2022) and H. axyridis (Rondoni et al., 2021), antennal
232
transcriptomes have reported the presence of odorant receptors and their chemosensory role in
233
prey recognition and searching behaviour. Studies involving the role of family-specific
234
chemical profiles in kin recognition have also been reported (Wong et al, 2014; Weiss and
235
Schneider, 2021). Thus, the first instars cannibalising the eggs first encountered either sibling
236
or non-sibling eggs might be attributed to the recognition of similar egg surface chemicals that
237
lowers the risk of consuming toxic food and increases the chances of survival by overcoming
238
the initial period of vulnerability, however, it does not confirm the presence of kin recognition
239
ability in any of the life stages except the third instars which cannibalised higher percentage of
240
non-sibling eggs. Previous investigations in P. dissecta and C. transversalis have also shown
241
stage-specific cannibalistic responses towards sibling and non-sibling larvae, where third
242
instars avoided sibling cannibalism while fourth instars indiscriminately cannibalised both
243
sibling and non-sibling eggs (Pervez et al., 2005). Joseph et al. (1999) in a study on H. axyridis
244
have reported that third instars avoided cannibalism of related victims and took longer and
245
showed higher encounter rates to cannibalise related victims than unrelated victims. The
246
highest percentage of eggs cannibalised by fourth instars may be attributed to increased
247
evolutionary survival pressure and higher energetic needs required for pupation (Khan et al.,
248
2003; Jafari, 2012; Khan and Yoldas, 2018a, b). The results suggest that possibly the nutritional
249
requirements of larval instars vary based on the developmental stage which in turn regulates
250
predatory prey consumption. In contrast, Agarwala and Dixon (1993b) in a study on A.
251
bipunctata, have reported that female and young larvae are able to recognize kin, however,
252
third instars showed no reluctance to eat second instar siblings.
253
Furthermore, the first encounters with the victim and the first cannibalistic attack on the
254
victim by each life stage were independent of the degree of relatedness, indicating that these
255
first encounters and consumptions were random events. However, both encounter duration and
256
latency to cannibalise were significantly different for different life stages of the cannibal,
257
suggesting that different life stages took different time durations before cannibalising eggs.
258
Also, the differences in encounter durations among different life stages may be attributed to
259
the extent of the mobility of the life stages and their need to procure food. Previous studies
260
have shown that larval stages are known to alter their movement patterns following feeding as
261
well as the area they transverse per unit time based on their size and age (Ferran and Dixon,
262
1993). For instance, the first instars would confine their search area to their immediate vicinity
263
since they are the most critical developmental stage and with limited mobility, they typically
264
stay close to the egg clutch. Post hatching, they first feed on their egg case and later on their
265
neighbouring unhatched sibling eggs, which provide them with the energy necessary to search
266
for food (Dixon, 2000; Hodek et al., 2012).
267
In conclusion, larval instars and adults prefer to cannibalise non-sibling eggs in M.
268
sexmaculatus. The presence of kin recognition mechanism and discrimination among sibling
269
and non-sibling eggs might play a beneficial role through increasing the inclusive fitness by
270
decreasing the cannibalistic incidences among siblings and percent egg consumption increases
271
with advancement of stage owing to the increased nutritional requirement and survival
272
pressure.
273
Acknowledgments: TY gratefully acknowledges CSIR, New Delhi, India, for Senior Research
274
Fellowship, (09/107(0405)/2019-EMR-I) dated October 20, 2020. GM is thankful to
275
Department of Higher Education, Government of Uttar Pradesh, India for providing financial
276
assistance under the Centre of Excellence programme.
277
Conflict of Interest
278
The authors declare that they have no conflict of interest.
279
Data availability statement: The datasets generated during and/or analysed during the current
280
study are available from the corresponding author on reasonable request.
281
References
282
Agarwala, B. K., & Dixon, A. F. G. (1993a). Kin recognition: egg and larval cannibalism in
283
Adalia bipunctata (Coleoptera: Coccinellidae). European Journal of Entomology, 90, 45–
284
50.
285
Agarwala, B. K., & Dixon, A. F. G. (1993b). Why do ladybirds lay eggs in clusters?. Functional
286
Ecology, 541-548.
287
Agarwala, B. K., Bhattacharya, S., & Bardhanroy, P. (1998). Who eats whose eggs? intra-
288
versus inter-specific interactions in starving ladybird beetles predaceous on aphids.
289
Ethology Ecology and Evolution, 10(4), 361–368.
290
Anten, N. P. R., & Chen, B. J. W. (2021). Detect thy family: Mechanisms, ecology and
291
agricultural aspects of kin recognition in plants. Plant Cell and Environment, 44(4), 1059–
292
1071.
293
Bortolotti, G. R., Wiebe, K. L., & Iko, W. M. (1991). Cannibalism of nestling American
294
kestrels by their parents and siblings. Canadian Journal of Zoology, 69(6), 1447–1453.
295
Bourke, A. F. G. (2014). Hamilton’s rule and the causes of social evolution. Philosophical
296
Transactions of the Royal Society B: Biological Sciences. 369(1642), 20130362.
297
Carazo, P., Perry, J. C., Johnson, F., Pizzari, T., & Wigby, S. (2015). Related male Drosophila
298
melanogaster reared together as larvae fight less and sire longer lived daughters. Ecology
299
and Evolution, 5(14), 2787–2797.
300
Chiu, J. M. Y., Shin, P. K. S., Wong, K. P., & Cheung, S. G. (2010). Sibling cannibalism in
301
juveniles of the marine gastropod Nassarius festivus (Powys, 1835). Malacologia, 52(1),
302
157–161.
303
Chung, M., Wang, M. Y., Huang, Z., & Okuyama, T. (2020). Diverse sensory cues for
304
individual recognition. Development Growth and Differentiation, 62(9), 507–515.
305
Clune, J., Goldsby, H. J., Ofria, C., & Pennock, R. T. (2011). Selective pressures for accurate
306
altruism targeting: Evidence from digital evolution for difficult-to-test aspects of inclusive
307
fitness theory. Proceedings of the Royal Society B: Biological Sciences, 278(1706), 666–
308
674.
309
Dixon, A. F. G. (2000). Insect predator–prey dynamics, ladybird beetles and biological
310
control. Cambridge University Press, Cambridge, UK.
311
Dobler, R., & Kölliker, M. (2010). Kin-selected siblicide and cannibalism in the European
312
earwig. Behavioral Ecology, 21(2), 257–263.
313
Dugas, M. B., McCormack, L., Gadau, A., & Martin, R. A. (2016). Choosy cannibals
314
preferentially consume siblings with relatively low fitness prospects. American
315
Naturalist.
316
Ferran, A., & Dixon, A. F. G. (1993). Foraging behaviour of ladybird larvae (Coleoptera:
317
Coccinellidae). European Journal of Entomology, 90(4), 383–402.
318
Gardner, A., & Welch, J.J. (2011). A formal theory of the selfish gene. Journal of Evolutionary
319
Biology, 24(8), 1801-13.
320
Gilbert, J. J. (2017). Non-genetic polymorphisms in rotifers: Environmental and endogenous
321
controls, development, and features for predictable or unpredictable environments.
322
Biological Reviews, 92(2), 964–992.
323
Henkel, S., & Setchell, J. M. (2018). Group and kin recognition via olfactory cues in
324
chimpanzees (Pan troglodytes). Proceedings of the Royal Society B: Biological Sciences,
325
285(1889).
326
Hodek, I, Emden, H. F., & Honek, I. (2012). Ecology and behaviour of the ladybird beetles
327
(Coccinellidae). Oxford, United Kingdom: Wiley-Blackwell.
328
Jafari, R. (2012). Cannibalism in Hippodamia variegata Goeze (Coleoptera: Coccinellidae)
329
under laboratory conditions. International Symposium: Current Trends in Plant
330
Protection - Proceedings, 544–549.
331
Johnson, J. C., Kitchen, K., & Andrade, M. C. B. (2010). Family affects sibling cannibalism in
332
the black widow spider, Latrodectus hesperus. Ethology, 116(8), 770–777.
333
Joseph, S. B., Snyder, W. E., & Moore, A. J. (1999). Cannibalizing Harmonia axyridis
334
(Coleoptera: Coccinellidae) larvae use endogenous cues to avoid eating relatives. Journal
335
of Evolutionary Biology, 12(4), 792–797.
336
Kalamara, M., Spacapan, M., Mandic-Mulec, I., & Stanley-Wall, N. R. (2018). Social
337
behaviours by Bacillus subtilis: quorum sensing, kin discrimination and beyond.
338
Molecular Microbiology, 110(6), 863–878.
339
Khan, M. H., & Yoldaş, Z. (2018). Investigations on the cannibalistic behavior of ladybird
340
beetle Coccinella septempunctata L. (Coleoptera: Coccinellidae) under laboratory
341
conditions. Turkish Journal of Zoology, 42(4), 432–438.
342
Khan, M. R., Khan, M. R., & Hussein, M. Y. (2003). Cannibalism and Interspecific Predation
343
in Ladybird Beetle Coccinella septempunctata (L.) (Coleoptera: Coccinellidae) in
344
Laboratory. Pakistan Journal of Biological Sciences, 6(24), 2013–2016.
345
Levis, N. A., & Ragsdale, E. J. (2022). Linking Molecular Mechanisms and Evolutionary
346
Consequences of Resource Polyphenism. Frontiers in Integrative Neuroscience, 16.
347
Lewis, Z., Heys, C., Prescott, M., & Lizé, A. (2014). You are what you eat: Gut microbiota
348
determines kin recognition in Drosophila. Gut Microbes, 5(4), 541–543.
349
Lightfoot, J. W., Dardiry, M., Kalirad, A., Giaimo, S., Eberhardt, G., Witte, H., & Sommer, R.
350
J. (2021). Sex or cannibalism: Polyphenism and kin recognition control social action
351
strategies in nematodes. Science Advances, 7(35).
352
Liu, X., Xia, J., Pang, H., & Yue, G. (2017). Who eats whom, when and why? Juvenile
353
cannibalism in fish Asian seabass. Aquaculture and Fisheries, 2(1), 1–9.
354
Lizé, A., McKay, R., & Lewis, Z. (2013). Gut microbiota and kin recognition. Trends in
355
Ecology and Evolution, 28(6), 325–326.
356
Lizé, A., McKay, R., & Lewis, Z. (2014). Kin recognition in Drosophila: The importance of
357
ecology and gut microbiota. ISME Journal, 8(2), 469–477.
358
Mateo, J. M. (2015). Perspectives: Hamilton’s legacy: Mechanisms of kin recognition in
359
humans. Ethology, 121(5), 419–427.
360
Mathiron, A. G. E., Pottier, P., & Goubault, M. (2019). Keep calm, we know each other: kin
361
recognition affects aggressiveness and conflict resolution in a solitary parasitoid. Animal
362
Behaviour, 151, 103–111.
363
Miranda, R. M., Espinoza, V., Dörner, J., Farías, A., & Uriarte, I. (2011). Sibling cannibalism
364
on the small octopus Robsonella fontaniana (d’Orbigny, 1834) paralarvae. Marine
365
Biology Research, 7(8), 746–756.
366
Modanu, M., Li, L. D. X., Said, H., Rathitharan, N., & Andrade, M. C. B. (2014). Sibling
367
cannibalism in a web-building spider: Effects of density and shared environment.
368
Behavioural Processes, 106, 12–16.
369
Omkar, & Bind, R. B. (2004). Prey quality dependent growth, development, and reproduction
370
of a biocontrol agent, Cheilomenes sexmaculata (Fabricius) (Coleoptera: Coccinellidae).
371
Biocontrol Science and Technology, 14, 665–673.
372
Omkar, Pervez, A., Gupta, A. K., Pervez, O. A., & Gupta, A. K. (2004). Role of surface
373
chemicals in egg cannibalism and intraguild predation by neonates of two aphidophagous
374
ladybirds, Propylea dissecta and Coccinella transversalis. Journal of Applied
375
Entomology, 128(9–10), 691–695.
376
Park, S., Jeong, J., & Park, D. (2005). Cannibalism in the Korean salamander ( Hynobius
377
leechii : Hynobiidae, caudata, amphibia) larvae . Integrative Biosciences, 9(1), 13–18.
378
Pener, M. P., & Simpson, S. J. (2009). Locust Phase Polyphenism: An Update. Advances in
379
Insect Physiology, 36, 1–272.
380
Penn, D. J., & Frommen, J. G. (2010). Kin recognition: An overview of conceptual issues,
381
mechanisms and evolutionary theory. Animal Behaviour: Evolution and Mechanisms, 55–
382
85.
383
Pereira, L. S., Agostinho, A. A., & Winemiller, K. O. (2017). Revisiting cannibalism in fishes.
384
Reviews in Fish Biology and Fisheries. Springer International Publishing.
385
Pervez, A., & Khan, M. (2021). Kin recognition by the adults of a biological control agent,
386
Propylea dissecta (Coleoptera: Coccinellidae). Journal of Biological Control, 34(3), 227-
387
230.
388
Pervez, Ahmad, Gupta, A. K., & Omkar. (2005). Kin recognition and avoidance of kin
389
cannibalism by the larvae of co-occurring ladybirds: A laboratory study. European
390
Journal of Entomology, 102(3), 513–518.
391
Pfennig, D. W. (2021). Phenotypic Plasticity & Evolution Causes, Consequences,
392
Controversies. Taylor & Francis.
393
Pfennig, D. W. (1999). Cannibalistic tadpoles that pose the greatest threat to kin are most likely
394
to discriminate kin. Proceedings of the Royal Society B: Biological Sciences, 266(1414),
395
57–61.
396
Pfennig, D. W., & Collins, J. P. (1993). Kinship affects morphogenesis in cannibalistic
397
salamanders. Nature, 362(6423), 836–838.
398
Pfennig, D. W., Reeve, H. K., & Sherman, P. W. (1993). Kin recognition and cannibalism in
399
spadefoot toad tadpoles. Animal Behaviour, 46(1), 87–94.
400
Pfennig, David W., Sherman, P. W., & Collins, J. P. (1994). Kin recognition and cannibalism
401
in polyphenic salamanders. Behavioral Ecology, 5(2), 225–232.
402
Rondoni, G., Roman, A., Meslin, C., Montagné, N., Conti, E., & Jacquin-Joly, E. (2021).
403
Antennal transcriptome analysis and identification of candidate chemosensory genes of
404
the harlequin ladybird beetle, Harmonia axyridis (Pallas) (coleoptera: Coccinellidae).
405
Insects, 12(3), 1–16.
406
Rosati, G., Giari, A., & Ricci, N. (1988). Oxytricha bifaria (ciliata, hypotrichida): General
407
morphology and ultrastructure of normal cells and giants. European Journal of
408
Protistology, 23(4), 343–349.
409
Schutt, B. (2017). Eat me : a natural and unnatural history of cannibalism.
410
Singh, S., Mishra, G., & Omkar (2019). Oviposition in aphidophagous ladybirds: effect of prey
411
availability and conspecific egg presence. International Journal of Tropical Insect
412
Science, 39(2), 107–114.
413
Sinotte, V. M., Conlon, B. H., Seibel, E., Schwitalla, J. W., de Beer, Z. W., Poulsen, M., &
414
Bos, N. (2021). Female-biased sex allocation and lack of inbreeding avoidance in
415
Cubitermes termites. Ecology and Evolution, 11(10), 5598–5605.
416
Sohail, M., Soomro, Q. A., Asif, M. U., Rauf, I., & Muhammad, R. (2021). Conspecific
417
neighbors and kinship influence egg cannibalism in the green lacewing, Chrysoperla
418
carnea (Stephens). Egyptian Journal of Biological Pest Control, 31(1).
419
Soler, J. J., Martín-vivaldi, M., Nuhlíčková, S., Ruiz-castellano, C., Mazorra-alonso, M.,
420
Martínez-renau, E., Svetlík, J., Eckenfellner, M., & Hoi, H. (2022). Avian sibling
421
cannibalism : Hoopoe mothers regularly use their last hatched nestlings to feed older
422
siblings. Zoological Research, 43(2), 265–274.
423
Tollrian, R. & Harvell, C. D., (1990). The evolution of inducible defence. (R. Tollrian & C. D.
424
Harvell, Eds.), Princeton University Press (Vol. 100). Princeton University Press.
425
Uveges, B., Basson, A. C., Móricz, Á. M., Bókony, V., & Hettyey, A. (2021). Chemical
426
defence effective against multiple enemies: Does the response to conspecifics alleviate
427
the response to predators? Functional Ecology, 35(10), 2294–2304.
428
Vander Meer, R. K., Breed, M. D., Winston, M. L., & Espelie, K. E. (2019). Pheromone
429
communication in social insects: ants, wasps, bees, and termites. CRC Press (Vol. 35).
430
Wall, D. (2016). Kin Recognition in Bacteria. Annual Review of Microbiology, 70, 143–160.
431
Walls, S. C., & Blaustein, A. R. (1995). Larval marbled salamanders, Ambystoma opacum, eat
432
their kin. Animal Behaviour, 50(2), 537–545.
433
Weiss, K., & Schneider, J. M. (2021). Family-specific chemical profiles provide potential kin
434
recognition cues in the sexually cannibalistic spider Argiope bruennichi. Biology Letters,
435
17(8), 20210260.
436
West, S. A., & Gardner, A. (2013). Adaptation and Inclusive Fitness. Current Biology, 23(13),
437
R577–R584.
438
Wong, J. W. Y., Meunier, J., Lucas, C., & Kölliker, M. (2014). Paternal signature in kin
439
recognition cues of a social insect: concealed in juveniles, revealed in adults. Proceedings
440
of the Royal Society B: Biological Sciences, 281(1793), 20141236.
441
Xie, J., Liu, T., Yi, C., Liu, X., Tang, H., Sun, Y., & Zhang, Y. (2022). Antenna-Biased Odorant
442
Receptor HvarOR25 in Hippodamia variegata Tuned to Allelochemicals from Hosts and
443
Habitat Involved in Perceiving Preys. Journal of Agricultural and Food Chemistry. 70(4),
444
1090-1100.
445
Zaviezo, T., Soares, A. O., & Grez, A. A. (2019). Interspecific exploitative competition
446
between Harmonia axyridis and other coccinellids is stronger than intraspecific
447
competition. Biological Control, 131, 62–68.
448
| 2022 | Effect of victim relatedness on cannibalistic behaviour of ladybird beetle, Fabricius (Coleoptera: Coccinellidae) | 10.1101/2022.09.30.510267 | null | null |
Graphsite: Ligand-binding site classification using Deep Graph Neural
Network
Wentao Shi1, Manali Singha2, Limeng Pu3, J. Ramanujam1,3, Michal Brylinski2,3*
1 Department of Electrical and Computer Engineering, Louisiana State University, Baton
Rouge, Louisiana, United States of America
2 Department of Biological Sciences, Louisiana State University, Baton Rouge, Louisiana,
United States of America
3 Center for Computation and Technology, Louisiana State University, Baton Rouge,
Louisiana, United States of America
* michal@brylinski.org
Abstract
Binding sites are concave surfaces on proteins that bind to small molecules called ligands. Types of
molecules that bind to the protein determine its biological function. Meanwhile, the binding process
between small molecules and the protein is also crucial to various biological functionalities. Therefore,
identifying and classifying such binding sites would enormously contribute to biomedical applications
such as drug repurposing. Deep learning is a modern artificial intelligence technology. It utilizes deep
neural networks to handle complex tasks such as image classification and language translation. Previous
work has proven the capability of deep learning models handle binding sites wherein the binding sites are
represented as pixels or voxels. Graph neural networks (GNNs) are deep learning models that operate on
graphs. GNNs are promising for handling binding sites related tasks - provided there is an adequate
graph representation to model the binding sties. In this communication, we describe a GNN-based
computational method, GraphSite, that utilizes a novel graph representation of ligand-binding sites. A
state-of-the-art GNN model is trained to capture the intrinsic characteristics of these binding sites and
classify them. Our model generalizes well to unseen data and achieves test accuracy of 81.28% on
classifying 14 binding site classes.
1/13
1
Introduction
1
Interactions of proteins with other molecules like peptides, neurotransmitters, nucleic acids, hormones,
2
metabolites, and other proteins have vital part in understanding the biological functions. Proteins are
3
basic building blocks and responsible for carrying out all biological functions in cellular environment. So,
4
identification of interaction between proteins and small molecules is crucial to understand how proteins
5
regulate different functions in a living cell [2]. The ligand binding site (also referred to as pocket) is a
6
groove or cavity in a protein where the small molecules or ligands can bind through interactions with
7
amino acids at that site [27]. Identification of off-targets binding can help scientists to repurpose the
8
existing drugs to cure some rare orphan diseases for which we do not have functional drugs available. So,
9
binding site prediction approaches can be used to find cures for rare diseases [10]. Therefore, binding site
10
prediction in structural biology is vitally important in the field of drug discovery and it can help predict
11
the novel drug targets. There are several available algorithms which can identify the ligand binding sites
12
on target protein structures such as eFindSite [7], Fpocket [20], and FTSite [25] etc. Besides that, the
13
ligand binding on protein depends on numerous factors of binding site. So, there are various methods
14
which account these factors such as conformational dynamics [1], druggability [16] and amino acid
15
compositions [31] of binding sites on target proteins. However, all these methods do not account for the
16
classification of binding sites depending on types of ligands.
17
Deep learning is an emerging machine learning technique. Deep learning-based models employ various
18
styles of multi-layer artificial neural networks to learn from data and make predictions. Deep learning
19
has achieved significant progress in computer vision applications such as object detection [13], face
20
recognition [28], and human pose estimation [35]. One of the keys to the success of those applications is
21
the convolutional neural network (CNN), which can learn hierarchical latent features from Euclidean
22
data (2D- and 3D images) by utilizing local trainable filters [3]. Such methodologies in computer vision
23
have inspired new works in structural biology in recent years. DeepDrug3D [26] achieves state-of-the-art
24
binding site classification performance by representing the binding sites as 3D images and deploying a
25
3D-CNN. DeeplyTough [30], which uses similar pocket representation as DeepDrug3D, implements
26
pocket-matching. DeepSite [15] is a binding site predictor that also forms similar 3D representations of
27
pockets by computing atomic-based pharmacophoric properties for each voxel. Other than 3D
28
representations, BionoiNet [29] projects pockets to 2D images that encode chemical properties, and a
29
2D-CNN is trained to perform classification.
30
Graph neural network (GNN) is another category of deep learning model that operates on graphs
31
which are non-Euclidean data. Over the recent years, GNNs have demonstrated encouraging
32
performance on applications such as text classification [12,18] and traffic prediction [22]. As for the field
33
of chemistry and biochemistry, GNNs are proven to be promising for a variety of applications including
34
predicting quantum property of an organic molecule [9], generating molecular fingerprints [5], predicting
35
protein interface [8], and predicting drug-target interaction [23]. These works are based on the idea that
36
molecular structures can be naturally interpreted as graphs. A typical example is the Lewis structure of
37
molecules where the atoms are treated as nodes and the chemical bonds are the undirected edges that
38
connect nodes.
39
In this communication, we describe a framework based on GNN to classify ligand-binding sites. A
40
novel graph representation of binding sites is developed and a GNN classifier is then trained to classify a
41
pocket dataset of 14 classes. Comparing with the methods that convert pockets to Euclidean data, the
42
process of converting to graphs is fast and lossless. So, the graphs can be generated on-the-fly and the
43
users only need to provide standard text files as input. Our implementation achieves state-of-the-art
44
performance and the followed case studies show that our model learns the underlying pattern of different
45
kinds of binding pockets.
46
2/13
A
B
W20
R174
H55
E14
Figure 1. Molecular structure and graph representation of a binding site. (A) The residues that interact
with the ligand of pocket 5x06F00. (B) The graph representation of 4 residues in A. Any atom pair that
has distance less than or equal to 4.5 ˚A is connected
2
Materials and Methods
47
2.1
Graph representation of biding sites
48
The pockets are transformed to graphs as the input of the classifier. The nodes of the graph are the
49
atoms, and an undirected edge is formed between two atoms if the distance between them is less than or
50
equal to 4.5 ˚A. We crafted 11 node features, 7 of them are spatial features, and the other 4 features are
51
chemical features. The spatial features are used to define the shape of the binding pocket, which are the
52
Cartesian coordinates (x, y, z) of the atoms, the spherical coordinates (r, theta, gamma) of the atoms,
53
and the solvent accessible surface area (SASA). We adopt the chemical features described in Bionoi [6],
54
which are charge, hydrophobicity, binding probability and sequence entropy. Fig 1 illustrates part of the
55
graph representation of a binding pocket. As can be seen in Fig 1B, each atom is connected to all the
56
neighboring atoms withing the radius of 4.5 ˚A. To distinguish the chemical bond-edges from the others,
57
we set the number of chemical bonds as the edge attribute. The edges with no chemical bonds have 0 as
58
their attributes and the edges on aromatic rings have 1.5 as their attributes.
59
2.2
Graph neural network
60
With the graph representation of the binding pockets, the binding site classification problem essentially
61
becomes a graph classification problem. A general graph classification framework that uses GNN can be
62
divided into three stages: message passing, graph readout, and classification. In addition to these three
63
stages, our model utilizes jumping knowledge connections [37] to let the model select information for
64
each node from different layers. Fig 2 illustrates the overall architecture of the GraphSite classifier. As
65
can be seen in Fig 2, the main body of the classifier is an embedding network which contains the
66
message passing layers, the jumping knowledge connections, and a global pooling layer which performs
67
the graph readout. The node features of input graph are updated by the message passing layers. The
68
outputs of all layers are processed by a max pooling layer that performs a feature-wise max pooling; the
69
intuition behind this it to let the model to learn the proper number of layers for each individual node;
70
this technology is known as the jumping knowledge [37]. The max pooling layer is followed by a global
71
3/13
1
2
e12
a12
…
Max
pool
Jumping knowledge
Global
pool
Fully-connected
layers
Classification
results
Embedding network
…
hω
A
B
C
D
E
Figure 2. The architecture of GraphSite classifier. (A) The input graph of a binding site. (B) A neural
network that computes the weight of message using the edge attribute as input. (C) The message passing
layers with Jumping Knowledge connections. (D) The global pooling layer which is the Set2Set model.
(E) The fully connected layers that generate classification results.
pooling layer, which reduces the dimension of node feature from n × d to d where n is the number of
72
nodes and d is dimension of the node feature. The output of the global pooling layer is a fixed-size
73
vector, and it is followed by fully connected layers to generate the final classification results.
74
2.2.1
Message passing
75
The message passing layers of GNNs update the node features by propagating information along edges.
76
From the perspective of each node, the information of its neighborhood is aggregated, and the updated
77
node features can reveal informative local patterns. As described in [41], most of the message passing
78
layers fall into the general form of neighborhood aggregation:
79
x(k)
i
= λ
�
x(k−1)
i
, aggrj∈N (i)ϕ
�
x(k−1)
i
,
x(k−1)
j
,
eij
��
,
(1)
where ϕ is a differentiable function that generates the message, aggr is a permutation-invariant function
80
(such as sum or max) that aggregates the messages, and λ is a differentiable function such as a
81
multi-layer perceptron (MLP). x(k)
i
is the output node feature of node i of layer k, x(k)
j
represents its
82
neighbor nodes, and eij is the edge attribute. To exploit both node features and edge feature of the
83
binding site graph, we develop a message passing layer which also falls into the general form described by
84
Equation 1, which is called neural weighted message (NWM):
85
x(k)
i
= hθ
(1 + ϵ) · x(k−1)
i
+
�
j∈N (i)
hω (eij) · x(k−1)
j
,
(2)
where hω is an MLP that takes the edge attribute as input and outputs a scaler as the weight of the
86
message, which is simply node feature j; ϵ is learnable scalar; hθ is another MLP that updates the
87
aggregated information. Note that the edge attributes are not updated during training, and they are the
88
same for all the layers. Fig 2B demonstrates an example of NWM: hω takes the edge attribute e12 as
89
input, generating a12 as the weight of message propagating from node 2 to node 1.
90
The NWM message passing rule can be regarded as an extension of the graph isomorphism network
91
(GIN) [36]. GIN is an expressive message passing model that is as powerful as the Weisfeiler-Lehman test
92
in distinguishing graph structures; we replace its sum aggregator with sum of weighted messages where
93
the weights are generated by a neural network hω. From another perspective, the NWM layer belongs to
94
the Message Passing Neural Network (MPNN) family [9]. The gated graph neural network (GG-NN) is
95
an MPNN family member and its message is formed by Aeijx(k)
j , where Aeij is a square transformation
96
4/13
matrix generated by an MLP which takes the edge attribute eij as input; if we put a restriction on the
97
matrix Aeij, such that it is a diagonal matrix and all elements on the diagonal are equal, the GG-NN
98
module becomes NWM. In fact, the neural message of GG-NN was one of our first design choices. In our
99
experiments, we found that regularizing GG-NN to NWM could help mitigate overfitting and NWM is
100
more computationally efficient. Therefore, we take NMM as our final design choice.
101
Finally, inspired by the idea that multiple aggregators can improve the expressiveness of GNNs [4],
102
we extend a single-channel NWM layer described by Equation 2 to a multi-channel NWM layer by
103
concatenating the outputs of multiple aggregators:
104
x(k)
i
= hθ
concatc∈Channels
(1 + ϵc) · x(k−1)
i
+
�
j∈N (i)
hωc (eij) · x(k−1)
j
,
(3)
where each pair of ϵc and hωc represents an aggregator learned as channel c, and C denotes the set of
105
channels. The aggregated node features are concatenated in their last dimension such that the
106
concatenated node features have the shape of n by d × |C| where d is the dimension of node feature.
107
Accordingly, the update neural network hθ now also acts as a reduction function that reduces the size of
108
node feature from d × |C| to d. Intuitively, the concatenation of multiple aggregators is analogous to
109
using multiple filters in CNN: each aggregator corresponds to a filter, and the concatenated output
110
corresponds to the output feature maps in a convolution layer in CNN.
111
2.2.2
Graph readout
112
The graph readout function reduces the size of graph to one node. This function should regard the
113
features of the nodes as a set, because there is no order among the nodes. i.e., the graph readout
114
function should be permutation invariant. The Set2Set [34] model is a global pooling function to perform
115
graph readout. Set2Set can generate fixed-sized embeddings for sets with various sizes, and it bears the
116
property of permutation invariance. It computes the global representation of the set by leveraging the
117
attention mechanism. Basically, a Long short-term memory (LSTM) [14] neural network recurrently
118
updates a global hidden state of the input set; during the recurrent process, the global hidden state is
119
used to compute the attention associated with each element in the set, and these attentions are in turn
120
used to update the global hidden state. After several such steps, the global graph representation is
121
formed by concatenating the global hidden state generated by the LSTM and the weighted sum of the
122
elements in the set.
123
2.2.3
Loss function
124
Instead of the cross-entropy loss, the focal loss [24] is used instead. As will be described in later section,
125
the dataset has imbalanced classes. Some classes such as ATP have much more data points than others.
126
Therefore, most of the data in a mini batch will come from these major classes and the cross-entropy loss
127
will be dominated by them. To mitigate this problem, the focal loss adds a damping factor (1 − pt)γ to
128
the cross-entropy loss:
129
FL (pt) = − (1 − pt)γ log (pt),
(4)
where pt is the predicted probability generated by the Softmax, and γ ≥ 0 is a tunable hyper-parameter.
130
With this damping factor, the dominating confident predictions with high probabilities will be suppressed
131
and the predictions with low probabilities will have higher weights. As a result, the dominated minority
132
classes with low probabilities will have higher weights, and the problem of imbalanced classes is improved.
133
2.3
Dataset
134
The dataset is generated by clustering the pockets according to their Tanimoto coefficients of the ligands,
135
because similar ligands bind to similar pockets. Note that identical pockets are removed from the
136
dataset. During our experiments, we found that some of the pocket clusters generated by this algorithm
137
5/13
Class
Label
0
ATP
1
Heme
2
Carbohydrate
3
Benzene ring
4
Chlorophyll
5
Lipid
6
Essential amino/citric/tartaric acid
7
S-adenosyl-L-homocysteine
8
CoenzymeA
9
Pyridoxal phosphate
10
Benzoic acid
11
Flavin mononucleotide
12
Morpholine ring
13
Phosphate
Table 1. The 14 labels of binding sites in the dataset.
are highly similar. We manually identified the type of ligands that bind to each class and found that due
138
to the large Tanimoto distance threshold in clustering, pockets from the same family are divided into
139
different clusters. For example, as illustrated in Fig 4, cluster 0 and cluster 9 are ATP-like pockets, and
140
cluster 3 and cluster 8 are both glucopyranose-related pockets. As a result, 30 largest clusters are
141
selected, and they are merged into 14 classes. The labels of the 14 classes are shown in Table 1.
142
3
Results and discussions
143
In this section, we first discuss the classification performance of GraphSite classifier along with the
144
baseline methods; then some interesting cases from the misclassified binding pockets are selected as case
145
studies. Finally, we test our model on unseen data which are uploaded to PDB after the curation of our
146
dataset.
147
3.1
Classification performance
148
Two GNN-based methods are evaluated: GraphSite and GIN. GIN uses a sum aggregator, so the edge
149
attributes are ignored. The purpose of having GIN as a baseline is to demonstrate the improvement of
150
NWM which utilizes edge attributes. The configurations of GraphSite and GIN are identical except the
151
architecture of GNN layers. Both models are trained with the Adam [17] optimizer for 200 epochs and
152
identical learning rate schedulers are used to half the learning rate at plateau. 25 experiments are
153
conducted for each model. In each experiment, each class is randomly divided into a training set and a
154
testing set with different random seeds. After training, the medium accuracies among the 25 experiments
155
on test set are used to evaluate the classification performance. In addition, docking and pocket matching
156
are also tested on the same classification task. We select SMINA [19], which is based on Auto-dock
157
Vina [32] as the docking tool. As for pocket matching, G-LoSA [21] is selected. Since there is no training
158
required for docking and pocket matching, the accuracies over the entire dataset are reported. For
159
docking, a label ligand is chosen manually for each class. For each prediction, the docking score of the
160
pocket is evaluated against all 14 label ligands, and the predicted class is the ligand with best docking
161
score. Pocket matching is conducted in a similar way: a label pocket is chosen for each class, and the
162
predicted class is the label pocket that has best matching score with the pocket to predict. Table 2
163
shows the classification performance. As shown in Table 2, GraphSite achieves the best overall
164
classification accuracy of 81.28%, along with a weighted F1-score of 81.66%. The accuracy is of GIN is
165
75.09%, and its weighted F1-score is 74.35%. The accuracy gain of 6.59% comes from replacing the GNN
166
layers of GIN into multi-channel NWM layers. On the other hand, docking and pocket matching are not
167
6/13
Model
Accuracy
Weighted
precision
Weighted
recall
Weighted
F1-score
Graphsite classifier
81.28%
82.33%
81.28%
81.66%
GIN classifier
75.09%
74.26%
75.09%
74.35%
SMINA
16.71%
43.45%
16.71%
16.10%
G-LoSA
14.76%
34.41%
14.76%
15.89%
Table 2. Classification performance.
working. The reasons can be multifold. First, using one fixed ligand/pocket for each class will decrease
168
the classification performance because they are not necessarily the “golden answer” for each particular
169
pocket. Second, the amount of computation required for docking and pocket matching makes it
170
impractical to run these algorithms exhaustively to maximize the classification accuracy.
171
Figure 3. Confusion matrix of the classification result of GraphSite on the test set
Fig 3 illustrates the confusion matrix on the test set of our model. As can be seen in Fig 3, each
172
7/13
number on the diagonal is a recall of a class; most of the classes are classified well except class 12 and
173
class 13. Class 12 contains morpholine rings, 17% of morpholine rings are misclassified as ATP, and 21%
174
of morpholine rings are misclassified as carbohydrates. Class 13 contains phosphate pockets, 26% of
175
which are misclassified as essential amino acids. The first reason for this is that the support of these two
176
classes in the dataset is low: only 1.77% binding pockets are morpholine rings and only 1.61% binding
177
pockets are phosphate. During training, more gradients will be generated for the majority classes and
178
the model will learn more from the majority classes; applying the Focal Loss only mitigate this problem
179
but cannot fix it completely. The second reason is that, in some of the cases, the binding moiety of the
180
ligand is similar to other types of ligands. For example, the binding moiety of some morpholine rings are
181
highly similar to ATP and carbohydrates. Therefore, the model is in fact making correct predictions
182
about the binding pockets for these cases.
183
4
Future work
184
Embedding
network
Embedding
network
Contrastive
loss
Graph embeddings
Figure 4. The Siamese-GraphSite architecture. This architecture takes a pair a of graph data as input
and it is optimized according to the contrastive loss such that graphs come from the same class are close
to each other and graphs from different classes are pushed away from each other.
The performance of Graphsite classifier indictates that the features of ligand-binding pockets are
185
extracted effectively from their graph representations. So, it is possible to extend the settings in this
186
project into other deep learning applications, such as metric learning and generative modeling. In the
187
next chapter, we describe a generative model based on Graphsite for drug discovery. Here, we explore a
188
metric learning model with a Siamese architecture [11] based on Graphsite. After training, the Siamese
189
network can generate embeddings of binding pockets for visualization and other machine learning
190
applications. As can be seen in Fig 4, the embedding network described previously takes a pair of graphs
191
as input and generate two graph embeddings; these embeddings are input of the contrastive loss [11]:
192
L (W, y, x1, x2) = 1
2 (1 − y) (dW )2 + 1
2 (y) (max (0, m − dW ))2 ,
(5)
where y is the label of a graph pair that 0 means a similar pair and 1 means a dissimilar pair; x1 and x2
193
are the input graph pair, W parameterizes the embedding network, dW is the Euclidean distance
194
between the graph embeddings, and m > 0 is a margin such that a pair contributes to the loss only if
195
their distance is within this margin. Intuitively, the contrastive loss is trying to train a model such that
196
8/13
Figure 5. t-SNE visualization of embeddings of selected clusters generated by the Siamese-GraphSite
model.
the embeddings from the same class are close to each other in the Euclidean space, and far away from
197
each other if they belong to different classes. Since the model is optimized to manipulate the embeddings
198
in the Euclidean space, the embeddings are ideal for distance-based applications such as t-SNE [33]
199
visualization and k-nearest neighbors. Figure 5 shows the t-SNE visulization of 8 classes from the
200
dataset. As can be seen, similar pockets are clustered together, and dissimilar pockets are separated
201
away from each other, which indicates that the graph Siamese model has learned effective embeddings
202
for the binding pockets. However, as the number of classes increases, the performance of the model
203
decreases significantly in our experiment. We list improving the performance of this metric learning
204
model as one of the future works of Graphsite.
205
5
Conclusion
206
In this communication, we describe GraphSite, a method to classify ligand-binding sites by modeling
207
ligand-binding sites as graphs and utilizing a GNN as the classifier. The trained model is able to capture
208
informative features of binding pockets, yielding state-of-the-art classification performance. The case
209
studies show that GraphSite successfully classified the binding sites independently of their ligands. Our
210
model is able to make meaningful prediction despite the noise in the dataset caused by the discrepancy
211
9/13
between the ligand and its binding moiety. There are several potential ways to improve or extend
212
GraphSite. First, compiling larger datasets with more classes will help training a more power model.
213
Second, exploring more meaningful node features of the binding site graph may also improve the
214
classification performance. Third, GraphSite can be extended to other deep learning-based applications
215
that involve binding sites. For example, it is possible to train a graph autoencoder to generate latent
216
embeddings of binding sites. Another potential application is to build a model to predict drug-target
217
interactions where the GNN layers of GraphSite can be used as the feature extractor of binding sites.
218
6
Supporting Information
219
• Graphsite is open-sourced and available at https://github.com/shiwentao00/Graphsite.
220
• The classifier implementation is open-sourced and available at
221
https://github.com/shiwentao00/Graphsite-classifier.
222
10/13
References
1. M. Araki, H. Iwata, B. Ma, A. Fujita, K. Terayama, Y. Sagae, F. Ono, K. Tsuda, N. Kamiya, and
Y. Okuno. Improving the accuracy of protein-ligand binding mode prediction using a molecular
dynamics-based pocket generation approach. Journal of computational chemistry,
39(32):2679–2689, 2018.
2. T. L. Blundell. Structure-based drug design. Nature, 384(6604 Suppl):23–26, 1996.
3. M. M. Bronstein, J. Bruna, Y. LeCun, A. Szlam, and P. Vandergheynst. Geometric deep learning:
going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017.
4. G. Corso, L. Cavalleri, D. Beaini, P. Li`o, and P. Veliˇckovi´c. Principal neighbourhood aggregation
for graph nets. arXiv preprint arXiv:2004.05718, 2020.
5. D. Duvenaud, D. Maclaurin, J. Aguilera-Iparraguirre, R. G´omez-Bombarelli, T. Hirzel,
A. Aspuru-Guzik, and R. P. Adams. Convolutional networks on graphs for learning molecular
fingerprints. arXiv preprint arXiv:1509.09292, 2015.
6. J. Feinstein, W. Shi, J. Ramanujam, and M. Brylinski. Bionoi: A voronoi diagram-based
representation of ligand-binding sites in proteins for machine learning applications. In
Protein-Ligand Interactions and Drug Design, pages 299–312. Springer, 2021.
7. W. P. Feinstein and M. Brylinski. efindsite: Enhanced fingerprint-based virtual screening against
predicted ligand binding sites in protein models. Molecular informatics, 33(2):135–150, 2014.
8. A. M. Fout. Protein interface prediction using graph convolutional networks. PhD thesis, Colorado
State University, 2017.
9. J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl. Neural message passing for
quantum chemistry. In International conference on machine learning, pages 1263–1272. PMLR,
2017.
10. R. G. Govindaraj, M. Naderi, M. Singha, J. Lemoine, and M. Brylinski. Large-scale
computational drug repositioning to find treatments for rare diseases. NPJ systems biology and
applications, 4(1):1–10, 2018.
11. R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping.
In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition
(CVPR’06), volume 2, pages 1735–1742. IEEE, 2006.
12. W. L. Hamilton, R. Ying, and J. Leskovec. Inductive representation learning on large graphs. In
Proceedings of the 31st International Conference on Neural Information Processing Systems, pages
1025–1035, 2017.
13. K. He, G. Gkioxari, P. Doll´ar, and R. Girshick. Mask r-cnn. In Proceedings of the IEEE
international conference on computer vision, pages 2961–2969, 2017.
14. S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780,
1997.
15. J. Jim´enez, S. Doerr, G. Mart´ınez-Rosell, A. S. Rose, and G. De Fabritiis. Deepsite:
protein-binding site predictor using 3d-convolutional neural networks. Bioinformatics,
33(19):3036–3042, 2017.
16. O. Kana and M. Brylinski. Elucidating the druggability of the human proteome with e findsite.
Journal of computer-aided molecular design, 33(5):509–519, 2019.
11/13
17. D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
18. T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks.
arXiv preprint arXiv:1609.02907, 2016.
19. D. R. Koes, M. P. Baumgartner, and C. J. Camacho. Lessons learned in empirical scoring with
smina from the csar 2011 benchmarking exercise. Journal of chemical information and modeling,
53(8):1893–1904, 2013.
20. V. Le Guilloux, P. Schmidtke, and P. Tuffery. Fpocket: an open source platform for ligand pocket
detection. BMC bioinformatics, 10(1):1–11, 2009.
21. H. S. Lee and W. Im. G-losa: An efficient computational tool for local structure-centric biological
studies and drug design. Protein Science, 25(4):865–876, 2016.
22. Y. Li, R. Yu, C. Shahabi, and Y. Liu. Diffusion convolutional recurrent neural network:
Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926, 2017.
23. J. Lim, S. Ryu, K. Park, Y. J. Choe, J. Ham, and W. Y. Kim. Predicting drug–target interaction
using a novel graph neural network with 3d structure-embedded graph representation. Journal of
chemical information and modeling, 59(9):3981–3988, 2019.
24. T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Doll´ar. Focal loss for dense object detection. In
Proceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017.
25. C.-H. Ngan, D. R. Hall, B. Zerbe, L. E. Grove, D. Kozakov, and S. Vajda. Ftsite: high accuracy
detection of ligand binding sites on unbound protein structures. Bioinformatics, 28(2):286–287,
2012.
26. L. Pu, R. G. Govindaraj, J. M. Lemoine, H.-C. Wu, and M. Brylinski. Deepdrug3d: Classification
of ligand-binding pockets in proteins with a convolutional neural network. PLoS computational
biology, 15(2):e1006718, 2019.
27. D. B. Roche, D. A. Brackenridge, and L. J. McGuffin. Proteins and their interacting partners: An
introduction to protein–ligand binding site prediction methods. International journal of molecular
sciences, 16(12):29829–29842, 2015.
28. F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition
and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition,
pages 815–823, 2015.
29. W. Shi, J. M. Lemoine, A.-E.-M. A. Shawky, M. Singha, L. Pu, S. Yang, J. Ramanujam, and
M. Brylinski. Bionoinet: ligand-binding site classification with off-the-shelf deep neural network.
Bioinformatics, 36(10):3077–3083, 2020.
30. M. Simonovsky and J. Meyers. Deeplytough: learning structural comparison of protein binding
sites. Journal of chemical information and modeling, 60(4):2356–2366, 2020.
31. J. Skolnick, M. Gao, A. Roy, B. Srinivasan, and H. Zhou. Implications of the small number of
distinct ligand binding pockets in proteins for drug discovery, evolution and biochemical function.
Bioorganic & medicinal chemistry letters, 25(6):1163–1170, 2015.
32. O. Trott and A. J. Olson. Autodock vina: improving the speed and accuracy of docking with a
new scoring function, efficient optimization, and multithreading. Journal of computational
chemistry, 31(2):455–461, 2010.
33. L. Van der Maaten and G. Hinton. Visualizing data using t-sne. Journal of machine learning
research, 9(11), 2008.
12/13
34. O. Vinyals, S. Bengio, and M. Kudlur. Order matters: Sequence to sequence for sets. arXiv
preprint arXiv:1511.06391, 2015.
35. S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh. Convolutional pose machines. In
Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages
4724–4732, 2016.
36. K. Xu, W. Hu, J. Leskovec, and S. Jegelka. How powerful are graph neural networks? arXiv
preprint arXiv:1810.00826, 2018.
37. K. Xu, C. Li, Y. Tian, T. Sonobe, K.-i. Kawarabayashi, and S. Jegelka. Representation learning
on graphs with jumping knowledge networks. In International Conference on Machine Learning,
pages 5453–5462. PMLR, 2018.
13/13
| 2021 | Graphsite: Ligand-binding site classification using Deep Graph Neural Network | 10.1101/2021.12.06.471420 | [
"Shi Wentao",
"Singha Manali",
"Pu Limeng",
"Ramanujam J.",
"Brylinski Michal"
] | creative-commons |
1
1
Title:
2
The absence of C-5 DNA methylation in Leishmania donovani allows DNA enrichment from complex
3
samples
4
Authors:
5
Cuypers B1, 2#, Dumetz F1*#, Meysman P2, Laukens K2, De Muylder G1, Dujardin J-C1, 3 and Domagalska
6
MA1
7
Affiliations:
8
1 Molecular Parasitology, Institute of Tropical Medicine, Antwerp, Belgium, 2 ADReM Data Lab,
9
Department of Computer Science, University of Antwerp, Antwerp, Belgium, 3 Department of
10
Biomedical Sciences, University of Antwerp, Antwerp, Belgium. * Present address: Merrick’s Lab,
11
Department of Pathology, University of Cambridge, Cambridge, UK # Contributed equally.
12
Corresponding author: mdomagalska@itg.be
13
Keywords:
14
Leishmania, Trypanosomatids, DNA-Methylation, Epigenomics, Whole Genome Bisulfite Sequencing,
15
DNA-Enrichment
16
17
2
Abstract
18
Cytosine C5 methylation is an important epigenetic control mechanism in a wide array of Eukaryotic
19
organisms and generally carried out by proteins of C-5 DNA methyltransferase family (DNMTs). In
20
several protozoans the status of this mechanism remains elusive, such as in Leishmania, the
21
causative agent of the disease leishmaniasis in humans and a wide array of vertebrate animals. In
22
this work, we show that the Leishmania donovani genome contains a C-5 DNA methyltransferase
23
(DNMT) from the DNMT6 subfamily, of which the function is still unclear, and verified its expression
24
at RNA level. We created viable overexpressor and knock-out lines of this enzyme and characterised
25
their genome-wide methylation patterns using whole-genome bisulfite sequencing, together with
26
promastigote and amastigote control lines. Interestingly, despite DNMT6 presence, we found that
27
methylation levels were equal to or lower than 0.0003% at CpG sites, 0.0005% at CHG sites and
28
0.0126% at CHH sites at genome scale. As none of the methylated sites were retained after manual
29
verification, we conclude that there is no evidence for DNA methylation in this species. We
30
demonstrate that this difference in DNA methylation between the parasite (no detectable DNA
31
methylation) and the vertebrate host (DNA methylation), allows enrichment of parasite versus host
32
DNA using Methyl-CpG-binding domain columns, readily available in commercial kits. As such, we
33
depleted methylated DNA from mixes of Leishmania promastigote and amastigote DNA with human
34
DNA, resulting in average Leishmania:human enrichments from 62x up to 263x. These results open a
35
promising avenue for unmethylated DNA enrichment as a pre-enrichment step before sequencing
36
Leishmania clinical samples.
37
3
Introduction
38
DNA methylation is an epigenetic mechanism responsible for a diverse set of functions across the
39
three domains of life, Eubacteria, Archeabacteria, and Eukaryota. In Prokaryotes, many DNA
40
methylation enzymes are part of so-called restriction modification systems, which play a crucial role
41
in their defence against phages and viruses. Prokaryotic methylation typically occurs on the C5
42
position of cytosine (cytosine C5 methylation), the exocyclic amino groups of adenine (adenine-N6
43
methylation) or cytosine (cytosine-N4 methylation) (1). In Eukaryotic species, DNA methylation is
44
mostly restricted to 5-methylcytosine (me5C) and best characterised in mammals, where 70-80% of
45
the CpG motifs are methylated (2). As such, DNA methylation controls a wide range of important
46
cellular functions, such as genomic imprinting, X-chromosome inactivation (in humans), gene
47
expression and the repression of transposable elements. Consequently, defects in genetic imprinting
48
are associated with a variety of human diseases and changes in DNA methylation patterns are
49
common hallmark of cancer (3,4). Eukaryotic DNA methylation can also occur at CHG and CHH
50
(where H is A, C or T) sites (5), which was considered to occur primarily in plants. However, studies
51
from the past decade demonstrate that CHG and CHH methylation are also frequent in several
52
mammalian cells types, such as embryonic stem cells, oocytes and brains cells (5-8).
53
Me5C methylation is mediated by a group of enzymes called C-5 DNA methyltransferases (DNMTs).
54
This ancient group of enzymes share a common ancestry and their core domains are conserved
55
across Prokaryotes and Eukaryotes (1). Different DNMT subfamilies have developed distinct roles
56
within epigenetic control mechanisms. For example, in mammals DNMT3a and DNMT3b are
57
responsible for de novo methylation, such as during germ cell differentiation and early development,
58
or in specific tissues undergoing dynamic methylation (9). In contrast, DNMT1 is responsible for
59
maintaining methylation patterns, particularly during the S phase of the cell cycle where it
60
methylates the newly generated hemimethylated sites on the DNA daughter strands (10). Some
61
DNMTs have also changed substrate over the course of evolution. A large family of DNMTs, called
62
DNMT2, has been shown to methylate the 38th position of different tRNAs
to yield ribo-5-
63
methylcytidine (rm5C) in a range of Eukaryotic organisms, including humans (11), mice (12),
64
Arabidopsis thaliana (13) and Drosophila melanogaster (14). Therefore, DNMT2s are now often
65
referred to as ‘tRNA methyltransferases’ or trDNMT and are known to carry out diverse regulatory
66
functions (15). However, in other Eukaryotic taxa, DNMT2 appears to be a genuine DNMT, as DNMT2
67
can catalyze DNA methylation in Plasmodium falciparum (16), and Schistosoma mansoni (17). In
68
Entamoeba histolytica both DNA and RNA can be used as substrates for DNMT2 (18,19). The
69
increase in available reference genomes of non-model Eukaryotic species has recently also resulted
70
4
in the discovery of new DNMTs, such as DNMT5, DNMT6 or even SymbioLINE-DNMT, a massive
71
family of DNMTs, so far only found in the dinoflagellate Symbiodinium (20).
72
Indeed, DNMT mediated C5 methylation has been shown to be of major functional importance in a
73
wide array of Eukaryotic species, including also protozoans such as Toxoplasma gondii and
74
Plasmodium (16,21). In contrast, studies have failed to detect any C5 DNA methylation in Eukaryotic
75
species such as Caenorhabditis elegans, Saccharomyces cerevisiae and Schizosaccharomyces pombe
76
(22,23). In many other protozoans, the presence and potential role of DNA-methylation remains
77
elusive. This is especially true for Leishmania, a Trypanosomatid parasite (Phylum Euglenozoa),
78
despite its medical and veterinary importance. Leishmania is the causative agent of the leishmaniasis
79
in humans and a wide variety of vertebrate animals, a disease that ranges from self-healing
80
cutaneous lesions to lethal visceral leishmaniasis.
81
Leishmania features a molecular biology that is remarkably different from other Eukaryotes. This
82
includes a system of polycistronic transcription of functionally unrelated genes (24). The successful
83
transcription of these cistrons depends at least on several known epigenetic modifications at the
84
transcription start sites (acetylated histone H3) and transcription termination sites (β-D-glucosyl-
85
hydroxymethyluracil, also called ‘Base J’), but little research has been done towards other epigenetic
86
modifications (25). We were therefore interested in the 5-C methylation status of Leishmania, which
87
has been poorly explored to date. In this context, a single study on a wide range of Eukaryotic
88
species lacking DNMT1 reported the absence of CG-specific methylation in Leishmania major,
89
however, using only a single sample of an unspecified life stage (26). The study also does not
90
comment on CHH and CHG specific methylation, which can be relevant as well. Contrastingly,
91
another manuscript demonstrated Me5C methylation in T. brucei, another Trypanosomatid species,
92
although at low levels (0.01 %) (27). To clarify the status of C-5 DNA methylation in Trypanosomatids
93
and Leishmania in particular, we present the first comprehensive study of genomic methylation in
94
Leishmania across different parasite life stages, making use of high-resolution whole genome
95
bisulfite sequencing.
96
5
Materials and Methods
97
In silico identification and phylogeny of putative DNMTs
98
To identify putative C-5 cytosine-specific DNA methylases in Leishmania donovani, we obtained the
99
hidden Markov model (hmm) for this protein family from PFAM version 32.0 (Accession number:
100
PF00145) (28). The hmm search tool of hmmer-3.2.1 (hmmer.org) was then used with default
101
settings to screen the LdBPKV2 reference genome for this hmm signature (29). The initial pairwise
102
alignment between the identified L. donovani and T. brucei C5 DNA MTase was carried out with T-
103
COFFEE V_11.00.d625267.
104
To construct a comprehensive phylogenetic tree of the C5 DNA MTase family, including members
105
found in Trypanosomatid species, we modified the approach from Ponts et al. (16). Firstly, we
106
downloaded the putative proteomes of a wide range of Prokaryotic and Eukaryotic species. These
107
species were selected to cover the different C5 DNA MTase subfamilies (1). Specifically, the following
108
proteomes were obtained: Trypanosoma brucei TREU92, Trypanosoma vivax Y486 and Leishmania
109
major Friedlin from TriTrypDB v41 (24,30,31), Plasmodium falciparum 3D7, and Plasmodium vivax
110
P01 from PlasmodDB v41 (32-34), Cryptosporidium parvum Iowa II and Cryptosporidium hominis
111
TU502 from CryptoDB v41 (35-37), Toxoplasma gondii ARI from ToxoDB v 41 (38,39), Euglena gracilis
112
Z1
(PRJNA298469)
(40),
Entamoeba
histolytica
HM-1:IMSS
(GCF_000208925.1)
(41),
113
Schizosaccharomyces pombe ASM294 (GCF_000002945.1) (42), Saccharomyces cerevisiae S288C
114
(GCF_000146045.2), Neurospora crassa OR74A (GCF_000182925.2) (43), Arabidopsis thaliana
115
(GCF_000001735.4), Drosophila melanogaster (GCF_000001215.4), Homo sapiens GRCh38.p12
116
(GCF_000001405.38), Bacillus subtilis 168 (GCF_000009045.1), Clostridium botulinum ATCC 3502
117
(GCF_000063585.1) (44), Streptococcus pneumoniae R6 (GCF_000007045.1) (45), Agrobacterium
118
tumefaciens (GCF_000971565.1) (46), Salmonella enterica CT18 (GCF_000195995.1) (47) and
119
Escherichia coli K12 (GCF_000005845.2) from NCBI, Ascobolus immersus RN42 (48) from the JGI
120
Genome Portal (genome.jgi.doe.gov) and Danio rerio (GRCz11) from Ensembl (ensembl.org).
121
All obtained proteomes were then searched with the hmm signature for C5 DNA MTases, exactly as
122
described above for L. donovani. All hits with an E-value < 0.01 (i.e. 1 false positive hit is expected in
123
every 100 searches with different query sequences) were maintained, and all domains matching the
124
query hmm were extracted and merged per protein. This set of sequences was aligned in Mega-X
125
with the MUSCLE multiple sequence alignment algorithm (49,50) and converted to the PHYLIP
126
format with the ALTER tool (51). Phage sequences and closely related isoforms were removed.
127
A maximum likelihood tree of this alignment was generated with RAxML version 8.2.10 using the
128
automatic protein model assignment algorithm (option: -m PROTGAMMAAUTO). RAxML was run in
129
6
three steps: Firstly, 20 trees were generated and only the one with the highest likelihood score was
130
kept. Secondly, 1000 bootstrap replicates were generated. In a final step, the bootstrap bipartions
131
were drawn on the best tree from the first round. The tree was visualised in Figtree v1.4.4
132
(https://github.com/rambaut/figtree/).
133
Culturing & DNA extraction for Bisulfite Sequencing
134
Promastigotes (extracellular life stage) of Leishmania donovani MHOM/NP/03/BPK282/0 cl4 (further
135
called BPK282) and its genetically modified daughter lines (see below) were cultured in HOMEM
136
(Gibco) supplemented with 20% (v:v) heat-inactivated foetal bovine serum at 26°C. Amastigotes
137
(intracellular life stage) of the same strain were obtained from three months infected golden Syrian
138
hamster (Charles Rivers) as described in Dumetz et al. (29) and respecting BM2013-8 ethical
139
clearance from Institute of Tropical Medicine (ITM) Animal Ethic Committee. Briefly, 5 week old
140
female golden hamsters were infected via intracardiac injection of 5.105 stationary phase
141
promastigotes. Three months post infection, hamsters were euthanised and amastigotes were
142
purified from the liver by Percol gradient (GE Healthcare) after homogenisation. T. brucei gambiense
143
MBA blood stream forms were obtained from OF-1 mice when the parasitaemia was at its highest,
144
according to ITM Animal Ethic Committee decision BM2013-1. Parasites were separated from the
145
whole blood as described in Tihon et al. (52). Briefly, the parasites were separated from the blood by
146
placing the whole blood on an anion exchanger Diethylaminoethyl (DEAE)-cellulose resin (Whatman)
147
suspended in phosphate saline glucose (PSG) buffer, pH 8. After elution and two washes on PSG,
148
DNA was extracted. DNA of L. donovani, both promastigotes and amastigotes, as well as T. brucei
149
was extracted using DNeasy Blood & Tissue kit (Qiagen) according manufacturer instructions.
150
Arabidopsis thaliana Col-0 was grown for 21 days under long day conditions, i.e. 16 hrs light and 8
151
hrs darkness. DNA was then extracted from the whole rosette leaves using the DNeasy Plant Mini Kit
152
(Qiagen).
153
Genetic engineering of L. donovani BPK282
154
We generated both an LdDNMT overexpressing (LdDNMT+) and null mutant line (LdDNMT-/-) of L.
155
donovani BPK282. All the PCR products generated to produce the constructs for LdDNMToverex and
156
LdDNMTKO were sequenced at the VIB sequencing facility using the same primer as for the
157
amplification. For LdDNMToverex, the overexpression construct, pLEXSY-DNMT, was generated by
158
PCR amplification of LdBPK_251230 from BPK282 genomic DNA using Phusion (NEB) and cloned
159
inside the expression vector pLEXSY-Hyg2 (JENA bioscience) using NEBuilder (NEB) according to
160
manufacturer’s instruction for primer design and cloning instructions (sup table for primers list).
161
Once generated, 10 µg of pLEXSY-DNMT was electroporated in 5.107 BPK282 promastigotes from
162
7
logarithmic culture using cytomix on a GenePulserX (BioRad) according to LeBowitz (1994) (53) and
163
selected in vitro by adding 50 μg/mL hygromycin B (JENA Bioscience) until parasite growth (54).
164
Verification of overexpression was carried out by qPCR on a LightCycler480 (Roche) using SensiMix
165
SYBR No-ROX (Bioline) on cDNA. Briefly, 108 logarithmic-phase promastigotes were pelleted, RNA
166
extraction was performed using RNAqueous-Micro total RNA isolation kit (Ambion) and quantified
167
by Qubit and the Qubit RNA BR assay (Life Technologies, Inc.). Transcriptor reverse transcriptase
168
(Roche) was used to synthesise cDNA following manufacturer’s instructions. qPCRs were run on a
169
LightCycler 480 (Roche) with a SensiMix SYBR No-ROX kit (Bioline); primer sequences available in
170
Supplementary Table S1. Normalisation was performed using two transcripts previously described
171
as stable in promastigotes and amastigotes in Dumetz et al. (2018) (55), LdBPK_340035000 and
172
LdBPK_240021200.
173
For the generation of LdDNMT-/-, a two-step gene replacement strategy was used: replacing the first
174
allele of LdBPK_250018100.1 by nourseothricin resistance gene (SAT) and the second allele by a
175
puromycin resistance gene (Puro). Briefly, each drug resistance gene was PCR amplified from pCL3S
176
and pCL3P using Phusion (NEB) and cloned between 300 bp of PCR amplified DNA fragments of the
177
LdBPK_250018100.1 5’ and 3’ UTR using NEBuilder (NEB) inside pUC19 for construct amplification in
178
E. coli DH5α (Promega) (cf. primer list in Supplementary Table S1). Each replacement construct was
179
excised from pUC19 using SmaI (NEB), dephosphorylated using Antarctic Phosphatase (NEB) and 10
180
µg of DNA was used for the electroporation in the same conditions as previously described to insert
181
the pLEXSY-DNMT. The knock-out was confirmed by whole genome sequencing.
182
Bisulfite sequencing and data analysis
183
For each sample, one microgram of genomic DNA was used for bisulfite conversion with
184
innuCONVERT Bisulfite All-In-One Kit (Analytikjena). Sequencing libraries were prepared with the
185
TruSeq DNA Methylation kit according to the manufacturer’s instructions (Illumina). The resulting
186
libraries were paired-end (2 x 100bp) sequenced on the Illumina HiSeq 1500 platform of the
187
University of Antwerp (Centre of Medical Genetics). The sequencing quality was first verified with
188
FastQC v0.11.4. Raw reads generated for each sample were aligned to their respective reference
189
genome with BSseeker 2-2.0.3 (56): LdBPK282v2 (29) for L. donovani, TREU927 (30) for T. brucei and
190
Tair10 (57) for the A. thaliana positive control. Samtools fixmate (option -m) and samtools markdup
191
(option -r) were then used to remove duplicate reads. CpG, CHG and CHH methylation sites were
192
subsequently called with the BS-Seeker2 ‘call’ tool using default settings and further filtered with our
193
Python3 workflow called ‘Bisulfilter’ (available at https://github.com/CuypersBart/Bisulfilter).
194
Genome-wide visualisation of methylated regions was then carried out with ggplot2 in R (58). In
195
8
Leishmania, the positions that passed our detection thresholds (coverage > 25, methylation
196
percentage > 0.8), were then manually inspected in IGV 2.5.0 (59).
197
Leishmania DNA enrichment from a mix of human and Leishmania DNA
198
To check whether the lack of detectable DNA methylation in Leishmania can be used for the
199
enrichment of Leishmania versus (methylated) human DNA, we carried out methylated DNA removal
200
on two types of samples: (1) An artificial mix of L. donovani BPK282/0 cl4 promastigote DNA with
201
human DNA (Promega) from 1/15 to 1/150000 (Leishmania:human) and (2) Linked promastigote and
202
hamster-derived amastigote samples from 3 clinical Leishmania donovani strains (BPK275, BPK282
203
and BPK026), which were generated in previous work (29). For this experiment, we used a 1/1500
204
artificial mix of promastigote DNA and human DNA (Promega) to reflect the median ratio found in
205
clinical samples. For each of the three biological replicates (strains), we carried out the experiment in
206
duplicate (technical replicates). All parasite DNA was extracted with the DNA (DNeasy Blood & Tissue
207
kit, Qiagen). Leishmania DNA (0.0017 ng/μL) was then enriched from the human DNA (25ng/μL)
208
using NEBNext Microbiome DNA Enrichment Kit (NEB) according to manufacturer instructions.
209
Evaluation of the ratio Leishmania/human DNA was performed by qPCR on LightCycler480 (Roche)
210
using SensiMix SYBR No-ROX (Bioline) and RPL30 primers provided in the kit to measure human DNA
211
and Leishmania CS primers (Cysteine synthase) (60).
212
9
Results
213
The Leishmania genome contains a putative C-5 DNA methyltransferase (DNMT)
214
Eukaryotic DNA methylation typically requires the presence of a functional C-5 cytosine-specific DNA
215
methylase (C5 DNA MTase). This type of enzymes specifically methylates the C-5 position of
216
cytosines in DNA, using S-Adenosyl methionine as a methyl-donor. To check for the presence of C5
217
DNA MTases in Leishmania donovani, we carried out a deep search of the parasite’s genome. In
218
particular, we used the LdBPKv2 reference genome (29) and searched the predicted protein
219
sequences of this assembly using the hidden-markov-model (hmm) signature of the C5 DNA MTase
220
protein family obtained from PFAM (PF00145) and obtained a single hit: the protein
221
LdBPK_250018100.1, (E-value: 2.7e-40). LdBPK_250018100.1 was already annotated as ‘modification
222
methylase-like protein’ with a predicted length of 840 amino acids. We will further refer to this
223
protein as LdDNMT. Interestingly, in another Trypanosomatid species, Trypanosoma brucei, the
224
homolog of this protein (Tb927.3.1360 or TbDNMT) has been previously been studied in detail by
225
Militello et al (27). Moreover, these authors showed that TbDNMT has all the ten conserved
226
domains that are present in functional DNMTs. We aligned TbDNMT with LdDNMT using T-Coffee
227
(Fig. 1) and found that these 10 domains are also present in LdDNMT, including also the putative
228
catalytic cysteine residue in domain IV.
229
Leishmania and Trypanosomatid C-5 DNA belong to the Eukaryotic DNMT6 family
230
To learn more about the putative function and evolutionary history of this protein, we wanted to
231
characterise the position of LdDNMT and those of related Trypanosomatid species within the DNMT
232
phylogenetic tree. Consequently, we collected the publicly available, putative proteomes of a wide
233
range of Prokaryotic and Eukaryotic species, searched them for the hmm signature of the C5-DNMT
234
family, aligned the identified proteins and generated a RAxML maximum likelihood tree. In total we
235
identified 131 putative family members in the genomes of 24 species (E-value < 0.01), including 4
236
Prokaryotic (Agrobacterium tumefaciens, Salmonella enterica, Escherichia coli and Clostridium
237
botulinum) and 20 Eukaryotic species. These Eukaryotic species were selected to contain organisms
238
from the Excavata Phylum (of which Leishmania is part) and a range of other, often better-
239
characterised Phyla as a reference. The Excavata species included 4 Trypanosomatids (Leishmania
240
donovani, Leishmania major, Trypanosoma brucei and Trypanosoma vivax), 1 other, non-
241
Trypanosomatid Euglenozoid species (Euglena gracilis) and 1 other non-Euglenozoid species
242
(Naegleria gruberi). The other Eukaryotic Phyla included in the analysis were: Apicomplexa
243
(Plasmodium vivax, Plasmodium falciparum, Cryptosporidium parvum, Cryptosporidium hominis),
244
10
Amoebozoa (Entamoeba histolytica), Angiosperma (Arabidopsis thaliana, Oryza sativa), Ascomycota
245
(Ascobolus immerses, Neurospora crassa) and Chordata (Homo Sapiens, Danio rerio) (Figure 2).
246
Our phylogenetic tree was able to clearly separate known DNMT subgroups, including DNMT1,
247
DNMT2, DNMT3, DRM (Domain rearranged methyltransferase), DIM and 2 groups of Prokaryotic
248
DNMTs (1,16,61). Interestingly, the tree also showed that Trypanosomatid DNMTs group together
249
and are part of the much less-characterised DNMT6 group, as has been previously described for
250
Leishmania major and Trypanosoma brucei (20). This group of DNMTs has also been found also in
251
diatoms (e.g. Thalasiosira) and recently in dinoflaggelates (e.g. Symbiodinium kawagutii and
252
Symbiodinium minutum), but its function remains elusive (20,62). The most closely related branch to
253
DNMT6 contains a group of bacterial DNMTs (here represented by Agrobacterium tumefaciens,
254
Salmonella enterica, Escherichia coli). This highlights that DNMT6 emerged from the pool of
255
Prokaryotic DNMTs independently from the groups previously mentioned. The fact that another
256
Euglenozoid, Euglena gracilis, has DNMT1, DNMT2, DNMT4 and DNMT5, while another Excavata
257
species, Naegleria gruberi has both a DNMT1 and a DNMT2, suggests that the ancestors of the
258
current Excavata species possessed a wide battery of DNMTs including also DNMT6. In the lineage
259
that eventually led to Trypanosomatids, these were all lost, except DNMT6.
260
Whole genome bisulfite sequencing reveals no evidence for functional C-5 methylation
261
As 1) we identified LdBPK_250018100.1 to be from the C5 DNA MTase family, 2) all 10 conserved
262
domains were present, we decided to check also for the presence and functional role of C5 DNA
263
methylation in L. donovani. Therefore, we assessed the locations and degree of CpG, CHG and CHH
264
methylation across the entire Leishmania genome and within the two parasite life stages:
265
amastigotes (intracellular mammalian life stage) and promastigotes (extracellular, insect life stage).
266
Amastigotes were derived directly from infected hamsters, while promastigotes were obtained from
267
axenic cultures. Promastigotes were divided in two batches, one passaged long-term in axenic
268
culture, the other passaged once through a hamster and then sequenced at axenic passage 3, thus
269
allowing us to study also the effect of long versus- short term in vitro passaging. Arabidopsis thaliana
270
and T. brucei were included as a positive control as the degree of CpG, CHG and CHH methylation in
271
A. thaliana is well known (63,64), while T. brucei is the only Trypanosomatid in which (low)
272
methylation levels were previously detected by mass spectrometry (27).
273
An overview of all sequenced samples can be found in Supplementary Table S2. All L. donovani
274
samples were sequenced with at least 30 million 100bp paired end reads (60 million total) per
275
sample resulting in an average genomic coverage of at least 94X for the Leishmania samples. The T.
276
brucei was sequenced with 69 million PE reads resulting in 171X average coverage and A. thaliana 27
277
11
million PE reads, resulting in 21X average coverage. Detailed mapping statistics can be found in
278
Supplementary Table S3.
279
We first checked for global methylation patterns across the genome. Interestingly, we could not
280
detect any methylated regions in Leishmania donovani promastigotes, both short (P3) and long-term
281
in vitro passaged, nor in hamster derived promastigotes or amastigotes (Fig. 3). Minor increases in
282
the CHH signal towards the start end of several chromosomes, were manually checked in IGV and
283
attributed to poor mapping in (low complexity) telomeric regions. This was in contrast to our
284
positive control, Arabidopsis thaliana, that showed clear highly methylated CpG, CHG and CHH
285
patterns across the genome. This distribution was consistent with prior results with MethGO
286
observed on Arabidopsis thaliana, confirming that our methylation detection workflow was working
287
(58).
288
In a second phase, we checked for individual sites that were fully methylated (>80% of the
289
sequenced DNA at that site) using BS-Seeker2 and filtering the results with our automated Python3
290
workflow. CpG methylation in all three biological samples for L. donovani was lower than 0.0003%,
291
CHG methylation lower than 0.0005% and CHH methylation lower than 0.0126% (Table 1,
292
Supplementary Table S4). However, when this low number of detected ‘methylated’ sites was
293
manually verified in IGV, they could all clearly be attributed to regions where BS-Seeker2 wrongly
294
called methylated bases, either because of poor mapping (often in repetitive, low complexity
295
regions) or of strand biases. In reliably mapped regions, there was clearly no methylation. Similarly,
296
we detected 0.0001% of CpG methylation, 0.0006% of CHG methylation and 0.0040% of CHH
297
methylation for T. brucei, which could all be attributed to mapping errors or strand biases. In A.
298
thaliana, our positive control, we detected 21.05% of CpG methylation, 4.04% of CHG methylation
299
and 0.31% of CHH methylation, which is similar as reported values in literature (65,66), and
300
demonstrates that our bioinformatic workflow could accurately detect methylated sites. We also
301
checked sites with a lower methylation degree (>40%), which gave higher percentages, but this
302
could be attributed to the increased noise level at this resolution (Supplementary Table S4). Indeed,
303
even when applying stringent coverage criteria (>25x) this approach is susceptible for false positive
304
methylation calls, as we are checking millions of positions (in case of Leishmania, more than 5.8
305
million CG sites, 3.9 million CHG sites and 9.3 million CHH sites).
306
To determine whether LdDNMT is essential and/or if it affects the C-5 DNA-methylation pattern, we
307
also sequenced an L. donovani DNMT knock-out (LdDNMT-/-) line as well as a DNMT overexpressor
308
(LdDNMT+). The successful generation of LdDNMT-/- and LdDNMT+ was verified by calculating their
309
LdDNMT copy number based on the sequencing coverage (Figure 4). Indeed, the copy number of the
310
12
LdDNMT gene in LdDNMT-/- was reduced to zero, while that of LdDNMT+ was increased to 78
311
copies. The overexpressor was also verified on the RNA level (Table 2) and showed a 2.5-fold higher
312
expression than the corresponding wild type. Although the LdDNMT+ initially seemed to have
313
slightly higher methylation percentages (Table 1, Supplementary Table S4), none of these
314
methylation sites passed our manual validation in IGV. Thus, we did not find evidence for
315
methylation in either of these lines. Additionally, the fact that the LdDNMT-/- line was viable shows
316
that LdDNMT is not an essential gene in promastigotes.
317
Absence of C5 DNA Methylation as a Leishmania vs host DNA enrichment strategy
318
The lack or low level of C5 DNA methylation opens the perspective for enriching Leishmania DNA in
319
mixed parasite- host DNA samples, based on the difference in methylation status (the vertebrate
320
host does show C5 DNA methylation). This could potentially be an interesting pre-enrichment step
321
before whole genome sequencing analysis of clinical samples containing Leishmania. Furthermore,
322
commercial kits for removing methylated DNA are readily available and typically contain a Methyl-
323
CpG-binding domain (MBD) column, which binds methylated DNA while allowing unmethylated to
324
flow trough.
325
To test this if these kits can be used for Leishmania, we first generated artificially mixed samples
326
using different ratios of L. donovani promastigote DNA with human DNA. Ratios were made starting
327
from 1/15 to 1/15000, which reflects the real ratio of Leishmania vs human DNA in clinical samples
328
(67). From these mixes, Leishmania DNA was enriched using NEBNext Microbiome DNA Enrichment
329
Kit (NEB) that specifically binds methylated DNA, while the non-methylated remains in the
330
supernatant. We observed an average 263 X enrichment of Leishmania versus human DNA (Figure
331
5). This ranged between 378x for the lowest dilution (removing 99.8% of the human DNA) to 164x
332
(removing 99.6% of the human DNA) in the highest diluted condition (1/15000 Leishmania:human).
333
Secondly, we wanted to test if enrichment via MBD columns worked equally well on L. donovani
334
amastigotes for (a) fundamental reasons, as an (indirect) second method to detect if there are any
335
methylation differences between promastigotes and amastigotes, and (b) practical reasons, as it the
336
(intracellular) life stage encountered in clinical samples. Therefore, we also carried out this
337
enrichment technique on 3 sets (3 strains) of hamster derived amastigotes and their promastigote
338
controls. Similarly as in the previous experiment, Leishmania-human DNA mixes were generated in a
339
1/1500 (Leishmania:human) ratio after which enrichment was carried out with the NEBNext
340
Microbiome DNA Enrichment Kit. The enrichment worked well for both life stages, the promastigote
341
samples were on average 76.22 ± 14.28 times enriched and the amastigote samples 61.68 ± 4.23
342
times (Table 3).
343
13
Discussion
344
With this work, we present the first comprehensive study addressing the status of DNA-methylation
345
in Leishmania.
346
We demonstrated that the Leishmania genome contains a C5-DNMT (LdDNMT) that contains all 10
347
conserved DNMT domains. We also showed the gene is expressed at the RNA level. As the C5-DNMT
348
family is diverse and several family members are known to have adopted (partially) distinct functions
349
during the course of evolution, we were particularly interested in the position of this DNMT within
350
the evolutionary tree of this family, as it could direct hypotheses about the function of this protein.
351
We found that LdDNMT is in fact a DNMT6, just as those found in L. major and T. brucei (20).
352
Interestingly, all other (non-Trypanosomatid) species studied so far had either multiple DNMT6
353
copies and/or other DNMT subfamily members in their genomes (20,62). Therefore,
354
Trypanosomatids might be a unique model species to further study the role of this elusive DNMT
355
subfamily, as there can be no interaction with the effects of other DNMTs.
356
The fact that our LdDNMT knock-out line (verified by sequencing) was viable shows that DNMT6 is
357
not essential for the survival of the parasite, at least in promastigotes and in our experimental
358
conditions. However, at the same time one might hypothesize that DNMT6 does offer a selective
359
advantage to the parasite. First of all, the sequence of DNMT core domains is extremely conserved
360
across the tree of life and this is no different from those that we encountered in Leishmania.
361
Secondly, Leishmania is characterised by a high genome plasticity and features extensive gene copy
362
number differences between strains (68,69). Therefore, one might speculate that the parasite would
363
have lost the gene long time ago if it did not provide any selective advantage.
364
In addition, we aimed to characterise the DNA-methylation patterns of the parasite’s genome.
365
Therefore, we carried out the first multi-life stage whole genome bisulfite sequencing experiment on
366
Leishmania and Trypanosomatids in general. We checked both the promastigote (both culture and
367
amastigote life stage). Surprisingly, we did not find any evidence for DNA methylation in L. donovani
368
even though we checked both for large, regional patterns (sensitive for low levels of methylation
369
over longer distances) and site-specific analyses (sensitive for high levels of methylation at individual
370
sites). This could either mean that there is indeed no DNA-methylation in these species, or that was
371
below our detection threshold. Regarding this detection threshold, two factors should be
372
considered. Firstly, bisulfite sequencing and analysis allows for the detection of specific sites that are
373
consistently methylated across the genomes of a mix of cells. For example, in our case, we looked
374
for sites that are methylated in at least 80% or 40% of the cases. Thus, if Leishmania consistently
375
methylates certain genomic positions, our pipeline would have uncovered this. However, if this
376
14
methylation would be more random, or occurring in only a small subset of cells, we would not be
377
able to distinguish this for random sequencing errors, and as such, we cannot exclude this possibility.
378
Secondly, bisulfite sequencing typically suffers from poor genomic coverage due to the harsh BS
379
treatment of the DNA (70). In our L. donovani samples we covered at least 30.14% of the CpG sites,
380
29.47% of the CHG sites and 24.23% of the CHH sites (even though having more than 90x average
381
coverage). However, as there are millions of CpG, CHG and CHH sites in the genome, the chance is
382
very small (0.75n, with n = number of methylated sites) that we would not have detected methylated
383
sites, even if present in low numbers.
384
In any case, it is hard to imagine that any of the typical Eukaryotic DNA methylation systems such as
385
genomic imprinting, chromosome inactivation, gene expression regulation and/or the repression of
386
transposable elements could be of significance with such low methylation levels. On the other hand,
387
given its phylogenetic position, it is perfectly possible that DNMT6 has changed its biological activity
388
and now carries out another function. Indeed, as we described above, a similar phenomenon was
389
observed with DNMT2 that switched it substrate from DNA to tRNA during the course of evolution
390
[16,17].
391
Correspondingly, we did not observe any detectable DNA methylation for T. brucei. These findings
392
are, however, in contrast to what has been reported before by Militello et al., who detected 0.01%
393
of 5MC in the T. brucei genome 28. Also, the methylated (orthologous) loci described in this paper
394
could not be confirmed in the current work. However, this is maybe not be surprising as the same
395
authors reported later that TbDNMT might in fact methylate RNA, as they identified methylated sites
396
in several tRNAs 71. This would indeed explain why we do not observe C5-DNA methylation in T.
397
brucei with high resolution, whole genome bisulfite sequencing, and further suggest that a similar
398
substrate switch to tRNA has occurred for DNMT6, just like has occurred for DNMT2. Further
399
functional characterisation of DNMT6 is required to verify this hypothesis.
400
From an applied perspective, this study opens new avenues for the enrichment of Trypanosmatid
401
DNA from clinical samples, which often have an abundance of host DNA. Indeed, depletion of
402
methylated DNA could be included as pre-enrichment step for existing enrichment approaches. For
403
example, our group has recently obtained excellent sequencing results of clinical samples using
404
SureSelect (97% of the samples for diagnostic SNPs, 83% for genome wide information for
405
sequenced samples), but was not able to sequence samples below 0.006% of Leishmania DNA
406
content (71). Perhaps the removal of methylated DNA could further enhance the sensitivity of this
407
method. In the case of Leishmania the technique could even be useful both from enrichments from
408
the mammalian hosts and the insect vector, as it was recently shown the phlebotomine vector also
409
15
carries Me5C in its genome (72). The depletion of methylated DNA as a pre-enrichment step before
410
whole genome sequencing has also been successfully used before for the parasite Plasmodium
411
falciparum (malaria) and shown to generate unbiased sequencing reads (73).
412
In conclusion, we demonstrated that the Leishmania genome encodes for a DNMT6, but DNA
413
methylation is either absent or present in such low proportion that it is unlikely to have a major
414
functional role. Instead, we suggest that more investigation at RNA level is required to address the
415
function of DNMT6 in Leishmania. The absence of DNA-methylation provides a new working tool for
416
the enrichment of Leishmania DNA in clinical samples, thus facilitating future parasitological studies.
417
Data Availability
418
Raw sequencing data is available in the Sequence Read Archive under project accession numbers
419
PRJNA560731 and PRJNA560871. Individual sample accession numbers are available in
420
Supplementary Table S2.
421
Acknowledgments
422
We thank Dr. Gaurav Zinta and Prof. Dr. Gerrit Beemster for providing us with the A. thaliana DNA,
423
as well as Prof. Dr. Philippe Büscher and Nicolas Bebronne for the blood stream form of T. brucei
424
gambiense MBA and Prof. Dr. Joachim Clos for the Leishmania expression vectors pCL3S an pCL3P.
425
This work was supported by the Interuniversity Attraction Poles Program of Belgian Science Policy
426
[P7/41 to JC.D.] and by the organisation “Les amis des Instituts Pasteur à Bruxelles, asbl” [F.D.]. The
427
computational resources and services used in this work were provided by the VSC (Flemish
428
Supercomputer Center), funded by the Research Foundation - Flanders (FWO) and the Flemish
429
Government – department EWI. This work was also supported by the Department of Economy,
430
Science and Innovation in Flanders ITM-SOFIB (SINGLE project, to J-C D). We thank the Center of
431
Medical Genetics at the University of Antwerp for hosting the NGS facility. BC is a post-doctoral
432
fellow funded from the FWO [12V5319N].
433
Author contributions
434
Designed the experiments: B.C., F.D., G.D.M., J-C.D., M.A.D. Performed the experiments: B.C., F.D.,
435
M.A.D. Analysed the data: B.C., F.D., P.M., K.L., M.A.D. Wrote the manuscript: B.C., F.D., P.M., K.L., J-
436
C.D, M.A.D. All authors reviewed and approved the final version of the manuscript.
437
Additional Information
438
The authors declare no competing interests.
439
16
References
440
1.
Jurkowski, T.P. and Jeltsch, A. (2011) On the evolutionary origin of eukaryotic DNA
441
methyltransferases and Dnmt2. PLoS One, 6, e28104.
442
2.
Li, E. and Zhang, Y. (2014) DNA methylation in mammals. Cold Spring Harb Perspect Biol, 6,
443
a019133-a019133.
444
3.
Vidal, E., Sayols, S., Moran, S., Guillaumet-Adkins, A., Schroeder, M.P., Royo, R., Orozco, M.,
445
Gut, M., Gut, I., Lopez-Bigas, N. et al. (2017) A DNA methylation map of human cancer at
446
single base-pair resolution. Oncogene, 36, 5648.
447
4.
Jin, Z. and Liu, Y. (2018) DNA methylation in human diseases. Genes & diseases, 5, 1-8.
448
5.
Guo, W., Chung, W.-Y., Qian, M., Pellegrini, M. and Zhang, M.Q. (2013) Characterizing the
449
strand-specific distribution of non-CpG methylation in human pluripotent cells. Nucleic acids
450
research, 42, 3009-3016.
451
6.
Lister, R., Pelizzola, M., Dowen, R.H., Hawkins, R.D., Hon, G., Tonti-Filippini, J., Nery, J.R., Lee,
452
L., Ye, Z., Ngo, Q.M. et al. (2009) Human DNA methylomes at base resolution show
453
widespread epigenomic differences. Nature, 462, 315-322.
454
7.
Tomizawa, S., Kobayashi, H., Watanabe, T., Andrews, S., Hata, K., Kelsey, G. and Sasaki, H.
455
(2011) Dynamic stage-specific changes in imprinted differentially methylated regions during
456
early mammalian development and prevalence of non-CpG methylation in oocytes.
457
Development (Cambridge, England), 138, 811-820.
458
8.
Xie, W., Barr, C.L., Kim, A., Yue, F., Lee, A.Y., Eubanks, J., Dempster, E.L. and Ren, B. (2012)
459
Base-resolution analyses of sequence and parent-of-origin dependent DNA methylation in
460
the mouse genome. Cell, 148, 816-831.
461
9.
Chédin, F. (2011) In Cheng, X. and Blumenthal, R. M. (eds.), Progress in Molecular Biology
462
and Translational Science. Academic Press, Vol. 101, pp. 255-285.
463
10.
Jones, P.A. and Liang, G. (2009) Rethinking how DNA methylation patterns are maintained.
464
Nature reviews. Genetics, 10, 805-811.
465
11.
Jurkowski, T.P., Meusburger, M., Phalke, S., Helm, M., Nellen, W., Reuter, G. and Jeltsch, A.
466
(2008) Human DNMT2 methylates tRNA(Asp) molecules using a DNA methyltransferase-like
467
catalytic mechanism. RNA, 14, 1663-1670.
468
12.
Tuorto, F., Liebers, R., Musch, T., Schaefer, M., Hofmann, S., Kellner, S., Frye, M., Helm, M.,
469
Stoecklin, G. and Lyko, F. (2012) RNA cytosine methylation by Dnmt2 and NSun2 promotes
470
tRNA stability and protein synthesis. Nature Structural &Amp; Molecular Biology, 19, 900.
471
13.
Burgess, A.L., David, R. and Searle, I.R. (2015) Conservation of tRNA and rRNA 5-
472
methylcytosine in the kingdom Plantae. BMC plant biology, 15, 199.
473
14.
Schaefer, M., Pollex, T., Hanna, K. and Lyko, F. (2009) RNA cytosine methylation analysis by
474
bisulfite sequencing. Nucleic acids research, 37, e12-e12.
475
15.
Jeltsch, A., Ehrenhofer-Murray, A., Jurkowski, T.P., Lyko, F., Reuter, G., Ankri, S., Nellen, W.,
476
Schaefer, M. and Helm, M. (2016) Mechanism and biological role of Dnmt2 in Nucleic Acid
477
Methylation. RNA biology, 14, 1108-1123.
478
16.
Ponts, N., Fu, L., Harris, E.Y., Zhang, J., Chung, D.-W.D., Cervantes, M.C., Prudhomme, J.,
479
Atanasova-Penichon, V., Zehraoui, E., Bunnik, E.M. et al. (2013) Genome-wide mapping of
480
DNA methylation in the human malaria parasite Plasmodium falciparum. Cell host &
481
microbe, 14, 696-706.
482
17.
Motorin, Y. and Grosjean, H. (1999) Multisite-specific tRNA:m5C-methyltransferase (Trm4) in
483
yeast Saccharomyces cerevisiae: identification of the gene and substrate specificity of the
484
enzyme. RNA, 5, 1105-1118.
485
18.
Jeltsch, A., Nellen, W. and Lyko, F. (2006) Two substrates are better than one: dual
486
specificities for Dnmt2 methyltransferases. Trends in biochemical sciences, 31, 306-308.
487
19.
Kaiser, S., Jurkowski, T.P., Kellner, S., Schneider, D., Jeltsch, A. and Helm, M. (2016) The RNA
488
methyltransferase Dnmt2 methylates DNA in the structural context of a tRNA. RNA biology,
489
14, 1241-1251.
490
17
20.
Ponger, L. and Li, W.H. (2005) Evolutionary diversification of DNA methyltransferases in
491
eukaryotic genomes. Molecular biology and evolution, 22, 1119-1128.
492
21.
Wei, H., Jiang, S., Chen, L., He, C., Wu, S. and Peng, H. (2017) Characterization of Cytosine
493
Methylation and the DNA Methyltransferases of Toxoplasma gondii. Int J Biol Sci, 13, 458-
494
470.
495
22.
Morselli, M., Pastor, W.A., Montanini, B., Nee, K., Ferrari, R., Fu, K., Bonora, G., Rubbi, L.,
496
Clark, A.T., Ottonello, S. et al. (2015) In vivo targeting of de novo DNA methylation by
497
histone modifications in yeast and mouse. eLife, 4, e06205.
498
23.
Greer, E.L., Blanco, M.A., Gu, L., Sendinc, E., Liu, J., Aristizábal-Corrales, D., Hsu, C.-H.,
499
Aravind, L., He, C. and Shi, Y. (2015) DNA Methylation on N6-Adenine in C. elegans. Cell, 161,
500
868-878.
501
24.
Ivens, A.C., Peacock, C.S., Worthey, E.A., Murphy, L., Aggarwal, G., Berriman, M., Sisk, E.,
502
Rajandream, M.A., Adlem, E., Aert, R. et al. (2005) The genome of the kinetoplastid parasite,
503
Leishmania major. Science (New York, N.Y.), 309, 436-442.
504
25.
van Luenen, H.G.A.M., Farris, C., Jan, S., Genest, P.-A., Tripathi, P., Velds, A., Kerkhoven,
505
R.M., Nieuwland, M., Haydock, A., Ramasamy, G. et al. (2012) Glucosylated
506
hydroxymethyluracil, DNA base J, prevents transcriptional readthrough in Leishmania. Cell,
507
150, 909-921.
508
26.
Huff, J.T. and Zilberman, D. (2014) Dnmt1-independent CG methylation contributes to
509
nucleosome positioning in diverse eukaryotes. Cell, 156, 1286-1297.
510
27.
Militello, K.T., Wang, P., Jayakar, S.K., Pietrasik, R.L., Dupont, C.D., Dodd, K., King, A.M. and
511
Valenti, P.R. (2008) African Trypanosomes Contain 5-Methylcytosine in Nuclear DNA.
512
Eukaryotic Cell, 7, 2012-2016.
513
28.
Bateman, A., Smart, A., Luciani, A., Salazar, G.A., Mistry, J., Richardson, L.J., Qureshi, M., El-
514
Gebali, S., Potter, S.C., Finn, R.D. et al. (2018) The Pfam protein families database in 2019.
515
Nucleic acids research, 47, D427-D432.
516
29.
Dumetz, F., Imamura, H., Sanders, M., Seblova, V., Myskova, J., Pescher, P., Vanaerschot, M.,
517
Meehan, C.J., Cuypers, B., De Muylder, G. et al. (2017) Modulation of Aneuploidy in
518
Leishmania donovani during Adaptation to Different In Vitro and In Vivo Environments and
519
Its Impact on Gene Expression. mBio, 8.
520
30.
Berriman, M., Ghedin, E., Hertz-Fowler, C., Blandin, G., Renauld, H., Bartholomeu, D.C.,
521
Lennard, N.J., Caler, E., Hamlin, N.E., Haas, B. et al. (2005) The genome of the African
522
trypanosome Trypanosoma brucei. Science (New York, N.Y.), 309, 416-422.
523
31.
Aslett, M., Aurrecoechea, C., Berriman, M., Brestelli, J., Brunk, B.P., Carrington, M.,
524
Depledge, D.P., Fischer, S., Gajria, B., Gao, X. et al. (2010) TriTrypDB: a functional genomic
525
resource for the Trypanosomatidae. Nucleic acids research, 38, D457-462.
526
32.
Aurrecoechea, C., Brestelli, J., Brunk, B.P., Dommer, J., Fischer, S., Gajria, B., Gao, X., Gingle,
527
A., Grant, G., Harb, O.S. et al. (2009) PlasmoDB: a functional genomic database for malaria
528
parasites. Nucleic acids research, 37, D539-543.
529
33.
Auburn, S., Böhme, U., Steinbiss, S., Trimarsanto, H., Hostetler, J., Sanders, M., Gao, Q.,
530
Nosten, F., Newbold, C.I., Berriman, M. et al. (2016) A new Plasmodium vivax reference
531
sequence with improved assembly of the subtelomeres reveals an abundance of pir genes.
532
Wellcome open research, 1, 4-4.
533
34.
Gardner, M.J., Hall, N., Fung, E., White, O., Berriman, M., Hyman, R.W., Carlton, J.M., Pain,
534
A., Nelson, K.E., Bowman, S. et al. (2002) Genome sequence of the human malaria parasite
535
Plasmodium falciparum. Nature, 419, 498-511.
536
35.
Puiu, D., Enomoto, S., Buck, G.A., Abrahamsen, M.S. and Kissinger, J.C. (2004) CryptoDB: the
537
Cryptosporidium genome resource. Nucleic acids research, 32, D329-331.
538
36.
Abrahamsen, M.S., Templeton, T.J., Enomoto, S., Abrahante, J.E., Zhu, G., Lancto, C.A., Deng,
539
M., Liu, C., Widmer, G., Tzipori, S. et al. (2004) Complete genome sequence of the
540
apicomplexan, Cryptosporidium parvum. Science (New York, N.Y.), 304, 441-445.
541
18
37.
Xu, P., Widmer, G., Wang, Y., Ozaki, L.S., Alves, J.M., Serrano, M.G., Puiu, D., Manque, P.,
542
Akiyoshi, D., Mackey, A.J. et al. (2004) The genome of Cryptosporidium hominis. Nature,
543
431, 1107-1112.
544
38.
Gajria, B., Bahl, A., Brestelli, J., Dommer, J., Fischer, S., Gao, X., Heiges, M., Iodice, J.,
545
Kissinger, J.C., Mackey, A.J. et al. (2008) ToxoDB: an integrated Toxoplasma gondii database
546
resource. Nucleic acids research, 36, D553-556.
547
39.
Lorenzi, H., Khan, A., Behnke, M.S., Namasivayam, S., Swapna, L.S., Hadjithomas, M.,
548
Karamycheva, S., Pinney, D., Brunk, B.P., Ajioka, J.W. et al. (2016) Local admixture of
549
amplified and diversified secreted pathogenesis determinants shapes mosaic Toxoplasma
550
gondii genomes. Nature communications, 7, 10147.
551
40.
Ebenezer, T.E., Zoltner, M., Burrell, A., Nenarokova, A., Novák Vanclová, A.M.G., Prasad, B.,
552
Soukal, P., Santana-Molina, C., O’Neill, E., Nankissoor, N.N. et al. (2019) Transcriptome,
553
proteome and draft genome of Euglena gracilis. BMC Biology, 17, 11.
554
41.
Lorenzi, H.A., Puiu, D., Miller, J.R., Brinkac, L.M., Amedeo, P., Hall, N. and Caler, E.V. (2010)
555
New assembly, reannotation and analysis of the Entamoeba histolytica genome reveal new
556
genomic features and protein content information. PLoS Negl Trop Dis, 4, e716.
557
42.
Wood, V., Gwilliam, R., Rajandream, M.A., Lyne, M., Lyne, R., Stewart, A., Sgouros, J., Peat,
558
N., Hayles, J., Baker, S. et al. (2002) The genome sequence of Schizosaccharomyces pombe.
559
Nature, 415, 871-880.
560
43.
Galagan, J.E., Calvo, S.E., Borkovich, K.A., Selker, E.U., Read, N.D., Jaffe, D., FitzHugh, W., Ma,
561
L.J., Smirnov, S., Purcell, S. et al. (2003) The genome sequence of the filamentous fungus
562
Neurospora crassa. Nature, 422, 859-868.
563
44.
Sebaihia, M., Peck, M.W., Minton, N.P., Thomson, N.R., Holden, M.T., Mitchell, W.J., Carter,
564
A.T., Bentley, S.D., Mason, D.R., Crossman, L. et al. (2007) Genome sequence of a proteolytic
565
(Group I) Clostridium botulinum strain Hall A and comparative analysis of the clostridial
566
genomes. Genome research, 17, 1082-1092.
567
45.
Hoskins, J., Alborn, W.E., Jr., Arnold, J., Blaszczak, L.C., Burgett, S., DeHoff, B.S., Estrem, S.T.,
568
Fritz, L., Fu, D.J., Fuller, W. et al. (2001) Genome of the bacterium Streptococcus
569
pneumoniae strain R6. Journal of bacteriology, 183, 5709-5717.
570
46.
Huang, Y.Y., Cho, S.T., Lo, W.S., Wang, Y.C., Lai, E.M. and Kuo, C.H. (2015) Complete Genome
571
Sequence of Agrobacterium tumefaciens Ach5. Genome announcements, 3.
572
47.
Parkhill, J., Dougan, G., James, K.D., Thomson, N.R., Pickard, D., Wain, J., Churcher, C.,
573
Mungall, K.L., Bentley, S.D., Holden, M.T. et al. (2001) Complete genome sequence of a
574
multiple drug resistant Salmonella enterica serovar Typhi CT18. Nature, 413, 848-852.
575
48.
Murat, C., Payen, T., Noel, B., Kuo, A., Morin, E., Chen, J., Kohler, A., Krizsan, K., Balestrini, R.,
576
Da Silva, C. et al. (2018) Pezizomycetes genomes reveal the molecular basis of
577
ectomycorrhizal truffle lifestyle. Nature ecology & evolution, 2, 1956-1965.
578
49.
Kumar, S., Stecher, G., Li, M., Knyaz, C. and Tamura, K. (2018) MEGA X: Molecular
579
Evolutionary Genetics Analysis across Computing Platforms. Molecular biology and
580
evolution, 35, 1547-1549.
581
50.
Edgar, R.C. (2004) MUSCLE: multiple sequence alignment with high accuracy and high
582
throughput. Nucleic acids research, 32, 1792-1797.
583
51.
Glez-Pe�a, D., G�mez-Blanco, D., Reboiro-Jato, M., Fdez-Riverola, F. and Posada, D.
584
(2010) ALTER: program-oriented conversion of DNA and protein alignments. Nucleic acids
585
research, 38, W14-W18.
586
52.
Tihon, E., Imamura, H., Dujardin, J.C., Van Den Abbeele, J. and Van den Broeck, F. (2017)
587
Discovery and genomic analyses of hybridization between divergent lineages of
588
Trypanosoma congolense, causative agent of Animal African Trypanosomiasis. Molecular
589
ecology, 26, 6524-6538.
590
53.
LeBowitz, J.H. (1994) Transfection experiments with Leishmania. Methods in cell biology, 45,
591
65-78.
592
19
54.
Morales, M.A., Watanabe, R., Dacher, M., Chafey, P., Osorio y Fortea, J., Scott, D.A.,
593
Beverley, S.M., Ommen, G., Clos, J., Hem, S. et al. (2010) Phosphoproteome dynamics reveal
594
heat-shock protein complexes specific to the Leishmania donovani infectious stage. Proc
595
Natl Acad Sci U S A, 107, 8381-8386.
596
55.
Dumetz, F., Cuypers, B., Imamura, H., Zander, D., D'Haenens, E., Maes, I., Domagalska, M.A.,
597
Clos, J., Dujardin, J.C. and De Muylder, G. (2018) Molecular Preadaptation to Antimony
598
Resistance in Leishmania donovani on the Indian Subcontinent. mSphere, 3, e00548-00517.
599
56.
Guo, W., Fiziev, P., Yan, W., Cokus, S., Sun, X., Zhang, M.Q., Chen, P.-Y. and Pellegrini, M.
600
(2013) BS-Seeker2: a versatile aligning pipeline for bisulfite sequencing data. BMC genomics,
601
14, 774.
602
57.
Berardini, T.Z., Reiser, L., Li, D., Mezheritsky, Y., Muller, R., Strait, E. and Huala, E. (2015) The
603
arabidopsis information resource: Making and mining the “gold standard” annotated
604
reference plant genome. genesis, 53, 474-485.
605
58.
Liao, W.W., Yen, M.R., Ju, E., Hsu, F.M., Lam, L. and Chen, P.Y. (2015) MethGo: a
606
comprehensive tool for analyzing whole-genome bisulfite sequencing data. BMC genomics,
607
16 Suppl 12, S11.
608
59.
Robinson, J.T., Thorvaldsdóttir, H., Wenger, A.M., Zehir, A. and Mesirov, J.P. (2017) Variant
609
Review with the Integrative Genomics Viewer. Cancer research, 77, e31-e34.
610
60.
Decuypere, S., Vanaerschot, M., Rijal, S., Yardley, V., Maes, L., de Doncker, S., Chappuis, F.
611
and Dujardin, J.C. (2008) Gene expression profiling of Leishmania (Leishmania) donovani:
612
overcoming technical variation and exploiting biological variation. Parasitology, 135, 183-
613
194.
614
61.
Cao, X., Aufsatz, W., Zilberman, D., Mette, M.F., Huang, M.S., Matzke, M. and Jacobsen, S.E.
615
(2003) Role of the DRM and CMT3 Methyltransferases in RNA-Directed DNA Methylation.
616
Current Biology, 13, 2212-2217.
617
62.
de Mendoza, A., Bonnet, A., Vargas-Landin, D.B., Ji, N., Li, H., Yang, F., Li, L., Hori, K.,
618
Pflueger, J., Buckberry, S. et al. (2018) Recurrent acquisition of cytosine methyltransferases
619
into eukaryotic retrotransposons. Nature communications, 9, 1341.
620
63.
Dubin, M.J., Zhang, P., Meng, D., Remigereau, M.S., Osborne, E.J., Paolo Casale, F., Drewe, P.,
621
Kahles, A., Jean, G., Vilhjalmsson, B. et al. (2015) DNA methylation in Arabidopsis has a
622
genetic basis and shows evidence of local adaptation. eLife, 4, e05255.
623
64.
Cokus, S.J., Feng, S., Zhang, X., Chen, Z., Merriman, B., Haudenschild, C.D., Pradhan, S.,
624
Nelson, S.F., Pellegrini, M. and Jacobsen, S.E. (2008) Shotgun bisulphite sequencing of the
625
Arabidopsis genome reveals DNA methylation patterning. Nature, 452, 215-219.
626
65.
Hsieh, T.F., Ibarra, C.A., Silva, P., Zemach, A., Eshed-Williams, L., Fischer, R.L. and Zilberman,
627
D. (2009) Genome-wide demethylation of Arabidopsis endosperm. Science (New York, N.Y.),
628
324, 1451-1454.
629
66.
Zemach, A., McDaniel, I.E., Silva, P. and Zilberman, D. (2010) Genome-wide evolutionary
630
analysis of eukaryotic DNA methylation. Science (New York, N.Y.), 328, 916-919.
631
67.
Verma, S., Kumar, R., Katara, G.K., Singh, L.C., Negi, N.S., Ramesh, V. and Salotra, P. (2010)
632
Quantification of Parasite Load in Clinical Samples of Leishmaniasis Patients: IL-10 Level
633
Correlates with Parasite Load in Visceral Leishmaniasis. PLOS ONE, 5, e10107.
634
68.
Cuypers, B., Berg, M., Imamura, H., Dumetz, F., De Muylder, G., Domagalska, M.A., Rijal, S.,
635
Bhattarai, N.R., Maes, I. and Sanders, M. (2018) Integrated genomic and metabolomic
636
profiling of ISC1, an emerging Leishmania donovani population in the Indian subcontinent.
637
Infection, Genetics and Evolution, 62, 170-178.
638
69.
Imamura, H., Downing, T., Van den Broeck, F., Sanders, M.J., Rijal, S., Sundar, S., Mannaert,
639
A., Vanaerschot, M., Berg, M., De Muylder, G. et al. (2016) Evolutionary genomics of
640
epidemic visceral leishmaniasis in the Indian subcontinent. eLife, 5.
641
20
70.
Olova, N., Krueger, F., Andrews, S., Oxley, D., Berrens, R.V., Branco, M.R. and Reik, W. (2018)
642
Comparison of whole-genome bisulfite sequencing library preparation strategies identifies
643
sources of biases affecting DNA methylation data. Genome Biol, 19, 33-33.
644
71.
Domagalska, M.A., Imamura, H., Sanders, M., Van den Broeck, F., Bhattarai, N.R.,
645
Vanaerschot, M., Maes, I., D’Haenens, E., Rai, K., Rijal, S. et al. (2019) Genomes of
646
intracellular <em>Leishmania</em> parasites directly sequenced from patients. bioRxiv,
647
676163.
648
72.
Bewick, A.J., Vogel, K.J., Moore, A.J. and Schmitz, R.J. (2016) The evolution of DNA
649
methylation and its relationship to sociality in insects. bioRxiv, 062455.
650
73.
Oyola, S.O., Gu, Y., Manske, M., Otto, T.D., O'Brien, J., Alcock, D., Macinnis, B., Berriman, M.,
651
Newbold, C.I., Kwiatkowski, D.P. et al. (2013) Efficient depletion of host DNA contamination
652
in malaria clinical sequencing. Journal of clinical microbiology, 51, 745-751.
653
654
655
21
Tables and Figures
656
Tables
657
Table 1: CpG, CHG and CHH methylation percentages in different Leishmania donovani lines (Ld),
658
Trypanosoma brucei and Arabidopsis thaliana (positive control).
659
CpG (%)
CHG (%)
CHH (%)
LdPro
0.0003
0.0005
0.0126
LdAmas
0.0001
0.0003
0.0073
LdHamPro
0.0002
0.0005
0.0113
LdDNMT+
0.0013
0.0026
0.0627
LdDNMT-/-
0.0002
0.0006
0.0079
Tbrucei
0.0001
0.0006
0.0040
Athaliana
21.0473
4.0401
0.3141
660
Table 2: qPCR estimation of LdBPK_251230 expression level (copy number) of Ldo-Pro and Ldo-
661
DNMToverex.
662
Ldo-Pro
LdDNMT+
RNA
1.53 ±0.2
3.78 ±0.3
663
Table 3: Enrichment (X) of Leishmania DNA in artificial mixtures of Leishmania and human DNA for
664
promastigotes and amastigotes of 3 clinical isolates (BPK026, BPK275 and BPK282). Enrichments
665
were carried out with the NEBNext Microbiome DNA Enrichment Kit (NEB).
666
BPK026
BPK275
BPK282
Average Enrichment (X)
St.Dev
Promastigotes
79.85
88.32
60.47
76.22
14.28
Amastigotes
64.83
56.87
63.33
61.68
4.23
667
668
669
670
Figures
671
672
Figure 1: Protein alignment of LdDNMT (LdBPK_250018100) and TbDNMT generated with T-coffee
673
picturing the similarities between the 10 homologous domains of C5 DNA methyltransferases. Black
674
highlights homology and the red character displays the position of the catalytic cysteine residue.
675
676
677
Figure 2: RAxML Maximum Likelihood tree showing the position of Trypanosomatid DNMT (DNMT 6)
678
within the DNMT family. Displayed branch bootstrap values are based on 1000 bootstraps.
679
680
681
682
683
684
685
Figure 3: CpG, CHG and CHH genome-wide methylation patterns in A) Leishmania donovani BPK282
686
(36 chromosomes), B) Trypanosoma brucei brucei TREU927 (11 chromosomes) and C) Arabidopsis
687
thaliana Col-0 (5 chromosomes). Data was binned over 10 000 positions to remove local noise and
688
variation.
689
A) L. donovani
B) T. brucei
C) A. thaliana
24
690
Figure 4: DNA/Gene copy number based on genomic sequencing depth on chromosome 25 position
691
465000-475000. Both the LdDNMT knock-out (LdDNMT-/-) and LdDNMT overexpressor line
692
(LdDNMT+) were successful with respectively 0 and 64 copies of the gene. The plot shows also that
693
the neighbouring genes LdBPK_250018000 and LdBPK_250018200 are unaffected and have the
694
standard disomic pattern.
695
696
Figure 5: Enrichment (X) of Leishmania DNA in artificial mixtures of Leishmania promastigote DNA
697
and human DNA, with the mixtures ranging from 1:15 to 1:15000 Leishmania:human DNA.
698
Enrichments were carried out with the NEBNext Microbiome DNA Enrichment Kit (NEB) and the
699
unmethylated Leishmania DNA was enriched on average 263 times.
700
| 2020 | The absence of C-5 DNA methylation in allows DNA enrichment from complex samples | 10.1101/747063 | [
"Cuypers B",
"Dumetz F",
"Meysman P",
"Laukens K",
"De Muylder G",
"Dujardin J-C",
"Domagalska MA"
] | creative-commons |
1
A parasite’s paradise: Biotrophic species prevail oomycete
1
community composition in tree canopies
2
Running title: Biotrophic oomycetes in tree canopies
3
Robin-Tobias Jauss1*, Susanne Walden2, Anna Maria Fiore-Donno2, Stefan Schaffer3,4,
4
Ronny Wolf3, Kai Feng2,5,6, Michael Bonkowski2, Martin Schlegel1,4
5
1 University of Leipzig, Institute of Biology, Biodiversity & Evolution, Talstraße 33, 04103 Leipzig
6
2 University of Cologne, Institute of Zoology, Terrestrial Ecology, Zülpicher Straße 47b, 50674 Köln
7
3 University of Leipzig, Institute of Biology, Molecular Evolution & Animal Systematics, Talstraße 33, 04103 Leipzig
8
4 German Centre for Integrative Biodiversity Research (iDiv) Halle Jena Leipzig, Deutscher Platz 5e, 04103 Leipzig
9
5 CAS Key Laboratory for Environmental Biotechnology, Research Center for Eco-Environmental Sciences, Chinese Academy
10
of Sciences, Beijing, China
11
6 College of Resources and Environment, University of Chinese Academy of Sciences, Beijing, China
12
*To whom correspondence should be addressed. E-mail: jauss@uni-leipzig.de
13
Abstract
14
Oomycetes (Stramenopiles, Protista) are among the most severe plant pathogens,
15
comprising species with a high economic and ecologic impact on forest ecosystems. Their
16
diversity and community structures are well studied in terrestrial habitats, but tree canopies
17
as huge and diverse habitats have been widely neglected. A recent study highlighted distinct
18
oomycete communities in the canopy region compared to forest soils when taking oomycete
19
abundances into account, in contrast to the homogeneity at the incidence level. It remains
20
however unknown if this homogeneity also leads to a functional homogenisation among
21
microhabitats. In this study, we supplemented functional traits to oomycete canopy and
22
ground communities, which were determined over a time period of two years with a
23
metabarcoding approach. Our results showed that even though most oomycetes occurred in
24
all habitats, a strong discrepancy between the strata and correspondingly the distribution of
25
2
oomycete lifestyles could be observed, which was constant over time. Obligate biotrophic
26
species, exclusively feeding on living host tissue, dominated the canopy region, implying tree
27
canopies to be a hitherto neglected reservoir for parasitic protists. Parasites highly
28
specialised on hosts that were not sampled could be determined in high abundances in the
29
canopy and the surrounding air, challenging the strict host dependencies ruled for some
30
oomycetes. Our findings further contribute to the understanding of oomycete ecosystem
31
functioning in forest ecosystems.
32
Keywords: protists, oomycetes, canopies, metabarcoding, parasites, forest ecosystems
33
1 INTRODUCTION
34
Some of the most devastating plant pathogens with worldwide economic and ecologic
35
relevance belong to the Oomycota, protists in the Stramenopiles within the SAR
36
superkingdom (Adl et al., 2019). They comprise several distinct lineages, i.a. the Pythiales,
37
Peronosporales and Saprolegniales (Marano et al., 2014) and occupy ecologically important
38
positions as saprotrophs and severe pathogens. The infamous oomycete Phytophthora
39
infestans causes one of the most destructive plant diseases, the potato late blight, and
40
initiated the great Irish famine in the late 1840’s with a million deaths and massive
41
emigration (Mizubuti & Fry, 2006). The ecological and economic impact of oomycetes has
42
led to an increased research interest on their community structures (Robideau et al., 2011;
43
Riit et al., 2016; Singer et al., 2016; Jauss et al., 2020b, 2020a; Fiore-Donno & Bonkowski,
44
2021), and, correspondingly, their pathogenicity and infection strategies (Rizzo & Garbelotto,
45
2003; Rizzo et al., 2005; Thines & Kamoun, 2010).
46
Three lifestyles are described for oomycetes: Saprotrophic species are free-living and feed
47
on dead and decaying matter (Lewis, 1973). They occupy key roles in the trophic upgrading
48
of terrestrial, marine and freshwater habitats (Marano et al., 2016). Although saprotrophy is
49
less common in oomycetes, it is believed to be the ancestral state of oomycete nutrition (F.
50
3
Martin et al., 2016; Spanu & Panstruga, 2017), while the majority of currently described
51
oomycetes are plant pathogens (Thines & Kamoun, 2010). The pathogenic lifestyles include
52
hemibiotrophy, characterised by an initial biotrophic phase later turning into a necrotrophic
53
phase after the death of the host (Fawke et al., 2015; Pandaranayaka et al., 2019), as well
54
as obligate biotrophy, which comprises species exclusively feeding on living host tissue
55
(Spanu & Kämper, 2010). Even though obligate biotrophic species usually do not actively kill
56
their host, they still damage the host by chlorosis, inflorescence and killing of seedlings, and
57
thus cause severe economic losses (Parkunan et al., 2013; Krsteska et al., 2014; Kamoun et
58
al., 2015).
59
Oomycete communities are well studied in terrestrial habitats, however, most studies focus
60
on soil and the rhizosphere (Arcate et al., 2006; Esmaeili Taheri et al., 2017; Sapp et al.,
61
2019; Fiore-Donno & Bonkowski, 2021). Recently, Jauss et al. (2020b) characterised
62
oomycete diversity and community composition in tree canopies, which are huge
63
ecosystems containing heterogeneous microhabitats and a large proportion of undescribed
64
diversity (Nadkarni, 2001). Albeit the same oomycetes were present on the ground and in
65
the canopy, communities inhabiting canopy habitats were significantly distinct from soil and
66
leaf litter communities in their abundances. The authors concluded that oomycete diversity in
67
forest ecosystems is shaped by deterministic microhabitat filtering, while a study by Jauss et
68
al. (2020a) could determine air dispersal and convective transport to be the stochastic
69
supplier and distributor of oomycetes among microhabitats and strata. However, the former
70
study only analysed one time point, while the latter study dealing with air samples could
71
show a strong temporal variability in community composition. Accordingly, seasonal
72
variability has been shown to influence protistan communities, to some extent, in several
73
studies (Nolte et al., 2010; Fiore-Donno et al., 2019; Fournier et al., 2020; Walden et al.,
74
2021). For cercozoan communities, Walden et al. (2021) could show annually reoccurring
75
succession patterns in the phyllosphere. This implied not only spatially, but also seasonally
76
structured cercozoan communities in tree canopies, although this was not reflected on a
77
4
functional scale. If seasonal variation is also reflected in the functional diversity of oomycetes
78
in forest ecosystems, however, remains elusive.
79
Accordingly, we supplemented functional traits and investigated the seasonal stability of
80
oomycete community composition in forest floors and tree canopies over a period of two
81
years. Our study tackles two hypotheses: (1) Oomycete communities vary not only in their
82
spatial distribution, but also in their seasonal composition, and (2) the deterministic
83
processes leading to differences in community composition between canopy and ground
84
habitats also shape the functional diversity and functional distribution among microhabitats.
85
2 MATERIAL AND METHODS
86
2.1 Sampling, DNA extraction and sequencing
87
Microhabitat samples were collected in two seasons over a period of two years, i.e. autumn
88
(October) 2017 and 2018 and spring (May) 2018 and 2019 in cooperation with the Leipzig
89
Canopy Crane (LCC) Facility in a floodplain forest in Leipzig, Germany (51.3657 N, 12.3094
90
E). Samples were obtained and processed as described in Jauss et al. (2020b). Briefly,
91
seven microbial microhabitat compartments related to tree surface were sampled in the
92
canopy at 20-30m height: Fresh leaves, dead wood, bark, arboreal soil and three cryptogam
93
epiphytes (lichen and two moss genera, Hypnum and Orthotrichum). In addition, two ground
94
samples (soil and leaf litter) were sampled. All microhabitat samples were taken with four
95
replicates, from three tree species with three replicates each. DNA extraction was performed
96
with the DNeasy PowerSoil kit (QIAGEN, Hilden, Germany) according to the manufacturer's
97
instruction. This procedure was performed on four sampling dates: October 2017 (Jauss et
98
al., 2020b), May 2018, October 2018 and May 2019 (this study). Oomycete-specific PCRs
99
and sequencing were performed as described in Jauss et al. (2020b) with tagged primers
100
5
designed by Fiore-Donno & Bonkowski (2021); the used primer tag combinations are
101
provided in Supplementary Table 1.
102
2.2 Sequence processing
103
Sequence processing and bioinformatics analyses followed the pipeline described in Jauss
104
et al. (2020b). Briefly, raw reads were merged using VSEARCH v2.10.3 (Rognes et al.,
105
2016) and demultiplexed with cutadapt v1.18 (M. Martin, 2011). Primer and tag sequences
106
were trimmed and concatenated sequencing runs were then clustered into operational
107
taxonomic units (OTUs) using Swarm v2.2.2 (Mahé et al., 2015). Chimeras were de novo
108
detected using VSEARCH. OTUs were removed from the final OTU table if they were
109
flagged as chimeric, showed a quality value of less than 0.0002, were shorter than 150bp, or
110
were represented by less than 0.005% of all reads (i.e. 368 reads). OTUs were first
111
taxonomically assigned by using BLAST+ v2.9.0 (Camacho et al., 2009) with default
112
parameters against the non-redundant NCBI Nucleotide database (as of June 2019) and
113
removed if the best hit in terms of bitscore was a non-oomycete sequence. Finer taxonomic
114
assignment was performed with VSEARCH on a custom oomycete ITS1 database (Jauss et
115
al., 2020b). The annotation was refined by assigning the species name of the best
116
VSEARCH hit to the corresponding OTU if the pairwise identity was over 95%, OTUs with
117
lower percentages were assigned higher taxonomic levels. Functional annotation was
118
performed on genus level with a custom python script, based on the oomycete functional
119
database published by Fiore-Donno & Bonkowski (2021). Samples with low sequencing
120
depth were removed by loading the final OTU table into QIIME 2 v2018.11 (Bolyen et al.,
121
2019) and retaining at least five samples per microhabitat and 15 samples per tree species
122
per sampling date, i.e. samples with at least 1172 reads. Additionally, the oomycete OTU
123
abundance matrix of air samples from Jauss et al. (2020a) was used for a comparison
124
between tree related microhabitats and the surrounding air from spring 2019, as these
125
samples were taken simultaneously.
126
6
2.3 Statistical analyses
127
All statistical analyses were conducted in R v3.5.3 (R Core Team, 2019). Alpha diversity
128
indices were calculated for each sample using the diversity function in the vegan package
129
(Oksanen et al., 2019). Non-metric multidimensional scaling was performed on the Bray-
130
Curtis dissimilarity matrix of the log transformed relative abundances (functions vegdist and
131
metaMDS in the vegan package, respectively), the same matrix was used for a
132
permutational multivariate analysis of variance (permANOVA) with the adonis function.
133
Partitioning and visualisation of relative abundances between canopy, soil and leaf litter was
134
performed with the ggtern package (Hamilton & Ferry, 2018). Determination of significantly
135
differentially abundant OTUs was performed with the DESeq2 package (Love et al., 2014).
136
All figures were plotted with the ggplot2 package (Wickham, 2016).
137
3 RESULTS
138
3.1 Taxonomic and functional annotation
139
We obtained 375 OTUs from 4,262,960 sequences. 77 OTUs (= 20.5% of all OTUs) showed
140
a sequence similarity of less than 70% to any known reference sequence. Plotting the
141
sequence similarity against reference sequences revealed similar patterns as previously
142
described by Jauss et al. (2020b), i.e., many OTUs showed a similarity of 97-100% to known
143
reference sequences, while additional peaks at ~75% and ~85% may indicate hitherto
144
undescribed oomycete lineages (Supplementary Figure 1).
145
Peronosporales and Pythiales dominated all microhabitats at all sampling events
146
(Supplementary Figure 2). Distribution of functional groups was relatively constant for all four
147
sampling events (Figure 1). Based on OTU presence/absence, the pattern was nearly
148
identical for all microhabitats (Figure 1A-D). Approximately 20% of all OTUs occupied a
149
hemibiotrophic lifestyle, 30% were determined to be obligate biotrophic, only few OTUs
150
7
belonged to saprotrophic species and the lifestyle of the remaining 50% of OTUs could not
151
be determined, mainly due to low sequence similarities to reference sequences. However,
152
when taking abundances of OTUs into account, the pattern clearly shifted. OTUs assigned to
153
obligate biotrophic species dominated canopy habitats, while ground habitats were more
154
dominated by hemibiotrophic species (Figure 1E-H).
155
Comparing the data from Spring 2019 (Figure 1D,H) with air samples previously published
156
by Jauss et al. (2020a) (Figure 2) revealed that the air surrounding canopy and ground
157
habitats was dominated by obligate biotrophic OTUs, irrespective of incidence or
158
abundance.
159
3.2 Abundance partitioning
160
3.2.1 Partitioning between Canopy, Soil and Leaf Litter
161
To further determine the distribution of functional groups together with the taxonomic
162
annotation, the relative abundances of each OTU were partitioned for canopy, soil and leaf
163
litter samples (Figure 3). Again, OTUs assigned to obligate biotrophic species dominated
164
canopy samples, while hemibiotrophic species were more evenly distributed or more
165
abundant in leaf litter and soil habitats. Albuginales were almost exclusively present in
166
canopy samples, Peronosporales dominated canopy and leaf litter samples, while Pythiales
167
showed a rather even distribution.
168
The relative abundances of the latter two orders were further partitioned into the four
169
sampling events (Supplementary Figure 3). Abundances of Pythiales were rather
170
homogenous and consistent throughout the seasons, while Peronosporales abundances
171
were more shifted to the canopy region in spring samples. In Autumn 2017, OTUs assigned
172
to the Peronosporales were almost exclusively present in canopy and leaf litter samples,
173
while the distribution in Autumn 2018 was more homogenous.
174
8
3.2.2 Differential Abundance Analysis
175
To determine which OTU abundances were significantly different between the two strata
176
ground and canopy as well as the two sampling seasons spring and autumn, a differential
177
abundance analysis was carried out (Figure 4, Supplementary Figure 4). Within the
178
Peronosporales, this revealed the genera Peronospora and Hyaloperonospora (obligate
179
biotrophic genera) to be the dominant taxa in canopy samples, while Phytophthora
180
(hemibiotrophic) species were significantly differentially abundant in ground samples (Figure
181
4). For the seasonal effect, more Peronospora species were differentially abundant in spring
182
samples compared to autumn samples (Supplementary Figure 4). Within the Pythiales, the
183
genera Pythium (hemibiotrophic) and Globisporangium (obligate biotrophic) were
184
significantly differentially abundant in ground samples. Most Pythiales, however, could not
185
be determined due to the low sequence similarity to reference sequences.
186
3.3 Alpha and beta diversity
187
Despite OTU richness being quite variable among microhabitats, Shannon diversity as well
188
as evenness were high and did not differ between the samplings (Supplementary Figure 5).
189
Beta diversity analyses revealed similar patterns for all seasons as well: the NMDS plot
190
(Figure 5) showed a large overlap of canopy inhabiting communities, which in turn did not
191
overlap with leaf litter and soil communities. This indicated distinct communities inhabiting
192
canopy and ground habitats, respectively, a pattern recurring in all samplings.
193
Variation in community composition was twice as high among microhabitats (R²=0.20) than
194
between canopy and ground (R²=0.11) or sampling dates (R²=0.10). Tree species (R²=0.05)
195
and season (R²=0.04) explained only a minor fraction of beta diversity (permANOVA, Table
196
1).
197
9
4 DISCUSSION
198
The most striking pattern of oomycete community composition is the distribution of obligate
199
biotrophic and hemibiotrophic species, with the former dominating canopy habitats and the
200
latter predominantly found in ground habitats (Figure 1). In a previous study, Jauss et al.
201
(2020b) proposed increasing functional diversity instead of increasing species richness with
202
increasing habitat diversity, as most OTUs were shared between all habitats irrespective of
203
specific strata or tree species. Here we supplemented functional traits of the detected OTUs,
204
which revealed that the observed diversity is driven by the lifestyle of the oomycetes.
205
Species occupying a hemibiotrophic lifestyle dominated the two ground habitats soil and leaf
206
litter. Hemibiotrophy is characterised by an initial biotrophic phase, which turns into a
207
necrotrophic phase (Fawke et al., 2015; Pandaranayaka et al., 2019). Oomycetes dwelling
208
the ground habitats are thus capable of feeding on the dead organic matter in the soil, leaf
209
litter and deadwood samples. Deadwood on the forest floor has already been shown to
210
harbour hemibiotrophic oomycetes (Kwaśna et al. 2017a; 2017b). In the canopy, however,
211
deadwood harbours only little hemibiotrophic species, as they are dominated by obligate
212
biotrophic species, like the other canopy habitats. The reason for this might be the high
213
number of obligate biotrophs in the other surrounding canopy habitats as well as in the air
214
(Figure 2). These samples might be overwhelmed by the passive influx of biotrophic species,
215
which are capable of surviving in the other, living, habitats, which would be an interplay
216
between stochastic and deterministic processes for community assembly.
217
Recent molecular studies analysing oomycete diversity determined similar patterns as
218
reflected in our study, i.e. soil habitats are dominated by hemibiotrophic species, mostly
219
members of the Pythiales (Sapkota & Nicolaisen, 2015; Riit et al., 2016; Fiore-Donno &
220
Bonkowski, 2021). Species of the genus Pythium were significantly differentially abundant in
221
our ground habitats. Habitats in the canopy, however, were dominated by the obligate
222
biotrophic genera Peronospora and Hyaloperonospora (Figure 4). Tree canopies have only
223
recently been subject to studies on microbial diversity (Jauss et al., 2020a, 2020b; Walden et
224
10
al., 2021; Herrmann et al., 2021), indicating tree canopies to be a hitherto neglected
225
reservoir for parasitic microorganisms. Species of the genus Hyaloperonospora are known to
226
be highly host-specific, infecting plant species of Brassicaceae and closely related families
227
(Lee et al., 2017 and references therein). However, none of our sampled trees and
228
microhabitats belong to the Brassicaceae or the order Brassicales. Yet, we observed a high
229
number of reads and OTUs assigned to the genus Hyaloperonospora in the microhabitat
230
samples in the canopy as well as in the air samples in both strata, while their number in
231
ground microhabitats is significantly depleted (Figure 4). This indicates a non-random
232
distribution of Hyaloperonospora species, as the air as a distribution mechanism should lead
233
to a more or less equal distribution in canopy and ground habitats. Here, they should not be
234
able to survive due to their high host specificity. But the domination in canopy samples
235
implies a capability of survival on hosts they are not specialised on. Thus, we tentatively
236
propose an even less strict host dependency for the genus Hyaloperonospora than already
237
suggested (Yerkes & Shaw, 1959; McMeekin, 1960; Dickinson & Greenhalgh, 1977).
238
The significant differential abundance in the canopy of several undetermined OTUs that can
239
only be assigned to the family Pythiaceae (Figure 4) indicates hitherto undescribed lineages,
240
specialised on the survival in the canopy. Members of the Pythiaceae can occupy all
241
lifestyles, from saprotrophy over hemibiotrophy to obligate biotrophy (Fawke et al., 2015;
242
Marano et al., 2016; Fiore-Donno & Bonkowski, 2021). If the OTUs in the canopy would
243
show an obligate biotrophic lifestyle, it would be in line with observations of the other
244
lineages in the canopy (Figure 1). Yet, the sequence similarity of these OTUs amounts to
245
only ca 80-85% to any reference sequence, thus we only tentatively draw conclusions about
246
their lifestyle.
247
A common pattern in microbial community ecology studies is a high seasonal variability
248
(Nolte et al., 2010; Fiore-Donno et al., 2019; Fournier et al., 2020; Walden et al., 2021).
249
Oomycete community compositions were in fact slightly, yet significantly distinct for every
250
sampling and correspondingly for every season (Table 1). This pattern is in line with
251
hypotheses proposed by Jauss et al. (2020a), that seasonal variation in air samples drives
252
11
the community composition in forest ecosystems. The environment, however, then selects
253
the species most adapted to the microhabitat, leading to overall similar community patterns
254
and microhabitat differences for every season (Figure 5). The seasonal changes in
255
microhabitat properties (e.g. temperature, moisture or habitat structure) thus affect all
256
habitats and communities equally. The season itself explained less variance in community
257
composition than the sampling dates (i.e., Autumn 2017 vs. Autumn 2018 etc.; Table 1),
258
suggesting that annual changes do not lead to similar community structures within
259
microhabitats in each season as an annual cycle per se, but rather indicate a high temporal
260
variability while preserving spatial diversity. Fournier et al. (2020) observed similar patterns,
261
concluding deterministic niche-based processes in microbial forest soil community assembly.
262
Implications are that ecosystem functioning of oomycete communities is not mainly affected
263
by seasonal fluctuations, but rather by microhabitat identity and, correspondingly, responses
264
of lifestyle to microhabitat filtering (Fiore-Donno & Bonkowski, 2021).
265
Conclusions
266
Both our hypotheses were confirmed in this study: Oomycetes show not only a spatial, but,
267
to a lesser extent, also a temporal variation in their communities. Within the temporal
268
variation however, the spatial variation is preserved, leading to overall similar community
269
patterns for every sampling date. Further, these deterministic processes also shape their
270
functional diversity in forest ecosystems. Our results indicate that tree canopies not only
271
offer numerous distinct habitats to microorganisms, but also serve as a reservoir for parasitic
272
species. Spatial diversity and correspondingly functional diversity drive the oomycete
273
community to a greater extent than temporal diversity. Thus, our findings contribute to future
274
studies on oomycete ecosystem functioning.
275
12
Funding
276
This work was supported by the Priority Program SPP 1991: Taxon-omics − New
277
Approaches for Discovering and Naming Biodiversity of the German Research Foundation
278
(DFG) with funding to MB (1907/19-1) and MS (Schl 229/20-1). We acknowledge support
279
from the Leipzig University Library for open access publishing.
280
Acknowledgements
281
The authors would like to thank Rolf Engelmann for his assistance with the field work by
282
operating the canopy crane, as well as the Leipzig Canopy Crane Platform of the German
283
Centre for Integrative Biodiversity Research (iDiv) for providing the site access and allowing
284
us to sample the trees from their field trial.
285
Conflict of Interest
286
None declared.
287
References
288
Adl, S. M., Bass, D., Lane, C. E., Lukeš, J., Schoch, C. L., Smirnov, A., Agatha, S., Berney,
289
C., Brown, M. W., Burki, F., Cárdenas, P., Čepička, I., Chistyakova, L., del Campo, J.,
290
Dunthorn, M., Edvardsen, B., Eglit, Y., Guillou, L., Hampl, V., … Zhang, Q. (2019).
291
Revisions to the Classification, Nomenclature, and Diversity of Eukaryotes. Journal of
292
Eukaryotic Microbiology, 66, 4–119. https://doi.org/10.1111/jeu.12691
293
Arcate, J. M., Karp, M. A., & Nelson, E. B. (2006). Diversity of Peronosporomycete
294
(oomycete) communities associated with the rhizosphere of different plant species.
295
Microbial Ecology, 51(1), 36–50. https://doi.org/10.1007/s00248-005-0187-y
296
Bolyen, E., Rideout, J. R., Dillon, M. R., Bokulich, N. A., Abnet, C. C., Al-Ghalith, G. A.,
297
Alexander, H., Alm, E. J., Arumugam, M., Asnicar, F., Bai, Y., Bisanz, J. E., Bittinger,
298
K., Brejnrod, A., Brislawn, C. J., Brown, C. T., Callahan, B. J., Caraballo-Rodríguez, A.
299
M., Chase, J., … Caporaso, J. G. (2019). Reproducible, interactive, scalable and
300
extensible microbiome data science using QIIME 2. Nature Biotechnology, 37(8), 852–
301
857. https://doi.org/10.1038/s41587-019-0209-9
302
Camacho, C., Coulouris, G., Avagyan, V., Ma, N., Papadopoulos, J., Bealer, K., & Madden,
303
13
T. L. (2009). BLAST+: Architecture and applications. BMC Bioinformatics, 10.
304
https://doi.org/10.1186/1471-2105-10-421
305
Dickinson, C. H., & Greenhalgh, J. R. (1977). Host range and taxonomy of Peronospora on
306
crucifers. Transactions of the British
Mycological Society, 69(1), 111–116.
307
https://doi.org/10.1016/s0007-1536(77)80121-6
308
Esmaeili Taheri, A., Chatterton, S., Gossen, B. D., & McLaren, D. L. (2017). Metagenomic
309
analysis of oomycete communities from the rhizosphere of field pea on the Canadian
310
prairies. Canadian Journal of Microbiology, 63(9), 758–768. https://doi.org/10.1139/cjm-
311
2017-0099
312
Fawke, S., Doumane, M., & Schornack, S. (2015). Oomycete Interactions with Plants:
313
Infection Strategies and Resistance Principles. Microbiology and Molecular Biology
314
Reviews, 79(3), 263–280. https://doi.org/10.1128/mmbr.00010-15
315
Fiore-Donno, A. M., & Bonkowski, M. (2021). Different community compositions between
316
obligate and facultative oomycete plant parasites in a landscape-scale metabarcoding
317
survey. Biology and Fertility of Soils, 57(2), 245–256. https://doi.org/10.1007/s00374-
318
020-01519-z
319
Fiore-Donno, A. M., Richter-Heitmann, T., Degrune, F., Dumack, K., Regan, K. M., Marhan,
320
S., Boeddinghaus, R. S., Rillig, M. C., Friedrich, M. W., Kandeler, E., & Bonkowski, M.
321
(2019). Functional traits and spatio-temporal structure of a major group of soil protists
322
(rhizaria: Cercozoa) in a temperate grassland. Frontiers in Microbiology, 10(JUN).
323
https://doi.org/10.3389/fmicb.2019.01332
324
Fournier, B., Samaritani, E., Frey, B., Seppey, C. V. W., Lara, E., Heger, T. J., & Mitchell, E.
325
A. D. (2020). Higher spatial than seasonal variation in floodplain soil eukaryotic
326
microbial
communities.
Soil
Biology
and
Biochemistry,
147,
107842.
327
https://doi.org/10.1016/j.soilbio.2020.107842
328
Hamilton, N. E., & Ferry, M. (2018). Ggtern: Ternary diagrams using ggplot2. Journal of
329
Statistical Software, 87(1), 1–17. https://doi.org/10.18637/jss.v087.c03
330
Herrmann, M., Geesink, P., Richter, R., & Küsel, K. (2021). Canopy Position Has a Stronger
331
Effect than Tree Species Identity on Phyllosphere Bacterial Diversity in a Floodplain
332
Hardwood Forest. Microbial Ecology, 81(1), 157–168. https://doi.org/10.1007/s00248-
333
020-01565-y
334
Jauss, R.-T., Nowack, A., Walden, S., Wolf, R., Schaffer, S., Schellbach, B., Bonkowski, M.,
335
& Schlegel, M. (2020a). To the canopy and beyond: Air samples reveal wind dispersal
336
as a driver of ubiquitous protistan pathogen assembly in tree canopies. BioRxiv,
337
2020.11.30.405688. https://doi.org/10.1101/2020.11.30.405688
338
Jauss, R.-T., Walden, S., Fiore-Donno, A. M., Dumack, K., Schaffer, S., Wolf, R., Schlegel,
339
M., & Bonkowski, M. (2020b). From forest soil to the canopy: Increased habitat diversity
340
does not increase species richness of Cercozoa and Oomycota in tree canopies.
341
Frontiers in Microbiology, 11, 3364. https://doi.org/10.3389/FMICB.2020.592189
342
Kamoun, S., Furzer, O., Jones, J. D. G., Judelson, H. S., Ali, G. S., Dalio, R. J. D., Roy, S.
343
G., Schena, L., Zambounis, A., Panabières, F., Cahill, D., Ruocco, M., Figueiredo, A.,
344
Chen, X. R., Hulvey, J., Stam, R., Lamour, K., Gijzen, M., Tyler, B. M., … Govers, F.
345
(2015). The Top 10 oomycete pathogens in molecular plant pathology. Molecular Plant
346
Pathology, 16(4), 413–434. https://doi.org/10.1111/mpp.12190
347
Krsteska, V., Dimeska, V., Stojkov, S., & Stojanoski, P. (2014). Peronospora Tabacina A.
348
The causing agent of blue mold disease on Tobacco. Bulgarian Journal of Agricultural
349
Science, 21(1), 126–133. http://eprints.uklo.edu.mk/312/
350
Kwaśna, H., Mazur, A., Kuźmiński, R., Jaszczak, R., Turski, M., Behnke-Borowczyk, J.,
351
Adamowicz, K., & Łakomy, P. (2017a). Abundance and diversity of wood-decay fungi in
352
managed and unmanaged stands in a Scots pine forest in western Poland. Forest
353
Ecology and Management, 400, 438–446. https://doi.org/10.1016/j.foreco.2017.04.023
354
Kwaśna, H., Mazur, A., Łabędzki, A., Kuźmiński, R., & Łakomy, P. (2017b). Communities of
355
fungi in decomposed wood of oak and pine. Forest Research Papers, 77(3), 261–275.
356
https://doi.org/10.1515/frp-2016-0028
357
Lee, J. S., Lee, H. B., Shin, H. D., & Choi, Y. J. (2017). Diversity, phylogeny, and host-
358
14
specialization of Hyaloperonospora species in Korea. Mycobiology, 45(3), 139–149.
359
https://doi.org/10.5941/MYCO.2017.45.3.139
360
Lewis, D. H. (1973). Concepts in fungal nutrition and the origin of biotrophy. Biological
361
Reviews, 48(2), 261–277. https://doi.org/10.1111/j.1469-185x.1973.tb00982.x
362
Love, M. I., Huber, W., & Anders, S. (2014). Moderated estimation of fold change and
363
dispersion for RNA-seq data with DESeq2. Genome Biology, 15(12), 550.
364
https://doi.org/10.1186/s13059-014-0550-8
365
Mahé, F., Rognes, T., Quince, C., de Vargas, C., & Dunthorn, M. (2015). Swarmv2: Highly-
366
scalable
and
high-resolution
amplicon
clustering.
PeerJ,
2015(12).
367
https://doi.org/10.7717/peerj.1420
368
Marano, A. V., Jesus, A. L., de Souza, J. I., Jerônimo, G. H., Gonçalves, D. R., Boro, M. C.,
369
Rocha, S. C. O., & Pires-Zottarelli, C. L. A. (2016). Ecological roles of saprotrophic
370
Peronosporales (Oomycetes, Straminipila) in natural environments. Fungal Ecology, 19,
371
77–88. https://doi.org/10.1016/j.funeco.2015.06.003
372
Marano, A. V, Jesus, A. L., Pires-Zottarelli, C. L. A., James, T. Y., Gleason, F. H., & de
373
Souza, J. I. (2014). Phylogenetic relationships of Pythiales and Peronosporales
374
(Oomycetes, Straminipila) within the “peronosporalean galaxy’’’.” In K. Jones, EBG and
375
Hyde, KD and Pang (Ed.), Freshwater fungi and fungal-like organisms (pp. 177–199).
376
WALTER
DE
GRUYTER
GMBH.
377
https://doi.org/https://doi.org/10.1515/9783110333480.177
378
Martin, F., Kohler, A., Murat, C., Veneault-Fourrey, C., & Hibbett, D. S. (2016). Unearthing
379
the roots of ectomycorrhizal symbioses. In Nature Reviews Microbiology (Vol. 14, Issue
380
12, pp. 760–773). Nature Publishing Group. https://doi.org/10.1038/nrmicro.2016.149
381
Martin, M. (2011). Cutadapt removes adapter sequences from high-throughput sequencing
382
reads. EMBnet.Journal, 17(1), 10. https://doi.org/10.14806/ej.17.1.200
383
McMeekin, D. (1960). The role of the oospores of Peronospora parasitica in downy mildew
384
of
crucifera.
Phytopathology,
50(2),
93-97
pp.
385
https://www.cabdirect.org/cabdirect/abstract/19601101892
386
Mizubuti, E. S. ., & Fry, W. E. (2006). Potato late blight. In The Epidemiology of Plant
387
Diseases (pp. 445–471). Kluwer Academic Publishers. https://doi.org/10.1007/1-4020-
388
4581-6_17
389
Nadkarni, N. M. (2001). Enhancement of forest canopy research, education, and
390
conservation in the new millennium. In K. E. Linsenmair, A. J. Davis, B. Fiala, & M. R.
391
Speight (Eds.), Tropical Forest Canopies: Ecology and Management (pp. 361–367).
392
Springer. https://doi.org/10.1007/978-94-017-3606-0_29
393
Nolte, V., Pandey, R. V., Jost, S., Medinger, R., Ottenwälder, B., Boenigk, J., & Schlötterer,
394
C. (2010). Contrasting seasonal niche separation between rare and abundant taxa
395
conceals the extent of protist diversity. Molecular Ecology, 19(14), 2908–2915.
396
https://doi.org/10.1111/j.1365-294X.2010.04669.x
397
Oksanen, J., Blanchet, F. G., Friendly, M., Kindt, R., Legendre, P., McGlinn, D., Minchin, P.
398
R., O’Hara, R. B., Simpson, G. L., Solymos, P., Stevens, M. H. H., Szoecs, E., &
399
Wagner,
H.
(2019).
vegan:
Community
Ecology
Package.
https://cran.r-
400
project.org/package=vegan
401
Pandaranayaka, E. P. J., Frenkel, O., Elad, Y., Prusky, D., & Harel, A. (2019). Network
402
analysis exposes core functions in major lifestyles of fungal and oomycete plant
403
pathogens. BMC Genomics, 20(1), 1–15. https://doi.org/10.1186/s12864-019-6409-3
404
Parkunan, V., Gitaitis, R. D., Dutta, B., Langston, D. B., & Ji, P. (2013). An Epidemic of
405
Downy Mildew caused by Peronospora destructor on Vidalia Sweet Onions in Georgia
406
in 2012. Plant Health Progress, 14(1), 54. https://doi.org/10.1094/php-2013-0328-01-br
407
R Core Team. (2019). R: A Language and Environment for Statistical Computing.
408
https://www.r-project.org/
409
Riit, T., Tedersoo, L., Drenkhan, R., Runno-Paurson, E., Kokko, H., & Anslan, S. (2016).
410
Oomycete-specific ITS primers for identification and metabarcoding. MycoKeys, 14, 17–
411
30. https://doi.org/10.3897/mycokeys.14.9244
412
Rizzo, D. M., & Garbelotto, M. (2003). Sudden Oak Death: Endangering California and
413
15
Oregon Forest Ecosystems. Frontiers in Ecology and the Environment, 1(4), 197.
414
https://doi.org/10.2307/3868064
415
Rizzo, D. M., Garbelotto, M., & Hansen, E. M. (2005). Phytophthora ramorum: Integrative
416
Research and Management of an Emerging Pathogen in California and Oregon
417
Forests.
Annual
Review
of
Phytopathology,
43(1),
309–335.
418
https://doi.org/10.1146/annurev.phyto.42.040803.140418
419
Robideau, G. P., De Cock, A. W. A. M., Coffey, M. D., Voglmayr, H., Brouwer, H., Bala, K.,
420
Chitty, D. W., Désaulniers, N., Eggertson, Q. A., Gachon, C. M. M., Hu, C. H., Küpper,
421
F. C., Rintoul, T. L., Sarhan, E., Verstappen, E. C. P., Zhang, Y., Bonants, P. J. M.,
422
Ristaino, J. B., & André Lévesque, C. (2011). DNA barcoding of oomycetes with
423
cytochrome c oxidase subunit I and internal transcribed spacer. Molecular Ecology
424
Resources, 11(6), 1002–1011. https://doi.org/10.1111/j.1755-0998.2011.03041.x
425
Rognes, T., Flouri, T., Nichols, B., Quince, C., & Mahé, F. (2016). VSEARCH: A versatile
426
open source tool for metagenomics. PeerJ, 2016(10). https://doi.org/10.7717/peerj.2584
427
Sapkota, R., & Nicolaisen, M. (2015). An improved high throughput sequencing method for
428
studying oomycete communities. Journal of Microbiological Methods, 110, 33–39.
429
https://doi.org/10.1016/j.mimet.2015.01.013
430
Sapp, M., Tyborski, N., Linstädter, A., López Sánchez, A., Mansfeldt, T., Waldhoff, G.,
431
Bareth, G., Bonkowski, M., & Rose, L. E. (2019). Site-specific distribution of oak
432
rhizosphere-associated oomycetes revealed by cytochrome c oxidase subunit II
433
metabarcoding.
Ecology
and
Evolution,
9(18),
10567–10581.
434
https://doi.org/10.1002/ece3.5577
435
Singer, D., Lara, E., Steciow, M. M., Seppey, C. V. W., Paredes, N., Pillonel, A., Oszako, T.,
436
& Belbahri, L. (2016). High-throughput sequencing reveals diverse oomycete
437
communities in oligotrophic peat bog micro-habitat. Fungal Ecology, 23, 42–47.
438
https://doi.org/10.1016/j.funeco.2016.05.009
439
Spanu, P., & Kämper, J. (2010). Genomics of biotrophy in fungi and oomycetes-emerging
440
patterns. In Current Opinion in Plant Biology (Vol. 13, Issue 4, pp. 409–414). Elsevier
441
Current Trends. https://doi.org/10.1016/j.pbi.2010.03.004
442
Spanu, P., & Panstruga, R. (2017). Editorial: Biotrophic plant-microbe interactions. In
443
Frontiers in Plant Science (Vol. 8, p. 192). Frontiers Research Foundation.
444
https://doi.org/10.3389/fpls.2017.00192
445
Thines, M., & Kamoun, S. (2010). Oomycete-plant coevolution: Recent advances and future
446
prospects. In Current Opinion in Plant Biology (Vol. 13, Issue 4, pp. 427–433). Curr
447
Opin Plant Biol. https://doi.org/10.1016/j.pbi.2010.04.001
448
Walden, S., Jauss, R.-T., Feng, K., Fiore-Donno, A. M., Schaffer, S., Wolf, R., Schlegel, M.,
449
& Bonkowski, M. (2021). On the phenology of protists: Recurrent patterns reveal
450
seasonal variation of protistan (Rhizaria: Cercozoa, Endomyxa) communities in tree
451
canopies. BioRxiv, 2021.02.15.431229. https://doi.org/10.1101/2021.02.15.431229
452
Wickham, H. (2016). ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New
453
York. https://ggplot2.tidyverse.org
454
Yerkes, W. D., & Shaw, C. G. (1959). Taxonomy of the Peronospora species on Cruciferae
455
and
Chenopodiaceae.
Phytopathology,
49(8),
499-507
pp.
456
https://www.cabdirect.org/cabdirect/abstract/19601100570
457
458
16
Data Accessibility
459
Raw sequence data have been submitted to the European Nucleotide Archive (ENA)
460
database under the Bioproject number PRJEB37525, with accession numbers ERS4399744,
461
ERS5649966 and ERS5649967.
462
All figures, codes and detailed bioinformatic/statistical methods used in this study are
463
available at https://github.com/RJauss/ParasitesParadise.
464
Author contributions
465
MB and MS conceived the study. RW and StS designed the sampling and DNA extraction.
466
AMF-D contributed the primers and functional annotation of oomycetes. SW and R-TJ
467
conducted the sampling, DNA extraction and PCRs. KF assisted DNA extraction and PCRs.
468
RT-J performed the bioinformatic and statistical analyses and drafted the manuscript. All
469
authors contributed to and approved the final version.
470
Tables
471
Table 1: Results of permutational multivariate analysis of variance (permANOVA) from the
472
adonis function. Factors were used independently with the default of 999 permutations. Season
473
provides the two factors Autumn and Spring, while Sampling Date corresponds to the specific time
474
points of sampling, i.e. Autumn 2017, Spring 2018 etc.
475
Df
SumsOfSqs
F value
R2
p
Tree Species
2
5.18
7.95
0.05
0.001
Microhabitat
8
20.45
9.12
0.20
0.001
Stratum
1
10.78
35.20
0.11
0.001
Season
1
4.00
12.15
0.04
0.001
Sampling Date
3
10.32
11.10
0.10
0.001
476
17
Figures
477
Figure 1: Functional annotation of oomycete OTUs in canopy and ground habitats. (A-D)
478
Distribution of functional groups based on OTU presence/absence, i.e. the proportion of OTUs per
479
Lifestyle. (E-H) Distribution of functional groups when taking abundances into account. A = Arboreal
480
Soil, B = Bark, D = Deadwood, F = Fresh Leaves, H = Hypnum, Li = Lichen, O = Orthotrichum, S =
481
Soil, LL = Leaf Litter
482
Figure 2: Functional annotation of oomycete OTUs from Spring 2019. Microhabitat samples
483
based on OTU presence/absence (A) and OTU abundances (C) compared to air samples based on
484
OTU presence/absence (B) and OTU abundances (D). For microhabitat abbreviations, see Figure 1.
485
Figure 3: Ternary plot partitioning the relative abundances of OTUs between canopy, soil and
486
leaf litter. Each dot represents one OTU, sorted by taxonomic order and coloured by lifestyle.
487
Incertae sedis comprises families and genera not associated with any order, e.g. Lagenaceae or
488
Paralagenidium. The order Undetermined represents OTUs with sequence similarities of less than
489
70% to any reference sequence.
490
Figure 4: Differential abundance analysis between the two strata canopy (top panels) and
491
ground (bottom panels) sorted by taxonomic order. Each dot represents one significantly
492
differentially abundant OTU grouped by genus. Y-axis (log2FoldChange) gives the measurement of
493
the differential abundance.
494
Figure 5: Non-metric multidimensional scaling (NMDS) ordination of Bray-Curtis dissimilarity
495
matrices for canopy and ground microhabitats. Canopy microhabitat communities show a large
496
overlap along all sampling events. Ground habitat communities are strongly separated, indicating
497
unique exclusive communities compared to the canopy region, irrespective of the sampling season.
498
Supplementary Figures
499
Supplementary Figure 1: Sequence similarity of reads (top) and OTUs (bottom) per sampling
500
event to published reference sequences. 20.5% of all OTUs, corresponding to 3% of all reads, had
501
a similarity of less than 70% to any known reference sequence (not shown).
502
Supplementary Figure 2: Taxonomic assignment of OTUs per sampling and microhabitat. Black
503
line separates canopy and ground habitats. Distribution of taxonomic groups was similar for every
504
sampling, i.e. Pythiales and Peronosporales dominating all samples.
505
Supplementary Figure 3: Ternary plot partitioning the relative abundances of Peronosporales
506
and Pythiales per sampling event. Each dot represents one OTU.
507
Supplementary Figure 4: Differential abundance analysis between the two seasons spring (top
508
panels) and autumn (bottom panels) sorted by taxonomic order. Each dot represents one
509
significantly differentially abundant OTU grouped by genus. Y-axis (log2FoldChange) gives the
510
measurement of the differential abundance.
511
Supplementary Figure 5: Boxplot of alpha diversity indices for microhabitat communities per
512
sampling. Outliers are given by dots. Observed patterns show no strong variability over the four
513
sampling events.
514
18
Supplementary Tables
515
Supplementary Table 1: Primer tags used in this study. Given are the sample ID, forward and
516
(reverse complemented) reverse tag and the ENA sequencing run ID.
517
Lifeetyiea BME pewmioween BBD cesense tovreen
Coat fo)
foe)
[ee
NMVOE
0.50
0.25
0.00
-0.25
0.50
0.50
0.25
0.00
0.25
0.50
Autumn 2017
Spring 2018
micronavditat
| Arboreal Soil
1) Deadwood
1 Fresh Leaves
Hypnum
Lichen
Eh Orthotrichum
Leaf Litter
BB soil
TreeSpecies
_ Fraxinus excels
iz Quercus robur
mi 1 Tilia cordata
| 2021 | A parasite’s paradise: Biotrophic species prevail oomycete community composition in tree canopies | 10.1101/2021.02.17.431613 | [
"Jauss Robin-Tobias",
"Walden Susanne",
"Fiore-Donno Anna Maria",
"Schaffer Stefan",
"Wolf Ronny",
"Feng Kai",
"Bonkowski Michael",
"Schlegel Martin"
] | creative-commons |
1
Population-level survey of loss-of-function mutations revealed that background
dependent fitness genes are rare and functionally related in yeast
Elodie Caudal1, Anne Friedrich1, Arthur Jallet1, Marion Garin1, Jing Hou1,* and Joseph Schacherer1,2,*
1. Université de Strasbourg, CNRS, GMGM UMR 7156, Strasbourg, France
2. Institut Universitaire de France (IUF)
* Corresponding authors
E-mail: jing.hou@unistra.fr (J.H.), schacherer@unistra.fr (J.S.)
2
Abstract
In natural populations, the same mutation can lead to different phenotypic outcomes due to the genetic
variation that exists among individuals. Such genetic background effects are commonly observed,
including in the context of many human diseases. However, systematic characterization of these effects at
the species level is still lacking to date. Here, we sought to comprehensively survey background-
dependent traits associated with gene loss-of-function (LoF) mutations in 39 natural isolates of
Saccharomyces cerevisiae using a transposon saturation strategy. By analyzing the modeled fitness
variability of a total of 4,469 genes, we found that 15% of them, when impacted by a LoF mutation,
exhibited a significant gain- or loss-of-fitness phenotype in certain natural isolates compared to the
reference strain S288C. Out of these 632 genetic background-dependent fitness genes identified, a total of
2/3 show a continuous variation across the population while 1/3 are specific to a single genetic
background. Genes related to mitochondrial function are significantly overrepresented in the set of genes
showing a continuous variation and display a potential functional rewiring with other genes involved in
transcription and chromatin remodeling as well as in nuclear-cytoplasmic transport. Such rewiring effects
are likely modulated by both the genetic background and the environment. While background-specific
cases are rare and span diverse cellular processes, they can be functionally related at the individual level.
All background-dependent fitness genes tend to have an intermediate connectivity in the global genetic
interaction network and have shown relaxed selection pressure at the population level, highlighting their
potential evolutionary characteristics.
Keywords: background effect | fitness variability | transposition saturation | Saccharomyces cerevisiae
3
Introduction
The same mutation might show different phenotypic effects across genetically distinct individuals due to
standing genomic variation1–8. Such background effects have been described across species and impact
the phenotype-genotype relationship, including in the context of health and disease. Indeed, they have
been observed in multiple human Mendelian disorders, where individuals carrying the same causal
mutation can display a wide range of clinical symptoms, including variable severity and age-of-onset1,7,9–
12. The underlying origin of these background effects may be both intrinsic, i.e. due to interactions
between the causal variant and other genetic modifiers9–11 and/or extrinsic, i.e. due to environmental
factors12,13. To date, a handful of modifier genes have been found associated with human disorders, most
notably in cystic fibrosis11,12. However, such examples remain rare and anecdotal due to the low number
of sample cases in most human Mendelian diseases.
In recent years, several large-scale surveys in different model organisms such as the yeasts
Saccharomyces cerevisiae and Schizosaccharomyces pombe, the nematode Caenorhabditis elegans and
various human cell lines highlighted the broad influence of genetic backgrounds on the phenotypic
outcomes associated with loss-of-function mutations14–22. In yeast, a study comparing systematic gene
deletion collections in two laboratory strains, �1278b and S288C, showed that approximately 1% of all
genes (57/5100) can display background-dependent gene essentiality, i.e. where the deletion of the same
gene can be lethal in one background but not the other14. Several origins underlying such gene essentiality
have been identified, including genetic interactions between the mitochondrial genome and/or viral
elements with the nuclear genome23 as well as genetic interactions between the primary deletion gene and
background-specific modifiers24. While gene essentiality may be the most severe manifestation associated
with loss-of-function mutations, gain- and loss-of-fitness variation related to genetic backgrounds or
environmental conditions were also found in yeast5,25. For example, about 20% of yeast genes showed
background-dependent fitness variation under a wide range of growth conditions, including the presence
of various drugs, osmotic stress, and nutrient sources in 4 genetically diverse isolates25. However, all
these studies only include a limited number of genetic backgrounds and therefore cannot accurately
reflect the extend of the background effect at the species level.
Recently, a large collection of 1,011 S. cerevisiae isolates originated from various ecological and
geographical sources has been completely sequenced26, representing an incomparable resource to
systematically study the effects of genetic backgrounds at the species level. Several strategies have been
developed in S. cerevisiae to explore the impact of loss-of-function mutations, including systematic gene
deletions using homologous recombination14, gene-disruption using the CRISPR-Cas9 editing systems27,
repeated backcrosses25 and transposon mutagenesis28–30. Among these strategies, transposon mutagenesis
based on random excision and insertions are particularly attractive for exploring in parallel a large
number of genetically diverse individuals. This method relies on transposition events via a carrier
plasmid, which allow for the generation of millions of mutants carrying genomic insertions leading to
loss-of-function mutations30. Due to the random insertion patterns in each genetic background, these
methods do not depend on sequence homology as is the case for traditional PCR-based gene deletions and
CRISPR-Cas9 related strategies27,31, and they do not present the risk of inadvertently introducing
exogenous genomic regions as it might be the case for backcross-based strategies25.
4
Here, we selected over a hundred natural isolates broadly representative of the diversity S. cerevisiae
species, and performed transposon saturation analyses using the Hermes transposition system29. We
generated, sequenced and analyzed large pools of transposon insertion mutants and constructed a logistic
model to predict the fitness effects of gene loss-of-function based on the insertion densities in and around
of each annotated gene. Comparing the fitness prediction between the different isolates and the S288C
reference, we identified 632 background-dependent fitness genes, corresponding to approximately 15% of
the genome. Overall, they are functionally related, with members of the same protein complex of biological
process showing similar variability in each genetic background. They also tend to show an intermediate
level of integration in genetic networks compared to non-essential and essential genes, and might be under
positive or relaxed purifying selection at the population level.
5
Results
Generation of LoF mutant collections using the Hermes transposon system
To gain insight into fitness variation associated with loss-of-function mutations across different S.
cerevisiae genetic backgrounds, we performed transposon saturation assays in various natural isolates
using the Hermes transposon system. The Hermes transposon system has previously been adapted in yeast
to allow the selection of random insertion events in liquid culture, which makes this system particularly
suitable for parallel analyses of a large number of genetically diverse individuals29. This system is based
on a centromeric plasmid, which contains the Hermes transposase under the control of a modified
galactose inducible promoter GalS, as well as a transposon carrying a selectable marker (Figure 1A).
Briefly, for any strain of interest, the plasmid is first transformed into stable haploid cells and then
propagated in media containing galactose to induce excision and reinsertion of the transposon at random
locations in the genome, thereby generating a large pool of individuals with hundreds of thousands of
insertions along the genome (Figure 1A). After a recovery phase in rich medium, the genome of this pool
of mutants is recovered, then fragmentated and circularized (Figure 1A). Using PCR with outward facing
primers specifically targeting the transposon, a library that exclusively contains the insertion sites can be
constructed and then sequenced using standard Illumina methods (Figure 1A). In principle, transposon
insertions that cause severe fitness defects, for example those occurring in essential genes, will not be
recovered due to the competitive disadvantage compared to events occurring in genes which are not
essential. Analysis of insertion patterns along the genomes of different individuals therefore provides a
proxy for fitness variation related to loss-of-function mutations.
In addition to the S288C reference strain, we selected 106 isolates originated from various ecological and
geographical sources that are broadly representative of the species diversity (Figure 1B, Table S1). Stable
haploid variants of this set of isolates have been generated previously32,33 and are all capable of growing
in galactose medium. We have adapted the initial version of the Hermes transposon plasmid to carry a
hygromycin resistance marker instead of nourseothricin to ensure compatibility with selected strains,
which may carry either a KanMX or NatMX marker at the HO locus. Transposon insertion profiles for
each isolate were obtained as described (Figure 1A). We observed a marked variability in terms of
insertion efficiency across different genetic backgrounds, ranging from ~100 to ~300,000 unique insertion
sites (Figure S1A, Table S1). No discernible correlation between the genetic origin of the isolates and the
transposon insertion efficiency was observed (Table S1). We then compared the insertion preferences
between the S288C reference strain and the 106 natural isolates (Figure S1B). Insertion densities for
known sequence motifs29 were conserved across the different genetic backgrounds (Figure S1B).
Using insertion profiles and the annotation of gene essentiality in the S288C reference, we analyzed the
average insertion patterns in the promoters (-500 bp to ATG, 100 bp window), the coding DNA sequences
(CDS), and the terminators (STOP to +500 bp, 100 bp window) for all annotated essential vs. non-
essential genes and found that the number of insertion drops from -100 bp prior to CDS and extends to -
100 bp until the terminator region, with on average ~3 times fewer insertions within the CDS in the
essential genes compared to non-essential genes (Figure S1C-E). This pattern is consistent with the results
obtained in previous studies using the Hermes system29.
6
Modeling based on insertion patterns to identify background-dependent fitness genes
We constructed a logistic model that simultaneously takes into account transposon insertions that have
occurred in the genes and surrounding regions using the insertion profiles from the reference strain S288C
and the corresponding gene essentiality annotations as a binary classifier (Figure 1C, Table S2, See
Methods). We applied this model to the insertion profiles of all 107 diverse isolates (Figure 1D). For each
annotated ORF (approximately a total of 6,300 ORFs), a probability was calculated based on the model,
ranging from a value of 1, corresponding to most likely non-essential, to 0, corresponding to most likely
essential. Genomic regions with low insertion densities contribute to overall low predictive powers
(Figure S2A-B), which were subsequently removed. Due to the variability of insertion efficiency across
strains (Figure S1A), removal of regions with low insertion density has led to entire strain backgrounds
with few interpretable genes. By maximizing both the number of strains and genes that remained after
data imputation (Figure S2C-D), a total of 52 backgrounds and probability predictions for 4,469 genes
were retained for subsequent analyses (Table S3).
Large-scale genome duplications, including aneuploidies and endoreduplications, are frequently observed
in yeast experimental evolution34–36. Such events may hamper the accuracy of the modeled fitness effect
in the context of the transposon insertion assay, as genes in the duplicated region will all appear to be fit
due to insertions in a single copy of the gene. We searched for signals of large-scale genome duplications
by examining all annotated essential genes along the chromosomes in our set of isolates (Figure 1D,
Figure S3A). We detected endoreduplication events in 7 out of 52 strains where all chromosomes
appeared to be duplicated based on the high predicted probability values for all essential genes (Figure
S3A). In another set of 6 strains, essential genes showed an intermediate to high probability prediction but
not high enough to be confidently classified as non-essential. These 6 strains were then confirmed as a
mixture of haploid and diploid cells using flow cytometry. In addition to these whole genome events, we
also detected 3 strains with an aneuploidy of chromosome I (ACT, BKL and ACV), one strain with an
aneuploidy of chromosome XII (CPG) and one strain with an aneuploidy of chromosome XIV (CQA).
These aneuploidies were not present in the original isolate, with the exception of chromosome I
aneuploidies in ACT and ACV strains, highlighting the dynamics of genome instability in different
genetic backgrounds. All 13 strains with whole genome endoreduplication were entirely removed from
the dataset. We also excluded aneuploid chromosomes from the analysis (Table S3).
Next, we looked specifically at the probability predictions in the reference S288C. The final set of 4,469
genes includes 3,732 and 737 that were annotated non-essential and essential, respectively. Among the
genes annotated as non-essential, approximately 180 were predicted to be likely essential in our data
(Table S3), of which more than 70% correspond to slow grow or galactose-specific fitness defect genes.
For example, the hexose transporters HXT6/7 and genes involved in galactose metabolism are all
predicted to be likely essential, as expected by using our transposition saturation strategy (Figure S3B).
On the other hand, 26 genes annotated as essential were predicted to be likely non-essential, with a
predicted probability > 0.8 (Table S3). Among these, we found FUR1, HIP1 and SSY5, consisting of
amino acid transporters that are only essential in the multi-auxotrophic BY4741 background, isogenic to
the prototrophic S288C we used in our study (Table S3). We have also found genes where the essentiality
concerns only part of the ORF, i.e. the essential domains, as has also been observed in previous studies
7
using this transposon saturation strategy28 (Figure S3C). Notably, we found the RET2 and SRP14 genes,
which are also among the essential genes specific to the S288C background compared to �1278b in
systematic gene deletion collections14. Indeed, these domain essential effects are recaptured in our dataset
when comparing insertion patterns between S288C and �1278b (Figure S3D). In fact, background-
specific essential genes between S288C and �1278b that did not display severe fitness defect when
deleted in the non-essential background24, including S288C-specific essential genes (RET2, UBC1 and
SRP14), and �1278b-specific essential genes (SKI8, TMA108 and AAT2), all showed domain essential
effects and are all recaptured in our data (Figure S3D).
Overall, the predicted probability based on our logistic model can serve as a reasonable proxy for fitness
variation related to loss-of-function mutations. Modeled fitness (predicted probability for non-
essentiality) is more accurate at predicting non-essential/high fitness cases than essential/low fitness
cases, which may in part due to that non-essential genes are de facto slow growers in the context of our
experimental conditions, and in part due to bias in transposon insertion densities across genomic regions
and genetic backgrounds. Essentialities related to specific domains can be recaptured by the raw insertion
patterns but not by our modeled fitness values (Figure S3C-D). However, this effect is inherent to the
transposon saturation system and should not lead to differential fitness effect prediction in different
genetic backgrounds. Our final dataset consists of 39 isolates from various origins and predicted fitness
for 4,469 genes, which is analysed in more detail (Figure 1E, Table S3).
Environmental dependency of fitness variability associated with LoF mutations across backgrounds
We first performed a hierarchical clustering based on the predicted fitness values of 4,469 genes across
the 39 genetic backgrounds (Figure 2). Profile similarity based on the predicted fitness effects did not
correlate with the genetic origins of the isolates (Figure 2). Genes that are consistently essential in
different isolates clustered together and are enriched for essential biological processes, including
ribosome biogenesis, rRNA processing, DNA replication, protein transport and cell cycle (Figure 2).
Genes that are consistently non-essential in all backgrounds formed a large cluster without significant
enrichment for any specific biological process. Interestingly, several clusters of genes with variable
fitness effects were identified, displaying modular switches from fit to non-fit phenotypes across the
entire population. Gene enrichment analyses revealed genes involved in mitochondrial translation,
transcription regulation and general translational processes (Figure 2). A large proportion of these genes
with population-wide fitness variation consists of nuclear encoded mitochondrial genes involved in
respiration, which were expected to show a selective disadvantage in our pool of mutants that must grow
on galactose. This observation suggests that such general fitness variability may be environment-related
rather than background-specific per se. However, other biological processes in addition to respiration and
mitochondrial functions have also been enriched, for which the impact of environment vs. genetic
background on their fitness variability remains unclear.
To further characterize the background-dependent fitness variation, we systematically compared the
predicted fitness values for each gene in a given isolate with the predictions of the reference strain S288C.
A differential fitness score for each gene in each background was calculated by subtracting the predicted
fitness value in a given strain from the corresponding fitness prediction in the reference S288C. A
minimum absolute value of the differential fitness score of 0.5 was considered significant, which
8
corresponds to a bona fide reverse in the direction of being predicted as essential or non-essential
according to our logistic model. In total, 632 genes were identified with marked fitness variation, with
458 and 174 showing a loss-of-fitness (S288C healthy and background sick) and a gain-of-fitness (S288C
sick and background healthy) compared to the reference, respectively. The number of identified
differential fitness genes ranges from 8 (ACP) to 88 (BQH) for loss-of-fitness cases (with a median of
61), and from 6 (CGD) to 42 (AMF) for gain-of-fitness cases (with a median of 16) (Figure 3A). A total
of 163 out of all 632 hits are related to respiration and mitochondrial functions, representing ~20% to
~60% of loss-of-fitness hits depending on the genetic background (Figure 3A). Furthermore, these
respiration-related genes tend to impact more backgrounds on average than non-respiration related hits
(Figure 3B). These observations echoed what was shown on hierarchical clustering where mitochondrial
related genes were highly enriched in clusters with modular fitness variation in several backgrounds
(Figure 2). Again, due to the overrepresentation of these respiration-related genes and their continuous
fitness variation in the population, we suspect that these hits are likely to be impacted by the environment
(i.e. pooled competition in galactose media) in addition to any specific genetic backgrounds.
We then calculated the z-statistics for all variable fitness hits to distinguish those that are background-
specific from the others that are possibly related to the environment (Table S4). In principle,
environment-related cases are more likely to vary continuously in the population with a low z-statistics,
whereas cases that are truly specific to some genetic backgrounds should be outliers with a high z-statistic
score (|z| > 3) (Figure 3C). Of the set of 632 genes, we found 179 that are background specific, which
mainly impact a single genetic background (Table S4). These background-specific genes are rarer
compared to the environment-related group, with a median of 5 identified per isolate both loss- and gain-
of-fitness types combined (Figure 3A). Genes related to respiration and mitochondrial functions are not
overrepresented in this group (23/179 vs. 691/4469 in the background, Fischer’s exact test P-value =
0.82). No significant enrichment for any biological processes or molecular functions has been identified.
By contrast, respiration-related genes are significantly overrepresented in the remaining group (140/453
vs. 691/4469, Fischer’s exact test P-value = 1.6e-10, odds ratio = 2). Each of these 453 potentially
environment-related genes impact on average 6 genetic backgrounds.
Environment-related fitness variation reveals potential functional rewiring
While a large fraction of potentially environment-related hits correspond to genes known to be involved
in respiration, the majority of this group is involved in other biological processes. To explore the
functional relationships within this group, we calculated the pairwise correlations between these genes
using predicted fitness values across the 39 strain backgrounds (Figure 4A). We constructed a network
based on the profile similarities where the edges correspond to a Pearson’s correlation > 0.6 (correlation)
or < -0.6 (anti-correlation) (Figure 4B, Figure S4). In total, 292 out of the 453 environment-related hits
exceeded our stringent correlation cut-offs (Figure S4). The profile similarity and network structure
revealed two main subnetworks, which are correlated within the subgroup but are anti-correlated between
subgroups (Figure 4A-B). The first subgroup contains mainly respiration-related genes, in particular
genes involved in mitochondrial translation (Figure 4A, Figure S4A), which are anti-correlated with
genes involved in transcription regulation and chromatin remodeling (SPT7, SPT8, SWC4, SWC5, ARP6,
ARP7, SIN3, RKR1, YAF9, UME1, NGG1, CHD1, STH1, for example) as well as genes involved in
nuclear-cytoplasmic protein transfer (KAP120, KAP122, KAP123, NUP57, NUP100, NUP188, POM152,
9
NIC96, MLP1, for example) (Figure S4A). Many of these correlations have been found between members
of the same protein complexes. Several members of the transcription and nuclear transport subgroup are
also annotated as related to respiration (deletion leads to absence of respiration) although they are not
directly involved in mitochondrial function, such as SIN3, a general chromatin remodeler, and KAP123, a
karyopherin responsible for nuclear import of ribosomal proteins. In addition to this large network,
several small networks have also been detected (Figure S4B-D), including PMT1, PMT2 and GET2,
which are involved in ER related glycosylation and are known to have physical interactions. Functional
enrichments in the anti-correlated subgroups suggest a potential ‘rewire’ between mitochondrial
translation and transcription regulation/nuclear transport, where modular switched of fitness effects
associated with gene loss-of-function may occur in different genetic backgrounds.
Functional insights into background dependent fitness genes
To further explore the functional enrichments of fitness variation genes at the strain level, we annotated
genes in our dataset into 16 functional neighbourhoods according to SAFE37 and looked for enrichment in
different neighbourhoods (Figure 5A). For each neighbourhood, we calculated the odds ratio of
enrichment based on the number of hits annotated in the neighbourhood vs. the total number of hits, with
the size of the neighbourhood and the total number of genes as background (one-sided Fisher’s exact
test). Globally, background-specific hits are not enriched for most processes except for cell polarity (OR
= 1.49, P-value = 0.026). Environment-related hits are enriched for respiration and mitochondrial
functions (OR = 3.77, P-value = 4.16e-17), as well as transcription and chromatin regulation (OR = 1.53,
P-value = 0.002), nuclear cytoplasmic transport (OR = 2.14, P-value = 0.004) and DNA repair (OR =
1.55, P-value = 0.01) (Figure 5A, Table S4). When looking at the same neighbourhood enrichment at the
strain level, environment-related hits are enriched for mitochondrial functions in most genetic
backgrounds, with the exception of ACP and CLG strains, the latter of which has a predicted fitness
profile that was most similar to the reference S288C (Figure 2). A large fraction of isolates showed
significant enrichments for transcription and chromatin regulation as well as nuclear-cytoplasmic
transport (Figure 5A). These enrichments are consistent with the rewiring hypothesis based on the profile
similarity network analysis (Figure 4B). Indeed, by specifically looking at the annotated genes in these
functional neighbourhoods, we observed various degrees of rewiring depending on the backgrounds
(Figure 5B-C). In the reference S288C, loss-of-function for annotated genes in these three
neighbourhoods showed either high- or low-fitness predictions (Figure 5B, Figure S5A). While in other
genetic backgrounds, these predictions may be reversed as gain- or loss-of-fitness hits compared to
S288C, with profiles ranging from similar to S288C (CLG) to almost completely reversed (AMF) (Figure
5C). Most notably, such rewire could include either only mitochondrial-related genes, or with one or more
processes related to either transcription and chromatin regulation or nuclear-cytoplasmic transport (Figure
5C). Depending on the genetic background, different sets of genes within the same functional
neighbourhood could be involved, highlighting the dynamics of such rewire (Figure S5B).
Compared to environment-related genes, background-specific ones are rare, and tend to show little
functional enrichment, as expected. However, in cases where multiple genes are detected in the same
genetic background, some enrichments emerge (Figure 5A, Table S4). For example, in the strain BDH, 8
background-specific genes were detected with 3 annotated into one of the 16 functional neighbourhoods,
and two of which are involved in MVB sorting and pH dependent signaling (RIM8 & RIM101). Both
10
genes are non-essential in S288C but predicted as loss-of-fitness in the BDH background (Figure 5A). In
the strain AMF, 16 background-specific genes were detected with 11 annotated, among which 2 were
involved in protein degradation and turnover (VID28 & PRE3) and 3 were involved in glycosylation and
cell wall biogenesis (OST1, OPI3 & FAB1). These observations demonstrate that background-specific
fitness variation genes, while rare, can be functionally related and may involve multiple members of the
same protein complex or biological process.
Finally, as previously posited6, genes exhibiting background-dependent fitness variation tend to show an
intermediate level of connectivity in terms of genetic interactions (Figure 6A, Table S5) and an
intermediate functional similarity between interacting gene pairs compared to genes that are consistently
non-essential or essential (Figure 6B). Both environment-related and background-specific hits have the
same pattern. Interestingly, background-specific genes display higher non-synonymous to synonymous
substitution rates (dN/dS) than essential genes and non-essential genes (Figure 6C), indicating potential
positive or relaxed purifying selection on these genes at the population level. Overall, background-
specific fitness genes tend to be diverse yet can be functionally related within a single genetic
background. Genes with environment-related fitness variation share general evolutionary characteristics
with background-specific cases.
11
Discussion
To have a better insight into the background-dependent fitness variation associated with gene loss-of-
functions, we explored a large number of natural yeast isolates using a transposon saturation strategy. We
modeled fitness by considering transposon insertion densities within gene coding sequence and
surrounding regions. The comparison of the modeled fitness between different isolates and the reference
S288C allowed the identification of 632 genes displaying background-dependent phenotypes. The
majority of these cases (71,7%) showed continuous fitness variation across the population and is at least
partly related to the environment. By contrast, background-specific cases tend to be rare, with on average
5 genes per isolate. At the individual level, both environment-related and background-specific variable
fitness genes can be functionally related.
The impact of the environment on the background-dependent fitness genes can be supported by two main
observations. First, this set of genes was highly enriched for respiration and mitochondrial functions,
which is consistent with a fitness loss under prolonged growth in media with galactose as the sole carbon
source. Indeed, mitochondrial-related genes were also found to be background-dependent in a previous
study involving 4 different isolates under conditions with non-fermentable carbon sources25. Second,
these genes showed a continuous variation across the population. Further analyses highlighted that genes
involved in two biological processes, namely transcription & chromatin remodeling and nuclear-
cytoplasmic transport, are anticorrelated with genes involved in mitochondrial translation in terms of their
fitness profiles. These anticorrelations indicate a modular change in the relative fitness of genes involved
in these processes compared to the reference strain S288C. However, whether such rewiring effect is
exclusively related to respiration conditions or could represent a general background-dependency effect
remains difficult to disentangle due to the experimental conditions required for transposon saturation
analyses.
In a recent large-scale analysis of environment-dependent genetic interactions, it has been shown that
most interactions specific to an environmental condition are in fact part of the global genetic interaction
network that were exacerbated or attenuated in the tested condition38. Compared to genetic interactions
between pairs of gene deletion mutants, the background-dependent gene loss-of-function phenotype could
be considered as interactions between the loss-of-function gene and background-specific modifiers, which
are expected to share general properties to genetic interactions with deletion mutants. Indeed, we tested
the gene deletion phenotype for one of the environment-related genes involved in transcription and
chromatin remodeling, the BMH1gene (Figure S5C). This gene was identified as loss-of-fitness in
multiple genetic backgrounds compared to S288C in our study. Interestingly, the loss-of-fitness
phenotype was indeed confirmed on standard rich media, suggesting the environment-related fitness
variation genes could have a general effect. In addition, genes involved in chromatin remodeling were
also found to display background-dependent fitness effects in a previous study comparing S288C and a
natural isolate 3S5. These observations suggest that the rewiring effect could have implications beyond a
specific experimental condition.
Although transposon saturation strategy can be versatile to genetic diversity among isolates, this method
also presents some limitations. Among all the isolates initially tested, only about half showed a reasonable
level of insertion efficiency, highlighting the unexpected variability of transposon activity between
12
different individuals. This variability results in an underestimate of the number of genes associated with
background-dependent phenotypes. In addition, loss-of-function phenotypes that are related to specific
protein domains but not the entire ORF are difficult to identify, unless the insertion efficiency is
extremely high. The Hermes system, like all currently available saturation systems, requires step of a
transposon induction in the presence of galactose30. This competition effect in a non-fermentable carbon
source may complicate downstream analysis as the effects of environment vs. genetic background can be
difficult to unravel. New strategies that take into account these factors are still needed in order to get a
more precise view of background-dependent gene loss-of-function phenotypes at the species level.
13
Material and methods
Strains and growth conditions
A total of 106 isolates were selected from the 1,011 Saccharomyces cerevisiae collection26. A prototrophic
haploid strain FY5, isogenic to the reference strain S288C was also included. Haploid segregants derived
from the 106 natural isolates were obtained after HO deletion and tetrad dissection32,33. Detailed
descriptions of the strains can be found in Table S1. Strains were maintained at 30°C using YPD (1% Yeast
extract; 2% Peptone, 2% Dextrose) in liquid culture or solid plates (2% of agar). Transposon activity was
induced in YPGal (1% Yeast extract; 2% Peptone, 2% Galactose) with Hygromycin B (200 µg/mL).
Sporulation was induced on solid plates containing 1% of potassium acetate and 2% of agar.
Ploidy control
Ploidy was estimated by flow cytometry. Cells in exponential growth phase were washed in water, then 70%
ethanol and sodium-citrate buffer (50 mM, pH 7.5) followed by RNase A treatment (500 µg/mL). To avoid
cell aggregates, each sample was sonicated then the DNA was labelled with propidium iodide (16 µg/mL),
a fluorescent intercalating agent. DNA content was then quantified using the 488 nm excitation laser of the
Accuri C6 plus flow cytometer (BD Biosciences).
Cell transformation
Cells in exponential growth phase were chemically transformed using the EZ-Yeast Transformation Kit
(MP biomedicals). We incubated cells 30 minutes at 42°C with EZ-Transformation solution, carrier DNA
and either 100 ng of pSTHyg plasmid or 1 µg of PCR fragment. After regeneration in YPD, cells were
spread on solid YPD plate supplemented with Hygromycin B and incubated at 30°C until transformants
appeared.
Construction of the pSTHyg plasmid
In order to be compatible with our isolates already carrying either a nourseothricin or a kanamycin
resistance cassette, the nourseothricin cassette of the pSG36 plasmid29 was replaced by a hygromycin B
resistance cassette. The pSG36 plasmid was amplified in 2 fragments by PCR excluding the natMX cassette,
then assembled with the hphMX cassette amplified from p41 plasmid (Addgene #58547) with overlapping
regions using Gibson assembly. The new plasmid, pSTHyg was amplified in E. coli and extracted using the
GeneJET Plasmid Miniprep Kit (Thermo Scientific™). The construction was verified using enzymatic
digestion with KpnI and PvuI.
Generation of transposon insertion mutant pools
Each natural isolate was grown in liquid YPD medium and chemically transformed with 100 ng of pSTHyg
plasmid as described. From the selective transformation plates, a single clone was picked and grown in
30 ml of YPD supplemented in hygromycin B under agitation at 30°C until saturation (~ 24h). Cells were
then diluted at an OD of 0.05 in 50 ml of YPGal supplemented with hygromycin B to activate the
14
transposase and induce the transposition for 72h at 30°C. Two successive dilutions were then performed
for 24h at an OD of 0.5 in 100 ml of YPD then YPD supplemented with hygromycin B to enrich for cells
the transposon in their genome. The final 100ml culture was centrifuged, water-washed and 500 µl aliquots
of cells were frozen at -20°C.
Sequencing library preparation
In order to sequence the genomic regions with a transposon insertion, the genomic DNA of the pool of cells
carrying insertion events was extracted using the MasterPure™ Yeast DNA Purification Kit (Lucigen).
Cells were lysed using a lysis solution supplemented in zymolyase 20T (1.5 mg/ml). Proteins and cellular
debris were removed with the MPC Protein Precipitation Reagent and several RNase A treatments were
realized to eliminate RNA. Genomic DNA was then precipitate with ethanol. The pellet was washed twice
with 70% ethanol and resuspended in 80 µl of water. The gDNA sample integrity was controlled on 1%
agarose gel and quantified on Nanodrop and Qubit using the Qubit™ dsDNA BR Assay Kit (Invitrogen™).
2 x 2 µg of gDNA were digested in parallel with 50 units of DpnII (NEB #R0543L) and NlaIII (NEB
#R0125L) in 50 µl for 16h at 37°C. The enzymatic reactions were inactivated for 20 min at 65°C and DNA
fragments were ligated with 25 Weiss units of T4 Ligase (Thermo Scientific #EL0011) in a total volume of
400 µl for 6h at 22°C. Circular DNA were then precipitated overnight at -20°C with ethanol, salt (NaOAc
3M pH5.2) and glycogen. After an 70% ethanol wash, the DNA pellet was resuspended in 50µL of water.
The junction between the genomic region and the transposon insertion site was amplified on both DpnII
and NlaIII digested and re-circularized gDNA by PCR using outward-facing primers targeting the
transposon. The PCR products were controlled on 1% agarose gel and displayed variable sizes centred
around 750 bp. Nanodrop and Qubit using the Qubit™ dsDNA BR Assay Kit (Invitrogen™) quantifications
were then performed to pool the same amount of NlaIII-digested and DpnII-digested PCR products. For
each sample, at least 6 µg at minimum 30 ng/µl was then sent to the BGI (Beijing Genomics Institute) for
sequencing. In total, each sequencing run provided 1 Gb of 100 bp paired-end reads using Illumina Hi-Seq
4000 or DNBseq technologies.
Determination of transposon insertion sites
The reads that contained the amplified part of the transposon were selected and the corresponding 57 bp
sequence was trimmed with Cutadapt39 and the reads corresponding to the plasmid were discarded. The
cleaned reads were mapped to the S288C reference genome with the corresponding SNPs inferred for each
isolate26 with BWA40. The genomic position of an insertion site was defined as the first base pair aligned
on the genome after the transposon region. For each insertion site, the number of reads and their orientation
were obtained.
Modelling the fitness effect of gene loss-of-function based on transposon insertion profiles
The number of insertions in the promoter region (-100 bp to ATG), beginning of the coding region (-100
bp to +100 bp from ATG), the coding region, end of the coding region (-100 bp to +100 bp from stop-
codon) were normalized as insertion densities per 100 bp. Gene essentiality annotations were obtained
from SGD (phenotype “inviable”) exclusively for annotations with gene deletion in the S288C
background. Respiration related gene annotations were obtained from SGD with the phenotype
15
“respiration: absent” after gene deletion in S288C. Galactose-specific loss-of-fitness was determined in
Costanzo et al.38, with a stringent cut-off of < -0.2. A logistic model was constructed using the glm()
function from the R package “stats”, using insertion densities in the reference strain S288C, in the
promoter region (-100 bp to ATG), beginning of the coding region (-100 bp to +100 bp from ATG), the
coding region, end of the coding region (-100 bp to +100 bp from stop-codon), raw insertion number in
the coding region and gene sizes as predictors, and essentiality annotations as a binary classifier. Genes
displaying a slow growth phenotype41, genes with differential fitness defect in galactose media38, as well
as genes showing respiration defects were excluded. Genes that are localized in regions with low insertion
densities, i.e. less than 3 insertions in the terminator region (STOP to +300 bp) and less than 50 insertions
in a 10 kb region surrounding the gene (-5 kb before ATG and +5 kb after STOP) were also excluded. A
total of 4600 genes were included in the model corresponding to 867 essential genes and 3737 non-
essential genes (Table S2). 10-fold cross-validation was performed using the R package “caret”, with
trainControl() and train() functions, method = “glm”, family = “binomial”. Cross-validation results
showed an average accuracy of 0.88 with a Kappa 0.57 (Table S2). The predictive value for non-essential
labels is 0.91, contrasting to a lower predictive value of 0.70 for essential labels, indicating a better
accuracy in predicting non-essential genes using this model. This lower predictive power for essential
genes is more or less expected as the absence or low numbers of insertions could be linked to the overall
low insertion density in certain genomic regions, which is independent of gene essentiality. Imputations
for missing values were performed using the function impute.knn() in the R package “impute”, with k =
10, rowmax = 50% and colmax = 80%. Quantile normalization of the imputed matrix was performed
using normalize.quantiles() function in the R package “preprocessCore”. All fitness prediction data can be
found in Table S3.
Validation of the phenotypic consequence of BMH1 gene loss-of-function
Stable haploid isolates, FY5 and CIB were diploidized using the pHS2 plasmid (Addgene #81037)
containing the HO gene encoding the endonuclease responsible for mating type switching and a hygromycin
resistance cassette. The BMH1 gene was replaced with a Hygromycin B resistance cassette in the diploid
isolates. Sporulation was induced on potassium acetate medium in diploid isolates heterozygous for BMH1
gene deletion. Around 20 resulting tetrads were then dissected on YPD using a MSM 400 micromanipulator
(Singer Instrument). Each spore grew for 48h at 30°C and the colony size was captured with the camera of
the colony picker, PIXL (Singer Instrument). Colony size measurements were then analysed using custom
R scripts.
Data availability
All sequencing data related to this study were deposited to the European Nucleotide Archive (ENA) under
the accession number PRJEB45777.
16
Acknowledgments
We thank Agnès Michel for helpful suggestions throughout the project. This work was supported by the
European Research Council (ERC Consolidator Grant 772505). It is also part of the Interdisciplinary
Thematic Institute IMCBio, as part of the ITI 2021-2028 program of the University of Strasbourg, CNRS
and Inserm, supported by IdEx Unistra (ANR-10-IDEX-0002), and EUR (IMCBio ANR-18-EUR-0016)
under the framework of the French Investments for the Future Program. JS is a member of the Institut
Universitaire de France.
17
Figure legends
Figure 1- Summary of the Hermes transposon saturation procedure. (A) A centromeric plasmid
carrying the Hermes transposase and a transposon containing a hygromycin resistance marker (HygMX) is
transformed into a haploid isolate background. Random transposon insertions are induced and selected.
The mutant pool is then recovered and a PCR library that contains only the insertion sites is constructed
and sequence. (B) Distribution of the selected 107 isolates across the species. The neighbour-joining tree
was constructed using biallelic SNPs in the 1,011 yeast collection26. Selected strains are highlighted in
black. (C) A logistic model was constructed using insertion profiles in the reference strain S288C. Gene
essentiality annotations were used as a binary classifier, excluding those annotated as involved in
galactose metabolism, respiration and slow growth. (D) The logistic model was applied to insertion
patterns in the remaining 106 isolates. Large-scale genome duplications were detected by looking at
fitness predictions for all annotated essential genes along each chromosome. Low coverage regions were
removed then imputed using k-nearest-neighbour method. The imputed fitness matrix was then quantile
normalized. (E) The final dataset after imputation consists of 39 isolates and 4469 genes. Strains included
in the final dataset are highlighted in blue.
Figure 2- Hierarchical clustering of 4469 fitness predictions across 39 genetic backgrounds. The
distance matrix was calculated using the Euclidean distance method. The genetic origin of each isolate
was color-coded, and the total insertion numbers per isolate was represented by dot size under the origin
color code. Genes annotated as essential in the reference S288C are highlighted in black, and genes
annotated as either galactose or respiration related are highlighted in yellow on the sidebars. Genes within
duplicated chromosomes were removed (yellow bars on heatmap). Biological processes that are
enrichment in subclusters are annotated.
Figure 3- Number and distribution of background-dependent fitness variation genes. (A) Number of
hits detected in each genetic background. Genes annotated as galactose or respiration-related and
background-specific genes are color-coded as indicated. Strains are sorted according to the total number
of insertions. (B) The number of genetic backgrounds impacted by the detected hits. Top panel, gain-of-
fitness genes compared to S288C; bottom panel, loss-of-fitness genes compared to S288C. (C) Z-statistic
distribution for hits that impact different numbers of genetic backgrounds. A cut-off of |z-statistics| > 3 is
indicated with dotted lines.
Figure 4- Correlation analyses for environment-related hits. (A) Pairwise profile similarity based on
predicted fitness across 39 backgrounds. Distance matrix was based on pairwise Pearson’s correlation.
Gene essentiality annotations are indicated on the upper sidebar and genes annotated as involved in
galactose/respiration are indicated on the left sidebar. (B) Network based on profile similarity among
environment-related hits. Genes annotated as involved in galactose/respiration are colored in yellow.
Positive correlations (> 0.6) are represented as red edges and negative correlations (< -0.6) are
represented as blue edges. Complete network with annotated gene names can be found in Figure S4.
Figure 5- Functional enrichments and rewiring for background-dependent fitness genes. (A)
Enrichments across 16 functional neighbourhoods defined by SAFE37. Dot sizes represent odds ratios
between the number of hits in a given neighbourhood vs. the total number of hits detected, with the size of
18
the neighbourhood vs. the total number of genes in the dataset as background, using one-sided Fischer’s
exact test. Global enrichment for background-specific (blue) and environment-related (orange) hits are
presented on the left panel, and strain-centric enrichments are on the right panel. Enrichments with a p-
value < 0.05 are shown. Backgrounds highlighted by dashed lines corresponds to example rewiring
diagrams in (C). (B) Predicted fitness for genes annotated in Respiration/mitochondrial targeting,
Transcription and chromatin organization and nuclear-cytoplasmic transport in the reference S288C.
Genes in different processes are arranged by descending order of the modeled fitness. Detailed annotated
version of this diagram can be found in Figure S5A. (C) Example rewiring diagrams in other backgrounds
compared to the reference S288C. A switch from healthy to sick (loss-of-fitness) is indicated in blue and a
switch from sick to healthy (gain-of-fitness) is indicated in orange for any given gene in a given
background. The diagrams for all 38 isolates are shown in Figure S5B.
Figure 6- Evolutionary features associated with background-dependent fitness genes. (A) Genetic
interaction degrees derived from the yeast global genetic interaction network37 for non-essential,
background-specific, environment-related and essential gene categories. The number of genes annotated
in each category are indicated. (B) Functional co-annotation rates37 for different gene categories. The co-
annotation rate corresponds to the fraction of interaction partners that are annotated in the same biological
process as the primary gene37. (C) Mean non-synonymous vs. synonymous substitution rates (dN/dS)
across 1,011 natural yeast isolates using the YN00 method26. Comparisons between categories were
performed using T-test, and significance levels are as indicated, with ns: P-value > 0.05, *: P-value <
0.05, **: P-value < 0.01, ***: P-value < 0.001 and ****: P-value << 0.0001.
19
Supplemental figure legends
Figure S1- (A) Number of reads (y-axis, log10 scale) vs. number of unique insertion sites (x-axis, log10
scale) across 107 diverse isolates. (B) Insertion preference comparison between the reference S288C and
the other 106 selected isolates. Sequence motifs are on the x-axis and the percentage of reads with a given
motif are presented as color coded bars. Error-bars correspond to the standard deviation across different
isolates. (C) Insertion density comparison between essential and non-essential genes in S288C in the
promoter region. Average insertion numbers in the -500 bp to +200 bp region relative to ATG are shown
in 100 bp windows. (D) Insertion density comparison between essential and non-essential genes in S288C
in the coding region (CDS). Average insertion numbers in the relative fractions of a given CDS are
shown. (E) Insertion density comparison between essential and non-essential genes in S288C in the
terminator region. Average insertion numbers in the -200 bp to +500 bp region relative to the stop-codon
are shown in 100 bp windows.
Figure S2- (A) Predicted non-essential probabilities (y-axis) as a function of the number of insertions in
the terminator region (300 bp after stop-codon). Non-essential genes are in blue and essential genes in
red. (B) Predicted non-essential probabilities (y-axis) as a function of the number of insertions in a 10 kb
region surrounding the CDS (5 kb before ATG and 5 kb after stop-codon). Non-essential genes are in blue
and essential genes in red. (C) The number of strains retained as a function of cut-offs of the number of
interpretable genes after removing low coverage regions (less than 50 insertions in the surrounding 10kb
region and/or less than 3 insertions in the 300 bp terminator region). (D) Number of genes retained after
imputation as a function of cut-offs of the number of interpretable genes after removing low coverage
regions.
Figure S3- (A) Average non-essential probability or predicted fitness for every 10 successive essential
genes along all 16 chromosomes for 52 strains that passed the coverage cut-offs. Strain-side clustering
was based on predicted fitness for all genes. (B) Insertion profiles for gene related to galactose
metabolism that are annotated as non-essential in S288C but detected as essential/sick in all or a fraction
of the 39 strains in the final dataset. Chromosomal positions and gene orientations are schematically
presented on the x-axis and insertion profiles for each strain are presented as black vertical bars. (C)
Insertion profiles for essential genes predicted as non-essential in S288C. Shaded areas correspond to
potential essential protein domains. (D) Insertion profiles for genes previously shown background-
specific essentiality between S288C and �1278b. Domain-specific essentiality regions are indicated.
Figure S4- (A) Annotated network based on profile similarity as shown in Figure 4B. (B-D) Subnetworks
with significant correlations independent from the large subnetwork involving respiration-related hits.
Figure S5- (A) Predicted fitness for genes annotated in respiration/mitochondrial targeting, transcription
& chromatin organization and nuclear-cytoplasmic transport in the reference S288C with gene name
annotations. Related to Figure 5B. (B) Rewiring diagrams for all 38 isolates relative to the reference
S288C. Related to Figure 5C. (C) Example of functional rewire in a natural isolate CIB compared to the
reference S288C for a transcription-related hit BMH1. Relative fitness ratio (colony size for WT divided
by deletion of BMH1) is shown on the upper right panel. Colony sizes of BMH1 deletion vs. WT were
measured using tetrad dissection of hemizygous diploids. 5 tetrads are shown for each background.
20
Supplemental tables
TableS1- Description of isolates used in this study
TableS2- Model construction and evaluations. This table contains 4 tabs:
GenesInModel: 4600 ORFs and their essentiality annotations used to construct the logistic model.
Insertion numbers and densities within coding sequence and surrounding regions are included.
Insertion numbers calculated from S288C insertion profile.
ModelSummary: Features included in the logistic model and their coefficient.
CrossValidation: Summary of the cross validation results.
CMStat: Confusion matrix, prediction accuracy and precision for essential/non-essential labels.
TableS3- Raw and final dataset with predicted fitness. This table contains 3 tabs:
Raw_data_pred: All raw predicted fitness based on the logistic model for 107 isolates.
Pred_final_39: Predicted fitness for 39 isolates included in the final dataset. Raw, imputed and
quantile normalized predictions are shown.
Score_final_39: Differential fitness score by comparing the predicted fitness in a given isolate to
S288C.
TableS4- Background-dependent fitness variation genes identified in this study. This table contains 4
tabs:
Z-statistics: Z-statistics for each of the 632 hits, including the number of genetic backgrounds
impacted for each hit.
Hits_SAFE_annotation: Annotations for each hit into the 16 functional neighbourhoods according
to SAFE37.
Enrichment_global: Enrichment for all hits across 16 functional neighbourhoods
Enrichment_Strain: Enrichment for hits in a given genetic background across 16 functional
neighbourhoods.
TableS5- Genetic interaction degree and dN/dS values and gene classifications
21
References
1. Cooper, D. N., Krawczak, M., Polychronakos, C., Tyler-Smith, C. & Kehrer-Sawatzki, H. Where
genotype is not predictive of phenotype: towards an understanding of the molecular basis of reduced
penetrance in human inherited disease. Hum. Genet. 132, 1077–1130 (2013).
2. Chandler, C. H., Chari, S., Tack, D. & Dworkin, I. Causes and Consequences of Genetic Background
Effects Illuminated by Integrative Genomic Analysis. Genetics 196, 1321–1336 (2014).
3. Sackton, T. B. & Hartl, D. L. Genotypic Context and Epistasis in Individuals and Populations. Cell
166, 279–287 (2016).
4. Chow, C. Y. Bringing genetic background into focus. Nat. Rev. Genet. 17, 63–64 (2016).
5. Mullis, M. N., Matsui, T., Schell, R., Foree, R. & Ehrenreich, I. M. The complex underpinnings of
genetic background effects. Nat. Commun. 9, 3548 (2018).
6. Hou, J., van Leeuwen, J., Andrews, B. J. & Boone, C. Genetic Network Complexity Shapes
Background-Dependent Phenotypic Expression. Trends Genet. 34, 578–586 (2018).
7. Chen, R. et al. Analysis of 589,306 genomes identifies individuals resilient to severe Mendelian
childhood diseases. Nat. Biotechnol. 34, 531–538 (2016).
8. Fournier, T. & Schacherer, J. Genetic backgrounds and hidden trait complexity in natural
populations. Curr. Opin. Genet. Dev. 47, 48–53 (2017).
9. Chow, C. Y., Kelsey, K. J. P., Wolfner, M. F. & Clark, A. G. Candidate genetic modifiers of retinitis
pigmentosa identified by exploiting natural variation in Drosophila. Hum. Mol. Genet. 25, 651–659
(2016).
10. Steinberg, M. H. & Sebastiani, P. Genetic modifiers of sickle cell disease. Am. J. Hematol. 87, 795–
803 (2012).
11. Dorfman, R. Modifier gene studies to identify new therapeutic targets in cystic fibrosis. Curr. Pharm.
Des. 18, 674–682 (2012).
12. Cutting, G. R. Modifier genes in Mendelian disorders: the example of cystic fibrosis: Modifiers of
cystic fibrosis. Ann. N. Y. Acad. Sci. 1214, 57–69 (2010).
13. Williams, R. A., Mamotte, C. D. S. & Burnett, J. R. Phenylketonuria: an inborn error of
phenylalanine metabolism. Clin. Biochem. Rev. 29, 31–41 (2008).
14. Dowell, R. D. et al. Genotype to Phenotype: A Complex Problem. Science 328, 469–469 (2010).
15. Kim, D.-U. et al. Analysis of a genome-wide set of gene deletions in the fission yeast
Schizosaccharomyces pombe. Nat. Biotechnol. 28, 617–623 (2010).
16. Blomen, V. A. et al. Gene essentiality and synthetic lethality in haploid human cells. Science 350,
1092–1096 (2015).
17. Boutros, M. et al. Genome-wide RNAi analysis of growth and viability in Drosophila cells. Science
303, 832–835 (2004).
18. Hart, T. et al. High-Resolution CRISPR Screens Reveal Fitness Genes and Genotype-Specific Cancer
Liabilities. Cell 163, 1515–1526 (2015).
19. Vu, V. et al. Natural Variation in Gene Expression Modulates the Severity of Mutant Phenotypes.
Cell 162, 391–402 (2015).
20. Wang, T. et al. Identification and characterization of essential genes in the human genome. Science
350, 1096–1101 (2015).
21. Kamath, R. S. et al. Systematic functional analysis of the Caenorhabditis elegans genome using
RNAi. Nature 421, 231–237 (2003).
22. Paaby, A. B. et al. Wild worm embryogenesis harbors ubiquitous polygenic modifier variation. eLife
4, (2015).
23. Edwards, M. D., Symbor-Nagrabska, A., Dollard, L., Gifford, D. K. & Fink, G. R. Interactions
between chromosomal and nonchromosomal elements reveal missing heritability. Proc. Natl. Acad.
Sci. 111, 7719–7722 (2014).
22
24. Hou, J., Tan, G., Fink, G. R., Andrews, B. J. & Boone, C. Complex modifier landscape underlying
genetic background effects. Proc. Natl. Acad. Sci. 116, 5045–5054 (2019).
25. Galardini, M. et al. The impact of the genetic background on gene deletion phenotypes in
Saccharomyces cerevisiae. Mol. Syst. Biol. 15, (2019).
26. Peter, J. et al. Genome evolution across 1,011 Saccharomyces cerevisiae isolates. Nature 556, 339–
344 (2018).
27. Sadhu, M. J. et al. Highly parallel genome variant engineering with CRISPR-Cas9. Nat. Genet. 50,
510–514 (2018).
28. Michel, A. H. et al. Functional mapping of yeast genomes by saturated transposition. eLife 6, (2017).
29. Gangadharan, S., Mularoni, L., Fain-Thornton, J., Wheelan, S. J. & Craig, N. L. DNA transposon
Hermes inserts into DNA in nucleosome-free regions in vivo. Proc. Natl. Acad. Sci. U. S. A. 107,
21966–21972 (2010).
30. van Opijnen, T. & Levin, H. L. Transposon Insertion Sequencing, a Global Measure of Gene
Function. Annu. Rev. Genet. 54, 337–365 (2020).
31. Sharon, E. et al. Functional Genetic Variants Revealed by Massively Parallel Precise Genome
Editing. Cell 175, 544-557.e16 (2018).
32. Hou, J. et al. The Hidden Complexity of Mendelian Traits across Natural Yeast Populations. Cell
Rep. 16, 1106–1114 (2016).
33. Fournier, T. et al. Extensive impact of low-frequency variants on the phenotypic landscape at
population-scale. eLife 8, e49258 (2019).
34. Harari, Y., Ram, Y., Rappoport, N., Hadany, L. & Kupiec, M. Spontaneous Changes in Ploidy Are
Common in Yeast. Curr. Biol. 28, 825-835.e4 (2018).
35. Venkataram, S. et al. Development of a Comprehensive Genotype-to-Fitness Map of Adaptation-
Driving Mutations in Yeast. Cell 166, 1585-1596.e22 (2016).
36. Johnson, M. S. et al. Phenotypic and molecular evolution across 10,000 generations in laboratory
budding yeast populations. eLife 10, e63910 (2021).
37. Costanzo, M. et al. A global genetic interaction network maps a wiring diagram of cellular function.
Science 353, (2016).
38. Costanzo, M. et al. Environmental robustness of the global yeast genetic interaction network. Science
372, eabf8424 (2021).
39. Martin, M. Cutadapt removes adapter sequences from high-throughput sequencing reads.
EMBnet.journal 17, 10 (2011).
40. Li, H. & Durbin, R. Fast and accurate short read alignment with Burrows-Wheeler transform.
Bioinforma. Oxf. Engl. 25, 1754–1760 (2009).
41. Giaever, G. et al. Functional profiling of the Saccharomyces cerevisiae genome. Nature 418, 387–
391 (2002).
HygMX
L-TIR
R-TIR
Hermes transposase
GalSpr
pSTHyg
Isolate with Hermes
transposon system
Galactose induction
and competition
Pooled individuals
with random insertions
Pooled DNA
Fragmentation
Circularization
PCR library with
insertion sites
S288C (ref.)
Essential
Non-essential
Transposon insertion analyses
across
107 diverse genetic backgrounds
Relative
insertion
abundance
ATG
STOP
ORF
Feature
extractions
Prom.
CDS
End of
CDS
Logistic
regression
model
Probability
1
0
Gene annotations
Ess. by Chr.
Genetic backgrounds
Probability
Non-ess.
Ess.
Detecting
large-scale
duplications
1
0
Model fitting
for all 106
backgrounds
Aneuploidy
Whole-genome
duplication
Data
cleanup
Remove low
coverage data
Final dataset
39 diverse
genetic backgrounds
&
4469 genes
737 essentials vs 3732 non-essentials
(S288C annotation)
KNN
imputation
Quantile
normalization
A.
B.
D.
E.
C.
Figure 1
#Insertions
Clade
1. Wine/European
13. African palm wine
2. Alpechin
21. Ecuadorean
23. North American oak
26. Asian fermentation
3. Brazilian bioethanol
4. Mediterranean oak
M1. Mosaic region 1
M2. Mosaic region 2
M3. Mosaic region 3
Essentiality (S288c)
Non-essential
0.2
0.4
0.6
0.8
Probability (Non-Ess.)
#Insertions
100K
150K
200K
250K
300K
Unclassified
Essential
0.0
1.0
Gal/Resp. related (S288c)
Others
Galactose sick or Petite
Clade
Translation;
Ribo. biogen.;
rRNA processing
Mitochondrial
translation
No enrichment
Bioprocess
Ribo. biogen.;
rRNA processing;
DNA replication;
Protein transport;
Cell cycle
Transcription
regulation
AMF
ANG
BFP
BKL
ACV
CGD
ACT
CHM
AVI
BHH
CCD
ABP
ADD
APH
CGQ
CLG
BID
CQA
BVF
ACH
ABS
BQH
BNT
BCV
ADH
ANH
CLB
BBQ
ACP
CPG
BDH
ACN
CNT
ACF
CIB
AND
BMK
Gal/Resp.
Ess.
S288c
Ʃ1278b
Mitochondrial
translation
Figure 2
ANG
BVF
BFP
ABS
AMF
BID
ADH
ACH
BNT
CQA
BQH
CGD
BKL
BDH
BCV
CLB
ACV
ACT
ANH
CHM
BBQ
CIB
CNT
AND
ACP
Ʃ1278b
ACN
CPG
AVI
BMK
ADD
APH
BHH
ACF
ABP
CLG
CGQ
CCD
0
20
60
80
20
40
40
Number of genes with
background-dependent fitness
Genetic backgrounds
S288C Ess./Sick
S288C Non-ess./Healthy
0
20
60
10
20
30
count
0
50
150
0
10
20
count
0
58
132
Distribution of genes with
background-dependent fitness
Number of genetic backgrounds
Number of genetic backgrounds
S288C Ess./Sick
S288C Non-ess./Healthy
A.
B.
C.
30
Gal/Respiration related
Others
−6
−3
0
3
6
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20 21 23 24 25 26 27 28 31 32 33 34
Z-score of probability
S288C Ess./Sick
S288C Non-ess./Healthy
Background-specific healthy/non-essential
Background-specific sick/essential
Possibly Environment-related
Number of genetic backgrounds
Background-specific
Figure 3
A.
Network of environment-related
background-dependent fitness genes
Ess.
−0.5
0
0.5
Essentiality (S288c)
Non-essential
Essential
Gal/Resp. related (S288c)
Others
Galactose sick or Petite
Gal/Resp.
Profile similarity
Pearson’s correlation
Profile similarity of environment-related
genes across 39 backgrounds
Anti-correlation (R2 < -0.6)
Correlation (R2 > 0.6)
B.
Figure 4
B.
Metabolism
Cytokinesis
MVB sorting and pH dependent signaling
tRNA wobble modification
Protein degradation/turnover
Nuclear−cytoplasmic transport
mRNA & tRNA processing
rDNA & ncDNA processing
Vesicle traffic
Ribosome biogenesis
Respiration, oxidative phosphorylation,
mitochondrial targeting
DNA replication & repair
Glycosylation, protein folding/targeting,
cell wall biosynthesis
Mitosis & chromosome segregation
Transcription & chromatin organization
Cell polarity & morphogenesis
CCD
CGQ
CLG
ABP
ACF
BHH
APH
ADD
BMK
AVI
CPG
ACN
ACP
AND
CNT
CIB
BBQ
CHM
ANH
ACT
ACV
CLB
BCV
BDH
BKL
CGD
BQH
CQA
BNT
ACH
ADH
BID
AMF
ABS
BFP
BVF
ANG
Environment-
related
Background-
specific
Odds ratio
10
20
30
2
3
Ʃ1278b
Background-centric enrichments
Global enrichment
Odds ratio
P-value < 0.05
Background
Environment
A.
C.
Gain-of-fitness/S288C
Loss-of-fitness/S288C
S288C-like
Mitochondrial only
Mito+Nuclear transport
Mito+Transcription
S288C-reverse
0.25
0.50
0.75
Fitness
S288C modeled fitness
Nuclear−cytoplasmic transport
Respiration, mitochondrial targeting
Transcription & chromatin organization
AMF
BNT
CHM
CLG
ACN
Figure 5
B.
C.
Genetic interaction degrees
Co-annotation rate
dN/dS (YN00)
A.
Non−Essential
Background-
specific
Envr. related
Essential
ns
****
****
****
****
****
0
500
1000
ns
****
****
****
*
****
0.0
0.5
1.0
Non−Essential
Background-
specific
Envr. related
Essential
ns
ns
****
****
ns
****
0
1
2
Non−Essential
Background-
specific
Envr. related
Essential
2776
125
268
508
2776
125
268
508
2776
125
268
508
Figure 6
A.
3
4
5
6
7
3
4
5
Number of unique insertion sites (log10)
Number of reads (log10)
B.
C.
D.
E.
106 isolates
0
25
50
75
nnnnnnAn
nTnnnnAn
nTnnnnnn
Percentage of reads with the motif
S288C
Motifs
0
1
2
3
4
5
(−500,−400]
(−400,−300]
(−300,−200]
(−200,−100]
(−100,0]
(0,100]
(100,200]
Window (-500 bp to +200 bp from ATG)
Average insertion number
0
1
2
3
4
(0,0.2]
(0.2,0.4]
(0.4,0.6]
(0.6,0.8]
(0.8,1]
Fraction of CDS
Average insertion number
0
1
2
3
(−200,−100]
(−100,0]
(0,100]
(100,200]
(200,300]
(300,400]
(400,500]
Average insertion number
Window (-200 bp to +500 bp from STOP)
Annotation
Non-Ess
Ess.
Figure S1
A.
0.00
0.25
0.50
0.75
1.00
0
10
20
30
40
50
60
70
80
Probability
Annotation
Non-Ess
Ess.
[0,50]
(250,300]
(500,550]
(750,800]
(1.05e+03,1.1e+03]
(1.5e+03,1.55e+03]
(2e+03,2.05e+03]
(2.7e+03,2.75e+03]
Number of insertions in +/- 5kb of CDS
Annotation
Non-Ess
Ess.
0.00
0.25
0.50
0.75
1.00
Number of insertions in STOP+300 bp
Probability
B.
C.
25
50
75
1000
2000
3000
4000
5000
Number of genes per strain without low coverage
Number of strains retained
3000
4000
5000
6000
1000
2000
3000
4000
5000
Number of genes retained
Number of genes per strain without low coverage
D.
Figure S2
0.75
0.50
0.25
Probability
I
II
III
IV
V
VI
VII
VIII
IX
X
XI
XII
XIII
XIV
XV
XVI
A.
C.
727500
728000
YBR256C RIB5
498500
499500
YGR002C SWC4
257600
258000
258400
YIL051C MMF1
19000
19500
20000
20500
YML126C ERG13
1154000
1155000
1156000
YDR342C HXT7
1160000
1161000
YDR343C HXT6
B.
292500
292750
293000
293250
293500
YDL092W SRP14
817000
817500
YDR177W UBC1
250000
250500
251000
251500
252000
YFR051C RET2
90000
90500
91000
YGL213C SKI8
90000
91000
92000
93000
YIL137C TMA108
197000
197500
198000
YLR027C AAT2
D.
S288C-specific essential
Ʃ1278b-specific essential
Ʃ1278b
S288C
Ʃ1278b
S288C
Clathrin adaptor
E2-ubiquitin conjugating
Signal recognation particle
ERAP-1 like C-term
WD40
WD40
235000 236000 237000 238000
YOL051W GAL11
80000
81000
82000
YPL248C GAL4
BMK
AND
ANH
CIB
ACF
AVI
CNT
ACN
BDH
CHM
CGQ
APH
CCD
BHH
ADD
ABP
CLG
CPG
CLB
ADH
BCV
BBQ
AMF
BQH
ABS
ACH
BVF
BNT
BID
CQA
ACT
BKL
ACV
CGD
ACP
BFP
ANG
AKE−re2
AKE−re1
ABC
CNM
CMT
BAK
BFP
ANG
AKH
ABA
BAP
CIA
BFQ
Ʃ1278b
S288C
Figure S3
Anti-correlation (R2 < -0.6)
Correlation (R2 > 0.6)
Others
Galactose sick or Petite
A.
B.
C.
D.
AEP3
AAP1
AMD1
ALY1
ARC1
BEM2
BLM10
CBP2
ACC1
AAT2
AEP1
COQ1
COQ2
ACB1
BMH1
CBP3
COQ8
COR1
CAF130
CHS1
CSF1
COX15
BUD3
ADH1
ACK1
DIA4
ARP8
EPL1
AZF1
BPH1
CMR2
EFB1
FZO1
GCN1
AFG3
ESL1
GCN2
CCS1
FOL2
GEA2
EXO5
DAN4
FAB1
ASE1
HER1
HFA1
FUN26
IKI3
GEA1
IRR1
CRT10
BRN1
CDC48
DYN1
GEM1
ISM1
KAP120
LST4
GLN1
CHD1
COQ6
MCD4
MCM6
LAM1
GTF1
GUP1
MHR1
GYP6
IMD4
GLY1
MEF1
COX19
MIP1
LPD1
MRH4
LUC7
MRP1
MRP2
MRM1
MRP4
MRP7
MRPL11
MRPL13
MRPL16
FUS3
MRPL22
MDR1
MRPL3
MRPL24
MRPL28
MRPL37
MRPL4
MRPL49
MRPS12
ENV9
MSE1
MRF1
MSK1
MSM1
MRPS16
MGA2
MSH1
CCC2
ISD11
MSR1
MSS2
KAP123
MMS1
MRPL8
LDB18
OAR1
PET122
PET494
AUS1
GET2
PMT1
NUP188
PAN1
PMC1
MIT1
MNR2
LAA1
POM152
RAD26
POL2
NUP57
NAM2
RMD9
PAT1
RKR1
MSS51
BUR6
KAP122
RAP1
NGG1
RRN7
IRC19
PET130
RFC1
RSM19
PET111
PET8
QCR2
RTC6
RET1
RPL18A
MYO4
PDR18
RML2
RSM23
NST1
RRN6
SEC63
SPT16
ARP7
RPH1
SPT7
RPO41
SPT6
CCH1
IST2
MSB2
REC114
SRB8
IFM1
ARP6
SQS1
RSM24
BPT1
FUN30
NNF1
NUP100
SPO22
RPL4A
MUB1
STH1
TCB1
PHB1
SUV3
STT4
TEL1
ERG4
MET14
SYO1
RRN5
SNT1
TUP1
SKT5
NAM9
PLB1
SIN3
SSH1
TOR2
VAM6
VPS13
STB2
UBP15
PEX30
URA3
GNA1
NBL1
PKR1
YML6
MIS1
UBP2
SLS1
TDA5
YOR296W
CDC5
MET30
PTA1
VMA13
ESC1
TOF2
VPS35
YHR182W
ERG3
GPR1
KEX2
KRE5
KRE6
MEF2
MER1
MLP1
MRPL51
NAT1
NCE103
PAN3
PMT2
PRM2
PRO3
PTC1
PTK2
QRI5
RCY1
RPL35B
RRP6
RSM25
SCM3
SCO1
SFB3
SHE3
SIP4
SLM5
SNF4
SPT23
SPT8
SSM4
STE5
STI1
SWC4
SWC5
TFB4
THI3
TKL1
TMA19
TOM20
TOP1
TPS3
TRK1
TUF1
UBR2
UCC1
UGP1
UME1
URA4
URA7
VAS1
VPS34
VPS41
XRN1
YBR298C−A
YEL074W
YHL030W−A
YJL193W
YLR412C−A
YMR317W
YNL040W
YNR065C
YOR333C
YPK9
YPP1
YPR117W
YPT7
YTA7
ZPR1
Figure S4
Nuclear−cytoplasmic
transport
Respiration, oxidative phosphorylation,
mitochondrial targeting
Transcription &
chromatin organization
A.
B.
C.
BMH1
∆bmh1/BMH1 CIB
∆bmh1/BMH1 S288C
T1 T2 T3 T4 T5
T1 T2 T3 T4 T5
9.2e−06
1.0
1.2
1.4
CIB
S288C
Relative fitness ratio
WT/∆bmh1
∆bmh1
MPS3
KAP120
NDC1
NUP57
NIC96
MLP1
SHE3
KAP123
NUP157
NUP100
SYO1
LRP1
POM152
NUP188
MYO4
RRP6
PHB1
COQ8
QCR2
FZO1
COR1
AFG3
NAM2
RRG1
IFM1
MRPL10
ATP10
SPO22
PET111
ATP25
CBP3
MRPL28
MSF1
RSM25
AEP3
MRPL11
MRPL4
COQ2
COQ6
MDM38
MRPL22
PET494
MSE1
MRP1
MRPL49
MRPL13
CBP2
NAM9
MRPL36
MSS51
COX19
MRPL51
MRPL37
RTC6
SUV3
MRPL3
AEP1
ATP18
OXA1
RML2
MEF2
MRPL24
TUF1
MSR1
SLS1
YML6
MRP2
RSM23
MRPS12
MRP4
MRPS16
PET8
RSM19
MSK1
PET122
MGA2
IES6
GAL11
RIC1
MET30
SNT1
TFC3
STH1
SST2
RPB11
PHO23
ELP3
TOG1
SDS3
SRB8
AZF1
ARP7
MAC1
RET1
SIF2
SIN3
UME1
TFB4
BUR6
MED7
SNF4
GIS1
RAP1
EGD1
NUT1
SIP4
SPT7
RCO1
SWC5
TUP1
ELP2
EPL1
IRE1
SPT23
IKI3
ARP8
HAC1
ESS1
SOK2
THI3
IXR1
SPT8
CRT10
SFL1
RPH1
SIR3
SPT3
SWC4
ARP6
MHR1
CHD1
CAF130
RPO41
YAF9
RGT1
YTA7
NGG1
BMH1
MRM1
MRPL16
COQ1
MRPL8
MRF1
MEF1
MSS2
MRP7
GTF1
GEM1
EXO5
RSM24
MRPS5
CBP1
SCO1
ISM1
ABP
ABS
ACF
ACH
ACN
ACP
ACT
ACV
ADD
ADH
AMF
AND
ANG
ANH
APH
AVI
BBQ
BCV
BDH
BFP
BHH
BID
BKL
BMK
BNT
BQH
BVF
CCD
CGD
CGQ
CHM
CIB
CLB
CLG
CNT
CPG
CQA
Ʃ1278b
CIB
Figure S5
| 2021 | Population-level survey of loss-of-function mutations revealed that background dependent fitness genes are rare and functionally related in yeast | 10.1101/2021.08.25.457624 | [
"Caudal Elodie",
"Friedrich Anne",
"Jallet Arthur",
"Garin Marion",
"Hou Jing",
"Schacherer Joseph"
] | creative-commons |
QTL mapping in an interspecific sorghum population uncovers candidate regulators of
salinity tolerance
Ashley N. Hostetler1, Rajanikanth Govindarajulu1,2, Jennifer S. Hawkins1
West Virginia University1
53 Campus Drive
Department of Biology
West Virginia University
Morgantown, WV 26505
Eurofins Lancaster Labs2
601 E. Jackson St.
Richmond, VA 23219
Author for correspondence:
Ashley N. Hostetler
ahende11@mix.wvu.edu
53 Campus Drive
Department of Biology
West Virginia University
Morgantown, WV 26505
Keywords – aquaporins, genetic map, plasma-intrinsic proteins, recombinant inbred line,
salinity tolerance, Sorghum bicolor, Sorghum propinquum, stress tolerance index,
quantitative trait loci
Abstract
Salt stress impedes plant growth and disrupts normal metabolic processes, resulting in
decreased biomass and increased leaf senescence. Therefore, the ability of a plant to
maintain biomass when exposed to salinity stress is critical for the production of salt
tolerant crops. To identify the genetic basis of salt tolerance in an agronomically
important grain crop, we used a recombinant inbred line (RIL) population derived from
an interspecific cross between domesticated Sorghum bicolor (inbred Tx7000) and a wild
relative, Sorghum propinquum, which have been shown to differ in response to salt
exposure. One-hundred seventy-seven F3:5 RILs were exposed to either a control or salt
treatment and seven traits related to plant growth and overall health were assessed. A
high-density genetic map that covers the 10 Sorghum chromosomes with 1991 markers
was used to identify nineteen total QTL related to these traits, ten of which were specific
to the salt stress response. Salt-responsive QTL contain numerous genes that have been
previously shown to play a role in ionic tolerance, tissue tolerance, and osmotic tolerance,
including a large number of aquaporins.
Introduction
Soil salinity imposes abiotic stress on plants when soluble ions, such as Na+ and Cl-,
accumulate in the soil surrounding the root rhizosphere. Saline soils have been found to
affect more than 6% of total land and 20% of irrigated land (Rengasamy, 2006; Food and
Agriculture Organization (FAO), 2008, 2009; Hasegawa, 2013; FAO, 2017; Yang et al.,
2020). Although most of the lands with elevated salinity have arisen from natural causes,
anthropogenic factors have recently led to the increase of salinity in lands cultivated for
agriculture (Munns and Tester, 2008).
The level of salt in the soil directly affects the production of agriculturally important
crops, yet crop tolerance varies depending on species and genotype (Munns and Tester,
2008). Previous research has identified the key mechanisms of salt tolerance as: 1) ion
exclusion, which results in the absence of salt ions from the shoot of the plant, 2) tissue
tolerance, achieved by the sequestration of ions into specific tissues, cells, and subcellular
organelles, and/or 3) osmotic tolerance, defined as the ability to maintain water uptake
and growth despite lower water potential (Munns & Tester, 2008; Carillo et al., 2011;
Fan et al., 2015; Genc et al., 2016; Negrão et al., 2017; Munns et al., 2019). In order to
incorporate tolerant genotypes into breeding programs, it is first necessary to determine if
tolerance is a result of ionic tolerance, tissue tolerance, and/or osmotic tolerance;
however, the genetic basis of these mechanisms remains unknown.
Sorghum (Sorghum bicolor L. Moench) is a staple crop for food, fuel, and feed
production (Doggett, 1970, 1988), and is notable for varieties that are naturally drought
and salt tolerant (Boursier & Läuchli, 1990; Almodares & Sharif, 2007; Almodares et al.,
2007, 2008b,a; Mullet et al., 2014; Fracasso et al., 2016; McCormick et al., 2018; Guo et
al., 2018; Henderson et al., 2020). In a previous study, we measured the variation in
salinity tolerance across a diverse panel of Sorghum genotypes that included wild species,
domesticated S. bicolor landraces, and improved S. bicolor lines (Henderson et al.,
2020). Salinity tolerance was assessed as the maintenance of biomass following 12 weeks
of 75 mM salt exposure (sodium chloride, NaCl). After long-term exposure to salinity, S.
bicolor genotypes ranged from 30% - 95% in biomass maintenance when compared to
genotypes in non-saline (control) conditions. The genotype that had the greatest reduction
in biomass upon exposure to salt was the wild species S. propinquum, which maintained
only 5% of its aboveground biomass (Henderson et al., 2020). Additionally, we also
measured the accumulation of sodium and potassium in genotypes that spanned the
tolerance rankings. The results showed that sodium and potassium accumulation was
accession dependent and did not correlate tolerance ranking, suggesting that the salt
response in sorghum is reliant on multiple mechanisms. We concluded that sorghum
serves as a valuable resource for dissecting the various underlying genetic controls of
salinity tolerance.
The observed variation in biomass retention upon exposure to saline conditions
(Henderson et al., 2020) indicates that there is quantitative genetic variation in salinity
tolerance in Sorghum. In the work presented here, a recombinant inbred line (RIL)
population constructed from a cross between S. propinquum (95% biomass reduction)
and S. bicolor (Tx7000 - landrace durra; 5% biomass reduction) was used to dissect the
genetic underpinnings of salinity tolerance. We developed a high-density genetic map
from 177 F3:5 RILs and identified quantitative trait loci (QTL) associated with biomass-
related traits during salt exposure. These findings and this population establish a resource
that can be used to further dissect the underlying genetic basis of ionic tolerance, tissue
tolerance, and osmotic tolerance.
Materials and Methods
Plant Material
A RIL mapping population derived from an interspecific cross of S. propinquum and S.
bicolor (inbred Tx7000, landrace durra) was used to investigate the genetic
underpinnings associated with variation in salinity tolerance. The RIL population consists
of 177 F3:5 lines with 75% (132 RILs) of the individuals being F5, 18% (31 RILs) of the
individuals being F4, and 7% (14 RILs) of the individuals being F3. Each line was derived
by the single seed descent method (Brim, 1966; Snape & Riggs, 1975) as described in
Govindarajulu et al. (2020).
Experimental Conditions
In a controlled greenhouse room, five seeds of each RIL were sown in 5 cm x 5 cm x 5
cm planting plugs filled with metromix soil. Target germination conditions were 21℃,
75% humidity, and 4.5 vapor pressure deficit (VPD). During germination, seedlings were
misted regularly with non-saline tap water and watered with a 20-10-20 N-P-K fertilizer
(J.R. Peters, Inc., Allentown, PA, USA) diluted to 200 mg N L-1 every 4th day. Once all
plants reached at least the third leaf stage (V3) of development (approximately 32 days
post sowing), soil plugs were transplanted into 5 cm x 5 cm x 25 cm treepots (Stuewe and
Sons, Tangent, OR, USA) filled with silica sand #4. Target growth conditions were 27℃
day/23℃ night with 16 h ambient daylight and 75% humidity. Following transplant,
seedlings were watered to saturation with non-saline tap water daily for two weeks to
provide a period of establishment. All plants were fertilized twice weekly with a 20-10-
20 N-P-K fertilizer (J.R. Peters, Inc., Allentown, PA, USA) diluted to 200 mg N L-1 for
the remainder of the study. Following establishment, three of the five biological
replicates were randomly assigned to a 75 mM NaCl salt treatment and two of the five
biological replicates were randomly assigned to a 0 mM NaCl control treatment.
Seedlings were watered once daily, in accordance with their assignment, to complete
saturation for the duration of the experiment. Treatment began 51 days after planting and
plants were treated for a total of 45 days.
Phenotypic Measurements
The following phenotypes were measured for each of the 177 lines: height (cm), rank
score, root biomass (g), dead aboveground biomass (g), live aboveground biomass (g),
total aboveground biomass (g), and total biomass (g). Height (cm) was taken from the
base of the stem to the tip of the newest emerged leaf. Rank score was a qualitative score
that described overall leaf ‘greenness’, leaf health, and mortality, where plants that
displayed no signs of stress received a low rank score, and plants that were extremely
stressed or had died received a high rank score (Table 1). Rank score was assessed by the
same person throughout the entirety of study to minimize bias. All biomass
measurements were taken on tissue collected from a destructive harvest and dried in 65℃
for a minimum of 72 h. Root biomass (g) was the total belowground biomass collected.
Roots were rinsed in water to remove all dirt and sand. Dead aboveground biomass (g)
included all biomass (leaves, tillers, and/or stem) attached to the plant where more than
50% of the tissue was brown; whereas live aboveground biomass (g) included all biomass
attached to the plant that was more than 50% green suggesting it was alive. Total
aboveground biomass (g) was the sum of live and dead aboveground biomass, while total
biomass (g) included live, dead, and root biomass. Mortality was scored as 1 if plants
were alive and 0 if dead.
Phenotypes were measured at three time points: 0 days (51 days after planting, 0 days of
treatment, referred to as pre-treatment), 15 days (short term exposure), and 45 days (long
term exposure) after treatment began. Height was taken at 0, 15, and 45 days after
treatment began, with 0 days indicating immediately before treatment. Rank score was
taken at 15 and 45 days after treatment began. All biomass was collected between 45-50
days after treatment and was immediately dried.
The stress tolerance index (STI) value is an valuable metric when comparing genotypic
tolerance within a population (Negrão et al., 2017). The STI value for each trait was
calculated using the following formula, where Y is the phenotypic trait, control is the trait
measurement in 0 mM NaCl conditions, salt is the trait measurement in 75 mM NaCl
conditions, and control average is the population average of the trait in control conditions
(Negrão et al., 2017):
𝑆𝑡𝑟𝑒𝑠𝑠 𝑇𝑜𝑙𝑒𝑟𝑎𝑛𝑐𝑒 𝐼𝑛𝑑𝑒𝑥 = (
𝑌𝑐𝑜𝑛𝑡𝑟𝑜𝑙
𝑌 𝑐𝑜𝑛𝑡𝑟𝑜𝑙 𝑎𝑣𝑒𝑟𝑎𝑔𝑒) × (
𝑌𝑠𝑎𝑙𝑡
𝑌 𝑐𝑜𝑛𝑡𝑟𝑜𝑙 𝑎𝑣𝑒𝑟𝑎𝑔𝑒)
For each RIL, the control value was averaged across control replicates and an STI value
was calculated for each salt treated plant. The STI value accounts for the overall
performance of the population and compares each RIL’s ability to maintain performance
under stress conditions. RILs with large STI values indicate larger phenotypic values for
a given trait and are often considered tolerant depending on the phenotype.
Statistical analysis of phenotypic values
All statistical analyses were performed on the control population, the salt treated
population, and on STI values. Three biological replicates of each RIL in the salt
population, three biological replicates for STI values, and two biological replicates of
each RIL in the control population were considered for QTL analysis. All statistics and
graphing were completed using R version 4.0.2 (R Core Team, 2013).
Least square means for each phenotype in each population (control, salt, STI) were
calculated for every RIL. Normality was assessed using both a Shapiro-Wilk test in R and
Q-Q plots from the car package version 3.0.10 (Fox & Weisberg, 2019). Traits that were
not normally distributed were transformed (Supplementary Table S1). Transformed
values were used in statistical tests and in QTL analysis. Correlations of phenotypes
within each treatment (control or salt) were assessed via a Pearson’s Correlation analysis
in R using the PerformanceAnalytics package 2.0.4 (Peterson & Carl, 2019) (Figure S1).
To determine if there was a treatment effect, both a nonmetric multidimensional scaling
(NMDS) analysis (Julkowska et al., 2019) and an analysis of variance (ANOVA) was
performed. The NMDS ordination clustered individuals based on Bray–Curtis
dissimilarity measures when all phenotypes were considered. The dimcheckMDS
function in the goeveg package version 0.4.2 was used to generate stress values for each
dimension; two dimensions was deemed appropriate. The NMDS was generated using the
vegan package version 2.5.6 (Oksanen et al., 2019) in R. The NMDS was paired with an
analysis of similarity (ANOSIM) to statistically test the ordination results from the
NMDS. Here, we tested if RILs were more similar within or between treatments
(significance assessed at 𝛼 = 0.05). An analysis of variance (ANOVA) was used to test if
control and salt populations differed for individual phenotypes (significance assessed at 𝛼
= 0.05).
Genetic Map Construction and QTL analysis
The RIL population used in this study was generated as previously described in
Govindarajulu et al. (2020). This population has been successful in identifying genes that
are key regulators of tiller elongation in sorghum (Govindarajulu et al., 2020); however,
in this study, advanced lines were included, therefore requiring a new genetic map. A
new genetic map for the S. propinquum by S. bicolor (Tx7000) RIL population was
constructed as previously described in Govindarajulu et al. (2020). In summary, using
high-quality nuclear DNA, the parent plants (S. propinquum and S. bicolor) were
sequenced at 18x depth, while the RILs were sequenced at 2x depth. SNP data were
aligned to the masked Sorghum bicolor reference genome version 3.1 (Paterson, 2008).
Loci were called as S. propinquum (A), S. bicolor (B), or heterozygous (H) when SNPs
were analyzed with the GenosToABHPlugin in Tassel ver 5.0 (Bradbury et al., 2007).
The ABH formatted SNP data file was then used as input to SNPbinner (Gonda et al.,
2019), which calculated breakpoints (Govindarajulu et al., 2020). Breakpoints were
merged if they were shorter than 0.2% of the chromosome length. After removing
heterozygous bin markers and duplicate bin markers, the kosambi map function in R/qtl
(Broman et al., 2003) was used to construct a high density genetic map.
QTL analysis was performed in R using the qtl package version 1.46.2 (Broman et al.,
2003). QTL were first identified by a single interval mapping QTL model. Significant
LOD peak scores were determined by comparing LOD peak scores after a 1,000
permutation test (𝛼 = 0.05) (Churchill & Doerge, 1994). If QTL were detected by interval
mapping (IM), phenotypes were assessed via a multiple QTL model (MQM). The MQM
tested for additional additive QTL, refined QTL positions, and tested for epistasis.
Following MQM, a type III analysis of variance assessed the significance of fit for the
final model, the proportion of variance explained, and the additive effect. QTL with a
negative additive value indicated that the trait was negatively influenced by S. bicolor
alleles, whereas a positive additive value indicated that the trait was positively influenced
by S. bicolor alleles. Genes (Sorghum bicolor ver. 3.1) within a 1.0 logarithm of the odds
(LOD) confidence interval for each QTL were identified.
Aquaporin Enrichment Analysis
To determine if there was an enrichment of aquaporin genes (SbAQP) located within
detected QTL, we randomly sampled the Sorghum bicolor (version 3.1) genome 50 times
in R (R Core Team, 2013). Starting positions were determined by first randomly selecting
a chromosome followed by selecting a random starting position on that chromosome.
Chromosome selection was determined by multiplying the total number of chromosomes
(2n = 10) by a randomly generated number [runif(1)] between 0 and 1. The resulting
number was truncated to an integer and represented the starting chromosome number.
This same process was repeated to determine the starting location on the chromosome;
however, instead of multiplying a random number by the total number of chromosomes,
the random number (between 0 and 1) was multiplied by the size of the corresponding
chromosome. The result was again truncated and represented the starting location in the
genome (in Mb). Genes within a 5 Mb window (5 Mb downstream from the starting
location) were extracted (Sorghum bicolor version 3.1)
Results
A high density genetic map covers 10 Sorghum chromosomes with 1991 makers
The resequencing, SNP calling, and bin calling of the 177 RILs generated 4055 total bin
markers (Figure S2). After removing duplicate markers, the map covered the 10
Sorghum chromosomes with 1991 markers (Supplementary Table S2) and was 913.71
cM in length (Supplementary Table S3).
S. bicolor and S. propinquum exhibit a differential response to salt exposure
Previous work from our lab described the variation in biomass retention following 12
weeks of 75 mM NaCl in the domesticated, S. bicolor, and the wild progenitor, S.
propinquum (Henderson et al., 2020). In response to salt treatment S. bicolor, landrace
durra, maintained 95% of its biomass, had high STI values for live aboveground
biomass, and had low STI values for dead aboveground biomass, whereas S. propinquum
maintained only 5% of its biomass, ranked low for live aboveground biomass, and ranked
high for dead aboveground biomass. Variation in tolerance in S. bicolor was genotype
dependent, and when tolerance rankings were evaluated within a phylogenetic
framework, all of the accessions from the landrace durra ranked as tolerant (~95%
biomass retained). The findings from our previous study (e.g., S. bicolor, specifically the
landrace durra, was shown to be tolerant to salinity stress whereas S. propinquum was
shown to be sensitive) set the foundation for the parental genotypes of the RIL population
used here.
Overall plant health decreased in response to salt exposure
In both the control and salt treated populations, the following phenotypes were recorded:
height (cm), rank score, root biomass (g), dead aboveground biomass (g), live
aboveground biomass (g), total aboveground biomass (g), total biomass (g), and
mortality. With the exception of mortality, all phenotypes were significantly different
between the control and salt treated populations (in both short term and long-term
exposure) (Table 2). Generally, the response to salt exposure was characterized as shorter
plants with reduced live aboveground biomass, root biomass, total aboveground biomass,
total biomass (Table 2, Supplementary Tables S4). Further, plants had larger rank
scores and more dead aboveground biomass (Table 2, Supplementary Tables S4).
These significant differences between RILs in control conditions and salt treated
conditions, in addition to the clear and significant clustering (p < 0.001, ANOSIM
R=0.16) of treatments in the NMDS analysis (Figure 1), are indicative of an overall
decrease in plant health in response to salt.
In response to long-term salt exposure, QTL were identified as important regulators
of salt tolerance
Although there was variation in plant response among RILs in the control and salt treated
populations after short term salt exposure (DAT=15) (Table 2), there were no QTL
detected. Therefore, the following analyses focus on QTL detected after long-term salt
exposure (DAT=45). QTL analysis was performed on RILs in the control treated
population, RILs in the salt treated population, and on the stress tolerance index (STI)
values. The individual QTL analysis performed on the RILs in the control and salt treated
populations was to determine QTL that were shared. QTL that were shared between the
control and salt treated populations are related to trait architecture and not the salt
specific response. All genes detected in QTL are listed in Supplementary Table S5.
Total biomass (TB)
In control conditions, plants ranged from 1.98 g to 12.18 g of total biomass, with a mean
of 7.49 g; however, in response to salt treatment, the total biomass decreased on average
32% (0.66 g - 9.42 g with a mean of 5.11 g, Table 2, Supplementary Table S4). There
were no QTL detected when the control population was mapped. This is likely due to
limited variation in architecture and biomass between S. bicolor and S. propinquum when
plants are grown in space restricted pots (data not shown); however, because S. bicolor is
tolerant and S. propinquum is sensitive to salt exposure (Henderson et al., 2020), we
suspect that salt specific QTL for total biomass were identified (qTB45_4.S and
qTB45_4.STI) (Figure 2) because genotypes with S. propinquum alleles had a greater
reduction in biomass in response to treatment, whereas genotypes that contain S. bicolor
alleles maintained biomass. Both QTL (qTB45_4.S and qTB45_4.STI) were detected on
chromosome 4 and both had positive additive effects indicating the S. bicolor alleles
positively influence the total biomass (Table 3). This indicates that RILs with S. bicolor
alleles in the 61.70 Mb - 68.41 Mb region on chromosome 4 have more overall biomass
(live aboveground biomass, dead aboveground biomass, and root biomass) after long-
term exposure to NaCl. The QTL detected when STI values were mapped (qTB45_4.STI)
explained the greatest amount of phenotypic variation (PVE = 13.02) (Table 3). Of the
genes within qTB45_4.STI, candidate genes were associated with aquaporins, stress
response proteins, salt tolerant proteins, and transporters (Supplementary Table S6).
Total aboveground biomass (TAGB)
In control conditions, plants accumulated an average of 5.00 g of total aboveground
biomass (1.34 g - 8.56 g); however, in response to salt treatment, there was an average
20% decrease in total aboveground biomass (average = 3.98 g) (Table 2,
Supplementary Table S4). Similar QTL were detected when the total biomass and the
total aboveground biomass were mapped as would be expected given the high correlation
between total biomass and total aboveground biomass (Supplementary Figure S1).
There were no QTL detected in the control population; however the same two QTL
(qTB45_4.S and qTB45_4.STI) were detected in the salt population (qTAGB45_4.S) and
when STI values (qTAGB45_4.STI) were mapped (Figure 2). qTAGB45_4.STI co-
localized with qTB45_4.STI and qTAGB45_4.S co-localized with qTB45_4.S.
Therefore, the same candidate genes identified in the total biomass QTL were also
identified for total aboveground biomass.
Dead aboveground biomass (DAGB)
In control conditions, dead aboveground biomass ranged from 0.06 g to 1.61 g (mean of
0.55 g), whereas in salt conditions there was an average increase of 45% with a mean of
0.80 g (Table 2, Supplementary Table S4). Genotypes that accumulate more dead
aboveground biomass, like S. propinquum (Henderson et al., 2020), are sensitive
compared to genotypes that do not. When dead aboveground biomass was mapped, two
QTL were detected on chromosome 2 (qDAGB45_2.C and qDAGB45_2.STI) and one
QTL was detected on chromosome 9 (qDAGB45_9.C) (Figure 2). RILs with S.
propinquum alleles positively influenced the amount of dead aboveground biomass. Of
the genes within qDAGB45_2.STI, candidate genes were associated with aquaporins,
sodium transporters, potassium transporters, salt tolerant proteins, and leaf senescence
(Supplementary Table S6).
Live aboveground biomass (LAGB)
In control conditions, live aboveground biomass ranged from 1.20 g to 6.95 g with an
average of 4.45 g (Supplementary Table S4). In response to salt treatment, there was an
average decrease of 28% in live aboveground biomass with a range from 0.32 g to 6.22 g
and an average of 3.18 g (Table 2). Two QTL were detected (qLAGB45_4.C and
qLAGB45_4.STI) on chromosome 4 (Figure 2). Both QTL had positive additive effects,
indicating that S. bicolor alleles positively influence live aboveground biomass. The QTL
detected when STI values were mapped (qLAGB45_4.STI) explained 11.65 percent of
phenotypic variation (Table 3).
Root biomass (RB)
The root biomass of plants grown in control conditions ranged from 0.48 g to 6.38 g with
a population average of 2.50 g; however, in salt conditions, the root biomass was reduced
by an average of 55% and ranged from 0.09 g to 2.48 g with a population average of 1.13
g (Table 2, Supplementary Table S4). A single QTL was detected when STI values
were mapped (qRB45_4.STI) (Figure 2). qRB45_4.STI explained 11.40 percent of the
phenotypic variation and had an additive effect of 0.08, indicating that individuals with S.
bicolor alleles in this region positively influenced root biomass.
Rank score (RS)
Rank score was a qualitative measure used to describe overall plant health (Table 1).
There was an average 58% increase in rank score in response to treatment, indicating that
there was an overall decrease in the health of plants exposed to NaCl (Table 2). In
control conditions, the average rank score of the population was 2.01 (0.72-3.20)
indicating that most of the individuals in the population were beginning to show signs of
leaf tip curling; however, some individuals with dead leaves continued to produce new
leaves. In contrast, the rank score of the salt treated population averaged 3.19 with a
range of 1.68 to 4.55 (Supplementary Table S4). This suggests that the production of
new leaves was halted, most leaves were dead or began dying, and all individuals were
displaying signs of stress. When STI values were mapped, a single QTL was detected on
chromosome 4 (qRS45_4.STI). qRS45_4.STI is located at 62.46 Mb - 63.95 Mb with a
peak near 63.67 Mb (Figure 2). qRS45_4.STI explained 10.77 percent of the phenotypic
variation and also had a negative additive effect of 0.19, indicating that S. bicolor alleles
are associated with overall better plant health in stress conditions. Of the genes located
within qRS45_4.STI, there are several genes that encode aquaporins, ion channels, and
chaperone proteins (Supplementary Table S6).
Height (HT)
At the final recording, height in the control population ranged from 52.92 cm to 122.37
cm with a population average of 87.81 cm (Supplementary Table S4). In response to
treatment, there was an average of 16.38% decrease in height, ranging from 46.32 cm to
112.47 cm with a population average of 75.46 cm (Table 2, Supplementary Table S4).
Eight QTL were detected for height (Figure 2). Three QTL (qHT45_7.C, qHT45_7.S,
and qHT45_7.STI) were detected on chromosome 7 in approximately the same region
(58.66 Mb - 61.46 Mb) (Table 3). Two QTL (qHT45_9.C and qHT45_9.STI) were
detected on chromosome 9 in approximately the same region (54.37 Mb - 56.91 Mb).
Since these five QTL were detected in the control population and the salt population, we
suspect that these QTL are important in plant height in the absence of stress. Two
additional salt specific QTL were detected on chromosome 1 (qHT45_1.S and
qHT45_1.STI) (Figure 2). Further, a QTL was detected on chromosome 4
(qHT45_4.STI) when STI values were mapped (Figure 2). Several candidate genes were
identified within the salt-specific QTL, including genes associated with aquaporins,
potassium transporters, and stress response proteins (Supplementary Table S6).
Enrichment Analysis
A total of 4276 unique genes were located within 1.0 logarithm of the odds (LOD)
confidence interval of the QTL identified in this study (Table 3, Supplementary Table
S5). Of these, we observed numerous genes (13) that encode aquaporins. In order to
determine if this constitutes an enrichment of aquaporin genes, we performed a random
sampling of 50 independent 5 Mb segments (34% of the genome) from the S. bicolor
reference genome and recorded the number of aquaporins found in each segment. A total
of 9098 genes were identified, of which only 9 encode aquaporins.
Discussion
In the present study, we screened 177 F3:5 RILs derived from a cross between the inbred
S. bicolor (Tx7000; landrace durra) and its wild relative, S. propinquum, for
performance-related traits in saline conditions. Because of sorghum’s importance in
biofuel and forage production, salinity tolerance was assessed as the ability of the plant to
maintain traits related to growth and performance in response to salt treatment. This
tolerance can be achieved by various mechanisms including Na+ exclusion from the aerial
organs of the plant, overall tissue tolerance, and osmotic adjustment; however, Na+
exclusion can also result from reduced Na+ uptake, increased Na+ extrusion to the roots
and/or soil media, or increased retrieval from the shoot (Wu et al., 2019). Ultimately,
each of these tolerance mechanisms results in the maintenance of plant vigor similar to
those plants grown in optimal conditions. In this study, we identified QTL associated
with total biomass, total aboveground biomass, height, dead aboveground biomass, live
aboveground biomass, root biomass, and rank score, in a control population, a saline
treated population, and from STI values. Among the 19 QTL detected, ten were either 1)
unique to the STI values, 2) unique to the saline environment, and/or 3) explained more
than ten percent of the phenotypic variation (Table 3).
The data presented here, in combination with previous findings (Henderson et al. 2020),
collectively demonstrates that there is increased tolerance to salinity stress in S. bicolor
compared to S. propinquum. For dead aboveground biomass, live aboveground biomass,
total aboveground biomass, and total biomass, S. bicolor alleles were associated with
tolerance. For example, the negative additive effect for the QTL detected for dead
aboveground biomass indicates that S. bicolor alleles were associated with less
accumulation of dead aboveground biomass. Similarly, the QTL detected for live
aboveground biomass, total aboveground biomass, and total biomass all have positive
additive effects indicating that S. bicolor alleles promote continued growth in stressful
conditions. It is important to note that, in optimal conditions, S. propinquum produces
more aboveground biomass compared to S. bicolor (Govindarajulu et al., 2020).
Therefore, in response to salt, the ability for lines with S. bicolor alleles to perform
favorably supports the conclusion that S. bicolor possesses greater tolerance to salinity
stress. Further, these results suggest that S. bicolor is better at handling both osmotic and
ionic stress. With osmotic adjustment, increased water can be taken up by the plant to
support the production of new biomass and to limit necrosis due to cell dehydration,
while increased ionic tolerance results in decreased leaf senescence resulting in overall
greater aboveground growth.
Because salinity stress is the product of both osmotic and ionic factors, their respective
causes and consequences are often difficult to disentangle; however, these two stresses
are often temporal in their action. When salts initially begin to accumulate in the soil, the
osmotic potential of the soil water decreases, resulting in decreased water extraction by
plant roots. This osmotic stress causes a sudden, short term loss of water, cell volume,
and turgor from leaf cells. Plants that are tolerant to stress during the osmotic phase are
better able to modify long distance signaling, limit stomatal closure, osmotically adjust,
and continue cell expansion/lateral bud development, resulting in the continuation of both
above and belowground growth (Munns & Tester, 2008). One important mechanism that
plants utilize to overcome the osmotic phase of salinity stress is the production and
accumulation of compatible solutes such as amino acids (i.e. proline), amines, betaines,
organic acids, sugars, and polyols (Parihar et al., 2015), which aid in water acquisition
and maintenance of cell turgor. For the QTL detected in this study, we identified various
genes whose products are related to osmotic adjustment, including genes involved in
proline production, aquaporins, CDPKs (calcium-dependent protein kinases), sensing and
signaling, cell division, Na+/Ca2+ exchanger, leaf senescence, early response to
dehydration, heat shock proteins, vacuolar proton exchangers, potassium antiporters, and
stress response proteins.
Following osmotic stress, as soil salinity levels rise, plants begin to accumulate Na+ and
Cl- ions, which if not properly handled will become toxic within the leaves. The most
common phenotype associated with ionic stress is increased leaf necrosis. Therefore, we
used dead aboveground biomass and rank score as a proxy for ionic toxicity. For
qDAGB45_2.STI, S. propinquum alleles positively correlated with greater dead
aboveground biomass, possibly as a result of increased Na+ accumulation in the aerial
plant tissue. Similarly, qR.S45_4.STI also had a negative additive effect, indicating that
S. propinquum alleles positively influenced the rank score, suggestive of greater
susceptibility to ionic toxicity (Table 1). Candidate genes associated with ionic stress in
these QTL include: calcium-dependent protein kinases, LEA-like proteins, aquaporins,
heat shock proteins, Na+/H+ antiporters, WRKY transcription factors, K+ uptake, and
cation transporters (Supplementary Table S5).
Most interestingly, there were numerous genes within the two dead aboveground biomass
QTL that we considered informative of ionic stress, specifically genes associated with
Ca2+ sensing/signaling and Na+ transport, which are important in limiting cytoplasmic ion
toxicity. A notable candidate gene associated with ionic sensing and signaling identified
in qDAGB45_2.STI is CDPK (calcium-dependent protein kinase) (Sobic.002G114800).
CDPKs are a class of calcium sensors that, in response to most environmental stresses,
have been previously shown to mediate abiotic stress via calcium waves that signal
various physiological responses (Urao et al., 1994a,b; Knight et al., 1997; Cheng et al.,
2002; Delormel & Boudsocq, 2019). Further, two genes that encode Na+/H+ transporters
were identified as candidate genes. Na+/H+ transporters are especially important in
sodium exclusion from areas of the plant such as the cytoplasm and aerial organs. Lastly,
genes associated with potassium uptake and distribution were identified. Potassium (K+)
is an essential nutrient for plant growth and development (Maathuis, 2009; Ahmad &
Maathuis, 2014; Morton et al., 2019). Because of the similarity in size and structure of
Na+ and K+, both ions often share transport systems. K+, however, is essential for protein
synthesis (Jones et al., 1979; Blaha et al., 2000), enzymatic reactions (Bhandal & Malik,
1988), and signaling (Shabala, 2017), whereas Na+ is not. Therefore, maintaining high
K+/Na+ ratios is important for salinity tolerance (Chen et al., 2005, 2007; Cuin et al.,
2008; Shabala, 2013; Wu et al., 2018; Morton et al., 2019).
Further, in a previous study that characterized the aquaporin (AQP) gene family in S.
bicolor, SbAQP transcript abundance was affected by both salt and drought stress (Reddy
et al., 2015). Here, we identified 13 unique aquaporin genes, which encode TIP1;1,
TIP2;1, TIP3;1, PIP1;3, PIP1;4, PIP2;2, PIP2;6, PIP2;7, and PIP1;5 (Supplementary
Table S5). Aquaporins are well known for their role in the transport of water and other
neutral solutes (Sakurai et al., 2005; Alexandersson et al., 2005; Maurel et al., 2008; Liu
et al., 2015; Reddy et al., 2015; Kadam et al., 2017; Hasan et al., 2017). Emerging
evidence indicates that some aquaporins are also capable of coupling water and ion
transport, resulting in osmotic adjustment (Byrt et al., 2017). Specifically, the
Arabidopsis plasma membrane intrinsic proteins AtPIP2;1 (AT3G53420) and AtPIP2;2
(AT2G37170) have been shown to co-transport water and Na+, suggesting that they play
dual roles in nutrient transport and osmotic adjustment (Byrt et al., 2017; Kourghi et al.,
2017). In the presence of salts, PIP2 is transported from the plasma membrane, resulting
in significant reductions in root hydraulic conductance (Boursiac et al., 2005; Sutka et
al., 2011; Byrt et al., 2017). In addition, the ionic conductance of AtPIP2;1 has been
shown to be inhibited by divalent cations, specifically Ca2+, which is known to play an
essential role in intracellular signaling in plants, particularly in response to abiotic stress
(Maurel et al., 2008; Verdoucq et al., 2008; Byrt et al., 2017). Therefore, AtPIP2;1 may
constitute a mechanism of sensing and signaling during the salt stress response in plants.
In our study, we identified the tandemly arranged Sobic.002G125000,
Sobic.002G125200, Sobic.002G125300, and Sobic.002G125700, which encode four
copies of SbAQP2;6, lie within the salt-specific QTL (qDAGB_2.STI), and share ~85%
sequence similarity with AtPIP2;1 (Supplementary Table S5). Given their presence in
salt specific QTL, known response to abiotic stress, and high sequence similarity to
AtPIP2;1, we propose that these aquaporins play a critical role in maintaining water
balance, controlling ion transport, and possibly in sensing and signaling, as mechanisms
of salinity tolerance in sorghum.
Conclusions
In our previous study, which compared salt tolerance rankings among a diverse group of
Sorghum genotypes and species, we concluded that salinity tolerance was acquired early
during domestication, specifically in the durra landrace, and then lost in improved lines in
a lineage-specific manner (Henderson et al., 2020). Here, we detected numerous genes
associated with sensing, signaling, and transport of Na+ in salt specific QTL. Further, we
identified numerous genes that encode aquaporins detected within salt responsive QTL.
The results of this study provide insights into QTL important for each of the tolerance
mechanisms (ionic, tissue, and osmotic tolerance). This is the first study where
individuals of each tolerance category (1-tolerant to ionic and osmotic stress; 2-tolerant to
ionic stress but sensitive to osmotic stress; 3-sensitive to ionic stress and tolerant to
osmotic stress; or 4-sensitive to ionic and osmotic stress) are identified in a common
genetic background. Therefore, these findings and this population provide a foundation
for future studies aimed at the dissection of the genetic basis of salinity tolerance.
Data Availability Statement
Phenotype data and the binned genotype data used for QTL mapping can be found in
Supplementary Tables S7-S8. All data necessary for confirming the conclusions of the
article are present within the article, figures, and tables.
Acknowledgements
We would like to acknowledge the WVU Genomics Core Facility, Morgantown WV for
the support provided to help make this publication possible, and CTSI Grant #U54
GM104942 which in turn provides financial support to the Core Facility. The authors
wish to thank Ryan Percifield for assistance during data collection, Dr. Stephen DiFazio
and Dr. Sandra Simon for guidance in data analysis, Dr. Erin Sparks for reviewing this
manuscript, and the West Virginia University Evansdale Greenhouse for supplying space.
This work was partially funded by the Eberly College of Arts and Sciences research
award to Ashley N. Hostetler.
Author Contributions
A.N.H and J.S.H designed experiments; A.N.H and J.S.H managed the project; A.N.H.,
prepared the samples; A.N.H. and R.G. lead the data analysis; A.N.H, and J.S.H wrote
the manuscript with contributions from R.G.
References
Agre P. 2004. Aquaporin water channels (Nobel lecture). Angewandte Chemie
International Edition 43: 4278–4290.
Ahmad I, Maathuis FJM. 2014. Cellular and tissue distribution of potassium:
Physiological relevance, mechanisms and regulation. Journal of Plant Physiology 171:
708–714.
Alexandersson E, Fraysse L, Sjövall-Larsen S, Gustavsson S, Fellert M, Karlsson M,
Johanson U, Kjellbom P. 2005. Whole Gene Family Expression and Drought Stress
Regulation of Aquaporins. Plant Molecular Biology 59: 469–484.
Almodares A, Hadi MR, Ahmadpour H. 2008a. Sorghum stem yield and soluble
carbohydrates under different salinity levels. African Journal of Biotechnology 7: 4051–
4055.
Almodares A, Hadi MR, Dosti B. 2007. Effects of salt stress on germination percentage
and seedling growth in sweet sorghum cultivars. Journal of Biological Sciences 7: 1492–
1495.
Almodares A, Hadi MR, Dosti B. 2008b. The Effects of Salt Stress on Growth
Parameters and Carbohydrates Contents in Sweet Sorghum. Research Journal of
Environmental Sciences 2: 298–304.
Almodares A, Sharif ME. 2007. Effects of irrigation water qualities on biomass and
sugar contents of sugar beet and sweet sorghum cultivars. : 6.
Bhandal IS, Malik CP. 1988. Potassium Estimation, Uptake, and Its Role in the
Physiology and Metabolism of Flowering Plants. In: Bourne GH, Jeon KW, Friedlander
M, eds. International Review of Cytology. Academic Press, 205–254.
Blaha G, Stelzl U, Spahn CMT, Agrawal RK, Frank J, Nierhaus KH. 2000.
Preparation of functional ribosomal complexes and effect of buffer conditions on tRNA
positions observed by cryoelectron microscopy. In: Methods in Enzymology. Elsevier,
292–309.
Boursiac Y, Chen S, Luu D-T, Sorieul M, van den Dries N, Maurel C. 2005. Early
Effects of Salinity on Water Transport in Arabidopsis Roots. Molecular and Cellular
Features of Aquaporin Expression. Plant Physiology 139: 790–805.
Boursier P, Läuchli A. 1990. Growth Responses and Mineral Nutrient Relations of Salt-
Stressed Sorghum. Crop Science 30: 1226–1233.
Bradbury PJ, Zhang Z, Kroon DE, Casstevens TM, Ramdoss Y, Buckler ES. 2007.
TASSEL: software for association mapping of complex traits in diverse samples.
Bioinformatics 23: 2633–2635.
Brim CA. 1966. A modified pedigree method of selection in soybeans 1. Crop Science 6:
220–220.
Broman KW, Sen Saunak. 2009. A guide to QTL mapping with R/qtl. New York:
Springer.
Broman KW, Wu H, Sen S, Churchill GA. 2003. R/qtl: {QTL} mapping in
experimental crosses. Bioinformatics 19: 889–890.
Byrt CS, Zhao M, Kourghi M, Bose J, Henderson SW, Qiu J, Gilliham M, Schultz
C, Schwarz M, Ramesh SA, et al.2017. Non-selective cation channel activity of
aquaporin AtPIP2;1 regulated by Ca2+ and pH. Plant, Cell & Environment 40: 802–815.
Carillo P, Annunziata MG, Pontecorvo G, Fuggi A, Woodrow P. 2011. Salinity Stress
and Salt Tolerance. In: Abiotic Stress in Plants - Mechanisms and Adaptations.
Chen Z, Newman I, Zhou M, Mendham N, Zhang G, Shabala S. 2005. Screening
plants for salt tolerance by measuring K+ flux: a case study for barley. Plant, Cell &
Environment 28: 1230–1246.
Chen Z, Pottosin II, Cuin TA, Fuglsang AT, Tester M, Jha D, Zepeda-Jazo I, Zhou
M, Palmgren MG, Newman IA, et al.2007. Root plasma membrane transporters
controlling K+/Na+ homeostasis in salt-stressed barley. Plant Physiology 145: 1714–
1725.
Cheng S-H, Willmann MR, Chen H-C, Sheen J. 2002. Calcium Signaling through
Protein Kinases. The Arabidopsis Calcium-Dependent Protein Kinase Gene Family.
Plant Physiology 129: 469–485.
Churchill GA, Doerge RW. 1994. Empirical threshold values for quantitative trait
mapping. Genetics 138: 963–971.
Cuin TA, Betts SA, Chalmandrier R, Shabala S. 2008. A root’s ability to retain K+
correlates with salt tolerance in wheat. Journal of Experimental Botany 59: 2697–2706.
Delormel TY, Boudsocq M. 2019. Properties and functions of calcium-dependent
protein kinases and their relatives in Arabidopsis thaliana. New Phytologist 224: 585–
604.
Doggett H. 1970. Sorghum. Sorghum.
Doggett H. 1988. Sorghum. Harlow, Essex, England: Longman Scientific & Technical ;
Fan Y, Shabala S, Ma Y, Xu R, Zhou M. 2015. Using QTL mapping to investigate the
relationships between abiotic stress tolerance (drought and salinity) and agronomic and
physiological traits. BMC Genomics 16: 43.
FAO. 2017. FAO soils portal.
Food and Agriculture Organization (FAO). 2008. Land and plant nutrition
management service.
Food and Agriculture Organization (FAO). 2009. Land and plant nutrition
management service.
Fox J, Weisberg S. 2019. An {R} companion to applied regression. Thousand Oaks, CA:
Sage.
Fracasso A, Trindade L, Amaducci S. 2016. Drought tolerance strategies highlighted
by two Sorghum bicolor races in a dry-down experiment. Journal of Plant Physiology
190: 1–14.
Genc Y, Oldach K, Taylor J, Lyons GH. 2016. Uncoupling of sodium and chloride to
assist breeding for salinity tolerance in crops. New Phytologist 210: 145–156.
Gonda I, Ashrafi H, Lyon DA, Strickler SR, Hulse‐Kemp AM, Ma Q, Sun H, Stoffel
K, Powell AF, Futrell S, et al.2019. Sequencing-Based Bin Map Construction of a
Tomato Mapping Population, Facilitating High-Resolution Quantitative Trait Loci
Detection. The Plant Genome 12: 180010.
Govindarajulu R, Henderson AN, Xiao Y, Chaluvadi SR, Mauro-Herrera M,
Siddoway ML, Whipple C, Bennetzen JL, Devos KM, Doust AN, et al.2020.
Integration of high-density genetic mapping with transcriptome analysis uncovers
numerous agronomic QTL and reveals candidate genes for the control of tillering in
sorghum. bioRxiv: 2020.07.06.189548.
Guo YY, Tian SS, Liu SS, Wang WQ, Sui N. 2018. Energy dissipation and antioxidant
enzyme system protect photosystem II of sweet sorghum under drought stress.
Photosynthetica 56: 861–872.
Hachez C, Chaumont F. 2010. Aquaporins: a family of highly regulated multifunctional
channels. In: MIPs and their Role in the Exchange of Metalloids. Springer, 1–17.
Hasan SA, Rabei SH, Nada RM, Abogadallah GM. 2017. Water use efficiency in the
drought-stressed sorghum and maize in relation to expression of aquaporin genes.
Biologia Plantarum 61: 127–137.
Hasegawa PM. 2013. Sodium (Na+) homeostasis and salt tolerance of plants.
Environmental and Experimental Botany 92: 19–31.
Henderson AN, Crim PM, Cumming JR, Hawkins JS. 2020. Phenotypic and
physiological responses to salt exposure in Sorghum reveal diversity among domesticated
landraces. American Journal of Botany 107: 983–992.
Jamil, A., Riaz S., Ashraf M., Foolad MR. 2011. Gene expression profiling of plants
under salt stress. Critical Reviews in Plant Sciences 30: 435–458.
Jones WRG, Brady CJ, Speirs J. 1979. Ionic and osmotic relations in plant cells. In:
Recent advances in cereal biochemistry. London: Academic Press, 63–103.
Julkowska MM, Saade S, Agarwal G, Gao G, Pailles Y, Morton MJL, Awlia M,
Tester M. 2019. MVApp - Multivariate analysis application for streamlined data analysis
and curation. Plant Physiology 180: 1261–1276.
Kadam S, Abril A, Dhanapal AP, Koester RP, Vermerris W, Jose S, Fritschi FB.
2017. Characterization and Regulation of Aquaporin Genes of Sorghum [Sorghum
bicolor (L.) Moench] in Response to Waterlogging Stress. Frontiers in Plant Science 8.
Kearsey MJ. 1998. The principles of QTL analysis (a minimal mathematics approach).
49: 1619–1623.
Knight H, Trewavas AJ, Knight MR. 1997. Calcium signalling in Arabidopsis thaliana
responding to drought and salinity. The Plant Journal 12: 1067–1078.
Kourghi M, Nourmohammadi S, Pei JV, Qiu J, McGaughey S, Tyerman SD, Byrt
CS, Yool AJ. 2017. Divalent Cations Regulate the Ion Conductance Properties of
Diverse Classes of Aquaporins. International Journal of Molecular Sciences 18: 2323.
Liu P, Yin L, Wang S, Zhang M, Deng X, Zhang S, Tanaka K. 2015. Enhanced root
hydraulic conductance by aquaporin regulation accounts for silicon alleviated salt-
induced osmotic stress in Sorghum bicolor L. Environmental and Experimental Botany
111: 42–51.
Lynch M, Walsh B. 1999. Genetics and Analysis of Quantitative Traits. The Quarterly
Review of Biology 74: 225–225.
Maathuis FJ. 2009. Physiological functions of mineral macronutrients. Current Opinion
in Plant Biology 12: 250–258.
Maurel C, Verdoucq L, Luu D-T, Santoni V. 2008. Plant aquaporins: membrane
channels with multiple integrated functions. Annual Review of Plant Biology 59: 595–
624.
McCormick RF, Truong SK, Sreedasyam A, Jenkins J, Shu S, Sims D, Kennedy M,
Amirebrahimi M, Weers BD, McKinley B, et al.2018. The Sorghum bicolor reference
genome: improved assembly, gene annotations, a transcriptome atlas, and signatures of
genome organization. The Plant Journal 93: 338–354.
Morton MJL, Awlia M, Al‐Tamimi N, Saade S, Pailles Y, Negrão S, Tester M. 2019.
Salt stress under the scalpel – dissecting the genetics of salt tolerance. The Plant Journal
97: 148–163.
Mullet J, Morishige D, McCormick R, Truong S, Hilley J, McKinley B, Anderson R,
Olson SN, Rooney W. 2014. Energy Sorghum--a genetic model for the design of C4
grass bioenergy crops. Journal of Experimental Botany 65: 3479–3489.
Munns R, Day DA, Fricke W, Watt M, Arsova B, Barkla BJ, Bose J, Byrt CS, Chen
Z-H, Foster KJ, et al.2020. Energy costs of salt tolerance in crop plants. New
Phytologist: 1072–1090.
Munns R, Passioura JB, Colmer TD, Byrt CS. 2019. Osmotic adjustment and energy
limitations to plant growth in saline soil. New Phytologist.
Munns R. 2005. Genes and salt tolerance: bringing them together. New Phytologist 167:
645–663.
Munns R, Tester M. 2008. Mechanisms of salinity tolerance. Annual Review of Plant
Biology 59: 651–681.
Negrão S, Schmöckel SM, Tester M. 2017. Evaluating physiological responses of
plants to salinity stress. Annals of Botany 119: 1–11.
Oksanen J, Blanchet FG, Friendly M, Kindt R, Legendre P, McGlinn D, Minchin
PR, O’Hara RB, Simpson GL, Soylmos P, et al.2019. vegan: Community Ecology
Package.
Parihar P, Singh S, Singh R, Singh VP, Prasad SM. 2015. Effect of salinity stress on
plants and its tolerance strategies: a review. Environmental Science and Pollution
Research 22: 4056–4075.
Paterson AH. 2008. Genomics of Sorghum. International Journal of Plant Genomics
2008: e362451.
Peterson BG, Carl P. 2019. PerformanceAnalytics: Econometric Tools for Performance
and Risk Analysis.
R Core Team. 2013. R: A language and enviornment for satistical computing. Vienna,
Austria: R Foundation for Statistical Computing.
Reddy PS, Rao TSRB, Sharma KK, Vadez V. 2015. Genome-wide identification and
characterization of the aquaporin gene family in Sorghum bicolor (L.). Plant Gene 1: 18–
28.
Rengasamy P. 2006. World salinization with emphasis on Australia. Journal of
Experimental Botany 57: 1017–1023.
Sakurai J, Ishikawa F, Yamaguchi T, Uemura M, Maeshima M. 2005. Identification
of 33 Rice Aquaporin Genes and Analysis of Their Expression and Function. Plant and
Cell Physiology 46: 1568–1577.
Shabala S. 2013. Learning from halophytes: physiological basis and strategies to
improve abiotic stress tolerance in crops. Annals of Botany: mct205.
Shabala S. 2017. Signalling by potassium: another second messenger to add to the list?
Journal of Experimental Botany 68: 4003–4007.
Shrivastava P., Kumar R. 2015. Soil salinity: A serious environmental issue and plant
growth promoting bacteria as one of the tools for its alleviation. Saudi Journal of
Biological Sciences 22: 123–131.
Smith O., Nicholson WV., Kistler L., Mace E., Clapham A., Rose P., Stevens C., et
al. 2019. A domestication history of dynamic adaptation and genomic deterioration in
Sorghum. Nature Plants 5: 369–379.
Snape JW, Riggs TJ. 1975. Genetical consequences of single seed descent in the
breeding of self-pollinating crops. Heredity 35: 211–219.
Sutka M, Li G, Boudet J, Boursiac Y, Doumas P, Maurel C. 2011. Natural variation
of root hydraulics in Arabidopsis grown in normal and salt-stressed conditions. Plant
physiology 155: 1264–1276.
Urao T, Katagiri T, Mizoguchi T, Yamaguchi-Shinozaki K, Hayashida N, Shinozaki
K. 1994a. Two genes that encode Ca2+-dependent protein kinases are induced by
drought and high-salt stresses in Arabidopsis thaliana. Molecular and General Genetics
MGG 244: 331–340.
Urao T, Katagiri T, Mizoguchi T, Yamaguchi-Shinozaki K, Hayashida N, Shinozaki
K. 1994b. An Arabidopsis thaliana cDNA encoding Ca2+-dependent protein kinase.
Plant Physiology 105: 1461–1462.
Verdoucq L, Grondin A, Maurel C. 2008. Structure–function analysis of plant
aquaporin At PIP2; 1 gating by divalent cations and protons. Biochemical Journal 415:
409–416.
Wu H, Shabala L, Zhou M, Su N, Wu Q, Ul‐Haq T, Zhu J, Mancuso S, Azzarello E,
Shabala S. 2019. Root vacuolar Na+ sequestration but not exclusion from uptake
correlates with barley salt tolerance. The Plant Journal: 1–13.
Wu H, Zhang X, Giraldo JP, Shabala S. 2018. It is not all about sodium: revealing
tissue specificity and signalling roles of potassium in plant responses to salt stress. Plant
and Soil 431: 1–17.
Yang Z, Li J-L, Liu L-N, Xie Q, Sui N. 2020. Photosynthetic Regulation Under Salt
Stress and Salt-Tolerance Mechanism of Sweet Sorghum. Frontiers in Plant Science 10.
Figure Legends
Figure 1. Non-metric multidimensional scaling (NMDS) analysis paired with an
analysis of similarity (ANOSIM) reveals treatment clustering. A NMDS paired with
an ANOSIM reveals that individuals were more similar within a treatment than between
treatments (ANOSIM R=0.16, p<0.001).
Figure 2. Sorghum genetic map with QTL locations from 177F3:F5 RILs. QTL
detected from a high-density genetic map. Empty spaces are regions that were removed
because bins were either heterozygous or neighboring markers were identical (duplicate
markers). The genetic map position is shown on the y-axis. Horizontal lines represent
bins used as markers. Colored vertical lines show the position of each QTL for each trait
in control conditions, salt conditions, or when STI values were mapped.
Supplemental Figures
Figure S1. Pearson correlations on raw phenotypes and transformed phenotypes for
control and salt populations at 15 days and 45 days after treatment. (A) Control
population 15 days after treatment (B) Salt population 15 days after treatment (C) Control
population 45 days after treatment (D) Transformed control data 45 days after treatment
(E) Salt population 45 days after treatment
Figure S2. Sorghum genetic map after using a sliding window method to call bin
markers as AA (S. propinquum), BB (S. bicolor), or AB. (A) After the resequencing,
SNP calling, and bin calling, 4055 total bins were detected across 10 sorghum
chromosomes. (B) The 4055 total bins are illustrated on the x-axis and the 177 RILs are
illustrated on the y-axis. The red regions illustrate bins that were called S. propinquum
(AA); the green regions illustrate bins that were called S. bicolor (BB); the blue regions
illustrate regions that were called heterozygous (AB).
Tables
Table 1. Rank scoring parameters of plant vigor. Plant vigor was assessed on a scale
of 0 to 5 with 0 indicating no signs of stress and 5 indicating plant death.
Score Observation
0
No leaf signs
1
Some leaf and leaf tip curling
2
Severe leaf and leaf tip curling, few leaves elongated
3
Most leaves dead but still producing new leaves
4
Plant still alive but no new growth
5
Plant dead
Table 2. Summary of summary statistics and phenotypic averages for RILs in the control population and salt population. An
analysis of variance (ANOVA) was used to test if there was significant variation in response to salt treatment. In response to salt
exposure plants were shorter, had less live aboveground biomass, root biomass, total aboveground biomass, and total biomass.
Additionally, in response to salt exposure plants had larger rank scores and more dead aboveground biomass. Mortality was not
affected in response to salt treatment.
Phenotype
Days After
Treatment
Control
S.D.
Salt
S.D.
RDPB
Significant
(Control-Salt)/Control
HT
15
63.75
13.19
55.30
10.01
0.13 ***
RS
15
0.45
0.48
2.18
0.53
-3.81 ***
HT
45
87.81
14.36
75.46
11.50
0.14 ***
LAGB
45
4.45
1.31
3.18
1.01
0.28 ***
RB
45
2.50
1.05
1.13
0.45
0.55 ***
TAGB
45
5.00
1.49
3.98
1.19
0.20 ***
TB
45
7.49
2.21
5.11
1.51
0.32 ***
RS
45
2.01
0.53
3.19
0.44
-0.58 ***
DAGB
45
0.55
0.28
0.80
0.42
-0.45 ***
Mortality
45
1.00
0.00
1.00
0.03
0.00
Significant Codes: (*) 0.05 (**) 0.01 (***) 0.001
HT-height, RS-rank score, M-mortality, RB-root biomass (g), DAGB-dead aboveground biomass (g), LAGB-live aboveground biomass (g), TAGB-total aboveground
biomass (g), TB-total biomass (g), RDPB-relative decrease in plant biomass
Table 3. QTLs identified in the RIL population using transformed least square means in control conditions, salt conditions,
and with stress tolerance index values. The QTLs reported were identified when using Multiple QTL Mapping (MQM) in control
conditions (0 mM NaCl), salt conditions (75 mM NaCl), and with stress tolerance index (STI) values. QTLs are named using the
following system: q[Trait][DAT]_[Chr].[Treatment]
Trait
Trt
QTL Name
DAT
Chr
Position
(cM)
Bin (Max
LOD)
Lod
score
p-value
PVE
Additive
Start
Mb
Peak
Mb
End
Mb
TB
S
qTB45_4.S
45
4
63.7
62.29
3.73
3.90E-05
10.37
0.51
61.70
62.29
68.41
TB
STI
qTB45_4.STI
45
4
64.6
62.46
4.91
2.41E-06
13.40
0.09
62.06
62.46
64.38
TAGB
S
qTAGB45_4.S
45
4
83.4
67.29
3.40
8.65E-05
9.49
0.40
61.70
67.29
68.41
TAGB
STI
qTAGB45_4.STI
45
4
64
60.23
4.18
1.34E-05
11.55
0.09
61.91
60.23
67.44
DAGB
C
qDAGB45_2.C
45
2
72.7
65.42
4.91
1.63E-05
12.15
-0.08
64.40
65.42
67.54
DAGB
C
qDAGB45_9.C
45
9
69.1
56.32
3.98
1.32E-04
9.71
-0.06
55.47
56.32
57.91
DAGB
STI
qDAGB45_2.STI
45
2
73
62.28
4.21
1.24E-05
11.63
-0.20
13.85
62.28
67.54
LAGB
C
qLAGB45_4.C
45
4
71.5
64.27
3.32
1.05E-04
9.22
0.45
62.17
64.27
67.29
LAGB
STI
qLAGB45_4.STI
45
4
73
63.41
4.22
1.22E-05
11.65
0.10
62.06
63.41
67.49
RB
STI
qRB45_4.STI
45
4
64.6
62.46
4.13
1.53E-05
11.40
0.08
62.17
62.46
62.54
RS
STI
qRS45_4.STI
45
4
69.1
63.67
3.89
2.71E-05
10.77
-0.19
62.46
63.67
63.95
HT
C
qHT45_7.C
45
7
57.7
59.01
4.72
2.50E-05
11.18
-5.00
58.66
59.01
60.24
HT
C
qHT45_9.C
45
9
63.4
55.07
5.32
6.57E-06
12.71
-5.46
54.37
55.07
56.91
HT
S
qHT45_1.S
45
1
108
75.72
5.17
9.14E-06
12.57
-4.44
76.22
75.72
80.57
HT
S
qHT45_7.S
45
7
63.5
60.17
4.95
1.51E-05
11.99
-4.25
59.01
60.17
61.46
HT
STI
qHT45_1.STI
45
1
107.8
77.00
5.45
3.37E-03
10.63
-0.04
75.67
77.00
80.57
HT
STI
qHT45_4.STI
45
4
82.2
66.96
4.66
1.16E-02
9.00
0.05
66.42
66.96
67.49
HT
STI
qHT45_7.STI
45
7
63.5
60.17
7.16
1.93E-04
14.34
-0.05
59.01
60.17
60.24
HT
STI
qHT45_9.STI
45
9
63.4
55.07
4.43
1.66E-02
8.51
-0.04
54.70
55.07
56.68
DAGB-dead aboveground biomass, LAGB-live aboveground biomass, RB-root biomass, RS-rank score, HT-height, TB-total biomass, TAGB-total aboveground biomass; Trt-treatment, STI-stress tolerance
index, C-control, S-salt; P-Sorghum propinquum, B-Sorghum bicolor; DAT-days after treatment; Chr-chromosome, PVE-percent variance explained
�����
�����
����
����
����
�����
�����
����
����
NMDS Axis 1
NMDS1 Axis 2
Treatment
Control
Salt
ANOSIM R = 0.16
P < 0.001
Figure 1. Non-metric multidimensional scaling (NMDS) analysis paired with an analysis of similarity (ANOSIM) reveals treatment clustering.
120
100
80
60
40
20
0
1
2
3
4
5
6
7
8
9
10
●
●
HT (DAT = 45) Salt
DAGB Control
DAGB STI
LAGB Control
TAGB Salt
TB Salt
RS STI
RB STI
LAGB STI
TAGB STI
TB STI
HT (DAT = 45) Control
HT (DAT = 45) STI
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
140
Mapping position (cM)
Chromosome
Figure 2. Sorghum genetic map with QTL locations from 177F3:F5 RILs.
PIP2;6
PIP2;2
PIP1;5
PIP2;7
PIP1;3
TIP2;1
TIP1;1
TIP3;1
| 2020 | QTL mapping in an interspecific sorghum population uncovers candidate regulators of salinity tolerance | 10.1101/2020.08.05.238972 | [
"Hostetler Ashley N.",
"Govindarajulu Rajanikanth",
"Hawkins Jennifer S."
] | creative-commons |
1
Ecological specialization, rather than the island rule, explains morphological
diversification in an ancient radiation of geckos
Héctor Tejero-Cicuéndez1*#, Marc Simó-Riudalbas1*, Iris Menéndez2,3, Salvador Carranza1
1 Institute of Evolutionary Biology (CSIC-Universitat Pompeu Fabra), Passeig Marítim de la
Barceloneta 37-49, 08003 Barcelona, Spain.
2 Departamento de Geodinámica, Estratigrafía y Paleontología, Facultad de Ciencias
Geológicas, Universidad Complutense de Madrid, C/ José Antonio Novais 12, Madrid, 28040,
Spain.
3 Departamento de Cambio Medioambiental, Instituto de Geociencias (UCM, CSIC), C/Severo
Ochoa 7, Madrid, 28040, Spain.
* These authors contributed equally.
# Correspondence to be sent to: Institute of Evolutionary Biology (CSIC-Universitat Pompeu
Fabra), Passeig Marítim de la Barceloneta 37-49, 08003 Barcelona, Spain.
E-mail: hector.tejero@ibe.upf-csic.es
ABSTRACT
Island colonists are often assumed to experience higher levels of phenotypic diversification
than their continental sister taxa. However, empirical evidence shows that exceptions to the
familiar “island rule” do exist. In this study, we tested this rule using a nearly complete sampled
mainland-island system, the genus Pristurus, a group of sphaerodactylid geckos mainly
distributed across continental Arabia and Africa and the Socotra Archipelago. We used a
recently published phylogeny and an extensive dataset of morphological measures to explore
whether island and mainland taxa share the same morphospace or if they present different
dynamics of phenotypic evolution. Moreover, we used habitat data to examine if ecological
specialization is correlated with morphological change, reconstructing the ancestral habitat
states across the phylogeny to compare the level of phenotypic disparity and trait evolution
between habitats. We found that insular species do not present higher levels or rates of
morphological diversification than continental groups. Instead, habitat specialization provides
insight into the evolution of body size and shape in Pristurus. In particular, the adaptation to
exploit ground habitats seems to have been the main driver of morphological change, producing
the highest levels of disparity and evolutionary rates. Additionally, arboreal species show very
constrained body size and head proportions, suggesting morphological convergence driven by
habitat specialization. Our results reveal a determinant role of ecological mechanisms in
morphological evolution and corroborate the complexity of ecomorphological dynamics in
mainland-island systems.
Keywords
Body size; disparity; evolutionary rate; island colonization; morphospace; Pristurus geckos.
2
INTRODUCTION
The life history and population biology of mainland and insular taxa of a specific evolutionary
radiation are fundamentally distinct (Foster 1964; Baeckens and Van Damme 2020). In the
mainland, communities are often assumed to be complex and composed of many species that
share a long history of coevolution (Losos 2009). In such a scenario, most of the ecological
niches will be filled, and high levels of interspecific competition are expected (Losos and
Ricklefs 2009). These factors, together with higher predation pressures, will tend to limit niche
expansion and, consequently, morphological diversification (Yoder et al. 2010). In contrast,
insular groups are usually exposed to higher levels of ecological opportunity and thus, they can
occupy the new or relatively unexploited adaptive landscapes that islands provide (Schluter
2000; Harmon et al. 2008). As a result, island species may experience increased rates of
phenotypic diversification and higher levels of morphological disparity compared to mainland
taxa (Whittaker et al. 2008). However, empirical evidence outlines a more complex scenario
in which island colonists might not necessarily experience great levels of evolutionary
divergence (Rundell and Price 2009), depending on multiple extrinsic factors (mostly
modulated by the geography or geology of the island), as well as intrinsic factors (i.e., the
biological characteristics of the group concerned) (Losos 2009, 2010). Thus, ecological
specialization is expected to be more pronounced when island colonization results in an
expansion into novel ecological contexts, and such specialization might carry morphological
changes depending on species and system-specific factors (Schluter 2000; Losos and Ricklefs
2009).
Whether ecological specialization follows island colonization or not, the study of
habitat occupancy is essential to understand the evolution of associated traits and the
structuring of ecological communities (Mahler et al. 2010). In particular, specialization in
substrate use can promote morphological diversity through deterministic body size evolution
and diversification (Reynolds et al. 2016). Moreover, microhabitat use can be strongly
correlated with convergent phenotypic evolution resulting in recurrent ecomorphs beyond the
effect of history and clade membership (Moen et al. 2016).
Arid regions, generally considered relatively depauperate in terms of animal diversity,
have been successfully inhabited by some vertebrate groups and harbor especially high levels
of reptile diversity. Among reptiles, geckos are particularly prominent due to their outstanding
diversity in ecological features, exhibiting a wide variety of morphological and behavioral
adaptations (Gamble et al. 2015). Afro-Arabian geckos have been recently prominent in studies
of the role of arid biomes in generating biodiversity (Metallinou et al. 2012, 2015; Šmíd et al.
2015, 2017; Garcia-Porta et al. 2017; Machado et al. 2021). However, the outcomes of
morphological diversification in these animals have only been properly investigated within the
genus Hemidactylus, which is the best-studied Arabian reptile group with well-resolved
taxonomy and reliably reconstructed biogeographic history (Carranza and Arnold 2012;
Gómez-Díaz et al. 2012; Šmíd et al. 2013a, 2013b, 2015, 2017, 2020; Vasconcelos and
Carranza 2014). In particular, two recent studies including continental and insular taxa from
the Socotra Archipelago proved that the genus Hemidactylus conforms to the “island rule” at
least regarding body size evolution (Garcia-Porta et al. 2016a, 2016b). In contrast, the geckos
of the genus Pristurus, which have also colonized and diversified within the same archipelago,
seem to show lower rates of body size diversification than other insular genera, but also
compared to their continental relatives (Garcia-Porta et al. 2016a). Despite these preliminary
results, a more nuanced analysis of the morphological evolution in Pristurus, including
undescribed diversity, morphological and ecological data, is still lacking. Interestingly, besides
having colonized the Socotra Archipelago, Pristurus geckos occupy a variety of habitats,
including rocky and sandy surfaces, gravel plains, and trees (Arnold 2009; Badiane et al. 2014).
3
Here we use a recently inferred phylogenetic assessment of Afro-Arabian reptiles
including undescribed diversity, together with an extensive morphological sampling and
detailed ecological information, to explore the morphological evolution in Pristurus geckos.
Specifically, we test alternative scenarios of body size and shape evolution in this genus, to
determine the role of island colonization and ecological specialization in generating the
morphological diversity observed. The independent diversification of both insular and
continental taxa, the ecological and behavioral diversity, and the unique phenotypic dataset
compiled in this study, make this group of geckos an exceptional system to investigate keystone
dynamics in evolutionary biology such as the “island rule” and ecological adaptation, and their
impact on morphological evolution.
MATERIALS AND METHODS
Phylogenetic and ecological data
We used a recently published phylogenetic tree of Afro-Arabian squamates (Tejero-Cicuéndez
et al. 2021). This tree contains all the species of Pristurus for which there exists genetic data,
including some species currently in the process of being described, resulting in a total of 30
species. We extracted the Pristurus clade from the squamate tree, both for the consensus and
for 1,000 trees randomly selected from the posterior distribution generated in the cited study.
Using a sample of posterior trees allowed us to take into account the phylogenetic and temporal
uncertainty in the subsequent analyses.
Each species was defined as insular (present in the Socotra archipelago) or continental
(present in mainland Africa or Arabia). For habitat specialization, each species was categorized
based on substrate preferences into one of three groups: ground-dwelling, rock climber, or
arboreal (Arnold 1993, 2009). Additionally, the ground-dwelling species were divided into
“soft-ground” (sandy surfaces) and “hard-ground” (gravel plains) categories to further
characterize the morphological adaptations to each type of ground habitat. Nevertheless, the
disparity dynamics and rates of trait evolution were estimated with the original categorizations
(mainland - island and the three habitat states) which, due to the limited number of species,
were more appropriate for the analyses.
Ancestral reconstructions
We studied island colonization and habitat specialization through time by reconstructing
ancestral states across the phylogeny. First, we fit several models of character evolution across
the phylogeny in order to select the best-fit model for insularity and for habitat evolution. With
such a purpose, we used the function fitDiscrete from the R package geiger v2.0 (Pennell et al.
2014; R Core Team 2019). We fit three models: an equal-rates model (ER), a symmetrical
model (SYM), and an all-rates-different model (ARD). We selected the best-fit model in each
case based on the Akaike information criterion, correcting for small sample size (AICc; Akaike
1973). We then used the function make.simmap from the R package phytools (Revell 2012),
which simulates plausible stochastic character histories after fitting a continuous-time
reversible Markov model for the evolution of the character states assigned to the tips of the
tree. We run 1,000 simulations with the previously selected model of character evolution (ER
model for both traits). Additionally, we randomly selected 100 trees from the posterior
distribution, and we ran 100 stochastic character histories on each of them for both traits
(insularity and habitat).
4
Phenotypic data
For the 30 species included in the phylogenetic tree, a total of 697 specimens were examined
and measured, with a minimum of one, a maximum of 56, and a mean of 23 specimens per
species (Table S1). All vouchers were obtained from the following collections: Institute of
Evolutionary Biology (CSIC-UPF), Barcelona, Spain (IBE), Natural History Museum,
London, UK (BM), Museo Civico di Storia Naturale, Carmagnola, Turin, Italy (MCCI),
Università di Firenze, Museo Zoologico "La Specola", Firenze, Italy (MZUF), Oman Natural
History Museum (ONHM), Laboratoire de Biogéographie et Écologie des Vertébrés de l'École
Pratique des Hautes Etudes, Montpellier, France (BEV), and National Museum Prague, Czech
Republic (NMP). The following measurements were taken by the same person (MSR) using a
digital caliper with accuracy to the nearest 0.1 mm: snout-vent length (SVL; distance from the
tip of the snout to the cloaca), trunk length (TrL; distance between the fore and hind limb
insertion points), head length (HL; taken axially from tip of the snout to the anterior ear border),
head width (HW; taken at anterior ear border), head height (HH; taken laterally at anterior ear
border), humerus length (Lhu; from elbow to the insertion of the forelimb), ulna length (Lun;
from wrist to elbow), femur length (Lfe; from knee to the insertion of the hindlimb) and tibia
length (Ltb; from ankle to knee). Tail length was not measured because most of the specimens
had regenerated tails or had lost it.
Morphological differentiation
As body size and shape evolution might be affected by island colonization and/or ecological
specialization, we characterized the morphospace occupied by each species to compare the
effect of each trait on the morphological breadth and differentiation. This allowed us to
investigate whether island colonists are morphologically distinct from their mainland relatives
(the island rule) and likewise whether the differential habitat use is reflected in species
morphology. For each species, the mean of each morphological variable was calculated and
log10-transformed in order to improve normality and homoscedasticity prior to subsequent
analyses. We then performed a phylogenetic regression of each trait on snout-vent length (SVL)
to remove the effect of the body size on the other variables. The residuals of these regressions
were used to implement a phylogenetically controlled principal component analysis (pPCA)
using the functions phyl.resid and phyl.pca from the R package phytools with the method set
to ‘lambda’ (Revell 2012). We used the principal components representing 75% of the
cumulative proportion of variance as shape variables. Additionally, we performed a principal
component analysis (PCA) with the shape data from all the specimens measured, after
correcting for body size through regressions on SVL similarly to our processing of the species
data. We generated per-species boxplots of size and shape variation with the specimen data.
We used the function phenogram from the R package phytools (Revell 2012) to map and
visualize size and shape variation across the species trees, and we further visualized the 2D
shape morphospace. For all these visualizations, we categorized the species according to
insularity (mainland or island) and to habitat use (ground, rock or tree) separately, to have a
detailed perspective of the extent of the morphospace occupied by each category.
Exploring differences in phenotypic disparities
Since one of the main possible outcomes of island colonization and/or ecological
specialization is the increase in phenotypic disparity, we tested this assumption following
previous research on other geckos (Garcia-Porta et al. 2016b) and defining disparity as the
average squared Euclidean distance between all pairs of species in a group for a given
continuous variable (Harmon et al. 2008), in our case body size (SVL) and two variables of
shape (pPC1 and pPC2). We did this for both discrete traits (insularity and habitat use), with
the aim of testing whether disparity is higher in island species than in the mainland, and, in the
5
light of our results of morphological differentiation, whether ground-dwelling species are
significantly more disparate than species in the other habitats. We first calculated the observed
disparity ratios (island/continent and ground/no-ground) for each morphological variable,
using the function disparity from the R package geiger (Pennell et al. 2014). In the case of
higher disparity in the island or in ground species (disparity ratio island/mainland or ground/no-
ground higher than 1), we then compared the observed ratios with a null distribution of disparity
ratios obtained from 10,000 simulations of the evolution of a continuous character according
to a Brownian motion model across the phylogeny. These simulations were performed by
applying the function sim.char after estimating the empirical rate parameter for body size and
shape from the best-fit model (Brownian motion) with the function fitContinuous, both from
the package geiger (Pennell et al. 2014). This approach allowed us to test, in the case of an
observed higher disparity in island or ground species, whether this is a significant increase
considering the rate of evolution of the character, or rather this is not evidence of effectively
increased disparity.
Differences in tempos of phenotypic diversification
In order to test the effect of insularity and ecological specialization in the tempo of
phenotypic evolution, we fitted different models of character evolution across the phylogeny
in which the evolutionary rates of body size and shape might or might not differ between
categories (i.e., between island and mainland, and between ground, rock and tree habitats). For
body size, we used the R package OUwie (Beaulieu and O’Meara 2021) to fit three alternative
models: BM1 (Brownian motion single rate, i.e., assuming one single rate regime for all
lineages in the phylogeny), OU1 (Ornstein-Uhlenbeck single-rate value with a phenotypic
optimum and a selective pressure towards it), and BMS (Brownian motion multi-rate model,
with different rate values for each of the regimes specified, i.e., island different from mainland
lineages, and differences between habitats). Similarly, we studied the rates of phenotypic
evolution for body shape, but in this case we fitted multivariate models including the first two
principal components resulting from the phylogenetic PCA (pPC1 and pPC2; 77% of the
variance, see Results section). We used the package mvMorph (Clavel et al. 2015) to fit four
alternative models: BM1, OU1, and BMM (analogous to BMS in the OUwie package), and
OUM (multi-rate Ornstein-Uhlenbeck model). We fitted these models in the 1,000 stochastic
character maps generated for the consensus tree, and also in the 100 character maps on each of
the 100 posterior trees used to reconstruct ancestral states (see above Ancestral
reconstructions). We then selected the best-fit models based on the AICc distributions and
means, and we extracted the distributions of rate values estimated for each regime (island,
mainland, ground, rock and tree) in the multi-rate models, to unravel the effect of each trait in
morphological rates. All the visualizations of the disparity and phenotypic rate analyses were
built with the R packages ggplot2 (Wickham 2011), patchwork (Pedersen 2020), and cowplot
(Wilke 2020).
RESULTS
Ancestral reconstructions and morphological variation
The ancestral reconstructions following an equal-rates model (ER), with the
probabilities of each state in ancestral nodes (island and mainland; ground, rock, and tree) can
be found in the Supplementary Material (Fig. S1). The ecological reconstruction shows rocky
habitats as the ancestral state in Pristurus, with several transitions to arboreal habitats and one
colonization of the ground in the ancestor of the clade known as “Spatalura group” (Arnold
1993, 2009; Fig. S1B). One of the subclades of this group later colonized more compact, harder
6
substrates, shown in our more detailed analysis separating soft- and hard-ground species (Fig.
S1).
The pPCA analysis of body shape resulted in two first components explaining 77% of
the total variance: pPC1 (61% of the variance) mainly representing limb dimensions (variables
Lhu, Lun, Lfe, Ltb), and pPC2 (16% of the variance) mostly representing head proportions
(variables HL, HW, HH) (Fig. 1a; Fig. S2 and Table S2; see Fig. S3 for body size and shape
differentiation using the specimen data and PCA). The morphospace occupied by mainland
species is notably larger than that occupied by island species, and they overlap almost
completely (Fig. 1a left, Fig. S2). When visualizing the phylomorphospace along with habitat
categories, we observe a wide portion occupied by the ground-dwelling species, especially in
pPC2 (head dimensions) (Fig. 1a right). These eight species of the “Spatalura group”
essentially occupy almost as much of the morphospace as all the rest of the species together,
with morphologies specialized to arboreal habits localized in a narrow area, especially for head
proportions (Fig. 1a right). We found a similar pattern for body size. Body size variability of
island species is completely contained in the range occupied by mainland species (Fig. 1b left).
On the contrary, ground-dwelling species show a size variability higher than all the species
from other habitats together, being the largest and the smallest species of Pristurus specialized
to ground habitats (Fig. 1b right). As with head proportions, arboreal species have apparently
constrained body sizes, being restricted to specific intermediate values within the genus’ size
range.
When separating ground species into hard- and soft-ground habitats, we observed a
clear morphological differentiation, especially in body size. Hard-ground species seem to be
highly specialized, with some of the largest body sizes of the genus, long limbs and large heads
(Fig. S4).
7
Figure 1. Morphological variability in Pristurus, with insight from insularity (left) and habitat
use (right). a) Morphospace with phylogenetic relationship between the species, showing body
shape differentiation. b) Traitgram showing body size (SVL) through time on the summary
phylogenetic tree of Pristurus, mapped by the discrete categories of land occupancy (left) and
by ecological specialization (right). Photos (proportional to species’ SVL): Pristurus carteri
(top) and P. masirahensis (bottom), the largest and smallest species of the genus, respectively.
Phenotypic disparity
We found that the morphological disparity in the island is lower than in the mainland for the
three variables, with disparity ratios island/mainland below 1 (SVL: 0.88; pPC1: 0.52; pPC2:
0.53). When comparing disparity between ground and no-ground habitats, we found a higher
observed disparity in ground for body size (SVL) and head proportions (pPC2), with disparity
ratios ground/no-ground of 2.25 for SVL, 0.86 for pPC1, and 2.47 for pPC2. Furthermore, both
for size and head proportions, the increased disparity in ground habitats was significant
compared with the null distribution of simulated disparity ratios (psize = 0.03; phead = 0.01; Fig.
2).
8
Figure 2. Observed (red arrows) and simulated (gray bars) ratios of phenotypic disparity
between ground versus no-ground habitats. a) Body size disparity ratios. b) Head proportions
(pPC2) disparity ratios.
Rates of morphological evolution
For body size, a multi-rate Brownian motion model (BMS) was the best fit both for insularity
and for ecological specialization (lowest AICc; Fig. 3a), suggesting differences in the rate of
morphological evolution across discrete categories (i.e., island vs. mainland, and different
habitats). For body shape, however, we did not find evidence for differences in evolutionary
rates, being the single-rate Brownian motion model (BM1) the best fit both for limb (pPC1)
and head (pPC2) dimensions, although the overlap across all models was considerable (Fig.
3b).
Figure 3. AICc distributions from the model fitting for a) body size and b) shape evolution of
the genus Pristurus. These results correspond to model fitting on 100 stochastic character maps
(insularity in the top panels and habitat in the bottom) on 100 trees from the phylogenetic
posterior distribution. BM1: Brownian motion single rate. OU1: Ornstein-Uhlenbeck single
rate. BMS (OUwie) / BMM (mvMorph): Brownian motion multi-rate. OUM: Ornstein-
Uhlenbeck multi-rate. For body size, the best supported model is a Brownian motion with rate
9
heterogeneity across categories. For body shape, a single-rate Brownian motion model was the
best-fit, although there is an extensive overlap across all models.
We extracted the rates of body size evolution from the Brownian motion multi-rate
models, and we found that island species present lower rates than mainland species (Fig. 4a
top). For ecological specialization, we found increased rates of body size evolution in the
ground-dwelling species relative to the other habitats, with arboreal habitats showing the lowest
rates (Fig. 4a bottom). We also extracted per-category rates of body shape evolution according
to the BMM model in mvMorph, even though the multi-rate models were not the best
supported, and we found a similar scenario, in which shape evolution (both for limbs and for
head proportions) would be notably faster in ground-dwelling species (Fig. 4b bottom). Results
from the analyses performed with the 1,000 stochastic character maps on the consensus tree
and with the 100 maps on each of 100 posterior trees lead to the same conclusions, so we show
the ones from the posterior trees on the main text. Results from analyses with the consensus
tree can be found in the Supplementary Material (Fig. S5).
10
Figure 4. Rates of a) body size and b) body shape evolution in the genus Pristurus, extracted
from multi-rate Brownian motion models fitted on a total of 10,000 character maps (100
stochastic character maps on 100 posterior trees) of land occupancy (island vs. mainland) and
habitat use (ground, rock, and tree).
DISCUSSION
The present study represents the first comprehensive comparative work on the genus Pristurus,
including undescribed diversity and extensive morphological (size and shape) and ecological
data. We tested the relative roles of island colonization and ecological specialization in the
evolution of the phenotypic diversity observed within the genus. We did not find evidence for
the validity of the ‘island rule’ in this radiation of geckos, since island species do not present a
notably different morphology, higher disparity, nor increased rates of morphological evolution,
relative to species in the continent. On the contrary, ecological specialization emerges as a
determinant factor in generating morphological diversity, with the colonization of ground
habitats as the main driver of phenotypic divergence. Our results reveal a complex scenario in
which different morphological traits interact with ecological characteristics of the species in
different ways, suggesting a differential relevance of body size and shape proportions for the
adaptation to specific habitats.
The tendency of island taxa to diverge in morphology compared to their continental
relatives is a general pattern in terrestrial vertebrates, especially concerning body size
(Lomolino 2005; Benítez-López et al. 2021). In fact, recent studies on Afro-Arabian geckos
colonizing the Socotra Archipelago found support for this ‘island effect’, particularly in the
genus Hemidactylus (Garcia-Porta et al. 2016a, 2016b). Nevertheless, preliminary results on
Pristurus geckos failed to find this phenomenon in this genus (Garcia-Porta et al. 2016a). Here
we corroborated and extended those preliminary results, incorporating the most complete
phylogenetic and morphological sampling within Pristurus up to date. We did not find the
predicted effects of island colonization in phenotypic diversification. Even though one of the
insular clades (the one including P. insignis, P. insignoides and P. sp. 12) has effectively
undergone an increase in body size, one of the mainland species (P. carteri) is the largest of
the genus (Fig. 1b). There is also no apparent divergence in body shape, with limb and head
proportions of island species being similar to mainland species (Fig. 1a). Moreover, neither
size nor shape disparities observed in the Socotra Archipelago are higher than those of the
mainland species. Finally, our results failed to find another expected outcome of island
colonization, as is the increase in rates of phenotypic evolution. Species in the mainland show
higher rates of body size evolution (Fig. 4a), and our evolutionary model fitting showed no
support for differences in shape rates between mainland and island species (BM1 was the best-
fit model; Fig. 3a). When extracting the rate values for shape from the multi-rate Brownian
motion model, we found little or no difference between mainland and island taxa (Fig. 4b). The
lack of an island effect in Pristurus, opposed to other similar diversifications of geckos, may
indicate the existence of different ecological contexts even in the same physical settings, which
would imply different ecological opportunities in the same island (Losos 2010). Namely, some
life-history traits, such as being diurnal, may have limited niche expansion in Pristurus species
in the Socotra Archipelago as a result of ecological interactions with other lizards, while
nocturnality may have prevented other geckos such as Hemidactylus or Haemodracon from
suffering that kind of ecological pressure, resulting in a phenotypic divergence of island
colonists (Garcia-Porta et al. 2016b; Tamar et al. 2019b). This is consistent with results on
global insular vertebrate communities suggesting that the prevalence of the island rule is
subjected to system-specific ecological and environmental dynamics (Benítez-López et al.
2021). Furthermore, a recent study on the anole radiation in the Greater Antilles did not find
11
evidence for an island effect, and instead point to ecological opportunity and key innovations
as the drivers of the adaptive radiation (Burress and Muñoz 2021).
Following that reasoning, ecological specialization gives us a much more nuanced
insight on Pristurus phenotypic evolution. The relationship between habitat use and
morphological traits is well recognized in lizards (Goodman et al. 2008; Losos 2009; Higham
and Russell 2010). In fact, preceding observations on Pristurus geckos suggested that many
morphological changes might be functionally associated with shifts in ecology and behavior
(Arnold 2009). Our results are consistent with this notion and provide strong evidence that
novel ecological opportunities produced high levels of phenotypic disparity associated with
increased rates of trait evolution in some forms of Pristurus, particularly the species exploiting
ground habitats. Even though ground-dwelling species do not show an extremely divergent
body shape relative to species inhabiting other habitats (i.e., rocky and arboreal habitats; Fig.
1a), they do comprise the largest and the smallest sizes of the genus (Fig. 1b) and show some
extreme values of limb and head proportions (pPC1 and pPC2 respectively; Fig. 1a), as well as
they occupy a very large portion of the genus’ entire morphospace (Fig. 1a). This extreme
variability within ground-dwelling species is reflected in our disparity results. While ground
species do not present higher disparity in limb dimensions (pPC1; ratio ground/no-ground <
1), they have more than twice as much disparity as all the rest of the species in body size and
head proportions, with these ratios being highly significant compared to the null model
generated from simulations of character evolution (Fig. 2). Furthermore, we found increased
rates of body size evolution in ground-dwelling species, followed first by rocky and last by
arboreal habitats (Fig. 4a). Although rate heterogeneity across habitat categories was not the
best-fit model for body shape (Fig. 3b), the rate values extracted from the multi-rate Brownian
motion model show a similar pattern, especially for head proportions, with highest rates in
ground-dwelling species (Fig. 4b). Taken together, these results point to the existence of a
morphological response to the ecological context, especially in body size. This is consistent
with the idea that the main driver of morphological divergence, even in an island colonization
event, is habitat diversity (Lack 1976; Losos and Parent 2009). If habitat heterogeneity in the
Socotra Archipelago is lower than in mainland Africa and Arabia (e.g., no particularly large
gravel plains in Socotra), or if the access to those habitats is limited for Pristurus geckos in the
island and not in the mainland (e.g., due to ecological interactions), phenotypic evolution after
island colonization would not be as extreme as expected under the ‘island rule’ framework.
This could imply a tight relationship between morphology and structural habitat, a pattern
observed in other Arabian geckos such as Ptyodactylus or Asaccus, where niche conservatism
is associated with a very conserved morphology (Metallinou et al. 2015; Carranza et al. 2016;
Simó-Riudalbas et al. 2017, 2018; Tamar et al. 2019a). This would be further supported by the
fact that within ground habitats, species show a clear morphological segregation between ‘hard’
and ‘soft’ substrates, suggesting a particularly conspicuous specialization to the former (large
bodies, long limbs and large heads in the hard-ground species: P. carteri, P. collaris, P.
ornithocephalus; Fig. S4). Alternatively, the lack of an island effect might be explained by
climatic divergences replacing ecomorphological differentiation, or by a low morphological
evolvability (Garcia-Porta et al. 2016a).
Another interesting result is the morphological convergence of arboreal species,
especially in body size and head proportions, where they present intermediate values within a
very restricted range (Fig. 1). Consistently, we found notably reduced evolutionary rates in
body size and head proportions in these species with the multi-rate models (Fig. 4). This might
corroborate the idea of adaptive processes leading to a tight relationship between ecological
traits and phenotype, since this scenario is expected if a specific habitat constrains the
morphology towards optimum values of body size and shape (Moen et al. 2016).
12
Ultimately, our results provide evidence of the determinant role of habitat specialization
in phenotypic evolution. This has important implications for understanding the prevalence of
the island rule in the context of differential ecological opportunity and, combined with previous
results on other similar systems, shows the complex nature of the relationships between
ecological mechanisms and morphology and their reliance on system-specific dynamics. More
detailed ecological and morphological data (e.g., dietary habits, geometric morphometrics of
head shape) might help for a deeper understanding of the evolutionary dynamics of this and
other groups of arid-adapted lizards.
DATA AVAILABILITY STATEMENT
Data and R scripts used for this study will be available in an online public repository.
COMPETING INTERESTS
The authors declare no competing interests.
ACKNOWLEDGEMENTS
We are very grateful to J. Roca, M. Metallinou, K. Tamar, J. Šmid, R. Vasconcelos, R. Sindaco,
F. Amat, Ph. de Pous, L. Machado, J. Garcia-Porta, J. Els, T. Mazuch, T. Papenfuss, B. Burriel
and all the people from the Environment Authority, Oman, for their help in different aspects of
the work. This work was supported by grants CGL2015-70390-P (MINECO/FEDER, UE) and
PGC2018-098290-B-I00 (MCIU/AEI/FEDER, UE), Spain and grant 2017-SGR-00991 from
the Secretaria d’Universitats i Recerca del Departament d’Economia i Coneixement de la
Generalitat de Catalunya to SC. H.T.-C. was funded by an FPI grant (BES-2016-078341)
(MINECO/AEI/FSE), Spain. I.M. was funded by a predoctoral grant from the Complutense
University of Madrid (CT27/16-CT28/16), Spain. M.S.-R was funded by an FPI grant (BES-
2013-064248) (MINECO/AEI/FSE), Spain.
REFERENCES
Akaike H. 1973. Information theory and an extension of the maximum likelihood principle. In:
Petrov B., Csaki F., editors. Second International Symposium on Information Theory.
Budapest (Hungary): Akademia Kiado. p. 267-281.
Arnold E.N. 1993. Historical changes in the ecology and behaviour of semaphore geckos
(Pristurus, Gekkonidae) and their relatives. Journal of Zoology. 229:353–384.
Arnold E.N. 2009. Relationships, evolution and biogeography of Semaphore geckos, Pristurus
(Squamata, Sphaerodactylidae) based on morphology. Zootaxa. 21:1–21.
Badiane A., Garcia-Porta J., Červenka J., Kratochvíl L., Sindaco R., Robinson M.D., Morales
H., Mazuch T., Price T., Amat F., Shobrak M.Y., Wilms T., Simó-Riudalbas M.,
Ahmadzadeh F., Papenfuss T.J., Cluchier A., Viglione J., Carranza S. 2014.
Phylogenetic relationships of Semaphore geckos (Squamata: Sphaerodactylidae:
Pristurus) with an assessment of the taxonomy of Pristurus rupestris. Zootaxa.
3835:33–58.
Baeckens S., Van Damme R. 2020. The island syndrome. Current Biology. 30:R338–R339.
13
Beaulieu J.M., O’Meara B. 2021. OUwie: analysis of evolutionary rates in an OU framework.
R package, version 2.6.
Benítez-López A., Santini L., Gallego-Zamorano J., Milá B., Walkden P., Huijbregts M.A.J.,
Tobias J.A. 2021. The island rule explains consistent patterns of body size evolution in
terrestrial vertebrates. Nature Ecology and Evolution. 5:768–786.
Burress E.D., Muñoz M.M. 2021. Ecological Opportunity from Innovation, not Islands, Drove
the Anole Lizard Adaptive Radiation. Systematic Biology. Advance online publication.
https://doi.org/10.1093/sysbio/syab031.
Carranza S., Arnold E.N. 2012. A review of the geckos of the genus Hemidactylus (Squamata:
Gekkonidae) from Oman based on morphology, mitochondrial and nuclear data, with
descriptions of eight new species. Zootaxa. 95:1–95.
Carranza S., Simó-Riudalbas M., Jayasinghe S., Wilms T., Els J. 2016. Microendemicity in the
northern Hajar Mountains of Oman and the United Arab Emirates with the description
of two new species of geckos of the genus Asaccus (Squamata: Phyllodactylidae).
PeerJ. 4:e2371.
Clavel J., Escarguel G., Merceron G. 2015. mvMORPH: An R package for fitting multivariate
evolutionary models to morphometric data. Methods in Ecology and Evolution.
6:1311–1319.
Foster J.B. 1964. Evolution of mammals on islands. Nature. 202:234–235.
Gamble T., Greenbaum E., Jackman T.R., Bauer A.M. 2015. Into the light: diurnality has
evolved multiple times in geckos. Biological Journal of the Linnean Society. 115:896–
910.
Garcia-Porta J., Morales H.E., Gómez-Díaz E., Sindaco R., Carranza S. 2016a. Patterns of
diversification in islands: A comparative study across three gecko genera in the Socotra
Archipelago. Molecular Phylogenetics and Evolution. 98:288–299.
Garcia-Porta J., Simó-Riudalbas M., Robinson M., Carranza S. 2017. Diversification in arid
mountains: biogeography and cryptic diversity of Pristurus rupestris rupestris in
Arabia. Journal of Biogeography. 44:1694–1704.
Garcia-Porta J., Šmíd J., Sol D., Fasola M., Carranza S. 2016b. Testing the island effect on
phenotypic diversification: insights from the Hemidactylus geckos of the Socotra
Archipelago. Scientific Reports. 6:23729.
Gómez-Díaz E., Sindaco R., Pupin F., Fasola M., Carranza S. 2012. Origin and in situ
diversification in Hemidactylus geckos of the Socotra Archipelago. Molecular Ecology.
21:4074–4092.
Goodman B.A., Miles D.B., Schwarzkopf L. 2008. Life on the rocks: Habitat use drives
morphological and performance evolution in lizards. Ecology. 89:3462–3471.
Harmon L.J., Melville J., Larson A., Losos J.B., Oakley T. 2008. The Role of Geography and
14
Ecological Opportunity in the Diversification of Day Geckos (Phelsuma). Systematic
Biology. 57:562–573.
Higham T.E., Russell A.P. 2010. Divergence in locomotor performance, ecology, and
morphology between two sympatric sister species of desert-dwelling gecko. Biological
Journal of the Linnean Society. 101:860–869.
Lack D. 1976. Island biology illustrated by the land birds of Jamaica. Berkeley: University of
California Press.
Lomolino M.V. 2005. Body size evolution in insular vertebrates: Generality of the island rule.
Journal of Biogeography. 32:1683–1699.
Losos J. 2009. Lizards in an evolutionary tree: Ecology and adaptive radiation of anoles.
University of California Press.
Losos J.B. 2010. Adaptive radiation, ecological opportunity, and evolutionary determinism.
American Naturalist. 175:623–639.
Losos J.B., Parent C.E. 2009. The speciation-area relationship. In: Losos J.B., Ricklefs R.E.,
editors. The Theory of Island Biogeography Revisited. Princeton University Press. p.
415–438.
Losos J.B., Ricklefs R.E. 2009. Adaptation and diversification on islands. Nature. 457:830–
836.
Machado L., Salvi D., James Harris D., Brito J.C., Crochet P.A., Geniez P., Ahmadzadeh F.,
Carranza S. 2021. Systematics, biogeography and evolution of the Saharo-Arabian
naked-toed geckos genus Tropiocolotes. Molecular Phylogenetics and Evolution.
155:106969.
Mahler D.L., Revell L.J., Glor R.E., Losos J.B. 2010. Ecological opportunity and the rate of
morphological evolution in the diversification of Greater Antillean anoles. Evolution.
64:2731–2745.
Metallinou M., Arnold E.N., Crochet P.-A., Geniez P., Brito J.C., Lymberakis P., Baha El Din
S., Sindaco R., Robinson M., Carranza S. 2012. Conquering the Sahara and Arabian
deserts: systematics and biogeography of Stenodactylus geckos (Reptilia: Gekkonidae).
BMC evolutionary biology. 12:258–258.
Metallinou M., Červenka J., Crochet P.A., Kratochvíl L., Wilms T., Geniez P., Shobrak M.Y.,
Brito J.C., Carranza S. 2015. Species on the rocks: Systematics and biogeography of
the rock-dwelling Ptyodactylus geckos (Squamata: Phyllodactylidae) in North Africa
and Arabia. Molecular Phylogenetics and Evolution. 85:208–220.
Moen D.S., Morlon H., Wiens J.J. 2016. Testing Convergence Versus History: Convergence
Dominates Phenotypic Evolution for over 150 Million Years in Frogs. Systematic
Biology. 65:146–160.
Pedersen T.L. 2020. Patchwork: The composer of plots. R package version 1.1.1.
15
Pennell M.W., Eastman J.M., Slater G.J., Brown J.W., Uyeda J.C., Fitzjohn R.G., Alfaro M.E.,
Harmon L.J. 2014. Geiger v2.0: An expanded suite of methods for fitting
macroevolutionary models to phylogenetic trees. Bioinformatics. 30:2216–2218.
R Core Team. 2019. R: A Language and Environment for Statistical Computing. Vienna,
Austria: R Foundation for Statistical Computing.
Revell L.J. 2012. phytools: An R package for phylogenetic comparative biology (and other
things). Methods in Ecology and Evolution. 3:217–223.
Reynolds R.G., Collar D.C., Pasachnik S.A., Niemiller M.L., Puente-Rolón A.R., Revell L.J.
2016. Ecological specialization and morphological diversification in Greater Antillean
boas. Evolution. 70:1882–1895.
Rundell R.J., Price T.D. 2009. Adaptive radiation, nonadaptive radiation, ecological speciation
and nonecological speciation. Trends in Ecology and Evolution. 24:394–399.
Schluter D. 2000. The ecology of adaptive radiation. Oxford University Press.
Simó-Riudalbas M., Metallinou M., de Pous P., Els J., Jayasinghe S., Péntek-Zakar E., Wilms
T., Al-Saadi S., Carranza S. 2017. Cryptic diversity in Ptyodactylus (Reptilia:
Gekkonidae) from the northern Hajar Mountains of Oman and the United Arab
Emirates uncovered by an integrative taxonomic approach. PloS ONE. 12:e0180397.
Simó-Riudalbas M., Tarroso P., Papenfuss T., Al-Sariri T., Carranza S. 2018. Systematics,
biogeography and evolution of Asaccus gallagheri (Squamata, Phyllodactylidae) with
the description of a new endemic species from Oman. Systematics and Biodiversity.
16:323–339.
Šmíd J., Carranza S., Kratochvíl L., Gvoždík V., Nasher A.K., Moravec J. 2013a. Out of
Arabia: A Complex Biogeographic History of Multiple Vicariance and Dispersal
Events in the Gecko Genus Hemidactylus (Reptilia: Gekkonidae). PLoS ONE.
8:e64018.
Šmíd J., Moravec J., Kratochvíl L., Gvoždík V., Nasher A.K., Busais S.M., Wilms T., Shobrak
M.Y., Carranza S. 2013b. Two newly recognized species of Hemidactylus (Squamata,
Gekkonidae) from the Arabian Peninsula and Sinai, Egypt. ZooKeys.:79–107.
Šmíd J., Moravec J., Kratochvíl L., Nasher A.K., Mazuch T., Gvoždík V., Carranza S. 2015.
Multilocus phylogeny and taxonomic revision of the Hemidactylus robustus species
group (Reptilia, Gekkonidae) with descriptions of three new species from Yemen and
Ethiopia. Systematics and Biodiversity. 13:346–368.
Šmíd J., Shobrak M., Wilms T., Joger U., Carranza S. 2017. Endemic diversification in the
mountains: genetic, morphological, and geographical differentiation of the
Hemidactylus geckos in southwestern Arabia. Organisms Diversity & Evolution.
17:267–285.
Šmíd J., Uvizl M., Shobrak M., Salim A.F.A., AlGethami R.H.M., Algethami A.R., Alanazi
A.S.K., Asubaie S.D., Busais S., Carranza S. 2020. Swimming through the sands of the
16
Sahara and Arabian deserts: Phylogeny of sandfish skinks (Scincidae, Scincus) reveals
a recent and rapid diversification. Molecular Phylogenetics and Evolution. 155:107012.
Tamar K., Mitsi P., Carranza S. 2019a. Cryptic diversity revealed in the leaf-toed gecko
Asaccus montanus (Squamata, Phyllodactylidae) from the Hajar Mountains of Arabia.
Journal of Zoological Systematics and Evolutionary Research. 57:369–382.
Tamar K., Simó-Riudalbas M., Garcia-Porta J., Santos X., Llorente G., Vasconcelos R.,
Carranza S. 2019b. An integrative study of island diversification: Insights from the
endemic Haemodracon geckos of the Socotra Archipelago. Molecular Phylogenetics
and Evolution. 133:166–175.
Tejero-Cicuéndez H., Patton A.H., Caetano D.S., Šmíd J., Harmon L.J., Carranza S. 2021.
Reconstructing Squamate Biogeography in Afro-Arabia Reveals the Influence of a
Complex and Dynamic Geologic Past. Systematic Biology. Advance online
publication. https://doi.org/10.1093/sysbio/syab025.
Vasconcelos R., Carranza S. 2014. Systematics and biogeography of Hemidactylus
homoeolepis Blanford, 1881 (Squamata: Gekkonidae), with the description of a new
species from Arabia. Zootaxa. 3835:501–527.
Whittaker R.J., Triantis K.A., Ladle R.J. 2008. A general dynamic theory of oceanic island
biogeography. Journal of Biogeography. 35:977–994.
Wickham H. 2011. ggplot2. Wiley Interdisciplinary Reviews: Computational Statistics. 3:180–
185.
Wilke C.O. 2020. cowplot: streamlined plot theme and plot annotations for ‘ggplot2’. R
package version 1.1.
Yoder J.B., Clancey E., Des Roches S., Eastman J.M., Gentry L., Godsoe W., Hagey T.J.,
Jochimsen D., Oswald B.P., Robertson J. 2010. Ecological opportunity and the origin
of adaptive radiations. Journal of evolutionary biology. 23:1581–1596.
| 2021 | Ecological specialization, rather than the island rule, explains morphological diversification in an ancient radiation of geckos | 10.1101/2021.07.30.454424 | [
"Tejero-Cicuéndez Héctor",
"Simó-Riudalbas Marc",
"Menéndez Iris",
"Carranza Salvador"
] | creative-commons |
1
Feasibility Analyses and Experimental Confirmation
of Dove Prism Based Dual-fiberscope Rotary Joint
Yuehan Liu1, Hyeon-Cheol Park2, Haolin Zhang2, and Xingde Li2
1Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA
2Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD 21205, USA
Contact email: xingde@jhu.edu
Abstract – Two-photon fluorescence microscopy has enjoyed its wide adoption in neuroscience. Head-
mounted miniaturized fiberscopes offered an exciting opportunity for enabling neural imaging in freely-
behaving animals with high 3D resolution. Here we propose a dual-fiberscope rotary joint based on a Dove
prism, for enabling simultaneous two-photon imaging of two brain regions with two fiberscopes in freely-
walking/rotating mice. Analytic proof has confirmed the key properties of a Dove prism. Feasibility
analyses and proof-of-concept experimental results have demonstrated the feasibility of such a rotary joint
for allowing two fiberscopes to rotate simultaneously while maintaining an excellent single-mode fiber-to-
fiber coupling for the excitation femtosecond laser. Fiberscopes with a dual-probe rotary joint offer an
exciting opportunity to explore neural network dynamics of multiple interconnected brain regions in freely-
walking rotating animals.
Keywords – Dove Prism, Dual-fiberscope Rotary Joint, Two-photon neuroimaging
1. INTRODUCTION
Brain activities involve neurons generating fast-propagating signals to encode and relay information within
dynamic neural networks. Neuroscientists aspire to obtain access to such networks with a high
spatiotemporal resolution, which will shed light on the fundamental working mechanisms of the brain.
Optical imaging, particularly two-photon fluorescence (TPF) microscopy, has played a significant role in
this endeavor [1, 2], which has exhibited multiple advantages such as high imaging resolution, 3D imaging
capability [3], deeper penetration depth [4], and the ability to simultaneously excite multiple fluorophores
with a single light source [5]. With the development of genetically encoded fluorescent calcium indicators
(such as e.g., GCaMP), bench-top two-photon (2P) microscopy has become one of the key platforms for
neural network imaging [6-8]. The past decade has witnessed many impressive progresses, from head-
restrained benchtop microscopy virtual navigation to the developments of large FOV microscopy for neuron
population imaging [9-11].
Another technological trend in this field is the miniaturization of imaging devices to enable real-time
imaging of freely behaving rodents [12]. Our group has developed the first, fully integrated, fiber-optic
scanning 2P endomicroscope for fluorescence and second harmonic generation imaging [13, 14]. By
introducing a customized double-clad fiber (DCF) of a pure silica single-mode core, label-free in vivo 2P
imaging at subcellular resolution has been achieved [15]. Along with a newly developed single-probe
optoelectrical commutator (OEC), in vivo functional neural dynamics imaging in freely-behaving mice has
been demonstrated [16-18]. In spite of all these exciting advances, tools for simultaneous 2P imaging over
multiple brain regions in freely-behaving animals (e.g., rodents) are still lacking.
The capability of simultaneous imaging over multiple interconnected brain regions, along with the option
for multicolor imaging, would provide an exciting opportunity for studying synergized functional
connectivity of involved neural networks, and provide unprecedented insight into the neural circuit
2
dynamics associated with various behaviors at population level with subcellular resolution. In addition, the
freely-moving style would minimize the differences between experimentally controlled actions and natural
behaviors, therefore, allowing precise examination of neural network functions. However, there are several
bottlenecks: The large footprint (i.e., 15x9x21 mm3) of the state-of-the-art 2P miniscopes [12] makes it very
challenging (if not impossible) to mount two 2P miniscopes over a very limited cortical area on a mouse
head. In addition, the weight of two miniscopes would impose a prohibitively heavy burden (> 4.90 g) onto
the mouse. One solution is to adopt the recently developed 2P fiberscopes which are extremely light (< 1 g
for each fiberscope head) and compact (2-2.4 mm in diameter). However, an advanced fiber-optic rotary
joint is needed for dual-probe imaging of freely walking/rotating rodents. Although a rotary joint for single-
mode core-to-core optical coupling from one stationary source fiber to another one rotating probe fiber has
been well-established, only very few vendors offer dual-fiber rotary joints at communication wavelengths
(e.g., 1300-1500 nm) but with very poor coupling efficiency and rotational coupling variation (See
Supplement 1). Due to a much smaller single-mode fiber core (~5 µm) at the 800-900 nm wavelength range,
the coupling is even more challenging. To the best of our knowledge, currently such a dual-fiberscope
rotary joint does not exist.
Here, we propose a dual-fiberscope rotary joint based on a Dove prism, allowing two fiberscopes to rotate
simultaneously while maintaining an excellent single-mode fiber-to-fiber coupling for the excitation
femtosecond laser from the stationary source fibers to the two fiberscopes. In this paper, we first present
analytic proofs to confirm the key properties and working principle of a Dove prism based dual-probe rotary
joint. We then report the feasibility of the dual-probe rotary joint using Ray-tracing analyses with the
fabrication tolerance/error of key parameters of a Dove prism taken into account. Finally we demonstrate
the initial proof-of-concept experimental results, confirming the feasibility of our proposed dual-probe
rotary joint.
2. DESIGN AND METHODS
A critical component for the rotary joint to accommodate two fiberscopes is the Dove prism, which is a
truncated right-angle prism with a base angle 𝛼 (usually 45°). For a given material, the length and aperture
size of the prism are designed based on the following Dove prism formula [19]:
𝐿
𝑎 =
1
sin(2𝛼) (1 + √𝑛2 − cos2𝛼 + sin𝛼
√𝑛2 − cos2𝛼 − sin𝛼
) .
(1)
Here 𝐿 is the length of the base (longest bottom face) of the prism, 𝑎 is the side length of the prism cross
section (or aperture), is the base angle and 𝑛 is the refractive index of the prism. Since the refractive 𝑛 is
wavelength dependent, the length 𝐿 of a Dove prism needs to be specifically chosen for a given wavelength
in order to achieve target performance. It is noted that a Dove prism can be used as an image inverter and
rotator. We first define the rotation axis (RA) of a Dove prism as the axis parallel to the prism base and
going through the center of the aperture (see Figure 1). A dove prism has several unique and very attractive
optical properties. 1) For an incident beam parallel to the RA, the exit beam from the prism (after
undergoing total internal reflection at the prism base) remains parallel to the RA; 2) For an incident beam
parallel to the RA, the distance of the exit beam to the RA remains the same as the distance of the incident
beam to the RA. This is crucial for fiber coupling of light since any lateral shift of the beam away from the
RA will affect the coupling efficiency of the beam into a single-mode fiber (or the single-mode core of a
DCF) [20]; 3) When the Dove prism is rotated by an angle 𝜃, the exit beam (i.e., the image of the stationary
incident beam parallel to the RA) rotates by an angle 2𝜃. This means the rotation of a fiberscope can be
compensated by rotating the Dove prism by a half angle so that the stationary incident beam can still be
3
coupled into the rotated fiberscope after going through the half-angle rotated Dove prism; 4) The optical
pathlength of any incident beams that are parallel with each other is a constant, which means a Dove prism
does not introduce an optical pathlength difference among these beams. This is very favorable for 2P
imaging since the material dispersion of the Dove prism for the incident femtosecond (fs) pulses can be
conveniently pre-compensated, e.g., by using a grating-prism (GRISM) pair (for simultaneous
compensation of the group velocity dispersion but for the third order dispersion as well) [16]. These key
properties make a Dove prism a viable choice for a dual-fiberscope rotary joint. Analytic proofs according
to geometric optics of these properties are provided in Supplement 2. Ray-tracing simulations by ZEMAX
also confirm these properties.
Figure 1. The Dove prism rotation axis (RA) is perpendicular to the prism aperture and also goes through
the center of the aperture as shown in the side view as well as the front view.
The design schematic of a Dove prism based dual-fiberscope rotary joint is shown in Figure 2. In essence,
two pairs of fiber-optic collimators (FCs) are used to couple light from the stationary source fibers to the
rotating probe fibers (or fiberscopes). A Dove prism is sandwiched between FC1&2 and FC3&4. Note that
the exit beam from the Dove prism rotates at twice the rate of the prism rotation. Therefore, the Dove prism
and FC3&4 are mounted on two separate coaxial rotation shafts. Once the two pairs of FCs (FC1→FC4,
FC2→FC3) and the two rotation axes are precisely aligned, a fiberscope rotation angle 2𝜃 can be
compensated by 𝜃 rotation of the Dove prism, and the two incident beams can thus be efficiently coupled
into the two fiberscopes through, the single-mode cores of the two DCFs in the fiberscopes [13]. The 2P
fluorescence photons collected by the any of two fiberscopes (mainly through the large outer cladding of
the DCF) can be separated by a dichroic mirror (DM) and then focused onto a photomultiplier tube (PMT)
for detection (Figure 2). Although the fluorescence wavelength is different from the designed working
wavelength (i.e., the 2P excitation wavelength) for the Dove prism and thus the TPF signals deviate from
the excitation beam paths, the large detection area of PMT is good enough for detection. Here we consider
GCaMP-based neural imaging as an example. According to Zemax simulations, the lateral shift of GCaMP
fluorescence (around 525 nm) beam from the 920 nm excitation beam is less than 0.2 mm, which is much
smaller than the photo cathode size in a PMT commonly used for 2P imaging.
4
Figure 2. Schematic of a dual-fiberscope rotary joint based on a Dove prism. SMF: single-mode fiber (from
the light source); FC: fiber collimator; DM: dichroic mirror; F: filter. The Dove prism is inserted into a hollow
shaft which is mounted through two bearings. Two fs laser incident beams from the stationary FC1&2 go
through the Dove prism and are coupled into rotary FC3&4 which are connected with two fiberscopes.
3. PRELIMINARY STUDIES AND RESULTS
Figure 3. (a) Schematic and (b) Photograph of preliminary experimental setup for a Dove prism based rotary
joint with one pair of fiber collimators (FCsta & FCrot). DP: Dove prism; B: bearing; S: shaft.
Ideally, the rotary joint with a Dove prism should provide excellent stability in optical coupling efficiency
at any rotational angle. However, misalignment of any optical components would result in optical
throughput variation over rotation [20]. In addition, an imperfect Dove prism itself with manufacturing
error/tolerance in geometry parameters (such as the length 𝐿 and/or the base angle ) would also impact
the coupling efficiency [21]. Furthermore, mismatch between the incident laser wavelength and the
designed wavelength for the Dove prism will lead to small but non-negligible lateral beam shift, which will
reduce the coupling efficiency and stability as well.
5
To investigate the feasibility of the dual-fiberscope rotary joint based on a Dove prism, proof-of-concept
experiments have been conducted. The most critical parameter to test is the coupling efficiency stability for
light coming from a stationary fiber collimator (i.e., FCsta in Figure 3a for the light from the laser) to the
rotating fiber (i.e., the fiberscope) and a fiber coupler after going through a half-angle rotating Dove prism
(see Figure 3a). Before a customized Dove prism with a proper length and aperture for a specific wavelength
becomes available, we selected an off-the-shelf one (PS992, Thorlabs) which was intended for 675 nm light.
We then chose a laser as the input light source available to us with a wavelength (668 nm) that is close to
the Dove prism working wavelength (675 nm). As shown in Figure 3, on the stationary side, FCsta (CFC11P-
B, Thorlabs) is connected to an x-y linear translation stage with a kinematic mount. On the rotary side, FCrot
(F240APC-B, Thorlabs) is mounted on a rotary kinematic mount, whose optical axis can be precisely
aligned parallel to its rotational axis. Once the two FCs and the Dove prism are well aligned, the incident
beam (668 nm) coming out of a stationary single-mode fiber SMF1 could be effectively coupled into the
single-mode fiber SMF2 at the rotary end with a minimum power fluctuation over rotation.
We first performed quantitative Ray-tracing analyses using ZEMAX where the fabrication tolerance/error
and wavelength mismatch for the off-the-shelf Dove prism were taken into account. The optical coupling
efficiency of the FC pair (FCsta & FCrot) at 668 nm was calculated as a function of angular and lateral
misalignment between two FCs. We concluded that to keep the coupling fluctuation below ±3% (which
would not impact the analyses of dynamic neural activities for two-photon fiberscopy brain imaging of
rodents [16]), the angular and lateral alignment tolerance over 360° rotation should be kept below ±2.4
mdeg and ±57.1 µm, respectively. Here, the Dove prism employed in our proof-of-concept experimentation
has a fabrication tolerance/error (i.e., ±0.15 mm in length and ±0.05°in base angle). Taking into
consideration both the wavelength mismatch (which is translated to a prism length mismatch) and the
fabrication tolerance of the Dove prism, a maximum lateral shift for the exit beam reaches 79.0 um,
corresponding to a normalized SMF coupling efficiency drop from an ideal 100.00% to 94.23% (see Figure
4a and b). This means the off-the-shelf Dove prism itself would cause approximately ±3% throughput
fluctuation even.
Figure 4. Preliminary feasibility studies of the performance of a Dove prism based rotary joint. (a)
Normalized SMF coupling efficiency versus lateral and angular misalignment based on quantitative Ray-
tracing analyses using Zemax. (b) Fabrication tolerance analyses by quantitative Ray-tracing. Blue curve:
normalized SMF coupling efficiency with a Dove prism sandwiched between two FCs. Red curve: lateral
shift of output beam caused by wavelength mismatch (for a Dove prism PS992 designed for 675 nm with a
laser used in experiments at 668nm) and fabrication errors of the Dove prism. (c) The measured normalized
optical throughput fluctuation over 180° rotation of FCrot.
Encouraged by the Ray-tracing analyses, we proceeded with experimental testing. The laser power
throughput from SMF1 to SMF2 was measured to be 72% which is excellent. A better than ±6% relative
fluctuation was achieved in the throughput (or coupling efficiency) over 180° rotation of FCrot
(accompanied by the compensating half angle rotation of the Dove prism over 90°) (see Figure 4). It is
6
noted that only 180° (rather than 360°) rotation of the FCrot is needed for testing owing to the rotational
symmetry. This excellent coupling or throughput stability was obtained even with non-precision bearings
and a non-tight-fit housing shaft available in our lab. It is noticed that the rotational fluctuation in the
coupling efficiency is about 2X as large as the simulation results, and this larger fluctuation was due to the
imperfect off-the-lab-shelf mechanical components (two ball bearings and rotating shaft) as well as the
mismatch between the intended wavelength for the generic Dove prism and the laser wavelength available
to us. The measured throughput fluctuation was a lightly less than ±6% (see Figure 4c). This translates to
±12% rotation-induced fluctuation in the fluorescence signal (∆F/F) during two-photon imaging of neural
dynamics, which is still considered acceptable since the relative dynamic change of neural activity related
two-photon fluorescence signal ∆F/F is generally greater than 50%.
4. CONCLUSION AND DISCUSSIONS
We have quantitatively (using Ray-tracing) and experimentally demonstrated the feasibility of a Dove prism
based rotary joint. We have analytically proved the unique optical properties of a Dove prism suited for a
dual-probe rotary joint. Quantitative Ray-tracing analyses support the feasibility of such a dual-probe rotary
joint. We further performed proof-of-concept experiments. Even without the use of precision mechanical
components and a prism of an unmatched length for the test laser wavelength, we were still able to achieve
a high coupling efficiency (~72%) and a fairly small rotational fluctuation (±6%).
Considering the quadratic dependence of the 2P fluorescence on the excitation intensity (𝐼2𝑃𝐹 ∝ 𝐼𝑒𝑥
2 ), a
variation in excitation laser intensity would induce a higher variation in the fluorescence signal. Assuming
the target rotation-induced fluorescence fluctuation is less than 10%, the acceptable excitation throughput
variation shall be maintained less than 5% over rotation. If needed, a smaller rotational variation in coupling
efficiency can be achieved by slightly sacrificing the coupling throughput, which can be compensated by
slightly increasing the input power from the laser.
For two-photon imaging of GCaMP based neural activities, a longer fs excitation wavelength at 920nm will
be used. The corresponding single-mode fiber core of the fiberscopes will be about 40% larger than the
SMF for 668 nm light used in the above experiments. The increased core diameter will help achieve stable
coupling. The use of a customized Dove prism designed for the exact wavelength of the excitation light
(i.e., 920nm) would reduce the rotational fluctuation in the coupling efficiency. In addition, the coupling
stability can also be improved by using precision bearings (as opposed to the ones we have in the lab) and
a shaft with a proper diameter for tight fit to the bearings’ inner diameter.
Although in the above experimentation we only considered one fiberscope (connected with FCrot) in the
above experiments, the same exercise can expand to a dual-probe configuration. Smaller kinematic mounts
can be used for the FCs or beam steering mirrors can be used in the beam paths to avoid potential beam
blocking by the mechanical components when two fiberscopes are connected to the dual-probe rotary joint.
Such a Dove-prism based rotary joint would enable for the first time simultaneous 2P imaging of two brain
regions in freely-walking/rotating mice. In principle, our design is not only restricted to neuroimaging of
rodents. Owing to the twist-free operation, it can be applied to non-human primates like rhesus macaque
and marmoset. We believe the system will open a new avenue for exploring neural network dynamics of
multiple interconnected brain regions associated with various behaviors.
ACKNOWLEDGEMENT: The authors are grateful for the partial support of this work by the Bisciotti
Foundation (Li and Park) and the National Institutes of Health under a grant R35CA209960 (Bhujwalla).
7
SUPPLEMENTAL MATERIALS
Supplement 1: Table Survey of off-the-shelf dual-channel fiber-optic rotary joints
Supplement 2: Analytic Proof of Dove Prism Properties with Geometric Optic
REFERENCES
[1]
W. Denk, J. H. Strickler, and W. W. Webb, "Two-photon laser scanning fluorescence microscopy," Science 248,
73-76 (1990).
[2]
K. Svoboda, and R. Yasuda, "Principles of two-photon excitation microscopy and its applications to
neuroscience," Neuron 50, 823-839 (2006).
[3]
W. R. Zipfel, R. M. Williams, and W. W. Webb, "Nonlinear magic: multiphoton microscopy in the biosciences,"
Nature biotechnology 21, 1369-1377 (2003).
[4]
F. Helmchen, and W. Denk, "Deep tissue two-photon microscopy," Nature methods 2, 932-940 (2005).
[5]
C. Xu, W. Zipfel, J. B. Shear, R. M. Williams, and W. W. Webb, "Multiphoton fluorescence excitation: new
spectral windows for biological nonlinear microscopy," Proceedings of the National Academy of Sciences 93,
10763-10768 (1996).
[6]
L. Tian, S. A. Hires, T. Mao, D. Huber, M. E. Chiappe, S. H. Chalasani, L. Petreanu, J. Akerboom, S. A.
McKinney, and E. R. Schreiter, "Imaging neural activity in worms, flies and mice with improved GCaMP
calcium indicators," Nature methods 6, 875-881 (2009).
[7]
R. P. Barretto, T. H. Ko, J. C. Jung, T. J. Wang, G. Capps, A. C. Waters, Y. Ziv, A. Attardo, L. Recht, and M. J.
Schnitzer, "Time-lapse imaging of disease progression in deep brain areas using fluorescence microendoscopy,"
Nature medicine 17, 223-228 (2011).
[8]
W. Yang, and R. Yuste, "In vivo imaging of neural activity," Nature methods 14, 349-359 (2017).
[9]
D. A. Dombeck, A. N. Khabbaz, F. Collman, T. L. Adelman, and D. W. Tank, "Imaging large-scale neural
activity with cellular resolution in awake, mobile mice," Neuron 56, 43-57 (2007).
[10] D. A. Dombeck, C. D. Harvey, L. Tian, L. L. Looger, and D. W. Tank, "Functional imaging of hippocampal
place cells at cellular resolution during virtual navigation," Nature neuroscience 13, 1433-1440 (2010).
[11] J. N. Stirman, I. T. Smith, M. W. Kudenov, and S. L. Smith, "Wide field-of-view, multi-region, two-photon
imaging of neuronal activity in the mammalian brain," Nature biotechnology 34, 857-862 (2016).
[12] W. Zong, R. Wu, S. Chen, J. Wu, H. Wang, Z. Zhao, G. Chen, R. Tu, D. Wu, and Y. Hu, "Miniature two-photon
microscopy for enlarged field-of-view, multi-plane and long-term brain imaging," Nature methods 18, 46-49
(2021).
[13] M. T. Myaing, D. J. MacDonald, and X. Li, "Fiber-optic scanning two-photon fluorescence endoscope," Optics
letters 31, 1076-1078 (2006).
[14] Y. Zhang, M. L. Akins, K. Murari, J. Xi, M.-J. Li, K. Luby-Phelps, M. Mahendroo, and X. Li, "A compact fiber-
optic SHG scanning endomicroscope and its application to visualize cervical remodeling during pregnancy,"
Proceedings of the National Academy of Sciences 109, 12878-12883 (2012).
[15] W. Liang, G. Hall, B. Messerschmidt, M.-J. Li, and X. Li, "Nonlinear optical endomicroscopy for label-free
functional histology in vivo," Light: Science & Applications 6, e17082-e17082 (2017).
[16] A. Li, H. Guan, H.-C. Park, Y. Yue, D. Chen, W. Liang, M.-J. Li, H. Lu, and X. Li, "Twist-free ultralight two-
photon fiberscope enabling neuroimaging on freely rotating/walking mice," Optica 8, 870-879 (2021).
[17] H. Guan, D. Li, H.-c. Park, A. Li, Y. Yue, Y. A. Gau, M.-J. Li, D. E. Bergles, H. Lu, and X. Li, "Deep-learning
two-photon fiberscopy for video-rate brain imaging in freely-behaving mice," Nature Communications 13, 1534
(2022).
[18] H. Guan, W. Liang, A. Li, Y.-T. A. Gau, D. Chen, M.-J. Li, D. E. Bergles, and X. Li, "Multicolor fiber-optic
two-photon endomicroscopy for brain imaging," Optics Letters 46, 1093-1096 (2021).
[19] H. Sar-El, "Revised Dove prism formulas," Applied optics 30, 375-376 (1991).
[20] C.-H. Chou, R.-J. Chen, H.-Y. Tsai, K.-C. Huang, and C.-C. Yang, "Optimization of Coupling Efficiency of
Fiber Optic Rotary Joint by Ray Tracing," in PHOTOPTICS (2020), pp. 71-75.
[21] D. Jia, C. Yu, W. Jing, H. Zhang, and Y. Zhang, "Effect of misalignment on rotating coupling efficiency of the
Dove prism," in Advanced Materials and Devices for Sensing and Imaging III(SPIE2008), pp. 128-136.
| 2022 | Feasibility Analyses and Experimental Confirmation of Dove Prism Based Dual-fiberscope Rotary Joint | 10.1101/2022.09.25.509388 | [
"Liu Yuehan",
"Park Hyeon-Cheol",
"Zhang Haolin",
"Li Xingde"
] | creative-commons |
1
Drosophila immunity: The Drosocin gene encodes two host defence pep-
1
tides with pathogen-specific roles
2
M.A. Hanson1*, S. Kondo2, and B. Lemaitre1*
3
1 Global Health Institute, School of Life Science, École Polytechnique Fédérale de Lausanne
4
(EPFL), Lausanne, Switzerland.
5
2 Invertebrate Genetics Laboratory, Genetic Strains Research Center, National Institute of Ge-
6
netics, Mishima, Japan
7
* Corresponding authors:
8
M. Hanson (mark.hanson@epfl.ch)
9
B. Lemaitre (bruno.lemaitre@epfl.ch)
10
1.1
Abstract
11
Antimicrobial peptides (AMPs) are key players in innate defence against infection in plants and animals.
12
In Drosophila, many host defence peptides are produced downstream of the Toll and Imd NF-κB path-
13
ways. Use of single and compound AMP mutations in Drosophila has revealed that AMPs can additively
14
or synergistically contribute to combat pathogens in vivo. However, these studies also revealed a high
15
degree of specificity, wherein just one AMP can play a major role in combatting a specific pathogen. We
16
recently uncovered a specific importance of the antibacterial peptide Drosocin for defence against En-
17
terobacter cloacae. Here, we show that the Drosocin locus (CG10816) is more complex than previously
18
described. In addition to its namesake peptide “Drosocin”, it encodes a second peptide generated from
19
a precursor via furin cleavage. We name this peptide “Buletin”, and show that it corresponds to the un-
20
characterized “Immune-induced Molecule 7” previously identified by MALDI-TOF. The existence of a
21
naturally occurring polymorphism (Thr52Ala) in the CG10816 precursor protein masked the identifica-
22
tion of this peptide previously. Using mutations differently affecting the production of these two
23
CG10816 gene products, we show that Drosocin, but not Buletin, contributes to the CG10816-mediated
24
defence against E. cloacae. Strikingly, we observed that Buletin, but not Drosocin, contributes to the
25
CG10816-mediated defence against Providencia burhodogranariea. Moreover, the Thr52Ala polymor-
26
phism in Buletin affects survival to P. burhodogranariea, wherein the Alanine allele confers better de-
27
fence than the Threonine allele. However, we found no activity of Buletin against either P. burhodogran-
28
ariea or E. coli in vitro. Collectively, our study reveals that CG10816 encodes not one but two prominent
29
host defence peptides with different specificity against different pathogens. This finding emphasizes the
30
complexity of the Drosophila humoral response consisting of multiple host defence peptides with spe-
31
cific activities, and demonstrates how natural polymorphisms found in Drosophila populations can af-
32
fect host susceptibility.
33
1.2
Introduction
34
The ability to rapidly combat pathogens is critical to organism health and survival. Or-
35
ganisms sense natural enemies through pattern recognition receptors, triggering the activation
36
of core immune signalling pathways. These pathways regulate the expression of a plethora of
37
immune effectors that provide a first line of innate defence. It was generally thought that innate
38
2
immune effectors act together as a cocktail to kill microbes. However recent studies have chal-
39
lenged this view revealing an unexpected high degree of specificity in the effector response to
40
infection [1–3].
41
Chief amongst immune effectors are antimicrobial peptides (AMPs), host-encoded anti-
42
biotics that exhibit microbicidal activities [1,2,4,5]. Insects, and particularly the genetically trac-
43
table model Drosophila, have been especially fruitful in identifying and characterizing AMP po-
44
tency and function [4,6–9]. In Drosophila, systemic infection triggers the expression of a battery
45
of antimicrobial peptides that are secreted into the hemolymph by the fat body to transform
46
this compartment into a potent microbicidal environment. This systemic AMP response is
47
tightly regulated by two signalling cascades: the Toll and Imd pathways. These two pathways
48
are similar to mammalian TLR and TNFalpha NF-κB signalling that regulate the inflammatory
49
response [10,11]. They are differentially activated by different classes of microbes. The Toll
50
pathway is predominantly instigated after sensing infection by Gram-positive bacteria and
51
fungi, while the Imd pathway is especially responsive to Gram-negative bacteria and some
52
Gram-positive bacteria with DAP-type peptidoglycan [11–13]. The expression of each AMP
53
gene is complex, receiving differential input from either pathway, with most AMPs being at least
54
somewhat co-regulated during the systemic immune response [14–16].
55
In Drosophila, several families of AMPs contribute downstream of Toll and Imd. This in-
56
cludes the Cecropin, Attacin, Diptericin, Defensin, Metchnikowin, Drosomycin, Baramicin, and
57
Drosocin gene families [1,3,4]. Other host defence peptide families include Daisho and Bo-
58
manin, which are important for defence, but in vitro killing activity is yet to be shown [17,18].
59
How these immune effectors contribute individually or collectively to host defence remains
60
poorly understood. Use of single and compounds mutants has revealed that defence against
61
some pathogens relies on the collective contributions of multiple AMP families. However recent
62
studies have also shown that single defence peptides can play highly specific and important
63
roles during infection. In one case, Diptericins are the critical AMP family for surviving infection
64
by Providencia rettgeri bacteria. This specificity is so remarkable that flies collectively lacking
65
five other AMP gene families nevertheless resist P. rettgeri infection like wild-type [6], while
66
even a single amino acid change in one Diptericin gene can cause pronounced susceptibility to
67
P. rettgeri [19]. Studies on Toll effector genes such as Bomanins, Daishos, or Baramicin A have
68
also found deletion of single gene families can cause strong susceptibilities against specific fun-
69
gal species [18,20], or mediate general defences against broad pathogen types [17,21]. Lastly,
70
loss of the gene Drosocin causes a specific and pronounced susceptibility to infection by
71
3
Enterobacter cloacae [6], agreeing with Drosocin peptide activity in vitro [22]. Unlike the exam-
72
ple with Diptericins and P. rettgeri, other AMPs also contribute collectively to defence against E.
73
cloacae [23].
74
Many AMP genes encode precursor proteins with multiple peptide products processed
75
by furin cleavage [20]. This was initially shown for the Apidaecin gene of honey bees, which
76
produces nine Apidaecin peptides from a single precursor [24]. Drosophila also encodes many
77
AMPs with polypeptide precursors. Examples include AMPs of the Attacin and Diptericin gene
78
families [25,26] or Baramicin A which encodes three kinds of unique peptide products on a sin-
79
gle precursor protein [20,27,28]. Meanwhile, the precursor protein of the nematode AMP
80
"NLP29" is cleaved into six similar Glycine-rich peptides [29,30]. To our knowledge, the inde-
81
pendent contributions of sub-peptides from a polypeptide AMP gene has so far never been ad-
82
dressed.
83
In this study, we reveal that the gene CG10816 encodes not only the antibacterial Droso-
84
cin peptide, but also another host defence peptide produced by furin cleavage of the Drosocin
85
precursor protein. We name this peptide Buletin, and show that it corresponds to IM7, an in-
86
ducible peptide first identified in 1998 by MALDI-TOF analysis whose gene counterpart was
87
never identified [31]. Using a new mutation affecting only the Drosocin peptide and not Buletin,
88
we show that these two peptides contribute independently to defence against different mi-
89
crobes. Survival analyses show that while Drosocin specifically affects defence against E. cloa-
90
cae, Buletin specifically affects defence against Providencia burhodogranariea. Moreover, a pre-
91
viously identified polymorphic site in Buletin (Thr52Ala described in [32]) mirrors the suscep-
92
tibility effect of Buletin deletion to P. burhodogranariea. We therefore uncover a striking exam-
93
ple where an AMP-encoding gene produces two peptides with distinct activities. The
94
CG10816/Drosocin gene is also an example of how an AMP polymorphism can significantly af-
95
fect the host defence against a specific microbe. Alongside recent findings using Diptericin and
96
P. rettgeri, our results highlight how AMP evolution is likely driven by differential activity
97
against ecologically-relevant microbes.
98
Results
99
For clarity of discussion: we will use the shorthand Drc (with a “c”, no italics) to refer to
100
the mature Drosocin peptide. Whenever possible, we will use CG10816 to refer to the Drosocin
101
gene (common shorthand Dro, with an “o”, italicized).
102
4
The Drosocin gene CG10816 encodes IM7
103
Previous proteomic analyses of hemolymph from infected Drosophila revealed several
104
Immune-induced Molecules (IMs) [31]. These molecules were annotated as IM1-IM24 accord-
105
ing to their mass, and over time each of these IMs was associated with a host defence peptide
106
gene [17,18,20,33]. At this point, only one of the 24 original IMs remains unknown: IM7. Previ-
107
ous efforts were unable to link this 2307 Da peptide to a gene in the Drosophila reference ge-
108
nome. However during our studies, we noticed that IM7 was absent in flies lacking 14 AMP
109
genes, indicating that it is likely produced by one of these genes [6,23]. We repeated these
110
MALDI-TOF proteomic experiments with hemolymph samples from flies carrying systematic
111
combinations of AMP mutations, ultimately honing in on the Drosocin-encoding gene
112
CG10816/Dro. Two independent CG10816/Dro mutants (DroSK4 and Dro-AttABSK2) both lack IM7
113
in MALDI-TOF peptidomic analysis (Fig. 1).
114
CG10816/Dro was initially identified as a single ORF gene encoding the Drc peptide. Drc
115
is an O-glycosylated Proline-rich peptide that binds bacterial DnaK/Hsp70 similar to other Pro-
116
line-rich insect AMPs [22,34–36]. Mature Drc requires O-glycosylation for activity, which in-
117
volves the biochemical linking of either mono- (MS), di- (DS), or rarely tri-saccharide (TS)
118
groups to the Threonine at position 11 of the Drc peptide [22,33]. These different O-glycosyla-
119
tions yield peptides with different mature masses of 2401, 2564, and 2767 Da (Drc-MS, -DS, and
120
-TS respectively). Unmodified Drc peptide has an expected mass of 2199 Da, which is not an
121
intuitive match for the 2307 Da peak of IM7, even considering other post-translational modifi-
122
cations. This suggests that another element of the CG10816/Dro gene encodes IM7.
123
5
124
Figure 1: The CG10816/Dro gene encodes a polypeptide including both Drc and IM7. A) Overview of the precursor
125
protein structure of the CG10816/Dro gene. The Thr52Ala polymorphism in IM7 was noted previously [32]. Here
126
we include an alignment of the CG10816 precursor protein between the Dmel_R6 reference genome and se-
127
quences from iso w1118, DroSK3, DroSK4, and DGRP-822 flies. B) MALDI-TOF proteomic data from immune-challenged
128
flies shows that both Drc (Drc-MS, Drc-DS) and the 2307 Da peak of IM7 is absent in DroSK4 and Dro-AttABSK2 flies.
129
The frameshift present in DroSK3 removes the Drc peptide, but does not prevent the secretion of IM7. Threonine-
130
encoding IM7 appears in DGRP-822 (2337 Da), alongside loss of the 2307 Da peak.
131
IM7 is the C-terminus product of the CG10816 precursor
132
It is puzzling that IM7 could not be annotated to the CG10816/Dro gene given that the
133
nucleotide sequence has been known for decades. One previous study noted that the
134
CG10816/Dro gene was likely cleaved at a furin-like cleavage site, and had a small undescribed
135
C-terminal peptide [25]. Lazzaro and Clark [32] further described a polymorphism in the
136
CG10816/Dro gene encoding either a Threonine or Alanine at residue 52 in the C-terminus of
137
DGRP-822
MKFTIVFLLLACVFAMAVA TP GKPRPYSPRPTSHPRPIRV
EALAIEDHLTQAAIRPPPILPA
MKFTIVFLLLACVFAMAVA TP
DroSK4
MKFTIVFLLLACVFAMAVA TP GKPRPYSPRPTSHPRPIRV
EALAIEDHLAQAAIRPPPILPA
iso w1118
MKFTIVFLLLACVFAMAVA TH SVA
SHPRPIRV
EALAIEDHLAQAAIRPPPILPA
DroSK3
Dmel_R6
MKFTIVFLLLACVFAMGVA TP GKPRPYSPRPTSHPRPIRV
EALAIEDHLTQAAIRPPPILPA
MKFTIVFLLLACVFAM(A/G)VATPGKPRPYSPRPTSHPRPIRVRREALAIEDHL(T/A)QAAIRPPPILPA
Signal peptide
Mature Drosocin
IM7 peptide
Furin
DP
RR
RR
RR
RR
RQA … encodes 59-residue nonsense peptide … RSNLF
A
B
DroSK4
iso w1118
Dro-AttABSK2
iso w1118
IM7-A
Drc-MS
Drc-DS
IM12
IM13
IM7-A
IM7-A
IM7-A
Drc-MS
Drc-DS
IM12
IM13
IM10
Drc-MS
Drc-DS
IM12
IM13
IM10
Drc-MS
Drc-DS
IM12
IM13
IM10
IM10
IM7-A
DroSK3
iso w1118
DGRP-822
iso w1118
IM7-T
Drc-DS
6
the precursor protein sequence (Thr52Ala). The D. melanogaster reference genome encodes
138
the Threonine version of this polymorphism. Using the sequence of the reference genome, the
139
CG10816 C-terminus mature mass would be 2337 Da without considering post-translational
140
modifications. If we instead substitute an Alanine at this site, the predicted mass of the CG10816
141
C-terminus becomes 2307 Da, exactly matching the observed mass of IM7. We confirmed that
142
our wild-type DrosDel isogenic genetic background encodes an Alanine allele both by Sanger
143
sequencing and LC-MS proteomics (data not shown). We next performed MALDI-TOF on the
144
hemolymph of flies from DGRP strain 822 (DGRP-822), which encodes a Threonine in its C-ter-
145
minus. Exactly matching prediction, DGRP-822 flies lack the 2307 Da IM7 peak, and instead have
146
a 2337 Da peak that appears after infection (Fig. 1B).
147
Serendipitously, while generating CG10816/Dro mutants using CRISPR-Cas9 we recov-
148
ered a complex aberrant locus (DroSK3) that disrupts 11 amino acid residues of the mature Drc
149
peptide, including its critical O-glycosylated Threonine (Fig. 1A). However the DroSK3 deletion
150
later continues in the same reading frame, including the RVRR furin cleavage site and C-termi-
151
nus. Thus we suspected that the C-terminal peptide would be secreted normally in DroSK3 flies.
152
When we ran MALDI-TOF analysis on immune-induced hemolymph from DroSK3 flies, we recov-
153
ered a signal that all but confirmed the identity of the CG10816 C-terminus: DroSK3 flies lacked
154
the Drc-MS and Drc-DS peaks, but the 2307 Da peak corresponding to IM7 remained immune-
155
inducible (Fig. 1B).
156
Taken together, we reveal that CG10816 encodes two peptides: Drc and IM7, which are
157
produced from a precursor protein by cleavage at a canonical furin cleavage site. IM7 is a 22-
158
residue peptide with a net anionic charge (-1.9 at pH = 7) that does not share overt similarity
159
with Drc (+5.1 at pH = 7), though both peptides are Proline-rich. A naturally occurring poly-
160
morphism previously obscured the annotation of IM7 as a CG10816 gene product. This analysis
161
was greatly facilitated by the use of newly-available AMP mutations. We name this C-terminal
162
peptide Buletin (Btn) after Philippe Bulet, whose dedicated efforts in the 1980s-1990s charac-
163
terized many of the Drosophila AMPs including Drosocin [4,22,37].
164
Drc, but not Btn, is responsible for the CG10816-mediated defence against Enterobacter
165
cloacae
166
Previous studies have suggested that flies lacking just the CG10816/Dro locus can resist
167
infection by most bacteria, but are specifically susceptible to infection by E. cloacae [6], and also
168
somewhat E. coli [38] and Providencia burhodogranariea [6]. The fact that CG10816 encodes not
169
7
one but two peptides raises the question of the specific contribution of these two peptides to
170
CG10816 effects. Therefore, we took advantage of the DroSK3 and DroSK4 mutations that differ-
171
ently affect the Drc and Btn peptides (Fig. 1A) to explore the respective role(s) these peptides
172
play by comparing the survival of these mutants to different infections. We focused our screen
173
on a panel of Gram-negative bacteria of interest: E. cloacae β12 bacteria that CG10816/Dro mu-
174
tants are specifically susceptible to [6,23], a recently-isolated Acetobacter sp. that can kill AMP
175
mutant flies [39], E. coli 1106 suggested to be affected by CG10816/Dro [22,38], and P. burhod-
176
ogranariea strain B where CG10816/Dro was shown to contribute to defence alongside other
177
AMPs [6]. All experiments were performed with wild-type and mutant flies that were
178
isogenized in the DrosDel genetic background according to Ferreira et al. [40].
179
We found that individual CG10816/Dro mutants (both DroSK3 and DroSK4) were not
180
overtly susceptible to infection by E. coli 1106 or Acetobacter sp. ML04.1 (Fig. S1). We could also
181
repeat our previous findings that DroSK4 and Dro-AttABSK2 flies were highly susceptible to E. clo-
182
acae infection, causing 40-50% mortality by 3 days after infection. Importantly, use of DroSK3
183
flies that lack Drc but produce Btn confirms that this susceptibility is principally caused by a
184
loss of Drc peptide and not Btn (Fig. 2A): DroSK4 and Dro-AttABSK2 flies lacking both Drc and Btn
185
were only slightly more susceptible than DroSK3 flies lacking Drc alone, a difference that was not
186
statistically significant (DroSK4 and Dro-AttABSK2 comparisons to DroSK3, p > .05 in both cases).
187
Thus, comparison of mutants lacking Drc, or both Drc and Btn, reveals that the CG10816-
188
mediated defence against E. cloacae is specifically mediated by the Drc peptide. Meanwhile Btn
189
does not seem to contribute to defence against this bacterial infection in a significant way.
190
Btn, but not Drc, is important for survival to P. burhodogranariea infection
191
We previously found that CG10816 could contribute to defence against P. burhodogran-
192
ariea synergistically alongside Diptericins and Attacins [6]. We next assessed the contribution
193
of our different CG10816/Dro gene mutants to defence against P. burhodogranariea. To our
194
surprise, the presence or absence of Btn causes a pronounced survival difference after infec-
195
tion by P. burhodogranariea: DroSK3 flies that still produce Btn survive as wild type, while
196
DroSK4 or Dro-AttABSK2 flies suffer significantly increased mortality (Fig. 4B). This trend is the
197
opposite of what is observed after infection with E. cloacae: Drc
198
8
199
Figure 2: Mutations affecting Buletin cause a specific susceptibility to P. Burhodogranaria. A) DroSK3 flies suc-
200
cumb to infection by E. cloacae slightly later than either DroSK4 or Dro-AttABSK2 flies that lack both Drc and Btn. The
201
ultimate rate of mortality is comparable (p > .05 in comparisons between these various Dro mutants). B) Drosocin
202
mutants that retain Btn (DroSK3) survive infection by P. burhodogranariea better than flies lacking both Drc and
203
Btn (DroSK4, DroAttSK2). C) Wild-type flies with the Threonine allele of the Btn Thr52Ala polymorphism phenocopy
204
the effect of Btn deletion compared to Alanine-encoding iso w1118 in defence against P. burhodogranariea.
205
does not play an important role in defence against P. burhodogranariea, but Btn does. As em-
206
phasized by the greater susceptibility of AMP-deficient ΔAMP14 and Imd-deficient RelE20 flies
207
(Fig. 2B), Btn deficiency explains only part of the susceptibility to P. burhodogranariea. This is
208
consistent with our previous study, which showed that CG10816/Dro contributes to defence
209
against this bacterium alongside the contributions of Diptericin and Attacin genes.
210
Collectively, our study shows that the CG10816/Dro locus encodes two host-defence
211
peptides with distinct activities in vivo. This reinforces the notion that innate immune effectors
212
can have very specific roles in vivo.
213
The Thr52Ala polymorphism affects Btn activity against P. burhodogranariea in vivo
214
The existence of a Threonine/Alanine polymorphic residue in Btn in natural fly popula-
215
tions suggests an arms race between Btn and naturally occurring pathogens. Such polymor-
216
phisms are common in AMP genes, and are proposed to reflect host-pathogen coevolutionary
217
selection [41,42]. The P. burhodogranariea strain used in this study was originally isolated
218
from the hemolymph of wild-caught flies [43], suggesting it is an ecologically relevant microbe
219
to D. melanogaster. This prompted us to investigate the contribution of this polymorphism in
220
defence against P. burhodogranariea. We next isolated a Btn-Threonine allele (BtnThr) that we
221
introgressed into the DrosDel background over seven generations. We infected isogenic BtnThr
222
and BtnAla (i.e. iso w1118) flies with P. burhodogranariea to determine if the Btn polymorphism
223
0
1
2
3
4
5
6
7
0
25
50
75
100
P. burhodogranariea B,
OD = 10, 25°C
nexp = 4
0
1
2
3
4
5
6
7
0
25
50
75
100
P. burhodogranariea B,
OD = 10, 25°C
nexp = 4
0
1
2
3
4
5
6
7
0
25
50
75
100
E. cloacae β12,
OD = 200, 25°C
nexp = 4
iso w1118
DroSK3
DroSK4
DroAttSK2
RelE20
ΔAMP14
*
iso BtnThr
B
A
Legend:
Percent survival
Time (days)
C
*
Drc :
Btn
:
-
+ (Ala)
+
+ (Ala)
-
-
-
-
+
+ (Thr)
9
impacts survival. In these experiments, iso BtnThr flies suffered a ~15% increase in mortality
224
compared to iso w1118 flies with BtnAla (Fig. 2C, p = .037). The Cox survival hazard ratio is a
225
measure of effect size. The hazard ratio of DroSK4 vs. DroSK3 flies (Fig. 2B) and iso BtnThr vs. iso
226
w1118 (Fig. 2C) is nearly-identical (hazard ratios: DroSK4-DroSK3 = 0.590, BtnThr-iso w1118: =
227
0.584). Thus the effect size of changing the Btn allele from Alanine to Threonine causes the
228
same hazard ratio difference as the effect of Btn deletion.
229
We therefore uncover an important role of Btn in defence against P. burhodogranariea,
230
and reveal that the Btn Thr52Ala polymorphism impacts survival against this ecologically rel-
231
evant pathogen. Alongside the effect of a polymorphism in Diptericin on survival to P. rettgeri
232
[19], here we provide a second example of how a polymorphic residue in an AMP gene signifi-
233
cantly impacts survival.
234
Discussion
235
Here we show that the CG10816/Dro gene encodes two peptides with distinct activities
236
in vivo. Buletin was not annotated previously due to a polymorphism that masked the identity
237
of this second peptide. Most immune studies have used Drosophila strains that encode the BtnAla
238
allele (e.g. Oregon-R [31], w1118 [44], DrosDel [6], or Canton-S [45]), while the D. melanogaster
239
reference genome encodes the BtnThr allele. The gene CG10816 produces a precursor protein
240
cleaved in two locations: i) after the signal peptide at a two-residue dipeptidyl peptidase site
241
that is nibbled off of the N-terminus of mature Drc (Fig. S3, similar sites noted in [20,46]), and
242
ii) at a furin cleavage motif that separates the Drc and Btn peptides (“RVRR” in CG10816). Both
243
cleavage motifs are common in AMP genes, including Drosophila Attacins, Defensins, Dipteri-
244
cins, and Baramicins, which all encode mature peptides separated by furin cleavage sites
245
[1,20,25].
246
The CG10816/Dro gene is restricted to the genus Drosophila [47]. However phylogenetic
247
inference for AMPs is difficult due to their short size [48,49], and functional analogues of the
248
Drc peptide that may share an evolutionary history are described in many holometabolous in-
249
sects [50]. It is therefore noteworthy that the range of Buletin is far more restricted: Buletin-
250
like peptides are found only in Dro genes of fruit flies ranging from the Melanogaster to Obscura
251
groups, and not in outgroup Drosophila species (Fig. S2). The Buletin peptide is therefore an
252
evolutionary novelty of the CG10816 gene C-terminus. The Thr52Ala polymorphism in Buletin
253
is likely maintained by balancing selection [42], similar to a trade-off between alternate alleles
254
of Diptericin in defence against the related bacterium Providencia rettgeri [19]. The apparent
255
10
cost of the Thr52Ala polymorphism to surviving infection by P. burhodogranariea suggests an
256
evolutionary trade-off between defence against this bacterium and some other function.
257
The Drc and Btn peptides are not homologous, although both are rich in Proline residues.
258
However Drosocin is O-glycosylated and has a strong cationic charge (+5.1 at pH = 7), while
259
Buletin is unmodified and has a net anionic charge (-1.9 at pH = 7). AlphaFold predicts Buletin
260
to have an a-helical structure [51]. We screened for Buletin activity in vitro diluted in LB ac-
261
cording to Wiegand et al. [52]. However in our conditions, we found no effect of Buletin using
262
either BtnThr or BtnAla against P. burhodogranariea or E. coli, even when co-incubated with sub-
263
lethal concentrations of Cecropin (Sigma) (Fig. S4). It is possible that Buletin contributes to host
264
defence alongside a co-factor, or protects the host from a virulence factor secreted by P. burhod-
265
ogranariea. We do not wish to rule out a direct action of Btn on bacteria though, as our in vitro
266
conditions could have been sub-optimal for revealing an antimicrobial effect. For instance, an
267
anionic AMP of the Greater wax moth synergizes with Lysozyme to kill E. coli [53], and AMPs
268
can act synergistically in vitro through complimentary mechanisms of action [26,36,54,55].
269
While in vitro approaches are a powerful demonstration for AMP function, we are realizing
270
more and more that this is not sufficient to understand peptide activity in vivo. For example,
271
the activity of azithromycin antibiotic changes 64-fold if tested in standard in vitro conditions
272
or with the addition of human serum [56]. Likewise Bomanin peptides do not display activity
273
in vitro, but Bomanin-deficient hemolymph loses Candida-killing activity [21]. While AMPs
274
were first identified for their potent microbicidal activity in vitro [4,9,57], recent studies in Dro-
275
sophila have recovered striking specificity of AMPs in defence in vivo that was never predicted
276
from in vitro analyses [6,18,19]. These results suggest both in vitro and in vivo approaches are
277
necessary to shed light on host defence peptide activity.
278
It is striking that the Threonine/Alanine polymorphism in Buletin affects the fly defence
279
against P. burhodogranariea. This polymorphism is found in wild populations of D. melano-
280
gaster, and at high frequencies in the Drosophila Genetic Reference Panel: 29% Threonine, 64%
281
Alanine, 7% unknown at DGRP allele 2R_10633648_SNP [32,58]. A polymorphism in Diptericin
282
A causes a profound susceptibility to defence against Providencia rettgeri [19], and similar pol-
283
ymorphisms are found in various AMP genes of flies [41,42] and other AMP genes from animals
284
including fish, birds, and humans [59–61]. We now add our study on Buletin and P. burhodo-
285
granariea to the building evidence that such polymorphisms can have major impacts on micro-
286
bial control. The existence of polymorphisms in AMP genes could have important implications
287
on the survival of species. For instance: we might wonder if inbreeding in honeybees could have
288
11
fixed disadvantageous AMP alleles contributing to colony collapse disorder [62]. Reduced AMP
289
expression is also associated with conditions like psoriasis [63] or susceptibility to Pseudomo-
290
nas aeruginosa infections in cystic fibrosis patients [64,65]. A targeted screen has even sug-
291
gested polymorphisms in human ß-Defensins correlate with atopic dermatitis [66]. Could poly-
292
morphisms in human AMPs help explain predisposition to similar infectious syndromes?
293
Conclusion
294
By uncovering a novel host defence peptide, our study contributes to a growing body of
295
literature establishing the Drosophila systemic infection model as boasting the unique ability to
296
reveal specific interplay of host effector-pathogen interactions. This mode of infection allows
297
the use of the fly hemolymph as an arena to monitor pathogen growth in the presence of effec-
298
tors, with fly survival as a rapid readout. While previous studies in vitro have suggested fly
299
AMPs had generalist activities, use of specific mutations affecting individual AMP genes has now
300
revealed specific relationships between host and pathogen. Early in vitro studies would never
301
have predicted the highly specific requirement for only single peptides in defence against spe-
302
cific pathogens. Taking lessons from the fly, it should be of significant interest to characterize
303
the differential activity of AMP polymorphisms in humans and other animals, which could re-
304
veal critical risk factors for infectious diseases.
305
1.3
Materials and Methods
306
Fly genetics
307
Genetic variants were isogenized into the DrosDel isogenic background over 7 genera-
308
tions as described in [40]. The specific mutations studied here were sourced as follows: the
309
DroSK3 mutation was generated by CRISPR-Cas9 via gRNA injection as described in [67]. The
310
DroSK3 sequence was validated by Sanger sequencing and the nucleotide and translated se-
311
quence is shown in Figure S3A. DroSK3 flies encode a truncated version of the Drc peptide lacking
312
its critical Threonine needed for O-glycosylation, and we could detect variants of this truncated
313
Drc peptide in MALDI-TOF spectra with variable degradation of the N-terminus (Fig. S3A-B).
314
The BtnThr allele used in this study was originally detected in DefSK3 flies from Parvy et al. [68]
315
by virtue of mutation-specific MALDI-TOF proteomics while screening for possible source
316
genes of IM7. After isogenization, iso BtnThr flies were confirmed to have a wild-type Defensin
317
gene by PCR. Sequence comparisons were made using Geneious R10.
318
Microbe culturing conditions for infections
319
12
Bacteria were grown to mid-log phase shaking at 200rpm in their respective growth
320
media (Luria Bertani, MRS+Mannitol) and temperature conditions, and then pelleted by cen-
321
trifugation to concentrate microbes. Resulting cultures were diluted to the desired optical den-
322
sity at 600nm (OD) for survival experiments, which is indicated in each figure. The following
323
microbes were grown at 37°C: Escherichia coli strain 1106 (LB), Providencia rettgeri (LB). The
324
following microbes were grown at 29°C: Providencia burhodogranariea (LB) and Acetobacter
325
sp. ML04.1 (MRS+Mannitol).
326
In vitro antibacterial assays
327
Both the BtnThr and BtnAla versions of the 22-residue IM7 peptide were synthesized by
328
GenicBio to a purity of >95%, and silk moth Cecropin A was provided by Sigma-Aldrich at a
329
purity of ≥97%. Peptide preparations were verified by HPLC. Peptides were dissolved in water,
330
and concentrations verified by a combination of BCA assay and Nanodrop A205 readings along-
331
side a BSA standard curve. We screened Btn for activity against both P. burhodogranariea and
332
E. coli alone at 100µM-1mM, or at 100µM in combination with serially diluted Cecropin concen-
333
trations spanning the Cecropin MIC (10µM-0.1µM). Microbes were allowed to grow to log-
334
growth phase, at which point they were diluted to OD = 0.0005 in LB, and then 80μL of this
335
dilute culture was added to 20μL of water or peptide mix to reach desired concentrations in a
336
96-well plate. Bacteria-peptide solutions were left overnight at room temperature and checked
337
for growth the next morning, and in one experiment optical density at 600nm was recorded
338
every ten minutes using a TECAN plate reader (Fig. S4).
339
Using these conditions, we found an MIC for Cecropin A against E. coli 1106 of ~1µM,
340
agreeing with previous E. coli literature [69]. We found an MIC of Cecropin A against P. burhod-
341
ogranariea of ~5µM, though even 0.63µM delays growth by ~3 hours compared to no-peptide
342
controls (Fig. S4). Even at 1mM, neither the BtnThr nor BtnAla showed any growth inhibition
343
alone, and 100µM peptide combinations with Cecropin A showed no reduction of MIC over Ce-
344
cropin A alone. 100µM represents the upper limit of AMP concentration in fly hemolymph after
345
infection [70], and the concentration of Btn in vivo is likely much lower than this based on
346
MALDI-TOF relative peak intensities [6,20,31,33]. As we tested Btn alone at 1mM, and at 100µM
347
Btn + Cecropin across the Cecropin MIC range, we find that at least in our conditions using LB
348
as diluent, Btn does not display in vitro activity.
349
Survival experiments
350
13
Survival experiments were performed as previously described [6], with 20 flies per vial
351
with total replicate experiment number reported within figures (nexp). ~5 day old males were
352
used in experiments, pricked in the thorax at the pleural sulcus. Flies were flipped thrice
353
weekly. Statistical analyses were performed using a Cox proportional hazards (CoxPH) model
354
in R 3.6.3.
355
Proteomic analyses
356
Raw hemolymph samples were collected from immune-challenged flies for MALDI-TOF
357
proteomic analysis as described previously [6,31]. In brief, hemolymph was collected by capil-
358
lary and transferred to 0.1% TFA before addition to acetonitrile universal matrix. Representa-
359
tive spectra are shown. Peaks were identified via corresponding m/z values from previous
360
studies [20,33]. Spectra were visualized using mMass, and figures were additionally prepared
361
using Inkscape v0.92.
362
Author contributions:
363
MAH performed bioinformatic analyses and planned and performed infection experiments. BL
364
supervised the project and MAH and BL wrote the manuscript. SK generated and supplied
365
DroSK3 flies.
366
Acknowledgements:
367
This research was supported by Sinergia grant CRSII5_186397 and Novartis Foundation
368
532114 awarded to Bruno Lemaitre. We would like to thank Adrien Schmid and Jonathan Pittet
369
of the EPFL Proteomics Core Facility (PCF) for their technical expertise.
370
References
371
1.
Hanson MA, Lemaitre B. 2020 New insights on Drosophila antimicrobial peptide function in host defense and beyond. Curr.
372
Opin. Immunol. 62, 22–30. (doi:10.1016/j.coi.2019.11.008)
373
2.
Lazzaro BP, Zasloff M, Rolff J. 2020 Antimicrobial peptides: Application informed by evolution. Science 368. (doi:10.1126/sci-
374
ence.aau5480)
375
3.
Lin SJH, Cohen LB, Wasserman SA. 2020 Effector specificity and function in Drosophila innate immunity: Getting AMPed and
376
dropping Boms. PLoS Pathog 16, e1008480. (doi:10.1371/journal.ppat.1008480)
377
4.
Imler J-L, Bulet P. 2005 Antimicrobial peptides in Drosophila: structures, activities and gene regulation. Chemical immunology
378
and allergy 86, 1–21. (doi:10.1159/000086648)
379
5.
Mookherjee N, Anderson MA, Haagsman HP, Davidson DJ. 2020 Antimicrobial host defence peptides: functions and clinical
380
potential. Nat Rev Drug Discov (doi:10.1038/s41573-019-0058-8)
381
6.
Hanson MA, Dostálová A, Ceroni C, Poidevin M, Kondo S, Lemaitre B. 2019 Synergy and remarkable specificity of antimicrobial
382
peptides in vivo using a systematic knockout approach. eLife 8. (doi:10.7554/elife.44341)
383
14
7.
Hultmark D, Engström A, Andersson K, Steiner H, Bennich H, Boman HG. 1983 Insect immunity. Attacins, a family of antibacterial
384
proteins from Hyalophora cecropia. The EMBO Journal (doi:10.1002/j.1460-2075.1983.tb01465.x)
385
8.
Rolff J, Schmid-Hempel P. 2016 Perspectives on the evolutionary ecology of arthropod antimicrobial peptides. Philosophical
386
Transactions of the Royal Society B: Biological Sciences 371. (doi:10.1098/rstb.2015.0297)
387
9.
Steiner H, Hultmark D, Engström \AA, Bennich H, Boman HG. 1981 Sequence and specificity of two antibacterial proteins in-
388
volved in insect immunity. Nature 292, 246–248. (doi:10.1038/292246a0)
389
10. Buchon N, Silverman N, Cherry S. 2014 Immunity in Drosophila melanogaster - from microbial recognition to whole-organism
390
physiology. Nature Reviews Immunology 14. (doi:10.1038/nri3763)
391
11. Lemaitre B, Hoffmann J. 2007 The Host Defense of Drosophila Melanogaster. Annual Review of Immunology 25, 697–743.
392
(doi:10.1146/annurev.immunol.25.022106.141615)
393
12. Kaneko T et al. 2004 Monomeric and polymeric gram-negative peptidoglycan but not purified LPS stimulate the Drosophila IMD
394
pathway. Immunity 20, 637–649. (doi:10.1016/s1074-7613(04)00104-9)
395
13. Leulier F, Parquet C, Pili-Floury S, Ryu J-H, Caroff M, Lee W-J, Mengin-Lecreulx D, Lemaitre B. 2003 The Drosophila immune
396
system detects bacteria through specific peptidoglycan recognition. Nat Immunol 4, 478–484. (doi:10.1038/ni922)
397
14. De Gregorio E, Spellman PT, Tzou P, Rubin GM, Lemaitre B. 2002 The Toll and Imd pathways are the major regulators of the
398
immune response in Drosophila. EMBO Journal 21, 2568–2579. (doi:10.1093/emboj/21.11.2568)
399
15. Lemaitre B, Reichhart JM, Hoffmann JA. 1997 Drosophila host defense: differential induction of antimicrobial peptide genes
400
after infection by various classes of microorganisms. Proceedings of the National Academy of Sciences of the United States of
401
America 94, 14614–9. (doi:10.1073/pnas.94.26.14614)
402
16. Tanji T, Hu X, Weber ANR, Ip YT. 2007 Toll and IMD Pathways Synergistically Activate an Innate Immune Response in Drosophila
403
melanogaster. Molecular and Cellular Biology 27, 4578–4588. (doi:10.1128/MCB.01814-06)
404
17. Clemmons AW, Lindsay SA, Wasserman SA. 2015 An Effector Peptide Family Required for Drosophila Toll-Mediated Immunity.
405
PLoS Pathogens 11. (doi:10.1371/journal.ppat.1004876)
406
18. Cohen LB, Lindsay SA, Xu Y, Lin SJH, Wasserman SA. 2020 The Daisho Peptides Mediate Drosophila Defense Against a Subset of
407
Filamentous Fungi. Front Immunol 11, 9. (doi:10.3389/fimmu.2020.00009)
408
19. Unckless RL, Howick VM, Lazzaro BP. 2016 Convergent Balancing Selection on an Antimicrobial Peptide in Drosophila. Current
409
Biology 26, 257–262. (doi:10.1016/j.cub.2015.11.063)
410
20. Hanson MA, Cohen LB, Marra A, Iatsenko I, Wasserman SA, Lemaitre B. 2021 The Drosophila Baramicin polypeptide gene pro-
411
tects against fungal infection. PLoS Pathog 17, e1009846. (doi:10.1371/journal.ppat.1009846)
412
21. Lindsay SA, Lin SJH, Wasserman SA. 2018 Short-Form Bomanins Mediate Humoral Immunity in Drosophila. J Innate Immun 10,
413
306–314. (doi:10.1159/000489831)
414
22. Bulet P, Urge L, Ohresser S, Hetru C, Otvos L. 1996 Enlarged scale chemical synthesis and range of activity of drosocin, an O-
415
glycosylated antibacterial peptide of Drosophila. European Journal of Biochemistry 238, 64–69. (doi:10.1111/j.1432-
416
1033.1996.0064q.x)
417
23. Carboni A, Hanson MA, Lindsay SA, Wasserman SA, Lemaitre B. 2021 Cecropins contribute to Drosophila host defence against
418
fungal and Gram-negative bacterial infection. Genetics , iyab188. (doi:10.1093/genetics/iyab188)
419
24. Casteels-Josson K, Capaci T, Casteels P, Tempst P. 1993 Apidaecin multipeptide precursor structure: a putative mechanism for
420
amplification of the insect antibacterial response. The EMBO journal 12, 1569–78. (doi:10.1002/j.1460-2075.1993.tb05801.x)
421
25. Hedengren M, Borge K, Hultmark D. 2000 Expression and evolution of the Drosophila attacin/diptericin gene family. Biochemical
422
and biophysical research communications 279, 574–81. (doi:10.1006/bbrc.2000.3988)
423
26. Rabel D, Charlet M, Ehret-Sabatier L, Cavicchioli L, Cudic M, Otvos L, Bulet P. 2004 Primary Structure and in Vitro Antibacterial
424
Properties of the Drosophila melanogaster Attacin C Pro-domain. Journal of Biological Chemistry 279, 14853–14859.
425
(doi:10.1074/jbc.M313608200)
426
15
27. Hanson MA, Lemaitre B. 2021 Repeated subfunctionalization of a modular antimicrobial peptide gene for neural function.
427
(doi:10.1101/2021.02.24.432738)
428
28. Huang J, Lou Y, Liu J, Bulet P, Jiao R, Hoffmann JA, Liegeois S, Li Z, Ferrandon D. 2020 The BaramicinA gene is required at several
429
steps of the host defense against Enterococcus faecalis and Metarhizium robertsii in a septic wound infection model in Dro-
430
sophila melanogaster. (doi:10.1101/2020.11.23.394809)
431
29. Lezi E et al. 2018 An Antimicrobial Peptide and Its Neuronal Receptor Regulate Dendrite Degeneration in Aging and Infection.
432
Neuron 97, 125-138.e5. (doi:10.1016/j.neuron.2017.12.001)
433
30. UniProt Consortium. 2021 UniProt: the universal protein knowledgebase in 2021. Nucleic Acids Res 49, D480–D489.
434
(doi:10.1093/nar/gkaa1100)
435
31. Uttenweiler-Joseph S, Moniatte M, Lagueux M, Van Dorsselaer a, Hoffmann J a, Bulet P. 1998 Differential display of peptides
436
induced during the immune response of Drosophila: a matrix-assisted laser desorption ionization time-of-flight mass spectrom-
437
etry study. Proceedings of the National Academy of Sciences of the United States of America 95, 11342–11347.
438
(doi:10.1073/pnas.95.19.11342)
439
32. Lazzaro BP, Clark AG. 2003 Molecular population genetics of inducible antibacterial peptide genes in Drosophila melanogaster.
440
Mol Biol Evol 20, 914–923. (doi:10.1093/molbev/msg109)
441
33. Levy F, Rabel D, Charlet M, Bulet P, Hoffmann JA, Ehret-Sabatier L. 2004 Peptidomic and proteomic analyses of the systemic
442
immune response of Drosophila. Biochimie 86, 607–616. (doi:10.1016/j.biochi.2004.07.007)
443
34. Kragol G, Lovas S, Varadi G, Condie BA, Hoffmann R, Otvos L. 2001 The antibacterial peptide pyrrhocoricin inhibits the ATPase
444
actions of DnaK and prevents chaperone-assisted protein folding. Biochemistry 40, 3016–3026. (doi:10.1021/bi002656a)
445
35. Otvos L, O I, Rogers ME, Consolvo PJ, Condie BA, Lovas S, Bulet P, Blaszczyk-Thurin M. 2000 Interaction between heat shock
446
proteins and antimicrobial peptides. Biochemistry 39, 14150–14159. (doi:10.1021/bi0012843)
447
36. Rahnamaeian M, Cytryńska M, Zdybicka-Barabas A, Vilcinskas A. 2016 The functional interaction between abaecin and pore-
448
forming peptides indicates a general mechanism of antibacterial potentiation. Peptides 78, 17–23. (doi:10.1016/j.pep-
449
tides.2016.01.016)
450
37. Bulet P, Dimarcq JL, Hetru C, Lagueux M, Charlet M, Hegy G, Van Dorsselaer A, Hoffmann JA. 1993 A novel inducible antibacterial
451
peptide of Drosophila carries an O-glycosylated substitution. J Biol Chem 268, 14893–14897.
452
38. Sanchez Bosch P et al. 2019 Adult Drosophila Lack Hematopoiesis but Rely on a Blood Cell Reservoir at the Respiratory Epithelia
453
to
Relay
Infection
Signals
to
Surrounding
Tissues.
Developmental
Cell
51,
1–17.
454
(doi:https://doi.org/10.1016/j.devcel.2019.10.017)
455
39. Marra A, Hanson MA, Kondo S, Erkosar B, Lemaitre B. 2021 Drosophila antimicrobial peptides and lysozymes regulate gut mi-
456
crobiota composition and abundance. (doi:10.1101/2021.03.19.436153)
457
40. Ferreira ÁG, Naylor H, Esteves SS, Pais IS, Martins NE, Teixeira L. 2014 The Toll-Dorsal Pathway Is Required for Resistance to Viral
458
Oral Infection in Drosophila. PLoS Pathogens 10. (doi:10.1371/journal.ppat.1004507)
459
41. Chapman JR, Hill T, Unckless RL. 2019 Balancing selection drives maintenance of genetic variation in Drosophila antimicrobial
460
peptides. Genome Biology and Evolution 11, 2691–2701. (doi:https://doi.org/10.1093/gbe/evz191)
461
42. Unckless RL, Lazzaro BP. 2016 The potential for adaptive maintenance of diversity in insect antimicrobial peptides. Philosophical
462
Transactions of the Royal Society B: Biological Sciences 371, 20150291. (doi:10.1098/rstb.2015.0291)
463
43. Juneja P, Lazzaro BP. 2009 Providencia sneebia sp. nov. and Providencia burhodogranariea sp. nov., isolated from wild Drosoph-
464
ila melanogaster. International Journal of Systematic and Evolutionary Microbiology 59, 1108–1111. (doi:10.1099/ijs.0.000117-
465
0)
466
44. Karlsson C, Korayem AM, Scherfer C, Loseva O, Dushay MS, Theopold U. 2004 Proteomic Analysis of the Drosophila Larval He-
467
molymph Clot. Journal of Biological Chemistry 279, 52033–52041. (doi:10.1074/jbc.M408220200)
468
45. Hedengren M, \AAsling B, Dushay MS, Ando I, Ekengren S, Wihlborg M, Hultmark D. 1999 Relish, a central factor in the control
469
of humoral but not cellular immunity in Drosophila. Molecular Cell 4, 827–837. (doi:10.1016/S1097-2765(00)80392-5)
470
16
46. Waumans Y, Baerts L, Kehoe K, Lambeir A-M, De Meester I. 2015 The Dipeptidyl Peptidase Family, Prolyl Oligopeptidase, and
471
Prolyl Carboxypeptidase in the Immune System and Inflammatory Disease, Including Atherosclerosis. Front. Immunol. 6.
472
(doi:10.3389/fimmu.2015.00387)
473
47. Hanson MA, Hamilton PT, Perlman SJ. 2016 Immune genes and divergent antimicrobial peptides in flies of the subgenus Dro-
474
sophila. BMC evolutionary biology 16, 228. (doi:10.1186/s12862-016-0805-y)
475
48. Hanson MA, Lemaitre B, Unckless RL. 2019 Dynamic Evolution of Antimicrobial Peptides Underscores Trade-Offs Between Im-
476
munity and Ecological Fitness. Frontiers in Immunology 10, 2620. (doi:doi: 10.3389/fimmu.2019.02620)
477
49. Lin Y, Cai Y, Liu J, Lin C, Liu X. 2019 An advanced approach to identify antimicrobial peptides and their function types for penaeus
478
through machine learning strategies. BMC Bioinformatics 20, 291. (doi:10.1186/s12859-019-2766-9)
479
50. Li W, Tailhades J, O’Brien-Simpson NM, Separovic F, Otvos L, Hossain MA, Wade JD. 2014 Proline-rich antimicrobial peptides:
480
Potential therapeutics against antibiotic-resistant bacteria. (doi:10.1007/s00726-014-1820-1)
481
51. Jumper J et al. 2021 Highly accurate protein structure prediction with AlphaFold. Nature (doi:10.1038/s41586-021-03819-2)
482
52. Wiegand I, Hilpert K, Hancock REW. 2008 Agar and broth dilution methods to determine the minimal inhibitory concentration
483
(MIC) of antimicrobial substances. Nature Protocols (doi:10.1038/nprot.2007.521)
484
53. Zdybicka-Barabas A, Mak P, Klys A, Skrzypiec K, Mendyk E, Fiołka MJ, Cytryńska M. 2012 Synergistic action of Galleria mellonella
485
anionic peptide 2 and lysozyme against Gram-negative bacteria. Biochim Biophys Acta 1818, 2623–2635.
486
(doi:10.1016/j.bbamem.2012.06.008)
487
54. Peng S, Yang M, Sun RN, Liu Y, Wang W, Xi Q, Gong H, Chen C. 2018 Mechanism of actions of Oncocin, a proline-rich antimicrobial
488
peptide, in early elongation revealed by single-molecule FRET. (doi:10.1007/s13238-017-0495-2)
489
55. Yu G, Baeder DY, Regoes RR, Rolff J. 2016 The More The Better? Combination Effects of Antimicrobial Peptides. Antimicrobial
490
Agents and Chemotherapy , AAC.02434-15. (doi:10.1128/AAC.02434-15)
491
56. Belanger CR, Lee AH-Y, Pletzer D, Dhillon BK, Falsafi R, Hancock REW. 2020 Identification of novel targets of azithromycin activity
492
against Pseudomonas aeruginosa grown in physiologically relevant media. Proc Natl Acad Sci USA 117, 33519–33529.
493
(doi:10.1073/pnas.2007626117)
494
57. Zasloff M. 1987 Magainins, a class of antimicrobial peptides from Xenopus skin: isolation, characterization of two active forms,
495
and partial cDNA sequence of a precursor. Proc Natl Acad Sci U S A 84, 5449–5453. (doi:10.1073/pnas.84.15.5449)
496
58. Mackay TFC et al. 2012 The Drosophila melanogaster Genetic Reference Panel. Nature 482, 173–8. (doi:10.1038/nature10811)
497
59. Hellgren O, Sheldon BC, Buckling A. 2010 In vitro tests of natural allelic variation of innate immune genes (avian ??-defensins)
498
reveal functional differences in microbial inhibition. Journal of Evolutionary Biology 23, 2726–2730. (doi:10.1111/j.1420-
499
9101.2010.02115.x)
500
60. Hellgren O, Sheldon BC. 2011 Locus-specific protocol for nine different innate immune genes (antimicrobial peptides: β-defen-
501
sins) across passerine bird species reveals within-species coding variation and a case of trans-species polymorphisms. Molecular
502
Ecology Resources 11, 686–692. (doi:10.1111/j.1755-0998.2011.02995.x)
503
61. Hollox EJ, Armour JAL. 2008 Directional and balancing selection in human beta-defensins. BMC Evolutionary Biology 8.
504
(doi:10.1186/1471-2148-8-113)
505
62. Fürst M a., McMahon DP, Osborne JL, Paxton RJ, Brown MJF. 2014 Disease associations between honeybees and bumblebees
506
as a threat to wild pollinators. Nature 506, 364–366. (doi:10.1038/nature12977)
507
63. Marcinkiewicz M, Majewski S. 2016 The role of antimicrobial peptides in chronic inflammatory skin diseases.
508
(doi:10.5114/pdia.2015.48066)
509
64. Bergsson G, Reeves EP, McNally P, Chotirmall SH, Greene CM, Greally P, Murphy P, O’Neill SJ, McElvaney NG. 2009 LL-37 Com-
510
plexation with Glycosaminoglycans in Cystic Fibrosis Lungs Inhibits Antimicrobial Activity, Which Can Be Restored by Hypertonic
511
Saline. The Journal of Immunology 183, 543–551. (doi:10.4049/jimmunol.0803959)
512
65. Goldman MJ, Anderson GM, Stolzenberg ED, Kari UP, Zasloff M, Wilson JM. 1997 Human β-defensin-1 is a salt-sensitive antibiotic
513
in lung that is inactivated in cystic fibrosis. Cell 88, 553–560. (doi:10.1016/S0092-8674(00)81895-4)
514
17
66. Prado-Montes de Oca E, García-Vargas A, Lozano-Inocencio R, Gallegos-Arreola MP, Sandoval-Ramírez L, Dávalos-Rodríguez NO,
515
Figuera LE. 2007 Association of β-Defensin 1 Single Nucleotide Polymorphisms with Atopic Dermatitis. Int Arch Allergy Immunol
516
142, 211–218. (doi:10.1159/000097023)
517
67. Kondo S, Ueda R. 2013 Highly Improved gene targeting by germline-specific Cas9 expression in Drosophila. Genetics 195, 715–
518
721. (doi:10.1534/genetics.113.156737)
519
68. Parvy J-P, Yu Y, Dostalova A, Kondo S, Kurjan A, Bulet P, Lemaitre B, Vidal M, Cordero J. 2019 The antimicrobial peptide defensin
520
cooperates with tumour necrosis factor to drive tumour cell death in Drosophila. eLife , eLife 2019;8:e45061.
521
(doi:10.7554/eLife.45061)
522
69. Silvestro L, Weiser JN, Axelsen PH. 2000 Antibacterial and antimembrane activities of cecropin A in Escherichia coli. Antimicrob
523
Agents Chemother 44, 602–607. (doi:10.1128/AAC.44.3.602-607.2000)
524
70. Fehlbaum P, Bulet P, Michaut L, Lagueux M, Broekaert WF, Hetru C, Hoffmann JA. 1994 Insect immunity: Septic injury of dro-
525
sophila induces the synthesis of a potent antifungal peptide with sequence homology to plant antifungal peptides. Journal of
526
Biological Chemistry 269, 33159–33163.
527
528
18
Supplementary Figures
529
530
Figure S1: CG10816/Dro mutants are not susceptible to E. coli 1106 or Acetobacter sp. ML04.1 infection. RelE20
531
mutants deficient for Imd signalling and ΔAMP14 flies lacking seven AMP gene families, which includes
532
CG10816/Dro deletion, both succumb to these infections, as found previously [6,23,39].
533
0
1
2
3
4
5
6
7
0
25
50
75
100
E. coli 1106,
OD = 200, 29°C
nexp = 2
0
2
4
6
8
10 12 14
0
25
50
75
100
Acetobacter sp. ML04.1
OD = 150, 25°C
nexp = 2
iso w1118
DroSK3
DroSK4
DroAttSK2
ΔAMP14
RelE20
Percent survival
Time (days)
Legend:
19
534
Figure S2: Alignment of Drosocin proteins encoded by various Drosophila species. Buletin-like C-terminus pep-
535
tides are found in D. pseudoobscura, D. suzukii, and D. melanogaster Dro genes. In D. willistoni and subgenus
536
Drosophila flies, Buletin-like peptides are not found. Full precursor protein sequences are shown for each species.
537
Uniquely the D. neotestacea and D. innubila Dro genes encode multiple Drc peptides in tandem between furin
538
cleavage sites (red boxes at top of alignment) [47]. These furin sites are usually followed by dipeptidyl peptidase
539
sites (yellow boxes at top of alignment), similar to the tandem repeat structure of honeybee Apidaecin and Dro-
540
sophila Baramicin [20,24].
541
20
542
Figure S3: The DroSK3 mutation deletes the Drc peptide N-terminus, but a truncated Drc peptide is still secreted.
543
A) Alignment and annotation of the nucleotide and mature peptide products of CG10816 in wild-type and DroSK3
544
mutant flies. The DroSK3 mutation causes a net deletion of 24 nucleotides, and an additional 4 codons (12 nucleo-
545
tides) are changed. DroSK3 flies produce Buletin, but also a truncated version of Drc lacking critical residues for
546
activity such as the O-glycosylated Threonine (changed to Alanine, orange critical residues). B) MALDI-TOF spectra
547
show that DroSK3 flies have unique peptides corresponding to different versions of the DroSK3 truncated Drc peptide
548
with progressively degraded N-termini (MALDI-TOF reflectron mode). This confirms that a truncated Drc peptide
549
is produced and secreted, though it lacks the critical PRPT motif needed for O-glycosylation. This truncated Drc
550
peptide also lacks dipeptidyl peptidase activity as it is mutated in the CG10816 dipeptidyl peptidase site “TP” à
551
“TH” (yellow/grey annotations in A). As a consequence, it is apparently secreted at full length after the signal
552
peptide, only to be progressively degraded from the N-terminus in the hemolymph.
553
SVASHPRPIRV
1218.4 Da predicted
HSVASHPRPIRV
1355.6 Da predicted
iso w1118
DroSK3
iso w1118
DroSK3
iso w1118
DroSK3
THSVASHPRPIRV
1456.7 Da predicted
relative intensity (%)
A
B
1218.7
1355.7
1456.8
1351.6
21
554
Figure S4: Representative experiment of Cecropin A and Buletin in vitro activity against P. burhodogranariea.
555
Bacteria were mixed with peptide in LB and allowed to grow shaking at room temperature overnight. Every 10
556
minutes, the absorbance at OD600 was recorded. Almost no growth was recorded in the observation period for
557
P. burhodogranariea in the presence of 5-10µM Cecropin A. This result is consistent with a separate experiment
558
where we monitored mixtures for bacterial growth only at the end (not shown), suggesting an MIC of Cecropin A
559
against P. burhodogranariea of ~5µM. We conclude that in these in vitro conditions, Buletin does not impact P.
560
burhodogranariea growth alone or in combination with pore forming Cecropin peptides.
561
0.05
0.15
0.25
0.35
0.45
0.55
0.65
0.75
0.85
0
10
20
30
40
50
60
70
80
90
100
OD600
# of 10 minute cycles
P. burhodogranariea growth in CecA and/or Buletin
10uM CecA
5uM CecA
2.5uM CecA
1.3uM CecA
0.6uM CecA
IM7ala 1mM
IM7thr 1mM
10Cec+IM7ala
5Cec+IM7ala
2.5Cec+IM7ala
1.3Cec+IM7ala
0.6Cec+IM7ala
10Cec+IM7thr
5Cec+IM7thr
2.5Cec+IM7thr
1.3Cec+IM7thr
0.6Cec+IM7thr
IM7ala100uM
IM7thr100uM
(+)
(-)
| 2022 | immunity: The gene encodes two host defence peptides with pathogen-specific roles | 10.1101/2022.04.21.489012 | [
"Hanson M.A.",
"Kondo S.",
"Lemaitre B."
] | creative-commons |
Divergence, gene flow and the origin of leapfrog geographic distributions: The
history of color pattern variation in Phyllobates poison-dart frogs
Running Head: Poison frog leapfrog distribution
Roberto Márquez1,2,*, Tyler P. Linderoth3,†, Daniel Mejía-Vargas2, Rasmus Nielsen3,4,5,
Adolfo Amézquita,2,‡ Marcus R. Kronforst1,‡
1Department of Ecology and Evolution, University of Chicago. Chicago, IL. 60637,
USA.
2Department of Biological Sciences, Universidad de los Andes. A.A. 4976, Bogotá,
D.C., Colombia.
3Department of Integrative Biology and Museum of Vertebrate Zoology, University of
California, Berkeley. Berkeley, CA. 94720, USA.
4Department of Statistics, University of California, Berkeley. Berkeley, CA. 94720,
USA.
5Center for GeoGenetics, University of Copenhagen, Copenhagen 1350, Denmark.
*Corresponding author. Department of Ecology and Evolution, University of
Chicago. 1101 East 57th St. Zoology 206. Chicago, IL. 60637. USA. Ph. 312-709-
8658. Email: rmarquezp@uchicago.edu.
†Current address: Department of Genetics, University of Cambridge, Downing Street,
Cambridge CB2 3EH, UK.
‡ Joint senior authors.
Submitted for consideration as an Original Article.
Márquez et al.
1
1
2
3
4
5
6
7
8
9
10
11
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
Abstract
The geographic distribution of phenotypic variation among closely related populations
is a valuable source of information about the evolutionary processes that generate and
maintain biodiversity. Leapfrog distributions, in which phenotypically similar
populations are disjunctly distributed and separated by one or more phenotypically
distinct populations, represent geographic replicates for the existence of a phenotype,
and are therefore especially informative. These geographic patterns have mostly been
studied from phylogenetic perspectives to understand how common ancestry and
divergent evolution drive their formation. Other processes, such as gene flow between
populations, have not received as much attention. Here we investigate the roles of
divergence and gene flow between populations in the origin and maintenance of a
leapfrog distribution in Phyllobates poison frogs. We found evidence for high levels
of gene flow between neighboring populations but not over long distances, indicating
that gene flow between populations exhibiting the central phenotype may have a
homogenizing effect that maintains their similarity, and that introgression between
“leapfroging” taxa has not played a prominent role as a driver of phenotypic diversity
in Phyllobates. Although phylogenetic analyses suggest that the leapfrog distribution
was formed through independent evolution of the peripheral (i.e. leapfrogging)
populations, the elevated levels of gene flow between geographically close
populations poise alternative scenarios, such as the history of phenotypic change
becoming decoupled from genome-averaged patterns of divergence, which we cannot
rule out. These results highlight the importance of incorporating gene flow between
Márquez et al.
2
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
populations into the study of geographic variation in phenotypes, both as a driver of
phenotypic diversity and as a confounding factor of phylogeographic inferences.
Key Words: Phylogeography, spatial population genetics, convergent evolution,
Dendrobatidae
Márquez et al.
3
57
58
59
60
61
Introduction
Geography has a strong influence on the diversification of closely related lineages,
since it largely mediates the level of gene flow between them (Huxley, 1942; Mayr,
1942). Therefore, studying the geographic distribution of phenotypic and genetic
variation among such lineages can generate valuable insights into the processes that
generate biological diversity. An intriguing pattern of geographic variation is the
“leapfrog” distribution, where phenotypically similar, closely related populations (of
the same or recently diverged species) are disjunctly distributed and separated by
phenotypically different populations to which they are also closely related (Chapman,
1923; Remsen, 1984). Such patterns have been reported in multiple taxa, such as birds
(e.g. Cadena, Cheviron, & Funk, 2010; Chapman, 1923; Norman, Christidis, Joseph,
Slikas, & Alpers, 2002; Remsen, 1984), flowering plants (Matsumura, Yokohama,
Fukuda, & Maki, 2009; Matsumura, Yokoyama, Tateishi, & Maki, 2006), and
butterflies (Brower, 1996; Emsley, 1965; Hovanitz, 1940; Sheppard, Turner, Brown,
Benson, & Singer, 1985). Since leapfrog patterns represent repeated instances of
similar phenotypes in space, they provide a rich opportunity to understand the
processes generating phenotypic geographic variation.
Two main hypotheses have been put forward to explain the origin of leapfrog
distributions (Norman et al., 2002; Remsen, 1984): First, the phenotypically similar,
geographically disjunct populations can owe their resemblance to recent common
ancestry (i.e. they are descendants of an ancestral population with the same
Márquez et al.
4
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
phenotype), and the disjunct range of “leapfrogging” forms is due to biogeographic
processes such as long-range migration or the extinction of geographically
intermediate populations. Second, the distribution of phenotypes may be due to
evolutionary convergence of populations with similar phenotypes, or divergence of
the central (intervening) populations from the ancestral phenotype. Clear phylogenetic
predictions can be drawn from these hypotheses: If phenotypic similarity among the
leapfrogging populations is due solely to recent common ancestry, then such
populations should be more closely related to one another than to geographically close
populations that display the intervening phenotype. If the geographic distribution of
phenotypes is due to convergent or divergent evolution then a correspondence
between phylogeny and phenotypes is not expected. In this case, however, ancestral
state reconstructions can identify whether the central or peripheral populations exhibit
derived (i.e. divergent) phenotypes. Accordingly, efforts to elucidate the evolutionary
mechanisms behind leapfrog distributions have mainly focused on inferring the
phylogenetic relationships among populations and using them to reconstruct the
evolution of the phenotype in question (e.g. Brower, 1996; Cadena et al., 2010;
Shun’Ichi Matsumura et al., 2009; Norman et al., 2002; Quek et al., 2010).
Although a cladogenetic description of population history can reveal a great deal
about the origin of leapfrog distributions, it is unable to capture some important
aspects of the diversification process. Among them is the extent of gene flow between
populations (or its absence), which can play an important role in the formation of
leapfrog distributions. For instance, reduced levels of genetic exchange between
Márquez et al.
5
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
populations with different phenotypes will promote the existence of such differences,
while introgressive hybridization between populations can facilitate phenotypic
similarity between them. Furthermore, if gene flow between geographically close
populations with different phenotypes is pervasive, it can homogenize previous
genetic divergence between these populations, decoupling the history of the
phenotype from genome-wide patterns of divergence (Hines et al., 2011; James,
Arenas-Castro, Groh, Engelstaedter, & Ortiz-Barrientos, 2020), which can complicate
inferences related to the origin of leapfrog distributions.
Here we examine the processes driving the origin of a leapfrog distribution present in
Phyllobates poison-dart frogs. This genus is found from Southern Nicaragua to
Western Colombia, and is composed of five nominal species (Myers, Daly, & Malkin,
1978; Silverstone, 1976): P. vittatus, P. lugubris, and P. aurotaenia, which exhibit a
bright dorsolateral stripe on a dark background, and P. terribilis and P. bicolor, which
display solid bright-yellow dorsal coloration (Fig. 1A). The two latter species exhibit
a leapfrog distribution in Western Colombia, separated by P. aurotaenia: P. bicolor
occurs on the slopes of the Western Andes, in the upper San Juan river basin, P.
aurotaenia in the lowlands along the San Juan and Atrato Drainages and onto the
Pacific coast, and P. terribilis along the Pacific coast south of the San Juan’s mouth
(Fig. 1C).
Early systematic work grouped P. terribilis and bicolor as sister species based on
morphological and ontogenetic characters (Maxson & Myers, 1985; Myers et al.,
Márquez et al.
6
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
1978). Although an early mitochondrial phylogeny supported these relationships
(Widmer, Lötters, & Jungfer, 2000), subsequent work has consistently recovered P.
terribilis and P. bicolor as non-sister taxa (Grant et al., 2006, 2017; Márquez,
Corredor, Galvis, Góez, & Amézquita, 2012; Santos et al., 2009), and even suggested
that P. aurotaenia may actually represent two separate lineages, one sister to P.
bicolor and the other to P. terribilis (Grant et al., 2017; Santos et al., 2009). Although
these studies only included 1-4 samples per Phyllobates species, and were based on
DNA sequences from a small number of markers (1-7 loci), their results are
compatible with convergent evolution giving rise to the leapfrog distribution.
In this study we aim to shed light on the evolutionary genetic and biogeographic
processes involved in the origin of the current geographic distribution of aposematic
coloration in Phyllobates poison frogs. Based on substantially increased sampling
across Colombian populations and thousands of genome-wide markers, we leverage
phylogenetics and spatial population genetics to 1) elucidate the extent of genetic
structure and evolutionary relationships among populations, and 2) evaluate the role
of gene flow between populations in the formation of the leapfrog distribution.
Materials and Methods
To obtain a representative sample of Colombian Phyllobates populations, we
conducted field expeditions to 23 localities throughout the genus’s range (Fig. 2B),
resulting in tissue (i.e. mouth swab, toe-clip, or liver) samples from 108 individuals
Márquez et al.
7
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
(Table S1). In addition, we obtained eight samples of P. vittatus and P. lugubris, (four
samples per species; Table S1) to serve as outgroups in our analyses. Both species are
distributed in Central America, and have been consistently found to be the sister
group of Colombian Phyllobates (Grant et al., 2006, 2017; Santos et al., 2009;
Widmer et al., 2000).
mtDNA Sequencing and analysis
To gain initial insight into the levels of genetic variation and structure among
populations we sequenced fragments of three mtDNA markers: 16S rRNA (16S;
569bp), Cytochrome Oxidase I (COI barcoding fragment; 658bp), and Cytochrome b
(Cytb; 699bp) from 74 individuals. We extracted DNA using either Qiagen DNeasy
spin columns or a salt precipitation protocol (Miller, Dykes, & Polesky, 1988), and
used primers 16Sar and 16Sbr (Palumbi et al., 1991), Chmf4 and Chmr4 (Che et al.,
2012), and CytbDen3-L and CytbDen1-H (Santos & Cannatella, 2011) to amplify the
16S, COI, and Cytb loci, respectively. Thermal cycling protocols consisted of 2 min at
95ºC, 30-35 cycles of 30 sec at 95ºC, 1 min at 45ºC and 1.5 min at 72ºC, and a final 5
min at 72ºC. PCR products were purified with ExoSAP (Affymetrix) and sequenced
in both directions using an ABI 3500 Genetic Analyzer (Applied Biosystems).
Chromatograms were assembled and visually inspected in Geneious R9 (Kearse et al.,
2012) to produce finalized consensus sequences.
Márquez et al.
8
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
We aligned our sequences and those available in GenBank (Table S1) using
MUSCLE (Edgar, 2004), and built mtDNA trees with PhyML 3.3 (Guindon et al.,
2010) and MrBayes 3.2.6 (Altekar, Dwarkadas, Huelsenbeck, & Ronquist, 2004;
Ronquist et al., 2012). MrBayes analyses consisted of 10 million iterations (two runs
with four chains each), sampling every 1,000 iterations, and discarding the first 2,500
trees (25%) as burnin. PhyML runs started from five different random trees, and used
SPR moves to search the tree space. Nodal support was evaluated using aBayes scores
(Anisimova, Gil, Dufayard, Dessimoz, & Gascuel, 2011). To obtain an estimate of
divergence times between mtDNA haplotypes, we inferred a time-calibrated tree
using BEAST v. 2.5.0. (Bouckaert et al., 2019). Based on results from previous work
(Santos et al., 2014), we set a log-normal prior with mean 8.13 million years (MY)
and standard deviation 1.2 MY (i.e. log(mean) = 2.12, log(s.d.) = 0.1) for the root age
of Phyllobates. We used a Calibrated Yule tree prior, and set default priors for all
other parameters, except for the clock rate mean and the Yule birth rate, which were
set to gamma(0.01, 1000). We ran the MCMC sampler for 100 million iterations,
sampling every 10,000, and generated a maximum clade credibility (MCC) tree using
Tree Annotator (distributed with BEAST) after discarding the first 5% of trees as
burnin. Mixing and stationarity of BEAST and MrBayes runs were evaluated visually
and based on effective sample sizes (ESS) using Tracer v. 1.5 (Rambaut &
Drummond, 2009). All mtDNA analyses were performed under partitioning schemes
and molecular evolution models chosen with PartitionFinder2 (Lanfear, Frandsen,
Wright, Senfeld, & Calcott, 2017).
Márquez et al.
9
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
Transcriptome-enabled exon capture
Based on the results of mtDNA analyses we chose 63 samples (60 ingroup, 3
outgroup) from 17 localities (Table S1) representing the range of observed mtDNA
variation among Colombian populations, and used them to perform transcriptome-
enabled exon capture (Bi et al., 2012; Hodges et al., 2007). Briefly, we designed a set
of DNA capture probes based on a transcriptome assembly and used them to enrich
sequencing libraries for a subset of the genome.
Transcriptome sequencing. We generated a transcriptome assembly from liver,
muscle, skin, and heart tissue of a single P. bicolor juvenile (NCBI BioSample
SAMN15546883). RNA was extracted using Qiagen RNeasy spin columns, and
pooled in equimolar ratios by tissue type to build a single cDNA library, which was
sequenced on an Illumina HiSeq 2000. We filtered and trimmed reads using
Trimmomatic v. 0.25 (Bolger, Lohse, & Usadel, 2014), and used Trinity (release
2013-02-25; Grabherr et al., 2011) to assemble them under default parameters, except
for the minimum contig length, which was increased to 250bp. Finally we collapsed
redundant contigs (e.g. alternative isoforoms) with CD-HIT-EST V.4.5.3 (Fu, Niu,
Zhu, Wu, & Li, 2012).
Enrichment probe design. We annotated our transcriptome using BLASTX (Altschul,
1997) against Xenopus tropicalis proteins (JGI 4.2.72), and used Exonerate (Slater &
Birney, 2005) to identify intron-exon boundaries in order to split transcripts into
Márquez et al.
10
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
individual exons. We then chose a final set of exons to enrich in the following way:
First we discarded those under 100bp, with GC content below 40% and above 70%, or
which overlapped by more than 10bp based on Exonerate annotations. Next, we
identified putatively repetitive elements and RNA-coding genes (e.g. rRNAs) in our
transcriptome assembly with RepeatMasker v. 4.0 (Smith, Hubley, & Green, 2013)
and BLASTn, respectively, and removed exons overlapping them. Finally, we blasted
our exon set against itself with BLASTn under default parameters, and whenever two
or more exons matched each other (e-value < 10-10), we retained only one of them.
This resulted in 38,888 exons (7.57Mb) that passed filters, which were used to design
1,943,120 100bp probes that were printed on two Agilent SureSelect custom 1M-
feature microarrays (3bp tiling).
DNA library preparation, target enrichment, and sequencing. We extracted DNA as
described above, and used a Diagenode Bioruptor to shear each extraction to a ~100-
500bp fragment distribution by performing 3-4 rounds of sonication (7min of 30s
on/off cycles per round). DNA libraries were built following Meyer & Kircher (2010),
except for bead cleanups, where we used a 1.6:1 ratio of beads to library (1.8:1 is
recommended) to obtain a slightly larger final fragment size distribution. Finished
libraries were combined in equimolar ratios into two 22.5 μg pools (one per array) for
target enrichment. Array hybridization was performed largely following Hodges et al.
(2009) with minor modifications: Each library pool was mixed with xGen Universal
P5 and P7 blocking oligonucleotides and a mixture of chicken, human, and mouse
COT-1 DNA. The two capture eluates were amplified separately by 18 cycles of PCR.
Márquez et al.
11
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
To reduce the propagation of PCR-induced errors, each eluate was amplified in four
parallel reactions. PCR products were pooled so that both captures were equally
represented, and sequenced on Ilumina HiSeq 2500 and 4000 machines.
Bioinformatic pipeline. De-multiplexed read files were filtered by collapsing PCR-
duplicate reads with SuperDeduper (Petersen, Streett, Gerritsen, Hunter, & Settles,
2015), trimming low quality bases and removing adapter contamination with
Trimmonatic (Bolger et al., 2014) and Skewer (Jiang, Lei, Ding, & Zhu, 2014) under
default parameters, except for the minimum read length, which was increased to 36
bp, and merging overlapping read pairs with FLASH (Magoč & Salzberg, 2011). To
generate a reference for read mapping, we combined all cleaned reads from the
ingroup species (i.e. P. terribilis, aurotaenia, and bicolor), and generated six de novo
assemblies with different kmer sizes (k = 21, 31, 41, 51, 61, and 71) using ABySS (J.
T. Simpson et al., 2009). We then merged the six assemblies using CD-HIT-EST and
Cap3 (Huang & Madan, 1999). Finally, we identified contigs that matched our target
exons using BLASTn, and retained only these for further analyses.
Reads from each sample were mapped to the reference using Bowtie2 v. 2.1.0
(Langmead & Salzberg, 2012), and outputs were sorted with Samtools v. 1.0 (Li et al.,
2009), de-duplicated with Picard v.1.8.4 (http://broadinstitute.github.io/picard), and
re-aligned around indels with GATK v. 3.3.0 (McKenna et al., 2010). We filtered our
data in the following ways: First, we performed a reciprocal blast using the methods
described above and removed any contigs with more than one match (e-value<10-10).
Márquez et al.
12
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
Second, we used ngsParalog (https://github.com/tplinderoth/ngsParalog) to identify
contigs with variants stemming from read mismapping due to paralogy and/or
incorrect assembly. This program uses allele frequencies to calculate a likelihood ratio
for whether the reads covering a site are derived from more than one locus in the
genome, while incorporating the uncertainty inherent in NGS genotyping. We
calculated p-values for these likelihood ratios based on a 50:50 mixed χ2 distribution
with one and zero degrees of freedom under the null, and removed any contigs with
significantly paralogous sites after Bonferronni correction (α = 0.05). Third, we
restricted all analyses to contigs covered by at least one read in at least 20 individuals,
bases with quality above 30, and read pairs mapping uniquely to the same contig (i.e.
proper pairs) with mapping quality above 20. Finally, we removed samples with less
than 2.5 million sites covered by at least one read after filtering. This resulted in a
dataset of 32,516 contigs (12.95 Mb) and 57 samples, which were used in all
downstream analyses.
Population Structure
To characterize genome-wide patterns of population differentiation we used our exon
capture dataset to perform Principal Component Analysis (PCA) of genetic
covariances calculated in PCangsd v.0.94 (Meisner & Albrechtsen, 2018), to estimate
admixture proportions (k = 2-9) in ngsAdmix v.32 (Skotte, Korneliussen, &
Albrechtsen, 2013), and to build a minimum-evolution tree in FastME v.2.1.5 (Lefort,
Desper, & Gascuel, 2015) using genetic distances estimated with ngsDist (Vieira,
Márquez et al.
13
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
Lassalle, Korneliussen, & Fumagalli, 2016). Nodal support for this tree was evaluated
using 500 bootstrapped distance matrices produced in ngsDist by sampling blocks of
10 SNPs. These three analyses used genotype likelihoods (GL) as input, which were
estimated in Angsd v.0.9.18 (Korneliussen, Albrechtsen, & Nielsen, 2014) at sites
covered by at least one read in at least 50% of the samples without filtering for
linkage disequilibrium. PCA and ngsAdmix analyses used one site per contig,
randomly chosen among those with minor allele frequencies above 0.05 that passed
the programs’ internal quality filters (5,634 sites). Genetic distance estimation for the
ME tree was restricted to variable sites (i.e. SNP p-value < 0.05; 84,218 sites). PCA
and ngsAdmix were run only on Colombian samples while the ME tree also included
outgroups.
Finally, we reconstructed a population graph to evaluate historical splits and mixtures
between sampling localities using Treemix (Pickrell & Pritchard, 2012). We called
genotypes using the HaplotypeCaller and GenotypeGVCFs tools of GATK v.3.3.0
under default parameters, except for the heterozygozity prior, minimum base quality,
and minimum variant-calling confidence, which were increased to 0.005, 30, and 20,
respectively, to accommodate for the multi-species nature of our dataset. We then
obtained allele counts for biallelic SNPs that were at least 1kb apart within each
contig (usually resulting in a single SNP per contig, since most contigs were under
1kb), and with at least 50% genotyping (20,275 SNPs), using Plink v.1.90 (Purcell et
al., 2007). In two cases, two nearby populations of the same color pattern (16.4 and
19.7 Km apart; Fig S1), which clustered closely in all other genetic structure analyses,
Márquez et al.
14
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
were merged into single demes for allele count estimation due to small sample sizes.
In addition, since we only had exon capture data for one P. vittatus individual, only P.
lugubris was used as outgroup in this analysis. We ran Treemix v.1.13 assuming m =
0–6 migration edges, and chose the optimal number of migration edges by performing
likelihood ratio tests in which we compared each value of m to the one immediately
smaller. P-values were calculated based on a χ2 distribution with two degrees of
freedom, since adding an extra edge adds two parameters (weight and direction of
migration) to the model. This approach recovered m=2 as the most likely scenario
(Table S2); results for m=0-6 are presented in Fig. S2.
Phylogenetic relationships between lineages
To reconstruct the phylogenetic relationships between Phyllobates lineages, we
inferred a species tree under the multispecies coalescent model, assuming independent
sites, as implemented in SNAPP (Bryant, Bouckaert, Felsenstein, Rosenberg, &
RoyChoudhury, 2012). SNAPP requires individuals to be assigned to operational
taxonomic units (OTUs) a priori. Given our small sample sizes for some localities, as
well as the evidence of gene flow between localities (see Results section), we took an
ad-hoc approach and grouped our sampling localities into eight geographically and
phenotypically coherent groups that showed evidence of being genetically distinct
entities (see locality colors in Fig. 2B). Briefly, each OTU contained individuals that
were geographically close, displayed the same color pattern, and showed evidence of
genetic clustering in population structure analyses. We did not require OTUs to be
Márquez et al.
15
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
fully reproductively isolated from each other. Further details on our OTU selection
criteria can be found in the online supplement.
For computational efficiency, SNAPP was run on a reduced version of the Treemix
dataset described above, restricted to SNPs genotyped for at least 75% of individuals
and at least one member of each OTU (5,938 SNPs). We ran the MCMC sampler
under default priors for 1,000,000 iterations, sampling every 250, and discarded the
first 150,000 as burnin. Stationarity and mixing were evaluated in Tracer (Rambaut &
Drummond, 2009) as detailed above, and the posterior tree distribution was
summarized as a maximum clade credibility (MCC) tree in TreeAnnotator. To obtain
estimates of divergence times between OTUs, we assumed a mutation rate of μ = 1e-9
mutations per site per year (Crawford, 2003; Sun et al., 2015), and a generation time
of one year (Phyllobates frogs are sexually mature at ~10-18 months after hatching;
Myers et al., 1978; R. Márquez pers obs.), and converted branch lengths to time units
as T = (τg/μ), where T is the divergence time in years, τ the branch length in
coalescent units, g the generation time, and μ the mutation rate (Bryant et al., 2012).
Phylogenetic Comparative Analyses
In order to evaluate whether the central or leapfrogging populations exhibit a derived
color pattern, we performed ancestral state reconstruction along the SNAPP MCC tree
using maximum parsimony (Fitch, 1971) in the R package phangorn (Schliep, 2011).
Aposematic coloration has been shown to co-evolve with several other traits, such as
Márquez et al.
16
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
body size, toxicity, and diet specialization in dendrobatid frogs (Pough & Taigen,
1990; Santos & Cannatella, 2011; Summers & Clough, 2001). Understanding
correlations between these traits within Phyllobates can shed light on how predation
pressures have affected the geographic distribution of color patterns in this group.
Therefore, we investigated the extent of correlated evolution between color pattern,
body size, and toxicity. We used the snout-to-vent length (SVL) as a proxy for body
size, and the average amount of batrachotoxin (BTX) in a frog’s skin as a proxy for
toxicity. BTX is the most abundant and toxic alkaloid found in Phyllobates skins
(Märki & Witkop, 1963; Myers et al., 1978). BTX levels were obtained from Table 2
of Daly, Myers, & Whittaker (1987), and SVL was measured from specimens in
natural history collections (193 specimens; Table S3). We used mean SVL values for
each lineage in analyses, and log-transformed BTX levels to attain normality of
residuals. Correlations between traits were evaluated using phylogenetic generalized
least squares regression (pGLS; Grafen, 1989; Martins & Hansen, 1997) with either
Brownian motion (Felsenstein, 1985), Lambda (Pagel, 1999), or Ornstein–Uhlenbeck
(Martins & Hansen, 1997) correlation structures. The best correlation structure was
chosen by performing pGLS with the three correlation structures and comparing the
fit of each model based on the AIC. Correlation structures were generated using the R
package ape (Paradis, Claude, & Strimmer, 2004), and regressions were performed in
the nlme package (Pinheiro, Bates, DebRoy, & Sarkar, 2017). In addition to the
highest clade credibility tree, we also conducted tests of phylogenetic correlations on
1,000 randomly selected trees from the post-burnin SNAPP posterior distribution to
account for phylogenetic uncertainty.
Márquez et al.
17
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
Spatial population genetics
We took a spatial population genetics approach to investigate the extent of divergence
and gene flow between populations in a spatially explicit way, aimed at understanding
the nature and drivers of genetic variation across the landscape. First we generated a
geo-genetic map of the Colombian Phyllobates populations using SpaceMix
(Bradburd, Ralph, & Coop, 2016). This consists of a bidimensional plot where the
distances between populations correspond to their expected geographic distances
under stationary isolation by distance (IBD). Differences between geographic and
geo-genetic locations therefore reflect historical rates of gene flow across the
landscape. Populations that exchange more alleles than expected under stationary IBD
are closer in geo-genetic than geographic space, and vice versa. For example,
populations separated by a topographic barrier will be further apart in geo-genetic
than geographic space. As input for SpaceMix we used allele counts generated as
detailed above (see Population Structure section), for sites that were variable among
Colombian individuals (8,093 sites). We then parameterized the full (“source_and
target”) SpaceMix model with an MCMC run comprised of 10 initial exploratory
chains (500,000 iterations each), followed by a 500,000,000 iteration “long” run,
which was sampled every 10,000 iterations. We used default prior settings, and
centered spatial (i.e. location) priors for each population at their sampling location.
Márquez et al.
18
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
For the two demes composed of individuals from nearby localities we used the
midpoint of the segment connecting both localities (Fig. S1).
SpaceMix accounts for the fact that a fraction of a population’s alleles may have been
acquired through recent long-range migration from another region of the map by
incorporating a long-range admixture proportion parameter, labelled w, which
represents the probability that an allele in a given population migrated recently from a
distant region. Since leapfrog distributions can, in principle, arise through
introgressive hybridization between disjunct populations, we evaluated the support of
our data for models with and without long-distance gene flow between populations
(ie. all w parameters set to zero vs. w allowed to vary). We did so by estimating the
Bayes Factor (Kass & Raftery, 1995) between both models using the Savage-Dickey
density ratio (Dickey & Lientz, 1970), which approximates the Bayes Factor between
nested models. Further details on this estimator and our implementation for SpaceMix
models can be found in the online supplement.
Next, we used EEMS (Petkova, Novembre, & Stephens, 2015) to identify areas of the
landscape where gene flow between populations is especially prevalent or reduced.
Briefly, this algorithm estimates the rate at which genetic similarity decays with
distance (i.e. the effective migration). Regions where this decay is quick or slow can
be interpreted as barriers or corridors of migration, respectively. We estimated mean
squared genetic differences between samples from genotype likelihoods in ATLAS
(Link et al., 2017), and used them as input for EEMS. We set the number of demes to
Márquez et al.
19
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
500, and averaged across 10 independent 10,000,000-step MCMC runs logged every
1,000 steps (20% burnin). Since our genetic dissimilarity matrix was inferred from
genotype likelihoods, specifying the number of SNPs used to compute the matrix
(required by EEMS) was not straightforward. We used the number of sites with SNP
p-value below 0.01, as calculated with Angsd (221,825 sites).
Finally, we assessed how attributes of the landscape influence genetic divergence
between populations. Based on the results of EEMS and SpaceMix analyses, we
evaluated the effect of three landscape features on genetic divergence: geographic
distance, differences in elevation, and the presence of the San Juan River as a
potential corridor of gene flow. To do so, we used the multiple matrix regression with
randomization (MMRR) approach proposed by Wang (2013), which is an extension
of multiple linear regression for distance matrices.
Our regression model consisted of genetic distance as a response variable and
geographic distance, difference in elevation, and the effect of the San Juan river as a
dispersal corridor as explanatory variables. As a proxy for genetic distance, we used
the linearized genome-wide weighted FST (FST/[1-FST]; J. Reynolds, Weir, &
Cockerham, 1983; Weir & Cockerham, 1984), estimated using Angsd based on 2D-
site frequency spectra (SFS). To maximize the number of sites used to estimate each
SFS, we included contigs with data for less than 20 individuals (but that passed all
other filters) in this analysis. We estimated geodesic distances among populations
based on GPS coordinates taken in the field using the pointDistance() function of the
Márquez et al.
20
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
raster R package (Hijmans, 2017), and calculated elevation differences based on
measurements taken in the field or extracted from Google Earth. To generate a proxy
for the San Juan river as a dispersal corridor we built a resistance layer where every
pixel overlapping the San Juan river had a value of 1 and all others had a resistance
value R, which made movement between two pixels along the San Juan R times more
likely than between two pixels outside the river. We then used this layer to calculate
least cost distances between populations with the costDistance() function of the
gdistance R package (Dijkstra, 1959; van Etten, 2017). Finally, we regressed the least
cost distance against the geodesic distance, and saved the model residuals as a
measure of the component of the resistance distance not explained by geographic
distance. These residuals were used as an explanatory variable in our model. Since
setting a biologically realistic value for R was not straightforward, we performed five
separate MMRR analyses using least-cost distances estimated with R = 2, 10, 20, 50,
and 100. The MMRR analysis was run using the script archived by Wang (2013;
https://doi.org/10.5061/dryad.kt71r) with 10,000 permutations to estimate p-values.
Results
Population structure among Colombian Phyllobates
As expected from a multi-species dataset, we found multiple genetically structured
clusters of individuals, which were largely concordant across analyses of the exon-
Márquez et al.
21
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
enrichment and mtDNA datasets (Fig. 2, Fig. S2-S3). However, these clusters align
much more closely with geography than either coloration or the current taxonomy: All
populations of P. terribilis grouped with the southern populations of P. aurotaenia,
while the northeastern populations of P. aurotaenia clustered closely with the
northern populations of P. bicolor. The southern P. bicolor and the P. aurotaenia
populations east and west of the Baudó mountains also formed independent clusters,
but their relationship to other lineages was less clear. Finally, two mtDNA sequences
from captive-bred P. aurotaenia of unknown origin (sequenced by Grant et al. [2006]
and Santos et al. [2009]) were sister to those from the southern populations of P.
bicolor in our genealogy (Fig. 2A). These results highlight the existence of several
previously unrecognized (i.e. cryptic) lineages. Notably, they reveal the existence of
three independent solid-yellow lineages, instead of two as previously thought, since
the populations currently classified as P. bicolor clustered as two clearly separate and
independent lineages. This points to an even greater discordance between coloration
phenotypes and genetic similarity than previously thought.
Phylogenetic relationships and divergence times.
The inferred species tree was generally consistent with our genetic structure results,
since tree topologies largely mirrored geography: Most OTUs were sister to close
geographic neighbors, and higher level relationships followed a north-south axis (Fig.
2-3). In addition, the three yellow lineages were recovered each as sister to a different
striped lineage. The topology of the SNAPP tree was largely concordant with those
Márquez et al.
22
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
obtained in mtDNA and Treemix analyses. We only found inconsistencies in the
placement of the P. aurotaenia populations from the eastern and western flanks of the
Serranía del Baudó: Mitochondrial haplotypes from these two populations were part
of a closely-related clade that also included all P. aurotaenia sequences from the
Atrato river, and this clade was sister to another one containing sequences from the
northern P. bicolor and the San Juan P. aurotaenia (Fig. 2A). Treemix also recovered
the eastern and western Baudó populations as sister taxa, but they were sister to the
rest of the Colombian populations (Fig. 2D and S2). Finally SNAPP recovered only
the western Baudó P. aurotaenia as sister to all other Colombian populations, while
the eastern Baudó P. aurotaenia was sister to the southern P. bicolor (Fig. 3).
Treemix inferred a migration edge from the base of the clade containing the
populations of P. bicolor and P. aurotaenia from the San Juan and Atrato drainages
into the eastern Baudó P. aurotaenia (Fig. 2D). Since Treemix reconciles instances
where a bifurcating tree model, such as the one used by SNAPP, does not fit the data
well by incorporating migration edges between branches of the tree, this result
suggests that these differences may be due to gene flow among populations.
Divergence time estimation based on the SNAPP tree revealed a Plio-Pleistocene
diversification of Phyllobates, and were generally concordant with previous estimates
(Santos et al., 2009, 2014), indicating that our mutation rate and generation time
assumptions are reasonable. The most recent common ancestor (MRCA) of
Phyllobates was placed at 5.1 million years ago (MYA), with subsequent cladogenesis
events from the late Pliocene to the Pleistocene (2.9-0.6 MYA; Fig. 3). These
Márquez et al.
23
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
divergence times were slightly older but within the 95% HPD intervals of those
estimated from mtDNA sequences (Fig. S4).
Comparative analyses
Ancestral state reconstructions found the striped phenotype to be ancestral to solid-
yellow (Fig. 4A). Phylogenetic regressions revealed a strong relationship between
color pattern and size, with solid-yellow lineages being significantly larger than
striped ones (Brownian Motion: β = 11.95, t = 9.92, df = 10, p = 9.03e-6; Fig. 4A),
but a much weaker relationship between coloration and toxicity (Ornstein–Uhlenbeck:
β = 1.19, t = 2.48, df = 5, p = 0.089; Fig. 4B). These results, suggest that at least two
co-evolving traits (solid yellow coloration and larger size), possibly related to
predator avoidance, are distributed in a leapfrog fashion in Phyllobates. Regressions
performed over a set of posterior trees instead of the summary tree resulted in effect
sizes and p-values centered around and qualitatively equivalent to those estimated
using the summary tree, showing that the above conclusions are robust to the
phylogenetic uncertainty present in our species tree reconstruction (Fig. S5).
Spatial Population Genetics
The effective migration surface estimated by EEMS revealed a corridor of migration
that matches the course of the San Juan river to a remarkable degree, considering that
this method is completely agnostic to the topography of the landscape (Fig. 6A). This
Márquez et al.
24
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
close match appears to lend strong support to to a high historical migration rate along
the San Juan. However, we note that this result should be interpreted considering the
sampling gap in the lower San Juan (See Fig. 5A). This corridor connects most of the
sampled P. aurotaenia populations and the northern P. bicolor, and could explain the
discordance between mtDNA and exon capture datasets in the phylogenetic placement
of P. aurotaenia populations from the Eastern and Western Baudó mountains.
Concordantly, SpaceMix estimated geo-genetic locations of populations along the San
Juan corridor that were much closer to one another than their actual geographic
positions: P. aurotaenia populations from the upper San Juan and Atrato drainages
and the northern P. bicolor converged to very close locations in the upper/mid San
Juan, overlapping considerably. The Baudó (east and west) and southern populations
of P. aurotaenia were also shifted towards this area, but to a lesser extent (Fig. 5B).
In addition EEMS estimated very low levels of migration in the area enclosing the
two southern P. bicolor populations, suggesting the existence of barriers to gene flow
around these populations. Interestingly, the geo-genetic location of these populations
was inferred north of its geographic location, past the mid San Juan cluster, and only
slightly overlapping with other populations (Fig. 5B). The estimated long-distance
admixture proportions were minimal for all populations (Fig. 5C), and the model with
these proportions fixed at 0 was overwhelmingly supported over one where they were
allowed to vary (Bayes factor = 1748).
Márquez et al.
25
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
In agreement with EEMS and SpaceMix results, the MMRR analysis found
significant effects of geographic distance, elevation differences, and the San Juan as a
migration corridor on genetic divergence between localities. Across the range of
resistance values used, geographic distance was the strongest predictor. However,
elevation differences and the San Juan as a barrier still had appreciable effects on
genetic divergence (Table 1).
Discussion
Our main goal in this study was to understand the evolutionary and biogeographic
processes that have shaped the leapfrog distribution of color pattern among
Phyllobates populations, focusing on the roles of genetic divergence and gene flow.
We found patterns of genetic structure and phylogenetic affinity between populations
that closely match geography, evidence for gene flow between geographically close
populations, especially along the San Juan river, and evidence against gene flow
between distant populations.
These results provide strong evidence against the hypothesis that introgression of
color pattern alleles between disjunct populations has played a role in generating the
geographic distribution of this trait. Instead, they suggest an important role for short-
range gene flow between neighboring populations. The high level of migration among
the central striped populations along the San Juan river suggests that allele movement
Márquez et al.
26
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
between these populations may have a homogenizing effect that maintains their
phenotypic similarity. In addition, we find evidence for a barrier to gene flow that
encloses the two sampled populations of the southern P. bicolor lineage, probably
associated with differences in elevation, which could be helping maintain the
phenotypic distinctiveness of this population. Conversely, the northern P. bicolor
populations showed a strong signature of gene flow with their neighboring striped
populations, suggesting that other forces, possibly selection, are maintaining the
phenotypic differences between these populations in the face of recurrent gene flow.
Nevertheless, to fully reject or accept these hypotheses, the history the alleles
underlying color pattern differences must be taken into account (Hines et al., 2011).
Our phylogenetic reconstructions are consistent with a scenario in which the disjunct
solid-yellow populations evolved their color patterns independently. However, the
high levels of gene flow between geographically proximal populations and the close
correspondence between phylogeny and geography lead us to suspect that, at least to
an extent, the recovered phylogenetic relationships could be a product of prevalent
gene flow between neighboring populations, and therefore may not reflect the history
of color pattern evolution. A recent simulation study showed that even moderate
levels of gene flow between geographic neighbors can confound phylogenetic
inferences of convergent evolution (James et al., 2020). This scenario seems
especially likely in the case of the northern populations of P. bicolor, given the
signature of gene flow with their nearby striped populations (e.g. the San Juan and
Atrato P. aurotaenia), but less so in the case of the southern P. bicolor, considering
Márquez et al.
27
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
the strong barriers to gene flow inferred around the populations of this lineage. For P.
terribilis we cannot favor either scenario, since we did not find strong evidence for or
against gene flow with its sister P. aurotaenia populations.
It is therefore plausible that convergent evolution and common ancestry have both
played a role in the origin of this leapfrog distribution. Teasing apart these two
scenarios is challenging if gene flow between neighboring populations is prevalent,
since high levels of genetic exchange between geographically close populations can
erode existing differentiation between them, leading to patterns of
genetic/phylogenetic affinity across the genome that mirror geography, regardless of
their previous history. In the case of leapfrog distributions, this means that, in the face
of persistent gene flow, peripheral populations will be closest to their phenotypically
distinct neighbors, even if their phenotypic similarity stems from common ancestry.
However, admixture between lineages is seldom uniform across the genome, since
selection (see below) can restrict gene flow at certain genomic regions (J. R. Turner,
Johnson, & Eanes, 1979; T. L. Turner, Hahn, & Nuzhdin, 2005; Wu, 2001). Such
regions can therefore preserve historic signatures that have been erased by gene flow
elsewhere in the genome. This is likely to be the case for loci underlying color pattern
variation in Phyllobates, especially in cases such as the Northern P. bicolor, where
phenotypic differences persist in spite of gene flow. Hence, the history of alleles at
these loci should provide unique insights into the history of this phenotype. Future
studies to identify such loci and understand their evolutionary history in relation to
our current results will be instrumental to uncover the demographic processes leading
Márquez et al.
28
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
to the current geographic and phylogenetic distribution of solid-yellow color pattern
in Phyllobates, since they will allow for much more explicit tests of the hypotheses
presented here.
Regardless of whether common ancestry or convergent evolution are at play in this
system, it seems clear that differential selective pressures on the striped and solid-
yellow populations have been involved in the origin and/or maintenance of the
geographic distribution of color patterns. Independent evolution of similar phenotypes
is many times promoted by similar changes in selective regimes (Darwin, 1859; Mayr,
1963; G. G. Simpson, 1953), and selection is required to maintain phenotypic
differences between populations in the face of gene flow (Endler, 1977). Two of the
three solid-yellow lineages (the northern and southern P. bicolor) occur at higher
elevations (~600-1500 m.a.s.l) than the rest of the genus (~0-500 m.a.s.l). It is
therefore possible that these mid-elevation habitats pose selective pressures (e.g.
predator communities or light environments) different from those of lowland forests,
which favor solid-yellow patterns over striped ones. The many known examples of
variation in coloration across altitudinal gradients lend support to this idea (e.g.
Köhler, Samietz, & Schielzeth, 2017; Rebelo & Siegfried, 1985; Reguera, Zamora-
Camacho, & Moreno-Rueda, 2014; Richmond & Reeder, 2002; Rios & Álvarez-
Castañeda, 2007). A similar situation could also be the case with P. terribilis, given
its distribution at the southern edge of the genus’s range, where it may also experience
different selective pressures from those faced by its closely related striped lineages.
Márquez et al.
29
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
The nature of color pattern variation in Phyllobates (i.e. solid vs striped) suggests that
differential predation pressures may be important for the origin/maintenance of solid-
yellow patterns. In aposematic species, advertisement signals with complex pattern
elements, such as stripes, have been shown to serve a distance-dependent purpose,
acting as conspicuous signals at short distances, while providing camouflage at long
distances (Barnett et al., 2017; Barnett & Cuthill, 2014; Tullberg, Merilaita, &
Wiklund, 2005). In contrast, bright, solid-colored signals remain conspicuous over a
much wider range of distances. For example, a recent study focusing on the poison
frog Dendrobates tinctorius found that striped patterns of this species are highly
detectable at close range, but become camouflaged when observed from further away.
Solid-yellow patterns, on the other hand, remained easily detectable over the whole
range of distances tested (Barnett, Michalis, Scott-Samuel, & Cuthill, 2018).
Therefore, it is likely that the striped and solid color patterns represent alternative
aposematic strategies that are advantageous under different environments and/or
predator communities.
The fact that we find a signature of correlated evolution between size and color
pattern is compatible with this idea, since larger aposematic signals have been shown
to be more detectable and memorable for predators (Forsman & Merilaita, 1999;
Gamberale & Tullberg, 1996). Accordingly, size and conspicuousness are positively
correlated among Dendrobatid poison frog species (Hagman & Forsman, 2003; Santos
& Cannatella, 2011). However, we do not find a comparable pattern for toxicity,
which has also been shown to co-vary with conspicuousness in poison frogs (Santos
Márquez et al.
30
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
& Cannatella, 2011; Summers & Clough, 2001). This could be an artifact of low
statistical power, since data are available only for one striped and two plain yellow
Colombian lineages, but we cannot rule out the possibility that toxicity is indeed
comparable between solid and striped populations. Furthermore, considering that
aposematism relies on avoidance learning, it is possible that, despite similar levels of
BTX, solid and striped populations differ in levels of palatability to predators. In any
case, a scenario where solid and striped populations are similarly toxic and/or
palatable is still compatible with predation pressures driving evolutionary
convergence, since all species are considerably toxic (Daly et al., 1987; Myers et al.,
1978). However, other explanations, such as geographic variation in mate preference
(R. G. Reynolds & Fitzpatrick, 2007; Summers, Symula, Clough, & Cronin, 1999;
Twomey, Vestergaard, & Summers, 2014; Yang, Richards-Zawacki, Devar, & Dugas,
2016), could also explain our results and cannot be ruled out.
It is worth noting, however, that the correlated evolution of body size and color
pattern could also be due to ontogenetic integration (Olson & Miller, 1958). Tadpoles
of all Phyllobates species are dark grey, and all of them develop a dorsolateral stripe
shortly before metamorphosis, which remains unchanged until adulthood in striped
lineages. Solid-yellow frogs, on the other hand, gradually lose dark pigmentation,
until the solid adult pattern is attained a few months after metamorphosis (Myers et
al., 1978). Therefore it is possible that, for example, the evolution of an extended
growth period could generate changes in both body size and color pattern. If this is the
case, then the concerted evolution of advertisement signal and body size would not
Márquez et al.
31
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
necessarily be evidence of striped and solid patterns representing alternative predator
avoidance strategies.
Our divergence time estimates indicate that the diversification of Phyllobates has
followed the Plio-Pleistocene history of the Central American and the Chocó
bioregions. The first cladogenesis event in our tree, which divides the Central
American and Chocoan taxa was inferred between 4.5-5.9 MYA, which coincides
with previously identified increases in faunal migration between Central and South
America at ~6 MYA (Bacon et al., 2015; Santos et al., 2009). Further branching
within South American lineages occurred later than 3 MYA, after both the Atrato
(Duque-Caro, 1990b, 1990a) and Tumaco (Borrero et al., 2012) basins emerged above
sea level to form the current Chocoan landscape. The Pleistocene was characterized
by recurrent climatic and environmental fluctuations, which have been proposed as
major drivers of neotropical rainforest biodiversity (Baker et al., 2020; Haffer, 1969;
Hooghiemstra & Van Der Hammen, 1998; Vanzolini & Williams, 1970). Although
the central Chocó has traditionally been regarded as a relatively stable Pleistocene
forest refuge throughout the Quaternary (Gentry, 1982; Haffer, 1967; Hooghiemstra
& Van Der Hammen, 1998), a notion supported by multiple palynological studies
(Behling, Hooghiemstra, & Negret, 1998; Berrío, Behling, & Hooghiemstra, 2000;
González, Urrego, & Martínez, 2006; Jaramillo & Bayona, 2000; Ramírez & Urrego,
2002; Urrego, Molina, Urrego, & Ramírez, 2006), there is some evidence of
fluctuations in sea level, temperature, fluvial discharge, and, to a lesser extent,
precipitation throughout the Quaternary in this region (González et al., 2006; Urrego
Márquez et al.
32
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
et al., 2006). Despite being less dramatic than those experienced by other tropical
forests (e.g. Amazonia), these fluctuations appear to have been related to changes in
vegetation, especially the extent of mangrove forests (González et al., 2006). This
may have promoted periodic retractions of Phyllobates populations towards the San
Juan, perhaps resulting in increased rates of gene flow among them. Future work to
understand how climatic fluctuations over the Quaternary have shaped the distribution
of suitable habitat for Phyllobates frogs should shed further light on the biogeographic
history of this genus in Northern South America.
Finally, our findings have broad implications for the systematics of Phyllobates. First
and foremost, this study provides definitive evidence that the populations currently
grouped under P. aurotaenia represent multiple independently-evolving lineages,
some of which have probably been reproductively isolated for enough time to warrant
recognition as separate species. Furthermore, we find that P. bicolor is comprised of
two well-structured lineages that may have evolved similar phenotypes independently.
In addition, we find highly variable levels of mtDNA divergence within P. lugubris
(0-4% 16S, 0.2-8% COI, and 0-5.7% Cytb uncorrected p-distances), which could also
be due to the existence of cryptic species.
It is, therefore, evident that a thorough revision of Phyllobates systematics is
warranted. At this point, however, we refrain from modifying the group’s current
taxonomy for several reasons. First, the complex interplay between divergence and
gene flow that we have found in this system makes species delimitation based on
Márquez et al.
33
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
genetic structure and coloration impractical. Therefore, integrating multiple lines of
evidence (e.g. coloration, genetic variation, alkaloid profiles, bioacoustic data, larval
and adult morphology) that allow us to draw stronger conclusions on the strength of
reproductive barriers between lineages is needed to disentangle species limits.
Second, although our study represents a substantial increase in geographic sampling,
there are still considerable gaps, such as the lower San Juan drainage, or the mid-
elevation forests south of the distribution of P. bicolor, that need to be considered.
The fact that the two captive-bred P. aurotaenia included in mtDNA analyses are
sister to the southern P. bicolor (Fig. 2A) suggests that we have not yet sampled the
full diversity of Phyllobates lineages in Colombia. Finally the holotype of P.
aurotaenia was collected in Condoto, Chocó, which is considerably distant from any
of our sampling localities (Fig. S6), and the type locality of P. bicolor is unknown
(Myers et al., 1978). This situation poses nomenclature issues, since, even if a robust
species delimitation were available, naming these species would not be
straightforward until the type specimens of P. bicolor and P. aurotaenia can be
confidently assigned to one of them. Further work with increased sampling, including
type specimens, and integrating multiple lines of evidence is therefore still needed to
generate a taxonomy for Phyllobates that more accurately represents the genus’s
evolutionary history.
Concluding remarks
Márquez et al.
34
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
Leapfrog distributions constitute geographic replicates for the occurrence of a
phenotype, and therefore provide important information about the origins of
phenotypic diversity among closely related lineages. Here we show that, despite
marked genetic structure and differentiation, there is considerable gene flow between
phenotypically similar populations at the center of a poison-dart frog leapfrog
distribution. This has probably been important for the origin and maintenance of the
geographic distribution of color patterns in this group. Furthermore, we found
instances of both reduced and increased levels of gene flow between neighboring
populations with different phenotypes, suggesting that in some cases reduced gene
exchange can contribute to the maintenance of phenotypic differences between
populations in a leapfrog distribution, while in others these differences actually persist
in the face of gene flow, probably due to local adaptation of different forms.
However, we are unable to answer a commonly addressed question about leapfrog
distributions: whether phenotypic differences between populations stem from
common ancestry or independent evolution. Even though our phylogenetic
reconstructions unambiguously suggest the latter on their own, our finding of
extensive gene flow among neighboring populations casts doubt on this conclusion.
Several other studies on the history of leapfrog distributions have obtained similar
phylogenies that align with geography instead of phenotypic similarity (Cadena et al.,
2010; Garcia-Moreno & Fjeldså, 1999; Norman et al., 2002; Quek et al., 2010; Toon,
Austin, Dolman, Pedler, & Joseph, 2012), leading to the view that leapfrog
distributions are often due to independent evolution. Our results, therefore, add to the
Márquez et al.
35
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
notion that processes such as pervasive gene flow (Hines et al., 2011; James et al.,
2020) or incomplete lineage sorting (Avise & Robinson, 2008) can decouple the
history of phenotypic change at a given trait from genome-wide patterns of
divergence, possibly leading to erroneous inferences of convergent evolution (Hahn &
Nakhleh, 2015).
Data accessibility
mtDNA sequences were uploaded to GenBank under accessions MT742690-
MT742754, MT749179-MT749246, and MT808222-MT808283. Raw Illumina reads
were uploaded to the NCBI SRA under BioProject ID PRJNA645960. The
assemblies, bam and vcf files, and body size data, as well as the code used for
analyses are available at https://doi.org/10.5061/dryad.8d4r3vd or as supplementary
material.
Acknowledgements
We thank Pablo Palacios-Rodríguez, José Alfredo Hernández, Carolina Esquivel,
Diana Galindo, Mabel González, and Fernando Vargas-Salinas for assistance in the
field, Lydia Smith, Valeria Ramírez-Castañeda, Alvaro Hernández, and Ke Bi for help
with molecular and bioinformatic procedures, Andrea Paz for advice on MMRR
analyses, and Alan Resetar (FMNH), Andrew Crawford and Alberto Farfán
(ANDES), Rayna Bell and Addison Wynn (USNM), Andrés Acosta and Carlos
Márquez et al.
36
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
Montaña (IAvH), John Taylor Rengifo (UTCh), Greg Schneider (UMMZ), and David
Kizirian (AMNH) for facilitating access to preserved specimens. Comments and
suggestions from Trevor Price, John Novembre, Valentina Gómez-Bahamon, John
Bates, Daniel Matute, Catalina Gonzalez, the Bates/Hackett lab, the Kronforst lab,
and four anonymous reviewers greatly improved this paper. We sincerely thank Lina
M. Arenas for allowing us to use her beautiful frog illustrations for our figures. This
work was funded by a Basic Sciences Grant from the Vice Chancellor of Research at
Universidad de los Andes, a Colombia Biodiversa Scholarship from the Alejandro
Angel Escobar foundation, a Pew Biomedical Scholarship, Neubauer Family funds
and a Steiner Award from the University of Chicago, and NSF grant DEB-1655336.
RM was partially supported by a Fellowship for Young Researchers and Innovators
(Otto de Greiff) from COLCIENCIAS. Computations were performed on the
University of Chicago’s Gardner HPC cluster, funded by NIH grant TR000430.
Tissue collections were authorized by permits No. 2194 and 1380 from the Colombian
Ministry of Environment and Authority for Environmental Licenses (ANLA).
Author contributions
R.M., A.A., and M.R.K. conceived the project, R.M., T.P.L, R.N., M.R.K, and A.A.
designed the research, R.M., A.A., R.N. and M.R.K. acquired funding, A.A., D.M-V.,
and R.M. collected samples, R.M. and T.P.L. generated the data, and R.M. analyzed
the data and wrote the paper with input from M.K. and edits from all authors.
Márquez et al.
37
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
Supplementary information
Supplementary note. Criteria used to select Operational Taxonomic Units (OTUs)
for phylogenetic inference with SNAPP.
Table S1. Information and accession numbers for samples used in this study. Locality
data is limited to prevent illegal traffic. Further details are available from the authors
upon request.
Table S2. Results of likelihood ratio tests performed on Treemix analyses run with m
= 0-6 migration edges.
Table S3. Information and snout-to-vent lengths of museum specimens measured for
comparative analyses.
Figure S1. Localities joined into a single deme for locality-level analyses.
Figure S2. Results of Treemix analyses run with m=0-6 migration edges.
Figure S3. Minimum-evolution tree based on genetic distances.
Figure S4. Mitochondrial DNA time tree inferred using BEAST 2.
Figure S5. Results of phylogenetic comparative analyses run on 1000 randomly-
drawn trees from the SNAPP posterior tree distribution.
Figure S6. Map of the type locality of P. aurotaenia.
Figure S7 and associated text. Details on our implementation of the Savage-Dickey
ratio to estimate Bayes Factors between nested SpaceMix models.
Márquez et al.
38
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
Tables
Table 1. Results from the multiple matrix regression with randomization (MMRR) analyses performed with multiple different resistance values for the
San Juan river as a corridor (see Methods for details). P-values were estimated using 10,000 permutations. Coef. = coefficient; Diff. = difference; Dist.
= distance; LC = least-cost.
Predictor
Resistance = 100
Resistance = 50
Resistance = 20
Resistance = 10
Resistance = 1
Coef.
t
p
Coef.
t
p
Coef.
t
p
Coef.
t
p
Coef.
t
p
Intercept
0.0325
0.5286
1
0.0325
0.5298
1
0.0327
0.5343
1
0.0338
0.5427
1
0.0369
0.5953
1
Geodesic Dist.
0.6419 10.0547
< 1e-4
0.6420
10.0618
0.0001
0.6423 10.0918
< 1e-4
0.6428
10.1470
< 1e-4
0.6483 10.0550
< 1e-4
San Juan LC Dist.
0.2386
4.0094
0.0137
0.2392
4.0241
0.0138
0.2420
4.0890
0.0118
0.2471
4.2051
0.0118
0.2239
3.7483
0.0126
Elevation Diff.
0.2886
4.5994
0.0032
0.2877
4.5917
0.0037
0.2852
4.5697
0.0023
0.2802
4.5231
0.0023
0.2271
3.6051
0.0129
r2 = 0.62, F = 48.13, p < 1e-4
r2 = 0.62, F = 48.21, p < 1e-4
r2 = 0.63, F = 48.61, p < 1e-4
r2 = 0.63, F = 49.33, p < 1e-4
r2 = 0.62, F = 46.61, p < 1e-4
862
863
864
865
866
867
Figures
Fig 1. A) Color pattern diversity, B) currently accepted phylogenetic relationships
(Grant et al., 2017), and C) geographic distribution of Phyllobates poison frogs.
Species distribution polygons were obtained from the IUCN red list of threatened
species website (https://www.iucnredlist.org/) and modified to fit natural history
collection records and our own observations.
Márquez et al.
40
868
870
871
872
873
874
875
876
877
Fig 2. Genetic structure among Phyllobates populations in western Colombia. A)
Maximum likelihood mtDNA genealogy inferred from 1926bp. B) Sampling localities
for this study. C) Principal component analysis plot based on the first three
components accounting for 29.4% of the variance. D) Treemix population graph
assuming 2 migration edges. E) Individual admixture proportions assuming 2-9
ancestral populations. Colors in A-D correspond to the operational taxonomic units
(OTU) used for phylogenetic analyses. Colors in E were chosen to loosely represent
these clusters.
Márquez et al.
41
879
880
881
882
883
884
885
886
887
Fig. 3 Phylogenetic relationships and divergence times among Phyllobates lineages
inferred using SNAPP. Divergence times assume a mutation rate of 10-9 mutations per
year and a generation time of one year. Each individual tree represents one sample
from the SNAPP posterior distribution. Clades present in more posterior trees have
higher posterior probabilities (i.e. higher nodal support). The color scheme is as in
Fig. 2.
Márquez et al.
42
889
890
891
892
893
894
895
896
Fig. 4 Evolutionary patterns of body size, color pattern, and toxicity in Phyllobates.
A) Maximum clade credibility tree derived from the SNAPP posterior distribution and
phylogenetic distribution of color pattern and snout-to-vent length (SVL) values
among lineages. Numbers on internodes represent clade posterior probabilities. B)
Phylogenetic biplot depicting the relationship between mean SVL and mean
batrachotoxin concentration. Grey boxes/points represent striped lineages while
yellow ones represent solid-yellow lineages.
Márquez et al.
43
898
899
900
901
902
903
904
905
906
907
908
909
910
Fig. 5 Gene flow among Phyllobates species across Western Colombia. A) Effective
migration surface estimated using EEMS. Cyan and brown areas of the map are those
where migration between demes is higher (cyan) or lower (brown) than expected
under isolation by distance. Grey lines depict the population grid and habitat outline
used by EEMS. B) Geo-genetic map inferred with SpaceMix. Ellipses represent the
Márquez et al.
44
912
913
914
915
916
95% Bayesian credible intervals around each population’s location on the geo-genetic
map, and colored dots represent actual sampling locations. Arrows connect sampling
and geogenetic locations. C) Density histograms of the posterior distributions of the
long-range admixture proportion parameters (w) from the SpaceMix model for each
population. Bars, points, and ellipses are colored by OTU as in Fig. 2. Histograms in
Fig. 5C are ordered by OTU
Márquez et al.
45
917
918
919
920
921
922
923
924
925
References
Altekar, G., Dwarkadas, S., Huelsenbeck, J. P., & Ronquist, F. (2004). Parallel
Metropolis coupled Markov chain Monte Carlo for Bayesian phylogenetic
inference. Bioinformatics, 20(3), 407–415. doi:10.1093/bioinformatics/btg427
Altschul, S. (1997). Gapped BLAST and PSI-BLAST: a new generation of protein
database search programs. Nucleic Acids Research, 25(17), 3389–3402.
doi:10.1093/nar/25.17.3389
Anisimova, M., Gil, M., Dufayard, J. F., Dessimoz, C., & Gascuel, O. (2011). Survey
of branch support methods demonstrates accuracy, power, and robustness of fast
likelihood-based approximation schemes. Systematic Biology, 60(5), 685–699.
doi:10.1093/sysbio/syr041
Avise, J. C., & Robinson, T. J. (2008). Hemiplasy: A new term in the lexicon of
phylogenetics. Systematic Biology, 57(3), 503–507.
doi:10.1080/10635150802164587
Bacon, C. D., Silvestro, D., Jaramillo, C., Smith, B. T., Chakrabarty, P., & Antonelli,
A. (2015). Biological evidence supports an early and complex emergence of the
Isthmus of Panama. Proceedings of the National Academy of Sciences of the
United States of America, 112(19), 6110–6115. doi:10.1073/pnas.1423853112
Baker, P. A., Fritz, S. C., Battisti, D. S., Dick, C. W., Vargas, O. M., Asner, G. P., …
Prates, I. (2020). Beyond Refugia: New insights on Quaternary climate variation
and the evolution of biotic diversity in tropical South America. In V. Rull & A.
C. Carnaval (Eds.), Neotropical Diversification: Patterns and Processes. Berlin:
Springer. doi:10.1007/978-3-030-31167-4
Barnett, J. B., & Cuthill, I. C. (2014). Distance-dependent defensive coloration.
Current Biology, 24(24), R1157–R1158. doi:10.1016/j.cub.2014.11.015
Barnett, J. B., Michalis, C., Scott-Samuel, N. E., & Cuthill, I. C. (2018). Distance-
dependent defensive coloration in the poison frog Dendrobates tinctorius ,
Dendrobatidae. Proceedings of the National Academy of Sciences of the United
States of America, 115(25), 6416–6421. doi:10.1073/pnas.1800826115
Barnett, J. B., Redfern, A. S., Bhattacharyya-Dickson, R., Clifton, O., Courty, T., Ho,
T., … Cuthill, I. C. (2017). Stripes for warning and stripes for hiding: Spatial
frequency and detection distance. Behavioral Ecology, 28(2), 373–381.
doi:10.1093/beheco/arw168
Márquez et al.
46
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
Behling, H., Hooghiemstra, H., & Negret, A. J. (1998). Holocene history of the Chocó
rain forest from Laguna Piusbi, southern Pacific lowlands of Colombia.
Quaternary Research, 50(3), 300–308. doi:10.1006/qres.1998.1998
Berrío, J. C., Behling, H., & Hooghiemstra, H. (2000). Tropical rain-forest history
from the Colombian Pacific area: A 4200-year pollen record from Laguna
Jotaordó. Holocene, 10(6), 749–756. doi:10.1191/09596830094999
Bi, K., Vanderpool, D., Singhal, S., Linderoth, T., Moritz, C., & Good, J. M. (2012).
Transcriptome-based exon capture enables highly cost-effective comparative
genomic data collection at moderate evolutionary scales. BMC Genomics, 13(1),
403. doi:10.1186/1471-2164-13-403
Bolger, A. M., Lohse, M., & Usadel, B. (2014). Trimmomatic: A flexible trimmer for
Illumina sequence data. Bioinformatics, 30(15), 2114–2120.
doi:10.1093/bioinformatics/btu170
Borrero, C., Pardo, A., Jaramillo, C. M., Osorio, J. A., Cardona, A., Flores, A., …
Castillo, H. (2012). Tectonostratigraphy of the Cenozoic Tumaco forearc basin
(Colombian Pacific) and its relationship with the northern Andes orogenic build
up. Journal of South American Earth Sciences, 39, 75–92.
doi:10.1016/j.jsames.2012.04.004
Bouckaert, R., Vaughan, T. G., Barido-Sottani, J., Duchêne, S., Fourment, M.,
Gavryushkina, A., … Drummond, A. J. (2019). BEAST 2.5: An advanced
software platform for Bayesian evolutionary analysis. PLoS Computational
Biology, 15(4), e1006650. doi:10.1371/journal.pcbi.1006650
Bradburd, G. S., Ralph, P. L., & Coop, G. M. (2016). A spatial framework for
understanding population structure and admixture. PLoS Genetics, 12(1), 1–38.
doi:10.1371/journal.pgen.1005703
Brower, A. V. Z. (1996). Parallel race formation and the evolution of mimicry in
Heliconius butterflies : A phylogenetic hypothesis from mitochondrial DNA
sequences. Evolution, 50(1), 195–221. doi:10.2307/2410794
Bryant, D., Bouckaert, R., Felsenstein, J., Rosenberg, N. A., & RoyChoudhury, A.
(2012). Inferring species trees directly from biallelic genetic markers: Bypassing
gene trees in a full coalescent analysis. Molecular Biology and Evolution, 29(8),
1917–1932. doi:10.1093/molbev/mss086
Cadena, C. D., Cheviron, Z. A., & Funk, W. C. (2010). Testing the molecular and
evolutionary causes of a ‘leapfrog’ pattern of geographical variation in
Márquez et al.
47
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
coloration. Journal of Evolutionary Biology, 24(2), 402–414.
doi:10.1111/j.1420-9101.2010.02175.x
Chapman, F. M. (1923). Mutation among birds on the genus Buarremon. Bulletin of
the American Museum of Natural History, 48, 243–278.
Che, J., Chen, H.-M., Yang, J.-Xi., Jin, J.-Q., Jiang, K., Yuan, Z.-Y., … Zhang, Y.-P.
(2012). Universal COI primers for DNA barcoding amphibians. Molecular
Ecology Resources, 12(2), 247–258. doi:10.1111/j.1755-0998.2011.03090.x
Crawford, A. J. (2003). Relative rates of nucleotide substitution in frogs. Journal of
Molecular Evolution, 57(6), 636–641. doi:10.1007/s00239-003-2513-7
Daly, J. W., Myers, C. W., & Whittaker, N. (1987). Further classification of skin
alkaloids from neotropical poison frogs (Dendrobatidae), with a general survey
of toxic/noxious substances in the Amphibia. Toxicon, 25(10), 1023–1095.
doi:10.1016/0041-0101(87)90265-0
Darwin, C. R. (1859). On the origin of species, by means of natural selection; or, The
preservation of favoured races in the struggle for life. London: John Murray.
Dickey, J. M., & Lientz, B. P. (1970). The weighted likelihood ratio, sharp hypotheses
about chances, the order of a Markov chain. The Annals of Mathematical
Statistics, 41(1), 214–226. doi:10.1214/aoms/1177697203
Dijkstra, E. W. (1959). A note on two problems in connexion with graphs.
Numerische Mathematik, 1(1), 269–271. doi:10.1007/BF01386390
Duque-Caro, J. H. (1990a). Neogene stratigraphy, paleoceanography and
paleobiogeography in northwest South America and the evolution of the Panama
seaway. Palaeogeography, Palaeoclimatology, Palaeoecology, 77(3–4), 203–
234. doi:10.1016/0031-0182(90)90178-A
Duque-Caro, J. H. (1990b). The Chocó block in the northwestern corner of South
America: Structural, tectonostratigraphic, and paleogeographic implications.
Journal of South American Earth Sciences, 3(1), 71–84. doi:10.1016/0895-
9811(90)90019-W
Edgar, R. C. (2004). MUSCLE: multiple sequence alignment with high accuracy and
high throughput. Nucleic Acids Research, 32(5), 1792–1797.
doi:10.1093/nar/gkh340
Emsley, M. G. (1965). The geographical distribution of the color-pattern components
of Heloconius erato and Heliconius melpomene with genetical evidence for the
systematic relationship between the two species. Zoologica, 49(15), 245–286.
Márquez et al.
48
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
Endler, J. A. (1977). Geographic Variation, Speciation, and Clines. Monographs In
Population Biology. Princeton: Princeton University Press.
Fitch, W. M. (1971). Toward defining the course of evolution: Minimum change for a
specific tree topology. Systematic Zoology, 20(4), 406. doi:10.2307/2412116
Forsman, A., & Merilaita, S. (1999). Fearful symmetry: Pattern size and asymmetry
affects aposematic signal efficacy. Evolutionary Ecology, 13(2), 131–140.
doi:10.1023/A:1006630911975
Fu, L., Niu, B., Zhu, Z., Wu, S., & Li, W. (2012). CD-HIT: Accelerated for clustering
the next-generation sequencing data. Bioinformatics, 28(23), 3150–3152.
doi:10.1093/bioinformatics/bts565
Gamberale, G., & Tullberg, B. S. (1996). Evidence for a peak-shift in predator
generalization among aposematic prey. Proceedings of the Royal Society B:
Biological Sciences, 263(1375), 1329–1334. doi:10.1098/rspb.1996.0195
Garcia-Moreno, J., & Fjeldså, J. (1999). Re-evaluation of species limits in the genus
Atlapetes based on mtDNA sequence data. Ibis, 141(2), 199–207.
doi:10.1111/j.1474-919X.1999.tb07542.x
Gentry, A. H. (1982). Phytogeographic patterns as evidence for a Chocó refuge. In G.
T. Prance (Ed.), Biological Diversification in the Tropics (pp. 112–136). New
York: Plenum Press.
González, C., Urrego, L. E., & Martínez, J. I. (2006). Late Quaternary vegetation and
climate change in the Panama Basin: Palynological evidence from marine cores
ODP 677B and TR 163-38. Palaeogeography, Palaeoclimatology,
Palaeoecology, 234(1), 62–80. doi:10.1016/j.palaeo.2005.10.019
Grabherr, M. G., Haas, B. J., Yassour, M., Levin, J. Z., Thompson, D. a, Amit, I., …
Regev, A. (2011). Full-length transcriptome assembly from RNA-Seq data
without a reference genome. Nature Biotechnology, 29(7), 644–652.
doi:10.1038/nbt.1883
Grafen, A. (1989). The phylogenetic regression. Philosophical Transactions of the
Royal Society B: Biological Sciences, 326(1233), 119–157.
doi:10.1098/rstb.1989.0106
Grant, T., Frost, D. R., Caldwell, J. P., Gagliardo, R., Haddad, C. F. B., Kok, P. J. R.,
… Wheeler, W. C. (2006). Phylogenetic systematics of dart-poison frogs and
their relatives (Amphibia: Athesphatanura: Dendrobatidae). Bulletin of the
Márquez et al.
49
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
American Museum of Natural History, 299, 1–262. doi:10.1206/0003-
0090(2006)299[1:PSODFA]2.0.CO;2
Grant, T., Rada, M., Anganoy-Criollo, M., Batista, A., Dias, P. H., Jeckel, A. M., …
Rueda-Almonacid, J. V. (2017). Phylogenetic systematics of Dart-Poison frogs
and their relatives revisited (Anura: Dendrobatoidea). South American Journal of
Herpetology, 12(s1), S1–S90. doi:10.2994/SAJH-D-17-00017.1
Guindon, S., Dufayard, J. F., Lefort, V., Anisimova, M., Hordijk, W., & Gascuel, O.
(2010). New algorithms and methods to estimate maximum-likelihood
phylogenies: Assessing the performance of PhyML 3.0. Systematic Biology,
59(3), 307–321. doi:10.1093/sysbio/syq010
Haffer, J. (1967). Speciation in Colombian forest birds west of the Andes. American
Museum Novitates, 2294, 1–57.
Haffer, J. (1969). Speciation in Amazonian Forest Birds. Science, 165(3889), 131–
137. doi:10.1126/science.165.3889.131
Hagman, M., & Forsman, A. (2003). Correlated evolution of conspicuous coloration
and body size in poison frogs (Dendrobatidae). Evolution, 57(12), 2904–2910.
doi:10.1111/j.0014-3820.2003.tb01531.x
Hahn, M. W., & Nakhleh, L. (2015). Irrational exuberance for resolved species trees.
Evolution, 70(1), 7–17. doi:10.1111/evo.12832
Hijmans, R. J. (2017). raster: Geographic Data Analysis and Modeling. R Package
Version 2.6-7.
Hines, H. M., Counterman, B. A., Papa, R., Albuquerque de Moura, P., Cardoso, M.
Z., Linares, M., … McMillan, W. O. (2011). Wing patterning gene redefines the
mimetic history of Heliconius butterflies. Proceedings of the National Academy
of Sciences of the United States of America. doi:10.1073/pnas.1110096108
Hodges, E., Rooks, M., Xuan, Z., Bhattacharjee, A., Benjamin Gordon, D., Brizuela,
L., … Hannon, G. J. (2009). Hybrid selection of discrete genomic intervals on
custom-designed microarrays for massively parallel sequencing. Nature
Protocols, 4(6), 960–974. doi:10.1038/nprot.2009.68
Hodges, E., Xuan, Z., Balija, V., Kramer, M., Molla, M. N., Smith, S. W., …
McCombie, W. R. (2007). Genome-wide in situ exon capture for selective
resequencing. Nature Genetics, 39(12), 1522–1527. doi:10.1038/ng.2007.42
Hooghiemstra, H., & Van Der Hammen, T. (1998). Neogene and Quaternary
development of the neotropical rain forest: The forest refugia hypothesis, and a
Márquez et al.
50
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
literature overview. Earth Science Reviews, 44(3–4), 147–183.
doi:10.1016/S0012-8252(98)00027-0
Hovanitz, W. (1940). Ecological color variation in a butterfly and the problem of
“protective coloration.” Ecology, 21(3), 371–380. doi:10.2307/1930846
Huang, X. Q., & Madan, A. (1999). CAP3: A DNA sequence assembly program.
Genome Research, 9(9), 868–877. doi:10.1101/gr.9.9.868
Huxley, J. S. (1942). Evolution The Modern Synthesis. New York City: Harper and
Brothers.
James, M. E., Arenas-Castro, H., Groh, J. S., Engelstaedter, J., & Ortiz-Barrientos, D.
(2020). Highly replicated evolution of parapatric ecotypes. BioRxiv,
2020.02.05.936401. doi:10.1101/2020.02.05.936401
Jaramillo, C., & Bayona, G. (2000). Mangrove distribution during the Holocene in
Tribugá Gulf, Colombia. Biotropica, 32(1), 14. doi:10.1646/0006-
3606(2000)032[0014:mddthi]2.0.co;2
Jiang, H., Lei, R., Ding, S. W., & Zhu, S. (2014). Skewer: A fast and accurate adapter
trimmer for next-generation sequencing paired-end reads. BMC Bioinformatics,
15(1), 1–12. doi:10.1186/1471-2105-15-182
Kass, R. E., & Raftery, A. E. (1995). Bayes factors. Journal of the American
Statistical Association, 90(430), 773–795.
doi:10.1080/01621459.1995.10476572
Kearse, M., Moir, R., Wilson, A., Stones-Havas, S., Cheung, M., Sturrock, S., …
Drummond, A. (2012). Geneious Basic: An integrated and extendable desktop
software platform for the organization and analysis of sequence data.
Bioinformatics, 28(12), 1647–1649. doi:10.1093/bioinformatics/bts199
Köhler, G., Samietz, J., & Schielzeth, H. (2017). Morphological and colour morph
clines along an altitudinal gradient in the meadow grasshopper
Pseudochorthippus parallelus. PLoS ONE, 12(12), 1–13.
doi:10.1371/journal.pone.0189815
Korneliussen, T. S., Albrechtsen, A., & Nielsen, R. (2014). ANGSD: Analysis of next
generation sequencing data. BMC Bioinformatics, 15(1), 1–13.
doi:10.1186/s12859-014-0356-4
Lanfear, R., Frandsen, P. B., Wright, A. M., Senfeld, T., & Calcott, B. (2017).
Partitionfinder 2: New methods for selecting partitioned models of evolution for
Márquez et al.
51
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
molecular and morphological phylogenetic analyses. Molecular Biology and
Evolution, 34(3), 772–773. doi:10.1093/molbev/msw260
Langmead, B., & Salzberg, S. L. (2012). Fast gapped-read alignment with Bowtie 2.
Nature Methods, 9(4), 357–359. doi:10.1038/nmeth.1923
Lefort, V., Desper, R., & Gascuel, O. (2015). FastME 2.0: A comprehensive,
accurate, and fast distance-based phylogeny inference program. Molecular
Biology and Evolution, 32(10), 2798–2800. doi:10.1093/molbev/msv150
Li, H., Handsaker, B., Wysoker, A., Fennell, T., Ruan, J., Homer, N., … Durbin, R.
(2009). The Sequence Alignment/Map format and SAMtools. Bioinformatics,
25(16), 2078–2079. doi:10.1093/bioinformatics/btp352
Link, V., Kousathanas, A., Veeramah, K., Sell, C., Scheu, A., & Wegmann, D.
(2017). ATLAS: Analysis Tools for Low-depth and Ancient Samples. BioRxiv,
105346. doi:10.1101/105346
Magoč, T., & Salzberg, S. L. (2011). FLASH: Fast length adjustment of short reads to
improve genome assemblies. Bioinformatics, 27(21), 2957–2963.
doi:10.1093/bioinformatics/btr507
Märki, F., & Witkop, B. (1963). The venom of the Colombian arrow poison frog
Phyllobates bicolor. Experientia, 19(8), 329–338. doi:10.1007/bf02152303
Márquez, R., Corredor, G., Galvis, C., Góez, D., & Amézquita, A. (2012). Range
extension of the critically endangered true poison-dart frog, Phyllobates terribilis
(Anura: Dendrobatidae), in western Colombia. Acta Herpetologica, 7(2), 341–
345. doi:10.13128/Acta_Herpetol-11387
Martins, E. P., & Hansen, T. F. (1997). Phylogenies and the comparative method: A
general approach to incorporating phylogenetic information into the analysis of
interspecific data. The American Naturalist, 149(4), 646–667.
doi:10.1086/286013
Matsumura, S., Yokohama, J., Fukuda, T., & Maki, M. (2009). Origin of the disjunct
distribution of flower colour polymorphism within Limonium wrightii
(Plumbaginaceae) in the Ryukyu Archipelago. Biological Journal of the Linnean
Society, 97(4), 709–717. doi:10.1111/j.1095-8312.2009.01253.x
Matsumura, S., Yokoyama, J., Tateishi, Y., & Maki, M. (2006). Intraspecific variation
of flower colour and its distribution within a sea lavender, Limonium wrightii
(Plumbaginaceae), in the northwestern Pacific Islands. Journal of Plant
Research, 119(6), 625–632. doi:10.1007/s10265-006-0022-7
Márquez et al.
52
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
Maxson, L. R., & Myers, C. W. (1985). Albumin evolution in tropical poison frogs
(Dendrobatidae): A preliminary report. Biotropica, 17(1), 50–56.
doi:10.2307/2388378
Mayr, E. (1942). Systematics and the Origin of Species from, the Viewpoint of a
Zoologist. New York: Columbia University Press.
Mayr, E. (1963). Animal Species and Evolution. Cambridge: Harvard University
Press.
McKenna, A., Hanna, M., Banks, E., Sivachenko, A., Cibulskis, K., Kernytsky, A., …
DePristo, M. A. (2010). The Genome Analysis Toolkit: A MapReduce
framework for analyzing next-generation DNA sequencing data. Genome
Research, 20(9), 1297–1303. doi:10.1101/gr.107524.110
Meisner, J., & Albrechtsen, A. (2018). Inferring population structure and admixture
proportions in low-depth NGS data. Genetics, 210(2), 719–731.
doi:10.1534/genetics.118.301336
Meyer, M., & Kircher, M. (2010). Illumina sequencing library preparation for highly
multiplexed target capture and sequencing. Cold Spring Harbor Protocols,
2010(6), pdb.prot5448-pdb.prot5448. doi:10.1101/pdb.prot5448
Miller, S. A., Dykes, D. D., & Polesky, H. F. (1988). A simple salting out procedure
for extracting DNA from human nucleated cells. Nucleic Acids Research, 16(3),
1215. doi:10.1093/nar/16.3.1215
Myers, C. W., Daly, J. W., & Malkin, B. (1978). A dangerously toxic new frog
(Phyllobates) used by Emberá Indians of western Colombia, with discussion of
blowgun fabrication and dart poisoning. Bulletin of the American Museum of
Natural History, 161, 307–366.
Norman, J. A., Christidis, L., Joseph, L., Slikas, B., & Alpers, D. (2002). Unravelling
a biogeographical knot: origin of the ‘leapfrog’ distribution pattern of Australo-
Papuan sooty owls (Strigiformes) and logrunners (Passeriformes). Proceedings
of the Royal Society B: Biological Sciences, 269(1505), 2127–2133.
doi:10.1098/rspb.2002.2136
Olson, E. C., & Miller, R. L. (1958). Morphological integration. Chicago: University
of Chicago Press.
Pagel, M. (1999). Inferring the historical patterns of biological evolution. Nature,
401(6756), 877–884. doi:10.1038/44766
Márquez et al.
53
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
Palumbi, S. R., Martin, A., Romano, S., McMillan, W. O., Stice, L., & Grabowski, G.
(1991). The simple fool’s guide to PCR. Honolulu: Department of Zoology,
University of Hawaii.
Paradis, E., Claude, J., & Strimmer, K. (2004). APE: Analyses of phylogenetics and
evolution in R language. Bioinformatics, 20(2), 289–290.
doi:10.1093/bioinformatics/btg412
Petersen, K. R., Streett, D. A., Gerritsen, A. T., Hunter, S. S., & Settles, M. L. (2015).
Super deduper, fast PCR duplicate detection in fastq files. In Proceedings of the
6th ACM Conference on Bioinformatics, Computational Biology and Health
Informatics - BCB ’15 (pp. 491–492). New York, New York, USA: ACM Press.
doi:10.1145/2808719.2811568
Petkova, D., Novembre, J., & Stephens, M. (2015). Visualizing spatial population
structure with estimated effective migration surfaces. Nature Genetics, 48(1),
94–100. doi:10.1038/ng.3464
Pickrell, J. K., & Pritchard, J. K. (2012). Inference of population splits and mixtures
from genome-wide allele frequency data. PLoS Genetics, 8(11), e1002967.
doi:10.1371/journal.pgen.1002967.s016
Pinheiro, J., Bates, D., DebRoy, S., & Sarkar, D. (2017). nlme: Linear and Nonlinear
Mixed Effects models. R Package Version 3.1-131.
Pough, F. H., & Taigen, T. L. (1990). Metabolic correlates of the foraging and social
behaviour of dart-poison frogs. Animal Behaviour, 39(1), 145–155.
doi:10.1016/S0003-3472(05)80734-1
Purcell, S., Neale, B., Todd-Brown, K., Thomas, L., Ferreira, M. A. R., Bender, D.,
… Sham, P. C. (2007). PLINK: A tool set for whole-genome association and
population-based linkage analyses. The American Journal of Human Genetics,
81(3), 559–575. doi:10.1086/519795
Quek, S.-P., Counterman, B. A., Albuquerque de Moura, P., Cardoso, M. Z.,
Marshall, C. R., McMillan, W. O., & Kronforst, M. R. (2010). Dissecting
comimetic radiations in Heliconius reveals divergent histories of convergent
butterflies. Proceedings of the National Academy of Sciences of the United
States of America, 107(16), 7365–7370. doi:10.1073/pnas.0911572107
Rambaut, A., & Drummond, A. J. (2009). Tracer v. 1.5. Available from:
Http://Beast.Bio.Ed.Ac.Uk/Tracer.
Márquez et al.
54
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
Ramírez, L. F., & Urrego, L. E. (2002). Vegetación holocénica en el delta del río San
Juan, Pacífico colombiano. In I. D. Correa & J. D. Restrepo (Eds.), Geologia y
Oceanografía del Dela del Río San Juan. Litoral Pacífico Colombiano. (1st ed.,
pp. 151–169). Medellín: Editorial Universidad EAFIT.
Rebelo, A. G., & Siegfried, W. R. (1985). Colour and size of flowers in relation to
pollination of Erica species. Oecologia, 65(4), 584–590.
doi:10.1007/BF00379677
Reguera, S., Zamora-Camacho, F. J., & Moreno-Rueda, G. (2014). The lizard
Psammodromus algirus (Squamata: Lacertidae) is darker at high altitudes.
Biological Journal of the Linnean Society, 112(1), 132–141.
doi:10.1111/bij.12250
Remsen, J. V. (1984). High incidence of “leapfrog” pattern of geographic variation in
andean birds: implications for the speciation process. Science, 224(4645), 171–
173. doi:10.1126/science.224.4645.171
Reynolds, J., Weir, B. S., & Cockerham, C. C. (1983). Estimation of the coancestry
coefficient: Basis for a short-term genetic distance. Genetics, 105(3), 767–779.
doi:10.1534/genetics.103.025387
Reynolds, R. G., & Fitzpatrick, B. M. (2007). Assortative mating in poison-dart frogs
based on an ecologically important trait. Evolution, 61(9), 2253–2259.
doi:10.1111/j.1558-5646.2007.00174.x
Richmond, J. Q., & Reeder, T. W. (2002). Evidence for parallel ecological speciation
in sincid lizards of the Eumeces skiltonianus species group (Squamata:
Scincidae). Evolution, 56(7), 1498–1513. doi:10.1111/j.0014-
3820.2002.tb01461.x
Rios, E., & Álvarez-Castañeda, S. T. (2007). Environmental responses to altitudinal
gradients and subspecific validity in pocket gophers (Thomomys bottae) from
Baja California Sur, Mexico. Journal of Mammalogy, 88(4), 926–934.
doi:10.1644/06-mamm-a-226r1.1
Ronquist, F., Teslenko, M., van der Mark, P., Ayres, D. L., Darling, A., Hohna, S., …
Huelsenbeck, J. P. (2012). MrBayes 3.2: Efficient Bayesian phylogenetic
inference and model choice across a large model space. Systematic Biology,
61(3), 539–542. doi:10.1093/sysbio/sys029
Santos, J. C., Baquero, M., Barrio-Amorós, C., Coloma, L. A., Erdtmann, L. K.,
Lima, A. P., & Cannatella, D. C. (2014). Aposematism increases acoustic
Márquez et al.
55
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
diversification and speciation in poison frogs. Proceedings of the Royal Society
B: Biological Sciences, 281(1796), 20141761. doi:10.1098/rspb.2014.1761
Santos, J. C., & Cannatella, D. C. (2011). Phenotypic integration emerges from
aposematism and scale in poison frogs. Proceedings of the National Academy of
Sciences of the United States of America, 108(15), 6175–6180.
doi:10.1073/pnas.1010952108
Santos, J. C., Coloma, L. A., Summers, K., Caldwell, J. P., Ree, R., & Cannatella, D.
C. (2009). Amazonian amphibian diversity is primarily derived from late
miocene andean lineages. PLoS Biology, 7(3), e56.
doi:10.1371/journal.pbio.1000056.st014
Schliep, K. P. (2011). phangorn: phylogenetic analysis in R. Bioinformatics, 27(4),
592–593. doi:10.1093/bioinformatics/btq706
Sheppard, P. M., Turner, J. R. G., Brown, K. S., Benson, W. W., & Singer, M. C.
(1985). Genetics and the evolution of Muellerian mimicry in Heliconius
butterflies. Philosophical Transactions of the Royal Society B: Biological
Sciences, 308(1137), 433–610. doi:10.1098/rstb.1985.0066
Silverstone, P. A. (1976). A revision of the poison-arrow frogs of the genus
Phyllobates Bibron in Sagra (Family Dendrobatidae). Natural History Museum
of Los Angeles County, Science Bulletin, 27, 1–53.
Simpson, G. G. (1953). The major features of evolution. New York: Columbia
University Press.
Simpson, J. T., Wong, K., Jackman, S. D., Schein, J. E., Jones, S. J. M., & Birol, I.
(2009). ABySS: A parallel assembler for short read sequence data. Genome
Research, 19(6), 1117–1123. doi:10.1101/gr.089532.108
Skotte, L., Korneliussen, T. S., & Albrechtsen, A. (2013). Estimating individual
admixture proportions from next generation sequencing data. Genetics, 195(3),
693–702. doi:10.1534/genetics.113.154138
Slater, G. S. C., & Birney, E. (2005). Automated generation of heuristics for
biological sequence comparison. BMC Bioinformatics, 6, 31. doi:10.1186/1471-
2105-6-31
Smith, A. F. ., Hubley, R., & Green, P. (2013). RepeatMasker Open-4.0.
Summers, K., & Clough, M. E. (2001). The evolution of coloration and toxicity in the
poison frog family (Dendrobatidae). Proceedings of the National Academy of
Márquez et al.
56
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
Sciences of the United States of America, 98(11), 6227–6232.
doi:10.1073/pnas.101134898
Summers, K., Symula, R., Clough, M., & Cronin, T. (1999). Visual mate choice in
poison frogs. Proceedings of the Royal Society B: Biological Sciences,
266(1434), 2141–2145. doi:10.1098/rspb.1999.0900
Sun, Y.-B., Xiong, Z.-J., Xiang, X.-Y., Liu, S.-P., Zhou, W.-W., Tu, X.-L., … Zhang,
Y.-P. (2015). Whole-genome sequence of the Tibetan frog Nanorana parkeri and
the comparative evolution of tetrapod genomes. Proceedings of the National
Academy of Sciences of the United States of America, 112(11), E1257–E1262.
doi:10.1073/pnas.1501764112
Toon, A., Austin, J. J., Dolman, G., Pedler, L., & Joseph, L. (2012). Evolution of arid
zone birds in Australia: Leapfrog distribution patterns and mesic-arid
connections in quail-thrush (Cinclosoma, Cinclosomatidae). Molecular
Phylogenetics and Evolution, 62(1), 286–295. doi:10.1016/j.ympev.2011.09.026
Tullberg, B. S., Merilaita, S., & Wiklund, C. (2005). Aposematism and crypsis
combined as a result of distance dependence: Functional versatility of the colour
pattern in the swallowtail butterfly larva. Proceedings of the Royal Society B:
Biological Sciences, 272(1570), 1315–1321. doi:10.1098/rspb.2005.3079
Turner, J. R., Johnson, M. S., & Eanes, W. F. (1979). Contrasted modes of evolution
in the same genome: allozymes and adaptive change in Heliconius. Proceedings
of the National Academy of Sciences of the United States of America, 76(4),
1924–1928. doi:10.1073/pnas.76.4.1924
Turner, T. L., Hahn, M. W., & Nuzhdin, S. V. (2005). Genomic islands of speciation
in Anopheles gambiae. PLoS Biology, 3(9), e285.
doi:10.1371/journal.pbio.0030285.st001
Twomey, E., Vestergaard, J. S., & Summers, K. (2014). Reproductive isolation
related to mimetic divergence in the poison frog Ranitomeya imitator. Nature
Communications, 5, 4749. doi:10.1038/ncomms5749
Urrego, L. E., Molina, L. A., Urrego, D. H., & Ramírez, L. F. (2006). Holocene
space–time succession of the Middle Atrato wetlands, Chocó biogeographic
region, Colombia. Palaeogeography, Palaeoclimatology, Palaeoecology, 234(1),
45–61. doi:10.1016/j.palaeo.2005.10.018
van Etten, J. (2017). R Package gdistance: Distances and routes on geographical grids.
Journal of Statistical Software, 76(13), 1–21. doi:10.18637/jss.v076.i13
Márquez et al.
57
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
Vanzolini, P. E., & Williams, E. E. (1970). South american anoles: the geographic
differentiation and evolution of the Anolis chrysolepis species group (Sauria,
Iguanidae). Arquivos de Zoologia, 19(1–2), 1–258. doi:10.11606/issn.2176-
7793.v19i3-4p125-298
Vieira, F. G., Lassalle, F., Korneliussen, T. S., & Fumagalli, M. (2016). Improving
the estimation of genetic distances from next-generation sequencing data.
Biological Journal of Linnean Society, 117(1), 139–149. doi:10.1111/bij.12511
Wang, I. J. (2013). Examining the full effects of landscape heterogeneity on spatial
genetic variation: A multiple matrix regression approach for quantifying
geographic and ecological isolation. Evolution, 67(12), 3403–3411.
doi:10.1111/evo.12134
Weir, B. S., & Cockerham, C. C. (1984). Estimating F-statistics for the analysis of
population structure. Evolution, 38(6), 1358. doi:10.2307/2408641
Widmer, A., Lötters, S., & Jungfer, K. H. (2000). A molecular phylogenetic analysis
of the neotropical dart-poison frog genus Phyllobates (Amphibia:
Dendrobatidae). Die Naturwissenschaften, 87(12), 559–562.
doi:10.1007/s001140050
Wu, C.-I. (2001). The genic view of the process of speciation. Journal of
Evolutionary Biology, 14(6), 851–865. doi:10.1046/j.1420-9101.2001.00335.x
Yang, Y., Richards-Zawacki, C. L., Devar, A., & Dugas, M. B. (2016). Poison frog
color morphs express assortative mate preferences in allopatry but not sympatry.
Evolution, 70(12), 2778–2788. doi:10.1111/evo.13079
Márquez et al.
58
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
| 2020 | Divergence, gene flow and the origin of leapfrog geographic distributions: The history of color pattern variation in poison-dart frogs | 10.1101/2020.02.21.960005 | [
"Márquez Roberto",
"Linderoth Tyler P.",
"Mejía-Vargas Daniel",
"Nielsen Rasmus",
"Amézquita Adolfo",
"Kronforst Marcus R."
] | creative-commons |
Harmoni: a Method for Eliminating Spurious
Interactions due to the Harmonic Components in
Neuronal Data
Mina Jamshidi Idajia,b,c,∗, Juanli Zhanga,d, Tilman Stephania,b, Guido
Noltee, Klaus-Robert M¨ullerc,f,g,h, Arno Villringera,i, Vadim V. Nikulina,j,k,∗∗
aNeurology Department, Max Planck Institute for Human Cognitive and Brain Sciences,
Leipzig, Germany
bInternational Max Planck Research School NeuroCom, Leipzig, Germany
cMachine Learning Group, Technical University of Berlin, Berlin, Germany
d Department of Neurology, Charit´e – Universit¨atsmedizin Berlin, Berlin, Germany
eDepartment of Neurophysiology and Pathophysiology, University Medical Center
Hamburg-Eppendorf, Hamburg, Germany
fDepartment of Artificial Intelligence, Korea University, Anam-dong, Seongbuk-gu, Seoul,
Republic of Korea
gMax Planck Institute for Informatics, Saarbr¨ucken, Germany
hGoogle Research, Brain Team
iDepartment of Cognitive Neurology, University Hospital Leipzig, Leipzig, Germany
jCentre for Cognition and Decision Making, Institute for Cognitive Neuroscience,
National Research University Higher School of Economics, Moscow, Russia
kNeurophysics Group, Department of Neurology, Charit´e-Universit¨atsmedizin Berlin,
Berlin, Germany
Abstract
Cross-frequency synchronization (CFS) has been proposed as a mechanism
for integrating spatially and spectrally distributed information in the brain.
However,
investigating CFS in Magneto- and Electroencephalography
(MEG/EEG) is hampered by the presence of spurious neuronal interactions
due to the non-sinusoidal waveshape of brain oscillations. Such waveshape
gives rise to the presence of oscillatory harmonics mimicking genuine neu-
ronal oscillations. Until recently, however, there has been no methodology
for removing these harmonics from neuronal data. In order to address this
long-standing challenge, we introduce a novel method (called HARMOnic
∗Corresponding author, jamshidi@cbs.mpg.de
∗∗Corresponding author, nikulin@cbs.mpg.de
Preprint submitted to bioRxiv
October 6, 2021
miNImization - Harmoni) that removes the signal components which can be
harmonics of a non-sinusoidal signal. Harmoni’s working principle is based
on the presence of CFS between harmonic components and the fundamen-
tal component of a non-sinusoidal signal. We extensively tested Harmoni
in realistic EEG simulations. The simulated couplings between the source
signals represented genuine and spurious CFS and within-frequency phase
synchronization. Using diverse evaluation criteria, including ROC analyses,
we showed that the within- and cross-frequency spurious interactions are sup-
pressed significantly, while the genuine activities are not affected. Addition-
ally, we applied Harmoni to real resting-state EEG data revealing intricate
remote connectivity patterns which are usually masked by the spurious con-
nections. Given the ubiquity of non-sinusoidal neuronal oscillations in elec-
trophysiological recordings, Harmoni is expected to facilitate novel insights
into genuine neuronal interactions in various research fields, and can also
serve as a steppingstone towards the development of further signal process-
ing methods aiming at refining within- and cross-frequency synchronization
in electrophysiological recordings.
1. Introduction
The importance of oscillatory neuronal activity has been demonstrated
by its association with cognitive, sensory, and motor processes in the brain
(Buzs´aki and Draguhn, 2004; Engel and Fries, 2010; Harris and Gordon, 2015;
Miller et al., 2010; Sadaghiani and Kleinschmidt, 2016). Various oscillatory
processes have to be integrated in order to support formation of behaviorally
relevant outputs based on a multitude of sensory and cognitive factors. This
neuronal integration is facilitated by complex spatial connectivity patterns
in the brain (Bullmore and Sporns, 2009; Nentwich et al., 2020). In this
context, phase-phase synchronization (PPS) has been hypothesized to repre-
sent a mechanism through which such spatially distributed information can
be integrated in the brain with a high temporal precision (Fries, 2015). Im-
portantly, PPS underlies not only spatially, but also spectrally distributed
interactions - so-called cross-frequency synchronization (CFS) (Canolty and
Knight, 2010; Jensen and Colgin, 2007; Nikulin and Brismar, 2006; Palva
et al., 2005; Palva and Palva, 2018a,b). Magneto- and Electroencephalog-
raphy (MEG/EEG) provide a unique opportunity to non-invasively study
these neuronal interactions in humans.
2
Since in the frequency domain analysis the kernel function is sinusoidal,
we often conceptualize oscillations as sinusoids. However, neural oscillations
with non-sinusoidal waveshape are abundant in human electrophysiological
recordings Cole and Voytek (2017). Such non-sinusoidality reflects complex
trans-membrane ion currents flowing though highly morphologically asym-
metric neurons (e.g. pyramidal cells) where inward and outward currents
are unlikely to balance each other with the exact temporal dynamics thus
leading to different shape of oscillations recorded with EEG/MEG/LFP (Lo-
cal field potential) (Jones et al., 2009). This ubiquity of the non-sinusoidal
waveform of brain oscillations has significant implications for the analysis of
brain connectivity.
A periodic signal can be decomposed into its harmonic components using
Fourier analysis. For the sake of clarity, we call the first harmonic the funda-
mental component and from here on by harmonics we mean the second and
higher harmonic components whose central frequencies are integer multiples
of the fundamental frequency. By band-pass filtering the signal around the
fundamental and harmonic frequencies, we can separate the respective com-
ponents, which are – by construction – CF synchronized to the fundamental
component (Hyafil, 2017; Scheffer-Teixeira and Tort, 2016). Additionally, if
the band-pass filters of the harmonics frequency are wide enough, a phase-
amplitude coupling (PAC) can be observed between the fundamental and
harmonic components (Giehl et al., 2021; Hyafil, 2017). Note that, as also
discussed in (Kramer et al., 2008), non-sinusoidal signals can be constructed
from the mixture of distinct sources with cross-frequency coupling. However,
in this work, we do not distinguish whether the non-sinusoidality originates
from signal mixing or the intrinsic waveshape of the signal. In the discussion
section, we elaborate on the effect of signal mixing.
In this manuscript, we address the effects of non-sinusoidal shape of the
brain oscillations on the observation of spurious interactions between the os-
cillatory brain activities. In spite of other spurious interactions (e.g. bias
of the data length), the spurious interactions due to the waveshape cannot
be determined by statistical methods. For example, our recently introduced
method for separating cross-frequency coupled sources cannot distinguish
sources with genuine interactions and those which are coupled because of
the higher frequency signal being the harmonic of the lower frequency one
(Idaji et al., 2020) because a harmonic-driven synchronization is not sta-
tistically distinguishable from a genuine coupling. Therefore, distinguishing
harmonic-driven and genuine interactions has currently gained more atten-
3
tion and still remains as a major challenge in the MEG/EEG connectivity
research (Giehl et al., 2021; Scheffer-Teixeira and Tort, 2016; Siebenh¨uhner
et al., 2020). The main reason of this challenge is that the connectivity anal-
ysis of MEG/EEG data is typically done using band-pass filtering, which
separates the fundamental and harmonic components of an oscillatory activ-
ity with a non-sinusoidal waveform. As a result, the observed within- and
cross-frequency synchronization between the components in the frequency
bands of the fundamental and harmonic frequencies can be mistakenly inter-
preted as genuine interaction. Figure 1 shows a schematic example where two
non-sinusoidal signals are synchronized. This coupling should be manifested
in the synchronization of the fundamental components, while the harmonic
components shape the waveform of the individual signals. However, the har-
monic components are also spuriously synchronized and additional CFS is
observed between and within the regions. Since these interactions (shown in
dashed lines in figure 1-B) are observed due to the waveform of the individual
signals, they are referred to as spurious, in contrast to genuine interactions.
The omnipresence of these spurious interactions in all human MEG/EEG
recordings makes the validity of the previously studied within- and cross-
frequency connectivity maps ambiguous.
There has been an attempt from Siebenh¨uhner et al. (2020) to discard
the potentially spurious connections from cross-frequency (CF) connectiv-
ity graphs based on the detection of ambiguous motifs in the connectivity
graphs. In that work, any CFS connection forming a triangle motif with
the local CFS and within-frequency inter-areal phase synchronization is con-
sidered as ambiguous and is discarded. However, such an approach cannot
disentangle the within-frequency spurious interactions in the harmonic fre-
quency bands, and is specific to the CF connectivity graphs. Furthermore,
this approach cannot distinguish cases of genuine couplings which form an
ambiguous motif. A more attractive approach, however, would remove or
suppress the data components that can be associated with the harmonics of
the periodic neuronal activity. Such an approach can provide the opportu-
nity of using the cleaned narrow-band data (in the frequency range of the
harmonics) for within-frequency and cross-frequency connectivity analyses.
In the current work, we introduce a novel, first-of-its-kind method for re-
moving effects of harmonics on the estimation of within- and cross-frequency
synchronization. Our method, called HARMonic miNImization (Harmoni),
is (to the best of our knowledge) the first existing signal processing tool for
suppressing higher harmonic components of a periodic signal, without band-
4
N1
N2
N1
N2
Beta Network
N1
N2
Alpha Network
Genuine alpha interaction
Spurious beta interaction
Spurious CF interaction
A
B
Figure 1: How non-sinusoidal shape of the neuronal oscillations impacts the connectivity
of brain regions. Panel A shows two non-sinusoidal oscillations with their fundamental
frequency in the alpha band. The second harmonics of these signals are located in the
beta band. As a byproduct of the coupling of the fundamental alpha components (the
solid line in panel B), the second harmonics are also coupled to each other, which results
in spurious interactions within the beta band (the dashed line in panel B) and across the
two frequency bands (dotted lines in panel B).
5
stop filtering or rejecting non-sinusoidally shaped signal components using
ICA or any other multi-variate decomposition.
We extensively tested Harmoni with realistic EEG simulations and show
that the spurious interactions are alleviated significantly, while the genuine
activities are not affected. Harmoni is then applied to resting-state EEG
(rsEEG) data and we show that the CFS connections mimicking genuine
interactions are suppressed, while many masked remote interactions are re-
covered.
2. Materials and Methods
2.1. Phase-Phase Synchronization
Phase-Phase Synchronization (PPS) can be defined for within-frequency
as well as for cross-frequency (CF) interactions. In order to define the within-
and cross-frequency synchronization indices, assume two complex narrow-
band signals x(t) = ax(t)ejφx(t), y(t) = ay(t)ejφy(t) ∈ C with central frequen-
cies fx and fy, respectively. Here, by narrow-band complex signal we mean
the analytic signal built using the Hilbert transform. Formally, if xH(t) is the
Hilbert transform of a narrow-band real signal xR(t) = ax(t) cos (φx(t)), then
x(t) = xR(t) + jxH(t) is the analytic signal of xR(t). In these formulations
the index R indicates that the signal is real valued and the index H denotes
a Hilbert transformed signal. Note that, another way to get the narrow-band
complex signals from a broad-band signal is complex wavelet transforms.
If fx = fy then x(t) and y(t) are two narrow-band signals in the same
frequency band. Their complex-valued coherence coh(x, y) ∈ C can be com-
puted from the following equation:
coh(x, y) = < ax(t)ay(t)ejφx(t)−jφy(t) >
�
< ax(t)2 >< ay(t)2 >
(1)
where < . > is the averaging operator over time and j = √−1 is the imagi-
nary number.
We use the absolute of the imaginary part of coherence (iCoh) (Nolte
et al., 2004) for estimating the connectivity between two signals in the same
frequency band. This prevents a lot of the within-frequency spurious inter-
actions due to signal mixing and volume conduction in EEG.
If nfx = mfy for m, n ∈ N, the cross-frequency synchronization (CFS,
known as m:n synchronization) of x(t) and y(t) can be quantified by m:n
6
absolute coherence cohm:n(x, y) ∈ R defined by the following equation:
cohm:n(x, y) = | < ax(t)ay(t)ejnφx(t)−jmφy(t) > |
�
< ax(t)2 >< ay(t)2 >
(2)
which is in principle similar to m:n phase locking value as:
plvm:n = | < ejnφx(t)−jmφy(t) > |
(3)
with the difference that in equation 2 the amplitudes of the signals are taken
into account and the phase estimations during higher amplitudes are weighted
higher. Giehl et al. (2021) have used a variant of equation 2. Equation 2
reduces to the absolute part of equation 1 for m = n = 1. In this work,
we are specifically interested in the case that m = 1 and n > 1, i.e. when
x(t) is a signal with central frequency fx and y(t) is a faster oscillation with
the central frequency fy = nfx.
In this case, coh1:n(x, y) = |coh(xn, y)|,
where xn(t) = ax(t)ejnφx(t) is built by multiplying the phase of x(t) by n, i.e.
accelerating x(t) by a factor of n.
CFS as defined by equation 2 has a real value between 0 and 1, with 0
corresponding to the lack of any phase synchronization between two com-
pletely independent signals and 1 for two perfectly synchronized time-series
with the same amplitude envelope.
2.2. Genuine vs. spurious interactions
The PPS and CFS indices of equations 1 and 2 have a bias based on
the length of the data time-series, i.e., two band-pass filtered random time-
series also show a value larger than 0. Therefore, a test of significance is
necessary for phase synchronization measures (Scheffer-Teixeira and Tort,
2016) in order to distinguish such spurious interactions when the data length
is not sufficient.
Another type of spurious interactions (which is not statistically discernible
from real interactions) is the interactions due to the waveshape of brain
signals. The reason is that harmonic components of a signal with a non-
sinusoidal shape have CFS to each other. As an illustrative example, figure
2 depicts a sawtooth-shaped signal and its fundamental and 7th harmonic
components. The 7th harmonic of this sawtooth-shaped signal has an almost
perfect 1:7 synchronization to the fundamental frequency (coh1:7 = 0.99).
Additionally, although it is not the focus of this manuscript, it is interesting
to note that when a non-sinusoidally shaped signal (here sawtooth-shaped)
7
6
12 18 24 30 36 42 48
Frequency (Hz)
FFT Magnitude
coh1:7=0.71
PAC=0.68
coh1:7=0.99
PAC=0.00
100 ms
main signal
1st harmonic
[5 - 7] Hz
7th harmonic
7th harmonic
[35 - 49] Hz
[41 - 43] Hz
Figure 2: A simulated sawtooth-shaped signal with the fundamental frequency equal to 6
Hz is depicted in the first row and the fundamental 6 Hz component (i.e. the 1st harmonic)
is shown in the second row. The 7th harmonic component filtered at a frequency window
with width of 2Hz is illustrated in third row. Additionally, the sawtooth signal was filtered
around the 7th harmonic frequency with a window size of 7Hz, depicted in the fourth row.
The magnitude of the fast Fourier transform (FFT) of each signal is depicted at its left side.
The CFS and PAC between the fundamental component and the two components with
central frequency of the 7th harmonic frequency are noted along the right side vertical lines.
The 7th harmonic on the third row shows a strong 1:7 synchronization to the fundamental
component (coh1:7 = 0.99) and no PAC. However, if filtered at a wider frequency band,
the harmonic component shown on the fourth row shows also a PAC with the fundamental
component. Note that the amplitude of the signals and their FFT magnitudes are scaled
arbitrarily for the sake of better illustration.
is filtered in a wider frequency range around the harmonic frequency, PAC
is observed between the harmonic and fundamental frequencies (in addition
to CFS). In this paper, however, our focus is on the n:m synchronizations.
The example of figure 2 shows that by band-pass filtering a single process
one can observe cross-frequency coupling between its different components,
although these components still represent the same complex signal. In the lit-
erature of cross-frequency coupling (Hyafil, 2017; Scheffer-Teixeira and Tort,
2016; Siebenh¨uhner et al., 2020; Giehl et al., 2021), such a coupling between
the components of a single process, or generally an interaction between two
signals where at least one of them is a higher harmonic of a non-sinusoidal
8
process is called spurious. This is usually in contrast to genuine interactions
between two signals representing two distinct processes where none of them
is a higher harmonic of a periodic signal. Formally, let x(t) = �
i x(i)(t)
and y(t) = �
i y(i)(t), i ∈ N be two n:m synchronized periodic oscillatory
processes, where x(i) and y(i) are the i-th harmonic components of x(t) and
y(t), respectively. The fundamental components (x(1) and y(1)) and higher
harmonics (x(i) and y(i) for i > 2) of each of these signals can be separated
from each other by band-pass filtering x(t) and y(t). The synchronization of
x and y implies that for any i1, i2 ∈ N, x(i1)(t) and y(i2)(t) are cross-frequency
synchronized. When assessing the synchronization of the narrow-band sig-
nals, we consider only the synchronization of fundamental components x(1)
and y(1) genuine. The synchronization of x(i1)(t) and y(i2)(t) for i1 > 1 or
i2 > 1 is harmonic-driven and is called spurious. Note that this does not
mean that the signal components are not synchronzed and the synchroniza-
tion value is non-zero because of insufficient number of data points or due
to filtering. By spurious interactions due to waveshape it is meant that any
coupling including higher harmonics is in fact mediated by the fundamen-
tal component of the respective non-sinusoidal signal. Figure 1 illustrates
various possible within- and cross-frequency spurious synchronizations due
to waveshape. In the next section we introduce an original signal processing
method for suppressing the harmonic-driven synchronizations in connectivity
analyses using electrophysiological data.
A final important note is that, as discussed in (Kramer et al., 2008), a
non-sinusoidal signal can be constructed from the mixing of distinct sources
with CFS or PAC. This is actually a major concern in electrophysiological
research even outside of connectivity topic. Although we do not account for
this issue in our analyses explicitly, we discuss it in the discussion section,
“Harmoni and signal mixing”.
2.3. HARMOnic miNImization (HARMONI)
Assume that z(t) = s(t) + ϵ(t), where s(t) is a a periodic signal with
the fundamental frequency of f0. ϵ(t) is additive noise or any other pro-
cess such as another oscillatory activity mixed with s(t). Harmoni aims at
removing the components of z(t) within a narrow frequency band around
nf0, n ∈ N, n ≥ 2 that have similar phase profile as the fundamental com-
ponent of s(t). For this purpose, we can write z(t) = xR(t) + yR(t) + ξ(t),
where xR(t) = ax(t) cos (φx(t)) and yR(t) = ay(t) cos (φy(t)) are the real-
valued contents (indicated by the index R) from frequency bands f0 and nf0,
9
0
10
20
30
40
50
PSD (dB)
Frequency (Hz)
Minimization
phase synchrony
+
PSD (dB)
10
30
20
40
Frequency (Hz)
2nd harmonic
fundamental frequency
coherence=0.47
coherence=0.07
noise
non-sinusoidal
signal
noisy non-sinusoidal signal
Accelerate by
a factor of 2
Hilbert
Hilbert
Figure 3: Harmoni is a method that removes
harmonics of a non-sinusoidal signal. The in-
puts are the band-pass filtered signals in the
frequency bands of the fundamental and har-
monic frequencies.
In this figure, the signal
is a non-sinusoidal alpha rhythm with fun-
damental and second harmonic frequencies of
10Hz and 20Hz, respectively. The band-pass
filtered signals at 10Hz and 20Hz are used as
inputs to the minimization block, which runs a
regression-like algorithm to find the best mul-
tiplier for removing the harmonic parts of y(t).
This is done by means of subtracting a scaled
version of xn(t) from y(t), where xn(t) is an
accelerated version of x(t) by multiplying its
phase by a factor of n (here n = 2).
The
output of Harmoni is a band-limited signal in
the harmonic frequency band (here 20Hz - the
second harmonic) where the harmonic compo-
nent is removed.
10
respectively. ξ(t) represents all other components of z(t) except xR(t) and
yR(t). Therefore, xR(t) and yR(t) are estimated using band-pass filtering z(t)
within the respective frequency bands of the fundamental and harmonic fre-
quencies. We define x(t) and y(t) as the analytical signals of xR(t) and yR(t)
built using the Hilbert transform and work with them in the next steps of
Harmoni. Note that x(t) and y(t) can be also generated by applying complex
wavelet transforms to z(t).
The fundamental component of a non-sinusoidal signal has 1:n synchro-
nization to its n-th harmonic component. Therefore, the phase information
of the harmonic components can be recovered from the phase of the funda-
mental component. Using x(t), Harmoni tries to remove the parts of y(t) that
are 1:n coupled to x(t), or equivalently 1:1 coupled to xn(t) = ax(t)ejnφx(t).
As mentioned above, the part of y(t) which is a harmonic of a component
in x(t) should be phase synchronized to xn(t). Therefore, we estimate the
harmonic part of y by λxn(t), λ ∈ C. ycorr(t) = y(t) − λxn(t) contains the
non-harmonic components of y(t), where ycorr(t) has a minimum possible
within-frequency synchronization to xn(t). The complex multiplier λ = cejφ
is estimated through the following optimization problem:
min
c,φ |coh
�
y(t) − λxn(t), xn(t)
�
|
for
λ = cejφ
Here, the phase of λ compensates the possible phase difference between the
harmonic and fundamental components. Figure 3 shows a schematic block
diagram of Harmoni.
Practically, we perform a grid-search procedure for
computing λ = cejφ, which is presented in algorithm 1. In practice, in a
connectivity pipeline, the activity of each brain site - that can be a region-
of-interest (ROI) or an electrode - is band-pass filtered within the two bands
of interest, namely f0 and nf0. Then Harmoni is applied on the data of each
sensor or ROI. In the next section, it will be described in detail how Harmoni
can be used in a connectivity analysis pipeline with electrophysiological data.
2.4. Connectivity pipeline in source space
Figure 4 shows a block-diagram of a connectivity pipeline, also imple-
menting Harmoni.
The first step is to band-pass filter the multi-channel
data within the frequency bands of interest f0 and nf0. For instance, if we
are interested in alpha and beta band, f0 = 10 and n = 2. Below, we will
elaborate upon the next steps.
11
2.4.1. Forward and inverse solutions
We used fsaverage standard head model and the three-layer boundary
element model (BEM) accompanied with MNE Python (Gramfort et al.,
2013, 2014). 64 electrodes (or a subset of it) with positions according to
the BioSemi cap were used and aligned to the MRI coordinates.
MNE-
Python was used to create a dipole grid on white matter surface with oct6
spacing between the grid points, resulting in 4098 sources per hemisphere.
The surface-based source space and the BEM solutions were then used for
computing a forward solution. An inverse solution with dipole directions nor-
mal to the cortical surface was computed with eLORETA inverse modelling
(Pascual-Marqui, 2007) with the regularization parameter equal to 0.05, and
the noise covariance equal to the covariance of 64 white-Gaussian signals with
equal duration to the data, which is an estimation of the identity matrix.
2.4.2. From sensor space to ROIs
The band-pass filtered multi-channel EEG data were projected to the
cortical surface using the computed inverse solution, resulting in ∼8000 re-
constructed surface sources. These sources were then grouped based on an
atlas into regions of interest (ROIs). We used the Desikan Killiany atlas with
68 ROIs (Desikan et al., 2006) for simulations and Schaefer atlas with 100
ROIs (Schaefer et al., 2018) for real data analysis. Singular value decom-
position (SVD) was then applied to the band-pass filtered time-series of the
sources of each ROI and a single time-series was computed per ROI. As a
result, the ∼8000 reconstructed cortical sources were translated to nROI ROI
times-series in each frequency band (here: nROI=number of ROIs in the used
atlas), which are ready for connectivity computations.
2.4.3. Harmoni
Although the ROI time series can be directly used for computing the
connectivity maps, we suggest to use Harmoni as an intermediate step in a
connectivity pipeline. Harmoni is applied on the signals of each ROI in the
two frequency bands of interest centered at f0 and nf0, which correspond
to the fundamental and the n-th harmonic frequencies. The output of the
algorithm is a signal in the frequency band of nf0 for which the harmonic
components are suppressed to a large extent. The ROI time series at f0 and
the Harmoni-corrected signals at nf0 are then passed to the next step for
computing the within- and cross-frequency synchronization maps.
12
nf0
f0
Band-pass Filtering
Multi-channel Data
Inverse-modelling
ROI time series
Connectivity
within-frequency
cross-frequency
ROI time series at f0
ROI time series at nf0
ROI-ROI
H
H
H
H
nf0 signals after Harmoni
H
band-pass Filter
Harmoni Algorithm
f0
centered at f0
Figure 4: The block-diagram of Harmoni pipeline in source space.
The multi-channel
signal is first band-pass filtered in the range of the fundamental frequency (f0) and the
harmonic frequency of interest (nf0). The narrow-band signals are mapped to the cortical
surface using the inverse solution and the ROI time series are extracted. The ROI signals
in the range of harmonic-frequency are then corrected with Harmoni and the potential
harmonic components are removed.
Finally, the ROI-ROI within- and cross-frequency
connectivity maps are computed. In this paper, without loss of generality and due to the
better SNRs, we set f0 = 10 and n = 2.
2.4.4. From ROIs’ time-series to connectivity maps
For both of the simulations and real data, after computing the ROI time
series and applying Harmoni on them, we computed a connectivity index
for each pair of the ROIs, resulting in an nROI × nROI graph. For within-
frequency connectivity (here in alpha and beta bands), we used the absolute
of imaginary part of coherence (iCoh) computed from the imaginary part of
equation 1 and for the cross-frequency synchronization we used the extension
of coherence for n:m coupling as in equation 2.
2.5. Simulations
2.5.1. Signals and SNR
The pipeline for producing signals and the definition of signal-to-noise
ratio (SNR) are similar to that of (Idaji et al., 2020). In this section we de-
scribe the procedure of simulating the signals and how SNR is defined in our
simulation pipelines. Note that in all places, band-pass filtering was carried
out using fourth-ordered Butterworth filters designed for the frequency band
of interest. The filtering was applied forward and backward in order to avoid
phase shift in data.
Additive noise: The time-series of the noise sources were produced with
the colornoise package (Patzelt, 2019) in Python by building a random signal
with a 1/f (pink) spectrum from a random white Gaussian noise.
13
Sinusoidal oscillations: Without loss of generality, in our simulations,
all of the time-series of the sinusoidal oscillatory sources were simulated in
alpha (8-12 Hz) and beta (16-24 Hz) frequency bands.
Independent sources (those without a synchronization to other source
signals) were generated by band-pass filtering white Gaussian noise in the
frequency band of interest. The analytic signals of these oscillations were
built using the Hilbert transform of them. For instance, if xR(t) is an alpha
oscillation produced by band-pass filtering white Gaussian noise within (8-
12) Hz and xH(t) is the Hilbert transform of xR(t), x(t) = xR(t) + jxH(t) is
the analytic signal of xR(t).
A source signal y(t) with 1:n synchronization to an oscillation x(t) was
simulated by phase-warping of x(t), i.e.:
x(t) = ax(t)ejφx(t)
y(t) = ay(t)ejnφx(t)+jφ0
(4)
where x(t) ∈ C is the analytic signal of an oscillation generated by band-pass
filtering white Gaussian noise around f0, y(t) ∈ C is the analytic signal of
an oscillation within a frequency band around nf0 and 1:n synchronized to
x(t), and φ0 is the phase difference of the two signals taken randomly from
a uniform distribution between [−π/2, π/2]. ay(t) is either equal to ax(t)
or equal to the envelope of another band-pass filtered white-Gaussian signal
in the same frequency band as y(t). For instance, if x(t) is an alpha band
oscillation and n = 2, y(t) is a beta band oscillation and 1:2 synchronized to
x(t). If ax(t) = ay(t), the 1:n synchronization of these two signals computed
from equation 2 is equal to 1. Note that in the case of ax(t) ̸= ay(t), the
interaction of x and y is for sure genuine. Therefore, for the simulation of
two genuinely (cross-frequency) synchronized sources, we used ax(t) ̸= ay(t).
The power of each oscillation is scaled based on the signal-to-noise (SNR)
ratio of the frequency band of interest (see below).
Non-sinusoidal oscillations: A non-sinusoidal signal s(t) = �
n s(n)(t),
n ∈ N with the fundamental frequency of f0 was generated by adding up
its fundamental component (or the first harmonic) s(1)(t) and the higher
harmonics components s(n)(t), n ≥ 2. In the following equations, s(1)(t) is an
oscillation at f0 produced by band-pass filtering a white Gaussian noise signal
and s(n)(t), n ≥ 2 is a 1:n synchronized oscillation produced by equation 4 to
14
be 1:n synchronized to s(1).
s(t) =
�
n
s(n)(t), n ∈ N
s(1)(t) = Re
�
a1(t)ejφ(t)�
s(n)(t) ∝ Re
�
a1(t)ejnφ(t)+jφn, n ≥ 2
�
(5)
where φn, n ≥ 2 are random numbers taken from a uniform distribution
between [−π/2, π/2].
Given a fundamental frequency of f0, let s1(t) = �
n s(n)
1 (t) be a simulated
non-sinusoidal oscillation based on equation 5 and s(1)
1 (t) = a1(t) cos
�
φ(t)
�
.
The following equations show how another non-sinusoidal signal s2(t) is sim-
ulated to be synchronized to s1(t):
s2(t) =
�
n
s(n)
2 (t), n ∈ N
s(1)
2 (t) = Re
�
a2(t)ejφ(t)+jψ1�
s(n)
2 (t) ∝ Re
�
a2(t)ejnφ(t)+jψn�
, n ≥ 2
(6)
where ψn, n ∈ N are random numbers taken from a uniform distribution
between [−π/2, π/2]. In equation 6, s(1)
2
is an oscillation with 1:1 synchro-
nization to s(1)
1
Note that the second harmonic is the strongest harmonic which is gen-
erally visible in real electrophysiological data.
Therefore, without loss of
generality, we only examine the removal of the second harmonic.
There-
fore, we simulated only the fundamental and the second harmonic. That
is, in our simulations, the non-sinusoidal source signals are simulated as
s(t) = s(1)(t) + s(2)(t) where s(1)(t) is an alpha oscillations and s(2)(t) is
the second harmonic in beta frequency band. After that, the amplitude of
s(1)(t) and s(2)(t) were re-scaled so that the SNR at each of alpha and beta
frequency bands for these signals are set to the desired value (see below).
Finally, s(1)(t) and s(2)(t) are added up together to generate s(t).
SNR: In realistic simulations, The SNR was defined as the ratio of the
mean power of the source signal in the sensor space divided by the mean
power of all pink noise sources in sensor space, filtered in the frequency band
of interest. In our realistic simulations, the SNR of alpha and beta bands
were set to 0dB and −10dB respectively.
15
For the toy examples, the SNR of a narrow-band source was defined as
the ratio of its power to the power of the pink noise, filtered in the frequency
band of interest. The SNR values at alpha and beta band were set to 5 dB
and −5 dB respectively.
2.5.2. Toy Examples
We used toy examples for initial assessment of the effect of Harmoni
on the interactions between two signals with non-sinusoidal components. We
used four scenarios for these toy examples, where the ground truth about the
existing genuine and spurious interactions between the simulated signals were
pre-defined. The left side of figure 5 depicts these scenarios schematically.
In each of the four scenarios, two signals zk(t), k = 1, 2 were simulated.
On the schemes of figure 5, z1(t) and z2(t) are depicted as shaded areas
in each scenario. In the rest of this section, the index k = 1, 2 refers to
these two signals. z1(t) and z2(t) were multi-band signals with components
in alpha and beta bands. In each scenario, specific ground truth genuine
interactions were simulated between the two signals, which produced known
spurious interactions, too. Harmoni was applied on each of the signals in
order to remove the beta-component which could be the harmonic component
of the alpha band component of the signal. The interactions between the two
signals were estimated using absolute within- and cross-frequency coherence
before and after Harmoni. We expected that Harmoni suppresses the spurious
interactions, but does not touch the genuine interactions. For each scenario,
50 runs with random seeds were carried out.
In all scenarios, the two signals z1(t) and z2(t) contained an alpha os-
cillation with non-sinusoidal waveshape. sk(t) = αk(t) + βk(t) is the non-
sinusoidal component of zk(t), where αk(t) represents the fundamental com-
ponent and βk its second harmonic, which is phase-synchronized to αk(t).
Below, the composition of z1 and z2 in all the four scenarios and their
genuine and spurious interactions are listed. Note that ξk(t) is the additive
1/f (pink) noise component of zk(t).
Scenario 1 (figure 5-A): zk(t) = sk(t) + ξk(t), k = 1, 2. The signal s1
was simulated using equation 5 and s2 was simulated to be synchronized to
s1 using equation 6. Therefore, a genuine interaction in alpha band between
the two signals was simulated. Additionally, a spurious interaction in beta
band, as well as spurious cross-frequency interactions between the two signals
were observed in the ground truth. Figure 15 shows exemplar signals of this
scenario.
16
A: Scenario1
B: Scenario 2
C: Scenario 3
D: Scenario 4
N1
N2
B1
A1
N1
N2
B1
B2
A1
A2
Beta Network
N1
N2
Alpha Network
N1
N2
A1
A2
Beta Network
B1
B2
N1
N2
Alpha Network
N1
N2
CFC Network
N1
N2
CFC Network
N1
N2
B1
A1
Toy Examples
Realistic Simulations
Scenario1
Scenario 2
Figure 5: Simulation scenarios. Toy examples: Two signals z1 and z2 were simulated
for each scenario, where various genuine and spurious synchronizations are present in
the ground truth. The solid lines show the simulated, genuine synchronizations, and the
dashed lines depict the spurious interactions observed in the ground-truth. Harmoni was
applied on each of the signals and the within- and cross-frequency synchronization for alpha
and beta bands were examined before and after Harmoni. In all scenarios, zk contained a
non-sinusoidally shaped component sk = αk + βk, where αk and βk are the fundamental
and second harmonic components of sk respectively. ˘βk, k = 1, 2 in scenarios 2 to 4 are
beta oscillations independent of sk, k = 1, 2 Realistic simulations: In the first row, each
dot shows a source and the connecting lines represent the synchronization of the source
signals. The sources with purple color and the letter N correspond to sources with non-
sinusoidal alpha oscillations having components in both alpha and beta frequency bands.
The blue color and letter B corresponds to sinusoidal beta band sources, and the red color
and letter A represent sinusoidal alpha frequency range sources. In the schematic brains
of rows 2 to 4, the ground truth alpha, beta, and CFS networks are depicted. While solid
lines depict genuine interactions, dashed lines show spurious interactions caused by non-
sinusoidal waveshape of the signals. In both of the toy examples and realistic simulations,
the main purpose of Harmoni is to suppress the spurious (dashed-line) connections, while
not affecting the genuine (solid-line) interactions.
17
Scenario 2 (figure 5-B): zk(t) = sk(t) + ˘βk(t) + ξk(t), k = 1, 2. s1 and
s2 were simulated as synchronized non-sinusoidal signals using equations 5
and 6 (similar to scenario 1). Each signal zk had an extra beta component
˘βk.
˘β1 and ˘β2 were simulated as narrow-band beta band oscillations and
synchronized to each other (with equation 4) but independent of sk, k = 1, 2.
In addition to the genuine integration between the z1 and z2 in beta band due
to the synchronization of ˘β1 and ˘β2, similar genuine and spurious interactions
as in scenario 1 were present in the ground truth. In figure 15 15 an example
of signals of this scenario is depicted (at the end of the manuscript).
Scenario 3 (figure 5-C): zk(t) = sk(t) + ˘βk(t) + ξk(t), k = 1, 2. s1 and
s2 were two independent non-sinusoidal oscillations (using equation 5) with
their fundamental and second harmonic components in alpha and beta band
respectively. ˘β1 and ˘β2 were two synchronized narrow-band beta oscillations
(using equation 4), which were independent of s1 and s2. As a result, no
CFS existed between z1 and z2 in the ground truth and the only genuine
interaction was a synchronization within beta band.
Scenario 4 (figure 5-D): zk(t) = sk(t) + ˘βk(t) + ξk(t), k = 1, 2. s1 and
s2 were two non-sinusoidal alpha oscillations simulated independently using
equation 5, and ˘β2 was a narrow-band beta oscillation 1:2 synchronized to
s1, i.e. ˘β2 was simulated to have 1:2 CFS to the alpha component of s1 (α1)
using equation 4. Therefore, in addition to the genuine CFS between z1 and
z2, a spurious synchronization within beta band between z1 and z2 existed
in the ground truth (i.e. between ˘β2 and β1). ˘β1 was a narrow-band beta
oscillations independent of s1, s2, and ˘β2.
Note that since there is no mixing between z1 and z2 in these simulations,
the absolute coherence was used for quantifying both the within- and cross-
frequency synchronizations.
2.5.3. Realistic simulations
Source positions. The oscillatory sources were located at the center of ran-
domly selected ROIs.
Additionally, the position of 50 pink noise sources
were selected randomly from the ∼8000 nodes of the source space grid. The
Desikan Killiany (DK) atlas was used.
Scalp EEG generation.. In order to generate the realistic multi-channel EEG
signal, oscillatory and noise signals in source space were mapped to the sensor
space using the forward solution with 64 electrodes according to BioSemi
EEG cap layout. 200 datasets were simulated by using random seeds.
18
Realistic simulation scenarios. The two scenarios depicted on the right side
of figure 5 were used for simulating realistic EEG data.
In scenario one, a pair of interacting non-sinusoidal source signals were
simulated using equations 5 and 6 with their fundamental frequency in alpha
band. Additionally, a pair of coupled sources in the beta band were generated
using equation 4 and n = 1. A pair of synchronized sinusoidal sources in
alpha band were simulated as well, by using equation 4 and n = 1.
In scenario 2, a pair of genuinely cross-frequency synchronized sources
were simulated using equation 4 with n = 2. In addition, a pair of synchro-
nized non-sinusoidal source signals were generated using equations 5 and 6.
Connectivity. The connectivity pipeline explained in detail above (also fig-
ure 4) was then applied to the simulated EEG data. As depicted in figure
5, each of these two scenarios include genuine and spurious interactions in
their ground-truth. By using Harmoni, we expect to suppress the spurious
interactions.
Evaluation criterion: ROC curve. Since the computed connectivity maps are
not binary values (while the ground truth connectivity is binary), we evaluate
the matching of computed connectivity maps and the ground truth using the
area under curve (AUC) of the receiver operating characteristic (ROC) curve
of the computed connectivity matrix. Figure 6 shows how true positive and
false positive values are computed. After thresholding the test graph (T)
with threshold level 0 ≤ p ≤ 1 (resulting in Tp), The true positive ratio
(TPR) and false positive ratio (FPR) corresponding to this threshold value
are computed as TPR(p) = Σi,jGijTp,ij
Σi,jGijTij
and FPR(p) = Σi,j∼GijTp,ij
Σi,j∼GijTij
, where
the subscripts ij indicates the (i, j)-th element of the adjacency matrix and
G is the ground-truth connectivity matrix. ∼G is the the 1’s complement of
G (i.e., all zeros are converted to 1 and vice-versa).
Using the TPR and FPR values for all the threshold level, an ROC curve
is built. The AUC of this curve reflects how well the computed connectivity
map matches the ground truth adjacency matrix of the graph corresponding
to the simulated connectivity.
The AUC of the ROC curve (AUC-of-ROC) was computed for each simu-
lation run before and after Harmoni and compared. We expected an increase
of AUC-of-ROC after Harmoni.
Additionally, for graphs where no true positives were expected (for exam-
ple the CFS network of scenario 1 or beta-band network of scenario 2) the
19
True Positive Ratio
False Positive Ratio
threshold p
AUC
TPR(p)
FPR(p)
B
Threshold
at p
p
0
1
Edge strength
1
0
Test Graph
Thresholded Test Graph
Mask
Ground Truth
Tp*G
Tp*~G
+
True Positive
TP(p)
False Positive
FP(p)
G
+
T
Tp
Figure 6: AUC of an ROC curve as an evaluation criterion for assessing the matching of
computed connectivity graphs and the ground truth ones. Panel A shows an exemplar
ROC curve.
In panel b, the procedure of computing the true positive (TP) and false
positive (FP) values corresponding to threshold level 0 ≤ p ≤ 1 is depicted. The true
positive ratio (TPR) and false positive ratio (FPR) corresponding to each threshold level
p is computed by TPR(p) = Σi,jGijTp,ij
Σi,jGijTij
and FPR(p) = Σi,j∼GijTp,ij
Σi,j∼GijTij
. The ij index
indicates the (i, j)-th element of the indexed matrix.
FPR curve was built as a curve of FPR vs. threshold. The AUC of this curve
(AUC-of-FPR) is a proxy of the amount of false positives. We expected a
drop of AUC-of-FPR after Harmoni.
2.6. Resting-state EEG
2.6.1. Data description
The resting-state EEG data from 81 subjects (20-35 years old, male,
right-handed) of an open-access database (LEMON) were used (Babayan
et al., 2019). The LEMON study was carried out in accordance with the
Declaration of Helsinki and the study protocol was approved by the ethics
committee at the medical faculty of the University of Leipzig.
The data
of each subject included 16 min resting-state recording with interleaved, 1-
min blocks of eyes-closed and eyes-open conditions. For this manuscript, we
used the data of the eyes-closed condition. The recordings were done with
a band-pass filter between 0.015 Hz and 1 kHz and a sampling rate of 2500
20
Hz.
For our analysis, we used the publicly available preprocessed data in the
database. The sampling rate was reduced to 250 Hz and the down-sampled
data were filtered within [1, 45] Hz with a fourth order Butterworth filter,
applied forward and backward. Then the data segments of eyes-open and
eyes-closed conditions were separated. Bad segments were removed manually
and ICA artifact rejection was employed to remove the noise components
relating to eye, heart, and muscle activity. Babayan et al. (2019) provide
detailed information about the data recording and preprocessing steps.
2.6.2. Connectivity
The pipeline in figure 4 was used, as simular to the simulated data connec-
tivity. Fourth-order Butterworth filters (applied forward-backward to avoid
phase shift) were used for filtering data in alpha band (8-12 Hz) and beta
band (16-24 Hz).
Similar to the connectivity pipeline described in detail
above (also figure 4), the band-pass filtered data were then projected onto
cortical source space using the inverse solution computed from fsaverage stan-
dard head, with 4098 vertices per hemisphere. Afterwards, a single time se-
ries was extracted (using SVD) for each ROI from the cortical sources within
that ROI. The Schaefer atlas (Schaefer et al., 2018) with 100 ROI and 7 Yeo
resting-state networks (Yeo et al., 2011) was used.
For each subject, the ROI-ROI connectivity for alpha-beta CFS was com-
puted before and after Harmoni, resulting in 100×100 connectivity adjacency
matrices. In order to make the connectivity graphs comparable before and
after Harmoni at the group level, the adjacency matrix of each subject was
z-scored before and after Harmoni. The z-scored matrices of the networks
before Harmoni were subtracted from the ones after Harmoni. Two-sided
paired t-tests was used for each connection to specify the links which were
changing significantly on group level. The Bonferroni method was used to
correct for multiple comparisons, i.e. the p-values were multiplied by 1002
and then the links with corrected pvalues > 0.05 were considered as signifi-
cant.
Asymmetry-index of CFS networks. In order to quantify the extent to which
the CFS adjacency matrices are asymmetric, we used the norm of the anti-
symmetric part of the adjacency matrix. For a given matrix A, the antisym-
metric part is defined as Aanti = 1
2(A − AT). We define
√
2||Aanti||/||A||
as an asymmetry-index. It can be shown that this index is between zero and
21
one, with zero value corresponding to a symmetric matrix and a value of one
for an antisymmetric matrix.
2.7. Depiction of CFS connectivity
We used a bipartite graph for the depiction of CFS networks. The CFS
networks have an asymmetric adjacency matrix and therefore, should be
depicted as directed graphs. We actually used a bipartite graph as a way of
illustrating a directed graph in a more comprehensive way.
A bipartite graph is a graph which has two sets of nodes and an edge
can only connect the vertices from different sets (i.e. alpha and beta sets
in our analysis) to each other.
In our case of CFS networks, each node
is a representative of a brain region and each set of nodes relates to the
activity of the brain regions in one of the frequency bands. Figure 7 shows
an illustrative example of such depiction for alpha-beta CFS. The upper and
lower node-sets represent the alpha and beta band activity of the ROIs of
interest, respectively. A link between node 1 from the upper set (alpha nodes)
with node 3 of the lower set (beta nodes) shows a CFS coupling between ROI
1 and 3. This connection would be the element (1,3) of the adjacency matrix
of the network. In a directed graph this edge would be an out-going edge for
node 1 and an in-coming edge for node 3.
In our illustration of the graph, each node can have a color, which shows
its centrality value. In this work, we did not use this feature and the node
colors are the label colors provided with the parcellation. For real data these
colors code the ROI’s Yeo resting-state network. Each edge is also color-
coded with the strength of the coupling that it represents. It can be the
absolute or relative strength of coupling.
2.8. Statistical Analysis
Two-sided paired t-tests were used for testing the difference of the mean
value of two paired samples. Specifically, the changes of the evaluation pa-
rameters in simulations (the AUC values) as well as real data (the change
in the connectivity values and the asymmetry-index) were tested before and
after Harmoni.
For testing the significance of the correlation of the initial value of a
parameter (before Harmoni) and its percentage change after Harmoni, we
used the correction method introduced in (Tu, 2016). Assume x is the base-
line value of a parameter of interest before Harmoni and y is its value after
Harmoni. The percentage change of this parameter is defined as (y − x)/x,
22
1
2
3
1
2
3
Nodes color-code
Edges color-code
Increasing value
Figure 7: Depicting CFS network as a
bipartite graph.
The nodes stand for
brain regions. While the upper set of
nodes represents the alpha activity in
the brain regions, the lower nodes are
for the beta activity in those regions.
When node 1 from alpha nodes (upper
nodes) is connected to node 3 of beta
nodes (lower nodes) it means that the
alpha activity in region 1 is coupled to
beta activity in node 3. The links are
color-coded based on the strength of the
coupling.
Additionally, each node in
each frequency band can have a color
which represents its centrality in that
frequency band.
which is mathematically coupled to x. Therefore, it would not be valid to
use the conventional statistical testing between the initial value and the per-
centage change and compare the observed correlation to zero.
Tu (2016)
suggests that the appropriate null value for the hypothesis test should be
r0 = −
�
1−rxy
2
rather than zero, where rxy is the Pearson correlation of x
and y. In this approach, the hypothesis test is H0 : rx,y/x+
�
1−rxy
2
= 0 versus
H1 : rx,y/x +
�
1−rxy
2
̸= 0. Finally, the expression for the z-test is suggested
to be z =
�
zr(r)−zr(ρ)
�
/
�
1/(n − 3), where zr(r) = 0.5ln((1+r)/(1−r)) is
the Fischer’s z transformation, r is the observed correlation coefficient, and
ρ is the correlation coefficient to be tested against.
3. Results
3.1. Simulations
Toy Examples. As the very first step, we used simplified simulations (toy
signals) to show that Harmoni is an effective algorithm for suppressing spu-
rious CFS and within-frequency interactions due to the non-sinusoidal shape
of the signals. In these simple simulations, where there are no complications
regarding source mixing or limitations of source reconstruction, the ground
truth about the interactions between the two simulated signals is known. In
fact, we were interested to validate two important properties of Harmoni: (1)
It suppresses the spurious interactions significantly, and (2) it does not affect
23
genuine interactions. In addition, these initial simulations serve as a demon-
stration for the main spurious interactions present due to non-sinusoidality.
In each of the four scenarios, two noisy multi-band signals zk(t), k = 1, 2
were simulated with components in alpha and beta band. Different genuine
interactions were simulated between the two signals, resulting in spurious
interactions as well.
Harmoni was applied to each of the two signals to
remove beta components associated with being a harmonic of alpha band
components, i.e. showing CFS with the alpha oscillation. The within- and
cross-frequency interactions were then estimated using absolute coherence to
investigate how they changed after using Harmoni and how these changes
were related to the ground truth.
Each scenario was simulated 50 times
with random seeds. Figure 8 depicts the boxplots of the strength of possible
within- and cross-frequency interactions between and within the two signals,
before and after Harmoni. The interactions in the schematic of each scenario
have the same color-code as their respective boxplots. The change of the
synchronization strength after Harmoni (in comparison to before Harmoni)
was tested with a two-sided paired t-test for each possible interaction, and
then corrected by the Bonferroni method.
In scenario one (figure 8-A), the two signals were synchronized non-
sinusoidal waves with their fundamental frequency in alpha band (i.e., zi(t) ≈
sk(t) + ξk(t) with sk(t) = αk(t) + βk(t) being the non-sinusoidal component
of zk(t). s1 and s2 were simulated to be synchronzied, i.e. α1 ↔ α2, where ↔
shows the synchronization). The CFS interaction between the two signals as
well as the interaction in beta band are by construction spurious. As shown
in figure 8-A, the within- and cross-frequency spurious coherence between
and within the two signals are successfully suppressed after Harmoni.
In scenario two (figure 8-B), each of the two signals contained another
beta component which was independent of the non-sinusoidal components,
but these components from z1 and z2 were simulated to be synchronized to
each other (i.e., zk(t) ≈ sk(t) + ˘βk(t), sk(t) = αk(t) + βk(t), with α1 ↔ α2,
˘β1 ↔ ˘β2). In this scenario, the CFS interaction is by construction spurious,
too. However, a part of the interaction between the two signals within the
beta band is genuine because of the interaction between ˘β1 and ˘β2. The
results in figure 8-B show that the CFS interactions are suppressed, and the
coherence between the beta components of the two signals does not have
any significant change, showing that the genuine beta synchronization is still
present..
Scenario three (figure 8-C) was similar to scenario two with the difference
24
that the non-sinusoidal oscillations from the two signals were not synchro-
nized (i.e., zk(t) ≈ sk(t) + ˘βk(t), sk(t) = αk(t) + βk(t), with ˘β1 ↔ ˘β2).
Therefore, no CFS between the two signals is observed. The boxplots in
figure 8-C show that the CFS within each signal is suppressed as expected
from the proper functioning of Harmoni, while CFS between the two signals
does not change, remaining at a negligible level. Importantly, the genuine
synchronization in beta-band does not change after Harmoni.
In scenario four (figure 8-D) zk(t) ≈ sk(t)+ ˘βk(t), sk(t) = αk(t)+βk(t) as
well. The ground truth interactions were set to α1 ↔ ˘β2. This setting results
in genuine CFS between the two signals. Figure 8-D shows that Harmoni
is robust: the genuine inter-signal CFS does not change, while the present
CFS within each signal as well as the spurious beta-band interaction drop
significantly. Additionally, the other CFS between the two signals which was
missing by construction, does not change and remains at a low value.
All in all the results of the above scenarios show that the spurious inter-
actions are suppressed by Harmoni, while the genuine interactions are not
changed.
Realistic EEG simulations. For the further evaluation of Harmoni, we de-
veloped an EEG simulation pipeline for generating realistic scalp EEG sig-
nals (details in the method section). The simulated EEG data consisted of
narrow-band sinusoidal source signals at alpha (8-12 Hz) and beta (16-24 Hz)
bands, as well as non-sinusoidal signals with fundamental frequency at alpha
band. The dipole positions were randomly selected from the center of 68 re-
gions of interest (ROIs) of Desikan Killiany atlas (Desikan et al., 2006). 1/f
(pink) noise data were also added to the generated source signals of interest.
All the source signals were forward modelled to generate realistic EEG. Two
scenarios (shown in figure 5) were used for generating the simulated EEG
signals. Both of the scenarios included coupled non-sinusoidal alpha sources.
In scenario one there were also within-frequency coupled narrow-band sinu-
soidal alpha and beta sources. In scenario two, in addition to the pair of
coupled non-sinusoidal sources, a genuine, remote cross-frequency coupled
pair of sinusoidal sources was simulated as well. As shown in figure 5, these
two scenarios have differential within- and cross-frequency network profiles.
We used the connectivity pipeline of figure 4 to compute the within-
frequency synchronization in beta band and the alpha-beta cross-frequency
synchronization maps.
As an illustrative example (figure 9) and a proof of principle, we first show
25
CFS within signal 2
CFS: alpha at signal 1
beta synchronization
to beta at signal 2
CFS: alpha at signal 2
to beta at signal 1
before
after
Harmoni
Harmoni
before
after
Harmoni
Harmoni
before
after
Harmoni
Harmoni
before
after
Harmoni
Harmoni
before
after
Harmoni
Harmoni
CFS within signal 1
before
after
Harmoni
Harmoni
before
after
Harmoni
Harmoni
after
Harmoni
Harmoni
before
after
Harmoni
Harmoni
before
after
Harmoni
Harmoni
before
before
after
Harmoni
before
after
Harmoni
Harmoni
after
Harmoni
Harmoni
before
after
Harmoni
Harmoni
before
after
Harmoni
Harmoni
before
Harmoni
Harmoni
Harmoni
after
Harmoni
Harmoni
Harmoni
Harmoni
after
Harmoni
Harmoni
Harmoni
Harmoni
before
after
before
after
before
before
after
before
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.00
0.05
0.10
0.15
0.20
0.25
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.25
0.00
0.05
0.10
0.15
0.20
***
***
***
***
***
***
***
***
***
***
***
*** p<1e-8 (corrected)
* p<0.05 (corrected)
***
***
***
*
A: Scenario1
coherence
coherence
coherence
coherence
B: Scenario 2
C: Scenario 3
D: Scenario 4
Figure 8: Performance of Harmoni on toy examples in 50 runs with random start seeds.
The left -side schemes are the simulation scenarios shown in figure 5. For all scenarios the
strength of each possible interaction is shown before and after Harmoni in the boxplots
in the same panel as the scenario scheme. The purple and blue color are associated with
the within-signal CFS, the two green colors are related to the inter-signal CFS values,
and finally the orange color is dedicated for the beta band synchronization among the
two signals. In all scenarios, two signals are simulated and each of them contains a non-
sinusoidal wave sk(t) = αk(t) + βk(t), k = 1, 2 with their fundamental component αk in
alpha band and their second harmonic βk in beta band. Scenario one: The boxplots show
that all of the within-signal CFS and the spurious interactions are suppressed significantly.
Scenario two: Only the beta-synchronization between the two signals does not change
significantly after Harmoni and stays at a large value due to the genuine synchronization
of ˘βk, k = 1, 2. Scenario three: The CFS within each signal is suppressed significantly,
the CFS values between the two signals do not change and have small values in general,
and importantly the beta-synchronization between the two signals stays almost the same
at a high value. Scenario four: a genuine CFS (light green) between the two signals is
simulated, which is not affected after Harmoni, while the spurious within-beta interactions
and the within-signal CFS are suppressed.
26
an example of scenario two. Two synchronized non-sinusoidal alpha source
signals were simulated with their corresponding sources in caudal middle-
frontal and inferior-parietal regions of right and left hemispheres, respectively.
In addition, two sinusoidal alpha and beta source signals, with CFS, were
simulated in the caudal middle-frontal and inferior-parietal regions of the left
and right hemispheres, respectively. The ground truth networks are shown in
figure 9-A. Afterwards, the source signals, along with random noise sources,
were projected to the sensor space and then the above-mentioned source space
pipeline was performed. Panel B of figure 9 depicts the top 1% connections
of the connectivity networks in alpha band as well as beta band and CFS
networks before and after Harmoni. The spurious beta and CFS connections
are suppressed.
Our main evaluation criterion for the realistic simulations was the area
under curve (AUC) of the receiver operating characteristic (ROC) curve and
the false positive ratio (FPR) curve. These curves were built by comparing
the adjacency matrix of the connectivity graphs before and after Harmoni
to their counterpart ground truth connectivity matrices. The ROC curve
was computed for the beta network in scenario one and the CFS network
in scenario two. The higher the AUC of ROC curve (AUC-of-ROC), the
more similar the connectivity matrix to the ground truth one. Figure 10
shows the results of evaluating the two scenarios of the simulation in 200
Monte Carlo simulations with random dipole positions. The increase of the
AUC-of-ROC in the left sides of panels A and B demonstrates a success
of Harmoni in both of the scenarios in correcting the connectivity maps in
the way that they are more similar to the ground truth. Consequently the
ratio of the true positive ratio (TPR) and FPR increases after Harmoni,
reflecting the suppression of spurious interactions (false positives) and not
affecting/increasing the genuine interactions (true positives). Moreover, the
percentage change of the AUC-of-ROC values decreases with the increase of
the initial value of AUC-of-ROC. That is, the closer the initial connectivity
map to the ground truth, the less correction Harmoni applies.
In other
words, if a network shows a lot of spurious interactions, then it is corrected
by Harmoni more strongly (see statistical analysis section in Methods for
quantifying this dependency in a statistically stringent manner). In addition,
at the left sides of both the panels of figure 10 the AUC of the FPR curves
(AUC-of-FPR) of the CF networks in scenario one, and the beta networks in
scenario two (where all the present interactions are spurious) decrease after
Harmoni (the second columns in figure 10-A and B), showing the suppression
27
0.0
0.2
0.4
0.6
0.8
1.0
anterior
posterior
Ventral view
Caudal view
CFS connectivity
non-sin alpha connectivity
geniune CFS
spurious CFS
A
B
non-sinusoidal alpha
Alpha-Beta connectivity
After Harmoni
Alpha-Beta connectivity
Before Harmoni
Beta connectivity
After Harmoni
Beta connectivity
Before Harmoni
Alpha connectivity
connectivity
Figure 9: An illustrative realistic simulation example, showing the effect of Harmoni in
suppressing the spurious interactions due to harmonics. Panel A depicts the ground truth,
where synchronized non-sinusoidal alpha sources were simulated in right caudal middle-
frontal and left inferior-parietal regions (red connecting line) and two cross-frequency
synchronized narrow-band alpha and beta sources were simulated in the left caudal middle-
frontal and right inferior-parietal regions (purple connection). The circular and bipartite
graphs depict the ground truth alpha and CFS networks. A bipartite graph allows to see
how different nodes from two networks, represented by horizontal bars, connect to each
other allowing non-symmetric connections - without using a directed graph. In the CFS
network, the dashed-lines represent the spurious interactions due the connectivity between
two non-sinusoidal signals, while the solid line represents the genuine interaction. Panel
B shows the top 1% connections of the within-frequency and cross-frequency networks
computed before and after Harmoni. The spurious beta connections and the spurious CFS
connections are suppressed. The glass brains were plotted with Brain Network viewer (Xia
et al., 2013) in MATLAB. The circular plots were generated with MNE Python (Gramfort
et al., 2013, 2014)
28
of the spurious interactions. The absolute value of the percentage change of
the AUC-of-FPR in these cases increases with the increase of the initial value.
This means that the more false positive links are present in the connectivity
maps, the more pronounced is the impact of Harmoni on the networks.
3.2. Harmoni on resting-state EEG data
Alpha oscillations recorded with resting-state EEG (rsEEG) are known
to have a non-sinusoidal waveshape in many brain areas. For example, the µ
rhythm in the somatomotor areas or visual alpha are well-known examples
of non-sinusoidal oscillations. This non-sinusoidal waveform is manifested
in the power spectral density (PSD) having a large peak at alpha and a
smaller peak at beta frequency band, together with 1:2 CFS between alpha
and beta bands. As an example from real data, figure 11 shows a segment
of a non-sinusoidal source signal extracted from the recordings of a subject’s
eyes-closed rsEEG from the LEMON dataset (Babayan et al., 2019). In this
case, the power spectrum of such signal shows two prominent peaks at the
fundamental frequency (11Hz) and its second harmonic frequency (22Hz).
Additionally, a third peak is visible at the third harmonic frequency as well
(33Hz). As indicated by the values of the cross-frequency coherence in the
figure, the harmonic components demonstrate CFS with the fundamental
frequency component.
We used rsEEG data from 81 subjects (data description in the Method
section) and applied Harmoni in order to disambiguate genuine from spurious
CFS alpha-beta interactions. Panel (A) of figure 12 illustrates the across-
subjects average of 1:2 alpha-beta synchronization at each cortical source
(i.e. a vertex on the cortical mantel). A very high 1:2 synchronization within
one cortical source is an indication of a non-sinusoidal waveshape of alpha
oscillation at the corresponding dipole. On average, the occipital, temporal
and central areas demonstrate the highest 1:2 alpha-beta synchronization.
This figure shows the ubiquity of harmonics in data and highlights the im-
portance of taking care of it in connectivity analysis. Note that although
we make the assumption that the 1:2 synchronization at a single source is a
harmonic-driven synchronization, we are fully aware that this can be a result
of residuals of signal mixing in source space. We explicitly address this point
in the discussion.
In order to compute the CFS connectivity networks, a similar data-
analysis pipeline as in the realistic simulations was used at the source space.
The rsEEG multi-channel data were mapped to 100 ROIs of the Schaefer
29
A: Scenario 1
B: Scenario 2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0.3
0.4
0.5
0.6
0.7
0.8
0.9
-20
0
20
40
60
80
100
120
140
before Harmoni
after Harmoni
r=-0.686
H0: r0=-0.262
p=0.0
AUC-of-ROC before Harmoni for beta
percentage change of AUC-of-ROC
AUC-of-ROC for beta connectivity
0.05
0.10
0.15
0.20
0.25
0.30
0.35
-80
-70
-60
-50
-40
-30
-20
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
AUC-of-FPR for CFS connectivity
before Harmoni
after Harmoni
0.4
AUC-of-FPR before Harmoni for CFS
percentage change of AUC-of-FPR
r=-0.813
H0: r0=-0.375
p=0.0
0.2
0.4
0.6
0.8
1.0
0.2
0.4
0.6
0.8
75
50
25
0
25
50
75
100
before Harmoni
after Harmoni
percentage change of AUC-of-ROC
AUC-of-ROC before Harmoni for CFS
AUC-of-ROC for CFS connectivity
p=0.0
r=0.595
H0: r0=-0.348
0.050
0.075
0.100
0.125
0.150
0.175
0.200
0.225
0.025
0.025
0.075
0.125
0.175
-120
-100
-80
-60
-40
-20
-0
-20
-40
0.225
r=0.861
H0: r0=-0.397
p=0
before Harmoni
after Harmoni
AUC-of-FPR for beta connectivity
AUC-of-FPR before Harmoni for beta
percentage change of AUC-of-FPR
Figure 10: Results of 200 realistic simulations according to scenario one (panel A) and
two (panel B) of figure 5. At the left side of panel A, the boxplots of the AUC-of-ROC
of beta connectivity before and after Harmoni are depicted, showing an increase after the
application of Harmoni. This indicates a successful correction of the network’s connections
after Harmoni in favor of suppressing the spurious interactions. Beneath the boxplots, the
scatter-plot of the percentage change vs. the AUC-of-ROC values for beta connectivity
before Harmoni is shown. The higher the initial AUC-of-ROC value (i.e. the more accurate
the initial connectivity map), the less difference between the AUC values before and after
Harmoni (i.e., the less the impact of Harmoni). At the right side of panel A the boxplots
of the AUC-of-FPR for the CFS connectivity are illustrated. Note that in scenario one the
whole CFS connectivity is spurious due to waveshape, which is to a great extent removed
by Harmoni (reflected in the decrease of the FPR). The bottom scatter-plot shows that the
percentage change increases as the AUC-of-FPR of the CFS network increases, meaning
that Harmoni has a larger effect on networks with more spurious interactions. Panel B
shows the results of scenario two, but for the AUC-of-ROC of the CFS network (the left
side) and the AUC-of-FPR of the beta connectivity (the right side). A similar outcome
as in scenario one is observed in scenario two: and increase in the AUC-of-ROC after
Harmoni for CFS networks, as well as a decrease in AUC-of-FPR for beta networks where
all the connections are spurious ones. The percentage-change scatter plots imply a similar
effect: the more spurious interactions in the simulated signals, the more corrections is
performed by Harmoni.
30
3rd harmonic
2nd harmonic
1st harmonic
original source signal
500ms
500ms
500ms
PSD (dB)
Frequency (HZ)
10
20
30
0
40
-5
-15
-25
-35
A
B
Coh1:2=0.65
Coh1:3=0.35
Coh2:3=0.27
500ms
1st harmonic
2nd harmonic
3rd harmonic
Figure 11: An example of a non-sinusoidal brain source signal. In panel A, a non-sinusoidal
brain oscillatory activity and its first three harmonics are shown along with the spatial
pattern of this activity. This source was extracted from eyes-closed rsEEG of a subject
of the LEMON dataset using independent component analysis (ICA) (extended InfoMax
ICA (Lee et al., 1999) with 32 components). Panel B shows the PSD of the non-sinusoidal
signal with the peaks at 11 Hz (first harmonic, or the fundamental frequency), 22 Hz
(second harmonic), and 33Hz (third harmonic).
The cross-frequency coherence of the
harmonic components and the fundamental component are reported as well. The largest
synchronization occurs between the first and second harmonic (coherence value of 0.65).
This is mainly due to the higher signal-to-noise ratio at these frequency bands.
31
atlas (Schaefer et al., 2018) with each ROI being assigned to one of the seven
resting-state Yeo networks, i.e. Default-mode network, Fronto-parietal, Lim-
bic, Ventral Attentional, Dorsal Attentional, Somatomotor, and Visual net-
works (Yeo et al., 2011).
Then, the components of beta activity at each
ROI that could potentially be a higher harmonic of alpha oscillations were
removed using Harmoni. Finally, the ROI-ROI alpha-beta CFS connectivity
networks, represented by 100 × 100 connectivity matrices were computed.
Figure 12-B and C show the across-subject mean connectivity graphs before
and after Harmoni over all subjects. In Panel B (CFS before Harmoni), the
dominating vertical links correspond to the local synchronization of the alpha
oscillations with their second harmonic (beta). This is an expected pattern
for the non-sinusoidal oscillations where both alpha and beta components
are generated at the same location and demonstrate spurious CFS. Panel C
shows that the application of Harmoni resulted in the unmasking of genuine
remote neuronal interactions which were previously under-emphasized due to
the presence of spurious cross-frequency connectivity. In order to be able to
compare the networks before and after Harmoni at the group level, the con-
nectivity matrices were z-scored for each subject and then these standardized
coherence scores before Harmoni were subtracted from the ones after Har-
moni, and paired two-sided t-tests (with Bonferroni correction of p-values)
were employed to specify the links which changed significantly after Harmoni.
Panel 12-D and E show the across-subject mean of the difference networks
for positive and negative links (only the significantly changing links). 12-D
depicts the connections which are more pronounced after Harmoni.
This
enhancement is observed for both inter and intra-hemispheric connections,
specifically between the visual cortices of the two hemispheres, between the
visual areas and the default mode and fronto-parietal regions. These effects
were achieved via the elimination of spurious connections which were driven
by harmonics. The presence of such harmonics masks the strength of the
genuine interactions which, however, become more pronounced after the ap-
plication of Harmoni. The presence of vertical lines and some cross-region
lines in figure 12-E illustrates that within-ROI CFS as well as many within-
hemispheric connections are significantly suppressed.
Importantly, Harmoni does not create any new connections, it rather leads
to a reweighing of the connections after the suppression of the spurious ones.
In order to validate this claim, we used paired t-tests to check whether the
across-subject mean of the weights of each connectivity link changes signif-
icantly after Harmoni. Accounting for multiple comparisons by Bonferroni
32
A
C: after Harmoni
B: before Harmoni
0
0.1
0.2
0.3
CFS strength
difference of zscores - positive
difference of zscores - negative
0.02
0.06
0.1
0.14
CFS strength
Coherence
-
D: after > before
E: after < before
0
-1
-2
-2.5
-0.5
-1.5
0
0.2
0.4
0.6
0.8
Visual
Somatomoto
Dorsal Attentional
Ventral Attentional
Limbic
Fronto-parietal
Default Mode
partly precuneus
V1 + part of cuneous
partly precuneus
Yeo Networks
Figure 12: Harmoni and rsEEG data. Panel (A) shows the across-subject average of 1:2
synchronization of the alpha and beta band activity over the cortex. If the 1:2 synchro-
nization is high at a given source, the second harmonic of the alpha activity may have
a large contribution to the beta activity. Panel (B) shows the bipartite illustration of
the mean CFS connectivity matrix. The nodes are sorted based on their assigned Yeo
resting-state network. The vertical links show the presence of CFS within a single region,
which is a sign of a synchronization due to waveshape (since this way they connect the
same region). Panel (C) is similar to panel (B), but for the data after the application
of Harmoni on beta band. The vertical links in the bipartite illustration are eliminated
and more inter-hemispheric connections emerged. Panel (D) shows the links which are
more pronounced after Harmoni, including more inter-hemispheric interactions. Panel (E)
shows the links which were suppressed by Harmoni. The networks of panels (D) and (E)
were computed by subtracting the z-scored coherence values before Harmoni from the ones
after Harmoni.
33
A
B
label index
label index
mean coherence difference
p-values (corrected)
Figure 13: Harmoni does not create new connections, i.e., an appearance of a synchroniza-
tion between two ROIs after Harmoni which was not present before Harmoni. Panel (A)
shows the significant across-subjects mean difference of the alpha-beta networks after and
before Harmoni (the coherence values before Harmoni were subtracted from the values
after Harmoni). All the values are ≤ 0, showing that the synchronization strengths drop
for all pairs of the ROIs on average. (B) The matrix of corrected p-values (Bonferroni cor-
rected) corresponding to the two-sided paired t-tests performed for each CFS connection
before and after Harmoni. The insignificant connections are not colored. All the signif-
icant changes indicated a decrease, −12.77 ≤ t(80) ≤ −4.9, p < 0.05 (after Bonferroni
correction).
correction, we found that all the significant changes were in the direction of a
decrease in the connectivity strength after Harmoni, −12.77 ≤ t(80) ≤ −4.9,
p < 0.05 (figure 13), which confirms that no new connection is produced by
Harmoni. Indeed, by suppressing the synchronizations that can mimic the
spurious interactions due to non-sinusoidal waveshape of alpha oscillations,
the ratio of the connectivity weights with respect to the maximum synchro-
nization is changed and therefore, some connection weights which previously
were in the low ranks move to higher percentiles of the connectivity weights
after the application of Harmoni. With this procedure, the dominant and
strongest connections change in the CFS network and we observe the net-
works in figure 12-B and C.
Another important feature of the MEG/EEG connectivity networks is
the symmetry of the adjacency matrix. All within-frequency or amplitude-
amplitude coupling networks are characterized by a symmetric adjacency
matrix. However, to the best of our knowledge, no study until so far inves-
tigated the presence of a similar pattern in the adjacency matrix for CFS
coupling which is strongly affected by the interactions due to higher har-
monics of non-sinusoidal shape of the signals. The CFS adjacency matrix
is by definition asymmetric. Actually, harmonic-driven spurious interactions
34
result in symmetric CFS matrix. In other words, if the alpha activity in
region i is coupled to the beta activity in region j, the (i,j)-th element of the
adjacency matrix is non-zero. If this coupling is due to the non-sinusoidal
shape of the waveform of the alpha-signals at both of these two regions, then
the beta activity in region i is also synchronized to the alpha activity in
region j, which results in a non-zero value at the (j, i)-th element of the
adjacency matrix. This decreases the extent to which the adjacency ma-
trix is asymmetric. Therefore, we reasoned that Harmoni should decrease
the extent to which the adjacency matrix of the CFS network is symmetric.
This idea was indeed confirmed as shown in figure 14-A with the boxplots
of an asymmetry-index (refer to Methods) of the CFS networks before and
after Harmoni for all subjects, where the asymmetry-index of the individual
CFS connectivity networks increases significantly after Harmoni (two-sided
paired t-test, t(80) ≈ 17.75, p ≈ 1.8e − 29). Furthermore, panel B of this
figure shows that the percentage change of the asymmetry-index significantly
decreases with the initial value of the index, pearson r=−0.75, p ≈ 0.0007
(with null hypothesis r=-0.55). In other words, Harmoni corrects the CFS
network more (resulting in a more asymmetric network), when there are more
potentially spurious interactions due to harmonics (i.e., the CFS network is
less symmetric). See Methods for the rigorous statistical treatment of this
analysis. Note that not all the harmonic-driven cross-frequency interactions
are reflected in the symmetry of the CFS network adjacency matrix.
4. Discussion
EEG and MEG techniques are becoming more and more frequently used
for the investigation of neuronal connectivity, owing to (1) their ability to
record neuronal activity directly, and 2) their refined temporal resolution in
a millisecond range which is required for the detection of subtle changes in
neuronal dynamics. In addition, the recent advancement of brain data anal-
ysis for mapping sensor recordings to the cortex has provided an opportunity
for computing the connectivity of different brain areas in source space. Yet,
connectivity analysis with MEG/EEG faces considerable challenges.
The
limited spatial resolution and spatial mixing of neural activity from differ-
ent regions hampers connectivity analysis. Additionally, the non-sinusoidal
shape of brain oscillations has been repeatedly highlighted as crucially af-
fecting the (mis)interpretation of underlying neuronal activity (Hyafil, 2017;
Lozano-Soldevilla, 2018). Because non-sinusoidality always implies a pres-
35
asymmetry-index of CFS networks
0.25
0.35
0.45
0.55 0.6
percentage change of asymmetry-index
0
20
40
60
80
0.25
asymmetry-index before Harmoni
0.3
0.35
0.45
0.4
0.5
before Harmoni
after Harmoni
A
B
r=-0.75
p=0.0006
H0: r0=-0.55
t(80)=17.75
p~10e28
Figure 14: The CFS networks of individual LEMON subjects becomes more asymmetric
after Harmoni. (A) the boxplots of the asymmetry-index of the CFS adjacency matrices
of all subjects shows that the asymmetry of the CFS adjacency matrices increases signif-
icantly after Harmoni. (B) The scatter-plot of the percentage change of the asymmetry-
index vs. the initial value of the index, i.e., before Harmoni.The less asymmetric the CFS
network (i.e., the more harmonic-driven symmetric connections), the more changes are
observed after Harmoni. The solid line shows the linear regression line and the blue shade
shows the result of a leave-one-out bootstrap.
ence of harmonics, these harmonics can often be mistakenly taken to repre-
sent genuine neuronal oscillations. Consequently, spurious interactions are
observed between harmonics of a non-sinusoidal oscillation and other neu-
ronal processes in the same frequency range, which in turn cannot be easily
disentangled from genuine interactions. This has been recognized earlier as
a major challenge for studying phase-amplitude coupling (PAC) in neuronal
data (Aru et al., 2015; Giehl et al., 2021; Jensen et al., 2016; Lozano-Soldevilla
et al., 2016; Zhang et al., 2021) as well as for n:m phase-synchronization
(Hyafil, 2017; Scheffer-Teixeira and Tort, 2016; Siebenh¨uhner et al., 2020).
In this work, we directly addressed the issue of spurious interactions due to
waveshape of oscillations and offer a solution for the assessment of phase
synchronization as one of the most important measures used for connectivity
analyses with brain electrophysiology (Marzetti et al., 2019; Nentwich et al.,
2020; Sadaghiani et al., 2021; Vidaurre et al., 2020).
Currently available measures for quantifying n:m phase-synchronization
(also referred to as cross-frequency synchronization - CFS) are not suitable for
differentiation between genuine and spurious interactions. Short data length,
filtering bias, and non-sinusoidal signal waveshape are being mentioned as
36
reasons for measuring spurious n:m phase-synchronization. Statistical tests
based on surrogate data can be used for disentangling spurious and genuine
phase-synchronization due to limited data points or filtering factor.
Yet,
these procedures cannot differentiate the genuine interactions from the spu-
rious ones due to the non-sinusoidality of oscillations (Scheffer-Teixeira and
Tort, 2016). The reason for this is that Fourier and narrow-band analysis is
the base of almost all current signal processing pipelines, where a signal is de-
composed into narrow frequency band components. Consequently, the higher
harmonics of a non-sinusoidal signal are analysed as representing genuine os-
cillations not directly relating to the fundamental frequency. In the context
of cross-frequency coupling, this can result in the observation of spurious in-
teractions which are mimicking genuine interactions and cannot be detected
by surrogate tests. Furthermore, the non-sinusoidal waveshape of oscillatory
brain signals produce spurious interactions in the within-frequency phase-
synchronization in the range of harmonic-frequency, as depicted schemati-
cally in figure 1.
Although the presence of spurious interactions in phase-synchronization
connectivity analysis of neurophysiological data has been largely acknowl-
edged by the community, there has been only very few attempts for providing
a potential solution for it. Palva et al. (2005) used the coincidence of cross-
frequency phase-phase and amplitude-amplitude coupling as the hallmark of
harmonic-driven CFS. This, however, is more a qualitative measure rather
than a quantitative one and can be less applicable to the inter-areal whole
brain connectivity analysis. In a recent paper, Siebenh¨uhner et al. (2020)
suggested a graph-theoretical analysis for discarding potential spurious CFS.
The authors employed a procedure of detecting ambiguous motifs in the CFS
graph combined with the within-frequency graphs of the fundamental and
harmonic frequencies of interest, and discarding the CFS interactions corre-
sponding to the links included in those motifs. This procedure, however, was
not validated using realistic MEG/EEG simulations. Such graph-based post-
processing of connectivity networks can in fact discard all the interactions
which mimic the motif of spurious interactions in the connectivity graphs.
However, due to the limited spatial resolution of MEG/EEG data, some of
the genuine interactions among the ROIs may still coincide with harmonic-
driven spurious interactions, as we show in figure 8-D. The graph motif of
such interactions is similar to the spurious interactions, depicted in figure 8-
A. Thus, a motif-discarding approach cannot distinguish the two cases of 8-A
and D and would label the CFS interaction as a spurious one. Moreover, this
37
graph-based correction method is applicable only to cross-frequency graphs,
while, as discussed in this study, the within-frequency interactions in the
harmonic frequency band may also include spurious interactions driven by
non-sinusoidal waveshape. Therefore, to the best of our knowledge, so far
there has been no method that can address the issue of spurious n:m inter-
actions due to waveshape via removing the harmonic components from the
neuronal signals.
A signal processing tool for dealing with harmonics in connectivity. In this
manuscript, we introduced the first signal processing tool for suppressing
spurious within- and cross-frequency synchronization due to non-sinusoidal
shape of the oscillatory activity in the brain. Our method significantly sup-
presses the spurious interactions, while at the same time not affecting genuine
interactions present in data. We first validated these two key properties using
simple, yet informing, simulations. They consisted of two signals with differ-
ent components interacting with each other, giving us a chance to evaluate
Harmoni’s performance in the presence of genuine and spurious interactions
in data. The results of these simulations (figure 8) showed that Harmoni
effectively suppresses spurious within- and cross-frequency interactions. Im-
portantly, this suppression did not affect the genuine interactions.
Realistic simulations: decrease in FPR, increase in AUC of ROC curve. In
order to comprehensively assess Harmoni’s performance, we used realistic
simulations where source mixing and limitations of source reconstruction are
present. Using the area under curve (AUC) of the receiver operating charac-
teristic (ROC) curve (figure 10), we showed that Harmoni increases the AUC
of ROC curve of connectivity networks where the ground truth included both
genuine and spurious interactions. This means that with Harmoni, it was
possible to uncover even weak connections that would have been masked by
spurious CFS otherwise. In the same direction as the results of the toy ex-
amples, the increase in AUC of ROC curve in realistic simulations indicates
that Harmoni does not affect genuine interactions (reflected in TPR) and
suppresses spurious interactions (i.e., false positives). In those simulations
where the ground truth connectivity networks were based on spurious in-
teractions only, Harmoni decreased the AUC of the FPR curve. Confirming
other results of the simulations, this result further demonstrates that spurious
interactions both for within-frequency and cross-frequency connectivity are
indeed suppressed significantly by Harmoni. This aspect of Harmoni is par-
ticularly important for the investigation of connectivity for beta oscillations
38
in the sensorimotor networks where comb-shaped mu oscillations are abun-
dant (Schaworonkow and Nikulin, 2019) and thus their harmonics in beta
frequency range should lead to spurious connectivity while merely reflecting
interactions at the base alpha frequency. Additionally, in studies addressing
the relationship of EEG and fMRI data, for example (Ritter et al., 2009),
Harmoni could contribute to the suppression of the effects of harmonic com-
ponents and disentangling the effect of harmonics and the genuine activity
in the same frequency band.
Moreover, given that our simulations were based on hundreds of runs with
different random locations of the sources, one can conclude that Harmoni is
applicable to a wide variety of source configurations in the cortex including
frontal, sensorimotor, and occipito-parietal areas.
Harmoni on resting-state EEG data. Real neuronal data are of a complex
nature and in most cases the ground truth of connectivity patterns is not
known. Therefore, the main validating stage of new methods is rather based
on simulations. However, any new method should also be applied to real data
to further extend its validity. For this purpose, we used resting-state EEG
(rsEEG) of 81 subjects from the LEMON database (Babayan et al., 2019).
We discussed how a symmetric adjacency matrix of a cross-frequency syn-
chronization network can reflect the presence of harmonics, and showed that
the adjacency matrices of the CFS networks become more asymmetric after
Harmoni. Additionally, we showed that Harmoni does not create new con-
nections which were not observed before the application of Harmoni. How-
ever, it changes the relative strength of the already existing connections by
suppressing spurious connectivity. Harmoni suppresses the CFS interactions
both within and between regions, as depicted in figure 12-E. Consequently,
other interactions, which were previously not ranked high due to the presence
of strong spurious interactions, become more pronounced after the applica-
tion of Harmoni. Although a detailed analysis of connectivity patterns of
rsEEG goes beyond the scope of the current study, below we illustrate a few
examples of the unmasked synchronization after the application of Harmoni.
In our data, only after the application of Harmoni, the visual cortical
areas appear to be interacting strongly with other regions, especially inter-
hemispherically. This in turn indicates that the interaction of the visual sys-
tem with other cortical areas is not based only on a relatively slow amplitude-
amplitude coupling as shown previously (Hipp and Siegel, 2015) but in fact
can demonstrate genuine millisecond-range functional interactions important
39
for the precise coordination of neuronal activity in the brain. Additionally,
Wang et al. (2008), in an resting-state fMRI study, found that the spon-
taneous activity in primary visual cortex is associated with the activity in
bilateral middle occipital gyrus, bilateral lingual gyrus, and bilateral cuneous
and precuneus suggesting that these spontaneous activities may be related to
visual imagery during resting-state. In our rsEEG data, the recovered inter-
hemispheric interactions between the visual networks after the application of
Harmoni can also be interpreted in this direction. Interestingly, figure 12-D
shows the influence of Harmoni in recovering remote interactions of alpha
and beta activity in ROIs overlapping with precuneus in both hemispheres -
precuneus is known as a critical region for visual imagery in memory recall
(Wang et al., 2008). Note that we also observed the emergence of precuneus
as an important region in cross-frequency interactions, as well as in the inter-
hemispheric interactions of visual cortices in our previous study (Idaji et al.,
2020) with similar data, where phase-phase synchronized sources were sepa-
rated with a multivariate source separation method.
Furthermore, figure 12-D illustrates intensified within- and inter-hemispheric
interactions of default mode network (DMN) and visual networks, especially
areas in the vicinity of V1. In line with our observation, in a recent paper,
Costumero et al. (2020) reported a connectivity of V1 with DMN as well as
posterior cingulate cortex in closed-eyes resting-state fMRI functional con-
nectivity, suggesting that this connectivity may reflect a brain configuration
associated with mental imagery.
Harmoni and signal mixing. Due to the limited spatial resolution of non-
invasive recordings, the activity of very close neuronal sources cannot be
disentangled when being recorded by non-invasive imaging techniques such
as MEG/EEG. Therefore, even at the source space, the observation of signals
with non-sinusoidal shapes in non-invasive recordings may be due to mixing of
distinct coupled sources with very close spatial locations. Using MEG/EEG,
such cases cannot be distinguished from single sources generating signals
with non-sinusoidal shapes. This limitation is also applicable to the Harmoni
connectivity pipeline, when applying it to MEG/EEG data. However, it is
important to note that, this problem is not a natural limitation of Harmoni.
If we have access to invasive LFP recordings where the spatial resolution can
be in the order of hundred of micrometers (Buzs´aki et al., 2012), Harmoni
can successfully resolve such cases.
The other aspect of spatial mixing relates to the leakage of spatially
40
distanced source signals to other locations, even after source reconstruction.
As a result, the synchronization observed at a single region (or even at a given
reconstructed cortical source) may be due to the synchronization between
distanced source signals which are spatially mixed and still could not be
fully disentangled with source separation or source reconstruction methods.
This, however, is again a general problem of data analysis in MEG/EEG
research and is not specific to Harmoni. Therefore, in some instances the
removal of harmonics in a ROI by Harmoni can lead to removing components
which were not a harmonic of a lower frequency in that region but rather
represents a leaked oscillatory activity from another coupled source. Yet, this
property can in fact be an advantage for Harmoni: It can remove some of the
spurious interactions which were present due to spatial leakage and uncover
the activity at the harmonic frequency, which was not a result of spatial
leakage of a coupled source. As an illustrative example for this property, in
panel A of figure 8, if β1 is not a harmonic of α1 but a leakage of a cross-
frequency coupled source different from s1, then the observed interaction
β1 − α2 would still be accounted as a spurious interaction. This interaction,
however, is successfully suppressed by Harmoni.
Finally, the mixing of background neuronal activity - known as 1/f noise
- and other noise sources with oscillatory activities affects the signal-to-noise
ratio (SNR) and consequently the estimation of the true phase of the os-
cillations. Using simulations in (Idaji et al., 2020), we showed how source
separation of cross-frequency coupled sources worsens with decreasing SNR.
Therefore, the phase estimates and consequently the n:m synchronization
suffer from noise contamination. Because of this issue, the synchronization
should be estimated with a sufficient amount of data points for MEG/EEG
recordings.
5. Code and data availability
The codes of Harmoni, simulating toy examples, as well as analysing
the simulated EEG and real data are available at github.com/harmonic-
minimization. EEG data is from LEMON dataset, which is a public database
(Babayan et al., 2019).
6. Author contributions
MJI: Conceptualization, Methodology, Software, Validation, Investiga-
tion, Formal analysis, Visualization, Project Administration, Writing - orig-
41
inal draft, Writing - review editing. JZ and TS: Software, Writing - review
editing. GN: Methodology, Writing - review editing. KRM: Writing - review
editing, Supervision. AV: Resources, Writing - review editing, Supervision.
VVN: Conceptualization, Methodology, Investigation, Project Administra-
tion, Writing - review editing, Supervision.
7. Competing interests
We declare no competing interests.
42
Toy example: Scenario 1
Toy example: Scenario 2
Figure 15: Examples of the composition of the two signals of scenario 1 and 2 of the toy
examples. In scenario 1 zk = sk + ξk, k = 1, 2 with sk = αk + βk. the noise components ξk
are not depicted. The solid lines show the simulated interactions, while the dashed lines
are the spurious interactions which are the by-products of the simulated ones.
43
References
References
Aru, J., Aru, J., Priesemann, V., Wibral, M., Lana, L., Pipa, G., Singer, W.,
Vicente, R., 2015. Untangling cross-frequency coupling in neuroscience.
Current Opinion in Neurobiology 31, 51–61. doi:10.1016/j.conb.2014.
08.002.
Babayan, A., Erbey, M., Kumral, D., Reinelt, J.D., Reiter, A.M., R¨obbig,
J., Schaare, H.L., Uhlig, M., Anwander, A., Bazin, P.L., et al., 2019. A
mind-brain-body dataset of mri, eeg, cognition, emotion, and peripheral
physiology in young and old adults. Scientific data 6, 1–21. doi:10.1038/
sdata.2018.308.
Bullmore, E., Sporns, O., 2009. Complex brain networks: graph theoretical
analysis of structural and functional systems. Nature reviews neuroscience
10, 186–198. doi:10.1038/nrn2575.
Buzs´aki, G., Anastassiou, C.A., Koch, C., 2012. The origin of extracellular
fields and currents—eeg, ecog, lfp and spikes. Nature reviews neuroscience
13, 407–420. doi:10.1038/nrn2575.
Buzs´aki, G., Draguhn, A., 2004. Neuronal oscillations in cortical networks.
science 304, 1926–1929. doi:10.1126/science.1099745.
Canolty, R.T., Knight, R.T., 2010. The functional role of cross-frequency
coupling. Trends in cognitive sciences 14, 506–515. doi:10.1016/j.tics.
2010.09.001.
Cole, S.R., Voytek, B., 2017. Brain oscillations and the importance of wave-
form shape. Trends in cognitive sciences 21, 137–149. doi:10.1016/j.
tics.2016.12.008.
Costumero, V., Bueichek´u, E., Adri´an-Ventura, J., ´Avila, C., 2020. Opening
or closing eyes at rest modulates the functional connectivity of v1 with
default and salience networks. Scientific reports 10, 1–10. doi:10.1038/
s41598-020-66100-y.
Desikan, R.S., S´egonne, F., Fischl, B., Quinn, B.T., Dickerson, B.C., Blacker,
D., Buckner, R.L., Dale, A.M., Maguire, R.P., Hyman, B.T., et al., 2006.
44
An automated labeling system for subdividing the human cerebral cortex
on mri scans into gyral based regions of interest. Neuroimage 31, 968–980.
doi:10.1016/j.neuroimage.2006.01.021.
Engel, A.K., Fries, P., 2010. Beta-band oscillations—signalling the status
quo? Current opinion in neurobiology 20, 156–165. doi:10.1016/j.conb.
2010.02.015.
Fries, P., 2015. Rhythms for cognition: communication through coherence.
Neuron 88, 220–235. doi:10.1016/j.neuron.2015.09.034.
Giehl, J., Noury, N., Siegel, M., 2021.
Dissociating harmonic and non-
harmonic phase-amplitude coupling in the human brain. NeuroImage 227,
117648. doi:10.1016/j.neuroimage.2020.117648.
Gramfort, A., Luessi, M., Larson, E., Engemann, D., Strohmeier, D., Brod-
beck, C., Goj, R., Jas, M., Brooks, T., Parkkonen, L., H¨am¨al¨ainen, M.,
2013. Meg and eeg data analysis with mne-python. Frontiers in Neuro-
science 7, 267. doi:10.3389/fnins.2013.00267.
Gramfort, A., Luessi, M., Larson, E., Engemann, D.A., Strohmeier, D., Brod-
beck, C., Parkkonen, L., H¨am¨al¨ainen, M.S., 2014. Mne software for pro-
cessing meg and eeg data. NeuroImage 86, 446 – 460. doi:10.1016/j.
neuroimage.2013.10.027.
Harris, A.Z., Gordon, J.A., 2015.
Long-range neural synchrony in be-
havior.
Annual review of neuroscience 38, 171–194.
doi:10.1146/
annurev-neuro-071714-034111.
Hipp, J., Siegel, M., 2015. Bold fmri correlation reflects frequency-specific
neuronal correlation. Current Biology 25, 1368–1374. doi:10.1016/j.cub.
2015.03.049.
Hyafil, A., 2017. Disharmony in neural oscillations. Journal of Neurophysi-
ology 118, 1–3. doi:10.1152/jn.00026.2017.
Idaji, M.J., M¨uller, K.R., Nolte, G., Maess, B., Villringer, A., Nikulin, V.V.,
2020. Nonlinear interaction decomposition (nid): A method for separa-
tion of cross-frequency coupled sources in human brain. NeuroImage 211,
116599. doi:10.1016/j.neuroimage.2020.116599.
45
Jensen, O., Colgin, L.L., 2007. Cross-frequency coupling between neuronal
oscillations. Trends in Cognitive Sciences 11, 267–269. doi:10.1016/j.
tics.2007.05.003.
Jensen, O., Spaak, E., Park, H., 2016.
Discriminating valid from spuri-
ous indices of phase-amplitude coupling. Eneuro 3. doi:10.1523/ENEURO.
0334-16.2016.
Jones, S.R., Pritchett, D.L., Sikora, M.A., Stufflebeam, S.M., H¨am¨al¨ainen,
M., Moore, C.I., 2009. Quantitative analysis and biophysically realistic
neural modeling of the meg mu rhythm: rhythmogenesis and modulation
of sensory-evoked responses. Journal of neurophysiology 102, 3554–3572.
doi:10.1152/jn.00535.2009.
Kramer, M.A., Tort, A.B., Kopell, N.J., 2008.
Sharp edge artifacts and
spurious coupling in eeg frequency comodulation measures.
Journal of
neuroscience methods 170, 352–357. doi:10.1016/j.jneumeth.2008.01.
020.
Lee, T.W., Girolami, M., Sejnowski, T.J., 1999. Independent component
analysis using an extended infomax algorithm for mixed subgaussian and
supergaussian sources. Neural computation 11, 417–441.
Lozano-Soldevilla, D., 2018.
Nonsinusoidal neuronal oscillations: bug or
feature?
Journal of neurophysiology 119, 1595–1598. doi:10.1152/jn.
00744.2017.
Lozano-Soldevilla, D., Ter Huurne, N., Oostenveld, R., 2016.
Neuronal
oscillations with non-sinusoidal morphology produce spurious phase-to-
amplitude coupling and directionality.
Frontiers in computational neu-
roscience 10, 87. doi:10.3389/fncom.2016.00087.
Marzetti, L., Basti, A., Chella, F., D’Andrea, A., Syrj¨al¨a, J., Pizzella, V.,
2019. Brain functional connectivity through phase coupling of neuronal
oscillations: A perspective from magnetoencephalography.
Frontiers in
Neuroscience 13, 964. doi:10.3389/fnins.2019.00964.
Miller, K.J., Schalk, G., Fetz, E.E., den Nijs, M., Ojemann, J.G., Rao,
R.P., 2010. Cortical activity during motor execution, motor imagery, and
imagery-based online feedback. Proceedings of the National Academy of
Sciences 107, 4430–4435. doi:10.1073/pnas.0913697107.
46
Nentwich, M., Ai, L., Madsen, J., Telesford, Q.K., Haufe, S., Milham, M.P.,
Parra, L.C., 2020. Functional connectivity of eeg is subject-specific, asso-
ciated with phenotype, and different from fmri. NeuroImage 218, 117001.
doi:10.1016/j.neuroimage.2020.117001.
Nikulin, V.V., Brismar, T., 2006. Phase synchronization between alpha and
beta oscillations in the human electroencephalogram. Neuroscience 137,
647–657. doi:10.1016/j.neuroscience.2005.10.031.
Nolte, G., Bai, O., Wheaton, L., Mari, Z., Vorbach, S., Hallett, M., 2004.
Identifying true brain interaction from eeg data using the imaginary part
of coherency. Clinical neurophysiology 115, 2292–2307. doi:10.1016/j.
clinph.2004.04.029.
Palva, J.M., Palva, S., 2018a. Functional integration across oscillation fre-
quencies by cross-frequency phase synchronization. European Journal of
Neuroscience 48, 2399–2406. doi:10.1111/ejn.13767.
Palva, J.M., Palva, S., Kaila, K., 2005. Phase synchrony among neuronal
oscillations in the human cortex. Journal of Neuroscience 25, 3962–3972.
doi:10.1523/JNEUROSCI.4250-04.2005.
Palva, S., Palva, J.M., 2018b. Roles of brain criticality and multiscale os-
cillations in temporal predictions for sensorimotor processing. Trends in
neurosciences 41, 729–743. doi:10.1016/j.tins.2018.08.008.
Pascual-Marqui, R.D., 2007. Discrete, 3D distributed, linear imaging meth-
ods of electric neuronal activity. Part 1: exact, zero error localization.
arXiv preprint arXiv:0710.3341 .
Patzelt, F., 2019. colornoise python package. https://pypi.org/project/
colorednoise/. Checked: 2020-06-24.
Ritter, P., Moosmann, M., Villringer, A., 2009. Rolandic alpha and beta
eeg rhythms’ strengths are inversely related to fmri-bold signal in primary
somatosensory and motor cortex. Human brain mapping 30, 1168–1187.
doi:10.1002/hbm.20585.
Sadaghiani, S., Brookes, M.J., Baillet, S., 2021. Connectomics of human
electrophysiology. PsyArXiv .
47
Sadaghiani, S., Kleinschmidt, A., 2016. Brain networks and α-oscillations:
structural and functional foundations of cognitive control. Trends in cog-
nitive sciences 20, 805–817. doi:10.1016/j.tics.2016.09.004.
Schaefer, A., Kong, R., Gordon, E.M., Laumann, T.O., Zuo, X.N., Holmes,
A.J., Eickhoff, S.B., Yeo, B.T., 2018.
Local-global parcellation of the
human cerebral cortex from intrinsic functional connectivity mri. Cerebral
cortex 28, 3095–3114. doi:10.1093/cercor/bhx179.
Schaworonkow, N., Nikulin, V.V., 2019. Spatial neuronal synchronization
and the waveform of oscillations: Implications for eeg and meg. PLoS Com-
putational Biology 15, e1007055. doi:10.1371/journal.pcbi.1007055.
Scheffer-Teixeira, R., Tort, A.B., 2016. On cross-frequency phase-phase cou-
pling between theta and gamma oscillations in the hippocampus. Elife 5,
e20515. doi:10.7554/eLife.20515.
Siebenh¨uhner, F., Wang, S.H., Arnulfo, G., Lampinen, A., Nobili, L., Palva,
J.M., Palva, S., 2020. Genuine cross-frequency coupling networks in human
resting-state electrophysiological recordings. Plos Biology 18, e3000685.
doi:10.1371/journal.pbio.3000685.
Tu, Y.K., 2016. Testing the relation between percentage change and baseline
value. Scientific reports 6, 1–8. doi:10.1038/srep23247.
Vidaurre, C., Haufe, S., Jorajur´ıa, T., M¨uller, K.R., Nikulin, V.V., 2020.
Sensorimotor functional connectivity: a neurophysiological factor related
to bci performance.
Frontiers in Neuroscience 14, 1278.
doi:10.3389/
fnins.2020.575081.
Wang, K., Jiang, T., Yu, C., Tian, L., Li, J., Liu, Y., Zhou, Y., Xu, L.,
Song, M., Li, K., 2008.
Spontaneous activity associated with primary
visual cortex: a resting-state fmri study.
Cerebral cortex 18, 697–704.
doi:10.1093/cercor/bhm105.
Xia, M., Wang, J., He, Y., 2013. Brainnet viewer: a network visualization
tool for human brain connectomics. PloS one 8, e68910. doi:10.1371/
journal.pone.0068910.
Yeo, B.T., Krienen, F.M., Sepulcre, J., Sabuncu, M.R., Lashkari, D.,
Hollinshead, M., Roffman, J.L., Smoller, J.W., Z¨ollei, L., Polimeni,
48
J.R., et al., 2011. The organization of the human cerebral cortex esti-
mated by intrinsic functional connectivity.
Journal of neurophysiology
doi:10.1152/jn.00338.2011.
Zhang, J., Idaji, M.J., Villringer, A., Nikulin, V.V., 2021. Neuronal biomark-
ers of parkinson’s disease are present in healthy aging. NeuroImage 243,
118512.
URL: https://www.sciencedirect.com/science/article/
pii/S1053811921007850, doi:10.1016/j.neuroimage.2021.118512.
49
Algorithm 1: Grid-search algorithm of Harmoni.
filter(., f0)
stands for band-pass filtering around f0. Hilbert(.) builds the ana-
lytic signal of its input using the Hilbert transform. Re(.) denotes
the real part of a complex number. std(.) stands for standard devi-
ation.
Input
: A signal z(t) ∈ R containing a non-sinusoidal component
with a fundamental frequency of f0
Frequency f0
Integer n (referring to the n-th harmonic)
Output: Harmonic-corrected signal ycorr(t) ∈ C centered at nf0
xR(t) = filter
�
z(t), f0
�
// band-pass filter around f0
x(t) = Hilbert
�
xR(t)
�
// the analytic signal of xR(t)
yR(t) = filter
�
z(t), nf0
�
// band-pass filter around nf0
y(t) = Hilbert
�
yR(t)
�
// the analytic signal of yR(t)
xn(t) = ax(t)ejnφx(t)
// accelerate x by a factor of n
xn(t) = xn(t)/std
�
Re(xn)
�
// normalize the power
˜y(t) = y(t)/std
�
Re(y)
�
for c = −1 to 1 with steps δc do
for φ = −π/2 to π/2 with steps δφ do
yres(t) = ˜y(t) − cxn(t)ejφ
cohc,φ = |coh
�
yres, xn
�
|
copt, φopt = argmin
c,φ
cohc,φ
// find the minimum
˜ycorr(t) = ˜y(t) − coptxn(t)ejφopt
ycorr(t) = ˜ycorr(t).std
�
Re(y)
�
// set the power of y
50
| 2021 | Harmoni: a Method for Eliminating Spurious Interactions due to the Harmonic Components in Neuronal Data | 10.1101/2021.10.06.463319 | [
"Idaji Mina Jamshidi",
"Zhang Juanli",
"Stephani Tilman",
"Nolte Guido",
"Müller Klaus-Robert",
"Villringer Arno",
"Nikulin Vadim V."
] | creative-commons |
ResFinderFG v2.0: a database of antibiotic resistance genes obtained by functional
1
metagenomics
2
Rémi Gschwind1,, Svetlana Ugarcina Perovic2, Marie Petitjean1, Julie Lao1, Luis Pedro Coelho2,
3
Etienne Ruppé1*
4
1 Université Paris Cité and Université Sorbonne Paris Nord, Inserm, IAME, F-75018 Paris,
5
France
6
2 Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University,
7
Shanghai, China
8
*Corresponding author:
9
Rémi GSCHWIND, PhD
10
INSERM UMR1137 IAME
11
Faculté de Médecine Bichat
12
16 rue Henri Huchard
13
75108 Paris, France
14
+33(0)658543545
15
remi.gschwind@inserm.fr
16
17
Abstract
18
Metagenomics can be used to monitor the spread of antibiotic resistance genes (ARGs). ARGs
19
found in databases such as ResFinder and CARD primarily originate from culturable and
20
pathogenic bacteria. However, ARGs composing the resistome of the human gut microbiota or
21
the environment remain understudied. Functional metagenomics is based on phenotypic gene
22
selection and can identify ARGs from non-culturable bacteria with a potentially low identity
23
shared with known ARGs. In 2016, the ResFinderFG v1.0 database was created to collect ARGs
24
from functional metagenomics studies. Here, we present the database second version,
25
ResFinderFG v2.0. Functional metagenomics studies were analyzed and DNA sequences
26
described were retrieved, deduplicated and annotated. Sequences were curated to include only
27
ARG sequences. ResFinderFG v2.0 was then compared to other databases for their relative
28
sensitivity in searches for ARGs in subcatalogs from the Global Microbial Gene Catalog
29
(GMGC). Fifty publications were considered, for a total of 23’764 ARGs identified from different
30
environments. After deduplication, annotation and curation, 3’913 ARGs were included. New
31
ARGs included are mainly glycopeptides/cycloserine or beta-lactams resistance genes identified
32
mostly in human-associated samples. Results of GMGC gene subcatalogs annotation showed
33
that ResFinderFG v2.0 detected comparable or higher ARG numbers than those detected with
34
other databases. Most of the unigene hits obtained were database-specific and ResFinderFG
35
v2.0
specific
unigene
hits
included
among
others:
glycopeptides/cycloserine,
36
sulofnamides/trimethoprim
resistance
genes
and
beta-lactamases
encoding
genes.
37
ResFinderFG v2.0 can be used to identify ARGs differing from those found in conventional
38
databases and therefore improve the description of resistomes.
39
Introduction
40
Antimicrobial resistance (AMR) is recognized as a global threat possibly leading to the lack of
41
efficient treatment against deadly infections1. From a genetic perspective, AMR is driven by
42
mutational events (e.g. fluoroquinolone resistance is driven by mutations in the topoisomerase-
43
encoding genes) and the expression of antibiotic resistance genes (ARGs). ARGs are
44
widespread in human- and animal-associated microbiomes, and in the environment2. Hence,
45
these microbial niches are now considered in a One Health manner3. Although not every ARG
46
represents a direct risk for human health4, genes are able to travel from one environment to
47
another by strain dissemination or horizontal gene transfer5. In this way, some ARGs represent a
48
risk as they may be transferred to pathogenic bacteria.
49
Identifying ARGs and assessing this risk is essential to better understand and putatively find
50
means to prevent their dissemination in pathogenic bacteria. To identify ARGs, culture-based
51
methods, PCR, qPCR6, genomic and metagenomic sequencing have been used. Metagenomics
52
makes it possible to sequence all the DNA from a sample and, thanks to sequences comparison
53
with specific databases, allows ARG identification in an environment or host. Several AMR
54
public reference databases7 exist such as CARD8 or ResFinder9. However, since the detection is
55
matching newly obtained sequences to ARG sequences in databases, only sequences that are
56
similar to previously-described ones will be detected with an acceptable degree of confidence.
57
Therefore, unknown ARGs, or ARGs sharing a low identity with ARGs included in the chosen
58
database may not be detected.
59
Although culturable and/or pathogenic bacteria only represent a small fraction of microbial
60
diversity, their genes make up the vast majority of the ARGs present in existing databases. In
61
order to detect new ARGs or low sequence similarity percentage ARGs, functional
62
metagenomics has been used10. This technique is based on phenotypic detection by expressing
63
exogenous DNA in an antibiotic-susceptible host. Using functional metagenomics, ARGs sharing
64
low amino acid identity to their closest homologue in NCBI11, or even not previously classified as
65
ARGs12, could be detected in human12–22, animal22–29, wastewater30–37 and other environmental
66
samples5,11,38–60. Despite being a laborious technique, genes described by functional
67
metagenomics are mainly absent in classical ARG databases. Two databases listing specifically
68
functionally identified ARGs were created: ResFinderFG v1.061 and FARME DB62. ResFinderFG
69
v1.0 (https://cge.food.dtu.dk/services/ResFinderFG-1.0/) was based on the data coming from 4
70
publications, while FARME DB includes data from 30 publications, mainly reporting
71
environmental genes which were not necessarily cured to include only ARGs sequences63. Here,
72
we report a new version of the ResFinderFG database, ResFinderFG v2.0, providing well-
73
curated data from functional metagenomics publications available until 2021 that include
74
environmental and host-associated samples.
75
Methods
76
Construction of ResFinderFG v.2.0
77
To retrieve publications using functional metagenomics for the identification of antibiotic
78
resistance genes, the 4 publications used to construct ResFinderFG v1.0 were first considered.
79
Then, all the publications which were cited by these 4 publications and all the publications that
80
cited one of these publications were collected. In addition, publications found with the following
81
terms on PubMed: “functional metagenomics” AND “antibiotic resistance”, were added to this
82
pool. After filtering out all the reviews, publications were screened one by one to check whether
83
functional metagenomics was actually used to study antibiotic resistance and whether insert
84
sequences described were available. Database construction and curation was then performed
85
as follows (Figure 1). Accession numbers describing insert DNA sequences functionally selected
86
using antibiotics were included and DNA sequences were retrieved using Batch Entrez. CD-
87
HIT64 was used to remove redundant DNA sequences and annotation of the remaining was done
88
using PROKKA v.1.1465. To specifically select insert DNA sequences with ARG annotations, a
89
representative pool of ARG annotations was obtained by applying the PROKKA annotation
90
process to the ResFinder v4.0 database. Resulting annotations were used as a reference to
91
specifically select insert DNA sequences containing an ARG. Accession number of the
92
remaining inserts were used to retrieve information on the insert DNA sequences, such as the
93
origin of the sample and the antibiotic used for selection. Additional filtering steps were added to
94
check the antibiotic used for selection and ARG annotation link, minimum gene size with at list
95
the median amino acid (aa) size of antibiotic resistance determinant (ARD) from the same ARD
96
family (260 aa for beta-lactamase, 378 for tetracycline efflux genes, 641 aa for tetracycline
97
resistance ribosomal protection genes, 178 aa for chloramphenicol acetyltransferase, 247 aa for
98
methyltransferase genes and 158 aa for dihydrofolate reductase genes) and the presence of one
99
unique annotation corresponding to an ARG on the insert DNA sequence. The database also
100
includes metadata (habitat categorization, antibiotic used for selection, ARG family) and ARO
101
annotation for each gene for comparability with other databases using ARO ontologies8.
102
103
Figure 1: ResFinderFG v.2.0 construction workflow.
104
Description of ResFinderFG v2.0
105
To assess the update of the ResFinderFG v2.0 database, the database was first compared in
106
terms of number of ARGs, ARG families and sample sources with ResFinderFG v1.0. ARG
107
families were categorized according to the antibiotic families they conferred resistance to:
108
glycopeptides/cycloserine,
sulfonamides/trimethoprim,
beta-lactams,
aminoglycosides,
109
macrolides-lincosamides-streptogramins, tetracyclines, phenicols and quinolones. Sample
110
sources were categorized as follows: aquatic, animal-associated, human-associated, plants-
111
associated, polluted environment and soil. Then, to detect the presence of ARGs in several gene
112
subcatalogs (human gut, soil and marine-freshwater) coming from the Global Microbial Gene
113
Catalog (GMGC, https://gmgc.embl.de/download.cgi2), ABRicate66 was run using default
114
parameters with different databases (ResFinderFG v2.0, ResFinder v4.0, CARD v3.0.8, ARG-
115
ANNOT v5, NCBI v3.6).
116
Data and code availability
117
All the computational steps and data used in the construction of the ResFinderFG v2.0 database
118
and the database itself are available on the following public GitHub repository:
119
https://github.com/RemiGSC/ResFinder_FG_Construction. The database was also deposited on
120
the Center of Genomic Epidemiology (CGE) server, where it can be used online
121
https://cge.food.dtu.dk/services/ResFinderFG/. Analysis processes for the description of
122
ResFinderFG
v2.0
are
accessible
on
the
following
public
GitHub
repository:
123
https://github.com/RemiGSC/ResFinder_FG_Analysis.
124
Results
125
Construction of ResFinderFG v2.0
126
A total of 50 publications using functional metagenomics to analyze ARG content were selected,
127
resulting in 23’776 accession numbers. CD-HIT identified 2’629 perfectly redundant insert
128
sequences (100% sequence identity). PROKKA identified 41’977 open reading frames (ORFs).
129
Among them, 7’787 ORFs matched with an ARG annotation of ResFinder v4.0 (228 unique ARG
130
annotations). Another 1’165 ORFs were removed because of a discordance between the
131
annotation and the antibiotic used for selection in the functional metagenomics experiment,
132
1’064 for an unexpected size relative to the ARG family, and 398 because more than one
133
putative ARG was present in the insert. A second round of CD-HIT was used to avoid
134
redundancy (100% sequence identity) in the ARG sequences and 3’913 ARGs remained and
135
form the database.
136
Comparison with ResFinderFG v1.0
137
First, the ARGs present in ResFinderFG v.2.0 were compared to the ones present in
138
ResFinderFG v1.0 (Figure 2). A total of 1’631 new ARGs were present in ResFinderFG v.2.0,
139
mainly due to new glycopeptides/cycloserine (+906 genes) and beta-lactams (+333 genes)
140
resistance genes. The glycopeptides/cycloserine resistance genes were mostly annotated as
141
homologues of D-Ala-D-X ligase. New beta-lactams antibiotics used for functional selection
142
compared to v1.0 were cefepime, meropenem and tazobactam. Regarding the sources of ARGs,
143
new ARGs mostly originated from human-associated samples (+1’333 genes).
144
145
Figure 2: a. Number of ARGs in the ResFinderFG v.1.0 and v.2.0 databases depending on a.
146
the antibiotic families involved; b. the sample sources.
147
ARG detection in several GMGC gene subcatalogs using ResFinderFG v.2.0 and other
148
databases
149
ABRicate (default parameters) was used to detect ARGs in GMGC human gut (Figure 3), soil
150
(Supplementary Figure 1a.) and aquatic (marine and freshwater) subcatalogs (Supplementary
151
Figure 1b.). Using ResFinderFG v2.0, 3’025, 211 and 129 unigene hits were obtained analyzing
152
human gut, soil and aquatic subcatalogs respectively. The 3 most frequently detected ARG
153
families in all gene catalogs were glycopeptides/cycloserine resistance genes (20.9 to 39.7% of
154
detected ARGs), sulfonamides/trimethoprim resistance genes (21.8 to 58.1% of detected ARGs)
155
and beta-lactamases encoding genes (7.9 to 25.6% of detected ARGs). Phenicols (up to 6.0% of
156
detected ARGs), aminoglycosides (up to 5.3%), cyclines (up to 6.2%) and macrolides/
157
lincosamides/streptogramins resistance genes (up to 0.03%) were also detected. Also,
158
ResFinderFG v2.0 provides habitat information on where a given ARG was first identified by
159
functional metagenomics. A majority of ARGs identified in the gut subcatalog (90.2%) were
160
indeed initially identified in the human gut by functional metagenomics (supplementary table 2).
161
In the soil gene subcatalog, 62.6% of ARGs detected were also genes identified initially in soil
162
with functional metagenomics. However, ARGs detected in the aquatic gene subcatalog were
163
primarily first identified by functional metagenomics in soil.
164
165
Figure 3: Number of unigene hits obtained analyzing GMGC human gut subcatalog using
166
several databases (ResFinder v4.0, NCBI v3.6, ARG-ANNOT v5, ResFinderFG v2.0 and CARD
167
v3.0.8) annotated by their antibiotic family. Others: bicyclomycin, beta-lactams, bleomycin,
168
disinfectant and antiseptic agents, fosfomycin, fusidic acid, multidrug, mupirocin, nitroimidazole,
169
nucleoside, peptide, rifampicin, streptothricin.
170
To compare ResFinderFG v2.0 to other databases, we ran the same ABRicate analysis of
171
GMGC gene subcatalogs using ResFinder v.4.0, CARD v3.0.8, ARG-ANNOT v5 and NCBI v3.6.
172
ResFinderFG v2.0 identified a comparable or even greater number of ARGs compared to other
173
databases. We observed that the most frequently observed ARG family depended on the
174
database used. In the human gut gene subcatalog, glycopeptides/cycloserine resistance gene
175
was the most frequent ARG family found by ResFinderFG v2.0 (39.7% of all unigene hits
176
obtained with ResFinderFG v2.0). In contrast, the beta-lactamase family was the top ARG family
177
with ARG-ANNOT (21.2%). NCBI and ResFinder detected mostly tetracycline resistance genes
178
(20.4 and 23.8% respectively). Finally, multidrug efflux pump unigene hits were the most
179
frequent using CARD (39.4%).
180
ResFinderFG v2.0 was the database with the highest fraction of database-specific hits, with
181
89.1% of specific unigene hits composed mainly by glycopeptides/cycloserine resistance genes
182
(D-alanine-D-alanine ligase ; supplementary table 3) and sulfonamides/trimethoprim resistance
183
genes (dihydrofolate reductase). By comparison, CARD had 73.7% of specific unigene hits,
184
mostly composed by gene encoding multidrug efflux pumps. Of note, 16.2% of unique CARD
185
specific multidrug efflux pump unigene hits found in the human gut were regulatory genes
186
(supplementary Table 3).
187
Between 2.6 and 4.2% of all unigene hits depending on the gene subcatalog analyzed were
188
shared by all the databases used. Beta-lactamases – encoding genes were the most prevalent
189
among them (ranging from 38.1 to 51.3% of the shared unigene hits), followed by, phenicols,
190
aminoglycosides and tetracyclines resistance genes. However, 25.1, 23.2 and 46.3% of beta-
191
lactamases, aminoglycosides and phenicols resistance genes respectively, were only detected
192
using ResFinderFG v2.0 (Figure 3; supplementary Figure 1).
193
Discussion
194
ResFinderFG v2.0 contains 3’913 ARGs which were described with functional metagenomics in
195
50 publications. Here, we showed that using ResFinderFG v2.0 enabled us to describe the
196
resistome with ARGs that were not detected by other databases. Notably, ResFinderFG v2.0
197
permitted a better description of sulfonamides/trimethoprim, glycopeptides/cycloserine resistant
198
genes and beta-lactamase encoding genes.
199
Exhaustive description of ARG content in the environment can be complicated since most ARG
200
databases are biased towards ARGs coming from culturable and/or pathogenic bacteria. One
201
way to detect genes that are not described in such databases, or that are too different from
202
described genes, is to use functional metagenomics: a laborious and low throughput method that
203
was used by only a few research groups10 which allows phenotypic identification rather than
204
sequence-based identification of ARGs. Yet, most of the ARGs characterized using functional
205
metagenomics were not deposited in ARG databases until the creation of ResFinderFG v1.0 in
206
201661. Since then, the database has not been updated but another database called FARME DB
207
was made including data coming from 30 publications62. Nevertheless, it contains all the inserts
208
sequences selected with functional metagenomics and therefore it also contains genes that are
209
not ARGs63. Therefore, we updated ResFinderFG v1.0 by including more publications using
210
functional metagenomics to characterize ARGs and we made a curation effort to ensure that the
211
sequences described are the unique ARGs responsible for the resistance phenotype in the initial
212
insert sequence.
213
ResFinderFG v2.0 includes more ARGs coming from human-associated samples12–22. For
214
example, characterization of the gut resistome with functional metagenomics showed that its
215
ARGs were not well described in ARG databases14. Inclusion of these ARGs is therefore
216
important for future metagenomic characterization of resistomes. Regarding the ARG family
217
concerned, most of the new ARGs included compared to ResFinderFG v1.0 are
218
glycopeptides/cycloserine
or
beta-lactams
resistance
genes.
Glycopeptides/cycloserine
219
resistance genes were selected using cycloserine, an antibiotic used in the therapy of
220
tuberculosis caused by multi resistant mycobacteria67. Beta-lactams resistant genes are of high
221
concern because beta-lactams antibiotics are widely used against priority pathogens68.
222
Using ResFinderFG v2.0, sulfonamides/trimethoprim, glycopeptides/cycloserine, beta-lactams,
223
phenicols, cyclines, quinolones, macrolides/lincosamides/streptogramins and aminoglycosides
224
resistance genes were evidenced studying three GMGC gene subcatalogs (human gut, soil and
225
aquatic). As expected, regarding their representation in the database, the most frequent unigene
226
hits were glycopeptide, cycloserine, sulfonamides/trimethoprim resistance genes. Analogous
227
analyses performed with other databases showed that ResFinderFG v2.0 detected a
228
comparable or higher number of ARGs depending on the other database used. Beta-lactamase
229
encoding genes were the most represented ARGs in unigene hits shared by all databases. Yet,
230
ResFinderFG v2.0 allowed the detection of a significant proportion of beta-lactamases encoding
231
genes which were not detected with other databases. It was expected since many publications
232
using functional metagenomics reported beta-lactamase encoding genes distant from the ones
233
described in ARG databases11,14,31,34,46,54,59 and a distant one has been evidenced recently from
234
soil samples39. Other antibiotic families were even more specifically associated with
235
ResFinderFG v2.0, such as sulfonamides/trimethoprim, phenicols, glycopeptides/cycloserine
236
resistance genes.
237
Our study has limitations, however. To ensure that genes included are true ARGs, we selected
238
only insert which had part of their sequence annotated as an ARG by PROKKA. Thus, ARGs not
239
identified by PROKKA may have been missed. Moreover, we did not recheck whether described
240
sequences were actually conferring resistance in vitro. Only the sequence corresponding to the
241
ARG annotation was included and we were not able to determine if the surrounding insert DNA
242
sequence was required to produce the resistant phenotype. Yet, since the original accession
243
numbers are available in each ResFinderFG v2.0 ARG sequence header, researchers can
244
easily obtain the complete insert DNA sequence to investigate.
245
Conclusion
246
ResFinderFG v2.0 is the new version of the ResFinderFG v1.0 database and includes 1’631
247
additional ARGs. This makes possible the detection of ARGs which would not be identified using
248
other currently used databases. Nevertheless, other databases also contain ARGs that are
249
absent in ResFinderFG v2.0. Therefore, to make an exhaustive description of the resistome of a
250
sample, ResFinderFG v2.0 should be used alongside other databases.
251
Acknowledgements
252
The authors are grateful to Frank Møller Aarestrup and Maja Weiss for hosting ResFinderFG
253
v2.0 on the Center for Genomic Epidemiology website, and to Andrew Bielski for English editing.
254
Conflicts of interest
255
All authors: none
256
Funding
257
This work was funded by the Joint Program Initiative for Antimicrobial Resistance (JPIAMR)
258
EMBARK (Establishing a Monitoring Baseline for Antimicrobial Resistance in Key environments)
259
project (International Development Research Centre, IDRC, grant 109304-001 to LPC, Agence
260
Nationale de la Recherche, ANR, grant ANR-19-JAMR-0004 to ER).
261
References
262
1.
WHO. Antimicrobial resistance. https://www.who.int/news-room/fact-
263
sheets/detail/antimicrobial-resistance.
264
2.
Coelho, L. P. et al. Towards the biogeography of prokaryotic genes. Nature 601, 252–
265
256 (2022).
266
3.
Mackenzie, J. S. & Jeggo, M. The One Health Approach—Why Is It So Important? Trop.
267
Med. Infect. Dis. 4, 88 (2019).
268
4.
Martínez, J. L., Coque, T. M. & Baquero, F. What is a resistance gene? Ranking risk in
269
resistomes. Nat. Rev. Microbiol. 13, 116–123 (2015).
270
5.
Forsberg, K. J. et al. The shared antibiotic resistome of soil bacteria and human
271
pathogens. Science 337, 1107–1111 (2012).
272
6.
Waseem, H. et al. Contributions and Challenges of High Throughput qPCR for
273
Determining Antimicrobial Resistance in the Environment: A Critical Review. Molecules 24, 163
274
(2019).
275
7.
Boolchandani, M., D’Souza, A. W. & Dantas, G. Sequencing-based methods and
276
resources to study antimicrobial resistance. Nat. Rev. Genet. 20, 356–370 (2019).
277
8.
Alcock, B. P. et al. CARD 2020: antibiotic resistome surveillance with the comprehensive
278
antibiotic resistance database. Nucleic Acids Res. 48, D517–D525 (2020).
279
9.
Zankari, E. et al. Identification of acquired antimicrobial resistance genes. J. Antimicrob.
280
Chemother. 67, 2640–2644 (2012).
281
10.
Dos Santos, D. F. K., Istvan, P., Quirino, B. F. & Kruger, R. H. Functional Metagenomics
282
as a Tool for Identification of New Antibiotic Resistance Genes from Natural Environments.
283
Microb. Ecol. 73, 479–491 (2017).
284
11.
Forsberg, K. J. et al. Bacterial phylogeny structures soil resistomes across habitats.
285
Nature 509, 612–616 (2014).
286
12.
Gibson, M. K. et al. Developmental dynamics of the preterm infant gut microbiota and
287
antibiotic resistome. Nat. Microbiol. 1, 16024 (2016).
288
13.
Moore, A. M. et al. Pediatric fecal microbiota harbor diverse and novel antibiotic
289
resistance genes. PloS One 8, e78822 (2013).
290
14.
Sommer, M. O. A., Dantas, G. & Church, G. M. Functional Characterization of the
291
Antibiotic Resistance Reservoir in the Human Microflora. Science 325, 1128–1131 (2009).
292
15.
Pehrsson, E. C. et al. Interconnected microbiomes and resistomes in low-income human
293
habitats. Nature 533, 212–216 (2016).
294
16.
Reynolds, L. J., Anjum, M. F. & Roberts, A. P. Detection of a Novel, and Likely Ancestral,
295
Tn916-Like Element from a Human Saliva Metagenomic Library. Genes 11, E548 (2020).
296
17.
Kintses, B. et al. Phylogenetic barriers to horizontal transfer of antimicrobial peptide
297
resistance genes in the human gut microbiota. Nat. Microbiol. 4, 447–458 (2019).
298
18.
Clemente, J. C. et al. The microbiome of uncontacted Amerindians. Sci. Adv. 1,
299
e1500183 (2015).
300
19.
Card, R. M. et al. Application of microarray and functional-based screening methods for
301
the detection of antimicrobial resistance genes in the microbiomes of healthy humans. PloS One
302
9, e86428 (2014).
303
20.
Cheng, G. et al. Functional screening of antibiotic resistance genes from human gut
304
microbiota reveals a novel gene fusion. FEMS Microbiol. Lett. 336, 11–16 (2012).
305
21.
Moore, A. M. et al. Gut resistome development in healthy twin pairs in the first year of life.
306
Microbiome 3, 27 (2015).
307
22.
Campbell, T. P. et al. The microbiome and resistome of chimpanzees, gorillas, and
308
humans across host lifestyle and geography. ISME J. 14, 1584–1599 (2020).
309
23.
Flórez, A. B., Vázquez, L. & Mayo, B. A Functional Metagenomic Analysis of Tetracycline
310
Resistance in Cheese Bacteria. Front. Microbiol. 8, 907 (2017).
311
24.
Versluis, D. et al. Sponge Microbiota Are a Reservoir of Functional Antibiotic Resistance
312
Genes. Front. Microbiol. 7, 1848 (2016).
313
25.
Wichmann, F., Udikovic-Kolic, N., Andrew, S. & Handelsman, J. Diverse antibiotic
314
resistance genes in dairy cow manure. mBio 5, e01017 (2014).
315
26.
Tian, B., Fadhil, N. H., Powell, J. E., Kwong, W. K. & Moran, N. A. Long-term exposure to
316
antibiotics has caused accumulation of resistance determinants in the gut microbiota of
317
honeybees. mBio 3, (2012).
318
27.
Martiny, A. C., Martiny, J. B. H., Weihe, C., Field, A. & Ellis, J. C. Functional
319
Metagenomics Reveals Previously Unrecognized Diversity of Antibiotic Resistance Genes in
320
Gulls. Front. Microbiol. 2, (2011).
321
28.
Allen, H. K. et al. Resident microbiota of the gypsy moth midgut harbors antibiotic
322
resistance determinants. DNA Cell Biol. 28, 109–117 (2009).
323
29.
Wang, X.-R. et al. Duck wastes as a potential reservoir of novel antibiotic resistance
324
genes. Sci. Total Environ. 771, 144828 (2021).
325
30.
McGivern, B. B., McDonell, R. K., Morris, S. K., LaPara, T. M. & Donato, J. J. Novel class
326
1 integron harboring antibiotic resistance genes in wastewater-derived bacteria as revealed by
327
functional metagenomics. Plasmid 114, 102563 (2021).
328
31.
Marathe, N. P. et al. Sewage effluent from an Indian hospital harbors novel
329
carbapenemases and integron-borne antibiotic resistance genes. Microbiome 7, 97 (2019).
330
32.
Cameron, A. et al. Functional screening for triclosan resistance in a wastewater
331
metagenome and isolates of Escherichia coli and Enterococcus spp. from a large Canadian
332
healthcare region. PloS One 14, e0211144 (2019).
333
33.
Zhang, L. et al. Novel clinically relevant antibiotic resistance genes associated with
334
sewage sludge and industrial waste streams revealed by functional metagenomic screening.
335
Environ. Int. 132, 105120 (2019).
336
34.
Marathe, N. P. et al. Functional metagenomics reveals a novel carbapenem-hydrolyzing
337
mobile beta-lactamase from Indian river sediments contaminated with antibiotic production
338
waste. Environ. Int. 112, 279–286 (2018).
339
35.
González-Plaza, J. J. et al. Functional Repertoire of Antibiotic Resistance Genes in
340
Antibiotic Manufacturing Effluents and Receiving Freshwater Sediments. Front. Microbiol. 8,
341
(2018).
342
36.
Munck, C. et al. Limited dissemination of the wastewater treatment plant core resistome.
343
Nat. Commun. 6, 8452 (2015).
344
37.
Uyaguari, M. I., Fichot, E. B., Scott, G. I. & Norman, R. S. Characterization and
345
quantitation of a novel β-lactamase gene found in a wastewater treatment facility and the
346
surrounding coastal ecosystem. Appl. Environ. Microbiol. 77, 8226–8233 (2011).
347
38.
Willms, I. M. et al. Discovery of Novel Antibiotic Resistance Determinants in Forest and
348
Grassland Soil Metagenomes. Front. Microbiol. 10, 460 (2019).
349
39.
Willms, I. M. et al. Novel Soil-Derived Beta-Lactam, Chloramphenicol, Fosfomycin and
350
Trimethoprim Resistance Genes Revealed by Functional Metagenomics. Antibiot. Basel Switz.
351
10, 378 (2021).
352
40.
Obermeier, M. M. et al. Plant resistome profiling in evolutionary old bog vegetation
353
provides new clues to understand emergence of multi-resistance. ISME J. 15, 921–937 (2021).
354
41.
Böhm, M.-E., Razavi, M., Flach, C.-F. & Larsson, D. G. J. A Novel, Integron-Regulated,
355
Class C β-Lactamase. Antibiot. Basel Switz. 9, E123 (2020).
356
42.
Park, K. S. et al. The novel metallo-β-lactamase PNGM-1 from a deep-sea sediment
357
metagenome: crystallization and X-ray crystallographic analysis. Acta Crystallogr. Sect. F Struct.
358
Biol. Commun. 74, 644–649 (2018).
359
43.
Wang, S. et al. Tetracycline Resistance Genes Identified from Distinct Soil Environments
360
in China by Functional Metagenomics. Front. Microbiol. 8, 1406 (2017).
361
44.
Lau, C. H.-F., van Engelen, K., Gordon, S., Renaud, J. & Topp, E. Novel Antibiotic
362
Resistance Determinants from Agricultural Soil Exposed to Antibiotics Widely Used in Human
363
Medicine and Animal Farming. Appl. Environ. Microbiol. 83, (2017).
364
45.
Pawlowski, A. C. et al. A diverse intrinsic antibiotic resistome from a cave bacterium. Nat.
365
Commun. 7, 13803 (2016).
366
46.
Gudeta, D. D. et al. Expanding the Repertoire of Carbapenem-Hydrolyzing Metallo-ß-
367
Lactamases by Functional Metagenomic Analysis of Soil Microbiota. Front. Microbiol. 7, 1985
368
(2016).
369
47.
Im, H., Kim, K. M., Lee, S.-H. & Ryu, C.-M. Functional Metagenome Mining of Soil for a
370
Novel Gentamicin Resistance Gene. J. Microbiol. Biotechnol. 26, 521–529 (2016).
371
48.
Hatosy, S. M. & Martiny, A. C. The ocean as a global reservoir of antibiotic resistance
372
genes. Appl. Environ. Microbiol. 81, 7593–7599 (2015).
373
49.
Perron, G. G. et al. Functional Characterization of Bacteria Isolated from Ancient Arctic
374
Soil Exposes Diverse Resistance Mechanisms to Modern Antibiotics. PLoS ONE 10, (2015).
375
50.
Udikovic-Kolic, N., Wichmann, F., Broderick, N. A. & Handelsman, J. Bloom of resident
376
antibiotic-resistant bacteria in soil following manure fertilization. Proc. Natl. Acad. Sci. U. S. A.
377
111, 15202–15207 (2014).
378
51.
Su, J. Q., Wei, B., Xu, C. Y., Qiao, M. & Zhu, Y. G. Functional metagenomic
379
characterization of antibiotic resistance genes in agricultural soils from China. Environ. Int. 65,
380
9–15 (2014).
381
52.
Cheng, G. et al. Identification of a novel fosfomycin-resistant UDP-N-acetylglucosamine
382
enolpyruvyl transferase (MurA) from a soil metagenome. Biotechnol. Lett. 35, 273 (2013).
383
53.
López-Pérez, M., Mirete, S., Jardón-Valadez, E. & González-Pastor, J. E. Identification
384
and modeling of a novel chloramphenicol resistance protein detected by functional
385
metagenomics in a wetland of Lerma, Mexico. Int. Microbiol. Off. J. Span. Soc. Microbiol. 16,
386
103–111 (2013).
387
54.
Vercammen, K. et al. Identification of a metagenomic gene cluster containing a new class
388
A beta-lactamase and toxin-antitoxin systems. MicrobiologyOpen 2, 674–683 (2013).
389
55.
McGarvey, K. M., Queitsch, K. & Fields, S. Wide variation in antibiotic resistance proteins
390
identified by functional metagenomic screening of a soil DNA library. Appl. Environ. Microbiol.
391
78, 1708–1714 (2012).
392
56.
Torres-Cortés, G. et al. Characterization of novel antibiotic resistance genes identified by
393
functional metagenomics on soil samples. Environ. Microbiol. 13, 1101–1114 (2011).
394
57.
Lang, K. S. et al. Novel florfenicol and chloramphenicol resistance gene discovered in
395
Alaskan soil by using functional metagenomics. Appl. Environ. Microbiol. 76, 5321–5326 (2010).
396
58.
Donato, J. J. et al. Metagenomic analysis of apple orchard soil reveals antibiotic
397
resistance genes encoding predicted bifunctional proteins. Appl. Environ. Microbiol. 76, 4396–
398
4401 (2010).
399
59.
Allen, H. K., Moe, L. A., Rodbumrer, J., Gaarder, A. & Handelsman, J. Functional
400
metagenomics reveals diverse beta-lactamases in a remote Alaskan soil. ISME J. 3, 243–251
401
(2009).
402
60.
Riesenfeld, C. S., Goodman, R. M. & Handelsman, J. Uncultured soil bacteria are a
403
reservoir of new antibiotic resistance genes. Environ. Microbiol. 6, 981–989 (2004).
404
61.
Munk, P. et al. Abundance and diversity of the faecal resistome in slaughter pigs and
405
broilers in nine European countries. Nat. Microbiol. 3, 898–908 (2018).
406
62.
Wallace, J. C., Port, J. A., Smith, M. N. & Faustman, E. M. FARME DB: a functional
407
antibiotic resistance element database. Database J. Biol. Databases Curation 2017, (2017).
408
63.
Lal Gupta, C., Kumar Tiwari, R. & Cytryn, E. Platforms for elucidating antibiotic
409
resistance in single genomes and complex metagenomes. Environ. Int. 138, 105667 (2020).
410
64.
Li, W. & Godzik, A. Cd-hit: a fast program for clustering and comparing large sets of
411
protein or nucleotide sequences. Bioinforma. Oxf. Engl. 22, 1658–1659 (2006).
412
65.
Seemann, T. Prokka: rapid prokaryotic genome annotation. Bioinforma. Oxf. Engl. 30,
413
2068–2069 (2014).
414
66.
Seemann, T. Abricate.
415
67.
Di Perri, G. & Bonora, S. Which agents should we use for the treatment of multidrug-
416
resistant Mycobacterium tuberculosis? J. Antimicrob. Chemother. 54, 593–602 (2004).
417
68.
WHO. Global Antimicrobial Resistance and Use Surveillance System (GLASS) Report:
418
2021. https://www.who.int/publications-detail-redirect/9789240027336.
419
420
| 2022 | ResFinderFG v2.0: a database of antibiotic resistance genes obtained by functional metagenomics | 10.1101/2022.10.19.512667 | [
"Gschwind Rémi",
"Perovic Svetlana Ugarcina",
"Petitjean Marie",
"Lao Julie",
"Coelho Luis Pedro",
"Ruppé Etienne"
] | creative-commons |
Multimodal brain imaging study of 19,825 participants reveals adverse effects of
moderate drinking
One Sentence Summary: Moderate alcohol intake, consuming two or more units of alcohol per
day, has negative effects on brain health.
Remi Daviet, PhD1, Gökhan Aydogan, PhD2, Kanchana Jagannathan, MS3, Nathaniel Spilka,
BA3, Philipp Koellinger, PhD4, Henry R. Kranzler, MD3,5, Gideon Nave, PhD1, Reagan R.
Wetherill, PhD3*
1Marketing Department, The Wharton School, University of Pennsylvania, Philadelphia, PA, USA
2Department of Economics, University of Zurich, Switzerland
3Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania,
Philadelphia, PA, USA
4Department of Economics, School of Business and Economics, Vrije Universiteit Amsterdam.
5Crescenz VAMC, Philadelphia, PA, USA
*Corresponding Author: Reagan R. Wetherill, Department of Psychiatry, 3535 Market Street,
Suite 500, Philadelphia, PA 19104
email: rweth@pennmedicine.upenn.edu
tel: +1 (215) 746-3953
2
ABSTRACT
Alcohol consumption can have significant deleterious consequences, including brain atrophy,
neuronal loss, poorer white matter fiber integrity, and cognitive decline, but the effects of light-
to-moderate alcohol consumption on brain structure remain unclear. Here we examine the
associations between alcohol intake and brain structure using structural, diffusion tensor, and
neurite orientation dispersion and density imaging data from 19,825 generally healthy
middle-aged and older adults from the UK Biobank. Systematically controlling for potential
confounds, we found that greater alcohol consumption was associated with lower global gray
and white matter volume, regional gray matter volume in cortical and subcortical areas, and
white matter fiber integrity and complexity. Post hoc analyses revealed that these associations
were non-linear. Our findings extensively characterize the associations between alcohol intake
and gray and white matter macrostructure and microstructure. Consuming two or more units of
alcohol per day, equivalent to one drink in some establishments, could have negative effects on
brain health, an important public health finding.
3
Converging lines of research provide compelling evidence that chronic, excessive
alcohol consumption is associated with global brain atrophy and regional brain changes.1–3
Recent meta-analyses of magnetic resonance imaging (MRI) findings show that individuals with
alcohol use disorder (AUD) have less global white matter volume (WMV)4 and less gray matter
volume (GMV) - both globally and locally in corticostriatal-limbic regions5 - than healthy controls.
Further, in a meta-analysis of pooled, multinational datasets from 33 imaging sites, individuals
with AUD had lower local thickness and surface area of the hippocampus, thalamus, putamen,
and amygdala than controls.6
In studies using diffusion-weighted MRI (dMRI), which allows a non-invasive
investigation of white matter microstructure via measures of water molecule diffusion,
individuals with AUD had lower fractional anisotropy (FA; the directional coherence of water
molecule diffusion) and greater mean diffusivity (MD; the magnitude of water molecule diffusion)
in the corpus callosum, frontal forceps, internal and external capsules, fornix, superior cingulate,
and longitudinal fasciculi than controls.1,7 However, because conventional dMRI measures (FA
and MD) are based on a simplistic model of brain tissue microstructure, they fail to account for
the complexities of neurite geometry.8 For example, the lower FA observed in individuals with
AUD1,7 may reflect lower neurite density and/or greater orientation dispersion of neurites, which
conventional dMRI measures do not differentiate.9,10 A key question raised by prior findings in
individuals with AUD that remains is whether, similar to heavy drinking, light-to-moderate
alcohol consumption adversely affects brain structure. Further, is the relationship between
alcohol intake and brain structure linear? In some studies of middle-aged and older adults,
moderate alcohol consumption was associated with lower total cerebral volume,11 gray matter
atrophy,12,13 and lower density of gray matter in frontal and parietal brain regions.13 However,
other studies have shown no association,14 and one study showed a positive association
between light-to-moderate alcohol consumption and GMV in older men.15 One interpretation of
these findings is that a U-shaped, dose-dependent association exists between alcohol use and
4
brain morphometry, with light-to-moderate drinking being protective against and heavy drinking
being a risk factor for lower GMV.15,16 However, these results are inconclusive, as a longitudinal
cohort study17 showed no difference in structural brain measures between abstinent individuals
and light drinkers, while moderate-to-heavy drinkers showed GMV atrophy in the hippocampi
and impaired white matter microstructure (lower FA, higher MD) in the corpus callosum.
The inconclusive nature of the evidence regarding the association between moderate
alcohol intake and brain structure may reflect the patchwork nature of the literature, which
consists of mostly small, unrepresentative studies with limited statistical power.18,19 Moreover,
most studies to date have not accounted for the effects of many relevant covariates and
therefore have yielded findings with limited generalizability. Potential confounds that may be
associated with individual differences in both alcohol intake and neuroanatomy include sex,20
body mass index (BMI),21 age,22 and genetic population structure (i.e., biological characteristics
that are correlated with environmental causes).23 Similar to other fields of research, progress in
this area may also be limited by publication bias.24
The current study used data from nearly 20,000 participants in the UK Biobank (UKB) to
characterize the associations between alcohol intake (i.e., mean units per day; one unit=10 ml
of pure ethanol) and brain structure (total GMV and WMV, regional GMV) and white matter
microstructure in the 27 major tracts (Fig. 1). We addressed the limitations of the existing
literature through a pre-registered analysis of multimodal imaging data from the UKB.25–27 The
UKB, a prospective cohort study representative of the United Kingdom (UK) population aged 40-
69 years, is the largest available collection of high-quality MRI brain scans, alcohol-related
behavioral phenotypes, and measurements of the socio-economic environment. Participants
self-reported their usual weekly alcohol consumption through a touch screen questionnaire,26
from which we calculated mean alcohol intake. A subsample of participants completed a brain
imaging scan session that included three structural modalities, resting and task-based fMRI, and
diffusion-weighted imaging.25–27 Importantly, the dMRI measures available in the UKB include
5
the conventional metrics of FA and MD, but also neurite orientation dispersion and density
imaging (NODDI).10 Such measures offer information on white matter microstructure and
Figure 1. Brain imaging regions of interest according to the Harvard-Oxford Atlas (top:
cortical regions) and AutoPtx (bottom: white matter tracts) from Cox and colleagues
(2019).
40
6
estimates of neurite density (i.e., intra-cellular volume fraction; ICVF), extracellular water
diffusion (i.e., isotropic volume fraction; ISOVF), and tract complexity/fanning (i.e., orientation
dispersion, OD). This allowed us to assess the nature of alcohol’s effects on white matter
microstructure in greater detail than any previous studies on the topic.
The richness and scale of the UKB dataset also enabled us to control for many important
confounds, including genetic population structure, and to estimate small effects accurately,
including those of moderate drinking on brain structure. Because regular moderate drinking is
the most common pattern of consumption in the UK, where 57% of adults, or an estimated 29.2
million individuals, reported drinking in the previous week,28 our findings have important public
health implications for the UK and other countries where alcohol is commonly consumed.
RESULTS
Characteristics of the 19,825 participants (52.5% female) are shown in Table 1.
Participants were healthy middle-aged and older adults.
We first estimated a series of linear regressions with alcohol intake as the main
explanatory variable of interest and imaging-derived phenotypes (IDPs) extracted by the UKB
brain imaging processing pipeline29 as the dependent variables, controlling for sex, age, age-
squared, age-cubed, height, total brain volume (grey+white matter, for volumetric data only), the
Townsend index of social deprivation measured at the zip code level30 and handedness. Family-
wise error (FWE) correction of the p-values was applied using the Holm method.31 This analysis
was pre-registered in the Open Science Network
(https://osf.io/trauf/?view_only=a3795f76c5a54830b2ca443e3e07c0f0).
We measured alcohol intake in log(1 + units/day) with one unit representing 10 ml of
pure ethanol and found that it was associated with lower global GMV (standardized β=-0.080
[95% CI -0.093 to -0.067], t=-12.17, p<1.0x10-6) and lower global WMV (standardized β=-0.044
[95% CI -0.059 to -0.028], t=-5.70, p<1.0x10-6). When controlling for total brain volume, we also
identified negative associations between alcohol intake and regional GMV in 16 brain regions
7
Table 1. Descriptive characteristics of the population
Study Sample
N = 19,825
Heavy Drinkers
n = 1,226
Abstainers
n = 1,527
Test
Statistic
Mean age (y) (SD)
62.7 (7.4)
62.7 (7.0)
63.3 (7.5)
t=-2.14*
Sex [n, (%) women]
10,406 (52.5)
488 (39.8)
1,058 (69.2)
z=15.47**
Population group (% white)
100
100
100
∙∙
Education (y) (SD)
13.5 (4.0)
13.7 (4.1)
13.4 (4.1)
t=1.87
Alcohol units per week (SD)
8.2 (8.2)
29.7 (9.4)
0.0
∙∙
BMI (SD)
26.6 (4.3)
27.3 (3.9)
27.2 (5.4)
t=0.76
Total GMV + WMV (cm3) (SD)
1,167.6 (111.3)
1,169.8 (104.8)
1,140.4 (109.9)
t=7.19**
Total GMV (cm3) (SD)
616.7 (55.3)
615.6 (52.3)
604.4 (55.2)
t=4.74**
Total WMV (cm3) (SD)
551.6 (62.0)
556.6 (59.1)
536.1 (60.1)
t=7.43**
Note: *p-value<0.05, ** p-value<0.001, z-statistic (proportions) and t-statistic (means) between
heavy drinkers and abstainers. BMI: body mass index, GMV: gray matter volume, WMV: white
matter volume, y: years
(standardized β range -0.048 to -0.020) (Supplementary Table 1) demonstrating local
associations that were above and beyond the global effects. Alcohol intake was also associated
with poorer white matter microstructure (lower FA, ICVF, and OD; higher MD and ISOVF)
(Supplementary Tables 2 and 3). Additional analyses that adjusted also for weight, BMI, and
educational attainment, revealed one additional association between alcohol intake and regional
GMV (left precentral gyrus, standardized β = -0.027 [95% CI -0.039 to -0.016], t = -4.076, p =
2.55 x 10-6) and an association between alcohol intake and ICVF in the bilateral anterior
thalamic radiation (left: standardized β = -0.027 [95% CI -0.041 to -0.013], t = -3.711, p = 2.07 x
10-4; right: standardized β = -0.028 [95% CI -0.043 to -0.014, t = -3.888, p = 1.01 x 10-4),
whereas the association between alcohol intake and the left lateral occipital cortex was no
longer statistically significant. The remaining regional associations with alcohol intake were, in
general, smaller than the global effects.
The strongest regional GMV effects identified above the global effects were in the
bilateral putamen (left: standardized β = -0.051 [95% CI -0.065 to -0.037], t = -7.087, p = 1.42 x
8
10-12; right: standardized β = -0.047 [95% CI -0.061 to -0.033], t = -6.664, p = 2.72 x 10-11) and
brain stem (standardized β = -0.033 [95% CI -0.045 to -0.020], t = -5.299, p = 1.18 x 10-7). Some
of our linear regressions showed positive associations between drinking and regional GMV
relative to the global effect. Specifically, greater alcohol intake was associated with greater
regional GMV in the bilateral pallidum (left: standardized β = 0.029 [95% CI 0.015 to 0.043], t =
3.938, p = 8.23 x 10-5 right: standardized β = 0.033 [95% CI 0.018 to 0.047], t = 4.473, p = 7.76 x
10-6), right inferior temporal gyrus (standardized β = 0.026 [95% CI 0.013 to 0.038], t = 4.040, p
= 5.35 x 10-5), and left lingual gyrus (standardized β = 0.023 [95% CI 0.013 to 0.034], t = 4.242,
p = 2.23 x 10-5). We estimated additional regression models to determine whether these
associations were positive in absolute value, or only relatively to the global effects. After
removing total brain volume as a control variable from the linear regression models, the
associations between alcohol intake and these regional GMV IDPs were no longer significant.
This finding suggests that the association between alcohol intake and brain structure is
negative, and likely occurs in stages over time, with alcohol intake affecting specific, perhaps
more vulnerable, brain regions before influencing other regions (e.g., bilateral pallidum).
In our analyses using dMRI IDPs, alcohol intake was associated with lower FA and
higher MD in the bilateral posterior thalamic radiation fibers and forceps minor (Supplementary
Table 2). Associations with thalamic radiation fibers (anterior, posterior, and superior) and the
forceps minor (Fmin) were among the largest in magnitude and found across all white matter
measures. As shown in Supplementary Tables 2 and 3, there were also associations between
alcohol intake and MD, ISOVF, ICVF, and/or OD in several association fibers [inferior fronto-
occipital fasciculus (IFOF), inferior longitudinal fasciculus (ILF); superior longitudinal fasciculus
(SLF); uncinate], and projection fibers [acoustic radiation (AR); forceps major (Fmaj);
corticospinal tract (CST); middle cerebellar peduncle (MCP)].
Pre-registered sensitivity analyses that re-estimated our regression models while
excluding heavy drinkers or non-drinkers altered the sign and/or magnitude of several of the
9
effects that we observed, suggesting that the associations vary non-linearly across the drinking
range. Thus, after we regressed out the effects of covariates (this time excluding brain volume,
in order to determine absolute effects), we grouped participants into five bins based on their
average daily drinking level (0, 0-1, 1-2, 2-3, 4+ units) and quantified the average levels of the
IDPs in each group (Fig. 2 and Supplementary Figs 1-5). The associations were mainly driven
by individuals who reported consuming at least two or more units of alcohol per day, with no
substantial effects of alcohol intake among individuals who reported consuming less than two
units/day. Figure 2 displays association patterns in the brain stem, left putamen, and left lateral
occipital cortex where individuals who consumed two or more units/day showed lower average
regional GMV. The most substantial association between alcohol intake and regional GMV
occurred among individuals who consumed four or more units/day. These individuals showed
FIGURE 2. Average regional gray matter volume based on daily drinking levels in the brain
stem, left putamen, and left lateral occipital cortex. Daily unit = 10 ml of pure ethanol.
10
lower average local GMV throughout the brain. The bilateral pallidum was the only regional
GMV IDP that did not show significant differences across average daily drinking level.
FIGURE 3. Average white matter water microstructure indices based on daily drinking levels
in the bilateral thalamic radiation fibers. Daily unit = 10 ml of pure ethanol. FA = fractional
anisotropy; MD = mean diffusion; ICVF = intracellular volume fraction; ISOVF = isotropic
volume fraction; OD = orientation diffusion; MD, ISOVF, and OD values are represented as
negative values.
11
We conducted similar analyses focusing on the five white matter microstructure
measures across 27 white matter tracts. In Figure 3 and Supplementary Figures 6 and 7, we
present the results for white matter tracts across the average daily drinking levels where the
mean residuals for three or more microstructure measures were significant. The majority of
associations between alcohol intake and white matter microstructure measures reflected less
healthy white matter -- that is a combination of lower FA, higher MD, lower ICVF, higher ISOVF,
and/or lower OD -- among individuals who reported consuming three or more units/day of
alcohol. These findings were evident across bilateral thalamic radiation fibers (ATR, PTR, STR)
(Fig. 3), association fibers (bilateral IFOF, bilateral ILF, bilateral SLF, right CingG, and right
UNC), and projection fibers (Fmin and MCP) (Supplementary Figs. 6 and 7). A positive
association between alcohol intake and FA was observed in the bilateral CST, where individuals
with greater alcohol intake had higher FA than non-drinkers who consumed less than two
units/day.
DISCUSSION
We conducted a multimodal brain imaging study of nearly 20,000 middle-aged and older
adults of European descent, a population sample that reported alcohol consumption across the
entire spectrum from abstinence to heavy drinking. The scale and granularity of the data
provided ample statistical power to identify small effects and explore non-linear dependencies
while accounting for important potential confounds. Associations between greater alcohol intake
and poorer brain health were small but significant across global brain measures and cortical and
subcortical gray matter and white matter microstructure. The comprehensiveness and sensitivity
of these findings add to our understanding of the associations between alcohol intake and brain
health in humans.
Although the link between alcohol intake and less healthy brain tissue was
predominantly driven by heavy drinkers, effects were also observed among individuals who
reported consuming two units/day of alcohol. This has important implications for
12
recommendations regarding safe drinking levels. In 2016, the UK Chief Medical Officers
published new “low-risk” alcohol consumption guidelines that advise limiting alcohol intake to 14
units per week32. One unit of alcohol is equivalent to 10 ml or 8 g of ethanol, which is contained
in 25 ml of 40% spirits, 250 ml of 4% beer, and 76 ml of 13% wine. Many drinking
establishments serve drinks that contain 35-50 ml of 40% spirits (1.4-2 units), 568 ml of 4% beer
(2.27 units), and 175 ml of 13% wine (2.30 units).33 Thus, in the UK, consuming just one
alcoholic drink (i.e., two units of alcohol) daily could have negative effects on brain health. This
has important public health implications insofar as 57% of UK adults, or an estimated 29.2
million individuals,28 endorsed past-week drinking.
Associations between measures of brain structure and alcohol intake were generally in
the expected direction, providing additional evidence of the negative effects of low-to-moderate
alcohol consumption on brain structure. Alcohol is a neurotoxic agent that induces brain
oxidative stress,34 alters neuroimmune response,35 damages myelin,7 and alters
neurotransmission and neurotransmitter systems.2 These alterations interfere with neural
function, resulting in cognitive impairments, and are likely associated with changes in dendritic
spine formation.36,37 Thus, it is not surprising that low-to-moderate alcohol consumption was
associated with less global GMV, global WMV, and regional GMV, and less healthy white matter
structure. Although the exact mechanisms of alcohol’s neurotoxic effects are still under
investigation, our findings provide the first evidence of an association between alcohol intake
and neurite orientation diffusion and density. Specifically, alcohol consumption was associated
with lower neurite density, lower tract complexity and greater water diffusion in thalamic
radiations and association fasciculi, and may reflect the effect of alcohol on myelin and axonal
fibers. Future investigations into the mechanisms underlying the neurotoxic effects of alcohol on
the brain, particularly among occasional binge drinkers (e.g. college students), are warranted.
Our findings also have implications for the design and analysis of future studies using
brain images in general population samples such as the UKB. A failure to account for the
13
effects of drinking, either by controlling for alcohol intake or excluding participants who drink
more than one drink (two units of alcohol) per day (which comprised 44% of our study
population), could introduce an unwelcome source of variance into the analysis. Furthermore,
while neuroimaging studies commonly examine linear relationships between brain features and
other explanatory variables, our results demonstrate that the linearity assumption underlying
most studies could be overly simple.
Our study is not without limitations and these provide opportunities for further research.
First, we relied on a sample of middle-aged individuals of European ancestry living in the UK.
We hope that future work will test the generality of our findings among individuals from other
populations, and in other age groups. It is reasonable to expect that the relationship we
observed would differ in younger individuals who have not experienced the chronic effects of
alcohol on the brain. An additional limitation stems from the self-reported measures of alcohol
intake in the UK Biobank, which covers only the past year. Such estimates do not adequately
reflect drinking prior to the past year and are susceptible to reporting and recall bias.38,39
In summary, in this comprehensive examination of the associations between alcohol
intake and brain macro- and micro-structure, we uncovered multiple associations. The
associations were most pronounced in heavy drinkers, yet some effects were observed among
individuals who reported consuming two or more units/day of alcohol. These findings provide an
extensive characterization of the associations between alcohol intake and gray and white matter
macrostructure and microstructure, and offer insights into the potential effects of light-to-
moderate alcohol consumption on brain architecture.
Methods
Sample
All UK Biobank (www.ukbiobank.ac.uk) participants provided written informed consent
and ethical approval was granted by the North West Multi-Centre Ethics committee. Our sample
14
comprised 19,825 individuals of European ancestry from the UKB database whose data were
available as of October 18, 2018. The number of participants included in each model decreased
when phenotype data were missing. All of the structural T1 MRI images that we used passed
the automated quality control of the UKB brain imaging processing pipeline.29 We ran additional
quality checks using the Computational Anatomy Toolbox (CAT; www.neuro.uni-jena.de/cat/) for
SPM (www.fil.ion.ucl.ac.uk/spm/software/spm12/), which resulted in 747 individuals who
exhibited substantial image inhomogeneity (i.e., overall volume correlation below two standard
deviations from the mean) being removed from the analysis.
Measures of alcohol consumption
Participants self-reported the number of units of alcohol (10 ml of pure ethanol)
consumed “in an average week” in several beverage categories in “units per week” (for frequent
drinkers) or “units per month” (for less frequent drinkers). The UKB assessment defined units of
alcohol as follows: a pint or can of beer/lager/cider = two units; a 25-ml single shot of spirits =
one unit; and a standard glass of wine (175 ml) = two units. The categories are red wine, white
wine/champagne, beer/cider, spirits, fortified wine, and “other”. Number of weekly units was
computed by summing the weekly number of units for all categories. When reported monthly,
the intake was converted to units per week by dividing by 4.3. Number of weekly units was
divided by seven to determine units per day.
MRI data acquisition
Participants were scanned using a Siemens Skyra 3T scanner (Siemens Healthcare,
Erlangen, Germany) using a standard 32-channel head coil, according to a freely available
protocol (http://www.fmrib.ox.ac.uk/ukbiobank/protocol/V4_23092014.pdf), documentation
(http://biobank.ctsu.ox.ac.uk/crystal/docs/brain_mri.pdf), and publication.40 As part of the
scanning protocol, high-resolution T1-weighted images, three-dimensional T2-weighted fluid-
attenuated inversion recovery (FLAIR) images, and diffusion data were obtained. High
resolution T1-weighted images were obtained using an MPRAGE sequence with the following
15
parameters: TR=2000ms; TE=2.01ms; 208 sagittal slices; flip angle, 8°; FOV=256 mm;
matrix=256×256; slice thickness=1.0mm (voxel size 1×1×1mm); total scan time=4min 54s. 3D
FLAIR images were obtained with the following parameters: TR=1800ms; TE=395.0ms; 192
sagittal slices; FOV=256mm; 256×256; slice thickness=1.05mm (voxel size 1.05×1×1mm); total
scan time=5min 52s. Diffusion acquisition comprised a spin-echo echo-planar sequence with
10 T2-weighted (b ≈ 0 s mm−2) baseline volumes, 50 b = 1000 s mm−2 and 50 b = 2000 s mm−2
diffusion weighted volumes, with 100 distinct diffusion-encoding directions and 2 mm isotropic
voxels; total scan time=6min 32s.
MRI data preprocessing
Structural imaging and diffusion data were processed by the UK Biobank team and
made available to approved researchers as imaging-derived phenotypes (IDPs); the full details
of the image processing and QC pipeline are available in an open access article.25,29 IDPs used
in analyses included total brain volume, gray matter volume, white matter volume, 139 regional
GMV IDPs derived using parcellations from the Harvard-Oxford cortical and subcortical atlases
and Diedrichsen cerebellar atlas (UKB fields 25782 to 25920), and tract-averaged measures of
fractional anisotropy (FA), mean diffusivity (MD), intra-cellular volume fraction (ICVF), isotropic
volume fraction (ISOVF), and orientation diffusion (OD). White matter measures were used from
the following white matter tracts: middle cerebellar peduncle (MCP), forceps major (FMaj),
forceps minor (FMin) and bilateral medial lemnisci, corticospinal tract (CST), acoustic radiation
(AR), anterior thalamic radiation (ATR), posterior thalamic radiation (PTR), superior thalamic
radiation (STR), superior longitudinal fasciculus (SLF), inferior longitudinal fasciculus (ILF) and
inferior fronto-occipital fasciculus (IFOF), and both the cingulate gyrus and parahippocampal
portions of the cingulum bundle. Individuals whose IDPs were more than four standard
deviations from the mean were excluded from analyses.
Statistical analyses
16
We pre-registered the analysis plan
(https://osf.io/trauf/?view_only=a3795f76c5a54830b2ca443e3e07c0f0). Our main analysis
(model A) estimated a linear regression with the IDPs as dependent variables and alcohol
intake (log(1 + units/week)) as the main independent variable of interest, controlling for sex,
age, age-squared, age-cubed, height, brain volume, the Townsend index of social deprivation
measured at the zip code level30 and handedness (right/left/ambidextrous; dummy-coded). To
control for genetic population structure, the models also included the first 40 genetic principal
components,41 and dummy-coded county of residence.42 All continuous variables (except age-
related variables) were standardized to a mean of 0 and a standard deviation of 1. We also
performed three sensitivity analyses. Model (B) included additional control for variables that are
potential downstream effects associated with alcohol intake: weight, body mass index (BMI),
and educational attainment.43 The two other models repeated the analysis of model (A), with
Model (C) excluding non-drinkers and model (D) excluding heavy drinkers (i.e., women who
reported consuming more than 18 units/week and men consuming more than 24 units/week).
Once we identified IDPs that were robustly associated with alcohol intake using linear
regression models, we investigated whether the associations were dose dependent. For
example, deleterious effects of alcohol on GMV of a specific brain region could occur only in
heavy drinkers. Hence, we binned participants in the following six categories based on average
alcohol intake: (1) abstainers (n=1,527), (2) individuals who drank less than one unit/day
(n=9,595) (3) individuals who drank between one (included) and two (excluded)
units/day (maximal amount recommended, n=5,189) (4) individuals who drank between two
(included) and three (excluded) units/day (n=2,215), (5) individuals who drink between three
(included) and four (excluded) units/day (n=805), and (6) individuals who drink at least four
units/day (n=568). We then calculated the mean IDP values (after regressing the influence of all
control variables specified in model A) and 95% confidence intervals (CI) around them.
Statistical significance
17
To control the family-wise error rate, we determined the significance thresholds for all
regressions using the Holm method31 and included the results from all IDPs.
18
References
1.
Bühler, M. & Mann, K. Alcohol and the human brain: a systematic review of different
neuroimaging methods. Alcohol. Clin. Exp. Res. 35, 1771–1793 (2011).
2.
Sutherland, G. T., Sheedy, D. & Kril, J. J. Neuropathology of alcoholism. Handb. Clin.
Neurol. 125, 603–615 (2014).
3.
Zahr, N. M., Kaufman, K. L. & Harper, C. G. Clinical and pathological features of alcohol-
related brain damage. Nat. Rev. Neurol. 7, 284–294 (2011).
4.
Monnig, M. A., Tonigan, J. S., Yeo, R. A., Thoma, R. J. & McCrady, B. S. White matter
volume in alcohol use disorders: a meta-analysis. Addict. Biol. 18, 581–592 (2013).
5.
Yang, X. et al. Cortical and subcortical gray matter shrinkage in alcohol-use disorders: a
voxel-based meta-analysis. Neurosci. Biobehav. Rev. 66, 92–103 (2016).
6.
Chye, Y. et al. Subcortical surface morphometry in substance dependence: An ENIGMA
addiction working group study. Addict. Biol. e12830 (2019).
7.
Pfefferbaum, A. et al. White matter microstructural recovery with abstinence and decline
with relapse in alcohol dependence interacts with normal ageing: a controlled
longitudinal DTI study. Lancet Psychiatry 1, 202–212 (2014).
8.
Churchill, N. W., Caverzasi, E., Graham, S. J., Hutchison, M. G. & Schweizer, T. A.
White matter during concussion recovery: Comparing diffusion tensor imaging (DTI) and
neurite orientation dispersion and density imaging (NODDI). Hum. Brain Mapp. 40,
1908–1918 (2019).
9.
Beaulieu, C. The basis of anisotropic water diffusion in the nervous system--a technical
review. NMR in Biomedicine: An International Journal Devoted to the Development and
Application of Magnetic Resonance In Vivo 15, 435–455 (2002).
19
10.
Zhang, H., Schneider, T., Wheeler-Kingshott, C. A. & Alexander, D. C. NODDI: practical
in vivo neurite orientation dispersion and density imaging of the human brain.
Neuroimage 61, 1000–1016 (2012).
11.
Paul, C. A. et al. Association of alcohol consumption with brain volume in the
Framingham study. Arch. Neurol. 65, 1363–1367 (2008).
12.
Mukamal, K. J., Longstreth, W. T., Jr, Mittleman, M. A., Crum, R. M. & Siscovick, D. S.
Alcohol consumption and subclinical findings on magnetic resonance imaging of the
brain in older adults. Stroke 32, 1939–1946 (2001).
13.
den Heijer, T. et al. Alcohol intake in relation to brain magnetic resonance imaging
findings in older persons without dementia. Am. J. Clin. Nutr. 80, 992–997 (2004).
14.
Gu, Y. et al. Alcohol intake and brain structure in a multiethnic elderly cohort. Clin. Nutr.
33, 662–667 (2014).
15.
Sachdev, P. S., Chen, X., Wen, W. & Anstey, K. J. Light to moderate alcohol use is
associated with increased cortical gray matter in middle-aged men: a voxel-based
morphometric study. Psychiatry Res. 163, 61–69 (2008).
16.
Anstey, K. J., Windsor, T. D., Rodgers, B., Jorm, A. F. & Christensen, H. Lower cognitive
test scores observed in alcohol abstainers are associated with demographic, personality,
and biological factors: the PATH Through Life Project. Addiction 100, 1291–1301 (2005).
17.
Topiwala, A. et al. Moderate alcohol consumption as risk factor for adverse brain
outcomes and cognitive decline: longitudinal cohort study. BMJ 357, j2353 (2017).
18.
Button, K. S. et al. Power failure: why small sample size undermines the reliability of
neuroscience. Nat. Rev. Neurosci. 14, 365–376 (2013).
19.
Gelman, A. & Carlin, J. Beyond Power Calculations: Assessing Type S (Sign) and Type
M (Magnitude) Errors. Perspect. Psychol. Sci. 9, 641–651 (2014).
20.
Fein, G. et al. Cortical and subcortical volumes in adolescents with alcohol dependence
but without substance or psychiatric comorbidities. Psychiatry Res. 214, 1–8 (2013).
20
21.
Bobb, J. F., Schwartz, B. S., Davatzikos, C. & Caffo, B. Cross-sectional and longitudinal
association of body mass index and brain volume. Hum. Brain Mapp. 35, 75–88 (2014).
22.
Thayer, R. E. et al. Negative and interactive effects of sex, aging, and alcohol abuse on
gray matter morphometry. Hum. Brain Mapp. 37, 2276–2292 (2016).
23.
Cardon, L. R. & Palmer, L. J. Population stratification and spurious allelic association.
Lancet 361, 598–604 (2003).
24.
Jennings, R. G. & Van Horn, J. D. Publication Bias in Neuroimaging Research:
Implications for Meta-Analyses. Neuroinformatics vol. 10 67–80 (2012).
25.
Miller, K. L. et al. Multimodal population brain imaging in the UK Biobank prospective
epidemiological study. Nat. Neurosci. 19, 1523–1536 (2016).
26.
Sudlow, C. et al. UK biobank: an open access resource for identifying the causes of a
wide range of complex diseases of middle and old age. PLoS Med. 12, e1001779
(2015).
27.
Nave, G., Jung, W. H., Karlsson Linnér, R., Kable, J. W. & Koellinger, P. D. Are Bigger
Brains Smarter? Evidence From a Large-Scale Preregistered Study. Psychol. Sci.
956797618808470 (2018).
28.
Office for National Statistics. Adult drinking habits in Great Britain, 2017. (2018).
29.
Alfaro-Almagro, F. et al. Image processing and Quality Control for the first 10,000 brain
imaging datasets from UK Biobank. Neuroimage 166, 400–424 (2018).
30.
Townsend, P., Phillimore, P. & Beattie, A. Health and deprivation: inequality and the
North. (Routledge, 1988).
31.
Holm, S. A Simple Sequentially Rejective Multiple Test Procedure. Scand. Stat. Theory
Appl. 6, 65–70 (1979).
32.
Alcohol Policy Team, Department of Health. How to keep health risks from drinking
alcohol to a low level Government response to the public consultation. (2016).
21
33.
Alcohol units | Alcohol Change UK. Alcohol Change UK
https://alcoholchange.org.uk/alcohol-facts/interactive-tools/check-your-drinking/alcohol-
units (2018).
34.
Qin, L. & Crews, F. T. NADPH oxidase and reactive oxygen species contribute to
alcohol-induced microglial activation and neurodegeneration. Journal of
Neuroinflammation vol. 9 (2012).
35.
Zahr, N. M., Mayer, D., Rohlfing, T., Sullivan, E. V. & Pfefferbaum, A. Imaging
neuroinflammation? A perspective from MR spectroscopy. Brain Pathol. 24, 654–664
(2014).
36.
Romero, A. M. et al. Chronic alcohol alters dendritic spine development in neurons in
primary culture. Neurotox. Res. 24, 532–548 (2013).
37.
Zhou, F. C. et al. Chronic alcohol drinking alters neuronal dendritic spines in the brain
reward center nucleus accumbens. Brain Res. 1134, 148–161 (2007).
38.
Boniface, S., Kneale, J. & Shelton, N. Drinking pattern is more strongly associated with
under-reporting of alcohol consumption than socio-demographic factors: evidence from a
mixed-methods study. BMC Public Health 14, 1297 (2014).
39.
Greenfield, T. K. & Kerr, W. C. Alcohol measurement methodology in epidemiology:
recent advances and opportunities. Addiction 103, 1082–1099 (2008).
40.
Cox, S. R. et al. Ageing and brain white matter structure in 3,513 UK Biobank
participants. Nat. Commun. 7, 13629 (2016).
41.
Price, A. L. et al. Principal components analysis corrects for stratification in genome-
wide association studies. Nat. Genet. 38, 904–909 (2006).
42.
Haworth, S. et al. Apparent latent structure within the UK Biobank sample has
implications for epidemiological analysis. Nat. Commun. 10, 333 (2019).
43.
Rietveld, C. A. et al. GWAS of 126,559 individuals identifies genetic variants associated
with educational attainment. Science 340, 1467–1471 (2013).
22
44.
Kranzler HR, Covault J, Feinn R, et al. Topiramate treatment for heavy drinkers:
moderation by a GRIK1 polymorphism. Am J Psychiatry 171, 445-52 (2015).
23
Acknowledgments
This research was carried out under the auspices of the Brain Imaging and Genetics in
Behavioral Research Consortium (https://big-bear-research.org/), using UK Biobank resources
under application 40830. The study was supported by funding from an ERC Consolidator Grant
to Philipp Koellinger (647648 EdGe), the National Institute on Alcohol Abuse and Alcoholism to
Reagan Wetherill (K23 AA023894), the Wharton Dean’s Research Fund, the Wharton
Neuroscience Initiative, and the VISN 4 Mental Illness Research, Education and Clinical Center
at the Crescenz VA Medical Center.
Author contributions
RD, PK, HRK, GN, and RW conceived of and designed the study. RD analyzed data. RD, GA,
KJ, PK, HRK, GN, and RRW interpreted data. RD and RRW wrote the paper. GA, NS, PK,
HRK, and GN critically edited the work. RRW edited the work. All authors approved the final
version to be submitted for publication and agree to be accountable for all aspects of this work.
Competing interests
HRK is a member of the American Society of Clinical Psychopharmacology’s Alcohol Clinical
Trials Initiative, which was supported in the last three years by AbbVie, Alkermes, Ethypharm,
Indivior, Lilly, Lundbeck, Otsuka, Pfizer, Arbor, and Amygdala Neurosciences and is named as
an inventor on PCT patent application #15/878,640 entitled: "Genotype-guided dosing of opioid
agonists," filed January 24, 2018. All other authors declare no competing interests.
Supplementary Materials for
Multimodal brain imaging study of 19,825 participants reveals adverse effects of
moderate drinking
Supplementary Table 1. Associations between alcohol intake and regional gray matter volume imaging-derived phenotypes
L/R
Model A
Model B
Model C
Model D
β
t
p
β
t
p
β
t
p
β
t
p
Frontal pole
L
-0.020
-4.552
5.34E-05
-0.020
-4.464
8.09E-06
-0.021
-3.941
8.15E-05
-0.020
-4.017
5.91E-05
Precentral gyrus
L
-0.027
-4.636
3.76E-06
-0.027
-4.706
2.55E-06
-0.033
-4.779
1.77E-06
∙∙
∙∙
∙∙
Precentral gyrus
R
-0.031
-5.355
8.66E-08
-0.032
-5.508
3.67E-08
-0.042
-6.162
7.34E-10
-0.026
-3.975
7.06E-05
Temporal pole
R
-0.024
-4.081
4.50E-05
-0.024
-4.015
5.97E-05
-0.030
-4.310
1.64E-05
∙∙
∙∙
∙∙
Inferior temporal
gyrus
R
0.025
3.893
9.93E-05
0.026
4.040
5.35E-05
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
Superior temporal
gyrus
L
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
-0.032
-4.310
1.64E-05
∙∙
∙∙
∙∙
Postcentral gyrus
L
∙∙
∙∙
∙∙
-0.023
-3.755
1.74E-04
-0.029
-3.995
6.50E-05
∙∙
∙∙
∙∙
Postcentral gyrus
R
-0.024
-3.933
8.41E-05
-0.025
-3.968
7.27E-05
-0.035
-4.802
1.58E-06
∙∙
∙∙
∙∙
Lateral occipital cortex
L
-0.021
-3.685
2.29E-04
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
Lateral occipital cortex
R
-0.247
-4.282
1.86E-05
-0.024
-4.106
4.04E-05
∙∙
∙∙
∙∙
-0.022
-3.394
6.90E-04
Cuneal cortex
L
-0.029
-4.369
1.25E-05
-0.027
-4.167
3.10E-05
-0.032
-4.110
3.97E-05
-0.028
-3.938
8.25E-05
Frontal orbital cortex
R
-0.021
-3.740
1.84E-04
-0.021
-3.702
2.15E-04
-0.027
-4.108
4.02E-05
∙∙
∙∙
∙∙
Lingual gyrus
L
0.025
4.446
8.81E-06
0.023
4.242
2.23E-05
∙∙
∙∙
∙∙
0.022
3.581
3.43E-04
Central opercular
cortex
R
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
-0.026
-3.905
9.45E-05
∙∙
∙∙
∙∙
Planum polare
L
-0.024
-4.048
5.18E-05
-0.026
-4.259
2.06E-05
-0.028
-3.895
9.87E-05
∙∙
∙∙
∙∙
Planum polare
R
-0.027
-4.701
2.61E-06
-0.028
-4.929
8.34E-07
-0.032
-4.703
2.58E-06
-0.023
-3.633
2.80E-04
Heschl’s gyrus
R
-0.024
-4.048
5.18E-05
-0.026
-4.387
1.16E-05
-0.026
-3.715
2.03E-04
∙∙
∙∙
∙∙
Putamen
L
-0.048
-6.711
1.99E-11
-0.051
-7.087
1.42E-12
-0.050
-5.946
2.79E-09
-0.041
-5.246
1.57E-07
Putamen
R
-0.043
-6.079
1.23E-09
-0.047
-6.665
2.72E-11
-0.047
-5.646
1.67E-08
-0.031
-3.909
9.29E-05
Pallidum
L
0.030
4.105
4.06E-05
0.029
3.938
8.23E-05
∙∙
∙∙
∙∙
0.029
3.930
8.52E-05
Pallidum
R
0.034
4.726
2.31E-06
0.033
4.473
7.76E-06
0.036
4.144
3.43E-05
0.036
4.741
2.14E-06
Amygdala
R
-0.024
-4.322
1.55E-05
-0.027
-4.878
1.08E-06
-0.032
-4.772
1.84E-06
∙∙
∙∙
∙∙
Brain stem
-0.033
-5.368
8.06E-08
-0.033
-5.299
1.18E-07
-0.034
-4.619
3.88E-06
-0.032
-4.949
7.53E-07
V Cerebellum
L
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.026
3.630
2.84E-04
Note. Reported β values are standardized. GMV: gray matter volume, IDP: imaging-derived phenotype, L/R: left/right, t: t-value; p: p-value. Model
A (n=19,825) estimated a linear regression with the IDPs as dependent variables and alcohol intake as the main independent variable of interest,
controlling for sex, age, age-squared, age-cubed, height, brain volume, the Townsend index of social deprivation measured at the zip code level,
handedness (right/ left/ ambidextrous; dummy-coded), the first 40 genetic principal components,1 and dummy-coded county of residence (p < 2.43
x 10-4).2 All continuous variables (except age-related variables) were standardized to a mean of 0 and a standard deviation of 1. Model B
(n=19,825) included additional control for variables that are potentially associated with alcohol intake: weight, body mass index, and educational
attainment (p < 2.34 x 10-4). Model C (n=18,298) repeated model A, excluding non-drinkers (p < 2.13 x 10-4). Model D (n=18,599) repeated model
A, excluding heavy drinkers (i.e., women who reported consuming more than 18 units/week and men who reported consuming more than 24
units/week3 (p < 7.25 x 10-4).
Supplementary Table 2. Associations between alcohol intake and white matter water molecular diffusion indices (FA and MD)
Tract
L/R
Model A
Model B
Model C
Model D
β
t
p
β
t
p
β
t
p
β
t
p
Fractional Anisotropy (FA)
CingG
R
-0.033
-4.354
1.35E-05
-0.034
-4.479
7.54E-06
-0.037
-4.067
4.78E-05
∙∙
∙∙
∙∙
PTR
L
-0.035
-4.584
4.58E-06
-0.037
-4.948
7.58E-07
-0.040
-4.509
6.55E-06
∙∙
∙∙
∙∙
PTR
R
-0.032
-4.182
2.91E-05
-0.033
-4.396
1.74E-05
-0.045
-4.899
9.71E-07
∙∙
∙∙
∙∙
Fmin
-0.031
-4.129
3.66E-05
-0.032
-4.272
1.95E-05
-0.034
-3.846
1.20E-04
∙∙
∙∙
∙∙
Mean Diffusivity (MD)
ILF
R
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.040
4.574
4.82E-06
∙∙
∙∙
∙∙
SLF
L
0.028
3.783
1.56E-04
0.028
3.834
1.26E-04
0.039
4.440
9.06E-06
∙∙
∙∙
∙∙
SLF
R
0.029
3.885
1.03E-04
0.029
3.901
9.61E-05
0.043
4.900
9.68E-07
∙∙
∙∙
∙∙
ATR
L
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.032
3.855
1.16E-04
∙∙
∙∙
∙∙
ATR
R
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.032
3.866
1.11E-04
∙∙
∙∙
∙∙
PTR
L
0.035
4.811
1.51E-06
0.038
4.682
2.87E-06
0.045
5.298
1.18E-07
∙∙
∙∙
∙∙
PTR
R
0.034
4.680
2.89E-06
0.033
4.587
4.52E-06
0.050
5.826
5.79E-09
∙∙
∙∙
∙∙
STR
L
0.037
5.102
3.39E-07
0.038
5.224
1.77E-07
0.055
6.453
1.13E-10
∙∙
∙∙
∙∙
STR
R
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.044
5.129
2.95E-07
∙∙
∙∙
∙∙
Fmin
R
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.034
3.788
1.53E-04
∙∙
∙∙
∙∙
MCP
0.030
3.889
1.01E-04
0.032
4.232
2.33E-05
0.034
3.764
1.68E-04
∙∙
∙∙
∙∙
Note. Reported β values are standardized. L/R: left/right, t: t-value; p: p-value. ATR, anterior thalamic radiation; BMI, body mass index; CingG,
cingulum gyrus; EA, educational attainment, FA, fractional anisotropy; MD, mean diffusivity; Fmin, forceps minor; ILF, inferior longitudinal
fasciculus; MCP, middle cerebellar peduncle; PTR, posterior thalamic radiation; SLF, superior longitudinal fasciculus; STR, superior thalamic
radiation. Model A (n=17,975) estimated a linear regression with the IDPs as dependent variables and alcohol intake as the main independent
variable of interest, controlling for sex, age, age-squared, age-cubed, height, brain volume, the Townsend index of social deprivation measured at
the zip code level, handedness (right/left/ambidextrous; dummy-coded), the first 40 genetic principal components,1 and dummy-coded county of
residence2 (p < 2.25 x 10-4). All continuous variables (except age-related variables) were standardized to a mean of 0 and a standard deviation of
1. Model B (n=17,975) included additional control for variables that are potentially associated with alcohol intake: weight, body mass index, and
educational attainment (p < 2.22 x 10-4). Model C (n=16,606) repeated model A, excluding non-drinkers (p < 1.97 x 10-4). Model D (n=16,873)
repeated model A, excluding heavy drinkers (i.e., women who reported consuming more than 18 units/week and men who reported consuming
more than 24 units/week2 (p < 7.25 x 10-4).
Supplementary Table 3. Associations between alcohol intake and neurite orientation dispersion and density imaging characteristics
Tract
L/R
Model A
Model B
Model C
Model D
β
t
p
β
t
p
β
t
p
β
t
p
Intracellular Volume Fraction (ICVF)
CingG
R
-0.030
-3.964
7.39E-05
-0.030
-3.979
6.94E-05
-0.035
-3.892
9.99E-05
∙∙
∙∙
∙∙
IFOF
L
-0.032
-4.223
2.42E-05
-0.032
-4.220
2.46E-05
-0.037
-4.144
3.42E-05
∙∙
∙∙
∙∙
IFOF
R
-0.033
-4.414
1.02E-05
-0.033
-4.348
1.38E-05
-0.038
-4.256
2.10E-05
∙∙
∙∙
∙∙
ILF
L
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
-0.034
-3.724
1.97E-04
∙∙
∙∙
∙∙
Unc
R
-0.029
-3.904
9.48E-05
-0.028
-3.766
1.67E-04
-0.034
-3.874
1.08E-04
∙∙
∙∙
∙∙
ATR
L
∙∙
∙∙
∙∙
-0.027
-3.711
2.07E-04
-0.035
-4.076
4.60E-05
∙∙
∙∙
∙∙
ATR
R
∙∙
∙∙
∙∙
-0.028
-3.888
1.01E-04
-0.036
-4.228
2.37E-05
∙∙
∙∙
∙∙
PTR
L
-0.034
-4.538
5.71E-06
-0.035
-4.659
3.19E-06
-0.042
-4.705
2.56E-06
∙∙
∙∙
∙∙
PTR
R
-0.034
-4.433
9.36E-06
-0.034
-4.433
9.37E-06
-0.045
-5.000
5.79E-07
∙∙
∙∙
∙∙
STR
L
-0.035
-4.632
3.65E-06
-0.035
-4.593
4.40E-06
-0.046
-5.182
2.22E-07
∙∙
∙∙
∙∙
STR
R
-0.031
-4.164
3.14E-05
-0.031
-4.095
4.23E-05
-0.044
-4.963
7.01E-07
∙∙
∙∙
∙∙
Fmin
-0.037
-4.906
9.38E-07
-0.036
-4.823
1.43E-06
-0.041
-4.602
4.22E-06
∙∙
∙∙
∙∙
Isotropic Volume Fraction (ISOVF)
ILF
R
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.039
4.488
7.23E-06
∙∙
∙∙
∙∙
SLF
R
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.035
4.036
5.45E-05
∙∙
∙∙
∙∙
ATR
L
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.035
4.200
2.69E-05
∙∙
∙∙
∙∙
PTR
L
0.037
5.145
2.70E-07
0.036
4.882
1.06E-06
0.047
5.539
3.27E-08
∙∙
∙∙
∙∙
PTR
R
0.036
4.949
7.51E-07
0.035
4.848
1.26E-06
0.051
5.974
2.36E-09
∙∙
∙∙
∙∙
STR
L
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
0.041
4.696
2.68E-06
∙∙
∙∙
∙∙
MCP
0.034
4.426
9.65E-06
0.037
4.915
8.94E-07
0.040
4.445
8.86E-06
∙∙
∙∙
∙∙
Tract Complexity/Fanning (OD)
IFOF
L
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
-0.036
-3.884
1.03E-04
-0.030
-3.522
4.29E-04
IFOF
R
-0.031
-4.047
5.22E-05
-0.032
-4.115
3.89E-05
-0.037
-4.050
5.14E-05
-0.029
-3.497
4.71E-04
SLF
L
-0.035
-4.476
7.65E-06
-0.035
-4.522
6.16E-06
-0.040
-4.395
1.11E-05
∙∙
∙∙
∙∙
SLF
R
-0.030
-3.912
9.16E-05
-0.031
-4.096
4.22E-05
-0.036
-4.030
5.60E-05
∙∙
∙∙
∙∙
STR
L
-0.027
-3.718
2.02E-04
-0.027
-3.787
1.53E-04
-0.033
-3.836
1.25E-04
-0.033
-3.901
9.62E-05
STR
R
-0.032
-4.486
7.32E-06
-0.033
-4.531
5.92E-06
-0.042
-4.836
1.34E-06
∙∙
∙∙
∙∙
AR
R
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
-0.039
-4.330
1.50E-05
∙∙
∙∙
∙∙
CST
L
-0.041
-5.517
3.50E-08
-0.044
-5.861
4.69E-09
-0.049
-5.484
4.22E-08
-0.038
-4.665
3.11E-06
CST
R
-0.049
-6.504
8.01E-11
-0.052
-6.906
5.14E-12
-0.053
-5.983
2.23E-09
∙∙
∙∙
∙∙
Fmaj
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
-0.042
-4.688
2.78E-06
∙∙
∙∙
∙∙
Fmin
-0.032
-4.235
2.30E-05
-0.033
-4.447
8.76E-06
∙∙
∙∙
∙∙
∙∙
∙∙
∙∙
Note. Reported β values are standardized. L/R: left/right, t: t-value; p: p-value. ATR, anterior thalamic radiation; BMI, body mass index; CingG,
cingulum gyrus; EA, educational attainment, FA, fractional anisotropy; MD, mean diffusivity; Fmin, forceps minor; ILF, inferior longitudinal
fasciculus; MCP, middle cerebellar peduncle; PTR, posterior thalamic radiation; SLF, superior longitudinal fasciculus; STR, superior thalamic
radiation. Model A (n=17,975) estimated a linear regression with the IDPs as dependent variables and alcohol intake as the main independent
variable of interest, controlling for sex, age, age-squared, age-cubed, height, brain volume, the Townsend index of social deprivation measured at
the zip code level, handedness (right/left/ambidextrous; dummy-coded), the first 40 genetic principal components,1 and dummy-coded county of
residence2 (p < 2.25 x 10-4). All continuous variables (except age-related variables) were standardized to a mean of 0 and a standard deviation of
1. Model B (n=17,975) included additional control for variables that are potentially associated with alcohol intake: weight, body mass index, and
educational attainment (p < 2.22 x 10-4). Model C (n=16,606) repeated model A, excluding non-drinkers (p < 1.97 x 10-4). Model D (n=16,873)
repeated model A, excluding heavy drinkers (i.e., women who reported consuming more than 18 units/week and men who reported consuming
more than 24 units/week3 (p < 7.25 x 10-4).
Supplementary Figure 1. Average regional gray matter volume in subcortical brain regions showing significant associations from
Model A (linear regression) based on daily drinking levels. Daily unit = 10 ml of pure ethanol.
Supplementary Figure 2. Average regional gray matter volume in frontal brain regionals showing significant associations in Model A
(linear regression) across average daily drinking levels. Daily unit = 10 ml of pure ethanol.
Supplementary Figure 3. Average regional gray matter volume in frontal, insular, and parietal brain regions showing significant
associations from Model A (linear regression) across average daily drinking levels. Daily unit = 10 ml of pure ethanol.
Supplementary Figure 4. Average regional gray matter volume in temporal brain regions showing significant associations in Model A
(linear regression) across average daily drinking levels. Daily unit = 10 ml of pure ethanol.
Supplementary Figure 5. Average regional gray matter volume in occipital and cerebellar brain regions showing significant
associations from Model A (linear regression) across average daily drinking levels. Daily unit = 10 ml of pure ethanol.
Supplementary Figure 6. Average white matter microstructure indices showing significant associations from Model A (linear
regression) across average daily drinking levels in association fibers. Daily unit = 10 ml of pure ethanol.
Supplementary Figure 7. Average white matter microstructure indices showing significant associations from Model 1 (linear
regression) across average daily drinking levels in projection fibers. Daily unit = 10 ml of pure ethanol.
References
1
Price, A. L. et al. Principal components analysis corrects for stratification in genome-wide association studies. Nat Genet 38,
904-909, (2006).
2
Haworth, S. et al. Apparent latent structure within the UK Biobank sample has implications for epidemiological analysis. Nat
Commun 10, 333, doi:10.1038/s41467-018-08219-110.1038/s41467-018-08219-1 [pii] (2019).
3
Kranzler, H. R. et al. Topiramate treatment for heavy drinkers: moderation by a GRIK1 polymorphism. Am J Psychiatry 171,
445-452, (2014).
| 2020 | Multimodal brain imaging study of 19,825 participants reveals adverse effects of moderate drinking | 10.1101/2020.03.27.011791 | [
"Daviet Remi",
"Aydogan Gökhan",
"Jagannathan Kanchana",
"Spilka Nathaniel",
"Koellinger Philipp",
"Kranzler Henry R.",
"Nave Gideon",
"Wetherill Reagan R."
] | creative-commons |
1
Title page
1
2
Neuronal mechanism of a BK channelopathy in absence epilepsy and movement disorders
3
Ping Dong1, Yang Zhang1, Mohamad A. Mikati2,3, Jianmin Cui4, Huanghe Yang1,2,*
4
1Department of Biochemistry, Duke University Medical Center, Durham, NC 27710, USA
5
2Department of Neurobiology, Duke University Medical Center, Durham, NC 27710, USA
6
3Department of Pediatrics, Duke University Medical Center, Durham, NC 27710, USA
7
4Department of Biomedical Engineering, Washington University in Saint Louis, Saint Louis, MO 63130,
8
USA.
9
* Corresponding author: Huanghe Yang. Email: huanghe.yang@duke.edu
10
11
Conflict of interest: The authors have declared that no conflict of interest exists.
12
2
Significance
13
Dysfunction of BK channels or BK channelopathy has been increasingly implicated in diverse
14
neurological disorders including epilepsy, movement, cognitive and neurodevelopmental disorders.
15
However, precision medicine to treat BK channelopathy is lacking. Here we characterized a mouse
16
model carrying a gain-of-function BK channelopathy D434G from a large family of patients with
17
absence epilepsy and involuntary movement disorders. The BK-D434G mice resemble the clinical
18
manifestations of absence seizures and exhibit severe motor defects. The hyperexcitability in BK-
19
D434G cortical neurons and cerebellar Purkinje cells underscores the neuronal mechanism of BK gain-
20
of-function induced absence epilepsy and movement disorders. The effectiveness of a BK channel
21
blocker on preventing absence seizures suggests that BK inhibition is a promising strategy to treat gain-
22
of-function BK channelopathy.
23
Abstract
24
A growing number of gain-of-function (GOF) BK channelopathy have been identified in patients with
25
epilepsy and paroxysmal movement disorders. Nevertheless, the underlying pathophysiology and
26
corresponding therapeutics remain obscure. Here we utilized a knock-in mouse model carrying human
27
BK-D434G channelopathy to investigate the neuronal mechanism of BK GOF in the pathogenesis of
28
epilepsy and movement disorders. We found that the BK-D434G mice manifest the clinical features of
29
absence epilepsy and exhibit severe motor deficits. BK-D434G mutation causes hyperexcitability of
30
cortical pyramidal neurons and cerebellar Purkinje cells, which contributes to the pathogenesis of
31
absence seizures and the motor defects, respectively. A BK channel blocker paxilline potently
32
suppresses BK-D434G-induced hyperexcitability and effectively mitigates absence seizures in mice.
33
Our study thus uncovered a neuronal mechanism of BK GOF in absence epilepsy and provided the
34
3
evidence that BK inhibition is a promising therapeutic strategy to mitigate BK GOF-induced
35
neurological disorders.
36
4
Introduction
37
KCNMA1 encodes the pore forming α subunit of the Ca2+- and voltage-activated large-
38
conductance BK type potassium channels that are widely expressed in the brain with high expression
39
levels in the cortex, cerebellar Purkinje cells, thalamus, hippocampus, basal ganglia, habenula, and
40
olfactory bulb (1-4). Owing to its large single channel conductance, its dual sensitivity to both voltage
41
and intracellular Ca2+ and its spatial proximity to voltage-gated Ca2+ channels (VGCCs) (4-9) (Fig. 1A),
42
BK channels play pivotal roles in shaping action potential repolarization, giving rise to fast after-
43
hyperpolarization (fAHP), controlling dendritic Ca2+ spikes and influencing synaptic transmission (1, 2,
44
10-12). Consistent with its importance in the nervous system, dysfunction of BK channels has been
45
implicated in the pathophysiology of various neurological disorders including epilepsy (12-16),
46
movement disorders (13, 15, 17-22), and neurodevelopmental and cognitive disorders such as
47
intellectual delay (15, 16, 18, 21, 23), autism spectrum disorder (14, 17, 21, 24), Fragile X syndrome
48
(25) and Angelman syndrome (26). How BK channels involve in such a diverse spectrum of
49
neurological disorders (27-29), however, remains largely elusive and demands in-depth studies.
50
KCNMA1 variants identified from human genetic analysis provide unique opportunities to
51
understand the neurological functions of BK channels (27, 29). The first KCNMA1-linked potassium
52
channelopathy D434G from a large family of patients with generalized epilepsy and/or paroxysmal
53
dyskinesia (13). Of the sixteen BK-D434G patients, nine individuals had absence epilepsy, which is
54
characterized by sudden, brief lapses of consciousness accompanied by behavioral arrest and distinctive
55
bilaterally synchronous spike-and-wave discharges (SWDs) at 2.5-4 Hz (30); twelve BK-D434G
56
patients developed paroxysmal nonkinesigenic dyskinesia (PNKD), which is an episodic movement
57
disorder characterized by attacks of hyperkinesia with intact consciousness (13). Interestingly, five BK-
58
D434G patients were affected by both absence epilepsy and PNKD. Subsequent biophysical
59
5
characterizations demonstrated that BK-D434G is a gain-of-function (GOF) mutation with enhanced
60
Ca2+ sensitivity (13, 31-33). It is intriguing why a GOF potassium channel mutation is associated with
61
epilepsy and dyskinesia, which is characterized by hyperexcitability and hypersynchronization in nature.
62
A growing number of human KCNMA1 variants have been identified over the past several years (27,
63
29). However, it is unknown 1) whether the KCNMA1 variants cause the associated neurological
64
disorders; 2) how the KCNMA1 variants affect neuronal activities at cellular level; 3) whether targeting
65
the mutant BK channels is effective to mitigate the associated neurological symptoms.
66
To address these questions, we characterized a knock-in mouse model carrying the BK-D434G
67
mutation. We found that the BK-D434G mice align with the clinical manifestations of absence seizures
68
and response to anti-absence medications. In vitro brain slice recordings revealed that the
69
hyperexcitability of the cortical pyramidal neurons contribute to BK-D434G-induced spontaneous
70
absence seizures. The effectiveness of paxilline (PAX), a BK channel specific blocker, on suppressing
71
BK-D434G-induced absence seizures in mice establishes the causal relationship of this GOF and
72
absence seizures. BK-D434G also induces hyperactivity in Purkinje cells and leads to functional and
73
morphological changes, all of which contribute to the observed motor defects. Our study not only
74
elucidate the cellular basis of the BK-D434G channelopathy in epilepsy and movement disorders, but
75
also demonstrate that BK inhibition can be a promising therapeutic strategy to mitigate BK GOF-
76
induced epilepsy.
77
Results
78
BK-D434G knock-in mice resemble clinical manifestations of absence epilepsy. A knock-in
79
mouse line carrying BK-D434G mutation was generated by homologous recombination (Fig. S1A and
80
Methods). The resulting animals were confirmed by both genotyping PCR (Fig. S1B) and genomic
81
sequencing (Fig. S1C). The heterozygous BK-D434G mutation (BKDG/WT) mice were viable and
82
6
survived into adulthood. However, only 17.3% of the offspring were homozygous BKDG/DG under the
83
BKDG/WT ×
84
BKDG/WT breeding scheme, which was significantly less than the expected 25% Mendelian inheritance
85
(P < 0.05, Chi-square = 6.495, Fig. S1D). This indicates that the BK-D434G mutation homozygosity
86
had some detrimental effects to the BKDG/DG mice.
87
Nine out of sixteen of the BK-D434G channelopathy patients had generalized epilepsy (13).
88
These patients typically had absence seizures with spike-wave-discharges (SWDs). We therefore used
89
simultaneous video-electroencephalogram (EEG) recording (Fig. 1B and Methods) to examine if the
90
knockin mice resemble the human BK-D434G patients’ clinical manifestations. We found that both the
91
heterozygous BKDG/WT and homozygous BKDG/DG mice, but not the BKWT/WT control mice exhibited
92
frequent episodes of spontaneous, generalized SWDs, each of which lasted for 0.5-10 seconds (Fig. 1C,
93
Table 1 and Movie S1). Power spectral analysis of the SWDs showed that the epileptic events of the
94
BK-D434G mice were composed of strong frequency bands of 3-8 Hz (Fig. 1D), which is comparable to
95
the typical SWD frequency range in other rodent models with absence seizures (Table 1) (34, 35).
96
Compared with the BKDG/WT mice, which had 54.1 ± 13.4 SWDs per hour, the homozygous BKDG/DG
97
mice showed dramatically increased incidences of SWDs (263.7 ± 34.8 SWDs/hour) (Fig. 1E). This
98
suggests that BK-D434G homozygosity can lead to more severe phenotypes, which may contribute to
99
the increased lethality of the BKDG/DG mice (Fig. S1D).
100
Combining the EEG recording with an unbiased, automatic video analysis on the total movement
101
on the freely moving mice (Fig. S2 and Methods), we demonstrated that the BK-D434G mice exhibited
102
frequent behavioral arrest during the SWDs onset (Movie S1 and Table 1), another hallmark of absence
103
epilepsy (36). Different from the BKWT/WT mice that constantly underwent alternating locomotive and
104
non-locomotive status (Fig. 1Ci), the BKDG/WT and BKDG/DG mice showed much higher incidences and
105
7
longer durations of non-locomotion (Fig. 1Cii and 1Ciii). After aligning the mouse locomotive activities
106
with the SWDs, we found that when SWDs developed, the mice were behaviorally arrested; whereas
107
when the mice were spared from SWDs, they were able to freely move around.
108
As the BK-D434G proband responded to valproate (13), we tested the effects of the first-line
109
anti-absence medicines valproate and ethosuximide (ESM) on our BKDG/DG mice. Administration of
110
valproate or ESM effectively suppressed the frequent SWDs in the animals for about an hour (Fig. 2).
111
Typical SWDs accompanied by behavioral arrest and responsiveness to the first-line anti-absence
112
seizure medicines (Table 1) explicitly demonstrated that the BK-D434G mice fully align with the
113
clinical manifestations of absence epilepsy from the human patients carrying the BK-D434G mutation
114
(13).
115
BK-D434G knock-in mice are susceptible to convulsant-induced tonic-clonic seizures. In addition to
116
absence seizures, two BK-D434G channelopathy patients were also reported to develop generalized
117
tonic-clonic seizures (13). We hypothesized that BK-D434G GOF mutation may increase the
118
susceptibility to develop tonic-clonic seizures. To test this, we administered pentylenetetrazole (PTZ), a
119
convulsant (37), to the BKWT/WT and BKDG/WT mice. We found that the low dosage of PTZ injection (40
120
mg/kg) induced generalized seizure (GS) stage (seizure score ≥ 4, see Methods for details) in all
121
BKDG/WT mice, whereas the same dosage of PTZ injection induced GS stage in only 42.9% of the
122
BKWT/WT mice (Fig. S3A). Furthermore, the BKDG/WT mice exhibited significantly increased seizure
123
scores (Fig. S3A), markedly prolonged GS duration (Fig. S3B) and dramatically reduced latency to GS
124
(Fig. S3C). Our characterizations indicated that the BK-D434G mice not only have spontaneous
125
seizures, but also are more vulnerable to PTZ-induced generalized tonic-clonic seizures.
126
Cortical pyramidal neurons of BKDG/WT mice show hyperexcitability. Cortical neurons play essential
127
roles in the pathogenesis of absence seizures (30); and BK channels are highly expressed in cortical
128
8
pyramidal neurons (2, 3). Therefore, we investigated whether BK-D434G GOF mutation alters the
129
membrane excitability of cortical pyramidal neurons. Our acute brain slice recording showed that the
130
cortical pyramidal neurons from the BKDG/WT mice exhibited hyperexcitability as evidenced by the
131
significantly increased action potential frequency compared with the BKWT/WT mice (Fig. 3, A-B and
132
Table S1). Single action potential analysis of the first spikes revealed that the BKDG/WT cortical neurons
133
exhibited much faster repolarization as evidenced by the significantly shortened action potential duration
134
(AP90) and augmented after-hyperpolarization amplitude (AHP) (Fig. 3, C-E). As K+ efflux through BK
135
channels contributes to fast after-hyperpolarization (fAHP) (11, 12), our observations of steeper
136
repolarization and enhanced AHP in the BKDG/WT neurons suggest that the BK-D434G GOF mutant
137
channels, which have a higher Ca2+ sensitivity (13, 31), can more efficiently hyperpolarize the
138
membrane following membrane depolarization and VGCC opening. The faster and stronger
139
hyperpolarization induced by BK-D434G would enable faster recovery of the voltage-gated sodium
140
channels (NaV) (Fig. 3F) and potentially facilitate the activation of the hyperpolarization-activated
141
cation (HCN) channels (11, 38). The rapid repriming of the NaV channels and enhanced HCN channel
142
activation collectively enables the cortical neurons to fire at a higher frequency. Taken together, our
143
electrophysiological characterization of the cortical pyramidal neurons from the BK-D434G mice
144
demonstrated a neuronal mechanism of BK GOF-induced hyperexcitability.
145
Pharmacological inhibition of BK channels suppresses BK-D434G-induced seizures. We next tested
146
whether pharmacological inhibition of BK channels can restore normal firing and suppress absence
147
seizures in the BK-D434G mice. PAX, a BK channels specific blocker, effectively suppressed the
148
hyperexcitability of cortical pyramidal neuron (Fig. 4, A and B and Table S1), markedly slowed down
149
membrane repolarization, prolonged AP90 and reduced AHP amplitude (Fig. 4, C-E). Consistent with
150
our brain slice recording, administration of PAX (0.35 mg/kg i.p.) eliminated the spontaneous SWDs of
151
9
the BKDG/DG mice and prevented their behavioral arrest for about 30 min (Fig. 4, F-H). Moreover, we
152
found the PAX also decreased the severity of the PTZ-induced seizures in the BK-D434G heterogeneous
153
mice (Fig. S4). Compared with the saline control, PAX administration significantly decreased seizure
154
scores (Fig. S4A), markedly reduced GS duration (Fig. S4B) and dramatically prolonged the latency to
155
GS (Fig. S4C). All these are consistent with the anti-convulsant effect of PAX on the rodent models of
156
epilepsy, including the PTZ-injected rodent models and an Angelman syndrome mouse model with
157
enhanced BK channel activity (39-41). Our in vitro and in vivo experiments thus explicitly showed that
158
pharmacological inhibition of BK channels can suppress absence seizures in the BK-D434G mice and
159
reduce their vulnerability to convulsant-induced seizure. Our findings not only further supported the
160
causal effect of BK-D434G GOF in epilepsy, but also demonstrated pharmacological inhibition as a
161
promising therapeutic strategy to mitigate BK GOF induced epilepsy.
162
BK-D434G mice exhibit severe locomotive defects. In addition to absence epilepsy, majority of the
163
patients (twelve out of sixteen) with BK-D434G mutation also had paroxysmal movement dyskinesia
164
(13). We thus performed a battery of locomotor tests to assess the potential motor defects of the knock-
165
in mice. We first used the open field test to evaluate their general locomotor activities (Fig. 5A). During
166
a 15-minute test, the total travel distance of the BKDG/DG mice were dramatically less than that of the
167
BKWT/WT and BKDG/WT mice (Fig. 5 A and B). Our balance beam test (Fig. 5C and Movie S2)
168
demonstrated that the BKDG/DG
mice took significantly longer time to traverse the balance beam (Fig.
169
5D) and had significant more incidences of hind-limb slips compared with the WT controls (Fig. 5E).
170
The BKDG/WT mice showed no defect on transverse time yet had a milder defect on the number of hind-
171
limb slips. Interestingly, the BKDG/WT mice occasionally but the BKDG/DG mice always used their tails to
172
maintain their balance on the beam (Fig. 5C and Movie S2). We next performed accelerated rotarod test,
173
which is a standard assay to evaluate impairment in rodent motor performance. Both the BKDG/WT and
174
10
the BKDG/DG
mice performed poorly on this more challenging motor task with significantly shorter
175
latency to fall (Fig. 5F and Movie S3). Compared with the BKDG/WT mice, the BKDG/DG mice showed
176
worst performance on rotarod. The severe defects observed during the accelerated rotarod test clearly
177
showed that the BK-D434G mice indeed have impaired motor functions. Several factors such as muscle
178
strength, motor learning and motor coordination may affect rotarod performance (42). To specifically
179
evaluate the motor coordination functions of the BK-D434G mice, we performed gait analysis utilizing
180
footprints (Fig. 5G). Our result revealed that the BKDG/DG mice had significantly shorter hind-limb stride
181
lengths than the BKWT/WT mice (Fig. 5H), suggesting that these mutant animals had severe defects on
182
motor coordination. Taken together, our multiple locomotor tests clearly demonstrated that the motor
183
functions of the BK-D434G mice were severely impaired.
184
Hyperexcitability of BK-D434G cerebellar Purkinje cells contributes to motor defects. BK
185
channels are highly expressed in Purkinje cells (PCs) and play a critical role in controlling PC
186
excitability (2). Genetic ablation of BK channels in murine PCs leads to cerebellar ataxia and impaired
187
motor coordination (43, 44) and some KCNMA1 channelopathy patients showed signs of cerebellar
188
atrophy (15, 17-20). Given all these facts, we set out to examine whether BK-D434G GOF in PC
189
contributes to the observed impairments in motor functions. By immunostaining with the PC marker
190
calbindin, we found that the adult BK-D434G mutant mice showed dramatic changes of their PC
191
morphology (Fig. 6A). The size of PC soma and the width of PC primary dendrites were significantly
192
enlarged in both the BKDG/WT and the BKDG/DG mice compared with the BKWT/WT mice (Fig. 6B and C),
193
indicating signs of PC hypertrophy in the BK-D434G mice.
194
Next, we conducted brain slice patch clamp recording on the PCs from the BKWT/WT and the
195
BKDG/WT mice (Fig. 6D-H). Similar to what we observed in cortical pyramidal neurons (Fig. 3), we
196
found that the BKDG/WT PCs had dramatically enhanced firing rate compared with the PCs of the
197
11
BKWT/WT mice (Fig. 6D and E). The subsequent single action potential waveform analysis showed that
198
the BKDG/WT PCs showed faster membrane repolarization (Fig. 6F, right) with significant reduction of
199
action potential duration (Fig. 6G) and increase of AHP amplitude (Fig. 6H). Consistent with our
200
observations in the cortical pyramidal neurons (Fig. 4), application of 10 µM PAX robustly reversed the
201
changes of single action potential waveform caused by the BK-D434G mutation and efficiently
202
suppressed the hyperactive PCs in the BKDG/WT mice (Fig. 6D-H and Table S1). Collectively, our
203
electrophysiological characterizations demonstrated that BK-D434G GOF can also induce
204
hyperexcitability in PCs by accelerating after-hyperpolarization and facilitating NaV channel
205
deinactivation. Sustained hyperexcitability in BK-D434G PCs may ultimately induce stress to the PCs,
206
leading to morphological changes and contributing to the observed locomotor defects in the BK-D434G
207
mutant mice.
208
Discussion
209
In this study, we show that the BK-D434G knock-in mice resembles the clinical manifestations
210
of generalized absence epilepsy observed in the BK-D434G patients (13). The BK-D434G mice
211
exhibited spontaneous SWDs, which can be suppressed by first line anti-absence medicines and BK
212
channel specific blocker PAX. These findings thus strongly support that BK-D434G GOF causes
213
absence seizures in the BK-D434G channelopathy patients.
214
Utilizing the BK-D434G knock-in mice, we uncovered the cellular pathophysiology of the GOF
215
BK mutation in inducing epilepsy and movement disorders. We found that BK-D434G causes
216
hyperexcitability in both cortical pyramidal neurons and cerebellar Purkinje cells, in which BK channels
217
are highly expressed (2, 3). BK channels are usually form protein complexes with VGCCs in the central
218
nervous system (9). With enhanced Ca2+ sensitivity, BK-D434G GOF mutation will be rapidly activated
219
following membrane depolarization and Ca2+ entry from the VGCCs. The enhanced BK channel activity
220
12
will accelerate fAHP as evidenced by significant shortening of ADP90 (Figs. 3C, 3D, 6F, 6G) and
221
enhanced amplitude of AHP (Figs. 3C, 3E, 6F, 6H). The accelerated fAHP in the BK-D434G neurons
222
can facilitate the recovery of NaV channels from inactivation and promote activation of HCN channels,
223
thereby increasing membrane excitability (Fig. 3F), which collectively leads to enhanced firing and
224
hyperexcitability (Figs. 3A-B and 6D-E).
225
Abnormal oscillatory rhythms within the cortico-thalamic system are generally believed to be
226
responsible for absence seizure ictogenesis (30, 36). The hyperexcitability of BK-D434G cortical
227
pyramidal neurons observed in this study supports the importance of cortical excitability in absence
228
seizure pathogenesis. Future studies need to be done to comprehensively characterize the excitabilities
229
of the different types of neurons in the cortico-thalamic system and illustrate the circuit basis of absence
230
seizure ictogenesis in the BK-D434G mice. It is interesting to observe that GABAergic Purkinje cells
231
from the BK-D434G mice are also hyperactive (Fig. 6D-H). One hypothesis to explain BK GOF-
232
induced hyperexcitability is that the GOF mutations would decrease the excitability of inhibitory
233
neurons, thereby leading to disinhibition of neuronal networks and subsequently hyperexcitability (13).
234
Our observation that the hyperexcitability of the BK-D434G Purkinje cells suggests that the inhibitory
235
neurons with high expression of BK GOF mutations, instead of reducing their excitability, would
236
increase their excitability. The enhanced GABA release will augment inhibitory inputs and switch the
237
downstream neurons in a circuit into a bursting mode, thereby causing hypersynchronization. In the
238
future, it is therefore important to elucidate the contributions BK GOF to membrane excitability in other
239
inhibitory neurons that have different BK channel expression levels.
240
In addition to having spontaneous absence seizures (Fig. 1), the BK-D434G mice are also more
241
susceptible to PTZ-induced tonic-clonic seizures (Fig. S3). It is likely that BK-D434G GOF may also
242
enhance the excitability of the other neurons outside of the cortico-thalamic system such as hippocampal
243
13
pyramidal neurons and dentate gyrus granule cells (12). Future investigations of the excitabilities of
244
these neurons from the BK-D434G mice will shine light on understanding the neuronal and circuit basis
245
of developing tonic-clonic seizures in some of the refectory and/or pharmaco-resistant absence seizure
246
patients (45).
247
Despite clinical applications of the first-line anti-absence medicines including ethosuximide
248
(ESM) and valproate since 1950s (46-48), 30% of absence epilepsy patients are pharmaco-resistant and
249
60% of them are affected by severe neuropsychiatric comorbidities, including attentional, mood,
250
cognitive and memory impairments (30, 49). While human genetics and animal models have shown that
251
VGCCs and GABAA receptor chloride channels contribute to the etiology of absence epilepsy (36, 50),
252
the contributions of potassium channels to absence epilepsy pathogenesis are still elusive. In this study,
253
we showed that PAX, a BK channel blocker, can effectively suppress BK-D434G induced
254
hyperexcitability and absence seizures (Fig. 4), as well as PTZ-induced tonic-clonic seizures (Fig. S4).
255
This is consistent with the previous findings that PAX can alleviate convulsant drug-induced
256
generalized epilepsy (39, 40) and spontaneous seizures in an Angelman syndrome mouse model with
257
enhanced BK channel activity (41). Our current study thus demonstrated that targeting BK channels
258
could be a novel strategy to mitigate absence epilepsy. Future investigations are needed to examine if
259
pharmacological inhibition of BK channels can be a general strategy to treat different BK GOF
260
channelopathy and could be used to treat pharmaco-resistant absence epilepsy. Of course, better BK
261
inhibitors also need to be developed because PAX’s anti-absence effect vanished in 30 minutes after
262
injection due to its poor pharmacokinetics (Fig. 4H) (51).
263
The BK-D434G mice showed severe locomotor defects as examined using open field, balance
264
beam, rotarod and gait analysis (Fig. 5), albeit no obvious sign of PNKD observed. This is different
265
from the clinical observation in which twelve out of sixteen BK-D434G patients had PNKD (13). This
266
14
discrepancy is likely due to the organism difference between mice and humans. It is also possible that
267
the frequent absence seizures in the BK-D434G mice complicate the detection of PNKD, which is not
268
trivial to monitor in mouse models (52). Nevertheless, the hyperexcitability and morphological changes
269
of the BK-D434G Purkinje cells explicitly demonstrated the involvement of the cerebellum in the
270
pathogenesis of PNKD. As BK-D434G induced hyperexcitability leads to Purkinje cell morphological
271
changes (Fig. 6A-C), we are not clear if acute administration of PAX or the first line anti-absence
272
medicine could mitigate the motor defects. Long-term drug treatment starting in early developmental
273
stages is needed to examine if BK inhibition can also be used to treat movement defects. Future
274
investigations using neuronal type-specific knockin of D434G in mice are needed to further dissect the
275
pathophysiological mechanism of BK-D434G in PNKD and develop corresponding therapies.
276
Taken together, the BK-D434G knock-in mice advanced our mechanistic understanding of the
277
pathophysiology of BK GOF in epilepsy and movement disorders. The mechanistic insights gained in
278
this study and our attempts to use PAX to treat absence seizures will shine light on developing novel
279
therapies to mitigate absence epilepsy and movement disorders, as well as designing precision medicine
280
to treat BK GOF channelopathy.
281
Methods
282
Origin of the mouse lines used. BK-D434G mutation mice were generated by homologous
283
recombination in embryonic stem cells and implanted in C57Bl/6J blastocysts using standard
284
procedures. The targeting vector was designed to flank the D434G mutation with a neomycin (Neo)
285
selection cassette with loxP sites after exon 10 of the KCNMA1 gene (Fig. S1A). Chimeric mice were
286
crossed to C57Bl/6J females (Jackson Labs). Germline transmission generated BK+BKD434G (BKDG/WT)
287
mice. Germline transmission was determined by genotyping PCR of mouse tail DNA (Fig. S1B), using
288
primers
pKCNMA1_genotyping
F1
(5’-GTGCCTAGAGGTGGCTGGGAATTAG-3’)
and
289
15
pKCNMA1_genotyping R1 (5’-CCTCTCCTACGGTGGTAAAGTATCC-3’) for the wildtype allele
290
(342 base pairs, bp) and the floxed allele (455 bp). The F1 hybrids were crossed to C57Bl/6J β-actin Cre
291
mice to excise the Neo cassette. The D434G mutations were confirmed by primers
292
pKCNMA1_sequencing
F2
(5’-GCTGAGTGGGGAGATGTATTGCTTC-3’)
and
293
pKCNMA1_sequencing R2 (5’-ACCTAAGGAGCCAGCACCAATCAT-3’). The BK-D434G mice
294
were then backcrossed to C57Bl/6J mice for five generations.
295
For all behavioral experiments, BKDG/WT males were bred with BKDG/WT females. Animals were housed
296
at a constant 24 °C in a 12 h light–dark cycle (lights on at 07:00) with ad libitum food and water. Both
297
males and females were used for in vivo and in vitro analysis. Mouse handling and usage were carried
298
out in a strict compliance with protocols approved by the Institutional Animal Care and Use Committee
299
at Duke University, in accordance with National Institutes of Health guidelines. PCR genotyping was
300
performed using tail DNA extraction.
301
Simultaneous Video-EEG Recording and Analysis. Mice with age of 1 to 6 months were anesthetized
302
with 1~2% isoflurane and mounted on a stereotaxic device (Kopf Instruments). A mouse
303
electroencephalogram (EEG) headstage (#8201, Pinnacle technology Inc., Lawrence, KS, USA) was
304
affixed to the skull with three screws, which served as differential recording leads on the frontal,
305
parietal, and cerebellar cortex. The headstage was subsequent secured to the skull by the dental cement
306
and the animal could recover for 5 days prior to EEG recording. EEG recordings were collected by a
307
preamplifier with 100x gain and high pass filtered at 1.0 Hz (#8200-SE, Pinnacle technology Inc.),
308
accompanied by spontaneous video monitoring on the top of the chamber (Logitech C920 HD Pro
309
Webcam, 24 frames per second). For drug treatment test on the BKDG/DG mice, a single dose of paxilline
310
(0.35 mg/kg, i.p.), ethosuximide (150 mg/kg, i.p.), valproate (200 mg/kg, i.p.) or saline control was
311
injected into the mice after 1 hour of recording as baseline. Data were acquired with an
312
16
analog�to�digital converter (PCI�6221; National Instruments, Austin, TX, USA) to a desktop
313
computer. A custom code written in MATLAB (MathWorks, Natick, MA, USA) was used to visualize
314
the raw EEG recording trace and plot the power spectra using the Fast Fourier Transform (FFT) within
315
the frequency range of 1-20 Hz. The numbers of SWD event were calculated using previously described
316
methods (53).
317
Video based motion analysis. The video-based motion was analyzed using a similar method previously
318
described (54). A custom-written MATLAB code was used to analyze the video recordings of freely
319
moving mice in the EEG recording chamber. The videos were first down-sampled to 1 frame per second,
320
and then converted to gray 8-bit images (Fig. S2A, upper panel). Since the mouse is darker than the
321
background in the gray images, we conducted the image segmentation ������� of the mouse at time t by
322
����� ���, ��� �
1, ��������,��� � �
0, ��������,��� � �
�
323
where ���,��� is the coordinate of the image, and � is the threshold, the value of which was empirically
324
set to be 10% of the darkest intensity (255) of the 8-bit image. A representative result of the
325
segmentation images is shown in Fig. S2A, middle panel.
326
To get the total movement of the mice over time, we obtained a subtracted image ������� by substrating
327
two sequential frames
328
��������, ��� � ������� ���, ��� � ����� ���, ���
where ������� is the next sequential frame of �����. The subtracted images were shown in Fig. S2A,
329
bottom panel. Pixels changed above the empirical threshold (300 pixels) were designated as a motion
330
status.
331
17
PTZ-induced seizure model. We performed intraperitoneal (i.p.) injection of 40 mg/kg
332
pentylenetetrazole (PTZ; Sigma, MO, USA) and then immediately placed animals in a chamber and
333
started video recording. PTZ-induced seizures were scored according to a modified Racine scale (55): 0,
334
normal behavior, no abnormality; 1, immobilization, lying on belly; 2, head nodding, facial, forelimb, or
335
hindlimb myoclonus; 3, continuous whole-body myoclonus, myoclonic jerks, tail held up stiffly; 4,
336
rearing, tonic seizure, falling down on its side; 5, tonic-clonic seizure, falling down on its back, wildly
337
rushing and jumping. 6: death. Score 4 and above are considered as generalized seizures. The latency to
338
develop generalized seizure and the duration of the generalized seizure was measured based on the
339
videos.
340
Electrophysiology. For the recording performed in brain slice, acute slice preparations were as
341
described previously (56). Briefly, BKWT/WT and BKDG/WT mice (postnatal day 15-24) were anesthetized
342
with isoflurane and decapitated. For the recording in different brain regions, the section orientation is
343
different. For the recording in the cortex, 300 µm coronal sections were prepared. For the cerebellar
344
Purkinje cells, 250 µm sagittal slices were prepared. The brain slices were cut in ice-cold NMDG aCSF
345
containing (in mM): 92 NMDG, 2.5 KCl, 1.2 NaH2PO4, 30 NaHCO3, 20 HEPES, 25 glucose, 5 sodium
346
ascorbate, 2 thiourea, 3 sodium pyruvate, 10 MgSO4·7H2O, 0.5 CaCl2·2H2O (Titrated pH to 7.3-7.4
347
using concentrated HCl). The slices were then incubated in HEPES holding solution (in mM): 92 NaCl,
348
2.5 KCl, 1.2 NaH2PO4, 25 NaHCO3, 20 HEPES, 25 glucose, 5 sodium ascorbate, 2 thiourea, 3 sodium
349
pyruvate, 2 MgSO4·7H2O, 2 CaCl2·2H2O) for 60-min at room temperature. After incubation, the slices
350
were transferred to a recording chamber and superfused (3 mL min-1) with artificial cerebrospinal fluid
351
(aCSF) at 33 �. (in mM): 124 NaCl, 2.5 KCl, 1.2 NaH2PO4, 24 NaHCO3, 5 HEPES, 12.5 glucose, 2
352
MgSO4·7H2O, 2 CaCl2·2H2O. All solutions used for electrophysiology were equilibrated with 95%
353
O2/5% CO2. Whole-cell recordings were performed with a MultiClamp 700B amplifier and sampled at
354
18
10 kHz using a Digidata1550A A/D converter. All data acquisition and analyses were performed using
355
the software pClamp 10.7 (Molecular Devices). For action potential recording, pipette resistance was 3-
356
7 MΩ when filled with an intracellular solution containing the following (in mM): 125 K-gluconate, 15
357
KCl, 10 HEPES, 2 Mg-ATP, 0.3 Na-GTP, 10 disodium phosphocreatine, and 0.2 EGTA, adjusted to pH
358
7.25 with KOH. After GΩ-seal and membrane break-through, the membrane resting potential was
359
monitored for 10 min until it is stabilized before recording of action potentials. For pharmacological
360
experiments, 10 µM paxilline was added to extracellular aCSF. AP90% duration was defined by action
361
potential duration of 90% repolarization. The fAHP size was measured as the difference between the
362
spike threshold and voltage minimum after the action potential. First interspike interval was the time
363
between the first and second action potential peaks. Input resistance (Rin) was calculated from voltage
364
deflections induced by rectangular hyperpolarizing current injections (0-100 pA). Membrane time
365
constant (τm) was obtained by fitting a single exponential function to these same hyperpolarizing voltage
366
deflections. Membrane capacitance (Cm) was calculated by dividing τm by Rin. AP amplitude was
367
calculated as the voltage difference between AP threshold and AP peak.
368
Histology. Mice were transcardially perfused with phosphate-buffered saline (PBS) followed by 4%
369
paraformaldehyde. The brain was removed and post-fixed in 4% paraformaldehyde overnight at 4°C and
370
dehydrated in 30% sucrose for 48 h. Sagittal section (50 μm) containing the cerebellum Purkinje cells
371
were collected by using a cryostat (Leica CM1900). The sections were rinsed 3 times with PBS for 10
372
min each and blocked with 5% goat serum and 0.3% Triton X-100 for 2 hours at room temperature and
373
incubated for overnight at 4°C with following primary antibodies: anti-calbindin (1:1000, mouse, Sigma
374
Aldrich, #C9848). After 3 rinses with PBS for 10 min, secondary antibodies (1:1000, conjugated with
375
Goat anti-Mouse Alexa 594, Thermo Fisher Scientific, A-11032) were incubated for 2 hours at room
376
temperature. Then the sections were washed 3 times with PBS for 10 min each and stained with DAPI
377
19
(1:10000 of 5 mg/mL, Sigma-Aldrich). Images were acquired using a Zeiss 780 inverted confocal
378
microscope. Representative images from at least three repeats.
379
Open field test. The mice were placed in a 45 × 45 cm arena composed of four white Plexiglas walls.
380
They could freely move in the arena for 15 min and their locomotion were continuously monitored by
381
video recording. Locomotor activities were evaluated as the distance traveled per 5 min and the total
382
distance by using a custom MATLAB code.
383
Balance beam test. Mice were given five training trials on an 80-cm long, 7-mm small round beam
384
elevated 30 cm above the table, as described previously (57). A video camera was placed 4-inch away
385
from the starting point, so the hindpaws slip could be easily recorded, whereas the opposite end of the
386
beam entered their home-cage with food pellets and bedding materials. The number of foot slips and
387
traversal time were measured as mice traversed the beam in a test trial 24 hours after training.
388
Accelerating Rotarod. The rotarod treadmills (ENV-577M, Med associates, St. Albans, VT, USA) was
389
used to asset the motor coordination of the mice. Before testing, all mice were trained on a fixed-speed
390
protocol at 4 rpm until they could stay on the rod for 30 s. On the same day as the training session, mice
391
were placed on the rotarod for four-10-minute trials with 30 mins rest between trials. In each trial, the
392
rotarod accelerated from 4 to 40 rpm at the rate of 1 rpm every 8 s, then remained at 40 rpm until the end
393
of the trial. The time until the mouse fell from the rod was recorded as the latency to fall. The
394
assessments were performed for four days.
395
Gait Analysis. The forepaws and hindpaws of the mice were painted with non-toxic red and blue inks,
396
respectively. After a two-minute habituation trial, each mouse could walk along a narrow, paper covered
397
runway. The length of each stride was measured.
398
20
Statistics. All the statistical analyses were performed using GraphPad Prism (GraphPad Software).
399
Sample number (n) values are indicated in the results section and Figure legends. All data are presented
400
as the mean ± standard error of the mean (s.e.m.). Sample sizes were chosen based on standards in the
401
field as well as previous experience with phenotype comparisons. No statistical methods were used to
402
predetermine sample size.
403
Author contributions: H.Y. and J.C. perceived the research. H.Y. supervised the project. H.Y. and P.D.
404
designed the experiments with critical help from J.C. and M.A.M. P.D. performed behavioral
405
experiments and brain slice recordings. Y.Z. and P.D. conducted immunofluorescence. P.D. and Y.Z.
406
conducted data analysis. P.D. and H.Y. wrote the manuscript.
407
Acknowledgments: We are grateful to Dr. Xuechu Zhen (Soochow University, China) for providing the
408
BK-D434G mice. We appreciate Drs. Dwight D. Koeberl and Arsen Hunanyan for their technical
409
assistance with locomotor behavioral tests. We also thank Drs. James O. McNamara, William Wetsel,
410
Pengfei Liang, Son Le, Trieu Le, and Zoe Shan for their critical comments on the manuscript. This work
411
was supported by the Duke Institute for Brain Sciences (to H.Y. and M.M.) and the American Epilepsy
412
Society Post-Doctoral Fellowship 693905 (to P.D.).
413
References
414
1.
Contet C, Goulding SP, Kuljis DA, & Barth AL (2016) BK Channels in the Central Nervous
415
System. International review of neurobiology 128:281-342.
416
2.
Sausbier U, et al. (2006) Ca2+ -activated K+ channels of the BK-type in the mouse brain.
417
Histochem Cell Biol 125(6):725-741.
418
3.
Knaus HG, et al. (1996) Distribution of high-conductance Ca(2+)-activated K+ channels in rat
419
brain: targeting to axons and nerve terminals. The Journal of neuroscience : the official journal
420
of the Society for Neuroscience 16(3):955-963.
421
4.
Salkoff L, Butler A, Ferreira G, Santi C, & Wei A (2006) High-conductance potassium channels
422
of the SLO family. Nat Rev Neurosci 7(12):921-931.
423
5.
Berkefeld H, et al. (2006) BKCa-Cav channel complexes mediate rapid and localized Ca2+-
424
activated K+ signaling. Science 314(5799):615-620.
425
21
6.
Turner RW & Zamponi GW (2014) T-type channels buddy up. Pflugers Archiv : European
426
journal of physiology 466(4):661-675.
427
7.
Latorre R, et al. (2017) Molecular Determinants of BK Channel Functional Diversity and
428
Functioning. Physiological reviews 97(1):39-87.
429
8.
Yang H, Zhang G, & Cui J (2015) BK channels: multiple sensors, one activation gate. Front
430
Physiol 6:29.
431
9.
Fakler B & Adelman JP (2008) Control of K(Ca) channels by calcium nano/microdomains.
432
Neuron 59(6):873-881.
433
10.
Yang H & Cui J (2015) “BK channels,” in Handbook of Ion Channels.
434
11.
Gu N, Vervaeke K, & Storm JF (2007) BK potassium channels facilitate high-frequency firing
435
and cause early spike frequency adaptation in rat CA1 hippocampal pyramidal cells. The Journal
436
of physiology 580(Pt.3):859-882.
437
12.
Brenner R, et al. (2005) BK channel beta4 subunit reduces dentate gyrus excitability and protects
438
against temporal lobe seizures. Nat Neurosci 8(12):1752-1759.
439
13.
Du W, et al. (2005) Calcium-sensitive potassium channelopathy in human epilepsy and
440
paroxysmal movement disorder. Nat Genet 37(7):733-738.
441
14.
Li X, et al. (2018) De novo BK channel variant causes epilepsy by affecting voltage gating but
442
not Ca(2+) sensitivity. Eur J Hum Genet 26(2):220-229.
443
15.
Yesil G, et al. (2018) Expanding the Phenotype of Homozygous KCNMA1 Mutations;
444
Dyskinesia, Epilepsy, Intellectual Disability, Cerebellar and Corticospinal Tract Atrophy. Balkan
445
medical journal 35(4):336-339.
446
16.
Tabarki B, AlMajhad N, AlHashem A, Shaheen R, & Alkuraya FS (2016) Homozygous
447
KCNMA1 mutation as a cause of cerebellar atrophy, developmental delay and seizures. Human
448
genetics 135(11):1295-1298.
449
17.
Zhang G, et al. (2020) A Gain-of-Function Mutation in KCNMA1 Causes Dystonia Spells
450
Controlled With Stimulant Therapy. Mov Disord.
451
18.
Liang L, et al. (2019) De novo loss-of-function KCNMA1 variants are associated with a new
452
multiple malformation syndrome and a broad spectrum of developmental and neurological
453
phenotypes. Hum Mol Genet 28(17):2937-2951.
454
19.
Du X, et al. (2020) Loss-of-function BK channel mutation causes impaired mitochondria and
455
progressive cerebellar ataxia. Proc Natl Acad Sci U S A 117(11):6023-6034.
456
20.
Buckley C, et al. (2020) Status Dystonicus, Oculogyric Crisis and Paroxysmal Dyskinesia in a
457
25 Year-Old Woman with a Novel KCNMA1 Variant, K457E. Tremor Other Hyperkinet Mov (N
458
Y) 10:49.
459
21.
Zhang ZB, Tian MQ, Gao K, Jiang YW, & Wu Y (2015) De novo KCNMA1 mutations in
460
children with early-onset paroxysmal dyskinesia and developmental delay. Mov Disord
461
30(9):1290-1292.
462
22.
Heim J, et al. (2020) Cataplexy in Patients Harboring the KCNMA1 p.N999S Mutation.
463
Movement disorders clinical practice 7(7):861-862.
464
22
23.
Bailey CS, Moldenhauer HJ, Park SM, Keros S, & Meredith AL (2019) KCNMA1-linked
465
channelopathy. The Journal of general physiology:jgp.201912457.
466
24.
Wu H, et al. (2020) Phenotype-to-genotype approach reveals head-circumference-associated
467
genes in an autism spectrum disorder cohort. Clin Genet 97(2):338-346.
468
25.
Deng PY, et al. (2013) FMRP regulates neurotransmitter release and synaptic information
469
transmission by modulating action potential duration via BK channels. Neuron 77(4):696-711.
470
26.
Sun AX, et al. (2019) Potassium channel dysfunction in human neuronal models of Angelman
471
syndrome. Science 366(6472):1486-1492.
472
27.
Miller JP, Moldenhauer HJ, Keros S, & Meredith AL (2021) An emerging spectrum of variants
473
and clinical features in KCNMA1-linked channelopathy. Channels.
474
28.
N'Gouemo P (2011) Targeting BK (big potassium) channels in epilepsy. Expert opinion on
475
therapeutic targets 15(11):1283-1295.
476
29.
Bailey CS, Moldenhauer HJ, Park SM, Keros S, & Meredith AL (2019) KCNMA1-linked
477
channelopathy. J Gen Physiol 151(10):1173-1189.
478
30.
Crunelli V, et al. (2020) Clinical and experimental insight into pathophysiology, comorbidity
479
and therapy of absence seizures. Brain 143(8):2341-2368.
480
31.
Yang J, et al. (2010) An epilepsy/dyskinesia-associated mutation enhances BK channel
481
activation by potentiating Ca2+ sensing. Neuron 66(6):871-883.
482
32.
Wang B, Rothberg BS, & Brenner R (2009) Mechanism of increased BK channel activation from
483
a channel mutation that causes epilepsy. The Journal of general physiology 133(3):283-294.
484
33.
Lee US & Cui J (2009) β subunit-specific modulations of BK channel function by a mutation
485
associated with epilepsy and dyskinesia. J Physiol 587(Pt 7):1481-1498.
486
34.
Clemente-Perez A, et al. (2017) Distinct Thalamic Reticular Cell Types Differentially Modulate
487
Normal and Pathological Cortical Rhythms. Cell reports 19(10):2130-2142.
488
35.
Sarkisova K & van Luijtelaar G (2011) The WAG/Rij strain: a genetic animal model of absence
489
epilepsy with comorbidity of depression [corrected]. Prog Neuropsychopharmacol Biol
490
Psychiatry 35(4):854-876.
491
36.
Crunelli V & Leresche N (2002) Childhood absence epilepsy: genes, channels, neurons and
492
networks. Nat Rev Neurosci 3(5):371-382.
493
37.
Borowicz KK, Piskorska B, Stepniak B, & Czuczwar SJ (2012) Effects of fluoxetine on the
494
anticonvulsant action of valproate and ethosuximide in mouse model of myoclonic convulsions.
495
Annals of agricultural and environmental medicine : AAEM 19(3):487-490.
496
38.
Shah MM (2014) Cortical HCN channels: function, trafficking and plasticity. J Physiol
497
592(13):2711-2719.
498
39.
Sheehan JJ, Benedetti BL, & Barth AL (2009) Anticonvulsant effects of the BK-channel
499
antagonist paxilline. Epilepsia 50(4):711-720.
500
40.
Mehranfard N, Gholamipour-Badie H, Motamedi F, Janahmadi M, & Naderi N (2014) The effect
501
of paxilline on early alterations of electrophysiological properties of dentate gyrus granule cells
502
23
in pilocarpine-treated rats. Iranian journal of pharmaceutical research : IJPR 13(Suppl):125-
503
132.
504
41.
Sun AX, et al. (2019) Potassium channel dysfunction in human neuronal models of Angelman
505
syndrome. Science 366(6472):1486.
506
42.
Brooks SP & Dunnett SB (2009) Tests to assess motor phenotype in mice: a user's guide. Nat
507
Rev Neurosci 10(7):519-529.
508
43.
Sausbier M, et al. (2004) Cerebellar ataxia and Purkinje cell dysfunction caused by Ca2+-
509
activated K+ channel deficiency. Proc Natl Acad Sci U S A 101(25):9474-9478.
510
44.
Chen X, et al. (2010) Disruption of the olivo-cerebellar circuit by Purkinje neuron-specific
511
ablation of BK channels. Proc Natl Acad Sci U S A 107(27):12323-12328.
512
45.
Kessler SK & McGinnis E (2019) A Practical Guide to Treatment of Childhood Absence
513
Epilepsy. Paediatr Drugs 21(1):15-24.
514
46.
Zimmerman FT & Burgemeister BB (1958) A new drug for petit mal epilepsy. Neurology
515
8(10):769-775.
516
47.
Goren MZ & Onat F (2007) Ethosuximide: from bench to bedside. CNS drug reviews 13(2):224-
517
239.
518
48.
Chapman A, Keane PE, Meldrum BS, Simiand J, & Vernieres JC (1982) Mechanism of
519
anticonvulsant action of valproate. Prog Neurobiol 19(4):315-359.
520
49.
Glauser TA, et al. (2010) Ethosuximide, valproic acid, and lamotrigine in childhood absence
521
epilepsy. N Engl J Med 362(9):790-799.
522
50.
Goldenberg MM (2010) Overview of drugs used for epilepsy and seizures: etiology, diagnosis,
523
and treatment. P & T : a peer-reviewed journal for formulary management 35(7):392-415.
524
51.
Tanner MR, et al. (2015) KCa1.1 Inhibition Attenuates Fibroblast-like Synoviocyte Invasiveness
525
and Ameliorates Disease in Rat Models of Rheumatoid Arthritis. Arthritis & Rheumatology
526
67(1):96-106.
527
52.
Peng Q, et al. (2019) The Rodent Models of Dyskinesia and Their Behavioral Assessment. Front
528
Neurol 10:1016.
529
53.
Van Hese P, Martens JP, Waterschoot L, Boon P, & Lemahieu I (2009) Automatic detection of
530
spike and wave discharges in the EEG of genetic absence epilepsy rats from Strasbourg. IEEE
531
Trans Biomed Eng 56(3):706-717.
532
54.
Liu D, et al. (2020) A common hub for sleep and motor control in the substantia nigra. Science
533
367(6476):440-445.
534
55.
Shimada T & Yamagata K (2018) Pentylenetetrazole-Induced Kindling Mouse Model. Journal
535
of visualized experiments : JoVE (136).
536
56.
Au - Ting JT, et al. (2018) Preparation of Acute Brain Slices Using an Optimized N-Methyl-D-
537
glucamine Protective Recovery Method. JoVE (132):e53825.
538
57.
Hunanyan AS, et al. (2015) Knock-in mouse model of alternating hemiplegia of childhood:
539
behavioral and electrophysiologic characterization. Epilepsia 56(1):82-93.
540
24
58.
Livingston S, Kajdi L, & Bridge EM (1948) The use of benzedrine and dexedrine sulfate in the
541
treatment of epilepsy. The Journal of pediatrics 32(5):490-494.
542
59.
Fletcher CF, et al. (1996) Absence epilepsy in tottering mutant mice is associated with calcium
543
channel defects. Cell 87(4):607-617.
544
60.
Noebels JL & Sidman RL (1979) Inherited epilepsy: spike-wave and focal motor seizures in the
545
mutant mouse tottering. Science 204(4399):1334-1336.
546
61.
Coenen AM, Drinkenburg WH, Inoue M, & van Luijtelaar EL (1992) Genetic models of absence
547
epilepsy, with emphasis on the WAG/Rij strain of rats. Epilepsy Res 12(2):75-86.
548
62.
Midzianovskaia IS, Kuznetsova GD, Coenen AM, Spiridonov AM, & van Luijtelaar EL (2001)
549
Electrophysiological and pharmacological characteristics of two types of spike-wave discharges
550
in WAG/Rij rats. Brain Res 911(1):62-70.
551
25
Fig. 1. BK-D434G knock-in mice have spontaneous absence seizures. (A) Ca2+- and voltage-
552
activated BK channels control the opening of voltage-gated Ca2+ channels (VGCCs or CaVs) through a
553
negative-feedback mechanism. (B) Schematic of the simultaneous video-electroencephalogram (EEG)
554
recordings of freely moving mice. (C) BKDG/WT (Cii) and BKDG/DG (Ciii) but not BKWT/WT mice (Ci) had
555
spontaneous spike-wave discharges (SWDs) and frequent behavior arrest. Top: raw EEG traces. Red
556
rectangles show the corresponding EEG traces on an expanded time scale. Middle: corresponding
557
spectrograms of the EEG traces. Bottom: video-based analysis of the total movement. The behavior
558
status is classified as motion state (red boxes) or immobile state (white boxes). See Methods for details.
559
(D) Summary of power spectral density of EEG recorded from BKWT/WT (n = 9), BKDG/WT (n = 9) and
560
BKDG/DG (n = 12) mice. Normalization was performed by averaging the power to the total recording
561
time. Two-way ANOVA, F(2,27) = 9.683, P = 0.0007. (E) Summary of the number of spontaneous SWDs
562
per hour for from BKWT/WT (n = 9), BKDG/WT (n = 9) and BKDG/DG (n = 12) mice. One-way ANOVA test,
563
26
F(2,27) = 11.57, P = 0.0002. * P < 0.05, ***P < 0.001, ****P < 0.0001. In all plots and statistical tests,
564
summary graphs show mean ± s.e.m.
565
27
Fig. 2. First-line anti-absence seizure medicines can abolish the absence seizure in BK-D434G
566
mouse. (A, C) Representative EEG traces, corresponding spectrograms, and total movement from the
567
BKDG/DG mice before and after the first-line anti-absence seizure medicines valproate (A) or ESM (C)
568
administration. (B, D) Summary of power spectral density of EEG recorded from BKDG/DG mice during
569
before and after valproate (B, 200 mg/kg, Two-way ANOVA, F(1,10) = 13.30, P = 0.0045. n = 6 mice) or
570
ESM (D, 150 mg/kg, Two-way ANOVA, F(1,12) = 66.10, P < 0.0001. n = 7 mice) administration. (E)
571
Time course of the drug effects of ethosuximide (ESM, orange), valproate (purple) and saline control
572
(black) on the spontaneous SWDs of the BKDG/DG mice. (Bin size = 5 min). The drug effects were
573
empirically divided into 4 different phases: baseline phase, 60 min prior to injection (grey box, two-way
574
repeated-measures ANOVA, F(2,17) = 1.176, P = 0.3324); early phase, 30 min post injection (yellow box,
575
two-way repeated-measures ANOVA, F(2,17) = 78.61, P < 0.0001); middle phase, from 35 to 80 min post
576
injection (blue box, two-way repeated-measures ANOVA, F(2,17) = 3.906, P = 0.0402); and late phase,
577
from 85 to 105 min post injection (green box, two-way repeated-measures ANOVA, F(2,17) = 0.5274, P =
578
0.5274). n = 7 mice per group for saline control, ESM, n = 6 mice for valproate administration. The
579
error bars indicate s.e.m., * P < 0.05, ** P < 0.01, *** P < 0.001, **** P < 0.0001.
580
581
28
582
Fig. 3. Whole-cell electrophysiology shows that the BK-D434G cortical pyramidal neurons are
583
hyperactive. (A) Representative evoked action potentials in cortical pyramidal neurons from BKWT/WT
584
and BKDG/WT mice. Firing was elicited by a 400 pA current injection for 1s. (B) Action potential
585
frequency of the first inter-spike interval (1st ISI) from the BKWT/WT and BKDG/WT cortical neurons. Two-
586
way repeated-measures ANOVA, F(1,42) = 6.640, P = 0.0136. (C) Representative single action potential
587
waveforms elicited by 400 pA current injection. Definitions of action potential parameters are labeled
588
with cyan and black dash line. AP90 was used to define as action potential duration of 90%
589
repolarization and AHP denotes after hyperpolarization. (D, E) BKDG/WT cortical neurons have shorter
590
action potential duration (D, two-tailed unpaired Student’s t-test, t42=3.111, P = 0.0033) and higher
591
amplitude of fast AHP (E, two-tailed unpaired Student’s t-test, t42=4.294, P = 0.0001) compared with
592
BKWT/WT neurons. n = 29 neurons from five BKWT/WT mice and n = 15 neurons from four BKDG/WT mice.
593
(F) The gain-of-function BK-D434G mutant channels enhance membrane repolarization and accelerate
594
de-inactivation of the voltage-gated sodium (NaV) channels, which enable the excitatory neurons to fire
595
at a higher frequency. * P < 0.05, ** P < 0.01, *** P < 0.001. In all plots and statistical tests, summary
596
graphs show mean ± s.e.m.
597
598
29
599
Fig. 4. Paxilline (PAX), a BK channel blocker, reduces the hyperactivity of the cortical pyramidal
600
neurons and suppresses the spontaneous absence seizures of the BK-D434G mice. (A)
601
Representative evoked action potentials in cortical pyramidal neurons from BKDG/WT mice. Firing was
602
elicited by 1s, 400 pA current injection, before and after application of 10 µM PAX. (B) Action
603
potential frequency of the first inter-spike interval (1st ISI) for the BKDG/WT cortical neurons before and
604
after application of PAX. Two-way repeated-measures ANOVA, F(1,14) = 4.799, P = 0.0459. n = 8
605
neurons from 3 mice. (C) Representative single action potential waveforms of the BKDG/WT cortical
606
30
neurons elicited by 400 pA current injection before and after application of PAX. (D, E) PAX broadens
607
action potential (AP) duration (D, two-tailed paired Student’s t-test, t7=4.591, P = 0.0025) and
608
suppresses after-hyperpolarization (AHP) in the cortical pyramidal cells from BKDG/WT mice (E, two-
609
tailed paired Student’s t-test, t7=5.581, P = 0.0008). n = 8 neurons from 3 mice. (F) Summary of power
610
spectral density of EEG recorded from BKDG/DG mice before and after PAX administration (n = 7 mice).
611
Two-way ANOVA, F(1,12) = 26.39, P = 0.0002. n = 7 mice. (G) Representative EEG traces,
612
corresponding spectrograms, and total movement from the BKDG/DG mice before (left panel) and after
613
0.35 mg/kg PAX (right panel) administration. (H) Time course of the drug effects of PAX (orange line)
614
and saline control (black dash line) on the spontaneous SWDs of the BKDG/DG mice. (Bin size = 5 min).
615
PAX is effective in the early phase, 30 min post injection (yellow box, two-way repeated-measures
616
ANOVA, F(1,12) = 42.57, P < 0.0001), but not in the later phases. n = 7 mice per group. The error bars
617
indicate s.e.m., * P < 0.05, *** P < 0.001, **** P < 0.0001. In all plots and statistical tests, summary
618
graphs show mean ± s.e.m.
619
31
Fig. 5. BK-D434G mice exhibit motor defects. (A) Representative animal track in the open-field
620
chamber of BKWT/WT, BKDG/WT and BKDG/DG mice. (B) BKDG/DG mice showed decreased locomotor
621
activity in 15 min open field test. Two-way repeated-measures ANOVA, F(2,23) = 9.974, P = 0.0008.
622
BKWT/WT, n = 9 mice, BKDG/WT, n = 12 mice, BKDG/DG, n = 5 mice. (C) Balance beam test of BK-D434G
623
mutation mice, the red arrows indicate the low position of the tails when hind-limb slips caused balance
624
loss were observed during the test. (D, E) BKDG/DG mice had significantly more hind-limb slips on the
625
balance beams (D, One-way ANOVA test, F(2,27) = 75.05, P < 0.0001) and took significantly longer to
626
traverse the balance beam (E, One-way ANOVA test, F(2,27) = 27.14, P < 0.0001) compared with
627
BKWT/WT controls. BKWT/WT, n = 15 mice, BKDG/WT, n = 12 mice, BKDG/DG, n = 3 mice. (F) Accelerating
628
rotarod latency to fall times for BK-D434G mutation mice over four days of testing compared with
629
control mice, showing a significant deficit for BK- D434G mice in this test. ****p < 0.0001, significant
630
difference between genotypes for those trials. Two-way repeated-measures ANOVA, F(2,28) = 19.56, P <
631
0.0001. BKWT/WT, n = 8 mice, BKDG/WT, n = 12 mice, BKDG/DG, n = 11 mice. (G) Representative images
632
of the gait patterns of the BKWT/WT, BKDG/WT and BKDG/DG mice, with forepaws are represented by red
633
paint and hind-paws by blue paint (scale bar, 2 cm). (H) Quantification reveals shortened stride length.
634
One-way ANOVA test, F(2,12) = 18.50, P = 0.0002. BKWT/WT, n = 5 mice, BKDG/WT, n = 4 mice,
635
BKDG/DG, n = 6 mice. In all plots and statistical tests, summary graphs show mean ± s.e.m., * p<0.05, **
636
p<0.01, ***p<0.001, ****p<0.0001.
637
32
Fig. 6. Cerebellar Purkinje cells (PCs) from the BK-D434G mutant mice are hyperactive, which
638
can be suppressed by paxilline. (A) Representative confocal images of PCs. The PCs from the
639
BKDG/WT and BKDG/DG mice show signs of hypertrophy. (B) The BKDG/WT and BKDG/DG mice have
640
enlarged PC somas. One-way ANOVA test, F(2,19) = 14.97, P = 0.0001, BKWT/WT, n = 6; BKDG/WT, n = 8;
641
BKDG/DG, n = 8. (C) The BKDG/WT and BKDG/DG mice have thickened main dendrite width. The dendrite
642
width is measured at the distance of one cell-soma diameter from the cell soma. One-way ANOVA test,
643
F(2,8) = 34.46, P = 0.0001, BKWT/WT, n = 3; BKDG/WT, n = 4; BKDG/DG, n = 4. (D) Representative evoked
644
action potentials in cerebellar PCs from BKWT/WT and BKDG/WT mice. Firing was elicited by 1s long 600
645
pA current injection, before and after application of 10 µM PAX. (E) Statics of the action potential
646
numbers of BKWT/WT and BKDG/WT PCs, before and after application of PAX. Two-way repeated-
647
measures ANOVA, F(3,20) = 20.06, P < 0.0001. (F) Representative single action potential waveforms
648
elicited by 600 pA current injection for BKWT/WT and BKDG/WT mice before and after 10 µM PAX. In the
649
right panel, BKWT/WT trace is presented in grey dash line for comparison with BKDG/WT. (G, H) BKDG/WT
650
33
PCs have shorter action potential duration (AP90) and higher amplitude of fast after-hyperpolarization
651
(fAHP) compared with BKWT/WT. PAX broadens AP duration (G, Two-way ANOVA, F(1,10) = 20.27, P
652
< 0.0001) and reduces fAHP (H, , Two-way ANOVA, F(1,10) = 75.31, P < 0.0001) of PCs from both
653
BKWT/WT and BKDG/WT mice. n = 6 neurons from 3 mice per group. In all plots and statistical tests,
654
summary graphs show mean ± s.e.m.. * p<0.05, ** p<0.01, ***p<0.001, ****p<0.0001.
655
656
34
657
Table 1. Comparison of the absence seizures in rodent models and human patients
658
659
Human Typical
Absence
seizure
BK-D434G
proband
BK-D434G
mouse
Tottering
mouse
WAG/Rij
rat
EEG
Onset age
3 years
< 6 months
< 4 weeks
3 weeks
> 75 days
Generalized synchronous
SWD
+
+
+
+
+
SWD frequency (Hz)
2.5-4
3-4
3-8
6-7
7-11
SWD duration (s)
4-20
N/A
0.5-10
0.3-10
1-45
Ictal behavior
Staring: myoclonus
+
+
+
+
+
Move during SWD
-
-
-
-
-
Pharmacology
ESM
+
N/A
+
+
+
Valproate
+
+
+
+
+
References
(36, 46, 49, 58)
(13)
Current study (59, 60)
(61, 62)
N/A: not applicable or not available
660
| 2021 | Neuronal mechanism of a BK channelopathy in absence epilepsy and movement disorders | 10.1101/2021.06.30.450615 | [
"Dong Ping",
"Zhang Yang",
"Mikati Mohamad A.",
"Cui Jianmin",
"Yang Huanghe"
] | null |
RESEARCH ARTICLE
Imaging-based evaluation of pathogenicity by novel DNM2 variants associated with
centronuclear myopathy
Kenshiro Fujise1, Mariko Okubo2,3, Tadashi Abe1, Hiroshi Yamada1, Kohji Takei1,
Ichizo Nishino2, *Tetsuya Takeda1 and Satoru Noguchi2
1Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, Okayama
University, Okayama, Japan; 2National Institute of Neuroscience, National Center of
Neurology and Psychiatry (NCNP), Kodaira, Tokyo, Japan; 3Department of Pediatrics,
The University of Tokyo, Tokyo, Japan.
*To whom correspondence should be addressed:
1. Tetsuya Takeda, PhD: Department of Biochemistry, Okayama University Graduate
School of Medicine, Dentistry and Pharmaceutical Sciences, Shikata-cho 2-5-1, Kita-ku,
Okayama 700-8558, Japan; E-mail: ttakeda@okayama-u.ac.jp; Tel. +81-86-235-7125;
FAX: +81-86-235-7126
Pathogenic evaluation of congenital myopathy
ABSTRACT
Centronuclear myopathy (CNM) is characterized clinically by muscle weakness and
pathologically by the presence of centralized nuclei and disarrangement of T-tubules in
muscle fibers. DNM2 which encodes a large GTPase dynamin 2 have been identified as a
causative gene for CNM. Nevertheless, the identification of DNM2 variants may not
always lead to the definitive diagnosis as their pathogenicity is often unknown.
In this study, by imaging T-tubule-like structures reconstituted in cellulo, we
demonstrated that aberrant membrane remodeling by mutant dynamin 2 is tightly
associated with gain-of-function features of DNM2 variants. This simple in cellulo assay
provided quantitative data required for accurately evaluating pathogenicity of reported
and novel DNM2 variants identified from CNM patients in our cohort. Our approaches
combining the in cellulo assay with clinical information of the patients enabled to explain
the course of a disease progression by pathogenesis of each variant in DNM2-associated
CNM.
KEYWORDS
centronuclear myopathy, DNM2, in cellulo assay, T-tubules, membrane remodeling
Pathogenic evaluation of congenital myopathy
INTRODUCTION
Centronuclear myopathy (CNM) is a congenital myopathy characterized clinically
by slowly progressive muscle weakness and pathologically by the presence of centralized
myonuclei, radial arrangement of sarcoplasmic strands of oxidative enzymes, and type 1
fiber predominance and hypotrophy(Jungbluth, Wallgren-Pettersson, & Laporte, 2008).
In addition, T-tubules (transverse tubules) and triads are disorganized on electron
microscopic examination(Al-Qusairi & Laporte, 2011). So far, seven genes, MTM1,
SPEG, BIN1, DNM2, RYR1, TTN and CCDC78, have been reported to be causative for
CNM(Agrawal et al., 2014; Jungbluth & Gautel, 2014; Majczenko et al., 2012; Romero,
2010). Diverse clinical manifestations among CNM patients are mainly attributed to the
different responsible genes and mutations. In DNM2, 26 pathogenic variants have been
reported to cause the autosomal dominant form of CNM with relatively mild and slowly
progressive symptoms(Biancalana et al., 2018; Bohm et al., 2012; Casar-Borota et al.,
2015; Hohendahl, Roux, & Galli, 2016). DNM2 encodes the ubiquitous isoform of
dynamin, dynamin 2, which is a GTPase essential for membrane fission in
endocytosis(Antonny et al., 2016; Ferguson & De Camilli, 2012). The defects in either
self-assembly or membrane binding ability of dynamin 2 should be the common cause
of DNM2-associated CNM since most of the reported pathogenic DNM2 variants were
missense in the stalk and pleckstrin homology (PH) domain.
A huge number of variants are now identified for various congenital diseases by
massive parallel sequencing technologies but pathogenicity is not unknown for many of
them mostly due to lack of functional testing. This is also the case with DNM2-associated
CNM as functional assays, such as GTPase measurement, self- assembly and membrane
Pathogenic evaluation of congenital myopathy
binding assays of the mutated dynamin, which may well determine pathogenicity of the
identified variants, are not always accessible without special settings.
In this study, pathogenicity of both reported and novel DNM2 variants associated
with CNM were systematically and quantitatively analysed by imaging T-tubule like
structures (TLS) reconstituted in cellulo. Our approaches combining the in cellulo assay
with clinical information clearly demonstrated strong correlation between genotypes,
cellulotypes and disease phenotypes in DNM2-associated CNM. These results suggest
that aberrant membrane remodeling by DNM2 variants is tightly linked to the
pathogenesis and prognosis of CNM.
Pathogenic evaluation of congenital myopathy
MATERIALS AND METHODS
Editorial Policies and Ethical Considerations
National Center of Neurology and Psychiatry (NCNP) has been functioning as a
referral center for muscle disease since 1978. All the samples and clinical data used in this
study were sent to NCNP from the physicians for diagnostic purposes (until March 2019).
Written consent was obtained from parents or guardians. This study was approved by the
Ethics Committee in NCNP.
Genetical and histological analyses of patients
Genetic variants were analysed using genomic DNA from serum or biopsied muscles
from 3933 cases which was suspected to be muscle diseases. Genetic analyses were
performed by targeted re-sequencing that covered all exonic regions and exon-intron
borders in DNM2 gene (2858 analyses) and/or by whole exome sequencing (2599
analyses) as described previously (Nishikawa, Mitsuhashi, Miyata, & Nishino, 2017).
Histological analysis was performed using muscle samples taken from the biceps brachii
and then frozen in isopentane cooled in liquid nitrogen as described previously (Okubo et
al., 2018).
Plasmid construction, cell culture and DNA transfection
All the expression constructs used in this study were generated using Gateway
Cloning Technology (Thermo Fisher Scientific). Entry clones of human dynamin 2
(NM_001005360) and human BIN1 isoform 8 (NM_004305.4) were prepared by B-P
recombination cloning of PCR products respectively amplified from
pcDNA3.1-GFP-Topo-hDNM2-WT (generous gift from P. Guicheney, UPMC) and
Pathogenic evaluation of congenital myopathy
pEGFP-mAmph2 (generous gift from P. De Camili, Yale University) using
corresponding primers (BIN1 fw:
5’-ggggacaagtttgtacaaaaaagcaggctgcatggcagagatgggcag-3’; BIN1 rv:
5’-ggggaccactttgtacaagaaagctgggtctgggaccctctcagtgaag-3’, Dynamin 2 fw:
5’-ggggacaagtttgtacaaaaaagcaggctgcatgggcaaccgcggga-3’, Dynamin 2 rv:
5’-ggggaccactttgtacaagaaagctgggtcgtcgagcagggatggc-3’) into pDONR201 vector.
Expression constructs of dynamin 2 and BIN1 were prepared by L-R recombination
cloning of their Entry clones into Destination vectors (generous gift from H. McMahon,
MRC-LMB) either for expressing proteins in mammalian cells (pCI vectors for
expressing FLAG-, or RFP-tagged proteins) or for bacterial protein expression (pET15b
for His-fusions and pGEX-6P-2 for GST-fusions) (generous gift from H. McMahon,
MRC-LMB).
C2C12 cells (ATCC CRL-1722) was grown in D-MEM (High Glucose) with
L-Glutamine, Phenol Red and Sodium Pyruvate (FUJIFILM Wako chemicals,
043-30085) supplemented with 10% fetal bovine serum (FBS) (Gibco, 12483, Lot
No.1010399) and Penicillin-Streptomycin (Gibco, 15140122) at 37 °C in 5% CO2. For
transfection of C2C12, 70% confluent cells in VIOLAMO VTC-P24 24-well plates (AS
ONE, 2-8588-03) were transfected with 0.5 �g expression plasmids using Lipofectamine
LTX with Plus Reagent (Thermo Fisher Scientific, 15338100). To examine consequences
of the expression of BIN1 or dynamin 2 in either wild type (WT) or mutant forms, cells
were fixed after 48 h of the transfection for phenotypic analyses.
Introduction of CNM mutations into dynamin 2
Pathogenic evaluation of congenital myopathy
Entry clones for the CNM mutants of dynamin-2 (R465W) was prepared by BP
recombination reaction of PCR products amplified from
pcDNA3.1-GFP-Topo-hDNM2-R465W (generous gift from P. Guicheney, UPMC)
using corresponding primers (Supplementary Table 1) into pDONR201. Entry clones for
other mutant dynamin 2 were prepared by introducing corresponding mutations into the
Entry clone of wild type human dynamin 2 using QuikChange Lightning Site-directed
Mutagenesis kit (Agilent Technologies, 210518) following manufacturer’s instruction.
Sense and antisense primers used for the site-directed mutagenesis are as follows.
E368K sense :5’-tcaatcgcatcttccacaagcggttcccatttgag-3’
E368K antisense: 5’-ctcaaatgggaaccgcttgtggaagatgcgattga-3’
R369Q sense: 5’-cgcatcttccacgagcagttcccatttgagctg-3’
R369Q antisense: 5’-cagctcaaatgggaactgctcgtggaagatgcg-3’
S619L sense: 5’-cagctggaaggccttgttcctccgagctg-3’
S619L antisense: 5’-cagctcggaggaacaaggccttccagctg-3’
G495R sense: 5’-ccatgaggacttcatcaggtttgccaatgccca-3’
G495R antisense: 5’-tgggcattggcaaacctgatgaagtcctcatgg-3’
V520G sense: 5’-gggagatcctggggatccgcagggg-3’
V520G antisense: 5’-cccctgcggatccccaggatctccc-3’
G624V sense: 5’-ttcctccgagctgtcgtctaccccgag-3’
G624V antisense: 5’-ctcggggtagacgacagctcggaggaa-3’
P294L sense: 5’-gggagtcgctgctggccctacgtag-3’
P294L antisense: 5’-ctacgtagggccagcagcgactccc-3’
R724H sense: 5’-ggacgacatgctgcacatgtaccatgccc-3’
R724H antisense: 5’-gggcatggtacatgtgcagcatgtcgtcc-3’
Pathogenic evaluation of congenital myopathy
Immunofluorescent microscopy and quantitative analysis of TLS
Primary antibodies used in this study were polyclonal rabbit anti-DDDDK tag (MBL,
PM020). The secondly antibody used in this study, Alexa Fluor 488-conjugated donkey
anti-Rabbit IgG (H+L) (A21206), was purchased from Thermo Fisher Scientific.
For immunostaining of C2C12, cells grown on coverslips were fixed with 4%
paraformaldehyde (EMS, 15710) in PBS for 15 min at room temperature. After washing
with PBSTB (PBS containing 0.1% Triton X-100, 1% BSA), the cells were
permeabilized and blocked with PBS containing 0.5% Triton X-100 and 3% BSA for 1 h
at room temperature. The samples were then incubated with primary antibodies diluted
1:1000 in PBSTB overnight at 4 °C in a humid chamber. After washing with PBSTB, the
cells were incubated with secondly antibodies diluted in PBSTB for 3 h at room
temperature. Then, the cells were washed with PBSTB and mounted in Fluoromount/Plus
(K048, Diagnostic BioSystems). Immunostained cells were observed under BX51
fluorescence microscope (OLYMPUS) and images were acquired with Discovery MH15
CMOS camera and ISCapture image acquisition software (Tucsen). All images were
analyzed using FIJI (Schindelin et al., 2012) and processed with Photoshop (Adobe).
Quantitative analysis of TLS
Quantitative analysis of the TLS was performed by FIJI as described previously
(Fujise et al., 2020). Firstly, background signal was subtracted from microscopic images
of BIN1-expressing cells (Rolling ball radius = 10 pixels). Then, the membrane tubules
were enhanced with FFT Bandpass Filter (Filter: large structures down to 5 pixels and up
to 3 pixels; Suppress stripes: None; Tolerance of direction: 5%). The membrane tubules
Pathogenic evaluation of congenital myopathy
were detected and binarized with Threshold command and the binarized membrane
tubules were skeletonized to be analyzed with Analyze Skeleton (2D/3D) plugin.
Membrane tubules with length between 0.5 and 2 �m were considered to be as “short”.
In vitro sedimentation assay using recombinant BIN1 and dynamin 2
Recombinant protein of BIN1 isoform 8 and dynamin 2 was expressed and purified
as GST fusion and His-tagged proteins, respectively as described previously (Fujise et al.,
2020).
In vitro sedimentation assay of dynamin 2 was performed as described previously
(Fujise et al., 2020)．In short, wild type or CNM mutant (E368K, R369Q, R465W and
S619L) dynamin 2 were diluted to 1 �M in reaction buffer (10 mM Hepes, 2 mM MgCl2,
100 mM NaCl, pH 7.5) and incubated for 5 min at 37 �. To induce disassembly, 1 mM
GTP was added to the preassembled dynamin 2 and incubated for 5 min at 37 �. The
samples were centrifuged at 230,000g for 10 min at 25 � using CS100GXL
ultracentrifuge and S120AT3 rotor (Eppendorf Himac Technologies) and resultant
supernatant and pellet were analyzed by SDS-PAGE followed by Coomasie Brilliant
Blue R-250 staining.
Dynamin GTPase activity
GTPase activity of dynamin 2 was determined by monitoring release of free
orthophosphate using malachite green assay as described previously (Fujise et al., 2020).
The malachite green reagent was prepared by mixing solution A (17 mg of Malachite
Green Carbinol base dye (229105, Merck) in 20 mL 1 N HCl) and Solution B (0.5 g
Ammonium molybdate (277908, Merck) in 7 mL 4 N HCl) with filling up to 50 mL by
Pathogenic evaluation of congenital myopathy
MilliQ water followed by filtration through 0.45 �m membrane (S-2504, KURABO). In
the assay, 0.2� �M dynamin in the presence of BIN1 at different molar ratio was mixed
with 1 mM GTP in GTPase reaction buffer (10 mM Hepes, 2 mM MgCl2, 50 mM NaCl,
pH 7.5) with or without 0.005 �g/�L lipid nanotubes and incubated for 5 min at 37 �.
After the reaction was stopped on ice for 10 min, 160 �L of malachite green reagent was
added to the 40 �L of the reaction mix in 96 well plate (442404, Thermo Fisher
Scientific). After 5 min shaking at 1200 rpm with Digital MicroPlate Genie Pulse
(Scientific Industries, Inc.), released orthophosphate was colorimetrically quantified by
measuring OD 650 nm using a microplate reader (SH-1000, CORONA ELECTRIC).
Statistical data analysis
Statistical data analysis was performed using Prism 8 (GraphPad Software) and
Excel (Microsoft). For all quantification provided, means and SEM are shown.
Statistical significance was determined using a two-sided t test and P values are shown
in the figures.
Data availability
All relevant data are included with the manuscript or available from the authors upon
request.
Pathogenic evaluation of congenital myopathy
RESULTS
Identification of SNVs from CNM patients by cohort analyses
We identified 17 sporadic patients with DNM2 variants in 3933 cases who were
suspected to have muscle diseases. Among these patients, 11 patients carried reported
variants (Supplementary Table 1), while 6 had novel missense variants (Supplementary
Table 2). In total, five novel heterozygous variants, c.1483G>A (p.G495R), c.1559T>G
(p.V520G), c.1871G>T (p.G624V), c.881C>T (p.P294L) and c.2171G>A (p.R724H),
were identified (Supplementary Table 2). The predicted substitutions in amino acid
residues occurred either at the unstructured loops in PH domain (Val520 and Gly624) and
stalk domains (Gly495) or at bundle signaling element domain, a flexible hinge between
the G-domain and stalk (Pro294 and Arg724) based on the crystal structure of human
dynamin 1 (Supplementary Fig. 1). Histological analyses of skeletal muscle biopsies
from the patients with reported variants exhibited typical myotubular myopathic features
(P2) or CNM pathological features (P1, P3-P11), including centrally placed nuclei,
peripheral halo, presence of radial sarcoplasmic strands, type 1 fiber predominance and
adipose tissue infiltration (Supplementary Fig. 2). Consistently, the typical
clinicopathological features of CNM were observed in patients with the novel variants
except for P16 and P17 (Supplementary Table 2).
CNM variants induced gain-of-function features of dynamin 2 in vitro
To characterize pathogenicity of reported DNM2 variants, we analyzed the mutant
dynamins (E368K, R369Q, R465W and S619L) for their self-assembly and GTPase
activities, both of which are essential for membrane fission by dynamin (Fujise et al.,
2020; Marks et al., 2001; Ramachandran et al., 2007; Wang et al., 2010; Warnock,
Pathogenic evaluation of congenital myopathy
Hinshaw, & Schmid, 1996). In the sedimentation assay, purified wild type and mutant
dynamin 2 self-assembled in the absence of GTP and more than 90% of proteins are
recovered in the precipitate (Fig. 1A and Supplementary Fig. 3, -GTP). Previous studies
demonstrated that CNM mutants of dynamin 2 self-assemble to form stable polymers
resistant to GTP hydrolysis-dependent disassembly (Fujise et al., 2020; Ramachandran et
al., 2007; Wang et al., 2010). Consistently, almost all the mutant dynamin 2 remained in
the precipitate even after GTP addition (Fig. 1A and Supplementary Fig. 3, E368K,
R369Q, R465W and S619L, +GTP). In contrast, more than 30% of self-assembled wild
type dynamin 2 were disassembled and recovered in the supernatant after GTP addition
(Fig. 1A and Supplementary Fig. 3, WT, +GTP).
Previous studies showed that elevated GTPase activity is a character of mutant
dynamin 2 with CNM variants (Chin et al., 2015; Fujise et al., 2020; Kenniston &
Lemmon, 2010; Wang et al., 2010). Consistent with the previous studies, E368K, R369Q,
and S619L mutants exhibited higher GTPase activities compared to that of wild type
dynamin 2 (Fig. 1B). However, statistical significance of the elevated GTPase activity
was not confirmed for R465W mutant (Fig. 1B). Previous studies showed that BIN1 is a
negative regulator of dynamin 2 (Cowling et al., 2017; Fujise et al., 2020). Consistently,
GTPase activity of wild type dynamin 2 was stoichiometrically inhibited by BIN1 (Fig.
1C and Supplementary Fig. 4, WT). In contrast, BIN1 failed to inhibit GTPase activities
of some mutant dynamin 2 (E368K and S619L), but, interestingly, those of other mutants
(R369Q and R465W) were inhibited by BIN1 (Fig. 1C and Supplementary Fig. 4). These
in vitro data suggest that CNM mutants generally exhibit elevated GTPase activity and
they are resistant to BIN1-mediated inhibition, although existence of the exceptional
Pathogenic evaluation of congenital myopathy
variants in GTPase activation and BIN1 sensitivity suggest limitation of the experimental
approaches.
Imaging-based evaluation of functional defects caused by genetic variants in DNM2
In C2C12 cells, FLAG-tagged wild type dynamin 2 formed very fine puncta, while all
the mutant dynamin 2 with reported variants either in the stalk (E368K, R369Q and
R465W) or in the PH domain (S619L) formed abnormally large puncta (Fig. 1D, F and
G). We previously demonstrated that co-expression of dynamin 2 and BIN1 in C2C12
cells induce TLS, membranous tubular structures mimicking T-tubules in skeletal
muscles (Fujise et al., 2020). Wild type dynamin 2 was recruited to BIN1 to induce
thicker and unevenly distributed TLS (Fig. 1E, DNM2 WT-FLAG). In contrast, all the
mutant dynamin 2 induced shorter dot-like TLS despite they are still colocalized with
BIN1 (Fig. 1E). Quantitative analyses showed that the number of shorter TLS (0.5-2 �m)
was increased in the presence of mutant dynamin 2 (Fig. 1H).
Dynamin 2 with CNM-associated novel variants induced shorter TLS
We next explored pathogenicity of the novel DNM2 variants that cause P294L,
G495R, V520G, G624V and R724H substitutions by analyzing their effects on TLS
formation. Similar to the reported CNM mutants, FLAG-tagged proteins of the novel
DNM2 variants, G495R, V520G and G624V formed aggregates in C2C12 cells (Fig. 2A,
C and D). Furthermore, these three mutants also induced significantly shorter TLS
compared to those with wild type dynamin 2 (Figs. 2B and E). Interestingly, two novel
DNM2 variants, P294L and R724H, formed neither aggregates (Fig. 2A, C and D) nor
aberrantly shorter TLS (Fig. 2B and E). These findings are compatible with the
Pathogenic evaluation of congenital myopathy
histological and clinical features of the patients P16 and P17, harboring c.881C>T
(P294L) and c.2171G>A (R724H) variants, in which atypical CNM histopathology with
low number of centrally placed nuclei devoid of radial strands were observed
(Supplementary Fig. 2).
Correlation of defective membrane tubulation and clinical phenotypes of patients
We hypothesized that the aberrant membrane remodeling activity by mutated
dynamin 2 were implication of pathogenicity by each variant and clinical severity of
CNM patients. Thus, we analyzed correlation between short TLS formation (Figs 1 and 2)
and the clinical parameters of the patients with DNM2 variants (ages of disease onset,
biopsy and disease duration) (Supplementary Table 1 and 2). Among these parameters,
ages of the disease onset and short TLS formation represented a linear correlation with
high correlation coefficient (r = -0.74) (Fig. 3A). In contrast, the disease duration and the
short TLS formation were not correlated (r = -0.1) (Figs. 3B). Thus, the defective
membrane remodeling can explain and predict the variant-dependent occurrences of
muscular weakness in DNM2-associated CNM. Importantly, P16 and P17 were
distributed far from the line on the graph (Fig. 3A). Based on these data together with the
pathological features, we concluded that novel variants, G495R, V520G and G624V, but
not P294L and R724H, are likely to be pathogenic variants. Interestingly, the results of in
cellulo experiments and ages of biopsy were also well correlated as plotted on an
exponential curve (r = -0.97) except for those with R465W variant (Figs. 3C). As
mentioned above, typical DNM2-associated CNM has milder and slowly progressive
symptoms and favorable prognosis, but the onset is at infantile to adolescence. Our
patients with reported variants were compatible to those features (Fig.3D). In contrast,
Pathogenic evaluation of congenital myopathy
the patients with novel variants were remarkably late-onset as a few previous reports
about late-onset DNM2-CNM patients (Fig. 3D).
Pathogenic evaluation of congenital myopathy
DISCUSSION
Recent advancement of the massive parallel sequencing technologies provides us
with an enormous amount of genomic data from patients of various congenital diseases.
Not surprisingly, pathogenicity of many of the identified variants is unknown, which
clearly shows the necessity of simple and fast methods for assaying the functional defects
caused by each variant.
In this study, we evaluated the pathogenicity of CNM-associated or -unassociated
DNM2 variants by imaging TLS formation in cultured cells and demonstrated that short
TLS formation in cellulo and the severity of symptoms are correlated with high
correlation coefficient (Fig. 3). This result suggests that the in cellulo assay in
combination with the genetical and clinicopathological diagnosis are powerful
approaches not only to determine pathogenicity of the genetic variants in DNM2, but also
to predict the disease severity. Since the imaging-based in cellulo assay is easily
accessible but provides with highly reproducible and quantitative results, it may
applicable to elucidate pathomechanisms of triadopathies accompanied with disorganized
T-tubules (Dowling, Lawlor, & Dirksen, 2014).
We analyzed the DNM2 variants identified from CNM patients in our cohort
(Supplementary Table.1 and 2). Both reported (E368K, R369Q, R465W and S619L) and
novel (G495R, V520G and G624V) variants formed abnormal aggregates and short TLS
(Figs. 1 and 2). These in cellulo phenotypes suggest that the novel variants, like reported
DNM2 variants, are responsible for gain-of-function features in self-assembly and
GTPase activity, both of which are essential for membrane fission by dynamin 2.
Furthermore, association between the deficits in TLS and clinical symptom of patients
suggests that p.G495R, p.V520G and p.G624V are likely to be pathogenic (Fig. 3).
Pathogenic evaluation of congenital myopathy
Interestingly, G495R locates in the hinge between stalk (middle domain) and PH domain,
whereas V520G and G624 are located in PH domain flanking stalk region based on the
structure of dynamin 1 monomer (Supplementary Fig. 1). Previous structural studies on
dynamin 1 and 3 demonstrated that PH domain is flipped back to interact with stalk and
GTPase domain to form inhibitory “closed” state which is released upon binding to
membrane by the PH domain (Faelber et al., 2011; Reubold et al., 2015). Thus, it is
possible that these novel CNM variants affect structures of PH and stalk domains to
impair the regulation of GTPase activity causing constitutively active GTPase state.
Future structural studies of these novel dynamin 2 mutants will explain detailed
mechanisms that cause their gain-of-function features linked to CNM pathogenesis.
In this study, we demonstrated correlation between the short TLS formation and ages
of the patients when their biopsy samples were collected (Fig.3C), suggesting possible
prediction of the DNM2 variant-associated prognosis of the disease. Interestingly,
p.R465W variant revealed unique features in both biological and clinical aspects: the
GTPase activity and BIN1-susceptibility of this mutant were similar to those of wild type,
and the patients have early age (infant to adolescence) at onset, and the relatively late age
at biopsy as classified in Group B (Fig. 3D). Further studies are required to elucidate the
precise pathogenesis by the R465W variant.
DNM2-associated CNM represents variable range of clinical phenotype with
association between genetic variants and clinical severities (Bohm et al., 2012). However,
most of the reported variants are associated with either early-onset severe phenotype (e.g.,
E368K, R369Q and S619L) or early-onset but with relatively mild phenotype (e.g.,
R465W). In contrast, only a few patients have been reported to develop late-onset disease.
In support of this notion, all the patients with the reported variants in our cohort had either
Pathogenic evaluation of congenital myopathy
early-onset severe phenotype with early muscle biopsy (Group A) or early-onset but
slowly progressive phenotype with late muscle biopsy (Group B) phenotypes. In contrast,
all of our novel variants, i.e., p.G495R, p.V520G and p.G624V, were associated with the
late-onset phenotype with late muscle biopsy (Group C) (Fig. 3D). Although the patients
in Group C were almost asymptomatic until the third to fourth decades of their life,
progression of muscle weakness after the onset was relatively rapid (Supplementary
Table 2).
Several therapeutic applications targeting mutated dynamin 2 have been developed
on animal studies (Buono et al., 2018; Trochet et al., 2018). Since a clinical trial using
investigational antisense medicine DYN101 are ongoing for DNM2-associated CNM
(NCT04033159) (https://www.clinicaltrials.gov/ct2/show/NCT04033159), establishing
accurate diagnosis of CNM patients is crucial. Our approach using simple in cellulo assay
together with genetical and clinicopathological analyses should contribute to precise
diagnosis, especially when muscle biopsy samples are unavailable for any reasons.
Furthermore, from the therapeutic point of view, early diagnosis by our simple assay also
help to determine the management of the patients.
FUNDING INFORMATION
This work was supported by JSPS KAKENHI, Grant numbers 18K07198,
19KK0180, grants from Wesco Scientific Promotion Foundation and Ryobi Teien
Memory Foundation for T.T. This work was also supported by Intramural Research Grant
(29-4, 2-5 for T.T. and I.N., 2-6, 30-9 for S.N.) for Neuronal and Psychiatric Disorders of
NCNP, and AMED under Grant Numbers JP19ek0109285h0003 for I.N. and S.N.. K.T.
Pathogenic evaluation of congenital myopathy
was supported by JSPS KAKENNHI, Grant number 19H03225. M.O. was supported by
Grant-in-Aid for JSPS Research Fellow Grant Number 19J12028.
Acknowledgements
The authors are thankful to P. Guicheney (UPMC), P. De Camili (Yale University)
and H. McMahon (MRC-LMB) for reagents.
Pathogenic evaluation of congenital myopathy
REFERENCES
Agrawal, P. B., Pierson, C. R., Joshi, M., Liu, X., Ravenscroft, G., Moghadaszadeh,
B., . . . Beggs, A. H. (2014). SPEG interacts with myotubularin, and its
deficiency causes centronuclear myopathy with dilated cardiomyopathy. Am J
Hum Genet, 95(2), 218-226. doi:10.1016/j.ajhg.2014.07.004
Al-Qusairi, L., & Laporte, J. (2011). T-tubule biogenesis and triad formation in skeletal
muscle and implication in human diseases. Skelet Muscle, 1(1), 26.
doi:10.1186/2044-5040-1-26
Antonny, B., Burd, C., De Camilli, P., Chen, E., Daumke, O., Faelber, K., . . . Schmid,
S. (2016). Membrane fission by dynamin: what we know and what we need to
know. EMBO J, 35(21), 2270-2284. doi:10.15252/embj.201694613
Biancalana, V., Romero, N. B., Thuestad, I. J., Ignatius, J., Kataja, J., Gardberg, M., . . .
Laporte, J. (2018). Some DNM2 mutations cause extremely severe congenital
myopathy and phenocopy myotubular myopathy. Acta Neuropathol Commun,
6(1), 93. doi:10.1186/s40478-018-0593-2
Bohm, J., Biancalana, V., Dechene, E. T., Bitoun, M., Pierson, C. R., Schaefer, E., . . .
Laporte, J. (2012). Mutation spectrum in the large GTPase dynamin 2, and
genotype-phenotype correlation in autosomal dominant centronuclear myopathy.
Hum Mutat, 33(6), 949-959. doi:10.1002/humu.22067
Buono, S., Ross, J. A., Tasfaout, H., Levy, Y., Kretz, C., Tayefeh, L., . . . Cowling, B. S.
(2018). Reducing dynamin 2 (DNM2) rescues DNM2-related dominant
centronuclear myopathy. Proc Natl Acad Sci U S A, 115(43), 11066-11071.
doi:10.1073/pnas.1808170115
Pathogenic evaluation of congenital myopathy
Casar-Borota, O., Jacobsson, J., Libelius, R., Oldfors, C. H., Malfatti, E., Romero, N. B.,
& Oldfors, A. (2015). A novel dynamin-2 gene mutation associated with a
late-onset centronuclear myopathy with necklace fibres. Neuromuscul Disord,
25(4), 345-348. doi:10.1016/j.nmd.2015.01.001
Chin, Y. H., Lee, A., Kan, H. W., Laiman, J., Chuang, M. C., Hsieh, S. T., & Liu, Y. W.
(2015). Dynamin-2 mutations associated with centronuclear myopathy are
hypermorphic and lead to T-tubule fragmentation. Hum Mol Genet, 24(19),
5542-5554. doi:10.1093/hmg/ddv285
Cowling, B. S., Prokic, I., Tasfaout, H., Rabai, A., Humbert, F., Rinaldi, B., . . . Laporte,
J. (2017). Amphiphysin (BIN1) negatively regulates dynamin 2 for normal
muscle maturation. J Clin Invest, 127(12), 4477-4487. doi:10.1172/JCI90542
Dowling, J. J., Lawlor, M. W., & Dirksen, R. T. (2014). Triadopathies: an emerging
class of skeletal muscle diseases. Neurotherapeutics, 11(4), 773-785.
doi:10.1007/s13311-014-0300-3
Faelber, K., Posor, Y., Gao, S., Held, M., Roske, Y., Schulze, D., . . . Daumke, O.
(2011). Crystal structure of nucleotide-free dynamin. Nature, 477(7366),
556-560. doi:10.1038/nature10369
Ferguson, S. M., & De Camilli, P. (2012). Dynamin, a membrane-remodelling GTPase.
Nat Rev Mol Cell Biol, 13(2), 75-88. doi:10.1038/nrm3266
Fujise, K., Okubo, M., Abe, T., Yamada, H., Nishino, I., Noguchi, S., . . . Takeda, T.
(2020). Mutant BIN1-Dynamin 2 complexes dysregulate membrane remodeling
in the pathogenesis of centronuclear myopathy. J Biol Chem.
doi:10.1074/jbc.RA120.015184
Pathogenic evaluation of congenital myopathy
Hohendahl, A., Roux, A., & Galli, V. (2016). Structural insights into the centronuclear
myopathy-associated functions of BIN1 and dynamin 2. J Struct Biol, 196(1),
37-47. doi:10.1016/j.jsb.2016.06.015
Jungbluth, H., & Gautel, M. (2014). Pathogenic mechanisms in centronuclear
myopathies. Front Aging Neurosci, 6, 339. doi:10.3389/fnagi.2014.00339
Jungbluth, H., Wallgren-Pettersson, C., & Laporte, J. (2008). Centronuclear
(myotubular) myopathy. Orphanet J Rare Dis, 3, 26.
doi:10.1186/1750-1172-3-26
Kenniston, J. A., & Lemmon, M. A. (2010). Dynamin GTPase regulation is altered by
PH domain mutations found in centronuclear myopathy patients. EMBO J,
29(18), 3054-3067. doi:10.1038/emboj.2010.187
Majczenko, K., Davidson, A. E., Camelo-Piragua, S., Agrawal, P. B., Manfready, R. A.,
Li, X., . . . Dowling, J. J. (2012). Dominant mutation of CCDC78 in a unique
congenital myopathy with prominent internal nuclei and atypical cores. Am J
Hum Genet, 91(2), 365-371. doi:10.1016/j.ajhg.2012.06.012
Marks, B., Stowell, M. H., Vallis, Y., Mills, I. G., Gibson, A., Hopkins, C. R., &
McMahon, H. T. (2001). GTPase activity of dynamin and resulting
conformation change are essential for endocytosis. Nature, 410(6825), 231-235.
doi:10.1038/35065645
Nishikawa, A., Mitsuhashi, S., Miyata, N., & Nishino, I. (2017). Targeted massively
parallel sequencing and histological assessment of skeletal muscles for the
molecular diagnosis of inherited muscle disorders. J Med Genet, 54(2), 104-110.
doi:10.1136/jmedgenet-2016-104073
Pathogenic evaluation of congenital myopathy
Okubo, M., Iida, A., Hayashi, S., Mori-Yoshimura, M., Oya, Y., Watanabe, A., . . .
Nishino, I. (2018). Three novel recessive DYSF mutations identified in three
patients with muscular dystrophy, limb-girdle, type 2B. J Neurol Sci, 395,
169-171. doi:10.1016/j.jns.2018.10.015
Ramachandran, R., Surka, M., Chappie, J. S., Fowler, D. M., Foss, T. R., Song, B. D.,
& Schmid, S. L. (2007). The dynamin middle domain is critical for
tetramerization and higher-order self-assembly. EMBO J, 26(2), 559-566.
doi:10.1038/sj.emboj.7601491
Reubold, T. F., Faelber, K., Plattner, N., Posor, Y., Ketel, K., Curth, U., . . . Eschenburg,
S. (2015). Crystal structure of the dynamin tetramer. Nature, 525(7569),
404-408. doi:10.1038/nature14880
Romero, N. B. (2010). Centronuclear myopathies: a widening concept. Neuromuscul
Disord, 20(4), 223-228. doi:10.1016/j.nmd.2010.01.014
Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V., Longair, M., Pietzsch, T., . . .
Cardona, A. (2012). Fiji: an open-source platform for biological-image analysis.
Nat Methods, 9(7), 676-682. doi:10.1038/nmeth.2019
Trochet, D., Prudhon, B., Beuvin, M., Peccate, C., Lorain, S., Julien, L., . . . Bitoun, M.
(2018). Allele-specific silencing therapy for Dynamin 2-related dominant
centronuclear myopathy. EMBO Mol Med, 10(2), 239-253.
doi:10.15252/emmm.201707988
Wang, L., Barylko, B., Byers, C., Ross, J. A., Jameson, D. M., & Albanesi, J. P. (2010).
Dynamin 2 mutants linked to centronuclear myopathies form abnormally stable
polymers. J Biol Chem, 285(30), 22753-22757. doi:10.1074/jbc.C110.130013
Pathogenic evaluation of congenital myopathy
Warnock, D. E., Hinshaw, J. E., & Schmid, S. L. (1996). Dynamin self-assembly
stimulates its GTPase activity. J Biol Chem, 271(37), 22310-22314. Retrieved
from https://www.ncbi.nlm.nih.gov/pubmed/8798389
Pathogenic evaluation of congenital myopathy
FIGURE LEGENDS
Figure 1. Gain of function features of mutant dynamin 2 in vitro and in cellulo. (A)
Quantitative analysis of the in vitro sedimentation assay. Relative amount of wild type
(WT) or mutant dynamin 2 (E368K, R369Q, R465W and S619L) either in the
supernatant (sup) or in the precipitate (ppt) with or without GTP (+ or -) are shown. (B)
GTPase activity of wild type (WT) and mutant dynamin 2 (E368K, R369Q, R465W and
S619L). Data are means ± SEM (n=3, N=3). (C) BIN1-mediated inhibition of dynamin 2
GTPase activity. Relative ratios of inhibited GTPase activities (∆GTPase activity) of wild
type (WT) and mutant dynamin 2 (E368K, R369Q, R465W and S619L) in the presence
of BIN1 (dynamin 2: BIN1 = 1:4 in molar ratio) are shown. Data are means ± SEM (n=3,
N=3). (D) Formation of aggregates by mutant dynamin 2. Localization of FLAG-tagged
wild type dynamin 2 (DNM2 WT-FLAG) or mutant dynamin 2 (DNM2 E368K-FLAG,
DNM2 R369Q-FLAG, DNM2 R465W-FLAG and DNM2 S619L-FLAG) in C2C12 cells
are shown. Scale bars are 10 �m. (E) TLS formation in the presence of wild type and
CNM mutant dynamin 2. Merged images of FLAG-tagged wild type (DNM2 WT-FLAG)
or mutant dynamin 2 (DNM2 E368K-FLAG, DNM2 R369Q-FLAG, DNM2
R465W-FLAG and DNM2 S619L-FLAG) (green) with BIN1-RFP (red) are shown.
Scale bars are 10 �m. (F) Enhanced aggregate formation by mutant dynamin 2. Size of
aggregates formed by either wild type or mutant dynamin 2 (shown in D) are measured
and their distribution is shown. Data are means ± SEM (n ≥ 891 aggregates in 10 cells).
(G) Increased number of aggregates by mutant dynamin 2. Average number of the
aggregates formed either by wild type or by mutant dynamin 2 per 100 μm2 of cell area
are shown. Data are means ± SEM (n ≥ 967 aggregates in ≥ 10 cells, N=3). (H)
Pathogenic evaluation of congenital myopathy
Quantification of the short TLS (0.5 ≤ 2 �m). Data are means ± SEM (n ≥ 936 TLS in ≥
10 cells, N=3).
Figure 2. TLS formation by novel CNM-associated dynamin 2 mutants. (A) Aggregate
formation by FLAG-tagged wild type dynamin 2 (DNM2 WT-FLAG) and five novel
mutant dynamin 2 (DNM2 G495R-FLAG, DNM2 V520G-FLAG, DNM2 G624V-FLAG,
DNM2 P294L-FLAG and DNM2 R724H-FLAG) in C2C12 cells. Scale bars are 10 �m.
(B) Defective TLS formation by novel mutant dynamin 2. Merged images of
FLAG-tagged wild type (DNM2 WT-FLAG) or novel mutant dynamin 2 (DNM2
G495R-FLAG, DNM2 V520G-FLAG, DNM2 G624V-FLAG, DNM2 P294L-FLAG and
DNM2 R724H-FLAG) (green) with BIN1-RFP (red) are shown. Scale bars are 10 �m.
(C) Distribution of the average diameter of aggregates formed by either wild type or
mutant dynamin 2. Data are means ± SEM (n ≥ 789 aggregates in 10 cells). (D) Average
number of aggregates formed by either wild type or mutant dynamin 2 per 100 μm2 of the
cell area. Data are means ± SEM (n ≥ 967 aggregates in ≥10 cells, N=3) (E)
Quantification of the short TLS (0.5 ≤ 2 �m) formed in the presence of wild type or novel
mutant dynamin 2. Data are means ± SEM (n ≥ 242 TLS in ≥ 6 cells, N ≥ 3).
Figure 3. Correlation analyses between TLS formation and clinical phenotypes. Scatter
plots for the presence of correlation between the short TLS formation and either disease
onset age (A), disease duration (B) or age at biopsy (C). r indicates correlation coefficient.
Non-pathogenic variants were shown with patient No in white square (P16 and P17). (D)
Relationship between age at biopsy and onset of the disease. Group A, B and C represent
severe phenotype (early onset and age at biopsy), slowly progressive phenotype (early
Pathogenic evaluation of congenital myopathy
onset and late biopsy) and late onset phenotype (late onset and age at biopsy),
respectively.
ppt
sup
2 t 2 8 °
os cos sos cs cs
Ayance esegi9V
eee
S88 3
(upupjowjow)
eee
a =
poe Ofed19
2
s
°
RI9Q RAGS —SBISL
368K
£ g s 8 .
I 192 204 (wate 5 $9) seinqny yoys yo sequinu
s 2 ¢
uurigos 20d
uogeBai66e jo soquinu eBesory
SS PES
ae38 °
sad (wirlz 5 $10) Seinqny Yous yo sequinu
Be Emme Bonar
(LU 99 Sed (wirtz s <9) saynqny poys jo saquinu
wri, 40d
Q uopebes66e jo soquinu oBesony
toaten (0.5< 2 ym) per col
3 @ 8
3 8
g
s
2
a
©) = =—s umber of short tut
°
3
$
2
a
2
3
number of short tubules (0.5<2um) per cell
8
°
° 20 40
age at biopsy (year-old)
$200
£
$150
=
4so0
i,
z
i
D
s
z .
3 .
z a
3 c
3 © GA95R, 5206, G624V
Arsssw
Mlexs8k, 23690, S619.
40 60
onset (year-old)
| 2021 | Imaging-based evaluation of pathogenicity by novel associated with centronuclear myopathy | 10.1101/2021.03.01.433478 | [
"Fujise Kenshiro",
"Okubo Mariko",
"Abe Tadashi",
"Yamada Hiroshi",
"Takei Kohji",
"Nishino Ichizo",
"Takeda Tetsuya",
"Noguchi Satoru"
] | null |
1
The branched chain aminotransferase IlvE promotes growth, stress resistance and
1
pathogenesis of Listeria monocytogenes
2
Karla D. Passalacquaa*#, Tianhui Zhoua*, Tracy A. Washingtonb, Basel H. Abuaitaa, Abraham
3
L. Sonensheinb, Mary X.D. O’Riordana#
4
5
aUniversity of Michigan Medical School, Department of Microbiology & Immunology, Ann
6
Arbor, MI
7
bTufts University School of Medicine, Department of Molecular Biology & Microbiology,
8
Boston, MA
9
10
Running title: Branched chain fatty acids in Listeria
11
12
#Address correspondence to: kpassal@umich.edu; oriordan@umich.edu
13
14
*Karla D. Passalacqua and Tianhui Zhou contributed equally to this work. Author order was
15
determined on the basis of seniority.
16
17
2
ABSTRACT
18
The bacterial plasma membrane is a key interface during pathogen-host interactions, and
19
membrane composition enhances resistance against host antimicrobial defenses. Branched chain
20
fatty acids (BCFAs) are the major plasma membrane component in the intracellular Gram-
21
positive pathogen Listeria monocytogenes (Lm) and BCFA metabolism is essential for Lm
22
growth and virulence. BCFA synthesis requires branched chain amino acids (BCAAs), and the
23
BCAA Isoleucine (Ile) is a necessary substrate for the predominant membrane anteiso-BCFAs
24
(ai-BCFAs) as well as an environmental signal for virulence regulation in Lm. In this study, we
25
explored how two proteins that metabolize or sense Ile contribute to Lm growth, BCFA
26
metabolism, and virulence. The IlvE aminotransferase incorporates Ile into ai-BCFAs, while
27
CodY is an Ile-sensing regulator that coordinates BCAA synthesis and virulence gene
28
expression. Analysis of deletion mutants lacking IlvE (ilvE) or CodY (codY) revealed a major
29
role for IlvE under nutrient restriction and stress conditions. Cultures of the ilvE mutant
30
contained proportionally less ai-BCFAs relative to wild type, while of the codY mutant had a
31
lower proportion of ai-BCFAs in stationary phase, despite containing more cell-associated Ile.
32
Both ilvE and codY mutants required exogenous Ile for optimal growth, but the ilvE mutant
33
had an absolute requirement for Valine and Leucine when Ile was absent. IlvE was also
34
necessary for resistance to membrane stress, cell-to-cell spread, infection of primary
35
macrophages, and virulence in mice. Our findings implicate IlvE as an integral aspect of Lm
36
stress resistance and emphasize the central importance of Ile in Lm growth and virulence.
37
38
3
INTRODUCTION
39
The bacterial plasma membrane is a key interface of pathogen-host interactions and an
40
important intrinsic barrier to host antimicrobial defenses. Situated just beneath and intimately
41
connected to the bacterial cell wall, the plasma membrane is a crucial structure of the bacterial
42
cell surface; thus, the interface between bacterium and host cell is of particular importance for
43
intracellular pathogens such as Listeria monocytogenes (Lm), the causative agent of listeriosis
44
(1-3). During its infectious cycle, Lm enters mammalian cells and traverses through a range of
45
cellular locales, each with distinct nutrient availability, redox state, and host antibacterial
46
mechanisms (4). Here, the bacterial membrane serves as an environmental sensor and a
47
defensive structure central to intracellular survival and replication (5). Therefore, exploring Lm
48
membrane dynamics is central for elucidating virulence strategies of this important pathogen.
49
As in many Gram-positive bacteria, including Staphylococcus aureus, the Lm plasma
50
membrane is predominantly composed of branched chain fatty acids (BCFAs), a structural
51
feature important for bacterial integrity against multiple stresses and during pathogenesis (6-14).
52
Odd numbered (C15, C17) anteiso-BCFAs (ai-BCFAs) are the most abundant form of BCFA in
53
the Lm membrane, and the ability of Lm to thrive in cold temperatures is due in large part to
54
high ai-BCFA content that enhances membrane fluidity (15-17). To optimize membrane fluidity
55
in different environments, Gram-positive bacteria alter the ratio of ai-BCFAs to iso-BCFAs,
56
where ai-BCFAs contribute to higher fluidity due to positioning of the terminal methyl groups on
57
the acyl chains (18). Because BCFA synthesis depends on the acquisition and/or biosynthesis of
58
branched chain amino acids (BCAAs: Isoleucine, Leucine and Valine [Ile, Leu, Val]) (7, 19),
59
membrane remodeling and BCAA metabolism are tightly linked. While Lm is fully capable of
60
synthesizing BCAAs de novo, exogenous BCAAs are required for optimal growth due in part to
61
4
high demand for BCFAs in the membrane and the activity of a ribosome-mediated attenuator that
62
limits BCAA synthesis (20, 21). During infection, Lm replicates inside host cells where BCAAs
63
and other nutrients are limited and may be actively withheld from bacteria by host defense
64
mechanisms (22, 23). Therefore, the ability of Lm to acquire host BCAAs and to make de novo
65
BCAAs, especially Ile, to generate membranes with high BCFA levels is critical to Lm
66
pathogenesis.
67
Branched chain amino acid aminotransferase (BCAT) enzymes initiate bacterial BCFA
68
synthesis by converting BCAAs into branched chain -keto acids. Downstream of BCAT,
69
branched chain -keto dehydrogenase enzymes (BKD) produce acyl coenzyme A (CoA)
70
molecules that are the primers for fatty acid synthesis (Fig. 1A) (9). While BKD is essential for
71
BCFA metabolism and protection from host immune defenses such as antimicrobial peptides
72
(12, 13), the Lm BCAT IlvE is required for resistance to the compound trans-cinnamaldehyde, a
73
small molecule with anti-microbial properties (14, 24). Additionally, the transcriptional regulator
74
CodY, which senses BCAA and GTP levels, plays a major role in coordinating BCAA
75
metabolism with virulence gene expression (21, 25-28). Importantly, when Ile levels are high,
76
CodY inhibits de novo BCAA synthesis, and when Ile levels are low, this inhibition is relieved,
77
allowing the bacteria to synthesize vital BCAAs (7). Thus, the ability of Lm to sense and
78
regulate BCAA levels, particularly Ile, and to implement BCFA remodeling is an important
79
attribute for adaptation to changing environments, especially in stress conditions as found in the
80
mammalian host.
81
Due to the central requirement for Ile in promoting membrane integrity through
82
generation of ai-BCFAs and for engaging CodY regulatory activity, we hypothesized that
83
proteins involved in Ile metabolism are central to the ability of Lm to cause disease. Therefore,
84
5
we used a genetic approach to explore how the BCAT IlvE and the regulator CodY contribute to
85
membrane dynamics, growth and pathogenesis of Lm. Here we show that deficiency of either
86
IlvE or CodY can alter membrane fatty acid content, but bacteria lacking IlvE are very
87
susceptible to membrane stress and nutrient limitation and are less fit in in vitro and in vivo
88
infection models.
89
90
RESULTS
91
92
Membrane anteiso-BCFA generation requires IlvE and relies on CodY for homeostasis in
93
stationary phase in nutrient restricted medium
94
Isoleucine (Ile) is an essential metabolite for protein translation and for synthesis of the
95
high ai-BCFA membrane content of Lm (Fig. 1A), and the aminotransferase IlvE is predicted to
96
be the first enzyme that commits Ile into the biosynthetic pathway for odd-numbered (C15, C17)
97
ai-BCFAs. Low availability of Ile in the intracellular environment during infection is thought to
98
act as a signal for Lm to coordinate metabolism and virulence, mainly through the Ile-sensing
99
transcriptional regulator CodY (21, 26, 27). To characterize dynamics of Ile usage in BCFA
100
biosynthesis and virulence, we assessed deletion mutant strains lacking IlvE or CodY (ilvE and
101
codY mutants) (Table S1 and Methods).
102
Previously, an ilvE transposon-generated null mutant was shown to have extremely low
103
levels of ai-BCFAs when grown in rich, undefined BHI medium (14). Therefore, we predicted
104
that the ilvE strain created for this study would have substantially lower levels of ai-BCFA
105
when grown in a nutrient-limited medium (Fig. 1A). Additionally, we predicted that the codY
106
mutant would generate ai-BCFA levels equivalent to WT, since eliminating CodY inhibition of
107
6
de novo BCAA synthesis should increase bacterial BCAA levels, resulting in ample building
108
blocks for ai-BCFAs. Note that recent RNAseq analysis showed that the expression of the Ile,
109
Leu, Val-production operon increased substantially in the codY mutant in both rich medium
110
(BHI) and in nutrient-limited medium (29), but not when BCAAs were extremely limited;
111
moreover, the codY null mutation had no significant impact on ilvE transcription in any media
112
tested (T. A. Washington, A. L. Sonenshein and B. R. Belitsky, manuscript in preparation)
113
despite the fact that CodY has a relatively strong binding site upstream of ilvE (30), which could
114
also be a regulatory binding site for the locus upstream of ilvE. Therefore, to test the role of IlvE
115
and CodY in fatty acid metabolism, we measured total fatty acid content in WT, ilvE,
116
ilvE::ilvE+ (ilvE complemented) and codY strains grown to mid-logarithmic and stationary
117
phase in Listeria defined medium (LDM - contains seven amino acids including BCAAs) (29), a
118
nutrient-limited medium (Fig. 1B-C, Tables 1-2, Tables S3-S4).
119
The ilvE strain contained significantly lower proportions of ai-BCFAs in the total fatty
120
pool compared to WT (Fig. 1B-C). During both culture phases, WT cells had greater than 80%
121
ai-BCFAs, while ilvE cells had 30% ai-BCFAs in logarithmic phase and 21% ai-BCFAs in
122
stationary phase. Whereas WT cells had extremely low levels of iso-BCFAs (0.9 – 12%), the
123
ilvE strain included substantial levels odd-numbered iso-C15 and iso-C17 fatty acids (37%) and
124
even-numbered iso-C14 and iso-C16 (26% to 33%) (Tables 1 and 2). This indicates that in the
125
absence of IlvE, Lm must incorporate the other BCAAs (Leu and Val) into BCFAs. The
126
complemented strain ilvE::ilvE+ showed an almost identical fatty acid profile to WT in both
127
phases. Interestingly, the codY mutant differed markedly from WT in stationary phase. Here,
128
ai-BCFAs made up 63% of the fatty acid profile, and even-numbered iso-BCFAs increased to
129
22% (ten-fold higher than WT). These results suggest that CodY may contribute to membrane
130
7
BCFA remodeling when salvageable nutrients are depleted, and may repress genes involved in
131
iso-BCFA synthesis during stationary phase.
132
We conclude that IlvE is a major driver for ai-BCFA generation during Lm growth in
133
nutrient-limited medium; however, bacteria lacking IlvE were still able to generate more than
134
20% ai-BCFAs, suggesting the presence of another aminotransferase that is able to incorporate
135
Ile into ai-BCFA. Additionally, we conclude that CodY is involved in BCFA homeostasis during
136
stationary phase in LDM.
137
138
Bacteria lacking CodY harbor higher levels of BCAA compared to WT
139
CodY is a sensor of BCAAs in Gram-positive bacteria, particularly Ile, and controls
140
BCAA synthesis when these important metabolites are at low levels (31, 32). We initially
141
hypothesized that bacteria lacking CodY would constitutively synthesize BCAAs in addition to
142
acquiring exogenous BCAAs, and therefore should be well positioned to generate sufficient
143
levels of ai-BCFAs regardless of growth phase. However, the observation that the codY mutant
144
had lower levels of ai-BCFAs in stationary phase in LDM (Fig. 1C) prompted us to directly
145
measure levels of cell-associated BCAAs during growth in LDM. We therefore grew WT,
146
codY, ilvE and ilvE::ilvE+ strains in LDM to mid-logarithmic and stationary phase, removed
147
the extracellular medium, and assessed cell-associated BCAA content by mass spectrometry
148
(Fig. 1D-E). Unsurprisingly, codY lysates contained higher levels of all three BCAAs relative
149
to WT during logarithmic growth, with Ile being the highest (~2.5-fold higher relative to WT),
150
confirming the role of CodY for BCAA synthesis during nutrient-restriction. Notably, we
151
observed approximately three-fold more Ile in stationary phase cultures for the codY mutant
152
relative to WT, but similar levels of Leu and Val. Thus, although the codY strain in stationary
153
8
phase contains more Ile available for ai-BCFA compared to WT, this strain does not match WT
154
levels of Ile incorporation into ai-BCFAs. These data suggest that CodY may play a role in
155
membrane ai-BCFA homeostasis during stationary phase through an as yet undefined
156
mechanism.
157
Since bacteria lacking IlvE showed a severe reduction in ai-BCFA content in LDM, we
158
hypothesized that the ilvE mutant would harbor higher levels of cell-associated Ile during all
159
growth phases in LDM compared to WT due to the lack of incorporation of this amino acid into
160
ai-BCFAs. However, we observed similar levels of Ile in the ilvE mutant and in WT during
161
logarithmic growth, and wide variability of cell-associated Ile in the ilvE strain during
162
stationary phase (Fig. 1D-E). Also, Val was approximately half the level in the ilvE mutant
163
compared to WT in logarithmic and stationary phase (Fig. 1D-E), while the Leu level was less
164
than half of WT only in stationary phase. These data suggest that when IlvE is lacking, Lm uses
165
Val and Leu for BCFA metabolism.
166
167
Growth in BCAA-limiting conditions requires branched-chain aminotransferase IlvE
168
While fatty acid analysis represents relative levels of lipid species in a population of
169
cells, these data do not reveal differences in growth rate between strains. Therefore, we
170
examined the contributions of IlvE and CodY to bacterial growth in nutrient replete Brain-Heart
171
Infusion medium (BHI) and in nutrient-limited LDM. All growth experiments were initiated
172
using bacteria grown to mid-logarithmic phase in LDM. We hypothesized that because the
173
absence of CodY normally contributes to increased BCAA biosynthesis during Ile limitation
174
(26), codY bacteria would grow as well as, or better than, WT bacteria in LDM. We also
175
9
hypothesized that growth of the ilvE mutant would be slower than WT in nutrient-limited
176
medium due to its severe reduction in ai-BCFAs, a major membrane component for Lm.
177
In BHI, both the ilvE and the codY mutants grew equivalently to WT, showing that
178
IlvE and CodY are not essential for Lm growth in a nutrient rich environment (Fig. 2A and Table
179
3). In LDM containing all three BCAAs at 100 g/mL, the ilvE strain grew slightly more
180
slowly than WT, whereas the codY strain grew the same as, or slightly better than, WT (Fig.
181
2B). Although the ilvE culture reached the same maximum density as WT in LDM, its doubling
182
time during logarithmic growth was about 1.7-fold longer than WT (Table 3). These data reveal
183
that ai-BCFA synthesis through IlvE contributes to bacterial growth rate when nutrients are
184
limited. In LDM, the ilvE::ilvE+ complemented strain also grew more slowly than WT, despite
185
the fact that it was able to generate BCFA profiles similar to WT in this medium (Fig. 1B-C). We
186
therefore asked whether the ilvE gene is expressed at WT levels in the complemented strain.
187
Indeed, RT-qPCR of ilvE expression revealed lower transcript levels of this gene in the
188
complemented strain (Fig. 2C), particularly during exponential growth.
189
Fatty acid distributions (Fig. 1B-C) suggested that mutants lacking IlvE or CodY use Val
190
and Leu for synthesis of iso-BCFAs at higher levels than WT. Therefore, we asked how ilvE
191
and codY strains would grow when one or all three of the BCAAs are lacking in the growth
192
medium, despite having the ability to synthesize all three BCAAs de novo. In LDM lacking all
193
three BCAAs, all four strains grew poorly, with WT and ilvE::ilvE+ reaching the highest
194
optical density at 600 nm (OD600) of ~0.4, compared to all strains reading ~0.6 in LDM when
195
all BCAAs were present (Fig. 2B versus Fig. 3A). Additionally, the ilvE mutant exhibited large
196
variability when no BCAAs were supplied, while codY grew the poorest (Fig. 2B versus Fig.
197
3A). The fact that the codY mutant grew so poorly in medium with no exogenous BCAAs was
198
10
surprising considering that this mutant has no restriction on de novo synthesis of BCAAs. These
199
data support the semi-auxotrophic nature of Lm for BCAAs, highlighting the importance of
200
exogenous BCAAs for optimal growth and revealing a key role for IlvE and CodY when all
201
exogenous BCAAs are unavailable.
202
While IlvE is needed for enzymatic incorporation of Ile into BCFAs, CodY specifically
203
senses and binds cellular Ile (33, 34). Due to their specific relationships with Ile, we then asked
204
how ilvE and codY mutants would grow when exogenous Ile is lacking but when Val and Leu
205
are present. Interestingly, both ilvE and codY strains were similarly attenuated when only Ile
206
was lacking, reaching a lower maximum density compared to WT and ilvE::ilvE+ strains (Fig.
207
3B). Again, this was unexpected for the codY mutant, since we predicted that the codY strain
208
would have no growth defect in the absence of Ile due to its higher cell associated Ile
209
concentrations (Fig. 1D-E). These results reveal a complex role for CodY in Ile sensing and
210
BCAA homeostasis. We conclude that both IlvE and CodY are required for optimal bacterial
211
growth when exogenous Ile is absent.
212
When either all three BCAAs or only Ile were absent in the growth medium (Fig. 3A-B),
213
the ilvE and codY mutants were attenuated for growth to a similar degree. However, in media
214
containing exogenous Ile but lacking either of the other two BCAAs (Leu or Val), the two
215
mutants revealed unique growth phenotypes (Fig. 3C-D). In LDM containing Ile and one other
216
BCAA (Val or Leu), the codY mutant grew more robustly than WT, suggesting a dominant role
217
for Ile in Lm growth when CodY regulation is lacking. But the ilvE mutant showed strict
218
requirements for Leu and Val in the presence of Ile. When only Leu was absent (ie, Ile and Val
219
present), the ilvE mutant grew as it did in normal LDM (with all BCAAs, Fig. 2B) for about 12
220
hours, but reached stationary phase early and then had a decrease in OD600 (Fig. 3C). When
221
11
only Val was absent (ie, Ile and Leu present), the ilvE strain was entirely unable to grow (Fig.
222
3D), revealing an absolute requirement for Val when exogenous Ile is not incorporated into ai-
223
BCFAs by IlvE. The IlvE complemented strain was able to eventually reach a maximum density
224
in stationary phase similar to that of WT in all of these conditions (Fig. 3A-D), albeit at a slightly
225
slower rate. Thus, when IlvE is not present, Lm has an increased dependence on Val and Leu for
226
growth. These data show that while CodY is tightly linked to Ile sensing and homeostasis, IlvE
227
activity plays a key role in cellular homeostasis when any of the individual BCAAs are lacking
228
exogenously. Collectively, these growth trends indicate that BCAA levels are controlled at
229
multiple levels in Lm.
230
231
Listeria lacking IlvE exhibit decreased intracellular replication in macrophages and
232
reduced cell-to-cell spread
233
Having established that IlvE and CodY play a role in generating membrane ai-BCFAs
234
and in promoting optimal growth in BCAA-limited environments, we asked whether these
235
proteins specifically contribute to Lm pathogenesis. We hypothesized that the ilvE mutant
236
would be less efficient at intracellular growth in a cell culture infection model due to its
237
relatively slow growth during nutrient restriction. Previously, the codY mutant strain has shown
238
different behaviors in various in vitro macrophage infections models (25, 28). Since the codY
239
mutant in this study grew robustly in nutrient-limited LDM (Fig. 2B), we predicted that it would
240
grow similarly to WT in primary macrophages. We also considered that the codY mutant would
241
be deficient in cell-to-cell spread given its stationary phase reduction of ai-BCFAs.
242
We infected primary bone marrow-derived murine macrophages (BMDM) with Lm
243
strains prepared from mid-log phase LDM cultures and measured viable intracellular bacteria at
244
12
0, 4 and 8 hours post infection. At 4 and 8 h post-infection, intracellular growth of the ilvE
245
mutant was at least 1 log lower than WT (Fig. 4A). However, the ilvE strain showed a growth
246
rate increase after 4 h, suggesting that this strain may be able to adapt to the intracellular
247
environment. The ilvE::ilvE+ strain showed an intermediate phenotype, where intracellular
248
growth was less than that of WT but greater than that of the ilvE mutant. WT and codY strains
249
replicated within primary BMDM equivalently. We conclude that IlvE is required for optimal
250
growth in the nutrient-limited environment of macrophages, while CodY is not essential for
251
adaptation to intracellular growth within this cell type.
252
We then infected L929 cells with Lm strains prepared from mid-logarithmic LDM
253
cultures to assess the requirements for IlvE and CodY during multiple stages of intracellular
254
infection as measured by cell-to-cell spread (Fig. 4B-C). Plaques formed from infection with the
255
ilvE mutant were approximately 66% the size of WT-infected plaques (Fig. 4C). The
256
complemented ilvE::ilvE+ strain had a partially rescued plaque phenotype. We also observed
257
that plaques formed by the codY mutant were not significantly different from those of WT (Fig.
258
4C). Taken together, these data demonstrate that IlvE is a critical component for Lm intracellular
259
growth and cell-to-cell spread.
260
261
IlvE enhances bacterial survival in response to exogenous membrane stress
262
Membrane BCFA content underlies Lm resistance to various cell stresses such as pH,
263
small molecules, low temperature, and host-specific antimicrobial mechanisms (8, 10, 12, 13, 16,
264
17, 35). As a foodborne pathogen, Lm must survive the acidic stomach environment and resist
265
damage from host molecules such as bile. To investigate the role of Ile-dependent BCFA
266
metabolism in protecting Lm membrane integrity, we tested the ability of ilvE and codY
267
13
mutants to survive in the presence of membrane disrupting bile salts. We used a bile salt mixture
268
of cholic acid and deoxycholic acid, which are similar to the emulsifying bile acids in the
269
mammalian GI tract. We hypothesized that Lm lacking IlvE would be more susceptible to bile
270
salt stress than WT strains with a full complement of ai-BCFAs. Mid-logarithmic phase bacteria
271
grown in LDM were exposed to 0, 1, 2 and 4 mg/mL bile salts at 37C for 30 min and measured
272
by counting CFU (Fig. 5A). WT Lm showed decreasing viability with increasing bile salt
273
concentration, with a reduction in viability of almost 2 logs from 0 to 4 mg/mL. The ilvE
274
mutant strain showed a consistent 1-log decrease in viability compared to WT at each
275
concentration. The complemented strain ilvE::ilvE+ was slightly less viable at 1 mg/mL, but
276
was similar to WT at 2 and 4 mg/mL. Lastly, the codY mutant showed susceptibility to bile salt
277
stress similar to that of WT. We therefore conclude that IlvE promotes resilience against
278
membrane stress, likely through its role in populating the Lm membrane with ai-BCFAs.
279
280
IlvE is required for optimal infection of C57BL/6 mice
281
While in vitro infections can shed light on the intracellular growth capabilities of Lm,
282
they do not illuminate the more complex physiological dynamics of an animal infection. We
283
hypothesized that IlvE and CodY would contribute to pathogenesis in a mouse model of
284
infection, and that the IlvE would have more of an impact due to its constitutive role in
285
membrane fatty acid synthesis. We used a competitive index (CI) assay to measure the fitness of
286
Lm strains in C56BL/6 mice (36). Briefly, we injected mice intraperitoneally with a WT Lm
287
strain that is resistant to erythromycin (WT-ermr) combined with a mutant strain (test strain-
288
erms) in a 1:1 mixture (WT-ermr : test strain-erms). After 48 h, spleens and livers were removed
289
and bacteria plated on LB-agar with or without antibiotic to discern resistant (WTR) versus
290
14
sensitive (test strainS) bacteria and calculated the CI. The lower the CI, the less “competitive” the
291
test strain was compared to WT during infection.
292
In both spleen and liver (Fig. 5B-C), substantially fewer ilvE bacteria were recovered.
293
The mean CI for the ilvE strain in both organs was less than 0.2, indicating severe attenuation
294
in mice. Although the IlvE complemented strain grew better in mice than the deletion mutant, it
295
was recovered at lower levels than WT, suggesting that robust expression of ilvE is necessary for
296
optimal survival in a whole animal. Lastly, while bacteria lacking CodY showed a CI of ~0.5 in
297
mouse spleen, a CI of 0.1 in liver suggests that the liver environment is a more restrictive growth
298
milieu for the codY mutant. Overall, these data underline a major role for ai-BCFA metabolism
299
through IlvE for all aspects of Lm growth and virulence, with CodY playing a major role mainly
300
during Ile restriction and severe nutrient restriction.
301
302
DISCUSSION
303
304
The plasma membrane of Listeria monocytogenes (Lm) is a major structure of the
305
bacterial cell surface and a key interface with host cells (3). Understanding how Lm assembles
306
and remodels membranes to thrive within the host is key to our understanding of this important
307
pathogen. In this study, we explored how two Isoleucine (Ile) responsive proteins, the
308
aminotransferase IlvE and the regulator CodY, contribute to growth, plasma membrane
309
composition, and virulence of Lm. Our findings reveal a crucial role for IlvE in generation of
310
membrane ai-BCFAs, robust growth during nutrient limitation, protection from membrane stress,
311
and virulence in cell culture and in mice. Additionally, our work shows that CodY is involved in
312
modulating membrane ai-BCFA content during stationary phase, and that exogenous Ile is
313
15
required for bacterial growth when CodY is lacking. However, we observed that CodY is
314
relevant in the nutrient environment of the liver, but contributes less to bacterial fitness in the
315
spleen where Lm primarily replicates in macrophages. Collectively, our findings point to a
316
complex role for Ile usage through IlvE in promoting ai-BCFA membrane composition and also
317
highlight an important relationship between BCAA and ai-BCFA metabolic pathways for Lm
318
pathogenesis.
319
Anteiso-BCFAs are the major component of the Lm plasma membrane, and the
320
aminotransferase IlvE incorporates Ile into ai-BCFA biosynthesis (Fig. 1) (15, 18). Our main
321
finding that IlvE is a crucial element of Lm biology and virulence is supported first by the
322
observation that bacteria lacking this enzyme (ilvE) are severely restricted for growth under
323
multiple conditions of nutrient limitation. Since Lm is an intracellular pathogen, and the
324
intracellular environment is a nutrient-restricted medium (23), Lm must have strategies for
325
acquiring or synthesizing critical metabolites, such as BCAAs, during infection (2, 23). Notably,
326
IlvE was not required for optimal growth in rich-undefined medium, showing that Ile
327
incorporation into ai-BCFAs is not necessary when exogenous nutrients are in great abundance.
328
Rather, IlvE was critically needed for axenic growth during BCAA limitation, in particular when
329
only exogenous Valine or Leucine was unavailable, underscoring the central importance of Ile
330
for membrane metabolism.
331
Our results also highlight the complex nature of BCAA metabolism in Lm, which is
332
somewhat curious, since these bacteria are able to synthesize BCAA endogenously but still
333
require exogenous BCAAs for optimal growth (22, 26). Lm expresses BCAA biosynthetic genes
334
during infection (26), indicating BCAA limitation within cells. Recent investigation into this
335
phenomenon has revealed that while the Ile-binding regulatory protein CodY inhibits BCAA
336
16
synthesis when Ile is abundant, the bacteria also limit BCAA synthesis through Rli60 even when
337
CodY inhibition is relieved during Ile restriction, as within the host (21). These opposing
338
processes allow the bacteria to fine-tune Ile levels in order to satisfy BCAA requirements for
339
growth while also allowing virulence gene expression (21). We showed that bacteria lacking
340
CodY or IlvE were severely attenuated for growth when Ile was not available in the medium,
341
highlighting a central role for exogenous Ile during growth. But bacteria lacking CodY harbored
342
more cell-associated BCAAs during growth in nutrient-limited LDM, strongly suggesting a
343
constitutive increase in endogenous BCAA synthesis when CodY inhibition is completely
344
lacking. Thus, the codY mutant’s poor growth in the absence of Ile was unexpected, since these
345
bacteria have a greater Ile pool most likely due to de novo synthesis. Moreover, the highly robust
346
growth of the codY mutant when exogenous Ile and one other BCAA were available
347
underscores a vital role for exogenous Ile in the fine-tuning of BCAA metabolism through
348
CodY, perhaps through involvement in controlling BCAA transport as in Bacillus subtilis (27).
349
Collectively, these findings suggest that within the nutrient-restricted intracellular environment,
350
Lm must be able to access sufficient Ile for ai-BCFA synthesis through IlvE activity, but must
351
also sense relative Ile limitation such that CodY metabolic inhibition is relieved to support
352
endogenous BCAA generation for optimal growth.
353
Another line of evidence pointing to the critical nature of IlvE in Lm biology is its major
354
role in supporting production of resilient membranes during nutrient restriction at biological
355
temperatures (37C). The importance of ai-BCFA membrane content for resistance to cold has
356
been well established for Lm, and indeed Lm is able to modulate the percentage of ai-BCFAs in
357
response to temperature, salinity, and pH (8, 15-17). However, the Lm membrane is always
358
predominantly made up of Ile-primed odd-numbered ai-BCFAs, emphasizing the central
359
17
importance of the Ile-to-ai-BCFA biosynthetic pathway for this pathogen. Our demonstration
360
that bacteria lacking IlvE have greatly reduced ai-BCFA content and are sensitive to bile salts
361
directly implicates Ile usage by IlvE as a major player in synthesizing resilient bacterial
362
membranes. Within host cells, Lm is subjected to a variety of membrane-targeting host defenses
363
such as antimicrobial peptides, and ai-BCFAs have been shown to be important for resistance to
364
these mechanisms when the enzyme branched-chain -keto acid dehydrogenase (BKD),
365
downstream of IlvE, is lacking (13). While those stresses are experienced by Lm inside host
366
cells, Lm is a foodborne pathogen, and so must also survive the low pH of the stomach and the
367
high concentration of bile acids in the small intestine (37, 38). A lifestyle-specific evolution of
368
ai-BCFA metabolism is evident in the Gram-positive dental pathogen Streptococcus mutans,
369
which requires IlvE for acid tolerance, such as might be experienced in the oral cavity (39).
370
Thus, the contribution of IlvE for bile salt resistance in Lm reveals that a major need for Ile and
371
ai-BCFAs evolved as a fundamental physiological feature for surviving stress within the diverse
372
environments that this pathogen experiences. Further exploration into the mechanism of bile salt
373
resistance may reveal membrane structural features and bile salt transport mechanisms as playing
374
key roles.
375
The central importance for IlvE was also revealed by the severe attenuation of the ilvE
376
strain in cell culture and in a mouse model of listeriosis. As mentioned previously, the
377
intracellular milieu is nutrient-restricted and a site of antimicrobial assault. Thus, the decrease in
378
intracellular growth of Lm lacking IlvE after four hours of macrophage infection is likely due to
379
enhanced microbial killing, as was seen in the BKD mutant (13). However, it should be noted
380
that the ilvE mutant established growth macrophages between 4 and 8 hours, which may
381
indicate a regulatory stress response when Ile incorporation into ai-BCFAs is compromised. This
382
18
observation, combined with the fact that the ilvE mutant still had 20-30% ai-BCFAs during
383
growth in LDM, hints at the presence of another transaminase that can use Ile for ai-BCFA
384
synthesis. Different from what we observed, an Lm mutant lacking ilvE in a different parental
385
strain background was almost entirely lacking in ai-BCFAs when grown in rich medium, well
386
under 10% of fatty acid content (14), and this could mean that Lm has several regulatory
387
strategies for membrane homeostasis depending on the nutritional content of the growth medium.
388
However, the amount of ai-BCFAs that we observed in the absence of IlvE was not sufficient for
389
full virulence in a whole animal, highlighting the necessity of IlvE mediated ai-BCFA synthesis
390
for membranes during infection.
391
Lastly, our results also shed light on the complexity of CodY regulation, which in
392
addition to BCAA metabolism, is also known to be involved in nitrogen and carbon assimilation
393
and regulation of Lm virulence gene expression (21, 25, 27). Previous studies of codY mutants
394
in in vitro macrophage models have shown different results, where CodY was not required for
395
growth in a transformed macrophage line (25), but was required for optimal growth within
396
primary macrophages (26). In our study, we did not observe a defect in growth within primary
397
macrophages for the codY mutant. But note that while the codY mutant had an identical fatty
398
acid profile to WT during logarithmic growth, it showed a significant reduction in ai-BCFAs
399
during stationary phase: and for our macrophage experiments, we used codY cultures that were
400
prepared at mid-logarithmic phase grown in nutrient-limited medium. This parameter may have
401
poised the bacteria to be more resistant to macrophage killing during the brief, 8-hour duration of
402
the experiment, and this possibility is currently being explored. Regardless, our data are the first
403
to describe a role for CodY in Lm pathogenesis in a whole animal model, where the codY
404
mutant was attenuated predominantly in the mouse liver.
405
19
In this study, we determined that the branched-chain amino acid transaminase IlvE plays
406
a central role in the membrane dynamics of L. monocytogenes and is necessary for robust
407
replication during intracellular infection in vitro and in vivo. Collectively, our findings highlight
408
an intricate connection between BCAA and BCFA metabolism, and further support a model
409
where Ile is a key metabolite for bacterial growth and virulence, in particular through the activity
410
CodY. Future investigation into how Lm remodels its membrane during interactions with the
411
host will expand our understanding of how pathogens use this defining cellular structure to
412
enhance infection.
413
414
20
FIGURE LEGENDS
415
416
Figure 1. Changes in Fatty Acid and BCAA content in Lm lacking IlvE or CodY. (A)
417
Simplified overview of branched chain fatty acid (BCFA) biosynthesis in Gram-positive bacteria
418
(based on detailed diagram in (9)) showing pathways that incorporate branched chain amino
419
acids (BCAAs: Ile, Leu & Val). Red X represents points in pathways where deletion mutants
420
were used in this study. Colored arrows indicate pathways of individual BCAAs that are
421
incorporated into final BCFA isoforms (18). Purple text = enzyme names. (B and C) Graphs
422
represent the relative amounts of the major fatty acids as a percentage of total fatty acids
423
contained in Lm cultures of WT, ilvE, ilvE::ilvE+ and codY strains grown in nutrient
424
limiting medium (LDM) to (B) mid-logarithmic and (C) stationary phase. Graphs represent
425
combined data from three independent experiments. Graphs shown here and data in Tables 2 and
426
3 are the combined quantities of odd numbered (C15 and C17) or even-numbered (C14 and C16)
427
BCFAs. Individual numbered species (e.g., ai-C15 only) and all other fatty acids are in
428
Supplemental Tables S3 and S4. (D and E). Cultures of WT, ilvE, ilvE::ilvE and codY
429
strains grown in LDM to (D) mid-logarithmic and (E) stationary phase were analyzed by mass
430
spectrometry. Concentrations of BCAAs were normalized to total protein content and are shown
431
as ratios relative to WT. Error bars show the range of fold difference compiled from 2
432
independent experiments.
433
434
Figure 2. Growth of ilvE and codY mutants in rich and nutrient-limited medium.
435
Bacterial growth of WT (circles), ilvE (triangles), ilvE::ilvE+ (inverted triangles), and codY
436
(squares), was analyzed on a Bioscreen instrument. Samples were inoculated from recovered
437
21
frozen cultures that had been prepared in LDM to mid-logarithmic phase. Optical Density at 600
438
nm (OD600) was measured for 24 hours at 37C with shaking. Experiments include growth in
439
(A) Rich medium = Brain Heart Infusion (BHI) and (B) LDM containing amino acids at 100
440
g/mL. Data are compiled from three independent experiments with three technical replicates
441
per experiment. Each point is the mean with error bars representing the Standard Deviation. (C)
442
RT-qPCR analysis of ilvE expression in Lm grown in LDM to logarithmic (left) and stationary
443
(right) phase.
444
445
Figure 3. Growth of Lm in LDM with variable exogenous BCAAs. Bacterial growth of WT
446
(circles), ilvE (triangles), ilvE::ilvE+ (inverted triangles), and codY (squares), performed as
447
in Figure 2, but in LDM containing (A) no BCAAs, (B) no Ile (Val & Leu only), (C) no Leu (Ile
448
& Val only), and (D) no Val (Ile & Leu only). Data are compiled from three independent
449
experiments with three technical replicates per experiment. Each point is the mean with error
450
bars representing standard deviation.
451
452
Figure 4. IlvE is required for optimal growth in macrophages and for cell-to-cell spread in
453
cell culture. (A) Total CFU from survival assays of Lm infection of Bone Marrow Derived
454
Macrophages (BMDM) assessed at 0.5, 4 and 8h post-infection. Data are compiled from three
455
independent experiments showing mean and standard deviation. MOI = 1. (B-C) Plaque assay of
456
Lm grown in L9 fibroblasts. (B) Representative image of plaques formed by WT & ilvE
457
bacteria after 48h growth. (C) Average plaque diameters from experiments that included WT,
458
ilvE and ilvE::ilvE+ (left) or WT and codY (right). Numbers below graphs are the mean
459
22
plaque diameter with standard deviation compiled from three independent experiments. Two-
460
tailed t-test comparing mutants to WT, ****P<0.0001; ns = not significant.
461
462
Figure 5. IlvE is required for resistance to membrane stress in response to bile salts and for
463
survival in a mouse model of listeriosis. (A) Log-phase bacteria grown in LDM were added to
464
PBS with 0, 1, 2 and 4 mg/mL Bile Salts (Cholic acid-Deoxycholic acid sodium salt mixture)
465
and incubated at 37C for 30 minutes. Input for all samples was ~107 CFU/mL. Data are
466
compiled from three independent experiments. One-way ANOVA (non-parametric) with Dunn’s
467
multiple comparisons post-test comparing mutant strains to WT. ns = not significant; *P<0.05;
468
***P<0.001; ****P<0.0001. (B and C) Female C56BL/6 mice were infected with a 1:1 mixture
469
of erythromycin-sensitive test strains and erythromycin-resistant WT strain via intraperitoneal
470
injection. After 48h infection, (B) spleens and (D) livers were harvested and assessed for viable
471
CFU and competitive index (CI) was calculated as the ratio of Sensitive/Resistant CFU. Data
472
represent two independent experiments with total n=7 mice for all strains except WT, which was
473
n=8. LOD = limit of detection.
474
475
23
TABLES
476
477
Table 1. Fatty Acid Content of L. monocytogenes during Logarithmic Growth in LDM
478
Percent of Total Fatty Acid Content – Logarithmic Growth
Mean Percent (SD)
WT
ilvE
ilvE::ilvE+
codY
Other
0.95 (0.20)
7.41 (3.58)
6.72 (9.05)
1.91 (0.54)
anteisoC15:C17
88.56 (0.30)
****29.67 (13.54)
84.67 (8.76)
86.41 (1.44)
isoC15:C17
9.60 (0.30)
****37.06 (6.15)
6.85 (1.80)
7.43 (1.36)
isoC14:C16
0.89 (0.11)
****25.86 (9.50)
1.76 (0.61)
4.24 (0.86)
Two-Way ANOVA using Dunnett’s Multiple Comparisons Test (compare rows within columns)
479
compared to Wild Type (WT). ****P<0.0001.
480
481
24
Table 2. Fatty Acid Content of L. monocytogenes during Stationary Phase in LDM
482
Percent of Total Fatty Acid Content – Stationary Phase
Mean Percent (SD)
WT
ilvE
ilvE::ilvE
codY
Other
1.65 (0.44)
*8.86 (1.85)
1.53 (0.21)
4.84 (3.27)
anteisoC15:C17
83.87 (2.76)
****21.06 (4.14)
85.55 (3.53)
****62.75 (8.89)
isoC15:C17
12.26 (2.09)
****36.87 (1.23)
8.32 (2.23)
10.44 (2.69)
isoC14:C16
2.22 (0.66)
****33.22 (5.51)
4.60 (3.75)
****21.97 (3.03)
Two-Way ANOVA using Dunnett’s Multiple Comparisons Test (compare rows within columns)
483
compared to Wild Type (WT). *P<0.05; ****P<0.0001
484
Data are combined from n = 3 experiments
485
486
Table 3. 1Doubling times of L. monocytogenes strains in small volume growth analysis in
487
rich (BHI) and nutrient-limiting (LDM) medium.
488
WT
ilvE
ilvE::ilvE+
codY
Mean (SD) in minutes
BHI
65.9 (5.9)
71.0 (2.3)
59.5 (18.1)
63.8 (3.3)
LDM
118.5 (5.9)
196.1 (17.0)
204.9 (13.9)
110.9 (6.5)
1Calculated per (40). Data are combined from n = 4 independent experiments combined
489
490
491
25
MATERIALS AND METHODS
492
493
Bacteria, cell culture and media
494
Listeria monocytogenes strains used in this study are listed in Supplemental Table S1.
495
Wild Type (WT) L. monocytogenes is 10403S and all mutants indicated were created using this
496
parental background. Bacteria were grown in either BHI or LDM (29). Briefly, LDM contains
497
the following final concentrations: 50 mM MOPS/2 mM K2HPO4 (pH 7.5), 0.02%
498
MgSO4*7H2O, 0.5 mM Ca(NO3)2, 0.2% NH4Cl, 0.5% Glucose, 0.004% FeCl3/Na3-
499
Citrate*2H2O, 0.5 g/mL Riboflavin, 1 g/mL Thiamine-HCl, 0.5g/mL Biotin, 0.005g/mL
500
Lipoic Acid, 100 g/mL of the amino acids Isoleucine, Leucine, Valine, Methionine, Arginine,
501
Histidine-HCl, Cysteine-HCl. Bone marrow derived murine macrophages (BMDM) were
502
isolated from wild type C57BL/6 mice per standard conditions and frozen in liquid nitrogen. The
503
day before in vitro infections, cells were thawed, spun by centrifugation, and resuspended in
504
fresh DMEM-10 (Gibco DMEM #11995-065 with 4.5 g/L D-Glucose and 110 mg/L Sodium
505
Pyruvate, 10% Fetal Bovine Serum [HyClone], 1% HEPES [Gibco 1M 15630-080], 1% Non-
506
Essential Amino Acids [Gibco 100X 11140-050] and 1% L-Glutamine [Gibco 200 mM 25030-
507
081]).
508
509
Creation of mutant strains
510
The markerless, in-frame ilvE mutant was constructed using the pKSV7 recombination
511
plasmid (41) per standard conditions such that 1,020 base pairs of the coding sequence were
512
excised. The gene LMRG_02078 sequence in biocyc.org was used for mutant deletion method
513
design. The complemented strain ilvE::ilvE+ was constructed using the ilvE parental strain by
514
26
inserting the coding sequence for LMRG_02078, including 500 base pairs upstream of the start
515
codon, using the shuttle integration vector pPL2 (42) per standard procedures. Note that two
516
independent complemented strains were constructed, one with a FLAG tag inserted at the 5 end
517
of the gene (ilvE::ilvE-FL). Primer sequences are listed in Supplementary Table S2.
518
The codY null mutant was created by insertion-deletion of a spc gene originating from the
519
plasmid pJL73 (43). The entire codY coding sequence was replaced, in the same orientation, by
520
the spectinomycin resistance cassette using the shuttle vector pMAD (44) per standard
521
procedures. A more detailed description of construction of the codY null mutant, including
522
primer sequences, will be included in an upcoming manuscript prepared by T.A. Washington, B.
523
R. Belitsky, and A. L. Sonenshein.
524
525
Growth and survival analysis
526
Cultures of all strains were grown in liquid LDM or BHI medium to Optical Density 600
527
nm (OD600) 0.40 – 0.50 and frozen at -80C in 1 mL aliquots. Frozen stocks were titered for
528
viable bacteria, and on the day of experiments, aliquots were thawed at 37C for five minutes
529
and shaken at 37C in fresh medium for 30 minutes. Bacteria were then diluted 1:10 into fresh
530
medium and added to a Bioscreen C honeycomb 100-well plate in a 300 L volume in triplicate.
531
Plates were incubated at 37C for 24 hours with constant shaking at medium speed. OD600
532
readings were taken every 15 minutes on the Bioscreen C instrument. Growth was graphed in
533
Prism. Doubling times were calculated per (40) as follows: n = [log10(high OD600) – log10 (low
534
OD600)] / 0.3010 (where OD600 values are from exponentially dividing cells). Doubling time =
535
time between OD600 / n.
536
27
Survival during exposure to Bile Salts was performed as follows. Strains were thawed
537
from frozen stocks of bacteria grown to mid-log (OD600 ~0.45) in LDM, added to fresh LDM,
538
and shaken at 37C for 30 min. Bacteria (~107 bacteria/mL) were then added to 4 mL of PBS
539
containing Bile Salts (Sigma #48305) at 0, 1, 2 and 4 mg/mL. Tubes were shaken at 37C for 30
540
min and then serially diluted with plating on LB-agar plates.
541
542
Fatty Acid Content
543
Bacteria were grown in LDM to mid-log (OD600 0.4-0.5) and stationary phase (OD600
544
0.9 – 1.1), spun by centrifugation, washed 1X with PBS, spun again, and frozen at -20C. Cells
545
were sent on dry ice to Microbial ID for Whole Cell Fatty Acid Analysis. Experiments were
546
performed three times, independently. Results were combined and graphed in Prism 7 or 8 with
547
standard deviation.
548
549
Amino Acid Analysis
550
Strains grown on BHI agar were used to inoculate fresh liquid LDM and were grown to
551
mid-log (OD600 0.45 – 0.55) or stationary phase (OD600 > 0.8). Cultures (12 or 10 mL) were
552
spun by centrifugation and washed one time with 2 mL 150 mM Ammonium Acetate. Cells were
553
again spun by centrifugation, the supernatant was removed, and cell pellets were snap frozen in a
554
dry ice-ethanol bath. Cells were stored at -80C until delivery to the Michigan Regional
555
Comprehensive Metabolomics Resource Core (MRC2) at the University of Michigan and
556
analyzed for total amino acid content as follows. Briefly, cells were homogenized in 200 L of
557
extraction solvent (20% water, 80% 1:1:1 methanol:acetonitrile:acetone) containing 13C or 15N-
558
labeled amino acid internal standards. Samples were incubated at 4C for 10 min, vortexed, and
559
28
spun by centrifugation at 4C for 10 min at 14,000 rpm. Samples were diluted 20-fold and
560
transferred to autosampler vials for mass spectrometric analysis. Chromatographic separation of
561
underivatized amino acids was done using an Intrada Amino Acid column (Imtakt USA). Mobile
562
phases for separation were water:acetonitrile (8:2 v/v) containing 100 mM ammonium formate
563
(solvent A) and acetonitrile with 0.3% formic acid (solvent B). Flow rate was 0.6 ml/min, and
564
sample injection volume was 5 L. ESI-MS/MS data acquisition was performed in positive ion
565
mode on an Agilent 6410 LC-MS with MRM transitions programmed for both labeled and
566
unlabeled internal standards. A pooled plasma reference sample and “test pooled” sample were
567
included as quality controls. Calibration standards were prepared containing all 20 proteinogenic
568
amino acids at various concentrations and analyzed in replicate along with test samples. LC-MS
569
data were processed using MassHunter Quantitative Analysis software version B.07.00. Amino
570
acids were quantified as pmol/million cells (ascertained by serial dilution and plating) and as
571
pmol/g total protein using linear calibration curves generated form the standards listed above.
572
All peak areas in samples and calibration standards were first normalized to the peak area of the
573
internal standards.
574
575
In vitro bone marrow derived macrophage infections
576
Bone marrow derived macrophages (see Bacteria, cell culture and media) were thawed
577
and plated in 24-well tissue culture plates with 2.5 X 105 cells/well and allowed to recover
578
overnight (~18 hours) at 37C/5% CO2. Following recovery, medium was removed and replaced
579
with 500 L of DMEM (no antibiotics) containing bacteria (prepped as in Growth Curve
580
analysis) at Multiplicity of Infection (MOI) of one. BMDM with bacteria were incubated for 30
581
min at 37C/5% CO2 and then washed three times with warm DPBS++ (+Calcium and
582
29
+Magnesium Chloride – Gibco 14040). One mL fresh DMEM-10 with Gentamicin (50 g/mL)
583
was added to cells to kill extracellular bacteria. Cells were incubated for 0, 4 and 8 hours. At
584
time of harvest, cells were washed one time with DPBS++ and then incubated in 1 mL of 0.1%
585
Triton-X for 5 min. Cells were removed by scraping and pipetting and then transferred to 3.5
586
mL sterile double distilled water and vortexed for 10s. 500 L 10X PBS was added to promote
587
bacterial integrity. Samples were either directly plated or serially diluted and then plated on LB-
588
agar plates and incubated overnight at 37C. Experiments were done with three technical
589
replicates per experiment on three separate days. Data were compiled and graphed in Prism 7 or
590
8.
591
592
In vivo mouse experiments
593
Mouse experiments were performed with 6 to 7-week-old female BALB/c mice. Bacteria
594
were grown in BHI to OD600 0.50 and frozen in 1 mL aliquots. On the day of experiments,
595
bacteria were thawed and resuspended in 3 mL of fresh BHI and incubated with shaking for 1.5
596
hours at 37C. Bacteria were pelleted by centrifugation, washed one time with sterile PBS,
597
pelleted again, and then resuspended in 1 mL sterile PBS. Bacteria were serially diluted and
598
plated to ascertain original titer. Bacteria were then combined in the following strain
599
combinations in a 1:1 ratio to attain a concentration of 105 CFU of each strain per 100 L of
600
PBS. WT-ermr:WT-erms; WT-ermr:ilvE- erms; WT-ermr:ilvE::ilvE+- erms; WT-ermr:codY-
601
erms. Mice were injected peritoneally with 100 L of bacterial inoculum. Inocula for all strain
602
combinations were serially diluted and plated on LB-agar and LB-agar-erythromycin plates to
603
measure INPUT concentrations. Mice were then housed for 48 hours in biocontainment rooms
604
before sacrifice and harvest of spleens and livers. Spleens were homogenized in 1 mL sterile
605
30
PBS with 1.0 mm Zirconia/Silica beads (BioSpec 11079110z), and livers were homogenized in 5
606
mL sterile PBS with a handheld tissue homogenizer. Samples were serially diluted and plated in
607
duplicate on both non-antibiotic containing LB agar plates and LB agar plates containing
608
erythromycin. CFU/mL per gram of tissue were obtained for all samples sets post-harvest
609
(OUTPUT), and the number of antibiotic sensitive and resistant bacteria were obtained by [CFU
610
on LB plates] minus [CFU on erythromycin-containing plates] = sensitive bacteria. Ratios of
611
erm-sensitive to erm-resistant bacteria for both INPUT and OUTPUT were calculated, and the
612
Competitive index was calculated as OUTPUT ratio / INPUT ratio (36). Mouse experiments
613
were performed on two separate days with n = 3 and n = 4 mice per experiment.
614
615
L929 Plaque Assay
616
L929 cells (mouse fibroblast cells) were grown in DMEM-10 medium (see “cell culture”
617
above) medium and plated in 6-well tissue culture plates at 105 cells/well at 37C/5% CO2 until
618
cells were almost 100% confluent. On the day of experiments, medium was removed and
619
replaced with fresh medium containing Lm at MOI = 30, incubated for 1h at 37C/5% CO2, and
620
washed three times with DPBS++ (plus ions). A 1:1 agarose (1.4%):2X DMEM overlay was
621
then added to each well. Plates were incubated at 37C/5% CO2 until plaques were visible.
622
Neutral red mixed with PBS was added to the wells for 1 h to allow for visualization of plaques.
623
After plaques were visible, images of each plate were taken (with a ruler included in the picture),
624
and plaque diameter was measured in ImageJ using the ruler in millimeters (mm) as a standard.
625
At least ten plaques were measured in three separate wells for each of three independent
626
experiments performed on different days. Data were compiled, the mean and standard deviation
627
31
calculated, and the Student’s unpaired, two-tailed t-test was used to compare mutant strains to
628
WT. Data are shown as the mean plaque diameter percentage of WT per each experiment.
629
630
Gene expression via RT-qPCR
631
Bacteria were grown in LDM to mid-log (OD600 0.4-0.5) and stationary phase (OD600
632
0.9 – 1.1) and then spun by centrifugation. After lysis by bead-beating, total bacterial RNA was
633
isolated using either the “Quick-RNA Fungal/Bacterial Miniprep” (Zymo Research #R2014) or
634
the “FastRNA Blue Kit” (MPBio #116025-050). RNA was extracted per manufacturers’
635
protocols and treated with DNase. RNA was precipitated using isopropanol and quantitated on a
636
Nanodrop ND-1000 spectrophotometer. cDNA was made using 250 ng RNA with Invitrogen
637
SuperScript II RT per the manufacturer’s protocol. No-RT controls were created for each RNA
638
sample by omitting RT in cDNA prep. To measure relative gene expression, 1 L of cDNA was
639
used for SYBR green qPCR using Brilliant II SYBR Green QPCR Master Mix with Low ROX
640
(Agilent #600830) in Bio-Rad Hard Shell PCR 96-well plates (Bio-Rad #64201794) with all
641
cDNA preps done in duplicate, including all no-RT controls. Plates were run on a Bio-Rad
642
CFX96 Real-Time System Thermal Cycler with the following protocol: 10 min at 95C, 40X
643
cycle of [10s 95C; 45s 53C; 1 min 72C], 30s 95C, 30s 65C, 30s 95C. The L.
644
monocytogenes genes ilvE (LMRG_02078) and gyrA2 were measured, and gyrA2 was used as a
645
housekeeping gene for data normalization. Primer sequences are listed in Supplemental Table
646
S2. Changes in gene expression were calculated per the 2^Ct method (45) comparing mutants
647
to WT.
648
649
32
ACKNOWLEDGMENTS
650
This work was funded by NIH R01 AI109048. The authors acknowledge Microbial ID for Fatty
651
Acid analysis and the Michigan Regional Comprehensive Metabolomics Core at the University
652
of Michigan for amino acid analysis. Some Illustrations were partially generated in Biorender
653
(www.biorender.com).
654
655
656
657
33
REFERENCES
658
1.
Freitag NE, Port GC, Miner MD. 2009. Listeria monocytogenes - from saprophyte to
659
intracellular pathogen. Nat Rev Microbiol 7:623-628.
660
2.
Radoshevich L, Cossart P. 2018. Listeria monocytogenes: towards a complete picture of
661
its physiology and pathogenesis. Nat Rev Microbiol 16:32-46.
662
3.
King JE, Roberts IS. 2016. Bacterial Surfaces: Front Lines in Host-Pathogen
663
Interaction. Adv Exp Med Biol 915:129-156.
664
4.
Pizarro-Cerda J, Cossart P. 2018. Listeria monocytogenes: cell biology of invasion and
665
intracellular growth. Microbiol Spectr 6.
666
5.
Carvalho F, Sousa S, Cabanes D. 2014. How Listeria monocytogenes organizes its
667
surface for virulence. Front Cell Infect Microbiol 4:48.
668
6.
Raines LJ, Moss CW, Farshtchi D, Pittman B. 1968. Fatty acids of Listeria
669
monocytogenes. J Bacteriol 96:2175-2177.
670
7.
Kaiser JC, Heinrichs DE. 2018. Branching Out: Alterations in Bacterial Physiology and
671
Virulence Due to Branched-Chain Amino Acid Deprivation. MBio 9.
672
8.
Annous BA, Becker LA, Bayles DO, Labeda DP, Wilkinson BJ. 1997. Critical role of
673
anteiso-C15:0 fatty acid in the growth of Listeria monocytogenes at low temperatures.
674
Appl Environ Microbiol 63:3887-3894.
675
9.
Zhu K, Ding X, Julotok M, Wilkinson BJ. 2005. Exogenous isoleucine and fatty acid
676
shortening ensure the high content of anteiso-C15:0 fatty acid required for low-
677
temperature growth of Listeria monocytogenes. Appl Environ Microbiol 71:8002-8007.
678
10.
Giotis ES, McDowell DA, Blair IS, Wilkinson BJ. 2007. Role of branched-chain fatty
679
acids in pH stress tolerance in Listeria monocytogenes. Appl Environ Microbiol 73:997-
680
1001.
681
11.
Keeney K, Colosi L, Weber W, O'Riordan M. 2009. Generation of branched-chain
682
fatty acids through lipoate-dependent metabolism facilitates intracellular growth of
683
Listeria monocytogenes. J Bacteriol 191:2187-2196.
684
12.
Sun Y, O'Riordan MX. 2010. Branched-chain fatty acids promote Listeria
685
monocytogenes intracellular infection and virulence. Infect Immun 78:4667-4673.
686
13.
Sun Y, Wilkinson BJ, Standiford TJ, Akinbi HT, O'Riordan MX. 2012. Fatty acids
687
regulate stress resistance and virulence factor production for Listeria monocytogenes. J
688
Bacteriol 194:5274-5284.
689
14.
Rogiers G, Kebede BT, Van Loey A, Michiels CW. 2017. Membrane fatty acid
690
composition as a determinant of Listeria monocytogenes sensitivity to trans-
691
cinnamaldehyde. Res Microbiol 168:536-546.
692
15.
Chihib NE, Ribeiro da Silva M, Delattre G, Laroche M, Federighi M. 2003. Different
693
cellular fatty acid pattern behaviours of two strains of Listeria monocytogenes Scott A
694
and CNL 895807 under different temperature and salinity conditions. FEMS Microbiol
695
Lett 218:155-160.
696
16.
Zhu K, Bayles DO, Xiong A, Jayaswal RK, Wilkinson BJ. 2005. Precursor and
697
temperature modulation of fatty acid composition and growth of Listeria monocytogenes
698
cold-sensitive mutants with transposon-interrupted branched-chain alpha-keto acid
699
dehydrogenase. Microbiology 151:615-623.
700
17.
Singh AK, Zhang YM, Zhu K, Subramanian C, Li Z, Jayaswal RK, Gatto C, Rock
701
CO, Wilkinson BJ. 2009. FabH selectivity for anteiso branched-chain fatty acid
702
34
precursors in low-temperature adaptation in Listeria monocytogenes. FEMS Microbiol
703
Lett 301:188-192.
704
18.
Kaneda T. 1991. Iso- and anteiso-fatty acids in bacteria: biosynthesis, function, and
705
taxonomic significance. Microbiol Rev 55:288-302.
706
19.
Zhang YM, Rock CO. 2008. Membrane lipid homeostasis in bacteria. Nat Rev
707
Microbiol 6:222-233.
708
20.
Eisenreich W, Slaghuis J, Laupitz R, Bussemer J, Stritzker J, Schwarz C, Schwarz
709
R, Dandekar T, Goebel W, Bacher A. 2006. 13C isotopologue perturbation studies of
710
Listeria monocytogenes carbon metabolism and its modulation by the virulence regulator
711
PrfA. Proc Natl Acad Sci U S A 103:2040-2045.
712
21.
Brenner M, Lobel L, Borovok I, Sigal N, Herskovits AA. 2018. Controlled branched-
713
chain amino acids auxotrophy in Listeria monocytogenes allows isoleucine to serve as a
714
host signal and virulence effector. PLoS Genet 14:e1007283.
715
22.
Joseph B, Goebel W. 2007. Life of Listeria monocytogenes in the host cells' cytosol.
716
Microbes Infect 9:1188-1195.
717
23.
Brown SA, Palmer KL, Whiteley M. 2008. Revisiting the host as a growth medium.
718
Nat Rev Microbiol 6:657-666.
719
24.
Feyaerts J, Rogiers G, Corthouts J, Michiels CW. 2015. Thiol-reactive natural
720
antimicrobials and high pressure treatment synergistically enhance bacterial inactivation.
721
Innovative Food Science & Emerging Technologies 27:26-34.
722
25.
Bennett HJ, Pearce DM, Glenn S, Taylor CM, Kuhn M, Sonenshein AL, Andrew
723
PW, Roberts IS. 2007. Characterization of relA and codY mutants of Listeria
724
monocytogenes: identification of the CodY regulon and its role in virulence. Mol
725
Microbiol 63:1453-1467.
726
26.
Lobel L, Sigal N, Borovok I, Ruppin E, Herskovits AA. 2012. Integrative genomic
727
analysis identifies isoleucine and CodY as regulators of Listeria monocytogenes
728
virulence. PLoS Genet 8:e1002887.
729
27.
Lobel L, Sigal N, Borovok I, Belitsky BR, Sonenshein AL, Herskovits AA. 2015. The
730
metabolic regulator CodY links Listeria monocytogenes metabolism to virulence by
731
directly activating the virulence regulatory gene prfA. Mol Microbiol 95:624-644.
732
28.
Lobel L, Herskovits AA. 2016. Systems Level Analyses Reveal Multiple Regulatory
733
Activities of CodY Controlling Metabolism, Motility and Virulence in Listeria
734
monocytogenes. PLoS Genet 12:e1005870.
735
29.
Belitsky BR. 2014. Role of PdxR in the activation of vitamin B6 biosynthesis in Listeria
736
monocytogenes. Mol Microbiol 92:1113-1128.
737
30.
Biswas R, Sonenshein AL, Belitsky BR. 2020. Genome-wide Identification of Listeria
738
monocytogenes CodY-Binding Sites. Mol Microbiol doi:10.1111/mmi.14449.
739
31.
Brinsmade SR. 2017. CodY, a master integrator of metabolism and virulence in Gram-
740
positive bacteria. Curr Genet 63:417-425.
741
32.
Sonenshein AL. 2005. CodY, a global regulator of stationary phase and virulence in
742
Gram-positive bacteria. Curr Opin Microbiol 8:203-207.
743
33.
Levdikov VM, Blagova E, Colledge VL, Lebedev AA, Williamson DC, Sonenshein
744
AL, Wilkinson AJ. 2009. Structural rearrangement accompanying ligand binding in the
745
GAF domain of CodY from Bacillus subtilis. J Mol Biol 390:1007-1018.
746
35
34.
Levdikov VM, Blagova E, Young VL, Belitsky BR, Lebedev A, Sonenshein AL,
747
Wilkinson AJ. 2017. Structure of the Branched-chain Amino Acid and GTP-sensing
748
Global Regulator, CodY, from Bacillus subtilis. J Biol Chem 292:2714-2728.
749
35.
Sen S, Sirobhushanam S, Hantak MP, Lawrence P, Brenna JT, Gatto C, Wilkinson
750
BJ. 2015. Short branched-chain C6 carboxylic acids result in increased growth, novel
751
'unnatural' fatty acids and increased membrane fluidity in a Listeria monocytogenes
752
branched-chain fatty acid-deficient mutant. Biochim Biophys Acta 1851:1406-1415.
753
36.
Auerbuch V, Lenz LL, Portnoy DA. 2001. Development of a competitive index assay
754
to evaluate the virulence of Listeria monocytogenes actA mutants during primary and
755
secondary infection of mice. Infect Immun 69:5953-5957.
756
37.
White SJ, McClung DM, Wilson JG, Roberts BN, Donaldson JR. 2015. Influence of
757
pH on bile sensitivity amongst various strains of Listeria monocytogenes under aerobic
758
and anaerobic conditions. J Med Microbiol 64:1287-1296.
759
38.
Begley M, Gahan CG, Hill C. 2005. The interaction between bacteria and bile. FEMS
760
Microbiol Rev 29:625-651.
761
39.
Santiago B, MacGilvray M, Faustoferri RC, Quivey RG, Jr. 2012. The branched-
762
chain amino acid aminotransferase encoded by ilvE is involved in acid tolerance in
763
Streptococcus mutans. J Bacteriol 194:2010-2019.
764
40.
Moat AG. 2002. Microbial Physiology, 4th ed. Wiley-Liss.
765
41.
Smith K, Youngman P. 1992. Use of a new integrational vector to investigate
766
compartment-specific expression of the Bacillus subtilis spoIIM gene. Biochimie 74:705-
767
711.
768
42.
Lauer P, Chow MY, Loessner MJ, Portnoy DA, Calendar R. 2002. Construction,
769
characterization, and use of two Listeria monocytogenes site-specific phage integration
770
vectors. J Bacteriol 184:4177-4186.
771
43.
LeDeaux JR, Grossman AD. 1995. Isolation and characterization of kinC, a gene that
772
encodes a sensor kinase homologous to the sporulation sensor kinases KinA and KinB in
773
Bacillus subtilis. J Bacteriol 177:166-175.
774
44.
Arnaud M, Chastanet A, Debarbouille M. 2004. New vector for efficient allelic
775
replacement in naturally nontransformable, low-GC-content, gram-positive bacteria. Appl
776
Environ Microbiol 70:6887-6891.
777
45.
Livak KJ, Schmittgen TD. 2001. Analysis of relative gene expression data using real-
778
time quantitative PCR and the 2(-Delta Delta C(T)) Method. Methods 25:402-408.
779
780
WT
ΔilvE
ΔilvE
::ilvE
ΔcodY
0
25
50
75
100
125
Percent of total FA
WT
ΔilvE
ΔilvE
::ilvE
ΔcodY
0
25
50
75
100
125
Percent of total FA
B
C
Fatty Acids
Logarithmic Growth
Fatty Acids
Stationary Phase
89%
86%
85%
30%
21%
84%
85%
63%
Other
iso C15:C17
anteiso C15:C17
iso C14:C16
A
Ile
Leu
Val
Arg
0
1
2
3
4
Fold Difference [Amino Acid]
relative to WT
Ile
Leu
Val
Arg
0
1
2
3
4
Fold Difference [Amino Acid]
relative to WT
WT
ΔilvE
ΔilvE::ilvE
ΔcodY
Amino Acids
Logarithmic Growth
Amino Acids
Stationary Phase
D
E
Figure 1
Figure 1 Legend
Figure 1. Changes in Fatty Acid and BCAA content in Lm lacking IlvE or CodY. (A) Simplified
overview of branched chain fatty acid (BCFA) biosynthesis in Gram-positive bacteria (based on detailed
diagram in (9)) showing pathways that incorporate branched chain amino acids (BCAAs: Ile, Leu & Val).
Red X represents points in pathways where deletion mutants were used in this study. Colored arrows
indicate pathways of individual BCAAs that are incorporated into final BCFA isoforms (18). Purple text =
enzyme names. (B and C) Graphs represent the relative amounts of the major fatty acids as a percentage of
total fatty acids contained in Lm cultures of WT, ΔilvE, ΔilvE::ilvE+ and ΔcodY strains grown in nutrient
limiting medium (LDM) to (B) mid-logarithmic and (C) stationary phase. Graphs represent combined data
from three independent experiments. Graphs shown here and data in Tables 2 and 3 are the combined
quantities of odd numbered (C15 and C17) or even-numbered (C14 and C16) BCFAs. Individual
numbered species (e.g., ai-C15 only) and all other fatty acids are in Supplemental Tables S3 and S4. (D
and E). Cultures of WT, ΔilvE, ΔilvE::ilvE and ΔcodY strains grown in LDM to (D) mid-logarithmic and
(E) stationary phase were analyzed by mass spectrometry. Concentrations of BCAAs were normalized to
total protein content and are shown as ratios relative to WT. Error bars show the range of fold difference
compiled from 2 independent experiments.
A
Rich medium (BHI)
LDM: Ile, Leu, Val
C
Logarithmic growth
ilvE expression
Stationary phase
ilvE expression
WT
ΔilvE
ΔilvE:
:ilvE
ΔcodY
0.0
0.5
1.0
1.5
Fold change (2ΔΔCt)
relative to WT
WT
ΔilvE
ΔilvE:
:ilvE
ΔcodY
0.0
0.5
1.0
1.5
2.0
2.5
Fold change (2ΔΔCt)
relative to WT
B
0
2
4
6
8
10
12
14
16
18
20
22
24
26
0.0
0.5
1.0
1.5
2.0
time (hours)
Optical Density (600 nm)
WT
ΔcodY
ΔilvE
ΔilvE::ilvE
0
2
4
6
8
10
12
14
16
18
20
22
24
26
0.0
0.2
0.4
0.6
0.8
1.0
time (hours)
Optical Density (600 nm)
WT
ΔcodY
ΔilvE
ΔilvE::ilvE
Figure 2
Figure 2. Growth of ΔilvE and ΔcodY mutants in rich and nutrient-limited medium. Bacterial
growth of WT (circles), ΔilvE (triangles), ΔilvE::ilvE+ (inverted triangles), and ΔcodY (squares), was
analyzed on a Bioscreen instrument. Samples were inoculated from recovered frozen cultures that had
been prepared in LDM to mid-logarithmic phase. Optical Density at 600 nm (OD600) was measured for
24 hours at 37°C with shaking. Experiments include growth in (A) Rich medium = Brain Heart Infusion
(BHI) and (B) LDM containing amino acids at 100 µg/mL. Data are compiled from three independent
experiments with three technical replicates per experiment. Each point is the mean with error bars
representing the Standard Deviation. (C) RT-qPCR analysis of ilvE expression in Lm grown in LDM to
logarithmic (left) and stationary (right) phase.
0
2
4
6
8
10
12
14
16
18
20
22
24
26
0.0
0.2
0.4
0.6
0.8
1.0
time (hours)
Optical Density (600 nm)
WT
ΔcodY
ΔilvE
ΔilvE::ilvE
0
2
4
6
8
10
12
14
16
18
20
22
24
26
0.0
0.2
0.4
0.6
0.8
1.0
time (hours)
Optical Density (600 nm)
WT
ΔcodY
ΔilvE
ΔilvE::ilvE
0
2
4
6
8
10
12
14
16
18
20
22
24
26
0.0
0.2
0.4
0.6
0.8
1.0
time (hours)
Optical Density (600 nm)
WT
ΔcodY
ΔilvE
ΔilvE::ilvE
A
C
B
LDM: Ile, Leu, Val
D
LDM: Ile, Leu, Val
LDM: Ile, Leu, Val
LDM: Ile, Leu, Val
0
2
4
6
8
10
12
14
16
18
20
22
24
26
0.0
0.2
0.4
0.6
0.8
1.0
time (hours)
Optical Density (600 nm)
WT
ΔcodY
ΔilvE
ΔilvE::ilvE
Figure 2
Figure 3. Growth of Lm in LDM with variable exogenous BCAAs. Bacterial growth of WT
(circles), ΔilvE (triangles), ΔilvE::ilvE+ (inverted triangles), and ΔcodY (squares), performed as in
Figure 2, but in LDM containing (A) no BCAAs, (B) no Ile (Val & Leu only), (C) no Leu (Ile & Val
only), and (D) no Val (Ile & Leu only). Data are compiled from three independent experiments with
three technical replicates per experiment. Each point is the mean with error bars representing standard
deviation.
0
2
4
6
8
10
103
104
105
106
107
time (hours)
Log total CFU
WT
ΔcodY
ΔilvE
ΔilvE ::ilvE
Lm in BMDM
WT
ΔilvE
WT
ΔilvE ΔilvE:ilvE
0
25
50
75
100
125
Plaque diam %WT
ns
****
1.24
(0.25)
0.83
(0.20)
0.99
(0.19)
0.92
(0.24)
0.89
(0.23)
WT
ΔcodY
0
25
50
75
100
125
Plaque diam %WT
****
A
B
C
Figure 4
Figure 4. IlvE is required for optimal growth in macrophages and for cell-to-cell spread in cell
culture. (A) Total CFU from survival assays of Lm infection of Bone Marrow Derived Macrophages
(BMDM) assessed at 0.5, 4 and 8h post-infection. Data are compiled from three independent
experiments showing mean and standard deviation. MOI = 1. (B-C) Plaque assay of Lm grown in L9
fibroblasts. (B) Representative image of plaques formed by WT & ΔilvE bacteria after 48h growth.
(C) Average plaque diameters from experiments that included WT, ΔilvE and ΔilvE::ilvE+ (left) or
WT and ΔcodY (right). Numbers below graphs are the mean plaque diameter with standard deviation
compiled from three independent experiments. Two-tailed t-test comparing mutants to WT,
****P<0.0001; ns = not significant.
WTr:WTs
WTr:ΔcodYs
WTr:ΔilvEs
WTr:ΔilvE::ilvEs
0.001
0.01
0.1
1
10
competitive index
WTr:WTs
WTr:ΔcodYs
WTr:ΔilvEs
WTr:ΔilvE::ilvEs
0.001
0.01
0.1
1
10
competitive index
0
1
2
4
103
104
105
106
107
Bile Salts (mg/mL)
CFU/mL
WT
ΔilvE
ΔilvE::ilvE
ΔcodY
*
mean CI
mean CI
LOD
LOD
Lm in C57BL/6 mice
Spleen (48h)
Lm in C57BL/6 mice
Liver (48h)
[1.1]
[0.54]
[0.14]
[0.41]
[1.6]
[0.11]
[0.10]
[0.22]
B
C
A
****
ns
ns
ns
****
****
***
ns
ns
Figure 5
Figure 5. IlvE is required for resistance to membrane stress in response to bile salts and for survival
in a mouse model of listeriosis. (A) Log-phase bacteria grown in LDM were added to PBS with 0, 1, 2
and 4 mg/mL Bile Salts (Cholic acid-Deoxycholic acid sodium salt mixture) and incubated at 37°C for 30
minutes. Input for all samples was ~107 CFU/mL. Data are compiled from three independent experiments.
One-way ANOVA (non-parametric) with Dunn’s multiple comparisons post-test comparing mutant strains
to WT. ns = not significant; *P<0.05; ***P<0.001; ****P<0.0001. (B and C) Female C56BL/6 mice were
infected with a 1:1 mixture of erythromycin-sensitive test strains and erythromycin-resistant WT strain via
intraperitoneal injection. After 48h infection, (B) spleens and (D) livers were harvested and assessed for
viable CFU and competitive index (CI) was calculated as the ratio of Sensitive/Resistant CFU. Data
represent two independent experiments with total n=7 mice for all strains except WT, which was n=8.
LOD = limit of detection.
Figure 5 Legend
| 2020 | The branched chain aminotransferase IlvE promotes growth, stress resistance and pathogenesis of | 10.1101/2020.01.31.929828 | [
"Passalacqua Karla D.",
"Zhou Tianhui",
"Washington Tracy A.",
"Abuaita Basel H.",
"Sonenshein Abraham L.",
"O’Riordan Mary X.D."
] | null |
Topology and cleavage of astrotactins
1
Murine astrotactins 1 and 2 have similar membrane topology and mature via endoproteolytic cleavage
catalyzed by signal peptidase
Patricia Lara1, Åsa Tellgren-Roth1, Hourinaz Behesti2, Zachi Horn2, Nina Schiller1, Karl Enquist1,
Malin Cammenberg1, Amanda Liljenström1, Mary E. Hatten2,
Gunnar von Heijne1*, IngMarie Nilsson1*
1Department of Biochemistry and Biophysics, Stockholm University, SE-10691 Stockholm, Sweden
2Laboratory of Developmental Neurobiology, The Rockefeller University, New York, NY, USA 10065
Running title: Topology and cleavage of astrotactins
*To whom correspondence should be addressed: Gunnar von Heijne and IngMarie Nilsson, Department of
Biochemistry and Biophysics, Stockholm University, Svante Arrhenius väg 16C, SE-10691 Stockholm,
Sweden, Phone: +46-8-162590; E-mail: gunnar@dbb.su.se; ingmarie@dbb.su.se
Keywords: Astrotactin, topology, signal peptidase, neuronal migration, glycosylation, glycoprotein,
central nervous system, glia, synapse
ABSTRACT
Astrotactins 1 (Astn1) and Astn2 are
membrane proteins that function in glial-guided
migration, receptor trafficking and synaptic
plasticity in the brain, as well as in planar polarity
pathways in skin. Here, we used glycosylation
mapping and protease-protection approaches to
map the topologies of mouse Astn1 and Astn2 in
rough microsomal membranes (RMs), and found
that Astn2 has a cleaved N-terminal signal peptide
(SP), an N-terminal domain located in the lumen of
the
RMs
(topologically
equivalent
to
the
extracellular surface in cells), two transmembrane
helices (TMHs), and a large C-terminal lumenal
domain. We also found that Astn1 has the same
topology as Astn2 but we did not observe any
evidence of SP cleavage in Astn1. Both Astn1 and
Astn2 mature through endoproteolytic cleavage in
the second TMH; importantly, we identified the
endoprotease responsible for the maturation of
Astn1 and Astn2 as the endoplasmic reticulum
signal peptidase. Differences in the degree of
Astn1 and Astn2 maturation possibly contribute to
the higher levels of the C-terminal domain of Astn1
detected on neuronal membranes of the central
nervous system. These differences may also explain
the distinct cellular functions of Astn1 and Astn2,
such as in membrane adhesion, receptor trafficking,
and planar polarity signaling.
Astrotactins are vertebrate-specific integral
membrane glycoproteins known to play critical
roles in central nervous system (CNS) and skin
development (1-4). An understanding of the
function of Astn1 and Astn2 in the control of
neuronal migration and of synaptic function could
be important for treatment of human brain disorders
such as epilepsy and autism spectrum disorders.
Although the number of gene mutations that can
disrupt neuronal migration is large (5), Astn1 is one
of a few adhesion receptors shown to directly
function in migration (6).
In mouse, there are two astrotactin family
members, Astn1 and Astn2 (ASTN1 and ASTN2 in
humans). Astn1 is involved in glial-guided
neuronal migration early in development (1,3,6,7)
through the formation of an asymmetric complex
with N-cadherin (CDH2) in the glial membrane (6).
Astn2, which is 48% homologous to Astn1 and has
two isoforms, is abundant in migrating cerebellar
granule neurons where it forms a complex with
Topology and cleavage of astrotactins
2
Astn1, and regulates the trafficking of Astn1 during
migration (4). At later stages of development,
Astn2 regulates synaptic function by trafficking of
other
membrane
receptors,
including
the
Neuroligins and other synaptic proteins (8). A
recent structure of the C-terminal endodomain of
Astn2 shows distinctive features responsible for its
activity (9). Astn1 and Astn2 are believed to share
the same membrane topology, with a cleaved N-
terminal signal peptide (SP), two transmembrane
helices (TMHs), and a large extracellular C-
terminal domain (10). Both Astn1 and Astn2
undergo an endoproteolytic maturation step in
which an unknown protease cleaves the protein just
after the second TM segment, with the two
fragments remaining attached through a single
disulfide bond (10,11).
In the present work, we have mapped the
topologies of mouse Astn1 and Astn2 in rough
microsomal
membranes
using
glycosylation
mapping and protease-protection assays. We find
that Astn2 has a cleaved N-terminal SP, an N-
terminal domain located in the lumen of the RMs
(topologically equivalent to the extracellular
surface in cells), two TMHs, and a large C-terminal
lumenal domain. We further show that Astn1 has
the same topology as Astn2, but see no evidence of
SP cleavage for Astn1. Finally, we identify the
endoprotease responsible for the maturation of
Astn1 and Astn2 as signal peptidase, an ER-
localized enzyme that normally removes SPs from
secreted and membrane proteins.
Results
Predicted topologies of mouse Astn1 and
Astn2 – Topology predictions for mouse Astn1
(UniProtKB Q61137-1, splicing isoform 1) and
Astn2 (UniProtKB Q80Z10-3, splicing isoform 3)
produced by the TOPCONS server (12) agree with
the topology model for Astn2 derived from epitope
tagging and cell-surface staining (11), i.e., an N-
terminal signal peptide (SP) followed by two
transmembrane segments (TMH1 & 2) and a large
C-terminal extracellular domain, Fig. 1. In cells,
both Astn1 and Astn2 are cleaved by an
unidentified endoprotease into two fragments that
remain linked by a disulfide bond (11). Edman
sequencing of the two Astn2 fragments showed that
the N-terminal one starts at Gly52 (just after the
predicted signal peptide) and the C-terminal one at
Asn466 (corresponding to Asn414 in the isoform
analyzed here). For Astn1, the C-terminal fragment
starts at Ser402; no sequence could be obtained from
the N-terminal fragment in this case.
Topology mapping of mouse Astn1 – To
characterize the mouse Astn1 protein we used a
well-established in vitro glycosylation assay
(13,14) to determine the topology of the protein
when cotranslationally inserted into dog pancreas
rough microsomes (RMs). The transfer of
oligosaccharides
from
the
oligosaccharide
transferase (OST) enzyme to natural or engineered
acceptor sites for N-linked glycosylation (-Asn-
Xxx-Ser/Thr-Yyy, where Xxx and Yyy cannot be
Pro (15-18)) in a nascent polypeptide chain is a
characteristic protein modification that can only
happen in the lumen of the ER where the active site
of the OST is located (19,20). The topology of
Astn1 in RMs was also probed by treatment with
proteinase K, that can only digest parts of the
protein protruding from the cytosolic side of the
RMs (21).
To be able to investigate the topology of the
1,302-residues-long and heavily glycosylated
Astn1 protein, we selected to work with various
truncated versions of the full-length protein. This
was necessary both because in vitro translation of
such large proteins is inefficient, and because the
attachment of an oligosaccharide increases the size
of the protein by only 2-3 kDa, a shift that is too
small to be detectable by SDS-PAGE for the full-
length protein but can easily be visualized when
using truncated versions.
Truncated
versions
of
Astn1
were
expressed in vitro using the TNT® SP6 Quick
Coupled System supplemented with column-
washed dog pancreas rough microsomes (RMs)
(14,21). The glycosylation status was investigated
using SDS-PAGE, and truncated Astn1 versions
were
designed
such
that
differences
in
glycosylation patterns could be used to infer the
topology of the protein in the RM membrane.
As shown in Fig. 2A, Astn1 1-381, a
version that extends from the putative SP to the end
of the loop between TMH1 and TMH2, receives a
single glycan when translated in the presence of
RMs (compare lanes 1 and 2). Notably, there is no
sign of the SP being cleaved (which would reduce
the Mw of the protein by 2.6 kDa). Astn1 78-381
(lanes 3, 4) and Astn1 78-451 (lanes 5, 6) also
receive only a single glycan, while Astn1 78-470
(lanes 7, 8) is glycosylated on two sites (note that
Topology and cleavage of astrotactins
3
glycan acceptor sites are rarely if ever modified to
100% in the in vitro translation system, hence
molecules with both one and two added glycans are
visible on the gel). The second glycan addition
therefore must be on Asn453.
To determine whether the first glycan
addition is on Asn115 or Asn226 (Asn328 is too close
to TMH2 to be reached by the OST (22)), we
expressed Astn1 versions lacking the entire N-
terminal region, up to but not including TMH2, Fig.
2B. The two shorter versions were not glycosylated
at all when expressed in the presence of RMs, while
Astn1 160-470 was modified on a single
glycosylation site. The latter must be Asn453,
showing that neither Asn226 nor Asn328 become
glycosylated. We conclude that the putative SP in
Astn1 appears not to be cleaved by signal peptidase
and probably forms an N-terminal transmembrane
helix (TMH0), and that Astn1 has two segments
(residues 22-152 and 402-1,302) exposed to the
lumen of the RMs, and one segment (residues 174-
380) exposed to the cytosol. Further, since Asn115 is
glycosylated in all four constructs, it appears that
the N-terminal segment in the Astn1 constructs that
start at M78 can be translocated to the lumenal side
of the RMs even though it lacks the putative SP.
We further used a protease-protection
assay (21) to verify the proposed topology of Astn1.
In order that segments of Astn1 that are protected
from proteinase digestion by the RM membrane
would be of a convenient size for SDS-PAGE
separation, we first expressed Astn1 78-728. As
seen in Fig. 2C, the protein becomes glycosylated
(compare lanes 1 and 2) but it is difficult to
determine on how many sites. Interestingly, two
prominent bands at ~38 kDa (marked N) and ~36
kDa (marked C) were generated in the presence of
RMs
(lane
2),
suggesting
an
internal
endoproteolytic cleavage, in agreement with the
published Edman sequencing results that identified
a cleavage site between Ser401 and Ser402 (11). In
addition, a third band at ~65 kDa that appears to
receive a single glycan in the presence of RMs was
also seen (lanes 1 and 2). The latter would be
consistent with internal translation initiation at
Met160, and indeed comigrates with Astn1 160-728
(lane 4).
Proteinase K treatment of RMs carrying
Astn1 78-728 digests cytoplasmically accessible
parts of the protein and leaves only two protected
fragments:
one
of
identical
size
to
the
“endoproteolytic” 36 kDa band, and one at ~39 kDa
(lane 3). The two protease-protected fragments are
precisely what would be expected from the
topology derived from the glycosylation study: the
39 kDa band (marked C*) represents the fragment
381-728 generated when proteinase K digests the
cytosolic loop, and the 36 kDa band represents the
slightly smaller C-terminal fragment 402-728
generated by endoproteolytic cleavage near the C-
terminal end of TMH2. The expected protected N-
terminal fragment 78-181 is too small to be
resolved on the gel.
Similar results were obtained for Astn1
160-728. In addition to the full-length protein at
~65 kDa, two bands at ~36 kDa (marked C) and ~25
kDa (marked N) were seen in the presence of RMs
(compare lanes 5 and 6); EndoH treatment shifted
both the full-length band at ~65 kDa and the ~36
kDa band to a lower Mw, while the 25 kDa band
did not shift (lane 8). Consistent with the Astn1
160-728 results, the glycosylated ~36 kDa band
represents the same endoproteolytic C-terminal
fragment 402-728, while the unglycosylated 25
kDa
band
represents
the
N-terminal
endoproteolytic fragment 160-401.
Given the sequence context of the
endoproteolytic cleavage site (see Discussion), we
hypothesized that the responsible protease may be
signal peptidase. Indeed, inclusion of a signal
peptidase inhibitor (23) in the in vitro translation of
Astn1 160-728 completely inhibits the formation of
the ~36 kDa and ~25 kDa products (lane 11).
We conclude that Astn1 has the same
topology as previously proposed for Astn2, namely
with two lumenal domains (residues 22-152 and
173-1,302) and one cytosolic domain (residues
174-381). The putative SP appears to not to be
cleaved,
but
rather
forms
an
N-terminal
transmembrane helix (TMH0). We identify signal
peptidase as the enzyme responsible for the
endoproteolytic cleavage event at Ser401.
Topology mapping of mouse Astn2 – We
used the same glycosylation mapping approach to
determine the topology of the 1,300 amino acids-
long mouse Astn2 protein (splice isoform 3, lacking
exon 4 that encodes a 52 residues segment in the
domain between TMH1 and TMH2). Astn2 1-482
includes both the putative SP, the two predicted
transmembrane helices TMH1 and TMH2, and a
portion of the large C-terminal domain. A small
amount of glycosylated full-length product at ~56
Topology and cleavage of astrotactins
4
kDa, two weak bands at ~50 kDa that might
represent glycosylated and unglycosylated products
lacking the SP (which has a calculated Mw of 6.4
kDa), and a prominent product at ~43 kDa are seen
in the presence of RMs, Fig. 3A (lanes 2, 4, 5). The
latter is sensitive to EndoH digestion, and the two
bands at ~50 kDa collapse to the lower Mw form
upon the same treatment (lane 6). The glycosylated
43 kDa band fits the Mw expected for a product
resulting from removal of the signal peptide
(residues 1-51) and the endoproteolytic cleavage at
Asn413 observed by Edman sequencing (11) (note
that we use a different splice version of Astn2 that
lacks 52 residues in the cytosolic segment
compared to the one used in this reference). This
explains the limited amount of glycosylated full-
length product (lanes 2, 4, 5), since most of the
molecules that become glycosylated are cleaved
after the SP and/or TMH2, as seen in lane 6.
To confirm this interpretation, we also
analyzed Astn2 161-482 that lacks the putative SP.
As seen in Fig. 3B, Astn2 161-482 yields four
prominent bands when expressed in the presence of
RMs (lane 2): unglycosylated full-length product at
~37 kDa, singly- and doubly-glycosylated full-
length products at ~39 kDa and ~42 kDa, and a
smaller endoproteolytic product at ~35 kDa. EndoH
treatment collapses the ~39 kDa and ~42 kDa bands
to the size of the unmodified full-length product at
~37 kDa, and the ~35 kDa band to a smaller ~30
kDa band (lane 5). Similar to Astn1, addition of the
signal peptidase inhibitor to the in vitro translation
completely inhibits the formation of the ~35 kDa
endoproteolytic product (lane 3), and signal
peptidase inhibitor plus EndoH treatment of RM-
integrated
Astn2
161-482 leaves
only
the
unmodified full-length product (lane 7; for
unknown reasons, the signal peptidase inhibitor
makes bands run slightly higher in the gel).
These results are entirely consistent with
the proposed topology of Astn2 (11), and identify
signal peptidase as the enzyme responsible for the
endoproteolytic cleavage event at Asn413.
Discussion
Earlier work using epitope mapping of
Astn2 expressed in COS7 cells have shown that the
N- and C-termini are exposed on the cell surface,
while the domain between TMH1 and TMH2 can
only
be
immunodecorated
in
detergent-
permeabilized cells (11). Further, both Astn1 and
Astn2 were shown to be cleaved by an unknown
endoprotease into an N- and a C-terminal fragment,
and Edman sequencing of the C-terminal fragments
identified cleavage sites between Ser401-Ser402 in
Astn1 and Gly465-Asn466 in Astn2, just after TMH2.
In addition, for Astn2, Edman sequencing of the N-
terminal
endoproteolytic
fragment
indicated
removal of the putative SP (residues 1-51); no
sequence was obtained for Astn1, leaving open
whether or not the putative SP is cleaved in this
protein.
Here, we have confirmed and extended
these results for Astn1 and Astn2 using
glycosylation mapping and protease-protection
assays in a coupled in vitro transcription-translation
system supplemented with RMs. Our results for
Astn2 are in perfect agreement with those from the
earlier study: Astn2 has a cleaved N-terminal SP,
an N-terminal domain located in the lumen of the
RMs (topologically equivalent to the extracellular
surface in cells), two TMHs, and a large C-terminal
lumenal domain, Fig. 4. We find that Astn1 has the
same topology as Astn2 but see no evidence of SP
cleavage; rather, it seems that the putative N-
terminal SP in Astn1 remains a part of the protein,
presumably forming a third transmembrane helix
(TMH0).
We further show that an inhibitor of the
signal peptidase complex completely inhibits the
endoproteolytic cleavage of both Astn1 and Astn2.
The unknown endoprotease involved in the
maturation of Astn1 and Astn2 is thus signal
peptidase, the enzyme that cleaves SPs from
secretory and membrane proteins in the ER (24).
While it is uncommon that signal peptidase
catalyzes internal cleavage reactions of this kind in
cellular proteins, many viral polyproteins mature
through signal peptidase-catalyzed cleavages after
internal hydrophobic segments in the primary
translation product (25,26). Indeed, the SP cleavage
site and the cleavage site after TMH2 identified by
Edman sequencing in Astn2 are precisely the ones
predicted
by
the
SignalP
server
(27),
Supplementary Fig. S2.
The present findings raise the possibility
that higher levels of SP-mediated cleavage of Astn2
relative to Astn1 explain the higher levels of the
Astn1 C-terminus we previously detected on CNS
neuronal surface membranes by antibody labeling
and functional assays (6,8). This also likely
contributes to the apparently distinct functions of
Topology and cleavage of astrotactins
5
Astn1 as a membrane adhesion receptor that
functions in glial-guided migration (3,6,7), and of
Astn2 as an endolysosomal trafficking protein that
functions in both migration (4) and synaptic
function (8). Finally, the exceptionally long Astn2
SP hints at the possibility that, after cleavage, the
SP may have additional functions in the cell, as seen
for many other very long SPs (28). It will therefore
be of interest to determine whether the Astn2 SP
domain functions in receptor trafficking or planar
polarity signaling pathways.
Experimental procedures
Enzymes and chemicals – Unless otherwise
stated, all chemicals were from Sigma-Aldrich (St.
Louis, MO, US). Plasmid pGEM1, TNT® Quick
Coupled transcription/translation system, Rabbit
Reticulocyte lysate system and deoxynucleotides
were from Promega (Madison, WI, US). [35S]-Met
was from PerkinElmer (Boston MA, US). All
enzymes were from Fermentas (Burlington, Ontario,
CA) except Phusion DNA polymerase that was
from Finnzymes (Espoo, FI) and SP6 RNA
Polymerase from Promega. The QuikChange™
Site-directed Mutagenesis kit was from Stratagene
(La Jolla, CA, US) and oligonucleotides were from
Eurofins MWG Operon (Ebersberg, DE). All other
reagents were of analytical grade and obtained from
Merck (Darmdstadt, Germany).
DNA manipulations – The cDNAs of
mouse astrotactin 1 and 2 (Astn1 and Astn2) (1,302
respectively 1,300 amino acid residues; see
Supplementary Figure S1) were cloned into the
pRK5
vector
using
ClaI/SalI
(Astn1)
and
BamHI/XbaI (Astn2) sites. The DNA was then
transferred to the pGEMI vector (Promega) at
XbaI/SmaI sites together with a preceding Kozak
sequence (29), as previously described (13). To
create truncations in Astn1, deletions were made
between amino acid position 1-78 and 1-160, and
stop codons were introduced at positions 382, 452,
471, and 729. Astn2 truncations were created in the
same way, with a deletion between 1-161 and a stop
codon at 483. The Astn1 and Astn2 cDNAs were
amplified by PCR using the Phusion DNA
polymerase with appropriate primers, and site-
specific mutagenesis was performed using the
QuikChangeTM Site-Directed Mutagenesis Kit from
Stratagene. All mutants were confirmed by
sequencing of plasmid DNA at Eurofins MWG
Operon (Ebersberg, Germany) and BM labbet AB
(Furulund, Sweden).
In vitro expression – All Astn constructs
cloned in the pGEMI and pRK5 were transcribed
and translated in an in vitro the TNT® SP6 Quick
Coupled System from Promega. 150-200 ng DNA
template, 1 µl of [35S]-Met (5 µCi) and 0.5 µl
column-washed dog pancreas rough microsomes
(RMs) (tRNA Probes, US) (30) were added to 10 µl
of reticulocyte lysate at the start of the reaction, and
the samples were incubated for 90 min at 30 °C (21).
Proteinase K treatment – PK treatment was
performed by adding 1 µl of CaCl2 (200 mM) and
0.2 µl of Proteinase K (4.5 units/ µl) to the
translation reaction. After incubating on ice for 30
min, 1 ml of PMSF (20 mM ethanolic solution) was
added to inactivate PK and samples were further
incubated on ice for 5 min (21).
EndoH treatment – For endoglycosidase H
(EndoH) treatment 9 µl of the TNT reaction was
mixed with 1 µl of 10X glycoprotein denaturing
buffer. Following addition of 1 µl of EndoH
(500,000 units/ml; NEB, MA, US), 7 µl of dH2O
and 2 µl of 10X G3 reaction buffer, and the sample
was incubated at 37 °C for 1 h (31). Mock controls
were identical, but lacking EndoH.
SPI treatment – To demonstrate cleavage
by signal peptidase, the inhibitor SPI (N-
methoxysuccinyl-Ala-Ala-Pro-Val-
chloromethylketone from Sigma) was dissolved in
dimethyl sulfoxide (DMSO) and added to the
translation mix at a final concentration of 1.4 mM
as previously described (14,23,31-33).
Analysis and quantitations - Translation
products were analyzed under reducing conditions
by SDS-polyacrylamide gel electrophoresis, and
proteins were visualized in a Fuji FLA 9000
phosphorimager (Fujifilm, Tokyo, JP) using the
Image Reader FLA 9000/Image Gauge V 4.23
software (Fujifilm).
Topology and cleavage of astrotactins
6
Acknowledgements: We gratefully thank Prof. Arthur E. Johnson, Texas A&M University, for providing
dog pancreas microsomes.
Conflict of interest: The authors declare that they have no conflicts of interest with the contents of this
article.
Author contributions: Planned experiments: (PL, ÅT-R, NS, KE, MEH, GvH, IN), performed
experiments: (PL, ÅT-R, HB, ZH, NS, KE, MC, AL), analyzed data: (PL, ÅT-R, KE, MEH, GvH, IN),
wrote the paper: (MEH, GvH, IN).
Topology and cleavage of astrotactins
7
References
1.
Zheng, C., Heintz, N., and Hatten, M. E. (1996) CNS gene encoding astrotactin, which supports
neuronal migration along glial fibers. Science 272, 417-419
2.
Chang, H., Cahill, H., Smallwood, P. M., Wang, Y., and Nathans, J. (2015) Identification of
Astrotactin2 as a Genetic Modifier That Regulates the Global Orientation of Mammalian Hair
Follicles. PLoS Genet 11, e1005532
3.
Edmondson, J. C., Liem, R. K., Kuster, J. E., and Hatten, M. E. (1988) Astrotactin: a novel
neuronal cell surface antigen that mediates neuron-astroglial interactions in cerebellar
microcultures. J Cell Biol 106, 505-517
4.
Wilson, P. M., Fryer, R. H., Fang, Y., and Hatten, M. E. (2010) Astn2, a novel member of the
astrotactin gene family, regulates the trafficking of ASTN1 during glial-guided neuronal
migration. J Neurosci 30, 8529-8540
5.
Ross, M. E., and Walsh, C. A. (2001) Human brain malformations and their lessons for neuronal
migration. Annu Rev Neurosci 24, 1041-1070
6.
Horn, Z., Behesti, H., and Hatten, M. E. (2018) N-cadherin provides a cis and trans ligand for
astrotactin that functions in glial-guided neuronal migration. Proc Natl Acad Sci U S A 115,
10556-10563
7.
Adams, N. C., Tomoda, T., Cooper, M., Dietz, G., and Hatten, M. E. (2002) Mice that lack
astrotactin have slowed neuronal migration. Development 129, 965-972
8.
Behesti, H., Fore, T. R., Wu, P., Horn, Z., Leppert, M., Hull, C., and Hatten, M. E. (2018)
ASTN2 modulates synaptic strength by trafficking and degradation of surface proteins. Proc Natl
Acad Sci U S A 115, E9717-E9726
9.
Ni, T., Harlos, K., and Gilbert, R. (2016) Structure of astrotactin-2: a conserved vertebrate-
specific and perforin-like membrane protein involved in neuronal development. Open Biol 6
10.
Chang, H. (2017) Cleave but not leave: Astrotactin proteins in development and disease. IUBMB
Life 69, 572-577
11.
Chang, H., Smallwood, P. M., Williams, J., and Nathans, J. (2017) Intramembrane Proteolysis of
Astrotactins. J Biol Chem 292, 3506-3516
12.
Tsirigos, K. D., Peters, C., Shu, N., Kall, L., and Elofsson, A. (2015) The TOPCONS web server
for consensus prediction of membrane protein topology and signal peptides. Nucleic acids
research 43, W401-407
13.
Lundin, C., Nordstrom, R., Wagner, K., Windpassinger, C., Andersson, H., von Heijne, G., and
Nilsson, I. (2006) Membrane topology of the human seipin protein. FEBS Lett 580, 2281-2284
14.
Cuviello, F., Tellgren-Roth, A., Lara, P., Ruud Selin, F., Monne, M., Bisaccia, F., Nilsson, I., and
Ostuni, A. (2015) Membrane insertion and topology of the amino-terminal domain TMD0 of
multidrug-resistance associated protein 6 (MRP6). FEBS Lett 589, 3921-3928
15.
Mellquist, J. L., Kasturi, L., Spitalnik, S. L., and Shakin-Eshleman, S. H. (1998) The amino acid
following an asn-X-Ser/Thr sequon is an important determinant of N-linked core glycosylation
efficiency. Biochemistry 37, 6833-6837
16.
Shakin-Eshleman, S. H., Spitalnik, S. L., and Kasturi, L. (1996) The amino acid at the X position
of an Asn-X-Ser sequon is an important determinant of N-linked core-glycosylation efficiency. J
Biol Chem 271, 6363-6366
17.
Igura, M., and Kohda, D. (2011) Quantitative assessment of the preferences for the amino acid
residues flanking archaeal N-linked glycosylation sites. Glycobiology 21, 575-583
18.
Bano-Polo, M., Baldin, F., Tamborero, S., Marti-Renom, M. A., and Mingarro, I. (2011) N-
glycosylation efficiency is determined by the distance to the C-terminus and the amino acid
preceding an Asn-Ser-Thr sequon. Protein Sci 20, 179-186
19.
Johansson, M., Nilsson, I., and von Heijne, G. (1993) Positively charged amino acids placed next
to a signal sequence block protein translocation more efficiently in Escherichia coli than in
mammalian microsomes. Mol Gen Genet 239, 251-256
Topology and cleavage of astrotactins
8
20.
Kelleher, D. J., and Gilmore, R. (2006) An evolving view of the eukaryotic
oligosaccharyltransferase. Glycobiology 16, 47R-62R
21.
Lara, P., Ojemalm, K., Reithinger, J., Holgado, A., Maojun, Y., Hammed, A., Mattle, D., Kim,
H., and Nilsson, I. (2017) Refined topology model of the STT3/Stt3 protein subunit of the
oligosaccharyltransferase complex. J Biol Chem 292, 11349-11360
22.
Nilsson, I., and von Heijne, G. (1993) Determination of the Distance Between the
Oligosaccharyltransferase Active Site and the Endoplasmic Reticulum Membrane. J Biol Chem
268, 5798-5801
23.
Green, N., Fang, H., Miles, S., and Lively, M. O. (2002) Structure and function of the
endoplasmic reticulum signal peptidase complex. in The enzymes (Dalbey, R. E., and Sigman, D.
S. eds.), 3rd Ed., Academic press. pp 57-75
24.
Nyathi, Y., Wilkinson, B. M., and Pool, M. R. (2013) Co-translational targeting and translocation
of proteins to the endoplasmic reticulum. Biochim Biophys Acta 1833, 2392-2402
25.
Liljeström, P., and Garoff, H. (1991) Internally Located Cleavable Signal Sequences Direct the
Formation of Semliki Forest Virus Membrane Proteins from a Polyprotein Precursor. J Virol 65,
147-154
26.
Pene, V., Lemasson, M., Harper, F., Pierron, G., and Rosenberg, A. R. (2017) Role of cleavage at
the core-E1 junction of hepatitis C virus polyprotein in viral morphogenesis. PLoS One 12,
e0175810
27.
Petersen, T. N., Brunak, S., von Heijne, G., and Nielsen, H. (2011) SignalP 4.0: discriminating
signal peptides from transmembrane regions. Nature methods 8, 785-786
28.
Kapp, K., Schrempf, S., Lemberg, M.K., and Dobberstein, B. . (2009) Post-Targeting Functions
of Signal Peptides in Protein Transport into the Endoplasmic Reticulum (Zimmermann, R. ed.),
Landes Bioscience. 2000-2013. pp 1-11
29.
Kozak, M. (1989) Context effects and inefficient initiation at non-AUG codons in eucaryotic cell-
free translation systems. Mol Cell Biol 9, 5073-5080
30.
Walter, P., and Blobel, G. (1983) Preparation of microsomal membranes for cotranslational
protein translocation. Methods Enzymol 96, 84-93
31.
Lundin, C., Kim, H., Nilsson, I., White, S. H., and von Heijne, G. (2008) Molecular code for
protein insertion in the endoplasmic reticulum membrane is similar for N(in)-C(out) and N(out)-
C(in) transmembrane helices. Proc Natl Acad Sci U S A 105, 15702-15707
32.
Nilsson, I., Johnson, A. E., and von Heijne, G. (2003) How hydrophobic is alanine? J Biol Chem
278, 29389-29393
33.
Nilsson, I., Johnson, A. E., and von Heijne, G. (2002) Cleavage of a tail-anchored protein by
signal peptidase. FEBS Lett 516, 106-108
Topology and cleavage of astrotactins
9
FOOTNOTES
This work was supported by grants from the Knut and Alice Wallenberg Foundation (2012.0282) and the
Swedish Research Council (621-2014-3713) to GvH, the Swedish Cancer Foundation (15 0888) to GvH
and IMN, the Swedish Foundation for International Cooperation in Research and Higher Education
(STINT) (210/083(12); KU 2003-4674) to IMN, the Swedish Foundation for Strategic Research (SSF)
(A305:200) and the SSF-Infection Biology (2012(SB12-0026)) to IMN, the Eugene W. Chinery Trust to
MEH, and the Renate, Hans, and Maria Hofmann Trust to MEH.
The abbreviations used are: Astn, Astrotactin; RM, rough microsomes from dog pancreas; OST,
oligosaccharyl transferase; ER, endoplasmic reticulum; TM, transmembrane; TMH, transmembrane helix;
EndoH, endoglycosidase H; PK, proteinase K; SPI, signal peptidase inhibitor
Topology and cleavage of astrotactins
10
Figure 1. TOPCONS topology predictions. (A) Overview of the sequence of Astn1 with hydrophobic
segments (blue), potential acceptor sites for N-linked glycosylation (Y), and proteolytic cleavage sites
(red triangles) determined by Edman sequencing (11) marked. The TOPCONS topology prediction
(http://topcons.cbr.su.se) is given below. TOPCONS is a consensus predictor that collects data from the
other prediction servers listed in the panel. (B) Same for Astn2.
Topology and cleavage of astrotactins
11
Figure 2. Topology mapping of Astn1 and inhibition of endoproteolytic cleavage by an inhibitor of signal
peptidase. (A) The indicated truncated versions of Astn1 were translated in vitro with [35S]-Met in the
presence (+) or absence (-) of RMs, and analyzed under reducing conditions by SDS-PAGE.
Unglycosylated products are indicated by an open circle, singly glycosylated products by a filled circle,
and doubly glycosylated products by two filled circles. The glycosylated Asn residues are indicated by a
Topology and cleavage of astrotactins
12
red circle in the cartoon. (B) Same as in panel A. (C) Astn1 78-728 was translated in vitro with [35S]-Met
±RMs (lanes 1 and 2). RMs were subjected to proteinase K (PK) digestion (lane 3). The N- and C-
terminal fragments resulting from endoproteolytic cleavage between Ser401 and Ser402 are indicated (N,
C), as is the protease-protected C-terminal fragment (C*). RMs carrying Astn1 160-728 were subjected to
EndoH (EH) digestion (lanes 4-8). Note the shift in mobility for the full-length and C bands caused by de-
glycosylation (compare lanes 7 and 8). Astn1 160-728 was also translated in vitro with [35S]-Met in the
presence (+) or absence (-) of RMs and the signal peptidase inhibitor N-methoxysuccinyl-Ala-Ala-Pro-
Val-chloromethylketone (SPI), lanes 9-11.
Topology and cleavage of astrotactins
13
Topology and cleavage of astrotactins
14
Figure 3. Topology mapping of Astn2 and inhibition of endoproteolytic cleavage by an inhibitor of signal
peptidase. (A) Astn2 1-482 was translated in vitro with [35S]-Met in the presence (+) or absence (-) of
RMs, and analyzed under reducing conditions by SDS-PAGE (lanes 1 and 2). Unglycosylated products
are indicated by an open circle, and singly glycosylated products by a filled circle. Two cleavage products
potentially resulting from removal of the SP by signal peptidase are indicated by a bracket, and the N-
terminal endoproteolytic fragment is marked by *. EndoH digestion of RMs with Astn2 1-482 is shown in
lanes 3-6; note that the two products potentially generated by removal of the SP (bracket) coalesce into
one band and that the endoproteolytic fragment (N) shifts to a lower molecular weight upon de-
glycosylation (lane 6). (B) Astn2 161-482 was translated in vitro with [35S]-Met in the presence (+) or
absence (-) of RMs and the signal peptidase inhibitor N-methoxysuccinyl-Ala-Ala-Pro-Val-
chloromethylketone (SPI). After translation, RMs were further treated with EndoH (EH) or subjected to
mock treatment. The glycosylated Asn residues are indicated by a red circle in the cartoons.
Topology and cleavage of astrotactins
15
Figure 4. Topology and proteolytic modifications of Astn1 and Astn2. Signal peptidase cleaves both
Astn1 and Astn2 after TMH2, and also removes the SP from Astn2. The disulfide bridge that keeps the
two endoproteolytic fragments together is indicated.
>sp|O14525|ASTN1_HUMAN Astrotactin-1 OS=Homo sapiens OX=9606 GN=ASTN1 PE=2
SV=3. 1,302 aa
MALAGLCALLACCWGPAAVLATAAGDVDPSKELECKLKSITVSALPFLRENDLSIMHSPS
ASEPKLLFSVRNDFPGEMVVVDDLENTELPYFVLEISGNTEDIPLVRWRQQWLENGTLLF
HIHHQDGAPSLPGQDPTEEPQHESAEEELRILHISVMGGMIALLLSILCLVMILYTRRRW
CKRRRVPQPQKSASAEAANEIHYIPSVLIGGHGRESLRNARVQGHNSSGTLSIRETPILD
GYEYDITDLRHHLQRECMNGGEDFASQVTRTLDSLQGCNEKSGMDLTPGSDNAKLSLMNK
YKDNIIATSPVDSNHQQATLLSHTSSSQRKRINNKARAGSAFLNPEGDSGTEAENDPQLT
FYTDPSRSRRRSRVGSPRSPVNKTTLTLISITSCVIGLVCSSHVNCPLVVKITLHVPEHL
IADGSRFILLEGSQLDASDWLNPAQVVLFSQQNSSGPWAMDLCARRLLDPCEHQCDPETG
RREHRAAGECLCYEGYMKDPVHKHLCIRNEWGTNQGPWPYTIFQRGFDLVLGEQPSDKIF
RFTYTLGEGMWLPLSKSFVIPPAELAINPSAKCKTDMTVMEDAVEVREELMTSSSFDSLE
VLLDSFGPVRDCSKDNGGCSKNFRCISDRKLDSTGCVCPSGLSPMKDSSGCYDRHIGVDC
SDGFNGGCEQLCLQQMAPFPDDPTLYNILMFCGCIEDYKLGVDGRSCQLITETCPEGSDC
GESRELPMNQTLFGEMFFGYNNHSKEVAAGQVLKGTFRQNNFARGLDQQLPDGLVVATVP
LENQCLEEISEPTPDPDFLTGMVNFSEVSGYPVLQHWKVRSVMYHIKLNQVAISQALSNA
LHSLDGATSRADFVALLDQFGNHYIQEAIYGFEESCSIWYPNKQVQRRLWLEYEDISKGN
SPSDESEERERDPKVLTFPEYITSLSDSGTKHMAAGVRMECHSKGRCPSSCPLCHVTSSP
DTPAEPVLLEVTKAAPIYELVTNNQTQRLLQEATMSSLWCSGTGDVIEDWCRCDSTAFGA
DGLPTCAPLPQPVLRLSTVHEPSSTLVVLEWEHSEPPIGVQIVDYLLRQEKVTDRMDHSK
VETETVLSFVDDIISGAKSPCAMPSQVPDKQLTTISLIIRCLEPDTIYMFTLWGVDNTGR
RSRPSDVIVKTPCPVVDDVKAQEIADKIYNLFNGYTSGKEQQTAYNTLLDLGSPTLHRVL
YHYNQHYESFGEFTWRCEDELGPRKAGLILSQLGDLSSWCNGLLQEPKISLRRSSLKYLG
CRYSEIKPYGLDWAELSRDLRKTCEEQTLSIPYNDYGDSKEI
>sp|Q80Z10-3|ASTN2_MOUSE Isoform 3 of Astrotactin-2 OS=Mus musculus
OX=10090 GN=Astn2 (lacks exon 4) 1,300 aa.
MAAAGARRSPGRGLGLRGRPRLGFHPGPPPPPPPPLLLLFLLLLPPPPLLAGATAAAASR
EPDSPCRLKTVTVSTLPALRESDIGWSGARTGAAAGAGAGTGAGAGAAAAAASAASPGSA
GSAGTAAESRLLLFVRNELPGRIAVQDDLDNTELPFFTLEMSGTAADISLVHWRQQWLEN
GTLYFHVSMSSSGQLAQATAPTLQEPSEIVEEQMHILHISVMGGLIALLLLLLVFTVALY
AQRRWQKRRRIPQKSASTEATHEIHYIPSVLLGPQARESFRSSRLQTHNSVIGVPIRETP
ILDDYDYEEEEEPPRRANHVSREDEFGSQMTHALDSLGRPGEEKVEFEKKGGISFGRTKG
TSGSEADDETQLTFYTEQYRSRRRSKGLLKSPVNKTALTLIAVSSCILAMVCGNQMSCPL
TVKVTLHVPEHFIADGSSFVVSEGSYLDISDWLNPAKLSLYYQINATSPWVRDLCGQRTT
DACEQLCDPDTGECSCHEGYAPDPVHRHLCVRSDWGQSEGPWPYTTLERGYDLVTGEQAP
EKILRSTFSLGQGLWLPVSKSFVVPPVELSINPLASCKTDVLVTEDPADVREEAMLSTYF
ETINDLLSSFGPVRDCSRNNGGCTRNFKCVSDRQVDSSGCVCPEELKPMKDGSGCYDHSK
GIDCSDGFNGGCEQLCLQQTLPLPYDTTSSTIFMFCGCVEEYKLAPDGKSCLMLSDVCEG
PKCLKPDSKFNDTLFGEMLHGYNNRTQHVNQGQVFQMTFRENNFIKDFPQLADGLLVIPL
PVEEQCRGVLSEPLPDLQLLTGDIRYDEAMGYPMVQQWRVRSNLYRVKLSTITLSAGFTN
VLKILTKESSRDELLSFIQHYGSHYIAEALYGSELTCIIHFPSKKVQQQLWLQYQKETTE
LGSKKELKSMPFITYLSGLLTAQMLSDDQLISGVEIRCEEKGRCPSTCHLCRRPGKEQLS
PTPVLLEINRVVPLYTLIQDNGTKEAFKNALMSSYWCSGKGDVIDDWCRCDLSAFDASGL
PNCSPLPQPVLRLSPTVEPSSTVVSLEWVDVQPAIGTKVSDYILQHKKVDEYTDTDLYTG
EFLSFADDLLSGLGTSCVAAGRSHGEVPEVSIYSVIFKCLEPDGLYKFTLYAVDTRGRHS
ELSTVTLRTACPLVDDNKAEEIADKIYNLYNGYTSGKEQQTAYNTLMEVSASMLFRVQHH
YNSHYEKFGDFVWRSEDELGPRKAHLILRRLERVSSHCSSLLRSAYIQSRVDTIPYLFCR
SEEVRPAGMVWYSILKDTKITCEEKMVSMARNTYGETKGR
Supporting Information Figure S1. Amino acid sequences of the splice variants of Astn1 and
Astn2 used in this study. Hydrophobic regions identified by TOPPRED are shown in yellow,
potential acceptor sites for N-linked glycosylation in red, confirmed signal peptidase
cleavage sites in bold, and Met residues used as start codons in N-terminally truncated
versions in green.
Supporting Information Figure S2. SignalP 4.1 (Nature methods 8, 785-786 ) predicts signal peptidase-catalyzed
cleavage (the peak in the C-score) of Astn2 after the SP (Ala51-Gly52; top) and after TMH2 (Gly412-Asn413;
bottom). Both sites agree with results obtained by Edman sequencing of the N- and C-terminal fragments o
Astn2 (J Biol Chem 292, 3506-3516 ).
| 2019 | Murine astrotactins 1 and 2 have similar membrane topology and mature via endoproteolytic cleavage catalyzed by signal peptidase | 10.1101/493858 | [
"Lara Patricia",
"Tellgren-Roth Åsa",
"Behesti Hourinaz",
"Horn Zachi",
"Schiller Nina",
"Enquist Karl",
"Cammenberg Malin",
"Liljenström Amanda",
"Hatten Mary E.",
"von Heijne Gunnar",
"Nilsson IngMarie"
] | creative-commons |
The potential role of collagens in congenital Zika
syndrome: A systems biology approach
Renato S Aguiar1,2*, Fabio Pohl3*, Guilherme L Morais4*, Fabio CS Nogueira5,6*,
Joseane B Carvalho4*, Letícia Guida7*, Luis WP Arge4, Adriana Melo8, Maria EL
Moreira7, Daniela P Cunha7, Leonardo Gomes7, Elyzabeth A Portari7, Erika
Velasquez5, Rafael D Melani5, Paula Pezzuto1, Fernanda L de Castro1, Victor EV
Geddes1, Alexandra L Gerber4, Girlene S Azevedo8, Bruno L Schamber-Reis9,
Alessandro L Gonçalves1, Inácio Junqueira-de-Azevedo10, Milton Y Nishiyama-
Jr10, Paulo L Ho11, Alessandra S Schanoski11, Viviane Schuch3, Amilcar Tanuri1,
Leila Chimelli12, Zilton FM Vasconcelos7‡, Gilberto B Domont5‡, Ana TR
Vasconcelos4‡, Helder I Nakaya3‡.
1Departamento de Genética, Instituto de Biologia, Universidade Federal do Rio
de Janeiro, Rio de Janeiro, Brazil.
2Departamento de Biologia Geral, Instituto de Ciências Biológicas, Universidade
Federal de Minas Gerais, Belo Horizonte, Brazil.
3Department of Clinical and Toxicological Analyses, School of Pharmaceutical
Sciences, University of São Paulo, São Paulo, Brazil.
4National Laboratory of Scientific Computation, LNCC/MCTI, Petrópolis, Brazil.
5Proteomics Unit, Department of Biochemistry, Institute of Chemistry, Federal
University of Rio de Janeiro, Brazil
6Laboratory of Proteomics, LADETEC, Institute of Chemistry, Federal University
of Rio de Janeiro, Brazil
7Fernandes Figueira Institute, Fiocruz, Rio de Janeiro, Brazil.
8Instituto de Pesquisa Professor Amorim Neto, Campina Grande, Paraíba, Brazil.
9Faculdade de Ciências Médicas de Campina Grande, Núcleo de Genética
Médica, Centro Universitário UniFacisa, Campina Grande, Paraíba, Brazil
10Special Laboratory for Applied Toxinology, Butantan Institute, Brazil
11Bacteriology Laboratory, Butantan Institute, Brazil
12Laboratório de Neuropatologia, Instituto Estadual do Cérebro, Rio de Janeiro,
Brazil.
* These authors contributed equally to this work.
‡ Correspondence should be addressed to:
• Helder I Nakaya (hnakaya@usp.br)
• Ana Tereza Ribeiro de Vasconcelos (atrv@lncc.br)
• Gilberto B Domont (gilbertodomont@gmail.com)
• Zilton Farias Meira de Vasconcelos (zilton.vasconcelos@iff.fiocruz.br)
Abstract
Zika virus (ZIKV) infection during pregnancy could cause a set of severe
abnormalities in the fetus known as congenital Zika syndrome (CZS).
Experiments using animal models and in vitro systems significantly contributed
to our understanding of the physiopathology of ZIKV infection. However, the
molecular basis of CZS is not yet studied in humans. Here, we used a systems
biology approach to integrate transcriptomic, proteomic and genomic data from
post-mortem brains of neonates with CZS. We observed that collagen genes
were greatly reduced in CZS brains at both the RNA and protein levels and that
neonates with CZS have several polymorphisms in collagen genes associated
with osteogenesis imperfect and arthrogryposis. These findings were validated
using immunohistochemistry and collagen staining of ZIKV infected and non-
infected samples. Additionally, it was found that cell adhesion genes that are
essential for neurite outgrowth and axon guidance were up-regulated and thereby
confirmed the neuronal migration defects observed. This work provided new
insights into the underlying mechanisms of CZS and revealed host genes
associated with CZS susceptibility.
Keywords: Zika; Microcephaly; Systems Biology; Collagen
Introduction
Zika virus (ZIKV) infection during pregnancy is associated with several
neurological problems in the fetus1,2. Collectively, the set of abnormalities is
known as congenital Zika syndrome (CZS) and could involve microcephaly, brain
calcifications, ventriculomegaly, cortical malformations due to migration disorders
including
agyria/lissencephaly,
congenital
contractures,
and
ocular
abnormalities1,3,4. In adults, the most common symptoms of ZIKV infection are
fever, rash, arthralgia, conjunctivitis, and headache5. Although most pregnant
women exposed to ZIKV give birth to healthy babies, 0.3–15% of cases develop
CZS6. The frequency of infant deaths (miscarriages and perinatal deaths) is low
(~1% of CZS), and most of them present intrauterine akinesia syndrome
(arthrogryposis)7.
Several studies in vitro and with brain organoids and neurospheres
demonstrated that ZIKV could directly infect human neural progenitor cells8,
impairs the cortical development9, affects neuron migration impacting brain
size10, and promotes brain malformation11. Nevertheless, the molecular basis of
CZS and susceptibility genes associated with the most severe cases in human
newborns remains unknown.
Systems biology approaches were successfully applied to reveal the
molecular mechanisms associated with viral infection and vaccination12,13. By
integrating different types of omics data, systems biology provides a global
overview of the network of genes, transcripts, proteins, and metabolites involved
with a biological condition or perturbation14. When applied to human infectious
diseases, it could provide critical insights into the complex interplay between
pathogen and host, thereby leading to potential novel intervention strategies.
In this study, we have generated genomic, transcriptomic and proteomic
data from the blood and post-mortem brain samples of eight neonates with
confirmed ZIKV infection during pregnancy and with no congenital genetic
diseases nor another STORCH group vertical transmission. After three-layer
omics data integration, we highlighted the molecular pathways underlying
neurological damage. Systems biology combined with histopathological analysis
revealed that genes associated with matrix organization were dramatically down-
regulated in the brain of neonates with CZS, which could explain the neuronal
migration disorders and microcephaly attributed to ZIKV infection.
Results
Neonates with severe CZS
From October 2015 to July 2016 we followed a group of pregnant women
with symptoms of ZIKV exposition at distinct weeks of gestation and from two
endemic areas in Brazil—northeast (Campina Grande, Paraiba state) and
southeast (Rio de Janeiro, Rio de Janeiro state) regions. During this period, we
enrolled pregnant women who were referred to public healthcare with a history of
rash or fetus with central nervous system (CNS) abnormality confirmed by
ultrasonography or magnetic resonance imaging, as well as postnatal physical
examination suggestive of microcephaly.
We focused on eight neonates that had died in the first 48 hours
postpartum with severe arthrogryposis (Figure 1a). ZIKV genome was detected
in all cases during pregnancy by RT-PCR in clinical samples from mothers and
the neonates such as urine, plasma, amniotic fluid, placenta, and umbilical cord.
We also detected the virus genome through RT-PCR and in situ hybridization in
fetal post-mortem tissues (Figure 1b). Other microcephaly causes including
congenital genetic diseases, infection with arboviruses that circulate in the same
area (Dengue and Chikungunya) and teratogenic pathogens (STORCH) were all
excluded (Table S1).
Five out of eight cases of CSZ showed ZIKV exposition symptoms in the
first trimester of pregnancy corroborating with other reports15,16 that describe
increasing risk of microcephaly at the beginning of gestation (Figure 1a).
Microcephaly
was
observed
in
the
early
gestation
weeks
through
ultrasonography in all cases. However, the cephalic perimeter at birth was
considered normal (higher than 32 cm) in most of the neonates due to severe
ventriculomegaly/obstructive hydrocephalus. The brain usually collapsed after
removal of the skull during autopsy showing tiny brains in all cases (on average
66 grams; ranging from 7 to 180 grams). A detailed neuropathological description
of all cases has been previously reported10.
Figure 1. Clinical diagnoses and brain damage of deceased neonates with
CZS. (a) Gestation timeline for the eight neonates with CZS. The symptoms
include fever, exanthema, arthralgia, conjunctivitis, and headache in pregnant
women during gestation. (b) Zika genome detection by RT-PCR from post-
mortem brain samples expressed in CT values. (c) Lesions in the central nervous
system (CNS) of neonates with CZS investigated by prenatal ultrasound and MRI
examinations; *At birth only 3 cases (Z04,Z05, Z07) had microcephaly. The
others had normal or enlarged cephalic perimeter due to obstructive
hydrocephalus;
**Cerebellum
and
Brainstem
hypoplasia;
***cortical
malformations due to neuronal migration disturbance (agyria, polymicrogyria or
lissencephaly) (d) Brains from autopsies showing various degrees of lesions,
including collapse due to hydrocephalus and small brains with few gyri or agyria
and consistently with severe loss of CNS structures and congested
leptomeninges. The numbers of Zika cases are depicted as presented in Table
S1.
The brains with higher viral load (lower cycle threshold or CT values)
exhibited the most destructive patterns of CNS structures (Figure 1b and Table
S1). Macroscopic observations showed thickened and congested leptomeninges,
very thin parenchyma and corpus callosum, and asymmetric ventriculomegaly
(Figure 1d). Shallow sulci or agyria was prevalent in all cases (Figure 1d). The
hippocampus, basal ganglia, and thalami were usually not well identified and
malformed. Cerebellar hypoplasia was observed in all cases, with an irregular
cortical surface and calcification foci were detected macroscopically (Figure 1d).
The brainstem was deformed and hypoplastic in most of the cases.
The histopathological analysis confirmed the migration disturbances
represented by abnormal immature cell clusters along the white matter and over
pia mater (Table S2). An intense immune response to cell injury was observed in
all cases as demonstrated by the gliosis and inflammatory infiltrate (T-
lymphocytes and histiocytes) in the meninges, cerebral hemispheres, and spinal
cord (Tables S1 and S2). Reduction of the descending motor fibers was also
observed. The histopathological analysis also displayed a loss of motor nerve
cells in the spinal cord and atrophy of the skeletal muscle. These could explain
the intrauterine akinesia and consequent arthrogryposis observed in all cases
(Tables S1 and S2).
Transcriptome and Proteome analyses of CZS Brains
We utilized high-throughput sequencing and mass spectrometry
technologies to assess the changes in the transcriptome and proteome of CZS
brains (Z03, Z05, and Z08 in Figure 1) compared to the control brain
(Edwards´syndrome). Differential expression analysis revealed 509 genes
associated with CZS, of which 228 were up-regulated and 281 were down-
regulated in ZIKV-infected neonates (Figure S1a and Table S3). Among the
pathways enriched with up-regulated genes, we found the “Unblocking of NMDA
Receptor, Glutamate Binding, and Activation” and “Glutamate Neurotransmitter
Release Cycle” (Figure S1a and Table S3). These findings support our previous
in vitro work showing that the blockage of the NMDA receptor prevents the
neuronal death induced by ZIKV infection17. Among the pathways enriched with
down-regulated genes, we found collagen formation, glucose metabolism,
signaling by TGF-beta receptor complex, Class I MHC mediated antigen
processing and presentation, and amyloid fiber formation (Figure S1b). These
down-regulated genes indicate that ZIKV infection could affect immune-response
pathways, cellular metabolism and the very formation of connective tissue in the
brain.
Figure S1. Modulation of brain-expressed genes in neonates with CZS
evidenced by transcriptome. Genes and pathways up-regulated (a) or down-
regulated (b) in CZS compared to ZIKV negative control brain in the prefrontal
cortex. The pathways enriched by Over Representation Analysis (left) are present
in the outermost layer and the differentially expressed genes (right) found in these
pathways (up-regulated in red and down-regulated in blue). In the innermost
layer, the links indicate the pathway in which the genes were found. In the middle
layer, the colors in the heat map represent the pathway enrichment P-value
obtained by Over Representation Analysis, while the line graph represents the
log2 fold change value for each gene in the CZS samples relative to the control.
Cell adhesion genes—essential for neuronal migration and recruiting of
immune cells including NCAM receptors—are up-regulated in the CZS brains,
which corroborates the migration disturbance and inflammatory infiltration events
(CD8+ T-lymphocytes and CD68+ histiocytes) observed in the histopathological
analysis (Table S3). Collagen genes (COL1A1, COL1A2, COL3A1, COL5A1,
COL5A2, COL6A3, COL12A1, and COL14A1) essential for the development of
the brain and the blood-brain barrier18 are down-regulated in the CZS brains.
We subsequently investigated the protein levels in CZS brains compared
to the ZIKV negative control. The proteomic analysis identified 252 and 110
proteins up- and down-regulated in CZS brain respectively (Figure 2a and Table
S4). Furthermore, a set of proteins were exclusively detected either in brains with
CZS (714 proteins) or in the ZIKV negative control brain (79 proteins) (Figure 2a
and Table S4). Similar to the transcriptomic analysis, up-regulated proteins were
enriched for “glucose metabolism” and “L1CAM interactions,” whereas down-
regulated proteins were enriched for “extracellular matrix organization” and
“collagen formation” (Figure 2b and Table S4). Among the proteins down-
regulated in all three neonates with CZS compared to control, COL1A1, COL1A2,
PPIB, SERPINH1, and OGN were found. While PPIB is instrumental in collagen
trimerization, SERPINH1 is critical to collagen biosynthesis19. Additionally, the
functions of OGN in the extracellular matrix are related to collagen fibrillogenesis,
cell proliferation, and development, as well as osteoblast differentiation and bone
development20.
Figure 2. Proteins related to CZS and microcephaly. (a) The number of
differentially expressed proteins up- (red) and down-regulated (blue). Proteins
with adjusted p-value < 0.05 were considered differentially expressed. Proteins
expressed only in one condition were considered exclusively expressed proteins.
(b) Enrichment of functional pathways for proteins found in CZS brains. Red
represents the up-regulated proteins, and blue represents the down-regulated
proteins. Adjusted p-value (-log10) of Over Representation Analysis is indicated
by color intensity and circle size. (c) Protein-protein interactions for down-
regulated proteins. Blue nodes indicate the down-regulated proteins observed
and grey nodes indicate the additional proteins. The circle size represents the
node degree.
Protein-protein interaction data obtained from the STRING database was
used to assess the interacting proteins related to the brain’s connective tissue
(Figure 2c). The interactome showed the same pattern of the down-regulation of
essential proteins hubs involved in collagen formation (COL1A1 and COL1A2)
and adhesive glycoprotein that mediates cell-to-cell and cell-to-matrix
interactions (ITGA2B, NCAM, FNB, IGB1, and THBS1). LOX is down-regulated
in both transcriptome and proteome analysis and plays a key role in cross-linking
fibers of collagen and elastin. The down-regulation of collagen pathways in the
brain endothelia could partially explain the vascular problems and ischemia
events observed in CZS neonates. Once again, proteomics analysis validates
transcriptomics as well as the macroscopic and microscopic images showing the
modulation of proteins involved in brain architecture matrix and neuronal
migration disorders in ZIKV affected neonates. The interactome showed the
modulation of fibrinogen components (FGA, FGB, and FGG), which are
components of blood clots and are formed following vascular injury. These
findings relate to the intense blood congested leptomeninges found in the CSZ
brains (Figure 2c and 1d).
We subsequently cross-referenced the lists of genes and proteins that
were differentially expressed in CZS compared to the negative control and found,
respectively, 12 and 23 up- and down-regulated shared genes and proteins
(Figure 3a). The functions of several of these genes could provide insights into
ZIKV neuropathogenesis. For instance, NCAM1 is essential for neurite
outgrowth, COL1A1 and COL1A2 genes encode the alpha 1 and 2 chains of
collagen type I, and PRDX2 regulates the antiviral activity of T cells (Figure 3a).
For TTR and AGT genes, however, the levels of the RNA were higher in CZS
than in control (up-regulated at RNA level) whereas the protein levels were lower
in CZS (down-regulated at protein level). Similarly, eight genes were up-regulated
in proteomics but down-regulated in transcriptomics. These inverted patterns
between RNA and protein levels could partially be due to post-transcriptional
regulation mechanisms that include miRNAs. Thus, we checked whether genes
that were up-regulated in transcriptomics but not up-regulated in proteomics
dataset were known miRNA targets. Our in silico approach predicted that eight
miRNAs were induced upon infection and possibly involved regulation of genes
related to CZS (Figure 3b). Among them, mir-17-5p was already shown to be
induced by flavivirus infections, including ZIKV infection in astrocytes21.
Figure 3. Transcriptomics and proteomics interplay in CZS. (a) The overlap
between differentially expressed genes (DEGs) and differentially expressed
proteins (DEPs) in CZS. The fraction of up- and down-regulated genes/proteins
are represented by orange and violet bars respectively. The links represent the
overlap between both DEGs and DEPs. The dashed lines indicate overlapped
genes with arc correspondent colors. (b) miRNAs predicted to regulate the genes
up-regulated in the transcriptomic dataset but not up-regulated in proteomics
dataset. (c) Pathways enriched in transcriptomics and proteomics datasets.
Numbers in red and blue are pathways enriched with up- and down-regulated
genes respectively. Dashed lines indicate common pathways.
When considering the proteins that were exclusively found either in brain
samples with CZS or in the ZIKV negative control brain, 99 shared up- and down-
regulated genes and proteins were observed (Tables S3 and S4). These included
genes such as LOX, PSMF1, NCAN, TNR, and NRCAM, which are associated
with crosslinking of collagen and elastin, processing of class I MHC peptides,
modulation of cell adhesion and migration, and neuronal cell adhesion.
We also integrated transcriptomics and proteomics at the pathway level.
Gene Set Enrichment Analysis (GSEA) was performed using the mean
foldchange between CZS and control brains as ranks and the Reactome
pathways as gene sets. It was observed that a higher overlap between
transcriptomics and proteomics with 47 pathways significantly enriched for both
layers of information (Figure 3c). Furthermore, down-regulated pathways were
again related to extracellular matrix organization and collagen formation and point
to the central role of collagen in CZS outcome.
Genetic variants associated with CZS
Whole exome sequencing analysis identified several rare variants with
potential deleterious functions in five neonates (Z01, Z02, Z04, Z06, and Z07 in
Figure 1). Combining variants that are presented in the same gene, it was found
that 23 genes have at least one single nucleotide polymorphisms (SNP) in all five
neonates (Figure 4a and Table S5). Variants in genes associated with
extracellular matrix organization (collagen genes, FBN2, FBN3, and FN1) (Figure
4b and Table S5), as well as CNS development (PTPRZ1), immune system (C7,
C8A, IL4R, IL7, IRF3, and TLR2), muscular contraction and arthrogryposis
(PIEZO2, RYR1 and TTN), and Notch and Wnt signaling pathways (NOTCH3,
NOTCH4 and VANGL1) were also found (Table S5).
Figure 4. Single nucleotide polymorphisms in neonates with CZS. The
exome analysis of five CZS cases (a) Genomic map showing genes with SNPs
(MAF < 0.05 and CADD > 15) in three or more neonates with CZS. The outermost
layer represents the reference genome (GRCh38). In the middle layer, each row
represents genes with at least one SNP in three to five neonates. Dark brown
rows represent genes that contain variants in all five neonates. (b) Most
deleterious SNPs found in extracellular matrix genes.
Integration of 3 Omics data types
The three layers of biological information were ultimately integrated into a
network containing the gene variants and the RNAs and proteins differentially
expressed in CZS cases (Figure 5). Only three genes appeared associated with
CZS in all the layers—COL1A1, COL12A1, and PTPRZ1 (Figure 5a). The former
two collagen genes are related to the extracellular matrix organization. The latter
gene PTPRZ1 is instrumental to the differentiation of oligodendrocytes22 and
have been associated with schizophrenia23. In total, there were 1,628 genes
associated with CZS at either the genomic, transcriptomic, or proteomic level.
Protein-protein interaction (PPI) data was used to construct a network with 341
of these genes (Figure 5b). Network analysis revealed several modules
associated with proteasome degradation, axon guidance, the FGF signaling
pathway, and Parkinson’s disease (Figure 5b).
Figure 5. Integration of three molecular layers in post-mortem ZIKV-infected
samples infected. (a) Intersections between three layers of information
(genomics, transcriptomics, and proteomics) involved with CZS. Point diagram at
the bottom represents the intersections between layers. Bar plot shows the
number of genes in each intersection. A dashed line indicates the genes present
in all the layers. (b) Protein-protein interaction network of CZS-related genes and
their cellular pathways.
A more stringent analysis was performed considering only the 64 genes
that were identified associated with CZS in at least two omics analyses (Figure
5a). Subsequently, these genes were integrated into a PPI network (Figure S2).
Several central genes were observed—THBS1 promotes synaptogenesis24; DCN
regulates collagen fibrils and matrix assembly25, and CLU shifts blood-brain
barrier amyloid-beta drainage26.
Figure S2. A network of highly associated CZS-related genes. More stringent
criteria (at least two layers of biological information) was used to select the genes
to construct the protein-protein interaction network.
Since all the analyses indicate that collagen genes are down-regulated in
CZS brains, we performed a Gomori’s trichrome staining for total collagen in CZS
brain, as well as in a different set of Zika negative control brains. We observed a
reduction of collagen fibers in the CZS brains particularly in the adventitia of the
vessels compared to Zika negative controls at the same gestational age. This
reduction validated our transcriptome and proteome findings (Figure 6a). Next,
the presence of COL1A1 was investigated through immunostaining directly in the
brain tissues from CZS cases relative to the controls, which also showed less
COL1A1 in all the CZS cases (Figure 6b). This corroborated the role of collagen
isoforms in the neuropathogenesis associated with ZIKV infection in the brain
tissues (Figure 6b).
Figure 6. Reduction of collagen fibers in CZS cases compared with Zika
negative controls. (a) Histopathological analysis confirms that the CZS brains
have fewer collagen fibers compared to negative ZIKV control brains at the same
gestational age. The total collagen that stains in green with the Gomori's
Trichrome is less evident than in controls, particularly in the adventitia of the
vessels. (b) Immuno-histochemistry for collagen 1 also shows less collagen in
CZS brains. Controls and CSZ cases are depicted as tables S1 and S3.
Discussion
Our findings indicate that collagen genes and the extracellular matrix could
play a significant role in CZS. Reduced levels of fibronectin and collagen IV
increase the permeability of the blood-brain barrier27. Once this barrier is
transposed, ZIKV could reach developing neural progenitor cells and severely
disrupt the neural development. However, a more direct effect on fibroblast cells
in the surrounding vasculature28 could not be discarded, and this effect could be
a result of cell death or dysregulation of ECM expression or tissue deposition.
The unique description of ECM gene modulation and ZIKV were reported in
monkey experimental model during CNS viral persistence. Aid et al. showed that
viral loads and viral persistence were negatively correlated with ECM genes,
including collagen family genes29. Experiments using animal models indicate that
deficiency in collagen compromise vessel resistance30. Mutations in collagen IV
and fibronectin have induced impaired basement membranes or mesoderm
defects respectively31. Moreover, mutations on COL1A and COL4A1 caused
defects in the basal membrane, resulting in a weakening of the brain vessels,
arterial rupture and ischemic stroke32,33. Along with collagen isoforms, the down-
regulation of the LOX gene that is responsible for cross-linking collagen fibers to
elastin could potentialize the vascularity deficiency. This could explain the blood
congestion in leptomeninges observed in all the brain samples analyzed here.
Specifically, glycine mutations affecting exon 49 of the COL1A2 gene was
associated with an increased risk of intracranial bleeding34. Both collagen and
LOX genes are stimulated in glioblastoma cells, and the suppression of this
pathway by ZIKV infection could explain the decreasing of angiogenesis and anti-
cancer effects that several authors are exploring to treat glioblastoma35-37 with
ZIKV-like particles.
Mutations in type I collagen also affect the extracellular matrix by
decreasing the amount of secreted collagen(s) impairing molecular and
supramolecular assembly through the secretion of mutant collagen or by inducing
endoplasmic reticulum stress and the unfolded protein response38. Mutated
COL1A1 were also associated with osteogenesis imperfecta, a generalized
disorder of connective tissues that resembles the observed arthrogryposis
phenotype common to all cases included in this work39. Mutations in COL1A1/2
genes were associated with congenital brittle bones with the development of
microcephaly and cataracts, as observed in the most severe cases of CSZ40. A
dominant mutation in COL12A1 was also related to joint laxity41, a phenotype
often found in ZIKV-infected children42.
Cell-cell interaction is necessary for neuron migration through cortex
layers during neurodevelopment43. L1CAM family of cell adhesion molecules are
associated with neurite outgrowth and axon guidance44. In ZIKV-infected brains,
NCAM1 and NFASC were up-regulated both at the RNA and protein levels. In
addition, we found a rare variant in Neuronal Cell Adhesion Molecule gene
(NRCAM) that could corroborate with ZIKV-infected brains phenotypes. These
findings indicate that those genes/proteins could be the molecular basis for
neurons migration defects already described by our group10 and should lead to
CNS structural defects and reduction of cortical region observed in CZS
newborns.
Among the pathways enriched for the up-regulated genes in ZIKV-infected
samples, we found genes related to glutamate neurotransmitter release cycle and
unblocking of NMDA receptor, glutamate binding, and activation. Previous
experimental work revealed that NMDA receptor blockage has a protective effect
on ZIKV-induced cell death17. In addition to gene expression, it was found that
the genes associated with apoptosis were also up-regulated at the protein level.
This corroborates the increased cell death proposed to neural progenitor cell pool
and revealed by experimental data45.
Successful viral infection and disease must overcome the organism
immune response. Pleiotrophin (PTN) is a cytokine that modulates inflammation
in the CNS46. Additionally, PTN negatively regulates protein tyrosine
phosphatase zeta (PTPRZ1), which binds to developmental proteins such as
beta-catenin47. The results of this study showed that PTPRZ1 was up-regulated
in ZIKV-infected brains both in the RNA and protein levels. Impressively, this
gene also presented rare polymorphisms associated that raises the possibility of
PTN–PTPRZ1 regulatory dysregulation and genetically driven suppression of
neuroinflammation, which might result in a viral evasion mechanism. Considering
the exclusive proteins (Table S3), another gene found in all three omics layers
was NRCAM. This gene is a cell adhesion molecule that can interact with
PTPRZ148.
We also observed rare mutations in genes related to the immune system,
including IRF3 and IL4R. IRF3 plays a critical role in the innate immune response
against DNA and RNA viruses, driving the transcription of type-I IFN genes49.
Additionally, a mutation in the IRF3 gene was associated with increased
susceptibility to HSV-1 infection in the CNS in humans50. SNPs in the interleukin-
4 receptor (IL4R) were also associated with increased susceptibility to dengue51.
Our findings indicate that mutations in those genes could also confer increased
susceptibility to ZIKV infection and CZS.
Altogether, this work is the first to investigate the molecular basis of ZIKV
infection after vertical transmission using post-mortem brain samples. Despite the
small sample size, these brain samples are unique considering the decrease in
CZS cases worldwide. Our systems biology approach allowed us to unveil the
different layers of biological information associated with CZS.
Acknowledgments
We would like to thank Diego Santos and Heliomar Pereira Marcos who
performed the collagen staining and immunohistochemistry. R.S.A. and A.M. are
grateful to Biometrix and Dia.Pro Diagnostic Bioprobes, for the donation of ELISA
kits for this project.
Author contributions
EV and RDM performed the proteomics experiments, FCSN and GBD
developed the proteomics rational, did data search, physiological analysis. FP,
IJA, MYNJ, PLH, ASS, VS, HIN performed and analyzed the transcriptomics
experiments. GLM, JBC, ALG, LWPA performed and analyzed the exome
experiments. LC performed the neuropathological analysis (collagen staining and
immuno-histochemistry). RSA, FP, ZFMV, GBD, ATRV, HIN integrated the omics
datasets, interpreted the results and wrote the initial draft. All authors helped with
the writing of the manuscript.
Funding
This research was supported by (CNPq – grant# 440900/2016-6), and
(CAPES – grant# 88881.130757/2016-01) and (FINEP – grant# 01.16.0078.00).
HIN is supported by CNPq and the São Paulo Research Foundation (FAPESP;
grants 2017/50137-3, 2012/19278-6, and 2013/08216-2). FLC, VEVG, JBC,and
LWPA. are funded by CAPES (grant 001). GBD, has financial support from
grants 88887.130697 (CAPES) and 440613/2016-7 (CNPq). FCSN is supported
by FAPERJ (E-26/202.650/2018). ATRV is supported by CNPq (303170/2017-4)
and FAPERJ (26/202.903/20).
References
1
Melo, A. S. D. et al. Congenital Zika Virus Infection Beyond Neonatal Microcephaly. Jama
Neurol 73, 1407-1416, doi:10.1001/jamaneurol.2016.3720 (2016).
2
Mlakar, J. et al. Zika Virus Associated with Microcephaly. New Engl J Med 374, 951-958,
doi:10.1056/NEJMoa1600651 (2016).
3
Ventura, C. V., Maia, M., Bravo-Filho, V., Gois, A. L. & Belfort, R. Zika virus in Brazil and
macular atrophy in a child with microcephaly. Lancet 387, 228-228, doi:10.1016/S0140-
6736(16)00006-4 (2016).
4
Brasil, P. et al. Zika Virus Infection in Pregnant Women in Rio de Janeiro. N Engl J Med
375, 2321-2334, doi:10.1056/NEJMoa1602412 (2016).
5
Duffy, M. R. et al. Zika virus outbreak on Yap Island, Federated States of Micronesia. N
Engl J Med 360, 2536-2543, doi:10.1056/NEJMoa0805715 (2009).
6
Johansson, M. A., Mier-y-Teran-Romero, L., Reefhuis, J., Gilboa, S. M. & Hills, S. L. Zika
and the Risk of Microcephaly. N Engl J Med 375, 1-4, doi:10.1056/NEJMp1605367
(2016).
7
Moore, C. A. et al. Characterizing the Pattern of Anomalies in Congenital Zika Syndrome
for
Pediatric
Clinicians.
JAMA
Pediatr
171,
288-295,
doi:10.1001/jamapediatrics.2016.3982 (2017).
8
Tang, H. L. et al. Zika Virus Infects Human Cortical Neural Progenitors and Attenuates
Their Growth. Cell Stem Cell 18, 587-590, doi:10.1016/j.stem.2016.02.016 (2016).
9
Shao, Q. et al. Zika virus infection disrupts neurovascular development and results in
postnatal microcephaly with brain damage. Development 143, 4127-4136,
doi:10.1242/dev.143768 (2016).
10
Chimelli, L. et al. The spectrum of neuropathological changes associated with congenital
Zika virus infection. Acta Neuropathol 133, 983-999, doi:10.1007/s00401-017-1699-5
(2017).
11
Garcez, P. P. et al. Zika virus disrupts molecular fingerprinting of human neurospheres.
Sci Rep 7, 40780, doi:10.1038/srep40780 (2017).
12
Kwissa, M. et al. Dengue virus infection induces expansion of a CD14(+)CD16(+)
monocyte population that stimulates plasmablast differentiation. Cell Host Microbe 16,
115-127, doi:10.1016/j.chom.2014.06.001 (2014).
13
Nakaya, H. I. & Pulendran, B. Vaccinology in the era of high-throughput biology. Philos
Trans R Soc Lond B Biol Sci 370, doi:10.1098/rstb.2014.0146 (2015).
14
Avey, S. et al. Multicohort analysis reveals baseline transcriptional predictors of
influenza vaccination responses. Sci Immunol 2, doi:UNSP eaa14656
10.1126/sciimmunol.aal4656 (2017).
15
Hoen, B. et al. Pregnancy Outcomes after ZIKV Infection in French Territories in the
Americas. N Engl J Med 378, 985-994, doi:10.1056/NEJMoa1709481 (2018).
16
Krow-Lucal, E. R. et al. Association and birth prevalence of microcephaly attributable to
Zika virus infection among infants in Paraiba, Brazil, in 2015-16: a case-control study.
Lancet Child Adolesc 2, 204-213, doi:10.1016/S2352-4642(18)30020-8 (2018).
17
Costa, V. V. et al. N-Methyl-d-Aspartate (NMDA) Receptor Blockade Prevents Neuronal
Death Induced by Zika Virus Infection. MBio 8, doi:10.1128/mBio.00350-17 (2017).
18
Obermeier, B., Daneman, R. & Ransohoff, R. M. Development, maintenance and
disruption of the blood-brain barrier. Nat Med 19, 1584-1596, doi:10.1038/nm.3407
(2013).
19
Mala, J. G. & Rose, C. Interactions of heat shock protein 47 with collagen and the stress
response:
an
unconventional
chaperone
model?
Life
Sci
87,
579-586,
doi:10.1016/j.lfs.2010.09.024 (2010).
20
Deckx, S., Heymans, S. & Papageorgiou, A. P. The diverse functions of osteoglycin: a
deceitful dwarf, or a master regulator of disease? FASEB J 30, 2651-2661,
doi:10.1096/fj.201500096R (2016).
21
Kozak, R. A. et al. MicroRNA and mRNA Dysregulation in Astrocytes Infected with Zika
Virus. Viruses 9, doi:10.3390/v9100297 (2017).
22
Kuboyama, K. et al. Protein tyrosine phosphatase receptor type z negatively regulates
oligodendrocyte
differentiation
and
myelination.
PLoS
One
7,
e48797,
doi:10.1371/journal.pone.0048797 (2012).
23
Buxbaum, J. D. et al. Molecular dissection of NRG1-ERBB4 signaling implicates PTPRZ1
as a potential schizophrenia susceptibility gene. Mol Psychiatr 13, 162-172,
doi:10.1038/sj.mp.4001991 (2008).
24
Christopherson, K. S. et al. Thrombospondins are astrocyte-secreted proteins that
promote CNS synaptogenesis. Cell 120, 421-433, doi:DOI 10.1016/j.cell.2004.12.020
(2005).
25
Robinson, K. A. et al. Decorin and biglycan are necessary for maintaining collagen fibril
structure, fiber realignment, and mechanical properties of mature tendons. Matrix Biol
64, 81-93, doi:10.1016/j.matbio.2017.08.004 (2017).
26
Wojtas, A. M. et al. Loss of clusterin shifts amyloid deposition to the cerebrovasculature
via disruption of perivascular drainage pathways. P Natl Acad Sci USA 114, E6962-E6971,
doi:10.1073/pnas.1701137114 (2017).
27
Baeten, K. M. & Akassoglou, K. Extracellular matrix and matrix receptors in blood-brain
barrier formation and stroke. Dev Neurobiol 71, 1018-1039, doi:10.1002/dneu.20954
(2011).
28
Hamel, R. et al. Biology of Zika Virus Infection in Human Skin Cells. J Virol 89, 8880-8896,
doi:10.1128/JVI.00354-15 (2015).
29
Aid, M. et al. Zika Virus Persistence in the Central Nervous System and Lymph Nodes of
Rhesus Monkeys. Cell 169, 610-620 e614, doi:10.1016/j.cell.2017.04.008 (2017).
30
Kuo, D. S., Labelle-Dumais, C. & Gould, D. B. COL4A1 and COL4A2 mutations and disease:
insights into pathogenic mechanisms and potential therapeutic targets. Hum Mol Genet
21, R97-110, doi:10.1093/hmg/dds346 (2012).
31
Poschl, E. et al. Collagen IV is essential for basement membrane stability but dispensable
for initiation of its assembly during early development. Development 131, 1619-1628,
doi:10.1242/dev.01037 (2004).
32
Dittrich, R. et al. Connective tissue and vascular phenotype in patients with cervical
artery dissection. Neurology 68, 2120-2124, doi:10.1212/01.wnl.0000264892.92538.a9
(2007).
33
Weng, Y. C. et al. COL4A1 mutations in patients with sporadic late-onset intracerebral
hemorrhage. Ann Neurol 71, 470-477, doi:10.1002/ana.22682 (2012).
34
Faqeih, E., Roughley, P., Glorieux, F. H. & Rauch, F. Osteogenesis imperfecta type III with
intracranial hemorrhage and brachydactyly associated with mutations in exon 49 of
COL1A2. Am J Med Genet A 149A, 461-465, doi:10.1002/ajmg.a.32653 (2009).
35
Zhu, Z. et al. Zika virus has oncolytic activity against glioblastoma stem cells (vol 214, pg
3145, 2017). Journal of Experimental Medicine 214 (2017).
36
Chen, Q. et al. Treatment of Human Glioblastoma with a Live Attenuated Zika Virus
Vaccine Candidate. MBio 9, doi:10.1128/mBio.01683-18 (2018).
37
Mammoto, T. et al. Role of collagen matrix in tumor angiogenesis and glioblastoma
multiforme progression. Am J Pathol 183, 1293-1305, doi:10.1016/j.ajpath.2013.06.026
(2013).
38
Bateman, J. F., Boot-Handford, R. P. & Lamande, S. R. Genetic diseases of connective
tissues: cellular and extracellular effects of ECM mutations. Nat Rev Genet 10, 173-183,
doi:10.1038/nrg2520 (2009).
39
Marini, J. C. et al. Consortium for osteogenesis imperfecta mutations in the helical
domain of type I collagen: regions rich in lethal mutations align with collagen binding
sites for integrins and proteoglycans. Hum Mutat 28, 209-221, doi:10.1002/humu.20429
(2007).
40
Plotkin, H. Syndromes with congenital brittle bones. BMC Pediatr 4, 16,
doi:10.1186/1471-2431-4-16 (2004).
41
Zou, Y. et al. Recessive and dominant mutations in COL12A1 cause a novel
EDS/myopathy overlap syndrome in humans and mice. Hum Mol Genet 23, 2339-2352,
doi:10.1093/hmg/ddt627 (2014).
42
Siqueira, W. L., Moffa, E. B., Mussi, M. C. M. & Machado, M. A. D. M. Zika virus infection
spread through saliva - a truth or myth? Braz Oral Res 30, doi:ARTN e46
10.1590/1807-3107BOR-2016.vol30.0046 (2016).
43
Nadarajah, B., Alifragis, P., Wong, R. O. L. & Parnavelas, J. G. Neuronal migration in the
developing cerebral cortex: Observations based on real-time imaging. Cereb Cortex 13,
607-611, doi:DOI 10.1093/cercor/13.6.607 (2003).
44
Maness, P. F. & Schachner, M. Neural recognition molecules of the immunoglobulin
superfamily: signaling transducers of axon guidance and neuronal migration (vol 10, pg
19, 2007). Nat Neurosci 10, 263-263, doi:DOI 10.1038/nn0207-263b (2007).
45
Cugola, F. R. et al. The Brazilian Zika virus strain causes birth defects in experimental
models. Nature 534, 267-271, doi:10.1038/nature18296 (2016).
46
Fernandez-Calle,
R.
et
al.
Pleiotrophin
regulates
microglia-mediated
neuroinflammation. J Neuroinflamm 14, doi:ARTN 46
10.1186/s12974-017-0823-8 (2017).
47
Pariser, H., Ezquerra, L., Herradon, G., Perez-Pinera, P. & Deuel, T. F. Fyn is a downstream
target of the pleiotrophin/receptor protein tyrosine phosphatase beta/xi-signaling
pathway: Regulation of tyrosine phosphorylation of Fyn by pleiotrophin. Biochem Bioph
Res Co 332, 664-669, doi:10.1016/j.bbrc.2005.05.007 (2005).
48
Sakurai, T. et al. Induction of neurite outgrowth through contactin and Nr-CAM by
extracellular regions of glial receptor tyrosine phosphatase beta. J Cell Biol 136, 907-
918, doi:DOI 10.1083/jcb.136.4.907 (1997).
49
Liu, S. et al. Phosphorylation of innate immune adaptor proteins MAVS, STING, and TRIF
induces IRF3 activation. Science 347, aaa2630, doi:10.1126/science.aaa2630 (2015).
50
Andersen, L. L. et al. Functional IRF3 deficiency in a patient with herpes simplex
encephalitis. J Exp Med 212, 1371-1379, doi:10.1084/jem.20142274 (2015).
51
Useche, Y. M. et al. Association of IL4R-rs1805016 and IL6R-rs8192284 polymorphisms
with clinical dengue in children from Colombian populations. J Infect Public Health,
doi:10.1016/j.jiph.2018.08.009 (2018).
Supplementary Material and methods
Patients and neuroimaging studies
From June 2015 to July 2016, pregnant women presenting acute febrile
illness clinic with a rash, fetal central nervous system (CNS) abnormalities at
prenatal ultrasonography (US), and/ or postnatal microcephaly or other CNS
malformation that was believed to be characteristic of congenital infection were
referred to the Microcephaly Reference Center IPESQ in Campina Grande
(Paraíba, Brazil) or Instituto Fernandes Figueira – Fiocruz (Rio de Janeiro,
Brazil). This study includes imaging and autopsy data from an institutional review
board–approved study (52888616.4.0000.5693 and 52675616.0.000.5269) that
allowed for imaging and follow-up of presumed Zika virus infection in pregnant
women and their neonates. Written informed consent was obtained from the
pregnant women and/or the parents of neonates. Detailed demographic, medical,
and prenatal history information, as well as clinical findings, were entered into
case-report forms by multidisciplinary medical teams. All women were referred
for at least one fetal ultrasonography during gestation. The onset symptoms
included fever, exanthema, arthralgia, conjunctivitis, and headache in the
pregnant women during gestation. The CNS of eight neonates who died in the
first 48 h of life (two of them immediately after delivery), three from northeastern
(Campina Grande, Paraíba state) and five from southeastern of Brazil (Rio de
Janeiro) whose mothers reported typical symptoms of ZIKV infection until the
18th gestational week, were examined postmortem. Intrauterine fetal
development was followed with ultrasonography and fetal MRI. Just after birth,
the cephalic perimeter was measured and the percentile was calculated
according to the expected for the gestational age 1. Prenatal US was performed
by fetal medicine specialists using either a Voluson E8 unit (General Electric,
Milwaukee, Wis) with transvaginal probes or a Samsung XG or WS80 unit
(Samsung, Seoul, South Korea) with 2–9- MHz probes. MR imaging of the fetus
was performed with a 3-T Skyra unit (Siemens Healthcare, Erlangen, Germany)
or a 1.5-T Espree unit (Siemens Healthcare) with an eight-channel body coil and
standard acquisition protocols. Postnatal head CT was performed with a 16-
section CT scanner (Siemens Healthcare). Postnatal MR imaging was performed
with a 1.5-T Espree brain MR imaging unit (Siemens Healthcare). Brain tissue
images were acquired with a 64-channel multisection CT scanner (GE
Healthcare) and a 3-T MR imaging unit (Achieva; Philips, Best, the Netherlands).
Autopsies
Full autopsies were performed and the brains were fixed in 10% buffered
formalin. In the three cases from IPESQ, one hemisphere was stored in RNA later
and then frozen for virus RNA detection and transcriptome/proteomic analysis. In
seven cases, the whole spinal cords were also removed, four of them with dorsal
nerve ganglia (DRG). The upper cervical spinal cord was also sampled in two
other cases, one with DRG. Formalin-fixed brains were weighed and the
percentile was calculated according to the expected for the gestational age 1. In
addition, samples from skeletal muscle (paravertebral, psoas, diaphragm or
adjacent to the head of the femur) were taken and examined histologically in five
cases. After macroscopic examination, representative areas, including those with
macroscopic lesions, were processed for paraffin embedding and 5 μm
histological sections were stained with hematoxylin and eosin (H&E). The
neuropathological findings of these patients have been reported previously 2.
Brains of ZIKV, CHIKV, DENV or STORCH negative controls of the same
gestational age (30-41st) were obtained from Maternidade Escola - UFRJ (Rio
de Janeiro, Brazil) covered by the institutional review board-approved study
(1705093) and from Paraiba state. The death cause of correlate negative controls
cases was genetic (trisomy of chromosome 18), acute perinatal anoxia, or
complications of prematurity.
Zika virus diagnostic procedures
ZIKV RNA was investigated in the mothers or babies through RT-PCR
targeting the env gene as described by Lanciotti et al., 2008 3. ZIKV RNA was
detected in fluid samples including blood, urine, amniotic fluid obtained by
amniocentesis during gestation, or in other fluids after birth (amniotic fluid and/or
blood cord). ZIKV virus genome was also investigated postnatal in the autopsied
tissues (placenta, brain, and other organs). Viral RNA was extracted from 140 μl
fluids using QIAmp MiniElute Virus Spin (QIAgen, Hilden, Germany), following
the manufacturer’s recommendations. ZIKV RNA detection was performed using
One Step TaqMan RT-PCR (Thermo Fisher Scientific, Waltham, MA, United
States) on 7500 Real-time PCR System (Applied Biosystems, Foster City, CA,
United States) with primers, probes, and conditions as described elsewhere [4].
Fifty milligrams of frozen organs such as cerebral cortex, heart, skin, spleen,
thymus, liver, kidneys, lung, and placenta were disrupted using Tissuerupter ®
(QIAgen, Hilden, Germany) using 325 μl of RTL buffer from RNEasy Plus Mini Kit
(QIAgen, Hilden, Germany), following the manufacturer’s protocol. RNA
extraction was processed with Rneasy Plus Mini Kit (QiAgen, Hilden, Germany),
following the recommendations of the manufacturer. Real-time RT-PCR was
performed using 1 μg of total tissue RNA using One Step TaqMan RT-PCR
(Thermo Fisher Scientific, Waltham, MA, United States) as described above.
Dengue and Chikungunya virus infections were excluded in all cases (fluids and
tissues) either by RT-PCR using ZDC Trioplex kits (Bio-Manguinhos, Fiocruz, Rio
de Janeiro, Brazil) or serological ELISA for qualitative determination of IgM and
IgG (Kit XGen, Biometrix, Brazil and Euroimmum kit, Lübeck, Germany). Other
congenital pathogens including Syphilis, Cytomegalovirus, Herpes Virus 1/2,
Toxoplasma Gondii and Rubella Virus (STORCH) were discharged by serological
ELISA against IgM (Dia.Pro Diagnostic Bioprobes, Italy), following the
manufacturer’s recommendations.
ZIKV RNA in situ hybridization (ISH) was also investigated on formalin-
fixed paraffin embedded (FFPE) tissue sections of all brain tissues using two
commercial RNAscope Target Probes (Advanced Cell Diagnostics, Hayward,
CA, United States) catalog # 464531 and 463781 complementary to sequences
866-1763 and 1550-2456 of ZIKV genome, respectively. Pretreatment,
hybridization, and detection techniques were performed according to the
manufacturer’s protocols 2.
Collagen staining/Immunohistochemistry
For total collagen visualization, paraffin-embedded sections from formalin
fixed fragments of post-mortem brains were stained with the Gomori’s Trichrome
reagent.
From the leptomeninges, choroid plexus of ZIKV cases and controls
immuno-histochemical reactions were performed, using the following monoclonal
antibody (Sigma) and dilution: anti-collagen type 1, clone col-1, 1,1000. Briefly, 5
μm thick tissue sections were incubated in an oven at 37 °C for six hours,
deparaffinized in xylene and rehydrated by placing in decreasing concentrations
of alcohol and washed in distilled water. To enhance antigen retrieval, the tissue
sections were pretreated in a pressure cooker for 15 minutes in the solution 1/20
Declare (pH 6) / 1/100 Trilogy (pH9) in distilled water. To block endogenous
peroxidase activity, they were exposed to hydrogen peroxide, washed with
distilled water and rinsed in phosphate buffered saline (PBS) to stop enzymatic
digestion, then incubated with the primary antibody overnight at 4°C, rinsed in
PBS for 5 minutes and incubated with Polymer Hi-Def (horseradish peroxidase
system) for 10 minutes at room temperature preceded by several washes in PBS.
The peroxidase reaction was visualized with DAB substrate, rinsed in running
water; the sections were then counterstained with Meyer’s hematoxylin for 1
minute, washed in running tap water for 3 minutes, dehydrated in alcohol, cleared
in xylene and mounted in a resinous medium.
Library preparation and RNA-sequencing
Brain samples were frozen in RNALater (Ambion®) and stored at -80 ° C
until extraction. The tissue was broken and homogenized by TissueRupter®
(QIAGEN) and RNA extraction was performed with the RNAeasy Plus Mini® kit
(QIAGEN), following the protocol suggested by the manufacturer.
The integrity of RNA was evaluated using Agilent 2100 Bioanalyzer with
RNA 6000 Pico. Total RNA was quantified by Quant-iT ™ RiboGreen® RNA
reagent and Kit (Invitrogen, Life Technologies Corp.) and the cDNA library was
constructed following the SMARTer Stranded Total RNASeq Peak Input
Mammalian Kit protocol (Takara Bio USA). The size distribution of the cDNA
library was measured by 2100 Bioanalyzer and quantitated prior to sequencing
using Quant-iT™ PicoGreen® RNA reagent and Kit (Invitrogen, Life
Technologies Corp.). The libraries were diluted to 4nM with 15% PhiX. The cDNA
library was sequenced with MySeq System (Illumina®, San Diego, CA) using the
MiSeq Reagent kit (150 cycles, 2 x 75 paired-end).
Pre-processing and analysis of RNA-seq data
FASTQ
quality
control
was
performed
using
FastQC
tool
(https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Paired-end reads
were aligned to the human genome, ENSEMBL GRCh38.89, by STAR v.2.5.3.a,
an ultra-fast aligner 4. Then, aligned reads were quantified using featureCounts
v.1.5.3 5. Differentially expressed genes (DEGs) between control and infected
conditions were detected using DESeq2 v.1.16.1 R package 6 (adjusted p-value
< 0.1). Functional enrichment analysis was performed using the Reactome
pathways (https://reactome.org/) and the EnrichR tool7 for Over Representation
Analysis. The overlap between gene lists was performed using circlize 8 and
UpSetR 9 packages.
DNA extraction for Whole Exome Sequencing
DNA extraction and exome sequencing Genomic DNA was extracted from
the central nervous system. Exome sequencing libraries were prepared using
Illumina TruSeq® Exome Kit (8 rxn × 6plex). Sequencing was performed using
Illumina NextSeq® 500/550 High Output Kit v2 (150 cycles), generating 2x75 bp
paired-end reads.
Whole Exome Sequencing analysis
The quality of the exome libraries was evaluated using the FASTQC tool
(http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). The removal of
reads or fragments with low quality was performed by the Trimmomatic software
(http://www.usadellab.org/cms/?page=trimmomatic). The resulting high-quality
reads were aligned with the human genome as a reference (version GRCh38)
using Bowtie2 10 with very sensitive default preset (-D 20 -R 3 -N 1 -L 20 -i
S,1,0.50), except to one mismatch per seed region (-N 1). The optical duplicates
were marked with mark duplicates tool (http://broadinstitute.github.io/picard/).
Further, the Genome Analysis Toolkit (GATK) version 3.7 11 was used to call
Single Nucleotide Variants (SNVs), small insertions and deletions (INDELs). All
variants were annotated with the HaplotypeCaller following the GATK best
practices manual 12,13. The variant calls with a read coverage of ≤ 5 reads or a
MAP quality (MAPQ) of ≤ 30 were filtered out in order to avoid false positives.
The SnpEff 14 and SnpSift 15 version 4.3r tools were used to predict and annotate
the functional impact of variants, using the dbSNP (build 151) 16 and dbNSFP
(version 3.5) 17 databases. The variants with MAF (minor allele frequency) ≤ 5%
in at least one of the following databases (retrieved from dbNSFP database V3.5)
were
considered:
1000
Genomes
project
Phase
3
(http://www.internationalgenome.org/), ExAC (http://exac.broadinstitute.org/),
gnomAD
(http://gnomad.broadinstitute.org/),
TOPMed
(https://www.nhlbiwgs.org/),
ESP6500
(http://evs.gs.washington.edu/EVS/),
TwinsUK (http://www.twinsuk.ac.uk/), ALSPAC (http://www.bris.ac.uk/alspac/),
ABraOM (http://abraom.ib.usp.br/). We also verified the presence of variants in
GWAS catalog retrieved from https://www.ebi.ac.uk/gwas/ and CLINVAR -
release 20180603 (https://www.ncbi.nlm.nih.gov/clinvar/). The 1000 Genomes
project Phase 3, EXAC and gnomAD included African, Ad Mixed American, East
Asian, Europea, South Asian and Non-Finnish European populations. The
TOPMed and ESP6500 included cohorts from the United States. The TwinsUK
included old aged twins from the United Kingdom and ALSPAC included
European cohorts. The ABraOM is a variant repository comprising a cohort of
elderly Brazilians 18. We considered only the variants with CADD score ≥ 15 19
and used a set of functional effect predictors such as MetaSVM, FATHMM, LRT,
PROVEAN, Polyphen2-HDIV, Polyphen2-HVAR, MutationTaster, Mutation
Assessor and SIFT for variants prioritization 20. All variants of interest were
manually inspected with IGV tool 21.
Protein extraction
Approximately thirty milligrams of brain tissue was homogenized with 1.5
ml of extraction solution containing 5% of sodium deoxycholate (SDC), 0.75 mM
dithiothreitol (DTT), protease and phosphatase inhibitors (Roche) in a
TissueRuptor (QIAgen). After incubation for 20 min at 80 °C, the solution was
vortexed for 20 s and centrifuged for 30 min at 4 °C, 20,000 g. The pellet from
overnight precipitation of 400 μl of the supernatant with cold acetone (ratio 1: 4),
was washed two times with acetone, centrifuged for 15 min at 4 ° C, 20,000 g
and dried. After solubilizing with 7M urea / 2M thiourea with 2 % SDC we used
the Qubit® protein assay kit (Invitrogen) to measure protein content according to
the manufacturer’s instructions.
Enzymatic digestion
Reduction and alkylation of 100 μg of soluble proteins used 10 mM DTT
for 1 h at 30 oC and 40 mM IAA for 30 min at room temperature, in the dark.
Samples were diluted 1:10 with triethylammonium bicarbonate buffer (TEAB) 100
mM pH 8.5 and digested with trypsin (1:25, w/w) for 18 hours at 35 o C. Addition
of a final concentration of 1% TFA stopped digestion and two centrifugations for
15 min, 4 °C at 20,000 g removed SDC. Finally, samples were desalted in Macro
SpinColumns C18 (Harvard Apparatus) and dried in a vacuum concentrator
(Martin Christ). Peptides were suspended in 15 μl of formic acid 0.1% and
quantified by the Qubit ® protein assay as described by the manufacturer.
Nano-LC MS2 analysis
Each sample was analyzed four times (4 technical replicates) in an EASY
1000 - nLC (Thermo Scientific) coupled to a Q-Exactive Plus mass spectrometer
(Thermo Scientific). Two µg of the peptide mixture was loaded in a homemade 3
cm trap column, 200 µm I.D., 5 µm ReprosilPur C18 AQ (Dr. Maishy) beads and
fractionated in 20 cm Self-Pack PicoFrit analytical column (New Objective), 75
µm I.D., 3 µm ReprosilPur C18 AQ (Dr. Maishy). nLC gradient fractionation lasted
180 min and a flow-rate of 250 nL/min: 167 min from 5% to 40% of solvent B
(95% ACN/ 5% H2O / 0.1% formic acid); 5 min from 40% to 95% of solvent B; and
8 min in 95% of solvent B. Column and trap were equilibrated with solvent A (95%
H2O / 5% ACN / 0.1% formic acid) after each run for 15 and 2 min, respectively.
The instrument was set in the positive polarity and Full-MS/DD MS2 mode.
Selected full scan parameters were 1 microscan, 70,000 resolution at 200 m/z,
3e6 ions for AGC target, 50 ms maximum injection time and range of 375-2000
m/z. Top 20 DD-MS2 parameters were 17,000 resolution, 200 m/z, 1e5 ions for
AGC target, the maximum injection time of 100 ms, 1.2 Th of isolation window,
NCE of 30, minimum intensity threshold of 10,000 ions, and dynamic exclusion
of 60 s.
Proteomics analysis
For database search, raw data were processed using Proteome
Discoverer 2.1 (PD2.1) software (Thermo Scientific) and the SuperQuant strategy
performed by nodes MSn-Deconvolution and Complementary Finder as referred
in 22. Search performed against all reviewed human and virus entries present in
the UniProt Database (Jan/2017) used Sequest HT algorithm. Virus proteins
were not considered for the analyses. The parameters used for the search were
full tryptic peptides, two missed cleavages allowed, precursor mass tolerance of
10 ppm, 0.1 Da product ion mass tolerance, cysteine carbamidomethylation as
fixed modification, and methionine oxidation and protein N-terminal acetylation
as variable modifications. To estimate the False Discovery Rate (FDR) of <1%
we used the node Percolator present in the PD2.1 using maximum parsimony. A
cutoff score was established to accept a false-discovery rate (FDR) of 1% at the
protein and peptide level, and proteins were grouped in master proteins using the
maximum parsimony principle.
Quantification used the workflow node Precursor Ions Area Detector in
PD2.1. The peak area estimated by the Extracted Ion Chromatogram (XIC) for
the three most abundant distinct peptides of each protein were averaged and
used for relative quantification. Statistical analysis was carried out on Perseus
version 1.6.0.7. 23. Data was converted to log2 scale and normalized by
subtracting the converted protein area value (XIC) by the median of the sample
distribution. Only proteins with peak area averages present in at least three runs
were used for quantitative evaluation.
We used the limma R package 24 to identify the proteins that were up- or
down-regulated in CZS brains compared to the control brain. A cutoff Adjusted
P-value < 0.1 was used. Proteins detected in at least 2 CZS samples and not
detected in the control were considered exclusively expressed in CZS. Proteins
detected in the control and not detected in any of the CZS samples were
considered exclusively expressed in control. Functional enrichment analysis was
performed using the Reactome pathways (https://reactome.org/) and the EnrichR
tool7 for Over Representation Analysis.
Network analysis
Protein-protein interaction (PPI) networks and the miRNA-gene network
were generated using the NetworkAnalyst tool 25. Protein-protein interactions
(edges) were retrieved from STRING interactome with confidence score 900. The
miRNA-gene interaction data were collected from TarBase and miRTarBase
(validated interactions). We used the Minimum Network tool to include the seed
genes/proteins (i.e. DEGs or DEPs) as well as the essential non-seed
genes/proteins that keep the network connection. Cytoscape program 26 was also
used to visualize the networks.
References
1
Larroche, J.-C. Developmental pathology of the neonate. (Excerpta Medica ;
sole distributors for the U.S.A. and Canada, Elsevier/North-Holland, 1977).
2
Chimelli, L. et al. The spectrum of neuropathological changes associated with congenital
Zika virus infection. Acta Neuropathol 133, 983-999, doi:10.1007/s00401-017-1699-5
(2017).
3
Lanciotti, R. S. et al. Genetic and serologic properties of Zika virus associated with an
epidemic, Yap State, Micronesia, 2007. Emerg Infect Dis 14, 1232-1239,
doi:10.3201/eid1408.080287 (2008).
4
Dobin, A. et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics 29, 15-21,
doi:10.1093/bioinformatics/bts635 (2013).
5
Shin, J. et al. Single-Cell RNA-Seq with Waterfall Reveals Molecular Cascades underlying
Adult Neurogenesis. Cell Stem Cell 17, 360-372, doi:10.1016/j.stem.2015.07.013 (2015).
6
Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion
for RNA-seq data with DESeq2. Genome Biol 15, 550, doi:10.1186/s13059-014-0550-8
(2014).
7
Chen, E. Y. et al. Enrichr: interactive and collaborative HTML5 gene list enrichment
analysis tool. BMC Bioinformatics 14, 128, doi:10.1186/1471-2105-14-128 (2013).
8
Gu, Z., Gu, L., Eils, R., Schlesner, M. & Brors, B. circlize Implements and enhances circular
visualization in R. Bioinformatics 30, 2811-2812, doi:10.1093/bioinformatics/btu393
(2014).
9
Lex, A., Gehlenborg, N., Strobelt, H., Vuillemot, R. & Pfister, H. UpSet: Visualization of
Intersecting
Sets.
IEEE
Trans
Vis
Comput
Graph
20,
1983-1992,
doi:10.1109/TVCG.2014.2346248 (2014).
10
Langmead, B. & Salzberg, S. L. Fast gapped-read alignment with Bowtie 2. Nat Methods
9, 357-359, doi:10.1038/nmeth.1923 (2012).
11
McKenna, A. et al. The Genome Analysis Toolkit: a MapReduce framework for analyzing
next-generation
DNA
sequencing
data.
Genome
Res
20,
1297-1303,
doi:10.1101/gr.107524.110 (2010).
12
DePristo, M. A. et al. A framework for variation discovery and genotyping using next-
generation DNA sequencing data. Nat Genet 43, 491-+, doi:10.1038/ng.806 (2011).
13
Van der Auwera, G. A. et al. From FastQ data to high confidence variant calls: the
Genome Analysis Toolkit best practices pipeline. Curr Protoc Bioinformatics 43, 11 10
11-33, doi:10.1002/0471250953.bi1110s43 (2013).
14
Cingolani, P. et al. A program for annotating and predicting the effects of single
nucleotide polymorphisms, SnpEff: SNPs in the genome of Drosophila melanogaster
strain w(1118); iso-2; iso-3. Fly 6, 80-92, doi:10.4161/fly.19695 (2012).
15
Cingolani, P. et al. Using Drosophila melanogaster as a Model for Genotoxic Chemical
Mutational
Studies
with
a
New Program, SnpSift.
Front
Genet 3, 35,
doi:10.3389/fgene.2012.00035 (2012).
16
Sherry, S. T. et al. dbSNP: the NCBI database of genetic variation. Nucleic Acids Res 29,
308-311 (2001).
17
Liu, X., Jian, X. & Boerwinkle, E. dbNSFP: a lightweight database of human
nonsynonymous SNPs and their functional predictions. Hum Mutat 32, 894-899,
doi:10.1002/humu.21517 (2011).
18
Naslavsky, M. S. et al. Exomic variants of an elderly cohort of Brazilians in the ABraOM
database. Hum Mutat 38, 751-763, doi:10.1002/humu.23220 (2017).
19
van der Velde, K. J. et al. GAVIN: Gene-Aware Variant INterpretation for medical
sequencing. Genome Biol 18, 6, doi:10.1186/s13059-016-1141-7 (2017).
20
Liu, X., Wu, C., Li, C. & Boerwinkle, E. dbNSFP v3.0: A One-Stop Database of Functional
Predictions and Annotations for Human Nonsynonymous and Splice-Site SNVs. Hum
Mutat 37, 235-241, doi:10.1002/humu.22932 (2016).
21
Robinson, J. T. et al. Integrative genomics viewer. Nat Biotechnol 29, 24-26,
doi:10.1038/nbt.1754 (2011).
22
Gorshkov, V., Verano-Braga, T. & Kjeldsen, F. SuperQuant: A Data Processing Approach
to Increase Quantitative Proteome Coverage. Anal Chem 87, 6319-6327,
doi:10.1021/acs.analchem.5b01166 (2015).
23
Tyanova, S. et al. The Perseus computational platform for comprehensive analysis of
(prote)omics data. Nat Methods 13, 731-740, doi:10.1038/nmeth.3901 (2016).
24
Ritchie, M. E. et al. limma powers differential expression analyses for RNA-sequencing
and microarray studies. Nucleic Acids Res 43, e47, doi:10.1093/nar/gkv007 (2015).
25
Xia, J. G., Benner, M. J. & Hancock, R. E. W. NetworkAnalyst - integrative approaches for
protein-protein interaction network analysis and visual exploration. Nucleic Acids Res
42, W167-W174, doi:10.1093/nar/gku443 (2014).
26
Shannon, P. et al. Cytoscape: A software environment for integrated models of
biomolecular
interaction
networks.
Genome
Res
13,
2498-2504,
doi:10.1101/gr.1239303 (2003).
| 2019 | The potential role of collagens in congenital Zika syndrome: A systems biology approach | 10.1101/541268 | [
"Aguiar Renato S",
"Pohl Fabio",
"Morais Guilherme L",
"Nogueira Fabio CS",
"Carvalho Joseane B",
"Guida Letícia",
"Arge Luis WP",
"Melo Adriana",
"Moreira Maria EL",
"Cunha Daniela P",
"Gomes Leonardo",
"Portari Elyzabeth A",
"Velasquez Erika",
"Melani Rafael D",
"Pezzuto Paula",
"Cas... | creative-commons |
1
Genetic association of FMRP targets with
psychiatric disorders
Nicholas E Clifton1,2, Elliott Rees2, Peter A Holmans2, Antonio F. Pardiñas2, Janet C Harwood2,
Arianna Di Florio2, George Kirov2, James TR Walters2, Michael C O’Donovan2, Michael J
Owen2, Jeremy Hall1,2, Andrew J Pocklington2
1. Neuroscience and Mental Health Research Institute, Cardiff University, Cardiff, United
Kingdom.
2. MRC Centre for Neuropsychiatric Genetics and Genomics, Institute of Psychological
Medicine and Clinical Neurosciences, Cardiff University, Cardiff, United Kingdom.
Corresponding author: Jeremy Hall
NMHRI
Cardiff University
Hadyn Ellis Building
Maindy Road
Cardiff, CF24 4HQ
UK
+442920688342
hallj10@cardiff.ac.uk
2
ABSTRACT
Genes encoding the mRNA targets of Fragile X mental retardation protein (FMRP) are
enriched for genetic association with psychiatric disorders. However, many FMRP targets
possess functions that are themselves genetically associated with psychiatric disorders,
including synaptic transmission and plasticity, making it unclear whether the genetic risk is
truly related to binding by FMRP or is alternatively mediated by the sampling of genes better
characterised by another trait or functional annotation. Using published common variant,
rare coding variant and copy number variant data, we examined the relationship between
FMRP binding and genetic association with schizophrenia, major depressive disorder and
bipolar disorder. We then explored the partitioning of genetic association between
overrepresented functional categories. High-confidence targets of FMRP were enriched for
common schizophrenia risk alleles, as well as rare loss-of-function and de novo
nonsynonymous variants in cases. Similarly, through common variation, FMRP targets were
associated with major depressive disorder, and we present novel evidence of association with
bipolar disorder. These relationships could not be explained by membership of other
functional annotations known to be associated with psychiatric disorders, including those
related to synaptic structure and function. This study reinforces the evidence that targeting
by FMRP captures a subpopulation of genes enriched for genetic association with a range of
psychiatric disorders, across traditional diagnostic boundaries.
3
INTRODUCTION
Fragile X mental retardation protein (FMRP) binds selected mRNA species to repress their
translation (1–5). In the brain, FMRP is highly, and dynamically, expressed in neurons, where
it regulates the dendritic synthesis of a range of proteins (6,7), many of which are modulators
of synaptic plasticity (1). The loss of FMRP function causes Fragile X syndrome (8),
characterised by abnormal dendritic morphology, impaired learning and memory, autism and
a high prevalence of seizures (9).
The mRNA targets of FMRP have received additional attention from psychiatric research due
to their enrichment for genes harbouring risk to psychiatric disorders. A set of 842 high-
confidence FMRP targets, originating from a study by Darnell et al in 2011 (1), have been
reported to be enriched for genetic association with schizophrenia (10–16), autism (17–20)
and major depressive disorder (21). In the case of schizophrenia, not only is this association
robust across genome-wide association studies, but it is also mirrored across studies of
multiple types of genetic mutation (common and rare) conferring risk to the disorder (10–
16).
Whilst the case for the involvement of some FMRP targets in psychiatric disorders is now
unequivocal, it has been noted that FMRP targets represent long, brain-expressed transcripts
(22) with considerable overlap with other sets of genes enriched for genetic association with
psychiatric disorders, including those encoding synaptic proteins (1,23). This has led to
speculation that the association between psychiatric disorders and FMRP targets is driven not
by the property of being targets of FMRP per se, rather that it reflects association to one or
more functional sets of genes that also happen to be overrepresented in the FMRP target set
(22). Furthermore, FMRP targets were defined by applying a cut-off to a probabilistic scale of
4
FMRP binding (1), though the relationship between these binding statistics and genetic
association with psychiatric disorders has not been investigated.
In the present study, we aimed to 1) establish whether the association of FMRP target genes
with schizophrenia depends on binding confidence; 2) determine whether these associations
can be explained by the sampling of otherwise characterised or functionally-annotated genes;
and 3) demonstrate the extent to which FMRP targets are associated with risk across a range
of psychiatric disorders.
RESULTS
The relationship between FMRP binding confidence and enrichment for association with
schizophrenia
We investigated the enrichment for common variant association with schizophrenia in bins
of expressed (1) genes (N = 400 per bin) grouped by their ranking of mRNA-FMRP binding
confidence. These gene set association analyses were performed using MAGMA, in which
effects of gene size and SNP density are controlled for within a multiple regression model
(24). Bins containing genes with greater FMRP binding confidence were more enriched for
association with schizophrenia (Figure 1a), with only the top three bins being significantly
associated (bin 1: corrected P = 2.3 × 10-5; bin 2: corrected P = 1.5 × 10-5; bin 3: corrected P
= 0.030).
FMRP targets have likewise been associated with schizophrenia through rare genetic variants
(12–15). We used exome sequencing data to determine which bins of genes were associated
with schizophrenia through rare and de novo coding variants. In the case-control analysis of
rare loss-of-function variants, notably, the same top three bins enriched for GWAS signal were
5
the only bins to be significantly enriched for association with schizophrenia through rare loss-
of-function variants (bin 1: corrected P = 1.3 × 10-5; bin 2: corrected P = 0.0035; bin 3:
corrected P = 0.034) (Figure 1c). Only the topmost bin was associated through de novo
nonsynonymous variants (corrected P = 1.3 × 10-4) (Figure 1d).
Since risk to schizophrenia is also conferred through structural genetic variants (25–28) in the
form of deletions or duplications of large sections of DNA, we investigated whether CNVs
from patients with schizophrenia are enriched for genes within bins of probable FMRP targets
compared to control subjects. Following logistic regression analysis, no bins surpassed the
threshold for significance (Figure 1f) and the same was true if we examined deletions and
duplications separately (Supplementary Figure 1).
Refining schizophrenia association of FMRP targets through functionally defined subgroups
Many proteins translated from mRNA targets of FMRP have synaptic functions (1). In turn,
substantial evidence shows that genes encoding proteins with synaptic functions are enriched
for genetic association with schizophrenia(11–13,23,29,30). To further assess the importance
of FMRP targeting to the association of genes with schizophrenia, we separated the 842 FMRP
target genes, as determined by Darnell et al (1), into subgroups defined by overrepresented
functional categories.
Molecular pathways were derived using pathway analysis (Figure 2) with GO (Supplementary
Table 2) and MP terms (Supplementary Table 3). The resulting 189 GO terms and 118 MP
terms were refined to identify terms independently overrepresented among FMRP targets.
This procedure left a total of 35 independent overrepresented terms (Supplementary Table
4).
6
To assess the contribution to genetic association of the property “FMRP binding”, versus that
of these functional ontologies, we created a superset (N = 1596) of brain-expressed genes
which are included in at least one of the 35 functional terms overrepresented for FMRP
targets. FMRP targets from this set (N = 401) were strongly enriched for common variant
association (β = 0.29, corrected P = 3.7 × 10-6), whilst genes not targeted by FMRP (N = 1195)
were not (β = 0.066, corrected P = 0.13) (Table 1). FMRP targets that were not included in any
of the 35 terms (N = 438) were also significantly associated (β = 0.17, corrected P = 0.0063).
Thus, FMRP targets appear to capture schizophrenia associated genes from these functional
categories (when taken as a whole). The burden of rare loss-of-function variants in cases
showed the same pattern of association as the common variants, being only enriched in the
sets that included FMRP targets (Table 1), regardless of superset membership. However,
enrichment for de novo nonsynonymous mutations showed a different picture, with
significant association being observed only for the set of genes that were exclusive to FMRP
targets (Rate ratio = 1.58, corrected P = 9.2 × 10-4).
In comparisons of effect sizes from analyses of any type of genetic variant, FMRP targets
annotated by overrepresented functional terms were not more enriched for association with
schizophrenia than unannotated FMRP targets (common variants: P = 0.081; rare loss-of-
function variants: P = 0.25; de novo nonsynonymous variants: P = 0.88).
We next sought to determine from which of the individual overrepresented functional terms
FMRP targets capture genetic association with schizophrenia, and whether association is
further enriched within FMRP targets from any single overrepresented term, compared to the
complete FMRP targets set. Several functionally-defined subsets of FMRP targets were
significantly associated with schizophrenia through common variation (Table 2), whilst genes
7
not targeted by FMRP were not associated; with the exception of those belonging to the term,
calcium ion transmembrane transporter activity (Supplementary Table 4), although in this
instance the fraction targeted by FMRP was associated with a significantly greater effect size
(P = 0.0088). The calcium ion transmembrane transporter activity fraction of FMRP targets (N
= 25) remained significantly associated with schizophrenia even after conditioning on all
FMRP targets (Supplementary Table 4), implying that this functionally-defined subset of FMRP
targets is more enriched for association with schizophrenia than FMRP targets as a whole. No
other term captured FMRP targets with a significantly greater enrichment of genetic
association than the full FMRP targets gene set.
Rare loss-of-function variants from patients with schizophrenia were enriched in FMRP
targets from two terms (abnormal spatial learning, abnormal motor coordination/balance)
(Supplementary Table 5), whilst no association was found between rare coding variants in
non-targeted genes from each term and schizophrenia. None of these subsets harboured
significantly more enrichment for case variants than all FMRP targets.
None of the subsets tested captured a significant burden of case de novo nonsynonymous
variants (Supplementary Table 6).
Overall, these analyses suggest that the overrepresentation of FMRP targets drives genetic
association of these biological pathways with schizophrenia, rather than the reverse.
Genetic association of FMRP targets in other psychiatric disorders
Schizophrenia shares substantial genetic susceptibility with bipolar disorder and major
depressive disorder (31–34) and FMRP targets have been previously associated through
common variation with major depressive disorder (21). For comparison across disorders, we
8
tested the enrichment of FMRP targets bins for association with major depressive disorder
and bipolar disorder using common variant data from GWAS. In both sets of analyses, there
was a clear relationship between FMRP binding confidence and genetic association (Figure
3b,c). The topmost bin, containing genes most likely to be FMRP targets, was the most
strongly enriched for association with bipolar disorder (corrected P = 1.4 × 10-6) and major
depressive disorder (corrected P = 2.5 × 10-4).
We investigated functionally-annotated subgroups of FMRP targets for association with
bipolar disorder and major depressive disorder. Beyond background association from brain-
expressed genes, FMRP targets annotated for membership of the 35 overrepresented
pathways were strongly associated with bipolar disorder (β = 0.23, corrected P = 1.6 × 10-5)
and major depressive disorder (β = 0.21, corrected P = 1.6 × 10-5), whilst genes from the same
functional terms not targeted by FMRP harboured no significant association (bipolar disorder:
β = 0.037, corrected P = 0.38; major depressive disorder: β = 0.031, corrected P = 0.49)(Table
3). A similar picture was observed for individual overrepresented GO / MP terms. Following
multiple testing correction, FMRP targets were significantly associated with bipolar disorder
from 4 terms (calcium ion transmembrane transporter activity, abnormal nest building
behavior, abnormal spatial learning and abnormal seizure response to inducing agent).
Notably, the association of FMRP targets from these 4 terms was common to schizophrenia
and bipolar disorder. FMRP targets from 1 term (abnormal synaptic vesicle morphology) were
significantly associated with major depressive disorder (Supplementary Table 4). FMRP
targets belonging to the term abnormal nest building behavior (N = 12) were more highly
enriched for association with bipolar disorder than FMRP targets as a whole. No FMRP targets
9
were significantly more enriched for association with major depressive disorder than the full
FMRP targets set (Supplementary Table 4).
DISCUSSION
In this study we investigated the extent to which targeting by FMRP is related to genetic
association with psychiatric disorders. We show that genes with high probability of being
targets of FMRP are enriched for association with schizophrenia, bipolar disorder and major
depressive disorder. We also show that it is the property of being an FMRP target that
captures the genetic association, rather than membership of gene sets that happen to be
enriched for targets of FMRP.
Only bins of genes with the highest FMRP binding confidence were enriched for association
with schizophrenia through common variation, exome sequencing-derived rare variation and
exome sequencing-derived de novo rare variation. This same relationship was reflected in
analyses of bipolar disorder and major depressive disorder risk alleles. Our observations are
consistent with previous gene set analyses of FMRP targets in the context of schizophrenia
(11–15) and major depressive disorder (21), but whilst FMRP targets have been previously
linked to bipolar disorder through rare coding variants (35), our findings provide novel
evidence linking FMRP targets to bipolar disorder through common variation.
Despite the evidence implicating FMRP targets in psychiatric disorders (11–15), the
overrepresentation of long, brain-expressed genes with synaptic functions has led to
cautiousness over the validity of the link to FMRP (22). The methods used here, and previously
(11), correct for, or are unaffected by, gene length, allowing us more confidence in concluding
that the relationships between FMRP binding and association with schizophrenia, bipolar
10
disorder and major depressive disorder exist beyond any confounding effects of gene length.
Furthermore, whilst the associated genes were derived from expressed mRNAs in mouse
brain, the associations did not generalize to bins of brain-expressed genes with low FMRP
binding confidence.
Consistent with previous pathway analysis (1), we note that a substantial proportion of FMRP
targets have functions related to synaptic activity, anatomy or development. Studies of FMRP
function show that its activity is regulated in response to neuronal activity (36–39) and is an
important mediator of synapse development (40–42), synaptic plasticity (43–45), learning
and memory (46–48). Genetic and functional studies have highlighted the relevance of
perturbed synaptic plasticity in psychiatric disorders (12,30,49–53), although we find that the
risk conferred by variants affecting such pathways overrepresented among FMRP targets is
concentrated within the fraction of genes targeted by FMRP. Hence, despite the convergence
of psychiatric risk on synaptic pathways (12,23,51–54), the association of FMRP targets was
not attributed to these overrepresented annotations. Instead, it appears that there is a
degree of specificity to this risk, such that genes regulated locally by FMRP during activity-
induced synaptic plasticity, required for development or learning, are most relevant to
psychiatric disorder.
It should be noted that other synapse-related gene sets are enriched for association with
psychiatric disorders independently of FMRP targets. For instance, recent schizophrenia
common variant analyses show a few such independent associations of gene sets related to
synaptic function (11). Here we found that, whilst strongest for genes targeted by FMRP,
genes involved in calcium ion transmembrane transporter activity held independent
association with schizophrenia. Furthermore, the strongly associated, albeit small,
11
intersection between genes from this set and FMRP targets contained a stronger enrichment
of schizophrenia common variant association than FMRP targets (or indeed the GO term) as
a whole. This is consistent with previous evidence for association of calcium channels with
schizophrenia (10,11,13), yet additionally suggests that FMRP captures a subset of genes
related to calcium ion transport in which common variant association is concentrated.
FMRP binding confidence was not related to genetic association with schizophrenia through
CNVs. Whilst FMRP targets have been consistently implicated in schizophrenia from analyses
of all other types of genetic variant, studies of structural variation in schizophrenia have
shown only modest association of FMRP targets (28,30,55), although a deletion at 15q11.2
affecting the FMRP interacting protein, CYFIP1, which is required for the regulation of
translation by FMRP (5,56), is associated. Why we observe risk for schizophrenia affecting
FMRP targets being conferred through all variants except for CNVs is unclear, although these
analyses may be influenced by the difficulty of attributing CNV association to individual genes.
Our observations resonate with the growing body of literature challenging the biological
validity of viewing major psychiatric disorders as discrete entities with independent genetic
aetiology (57–60). There is considerable overlap between the genetic risk attributable to
schizophrenia, bipolar disorder and major depressive disorder (31–34). The present (and
published) findings highlight that FMRP targets are a point of shared heritability. Additional
evidence suggests that genetic association of FMRP targets may extend also to autism (17–
20) and attention-deficit hyperactivity disorder (61).
Our findings highlight a set of genes regulated through a common mechanism that harbour
risk across several psychiatric disorders. However, there is a degree of uncertainty as to
precisely which mRNAs are regulated by FMRP. Multiple studies have examined this, each
12
yielding overlapping, yet distinct sets of FMRP targets (1,4,62–65); some of the variability
likely originating from tissue-specificity. When performing pathway analyses with genomic
data, many studies, including this one, have obtained FMRP targets from an investigation of
mRNA-FMRP interaction sites in mouse cortical polyribosomes (1), in which membership was
assigned by applying a stringent cut-off to a continuous scale of binding confidence, likely
resulting in some false positives and more false negatives. Moreover, binding by FMRP may
not equate to translational repression in the cell, which requires additional contribution from
binding partners CYFIP1 and eIF4E, within a protein complex (5). Hence, this line of research
will benefit from further validation of FMRP-regulated protein synthesis in the context of
psychiatric pathology.
Our results serve to strengthen the evidence that a population of genes targeted by FMRP,
many of which have synaptic functions, are affected by genetic variation conferring risk to
psychiatric disorders, including schizophrenia, bipolar disorder and major depressive
disorder. We conclude that targeting by FMRP is currently the most suitable functional
annotation to reflect the origin of these associations and represents a common mode of
regulation for a set of genes contributing risk across several major psychiatric presentations.
MATERIALS AND METHODS
Gene sets
FMRP binding statistics for 30999 transcripts were obtained from Darnell et al (2011)
(Supplementary Table S2C) (1), a study of mRNA-FMRP interaction sites in mouse cortical
polyribosomes using crosslinking immunoprecipitation combined with high-throughput RNA
sequencing. We filtered the data to include only genes which were expressed (chi-square
score > 0), therefore selecting only those from which binding statistics could be obtained. We
13
converted Mouse Entrez IDs to human Entrez IDs via their shared HomoloGene ID, obtained
from Mouse Genome Informatics Vertebrate Homology database release 6.10
(HOM_AllOrganism.rpt, 8th January 2018). Genes that did not convert to a unique protein-
coding human homologue were excluded. The remaining 8595 genes were ranked by their
FMRP binding confidence P-value and the top 8400 were split into 21 bins of 400 genes to
determine the relationship between FMRP binding confidence and schizophrenia association.
Functional enrichment analyses were performed using the set of 842 FMRP targets (reported
FDR < 0.01 in Darnell et al, 2011) (1) that has been widely used in previous enrichment studies
(11,12).
Samples
Common variants
All genetic data were obtained from published case and control samples. Schizophrenia
common variant summary statistics were taken from the Pardiñas et al (2018) study, a meta-
analysis of genome-wide association studies (GWAS) (11) based on a sample of 40 675 case
and 64 643 control subjects. Bipolar disorder common variant summary statistics were
provided by a recent Psychiatric Genomics Consortium (PGC) GWAS (53), consisting of 20 352
cases and 31 358 controls from 32 cohorts of European descent. Major depressive disorder
common variant summary statistics were taken from a PGC meta-analysis of 135 458 cases
and 344 901 controls from seven independent cohorts of European ancestry (21).
Rare coding variants
Exome sequencing-derived rare coding variant data from a Swedish schizophrenia case-
control study (16) were obtained from the NCBI database of genotypes and phenotypes
(dbGaP). After excluding individuals with non-European or Finnish ancestry, and samples with
14
low sequencing coverage, we retained exome sequence in 4079 cases and 5712 controls for
analysis.
De novo coding variants
De novo mutations were derived (66) from previously published exome sequencing studies
of, collectively, 1136 schizophrenia-proband parent trios (12,67–74) (Supplementary Table 1).
Copy number variants
Copy number variant (CNV) data were compiled from the CLOZUK and Cardiff Cognition in
Schizophrenia samples (11 955 cases, 19 089 controls) (27,75), as well as samples from the
International Schizophrenia Consortium (3395 cases, 2185 controls) (76) and the Molecular
Genetics of Schizophrenia (2215 cases, 2556 controls) (77), giving a total of 17 565 case and
24 830 control subjects. Genotyping, CNV calling and quality control information can be found
in the original reports (25,27,30,49,76,77).
Gene set association analysis
Schizophrenia, bipolar disorder and major depressive disorder GWAS single nucleotide
polymorphisms (SNPs) were filtered to include only those with a minor allele frequency ≥
0.01. SNP association P-values were combined (SNP-wise Mean model) into gene-wide P-
values in MAGMA v1.06 (24), using a window of 35 kb upstream and 10 kb downstream of
each gene to include proximal regulatory regions. The European panel of the 1000 Genomes
Project (78) (phase 3) was used as a reference to account for linkage disequilibrium between
genes. Gene sets were tested for enrichment for association with each disorder using one-
tailed competitive gene set association analyses in MAGMA, which compares the mean
association of genes from the gene set to those not in the gene set, correcting for gene size
and SNP density. The default background was all protein-coding genes.
15
Case-control exome sequencing data were analysed using Hail (https://github.com/hail-
is/hail). We annotated variants using Hail’s Ensembl VEP method (version 86,
http://oct2016.archive.ensembl.org/index.html) and defined loss-of-function variants as
nonsense, essential splice site and frameshift annotations and nonsynonymous variants as
loss-of-function and missense annotations. For gene set enrichment tests, we focused on
ultra-rare singleton loss-of-function and nonsynonymous variants, that is those observed
once in all case-control sequencing data and absent from the non-psychiatric component of
ExAC (79). Enrichment statistics were generated using a Firth’s penalized-likelihood logistic
regression model that corrected for the first 10 principal components, exome-wide burden of
synonymous variants, sequencing platform and sex.
De novo variant gene set enrichment was evaluated by comparing the observed number of
de novo variants in a set of genes to that expected, which was based on the number of trios
analysed and per-gene mutations rates (80,81). Gene set enrichment statistics for de novo
variants were generated by using a two sample Poisson rate ratio test to compare the
enrichment of de novo variants within the gene set to that observed in a background set of
genes.
CNV analyses were restricted to CNVs at least 100 kb in size and covered by at least 15 probes.
Gene set association was tested by logistic regression, in which CNV case-control status was
regressed against the number of set genes overlapped by the CNV, with covariates: CNV size,
genes per CNV, study and chip type. To correct for P-value inflation, empirical P-values were
obtained by calculating the fraction of random size-matched sets of brain-expressed (1) genes
that yielded an association as or more significant.
Multiple testing was corrected for using the Bonferroni method.
16
Pathway analysis
For gene ontology enrichment analyses, functional annotations of each gene were compiled
separately from the Gene Ontology (GO) (82) and Mouse Genome Informatics (MGI)
Mammalian Phenotype (MP) (83) databases (July 4th 2018). GO annotations were filtered to
exclude genes with the following evidence codes: NAS (Non-traceable Author Statement), IEA
(Inferred from Electronic Annotation), and RCA (inferred from Reviewed Computational
Analysis). GO or MP terms containing fewer than 10 genes were then excluded. For all
pathway analyses, genes were restricted to those expressed (chi-square score > 0) in the
mouse brain tissue used by Darnell et al 2011 (1). Enrichment of FMRP targets for each
GO/MP term was assessed by Fisher’s exact tests, with the contrast group being all remaining
brain-expressed genes. Following separate Bonferroni correction for 8270 GO terms or 4606
MP terms, significantly (P < 0.01) overrepresented terms were subjected to a competitive
refinement procedure to resolve the effects of redundancy between terms. During
refinement, terms were re-tested for overrepresentation in FMRP targets following the
removal of genes from the term with the highest odds ratio in Fisher’s exact test. Terms that
were no longer significant upon re-test (unadjusted P > 0.01) were dropped. This was done
repeatedly, such that genes from the remaining term with the highest odds ratio on each
repeat were removed in addition to those removed on previous iterations.
In primary analyses of genetic association, brain-expressed (1) genes from all
overrepresented GO / MP terms (following refinement) were grouped together and divided
into those targeted and those not targeted by FMRP, and compared to a background of brain-
expressed genes. In secondary analyses, genes from each individual overrepresented term
were divided in the same way and tested for association using all protein-coding genes as a
17
comparator. P-values were Bonferroni corrected for the number of functional terms being
tested at each stage of analysis.
We performed a number of tests to investigate the relative enrichments for association
between two sets of genes, one a subset of the other. For common variant association, we
used the conditional analysis function provided by MAGMA. For rare or de novo coding
variants, we compared the effect sizes of the subset of genes with that of the larger set after
excluding members of the subset. For the rare coding variant case control analyses, this was
done by performing a z-test of beta values, whilst for de novo variant analyses, a two-sample
Poisson rate ratio test was used.
In cases where enrichment for genetic association was compared between non-overlapping
gene sets, a z-test of beta values (common and rare variants) or a two-sample Poisson rate
ratio test (de novo variants) was used.
ACKNOWLEDGEMENTS
This work was supported by Medical Research Council (MRC) grants MR/L010305/1 and
G0800509, a Wellcome Trust Strategic Award (100202/Z/12/Z), The Waterloo Foundation
‘Changing Minds’ programme, and Neuroscience and Mental Health Research Institute
(Cardiff University) core funding to NC.
We thank the Bipolar Disorder and Major Depressive Disorder workgroups of the Psychiatric
Genomics Consortium for providing summary statistics used in this study. We would also like
to thank the research participants and employees of 23andMe for making this work possible.
18
Exome sequencing datasets described in this manuscript were obtained from dbGaP at
http://www.ncbi.nlm.nih.gov/gap through dbGaP accession number phs000473.v2.p2.
Samples were provided by the Swedish Cohort Collection supported by the NIMH Grant No.
R01MH077139, the Sylvan C. Herman Foundation, the Stanley Medical Research Institute and
The Swedish Research Council (Grant Nos. 2009-4959 and 2011-4659). Support for the exome
sequencing was provided by the NIMH Grand Opportunity Grant No. RCMH089905, the
Sylvan C. Herman Foundation, a grant from the Stanley Medical Research Institute and
multiple gifts to the Stanley Center for Psychiatric Research at the Broad Institute of MIT and
Harvard.
Analyses of copy number variation described in this manuscript used datasets from the
Molecular Genetics of Schizophrenia (MGS; dbGAP phs000021.v3.p2 and phs000167.v1.p1)
and the International Schizophrenia Consortium (ISC). The CLOZUK and CLOZUK2 data sets
contain data obtained from outside sources: dbGaP phs000404.v1.p1, phs000187.v1. p1,
phs000303.v1.p1, phs000179.v3.p2, phs000421.v1.p, phs000395.v1.p1, phs000519.v1.p1
and the Wellcome Trust Case Control Consortium 2 study.
CONFLICT OF INTEREST STATEMENT
The authors declare no conflict of interest.
REFERENCES
1.
Darnell, J. C., Van Driesche, S. J., Zhang, C., et al. (2011) FMRP stalls ribosomal translocation on
mRNAs linked to synaptic function and autism. Cell, 146, 247–261.
2.
Laggerbauer, B., Ostareck, D., Keidel, E. M., et al. (2001) Evidence that fragile X mental
retardation protein is a negative regulator of translation. Hum. Mol. Genet.
3.
Li, Z., Zhang, Y., Ku, L., et al. (2001) The fragile X mental retardation protein inhibits translation
via interacting with mRNA. Nucleic Acids Res.
4.
Ascano, M., Mukherjee, N., Bandaru, P., et al. (2012) FMRP targets distinct mRNA sequence
19
elements to regulate protein expression. Nature, 492, 382–6.
5.
Napoli, I., Mercaldo, V., Boyl, P. P., et al. (2008) The Fragile X Syndrome Protein Represses
Activity-Dependent Translation through CYFIP1, a New 4E-BP. Cell, 134, 1042–1054.
6.
Christie, S. B., Akins, M. R., Schwob, J. E., et al. (2009) The FXG: A Presynaptic Fragile X Granule
Expressed in a Subset of Developing Brain Circuits. J. Neurosci.
7.
Stefani, G., Fraser, C. E., Darnell, J. C., et al. (2004) Fragile X Mental Retardation Protein Is
Associated with Translating Polyribosomes in Neuronal Cells. J. Neurosci.
8.
Verkerk, A. J. M. H., Pieretti, M., Sutcliffe, J. S., et al. (1991) Identification of a gene (FMR-1)
containing a CGG repeat coincident with a breakpoint cluster region exhibiting length variation
in fragile X syndrome. Cell, 65, 905–914.
9.
Bagni, C., Tassone, F., Neri, G., et al. (2012) Fragile X syndrome: Causes, diagnosis, mechanisms,
and therapeutics. Fragile X syndrome: Causes, diagnosis, mechanisms, and therapeutics. J. Clin.
Invest. (2012) .
10.
Ripke, S., Neale, B. M., Corvin, A., et al. (2014) Biological insights from 108 schizophrenia-
associated genetic loci. Nature, 511, 421–427.
11.
Pardiñas, A. F., Holmans, P., Pocklington, A. J., et al. (2018) Common schizophrenia alleles are
enriched in mutation-intolerant genes and in regions under strong background selection. Nat.
Genet., 50, 381–389.
12.
Fromer, M., Pocklington, A. J., Kavanagh, D. H., et al. (2014) De novo mutations in schizophrenia
implicate synaptic networks. Nature, 506, 179–84.
13.
Purcell, S. M., Moran, J. L., Fromer, M., et al. (2014) A polygenic burden of rare disruptive
mutations in schizophrenia. Nature, 506.
14.
Richards, A. L., Leonenko, G., Walters, J. T., et al. (2016) Exome arrays capture polygenic rare
variant contributions to schizophrenia. Hum. Mol. Genet., 25, 1001–1007.
15.
Leonenko, G., Richards, A. L., Walters, J. T., et al. (2017) Mutation intolerant genes and targets
of FMRP are enriched for nonsynonymous alleles in schizophrenia. Am. J. Med. Genet. Part B
Neuropsychiatr. Genet., 724–731.
16.
Genovese, G., Fromer, M., Stahl, E. A., et al. (2016) Increased burden of ultra-rare protein-
altering variants among 4,877 individuals with schizophrenia. Nat. Neurosci., 19, 1433–1441.
17.
Jansen, A., Dieleman, G. C., Smit, A. B., et al. (2017) Gene-set analysis shows association
between FMRP targets and autism spectrum disorder. Eur. J. Hum. Genet., 25, 863–868.
18.
Iossifov, I., Ronemus, M., Levy, D., et al. (2012) De Novo Gene Disruptions in Children on the
Autistic Spectrum. Neuron.
19.
Iossifov, I., O’Roak, B. J., Sanders, S. J., et al. (2014) The contribution of de novo coding
mutations to autism spectrum disorder. Nature.
20.
De Rubeis, S., He, X., Goldberg, A. P., et al. (2014) Synaptic, transcriptional and chromatin genes
disrupted in autism. Nature.
21.
Wray, N., Ripke, S., Mattheisen, M., et al. (2018) Genome-wide association analyses identify 44
risk variants and refine the genetic architecture of major depression. Nat. Genet., 50, 668–681.
22.
Ouwenga, R. L. and Dougherty, J. (2015) Fmrp targets or not: long, highly brain-expressed
genes tend to be implicated in autism and brain disorders. Mol. Autism, 6, 16.
23.
Hall, J., Trent, S., Thomas, K. L., et al. (2015) Genetic risk for schizophrenia: Convergence on
synaptic pathways involved in plasticity. Biol. Psychiatry, 77, 52–58.
20
24.
de Leeuw, C. A., Mooij, J. M., Heskes, T., et al. (2015) MAGMA: Generalized Gene-Set Analysis
of GWAS Data. PLoS Comput. Biol., 11.
25.
Rees, E., Walters, J. T. R., Georgieva, L., et al. (2014) Analysis of copy number variations at 15
schizophrenia-associated loci. Br. J. Psychiatry, 204, 108–114.
26.
Kirov, G., Grozeva, D., Norton, N., et al. (2009) Support for the involvement of large copy
number variants in the pathogenesis of schizophrenia. Hum. Mol. Genet., 18, 1497–1503.
27.
Rees, E., Walters, J. T. R., Chambert, K. D., et al. (2014) CNV analysis in a large schizophrenia
sample implicates deletions at 16p12.1 and SLC1A1 and duplications at 1p36.33 and CGNL1.
Hum. Mol. Genet., 23, 1669–1676.
28.
Marshall, C. R., Howrigan, D. P., Merico, D., et al. (2017) Contribution of copy number variants
to schizophrenia from a genome-wide study of 41,321 subjects. Nat. Genet., 49, 27–35.
29.
Kirov, G., Pocklington, A. J., Holmans, P., et al. (2012) De novo CNV analysis implicates specific
abnormalities of postsynaptic signalling complexes in the pathogenesis of schizophrenia. Mol.
Psychiatry, 17, 142–153.
30.
Pocklington, A. J., Rees, E., Walters, J. T. R., et al. (2015) Novel Findings from CNVs Implicate
Inhibitory and Excitatory Signaling Complexes in Schizophrenia. Neuron, 86, 1203–1214.
31.
Lee, S. H., Ripke, S., Neale, B. M., et al. (2013) Genetic relationship between five psychiatric
disorders estimated from genome-wide SNPs. Nat. Genet., 45, 984–94.
32.
Cardno, A. G. and Owen, M. J. (2014) Genetic relationships between schizophrenia, bipolar
disorder, and schizoaffective disorder. Schizophr. Bull., 40, 504–515.
33.
Owen, M. J. and O’Donovan, M. C. (2017) Schizophrenia and the neurodevelopmental
continuum:evidence from genomics. World Psychiatry, 16, 227–235.
34.
Doherty, J. L. and Owen, M. J. (2014) Genomic insights into the overlap between psychiatric
disorders: implications for research and clinical practice. Genome Med., 6, 29.
35.
Goes, F. S., Pirooznia, M., Parla, J. S., et al. (2016) Exome sequencing of familial bipolar disorder.
JAMA Psychiatry, 73, 590–597.
36.
Nakamoto, M., Nalavadi, V., Epstein, M. P., et al. (2007) Fragile X mental retardation protein
deficiency leads to excessive mGluR5-dependent internalization of AMPA receptors. Proc. Natl.
Acad. Sci.
37.
Wang, H., Wu, L. J., Kim, S. S., et al. (2008) FMRP Acts as a Key Messenger for Dopamine
Modulation in the Forebrain. Neuron.
38.
Wang, H., Morishita, Y., Miura, D., et al. (2012) Roles of CREB in the regulation of FMRP by
group i metabotropic glutamate receptors in cingulate cortex. Roles of CREB in the regulation
of FMRP by group i metabotropic glutamate receptors in cingulate cortex. Mol. Brain (2012) .
39.
Antar, L. N., Afroz, R., Dictenberg, J. B., et al. (2004) Metabotropic Glutamate Receptor
Activation Regulates Fragile X Mental Retardation Protein and Fmr1 mRNA Localization
Differentially in Dendrites and at Synapses. J. Neurosci., 24, 2648–2655.
40.
Wang, X., Zorio, D. A. R., Schecterson, L., et al. (2018) Postsynaptic FMRP Regulates
Synaptogenesis In Vivo in the Developing Cochlear Nucleus . J. Neurosci.
41.
Davis, J. K. and Broadie, K. (2017) Multifarious Functions of the Fragile X Mental Retardation
Protein. Multifarious Functions of the Fragile X Mental Retardation Protein. Trends Genet.
(2017) .
42.
Doll, C. A. and Broadie, K. (2015) Activity-dependent FMRP requirements in development of
21
the neural circuitry of learning and memory. Development.
43.
Shang, Y., Wang, H., Mercaldo, V., et al. (2009) Fragile X mental retardation protein is required
for chemically-induced long-term potentiation of the hippocampus in adult mice. J.
Neurochem.
44.
Huber, K. M., Gallagher, S. M., Warren, S. T., et al. (2002) Altered synaptic plasticity in a mouse
model of fragile X mental retardation. Proc. Natl. Acad. Sci. U. S. A., 99, 7746–7750.
45.
Eadie, B. D., Cushman, J., Kannangara, T. S., et al. (2012) NMDA receptor hypofunction in the
dentate gyrus and impaired context discrimination in adult Fmr1 knockout mice. Hippocampus.
46.
Yan, Q. J., Asafo-Adjei, P. K., Arnold, H. M., et al. (2004) A phenotypic and molecular
characterization of the fmr1-tm1Cgr fragile X mouse. Genes, Brain Behav.
47.
Michalon, A., Sidorov, M., Ballard, T. M., et al. (2012) Chronic Pharmacological mGlu5 Inhibition
Corrects Fragile X in Adult Mice. Neuron.
48.
Santos, A. R., Kanellopoulos, A. K. and Bagni, C. (2014) Learning and behavioral deficits
associated with the absence of the fragile X mental retardation protein: What a fly and mouse
model can teach us. Learning and behavioral deficits associated with the absence of the fragile
X mental retardation protein: What a fly and mouse model can teach us. Learn. Mem. (2014)
.
49.
Clifton, N. E., Pocklington, A. J., Scholz, B., et al. (2017) Schizophrenia copy number variants
and associative learning. Mol. Psychiatry, 22, 178–182.
50.
Duman, R. S., Aghajanian, G. K., Sanacora, G., et al. (2016) Synaptic plasticity and depression:
New insights from stress and rapid-acting antidepressants. Synaptic plasticity and depression:
New insights from stress and rapid-acting antidepressants. Nat. Med. (2016) .
51.
Howard, D. M., Adams, M. J., Shirali, M., et al. (2018) Genome-wide association study of
depression phenotypes in UK Biobank identifies variants in excitatory synaptic pathways. Nat.
Commun., 9, 1470.
52.
Howard, D. M., Adams, M. J., Clarke, T. K., et al. (2019) Genome-wide meta-analysis of
depression identifies 102 independent variants and highlights the importance of the prefrontal
brain regions. Nat. Neurosci., 22, 343–352.
53.
Stahl, E. A., Breen, G., Forstner, A. J., et al. (2019) Genome-wide association study identifies 30
loci associated with bipolar disorder. Nat. Genet., 51, 793–803.
54.
Rees, E., O’Donovan, M. C. and Owen, M. J. (2015) Genetics of schizophrenia. Curr. Opin. Behav.
Sci., 2, 8–14.
55.
Szatkiewicz, J., O ’dushlaine, C., Chen, G., et al. (2014) Copy number variation in schizophrenia
in Sweden. Mol. Psychiatry, 19, 762–773.
56.
DeRubeis, S., Pasciuto, E., Li, K., et al. (2013) CYFIP1 coordinates mRNA translation and
cytoskeleton remodeling to ensure proper dendritic Spine formation. Neuron, 79, 1169–1182.
57.
O’Donovan, M. C. and Owen, M. J. (2016) The implications of the shared genetics of psychiatric
disorders. The implications of the shared genetics of psychiatric disorders. Nat. Med. (2016) ,
22.
58.
Owen, M. J. (2014) New approaches to psychiatric diagnostic classification. New approaches
to psychiatric diagnostic classification. Neuron (2014) , 84, 564–571.
59.
Craddock, N. and Owen, M. J. (2010) The Kraepelinian dichotomy - Going, going... but still not
gone. The Kraepelinian dichotomy - Going, going... but still not gone. Br. J. Psychiatry (2010) ,
196, 92–95.
22
60.
Jablensky, A. (2016) Psychiatric classifications: Validity and utility. World Psychiatry.
61.
Thapar, A., Martin, J., Mick, E., et al. (2015) Psychiatric gene discoveries shape evidence on
ADHD’s biology. Mol. Psychiatry, 21, 1202–1207.
62.
Suhl, J. a, Chopra, P., Anderson, B. R., et al. (2014) Analysis of FMRP mRNA target datasets
reveals highly associated mRNAs mediated by G-quadruplex structures formed via clustered
WGGA sequences. Hum. Mol. Genet., 23, 1–41.
63.
Brown, V., Jin, P., Ceman, S., et al. (2001) Microarray identification of FMRP-associated brain
mRNAs and altered mRNA translational profiles in fragile X syndrome. Cell, 107, 477–487.
64.
Maurin, T., Lebrigand, K., Castagnola, S., et al. (2018) HITS-CLIP in various brain areas reveals
new targets and new modalities of RNA binding by fragile X mental retardation protein. Nucleic
Acids Res., 1–12.
65.
Miyashiro, K. Y., Beckel-Mitchener, A., Purk, T. P., et al. (2003) RNA cargoes associating with
FMRP reveal deficits in cellular functioning in Fmr1 null mice. Neuron, 37, 417–431.
66.
Rees, E., Carrera, N., Morgan, J., et al. (2019) Targeted Sequencing of 10,198 Samples Confirms
Abnormalities in Neuronal Activity and Implicates Voltage-Gated Sodium Channels in
Schizophrenia Pathogenesis. Biol. Psychiatry, 85, 554–562.
67.
Ambalavanan, A., Girard, S. L., Ahn, K., et al. (2016) De novo variants in sporadic cases of
childhood onset schizophrenia. Eur. J. Hum. Genet., 24, 944–948.
68.
Girard, S. L., Gauthier, J., Noreau, A., et al. (2011) Increased exonic de novo mutation rate in
individuals with schizophrenia. Nat. Genet., 43, 860–863.
69.
Guipponi, M., Santoni, F. A., Setola, V., et al. (2014) Exome sequencing in 53 sporadic cases of
schizophrenia identifies 18 putative candidate genes. PLoS One, 9.
70.
Gulsuner, S., Walsh, T., Watts, A. C., et al. (2013) Spatial and temporal mapping of de novo
mutations in schizophrenia to a fetal prefrontal cortical network. Cell, 154.
71.
McCarthy, S. E., Gillis, J., Kramer, M., et al. (2014) De novo mutations in schizophrenia implicate
chromatin remodeling and support a genetic overlap with autism and intellectual disability.
Mol. Psychiatry, 19, 652–658.
72.
Takata, A., Xu, B., Ionita-Laza, I., et al. (2014) Loss-of-Function Variants in Schizophrenia Risk
and SETD1A as a Candidate Susceptibility Gene. Neuron, 82, 773–780.
73.
Wang, Q., Li, M., Yang, Z., et al. (2015) Increased co-expression of genes harboring the
damaging de novo mutations in Chinese schizophrenic patients during prenatal development.
Sci. Rep., 5.
74.
Xu, B., Ionita-Laza, I., Roos, J. L., et al. (2012) De novo gene mutations highlight patterns of
genetic and neural complexity in schizophrenia. Nat. Genet., 44, 1365–1369.
75.
Rees, E., Kendall, K., Pardiñas, A. F., et al. (2016) Analysis of Intellectual Disability Copy Number
Variants for Association With Schizophrenia. JAMA Psychiatry, 73, 963–969.
76.
Stone, J. L., O’Donovan, M. C., Gurling, H., et al. (2008) Rare chromosomal deletions and
duplications increase risk of schizophrenia. Nature, 455, 237–241.
77.
Levinson, D. F., Duan, J., Oh, S., et al. (2011) Copy number variants in schizophrenia:
Confirmation of five previous finding sand new evidence for 3q29 microdeletions and VIPR2
duplications. Am. J. Psychiatry, 168, 302–316.
78.
Auton, A., Abecasis, G. R., Altshuler, D. M., et al. (2015) A global reference for human genetic
variation. Nature, 526, 68–74.
23
79.
Lek, M., Karczewski, K. J., Minikel, E. V, et al. (2016) Analysis of protein-coding genetic variation
in 60,706 humans. Nature.
80.
Samocha, K. E., Robinson, E. B., Sanders, S. J., et al. (2014) A framework for the interpretation
of de novo mutation in human disease. Nat. Genet., 46, 944–950.
81.
Ware, J. S., Samocha, K. E., Homsy, J., et al. (2015) Interpreting de novo Variation in Human
Disease Using denovolyzeR. Curr. Protoc. Hum. Genet.
82.
The Gene Ontology Consortium (2017) Expansion of the Gene Ontology knowledgebase and
resources. Nucleic Acids Res.
83.
Smith, C. L., Blake, J. A., Kadin, J. A., et al. (2018) Mouse Genome Database (MGD)-2018:
Knowledgebase for the laboratory mouse. Nucleic Acids Res., 46, D836–D842.
24
Figure 1 Schizophrenia association of gene sets ranked by FMRP binding confidence. All expressed
genes were ranked by FMRP binding confidence and grouped into 21 bins of 400 genes. Presented are
-log10(P), where the P-value is derived from gene set association analysis using the genetic variant type
shown. CNV analyses were corrected for P-value inflation using random size-matched sets of
expressed genes. Rare coding variants were derived from case-control exome sequencing studies of
schizophrenia and defined as variants observed once in all sequenced samples and never in the non-
psychiatric component of ExAC. Loss-of-function variants include nonsense, splice site and frameshift
mutations. Nonsynonymous variants include loss-of-function and missense variants. Dotted lines
represent a threshold for statistical significance after correction for 21 tests. SNPs, single nucleotide
polymorphisms; CNVs, copy number variants.
25
Figure 2 Pathway analysis workflow. Predominant functional subsets of FMRP targets were tested for
genetic association with psychiatric disorders. GO, gene ontology; MP, mammalian phenotype; FDR,
false discovery rate.
26
Figure 3 Genetic association of FMRP target bins with schizophrenia, bipolar disorder and major
depressive disorder. Shown are -log10(P-value) following common variant gene set association analysis
of 21 bins of 400 genes ranked by FMRP binding confidence. Dotted lines represent a threshold for
statistical significance after correction for 21 tests.
27
Table 1 Partitioning FMRP targets genetic association by overrepresented functional annotation. GO
and MP functional terms independently overrepresented among FMRP targets were merged, then
divided by FMRP targets membership. Genes not brain-expressed were removed. Background
association originating from brain expression was controlled for within gene set association analyses.
Shown are the resulting effect sizes (β or Rate Ratio) and P-values (P). For each variant type, P-values
were Bonferroni adjusted for 3 tests. SNPs, single nucleotide polymorphisms; LoF, loss-of-function;
NS, nonsynonymous.
Gene set
N
Common SNPs
Rare LoF
De novo NS
β
P
β
P
Rate Ratio
P
Genes exclusive
to functional terms
1195
0.066
0.13
0.010
1.0
0.98
1.0
Overlapping genes
401
0.29
3.7 × 10-6
0.43
3.5 × 10-5
1.27
0.17
Genes exclusive to
FMRP targets
438
0.17
0.0063
0.34
0.0023
1.58
9.2 × 10-4
28
Term
Genes not FMRP targets
Genes FMRP targets
N
Beta
P
Padj
N
Beta
P
Padj
Calcium ion transmembrane
transporter activity
(GO:0015085)
91
0.419
4.7 × 10-4
0.017
25
1.080
6.9 × 10-6
2.4 × 10-4
Abnormal motor
coordination/balance
(MP:0001516)
538
0.104
0.028
0.97
117
0.463
2.8 × 10-5
9.6 × 10-4
Abnormal seizure response
to inducing agent
(MP:0009357)
125
0.190
0.043
1.0
42
0.710
1.1 × 10-4
0.0038
Abnormal spatial learning
(MP:0001463)
141
0.161
0.057
1.0
61
0.569
1.4 × 10-4
0.0049
Growth cone
(GO:0030426)
50
0.245
0.077
1.0
27
0.854
1.7 × 10-4
0.0060
Abnormal nest building
behaviour
(MP:0001447)
15
0.265
0.22
1.0
12
1.290
2.0 × 10-4
0.0071
Abnormal excitatory
postsynaptic currents
(MP:0002910)
60
0.177
0.12
1.0
35
0.715
3.5 × 10-4
0.012
Axon part
(GO:0033267)
108
0.134
0.12
1.0
54
0.505
6.6 × 10-4
0.023
Table 2 GO and MP terms overrepresented among FMRP targets which capture a significant (Padj <
0.05) portion of the common variant genetic association with schizophrenia. Shown are effect sizes
(Beta) and P-values (P) in gene set association analysis of genes targeted, or not targeted, by FMRP.
P-values were Bonferroni adjusted (Padj) for 35 terms.
29
Gene set
N
Schizophrenia
Bipolar disorder
Major depressive
disorder
β
P
β
P
β
P
Genes exclusive
to functional terms
1195
0.066
0.13
0.037
0.38
0.031
0.49
Overlapping genes
401
0.29
3.7 × 10-6
0.23
1.6 × 10-5
0.21
9.7 × 10-5
Genes exclusive to
FMRP targets
438
0.17
0.0063
0.14
0.0074
0.15
0.0026
Table 3 Partitioning FMRP targets common variant association by overrepresented functional annotation.
Analyses were performed using a background of brain-expressed genes to account for background association.
Shown are the effect sizes (β) and P-values (P) from gene set association analyses using MAGMA. For each
disorder, P-values were adjusted for 3 genes sets using the Bonferroni method.
30
ABBREVIATIONS
FMRP
Fragile X mental retardation protein
mRNA
Messenger ribonucleic acid
MAGMA
Multi-marker Analysis of GenoMic Annotation
GWAS
Genome-wide association study
DNA
Deoxyribonucleic acid
CNV
Copy number variant
GO
Gene ontology
MGI
Mouse genome informatics
MP
Mammalian phenotype
CYFIP1
Cytoplasmic FMR1 interacting protein 1
PGC
Psychiatric genomics consortium
| 2020 | Genetic association of FMRP targets with psychiatric disorders | 10.1101/2020.02.21.952226 | [
"Clifton Nicholas E",
"Rees Elliott",
"Holmans Peter A",
"Pardiñas Antonio F.",
"Harwood Janet C",
"Di Florio Arianna",
"Kirov George",
"Walters James TR",
"O’Donovan Michael C",
"Owen Michael J",
"Hall Jeremy",
"Pocklington Andrew J"
] | creative-commons |
Induction of hierarchy and time through one-dimensional probability space with
certain topologies
Shun Adachi∗
Department of Microbiology, Kansai Medical University,
2-5-1 Shin-machi, Hirakata, Osaka 573-1010, JAPAN
(Dated: 13 September 2019)
Background: In a previous study, the authors utilized a single dimensional operationalization of species density
that at least partially demonstrated dynamic system behavior.
Purpose: For completeness, a theory needs to be developed related to homology/cohomology, induction of the
time dimension, and system hierarchies.
Method: The topological nature of the system is carefully examined and for testing purposes, species density
data for a wild Dictyostelia community data are used in conjunction with data derived from liquid-chromatography
mass spectrometry of proteins.
Results: Utilizing a Clifford algebra, a congruent zeta function, and a Weierstraß ℘ function in conjunction
with a type VI Painlev´e equation, we confirmed the induction of hierarchy and time through one-dimensional
probability space with certain topologies. This process also served to provide information concerning interactions
in the model.
Conclusions: The previously developed “small s” metric can characterize dynamical system hierarchy and in-
teractions, using only abundance data along time development.
CONTENTS
I. Introduction
1
II. Field Research & Experiments
2
II.1. Field Research
2
II.2. Experiments
2
II.2.1. Cell culture
2
II.2.2. Protein experiments
2
II.2.3. LC/MS
3
III. Results
3
III.1. General guidelines for topological evaluations 3
III.2. O ∼= ∆ case
3
III.3. O ∼= C case
3
III.4. O ∼= ˆC case
6
III.5. Congruent zeta function
7
III.6. Further consideration of 1+1 dynamics
10
III.7. ℘ as evaluations for interactions
10
IV. Discussion
12
Acknowledgments
12
References
13
I.
INTRODUCTION
In a previous study, the authors developed a system
whereby a static set of species density information can be
∗ S. Adachi: f.peregrinusns@mbox.kyoto-inet.or.jp
utilized to predict dynamics therein by extracting proba-
bilistic information [1]. We developed a new complex sys-
tem measure, “small s”, related to a probability space.
When Nk is the individual density for the k-th ranked
species and is approximated by a logarithmic distribu-
tion with parameters a, b with respect to the ranks of the
values of individual densities,
Nk = a − b ln k,
(1)
and
ℜ(s) =
ln N1
Nk
ln k (k ̸= 1), ℑ(s) = e
ℜ(s)
b
E(N),
(2)
where E(N) is averaged species density.
For k = 1,
ζ(s) = E(N)
N1
for species, where ζ(s) is a Riemann zeta
function.
Therefore, it appears doubtful why single-
dimensional information (Nk), with a topology labelled
by rank k, can induce a 3-dimensional system (a, b, ln k,
regarding Nk as free energy, the others as internal energy
or enthalpy, temperature, and entropy, respectively) of
an individual density, accompanied with an even addi-
tional time dimension. To explain this, first of all, we set
a 1-dimensional C∞ manifold with a topology as (B, O)
with s ∈ B. Inspired by the Bethe ansatz (e.g. [2]), we
set three different topologies isomorphic to ∆, C, andˆC
for further clarification of our model. These topologies
naturally invest a cohomology, time dimension, and hier-
archy to the system. Furthermore, we are able to define
a proper topology independently from moduli of mea-
surements with individual numbers and a Galois action
dependent on moduli of it in an evolutionary system
with hierarchy by Galois extension, such as biological
systems in this case.
For application to biological hi-
erarchies, this model is tested using protein abundance
data derived from liquid-chromatography mass spectrom-
etry (LC/MS) of HEK-293 cells and species density data
2
from a wild Dictyostelia community. Finally, we sought
to evaluate interactions of the constituents of biologi-
cal systems by invoking a Weierstraß ℘ function to es-
timate the strength of homo- and hetero-interactions.
These results serve to further justify our “small s” met-
ric to decipher system dynamics of interest. For exam-
ple, adapted, non-adapted (neutral), and disadapted (re-
pressed) proteins can be classified by expansion of the
model using a Clifford algebra.
Furthermore, utilizing
a congruent zeta function elucidates the contribution to
adaptive/disadaptive situations from each hierarchy.
II.
FIELD RESEARCH & EXPERIMENTS
II.1.
Field Research
Data concerning the number of individuals in each
species were obtained from natural (nonlaboratory) en-
vironments.
The sampling is described in [3].
Field
experiments were approved by the Ministry of the En-
vironment, Ministry of Agriculture, Forestry and Fish-
eries, Shizuoka Prefecture and Washidu Shrine (all in
Japan). The approval Nos. are 23Ikan24, 24Ikan72-32,
and 24Ikan72-57. Soil samples were obtained from two
point quadrats in the Washidu region of Izu in Japan.
The number of individual cellular slime molds per gram
of soil was determined by counting the number of plaques
cultivated from soil samples.
Species were identified
by morphology and the DNA sequences of 18S rRNA
genes. Samples were obtained monthly from May 2012
to January 2013 inclusive.
Relevant calculations were
performed using Microsoft Excel 16.16.13 and SageMath
8.8.
In
more
detail,
sampling
occurred
using
two
100 m2 quadrats in Washidu (35◦3′33′′N, 138◦53′46′′E;
35◦3′45′′N, 138◦53′32′′E). Within each 100 m2 quadrat,
nine sample points were established at 5 m intervals.
From each sampling point, 25 g of soil was collected.
Cellular slime molds were isolated from these samples as
follows. First, one sample from each site was added to
25 ml of sterile water, resuspended, and then filtrated
with sterile gauze. Next, 100 µl of each sample solution
was mixed with 100 µl HL5 culture medium containing
Klebsiella aerogenes and spread on KK2 agar. After two
days of storage in an incubator at 22 ◦C, the number of
plaques on each agar plate was enumerated and recorded.
Note that the number of plaques corresponds to the to-
tal number of living cells at any possible stage of the life
cycle.
That is, the niche considered here is the set of
propagable individuals of Dictyostelia; these are not ar-
ranged in any hierarchy or by stage in the life cycle. Also,
note that we did not examine the age or size structure of
organisms, since most of these were unicellular microbes.
Mature fruiting bodies consisting of cells from a single
species were collected along with information regarding
the number of plaques in the regions in which each fruit-
ing body was found. Finally, spores were used to inocu-
late either KK2 for purification or SM/5 for expansion.
All analyses were performed within two weeks from the
time of collection. The isolated species were identified
based on 18S rRNA (SSU) sequences, which were ampli-
fied and sequenced using PCR/sequencing primers, as de-
scribed in [4] and the SILVA database (http://www.arb-
silva.de/).
The recipes for the media are described at
http://dictybase.org/techniques/media/media.html.
II.2.
Experiments
II.2.1.
Cell culture
A human HEK-293 cell line from an embryonic kidney
was purchased from RIKEN (Japan). The sampling is
described in [5]. The original cultures were frozen on ei-
ther March 18, 2013 (3-year storage) or March 5, 2014
(2-year storage). They were subsequently used in exper-
iments between February and June 2016. The strain was
cultured in Modified Eagle’s Medium (MEM) + 10% fatal
bovine serum (FBS) + 0.1 mM nonessential amino acid
(NEAA) at 37 ◦C with 5% CO2. Subculturing was per-
formed in 0.25% trypsin and prior to the experiment, the
original cells from RIKEN were frozen following the stan-
dard protocol provided by RIKEN: in culture medium
with 10% dimethyl sulfoxide (DMSO), they were cooled
until reaching 4 ◦C at −2 ◦C/min, held at that tempera-
ture for 10 min, then cooled until reaching −30 ◦C at −1
◦C/min in order to freeze, held at that temperature for
10 min, then cooled again until reaching −80 ◦C at −5
◦C/min, and finally held at that temperature overnight.
The next day, they were transferred to storage in liquid
nitrogen.
II.2.2.
Protein experiments
The HEK-293 proteins were extracted using the stan-
dard protocol for the RIPA buffer (NACALAI TESQUE,
INC., Kyoto, Japan).
The sampling is described in
[5]. Approximately 106 harvested cells were washed once
in Krebs-Ringer-Buffer (KRB; 154 mM NaCl, 5.6 mM
KCl, 5.5 mM glucose, 20.1 mM HEPES pH 7.4, 25 mM
NaHCO3).
They were resuspended in 30 µl of RIPA
buffer, passed in and out through 21G needles for de-
struction, and incubated on ice for 1 h. They were then
centrifuged at 10,000 g for 10 min at 4 ◦C, followed by
collection of the supernatants. The proteins were quan-
tified using a Micro BCA Protein Assay Kit (Thermo
Fisher Scientific, Waltham, U.S.A.) and further process-
ing was performed using XL-Tryp Kit Direct Digestion
(APRO SCIENCE, Naruto, Japan). The samples were
solidified in acrylamide gel, washed twice in ultrapure wa-
ter, then washed three times in dehydration solution, and
finally dried. The samples were then processed using an
In-Gel R-CAM Kit (APRO SCIENCE, Naruto, Japan).
The samples were reduced for 2 h at 37 ◦C, alkylated
for 30 min at room temperature, washed five times with
3
ultrapure water, washed twice with destaining solution,
and then dried. The resultant samples were trypsinized
overnight at 35 ◦C. The next day, the dissolved digested
peptides were collected by ZipTipC18 (Merck Millipore,
Corp., Billerica, U.S.A.). The tips were dampened twice
with acetonitrile and equilibrated twice with 0.1% triflu-
oroacetic acid. The peptides were collected by ∼ 20 cy-
cles of aspiration and dispensing, washed twice with 0.1%
trifluoroacetic acid, and eluted by 0.1% trifluoroacetic
acid /50% acetonitrile with aspiration and dispensing
five times × three tips followed by vacuum drying. The
final samples were stored at −20 ◦C. Before undertak-
ing LC/MS, they were resuspended in 0.1% formic acid,
and the amounts were quantified by Pierce Quantita-
tive Colorimetric Peptide Assay (Thermo Fisher Scien-
tific, Waltham, U.S.A.). This protocol is published at
http://dx.doi.org/10.17504/protocols.io.h4qb8vw.
II.2.3.
LC/MS
LC/MS was undertaken by the Medical Research Sup-
port Center, Graduate School of Medicine, Kyoto Univer-
sity with a quadrupole–time-of-flight (Q-Tof) mass spec-
trometer TripleTOF 5600 (AB Sciex Pte., Ltd., Concord,
Canada). Standard protocols were followed. The load-
ing amount for each sample was 1 µg. We extracted the
quantitative data for the unused information for iden-
tified proteins using ProteinPilot 4.5.0.0 software (AB
Sciex Pte., Ltd., Concord, Canada). For further details
see [5].
III.
RESULTS
III.1.
General guidelines for topological evaluations
We start from a 1-dimensional C∞ manifold with a
topology, (B, O). Note that many aspects of (B, O) can
be explained by the inverse square law by drawing on
forces in the models below.
This partial topology of O means, for example, a reg-
ular automorphism on ∆, f(∆) = {eiθ z−α
1−¯αz; z ∈ B, θ ∈
R, α ∈ ∆} can explain anything emanating from the set
of f, for example, isomorphism to R3 space as shown
in [1], and explored in more detail below.
An appar-
ently neutral particle system introduced with hierarchies
by Galois extension could be Gal(Q(ζn)/Q) ∼= (Z/nZ)×
when ζn is a cyclotomic field.
If GCD(n, m) is 1,
Gal(Q(ζnm)/Q) ∼= Gal(Q(ζn)/Q) × Gal(Q(ζm)/Q). This
would lead to a Kummer extension decomposed to species
with p identity [1].
For a topology of C, f(C) = {az + b; z ∈ B, a, b ∈ C}
and isomorphic to R4, later indicated as (3 + 1) dimen-
sions with a time dimension. Obviously interaction of a
complex metric, e.g.
s2, w2 in [1], can induce a time
dimension.
For a topology of ˆC, f(ˆC) = { az+b
cz+d; z ∈
B, a, b, c, d ∈ C} and isomorphic to R6(R3 × R3), later
indicated by letting R4 compact by inducing a hierarchy
as in [1].
Fundamentally, a simply connected subregion without
holes such as a Riemann surface induced during hierar-
chization is isomorphic and holomorphic to either ∆, C,
or ˆC. Schwarz-Christoffel mapping enables a conformal
transformation from polygons to one of those regions, and
the Widely Applicable Information Criterion (WAIC) has
a central role as an analogy to logarithmic velocity in
fluid mechanics calculated from D [1]. Without singular-
ity, this is straightforward to consider and we focus on
the case for singular points. As in the Bethe ansatz [2],
a single dimension z with a particular topology is able to
induce both a (3+1)-dimensional system and hierarchies.
III.2.
O ∼= ∆ case
The Riemann-Roch theorem states
l(D) − l(K − D) = deg(D) − g + 1,
(3)
where D is a divisor, K is a canonical divisor, and g is
a genus number. Let TB be a bundle. An interaction,
TB ˜×TB := ∪
p∈B TpB × TpB, becomes a 3-dimensional
C∞ manifold. Let open base elements of the manifold
be x, y, z, and the planes on the bases be X, Y, Z. If we
consider interactions of these bases, the left term of Eq.
(3) is 3, from g = 10 and deg(D) = 12.
Let F
F(z) = q
∞
∏
n=1
(1 − qn)2(1 − q11n)2 =
∞
∑
n=1
c(n)qn
(4)
be a totally real number field of degree g over Q, and
K be a totally imaginary quartic extension of F.
Let
D and Dint be simple algebras over K with D = es/b.
Let G = GU(D, α) with α being a second kind involu-
tion of D. Take a 3-dimensional ℓ-adic system in which
WE = ℜ(s) = ℓ, D× = p = |D|E(ΣN), GLd(E) = v =
ln Nk/ ln p, where WE denotes the Weil group of center
E as a Langlands correspondence [6] [1] [5]. ℓ is obviously
an ´etale (crystalline) topology independent of moduli Nk,
in the sense that a homomorphism of Noetherian local
rings is unramified and flat, and the object is a localiza-
tion of a finitely generated algebra of the origin [1]. These
p(ℓ)-adic geometries are analogical to real differentiables
and Clifford-Klein geometries as calculated later.
The
O ∼= ∆ case visualizes both persistence homology p and
´etale cohomology l.
III.3.
O ∼= C case
A Minkowski metric small s [1] can be utilized for a
time developing model when sin, cos of the metric are
converted to sinh, cosh. However, for more detailed anal-
ysis, another Minkowski metric in our model could be
sM = [ℑ(s)2(∆ℑ(s))2−(∆a)2−(∆b)2−(∆ ln k)2]
1
2 . (5)
4
In this sense, the world line of a species is identical
and a different species is non-zero, discretely depending
on ∆ℑ(s).
When we take ds2
M = a(V1)ds2
M1, ds2
M1 =
a(V2)ds2
M2, and so on.
ds2
M = ds′2
M due to a Lorentz
transformation and ln(sM) = ∑∞
i=1 ln a(Vi) becomes a
module when 2dsM = 0. A set of species can thus be
characterized by this module of sM. A Lagrangian could
be
L = −ϕℑ(s),
(6)
and a Hamiltonian could be
H = −ϕℑ(s)2
√
ℑ(s)2 + (H(t)D)2
ℑ(s)2 − (H(t)D)2 .
(7)
We can consider D′ ∼= Dint, G′ ∼= Gint, and a time
dimension is induced by some admissible isomorphisms
(Proposition 2.5.6 in [7]).
Note that ‘temperature’ b
and root of time t are closely correlated by t = b arg D
[1]. Now consider the Poincar´e conjecture, where every
simply connected closed n-dimensional manifold WE is
homeomorphic to n-dimensional sphere Sn. Let a Morse
function be f : WE → [a, b], in which a, b are regular val-
ues. Let f have critical points p, p′ that correspond to in-
dexes λ, λ+1 as time. Consider that Sn−λ−1 and Sλ cross
at a single point; this indicates the status of present. The
exchange of Morse functions would result in no new crit-
ical point appearing and disappearance of critical points
p, p′ (h-cobordism theorem).
This is what happens at
the present state following the time arrow. Remark that
p, p′ are linked to a Hecke ring via non-trivial zero points
of Riemann zeta [1], fulfilling the condition of the Yang-
Baxter equation. Thus this phenomenon is closely related
to an analogy to quantum entanglement and face models
[8, 9]. Of course, in the case of species, as species still
exist, they will reappear with different p values in this
model.
In this sense, for any labelling of time points τ ′ ∈ TS∗,
a potential for the Petersson-Weil metric is as follows:
ωW P = d(σT (τ ∨⊥ τ∗) − σT (τ ∨⊥ τ ′)),
(8)
when ∨⊥ is a quasi-Fuchsian Kleinian group [10].
The
‘mating’ represents the coupling of times corresponding
to p, p′.
Now consider p, p′ as characteristics on a field k, as
in d = p = 0 in [5]. Let E be a singular hyperelliptic
curve of the system. Real D will be a tensor product
of an endomorphism of E on ¯k and Q, approximately.
The resultant D is a quaternion field on Q. Take a set
of ln N as an ℓ-adic rational Tate module as in [5]. D
will only ramify at p, p′ or a point at infinity (c.f. [11]).
This restricts the possible direction of the time arrow to
vanish p, p′ only.
Generally, for species, we draw a picture of time de-
velopment when the observer is at k = 1.
For other
observations, we can simply take k → k′ shifts for the
calculations. That is, we can take a cyclotomic field re-
lated to the number of kmax. In this sense, time in the
context of a complex metric can be utilized and the world
line is in web form branched at each cross-section of p and
p′, not in parallel as discussed in some studies. For mov-
ing one distinct world line to another, we need velocity
H(t)D > ℑ(s).
Next, shift from p to l = ℜ(s) following the method
outlined in [1], and simply consider a combinatory func-
tion in a probability space, Γ(s + 1) = sΓ(s). This is
an example of a shift map. If we take a function similar
to a Γ function, we can observe discrete time develop-
ment merely by multiplying a master s function if we
know the particular s. That is, adding a single fractal
dimension in the past world (subtracting a single dimen-
sion from the future world by an observation) results in
a simple multiplication of s and master Γ(s). Therefore,
only evaluating an s of interest is sufficient for this aim.
Similarly, consider the Maass form of the Selberg
zeta function in [1] as calculating the mode of species
dynamics.
Stirling’s approximation would be Γ(s) ≈
√
2π
s ( s
e)s exp( 1
12s), and considering a first-order approx-
imation of the exponent with (1 + 1/12s) can suitably
approximate the situation with superstring theory of 12
dimensions. For further approximation, we need addi-
tional dimensions. Jacobian mapping independent of a
path λ
Φ(p) = (
∫
λ
φ1, · · · ,
∫
λ
φg) ∈ Cg/tΩZ2g = J
∼(B)
(9)
is one choice. If we know the master Maass form as the
invariant form for ρG(cG) = cGIdW when IdW is an iden-
tity mapping of a system of interest (Stone-von Neumann
theorem; [12–15]), differential operation does not cause
any difference in the form. This ensures the condition
for a suitable D-module and the accompanying derived
category.
Thus we can adopt a modified microlocally
analytic b function as ∂b = i∂ as a substitute for the dif-
ferential operation; i.e., ∂2
b = −∂2, rotating the form in
the angle of π, and ∂4
b = ∂4 = i.d., reverting back to the
original orientation of the form. An Ornstein-Uhlenbeck
operator would be L = − ∑d
i=1 ∂∗
i ∂i = ∑d
i=1 ∂2
b . Set-
ting a bounded Baire function h on Rd and f as a so-
lution of Lf = h− < h >, < h >=
∫
Rd h(x)g(x)dx,
E(h(W))− < h >= E(Lf(W)) means a deviation from
the expected function h value in the future. The oper-
ator ∂b is thus characterized for an operator calculating
a future state. ∂2
b could be an element of a D-module
as D ◦ D = i.d. Then ∂b would develop to analogies to
energy or momentum, ∂b/∂t = E or −∂b/∂xk = pxk
as variations of operators. The π/2 rotation of ℑ(s) in
[1] is thus justified by the modified b function.
Con-
sidering (3 + 1) dimensions with an interaction of two 2-
dimensional particles, this theory and transactional inter-
pretation of quantum mechanics [16] are suitable. If we
regard ∂k
b , k ∈ Z as ideals of a finitely generated Jacobson
radical, Nakayama’s lemma shows maintaining identity
before and after the operation means the module is zero.
Therefore, in this finite case, everything is an observant
and at least an infinite generation is required to achieve
the values out of zeros. That means, if we see something,
5
time is infinite. Hironaka’s resolution of singularities at
characteristic 0 implies such a mating of p, p′.
To resolve such a master relation, consider a form of
“velocity” as v ∈ TB. Then take a 2-dimensional space
consisting of s ∈ B. s(v, t) = p(v)+tq(v) as in a Lagrange
equation. The Gauss curvature of this surface K ≤ 0.
K ≡ 0 is only achieved when TB is time-independent,
and this TB with K = 0 is the time-invariant bundle us-
able for TB ˜×TB calculation for a 3-dimensional system
and 6-dimensional hierarchies. Additionally, the Legen-
dre transformation of the above equation is X = v, Y =
tv −s, Z = t and {v −q(v)} dY
dv = Y +p(v). K = 0 means
v = 0 and s = p(0) is the required solution. Furthermore,
s can be regarded as a Dirac measure (w is a counterpart
of mass and s = w + 1), and s′ = −s can be regarded as
a Schwartz distribution. Although addition is allowed in
the distribution, generally multiplication is not (we will
illustrate that it is feasible later). However, setting the
differential as ∂2
b , it becomes first order with a minus sign
and differentiation by time: t2 is plausible. For instance,
∫ ∫
· · ·
V
∫
s∆φdt =
∫ ∫
· · ·
V
∫
φ[∆s]dt +
∫
· · ·
S
∫
s[dφ
dν ]dS
−
∫
· · ·
S
∫
φ[ ds
dν ]dS,
(10)
where φ is a distribution of interest, s ∈ S, and ν is a
differential by unit area. The first term on the right is
noise, the second is related to fractal structure, and the
third is oscillative behavior. Besides singular points, it
is regular. An entire function considering negative even
singular points of l − n regarding w = s − 1 would be
Zl =
Pf.wl−n
π(n−2)/22l−1Γ( l
2)Γ( l+2−n
2
),
(11)
where at the singular points, k ∈ Z≥0, Z−2k = □kw; □ =
(−1)( ∂4
b
∂x4
1 + ∂4
b
∂x4
2 +· · ·+
∂4
b
∂x4
n−1 − ∂4
b
∂t4 ). In the ∅ = ∂B case,
□Z2 = w, □kZ2k = w. This means, periodical popula-
tion bursting/collapsing by negative even w values [1].
For negative odd w values, chaos ensues (ˇSarkovski, Ste-
fan, Block theorem) [17]. Thus, adopting s, w is suitable
for applying a single-dimensional model.
s is a mea-
sure provided it is finite in bounded domains.
There-
fore, singular points reflect appearance/disappearance of
fractal structures. In summary, a topology O should be
({m = k} ⊂ N, {ε = b}, {Ω = a}) of Nk = a − b ln k in
[1]. For further details regarding distributions, see [18].
Now let E be an elliptic curve: y2 + y = x3 − x2 as in
[19]. This is equivalent to y(y + 1) = x2(x − 1). If we
consider (3 + 1)-dimensional N = 1 SU(2) without fluc-
tuation, x2 could be mass, (x − 1) could be a goldstino
as spontaneous breaking of supersymmetry, y could be 3-
dimensional fitness D with fluctuation, and y+1 could be
(3 + 1)-dimensional s [1]. The goldstino would represent
temporal asymmetry. In Gaussian ensembles, a complex
system GUE breaks time-reversal and a self-dual quater-
nion system GSE preserves it. Therefore y + 1 preserves
time symmetry and consequently the present y breaks
the symmetry.
t
�
Γ
�
Dt
Γ(t)
F (a,b,c;z)
�{
{
{
{
{
{
{
{
A Riemann scheme would uniformize the fitness space
as a hypergeometric differential equation.
Now consider
dY
dx = (A
x +
B
x − 1)Y,
(12)
A =
λ1 + λ3 + λ4 + λ5
λ2
0
0
λ3 + λ4 λ5
0
0
0
,
(13)
B =
0
0
0
0
0
0
λ1(λ1+λ3+λ5)
λ5
λ1λ2+λ2λ3+λ3λ5)
λ5
λ2 + λ4 + λ5
.
(14)
This will culminate in a generalized hypergeometric func-
tion 3F2 that satisfies a Fuchs-type differential equation
3E2. If we set proper region ∆ (13 different regions),
y(x) =
∫
∆
sλ1(s − 1)λ2tλ3(t − x)λ4(s − t)λ5dsdt.
(15)
x = 0, w = D, s = 1 would result in
y(0) =
∫
∆
sλ1wλ2tλ3+λ4{−(t − 1)}λ5dsdt.
(16)
λ1 = λ2 = λ3 = λ4 = λ5 = 1 would be E2 :
−
∫
y(y + 1)x2(x − 1)dxdy form, obviously the integral
of the interaction of two elliptic curves.
C = {s/b}
�
exp.
�
C/∧ = C×/DZ
C× = D
time reversal
�n
n
n
n
n
n
n
n
n
n
n
For consideration of an interacting 4-dimensional sys-
tem, let us take Painlev´e VI equations on a (3 + 1)-
dimensional basis with a single Hamiltonian [20] [21].
The Hamiltonian should be Hk = ∂k ln τ(t) = ∂kτ(t)
τ(t)
=
H(t)Nk =
Nk
E(ΣN) = ϕ when H(t) is a Hubble parame-
ter [22] [1]. τ(t) is thus an inverse of a Hubble param-
eter, and its kth boundary is a kth species. Note that
the 3-dimensional system represents the smallest possi-
ble number of dimensions whose associativity equations
become non-empty even in the presence of the flat iden-
tity. Furthermore, considering a fundamental group π1 of
C0,n := P1\{z1, ..., zn}, the dimension of representations
ρ of π1 in SL(2, C) is 2(n−3) [22]. If we would like to set
π1 as an ´etale topology with 0 dimension, n = 3. (3+1)-
dimensional semisimple Frobenius manifolds constitute a
6
subfamily of Painlev´e VI:
d2X
dt2 = 1
2( 1
X +
1
X − 1 +
1
X − t)(dX
dt )2
−(1
t +
1
t − 1 +
1
X − t)dX
dt
+X(X − 1)(X − t)
t2(t − 1)2
[(θ∞ − 1
2)2
+θ2
0
t
X2 + θ2
1
t − 1
(X − 1)2 + (θ2
t − 1
4) t(t − 1)
(X − t)2 ].
(17)
Recall that the above equation is related to a rank 2
system:
dΦ
dz = (A0
z +
At
z − t +
A1
z − 1)Φ,
(18)
or
dA0
dt
= [At, A0]
t
, dA1
dt
= [At, A1]
t − 1
(19)
with 4 regular singular points 0, t, 1, ∞ on P1. Also,
A0 + At + A1 = −A∞ = diag{−θ∞, θ∞}.
(20)
Note that the total sum of the matrix system is equal to
0. Assuming a 3-wave resonant system [23],
∂τu1 + c1∂xu1 = iγ1u∗
2u∗
3
∂τu2 + c2∂xu2 = iγ2u∗
3u∗
1
∂τu3 + c3∂xu3 = iγ3u∗
1u∗
2
(21)
(22)
(23)
An expansion of this model results in the h11
V = h12
ˆV mir-
ror symmetry relation for the Calabi-Yau threefolds. Re-
call that matrix Painlev´e systems of two interacting sys-
tems
t(t − 1)HMat
VI (α, β, γ, δ, ω; t; q1, p1, q2, p2)
= tr[Q(Q − 1)(Q − t)P 2
+{(δ − (α − ω)K)Q(Q − 1) + γ(Q − 1)(Q − t)
−(2α + β + γ + δ)Q(Q − t)}P + α(α + β)Q],
(24)
has 11 parameters.
Now let us convert a Painlev´e VI equation to a more
realizable form as in physics. The Painlev´e VI equation
is equivalent to
d2z
dτ 2 =
1
(2πi)2
3
∑
j=0
αj℘z(z + Tj
2 , τ)
(25)
where (α0, ..., α3) := (α, −β, γ, 1
2 − δ), (T0, ..., T3) =
(0, 1, τ, 1 + τ), and ℘ is the Weierstraß℘ function (The-
orem 5.4.1 of [20]). Furthermore, any potential of the
3-dimensional normalized analytic form
Φ(x0, x1, x2) = 1
2(x0x2
1 + x2
0x2) +
∞
∑
n=0
M(n)
n!
e
n+1
r+1 x1xn
2
(26)
can be expressed through a solution to the Painlev´e VI
equivalent with (α0, ..., α3) = ( 1
2, 0, 0, 0), that is,
d2z
dτ 2 = − 1
8π2 ℘z(z, τ).
(27)
When q = D = eiπτ, the Picard solution of the τ func-
tion on the 4 dimensions that corresponds to the c = 1
conformal field blocks in an Ashkin-Teller critical model
would be
τPicard(t) = const ·
qσ2
0t
t
1
8 (1 − t)
1
8
ϑ3(σ0tπτ ± σ1tπ|τ)
ϑ3(0|τ)
, (28)
where
the
Jacobi
theta
function
is
ϑ3(z|τ)
=
∑
n∈Z eiπn2τ+2inz; trMµMν = 2 cos 2πσµν when the pa-
rameter space of (θ0, θt, θ1, θ∞) is M [24] [25] [22] [26].
For other algebraic solutions, see [27].
Let us calcu-
late a Clifford algebra in an n = 3 system [28]. First,
let the representation (ρ, V ) of the algebra Cln fulfill
the condition ρ : Cln ∋ ϕ �−→ ρ(ϕ) ∈ End(V ) with
ρ(ϕ)ρ(ψ) = ρ(ϕψ). When n is odd, for example, 3, there
are nonequivalent representations:
ρ+ : Cl3 ≃ C(2) ⊕ C(2) ∋ (ϕ, ψ) → ψ ∈ End(C2), (29)
ρ− : Cl3 ≃ C(2) ⊕ C(2) ∋ (ϕ, ψ) → ϕ ∈ End(C2).
(30)
For example, let us calculate a complex v, v′ by ℜ(v) = v,
ℑ(v) = e(ℜ(v)/b)E(N), ℜ(v′) = Nk/ℑ(v), and ℑ(v′) =
e(ℜ(v′)/b)E(N) as in [5]. The next complex v′′ is ℜ(v′′) =
Nk/ℜ(v′) and ℑ(v′′) = e(ℜ(v′′)/b)E(N). We can calculate
v′′′ by the same operator as before. We denote this situ-
ation RRR. Graphing the calculated ℑ(v′′′) values with
their rank among 800 proteins permits classification into
3 groups demarcated based on slope values, namely, val-
ues below 1.01, between 1.01 to 2.00, and above 2.00
(Fig. 1). The 0.30 value of Filamin-A was excluded be-
cause it probably mostly reflects adapted proteins in fi-
broblasts (HEK-293). The irreducible representations in
the raw LC/MS data of [5] are 4-dimensional 1–2 (aver-
age 1.368 ± 0.004, 99% confidence) in non-adapted situa-
tions and 3-dimensional 1 (average 1.001511 ± 0.000006,
99% confidence) in adapted situations, respectively (Sup-
plemental Table 1).
The remainder are probably re-
pressed (disadapted) proteins. In tensor algebra TB :=
⊕∞
n=0 B⊗n, B = ⊕
i∈I RXi, x ∈ X, x ⊗ x − q(x) ∈
R ⊕ B⊗2, x is a single fractal dimension (= w), and
the fractal dimension of q(x) is 1/2, 1 for non-adapted
and adapted stages, respectively [1]. We are thus able
to calculate a characteristic number related to protein
adaptation.
III.4.
O ∼= ˆC case
For the species data set (Table I) [1], consider that a
sequential operation is an exact form. As in [5], setting
operation III as ℜ(v) = v, ℑ(v) = µl = e(ℜ(v)/b)E(N),
7
FIG. 1. ℑ(v′′′) values versus their ranks.
ℜ(v′)
=
E[l]
=
l
=
ln(Nk)/ ln(ℑ(v)),
ℑ(v′)
=
e(ℜ(v′)/b)E(N), ℜ(v′′)
=
ln(Nk)/ ln(ℑ(v′)), ℑ(v′′)
=
e(ℜ(v′′)/b)E(N), v′′′ by ℜ(v′′′) = ln(Nk)/ ln(ℑ(v′′)), and
ℑ(v′′′) = e(ℜ(v′′′)/b)E(N), we have ℜ(v) ≃ ℜ(v′′) ≃ 0,
ℑ(v) ≃ ℑ(v′′) ≃ 0, ℜ(v′) ≃ ℜ(v′′′) ≃ 0, ℑ(v′) ≃ ℑ(v′′′) ≃
0 (Table 1), suggesting that an actual/potential of species
creates an actual/potential appearance of the adapted hi-
erarchy above two layers. Recall that this is a short ex-
act sequence; the morphism ℑ becomes monomorphism
and ℜ(ln) becomes epimorphism. Furthermore, Imℑ is
equal to Kerℜ(ln).
Obviously there also exists a ho-
momorphism h : ℑ(v′) → ℜ(v′), h : ℜ(v′′) → ℑ(v′),
h : ℑ(v′′) → ℜ(v′′) or h : ℜ(v′′′) → ℑ(v′′), and the
short exact sequence is a split. These are abelian groups
and ℜ(v′) ≃ ℑ(v) ⊕ ℑ(v′), ℑ(v′) ≃ ℜ(v′) ⊕ ℜ(v′′),
ℜ(v′′) ≃ ℑ(v′) ⊕ ℑ(v′′), ℑ(v′′) ≃ ℜ(v′′) ⊕ ℜ(v′′′). The
data show that an actual layer is a direct sum of a po-
tential layer below and a potential layer. The data also
show that a potential of the layer is a direct sum of a real
layer and a layer above the layer. Finally, defining a Ga-
lois action Gal(L/K), actions defined by ℜ(v′)/ℑ(v) ≃
ℑ(v′), ℑ(v′)/ℜ(v′) ≃ ℜ(v′′), ℜ(v′′)/ℑ(v′) ≃ ℑ(v′′), and
ℑ(v′′)/ℜ(v′′) ≃ ℜ(v′′′) are all Galois, achieving our goal
for defining proper Galois actions with a topology of v for
biological hierarchies. A species is thus likely to emerge
from the interaction of species.
ℜ(v)
ℑ
�
I(ℜ) � ℜ(v′)
ℑ
�
I(ℜ) � ℜ(v′′)
ℑ
�
I(ℜ) � ℜ(v′′′)
ℑ
�
ℑ(v)
ℜ(ln)
x
x
x
�x
x
x
I(ℑ)
� ℑ(v′)
ℜ(ln)
w
w
w
�w
w
w
I(ℑ)
� ℑ(v′′)
ℜ(ln)
v
v
v
�v
v
v
I(ℑ)
� ℑ(v′′′)
For species [1], consider that a sequential operation in
TABLE I. N values.
N
P. pallidum (WE) D. purpureum (WE) P. violaceum (WE)
May
0
76
0
June
123
209
52
July
1282
0
0
August
1561
0
0
September 901
107
0
October
1069
35
0
November 60
0
101
December 190
0
0
January
29
0
0
N
P. pallidum (WW) D. purpureum (WW) P. violaceum (WW)
May
0
83
0
June
147
0
0
July
80
215
320
August
1330
181
0
September 809
77
649
October
799
0
107
November 336
0
0
December 711
0
0
January
99
0
0
WE: Washidu East quadrat; WW: Washidu West quadrat
(please see [3]). Scientific names of Dictyostelia species: P.
pallidum: Polysphondylium pallidum; D. purpureum:
Dictyostelium purpureum; and P. violaceum:
Polysphondylium violaceum. N is number of cells per 1 g of
soil. Species names for Dictyostelia represent the
corresponding values. Red indicates ℜ(s) values of species
that were approximately integral numbers greater than or
equal to 2.
the previous sections is an exact form. As in [5], setting
an operation III, we have ℜ(v) ≃ ℜ(v′′) ≃ 0 and ℑ(v) ≃
ℑ(v′′) ≃ 0, but no further (Table 1), suggesting that
an actual/potential of species creates an actual/potential
appearance of the adapted hierarchy above two layers,
which diminishes in the three layers above. This might
reflect effects from different time scales among different
layers [3]. Similar to the previous section, ℜ(v′) ≃ ℑ(v)⊕
ℑ(v′) and ℑ(v′) ≃ ℜ(v′) ⊕ ℜ(v′′).
From the III morphisms, we can draw a short exact
sequence corresponding to ℜ(v) → ℑ(v) → l = ℜ(v′) →
l × (ℑ(s) = ℑ(v′)) → ℜ(v) = ℜ(v′′),
0 → A(u)
ι→ B(u)
sp
→ C(u ×
√
−1S∗) → 0,
(31)
regarding g = l as a specific spectrum of the Schwartz
distribution (or Sato hyperfunction [29, 30]) of a micro-
function sp g [31, 32]. Not only addition, but also multi-
plication is feasible for −s in this regard.
III.5.
Congruent zeta function
Hereafter we will adhere to the situation where O ∼= ˆC.
For the other aspect, instead of ℑ(v′), we can consider
Z/lZ, by 1/l-powered ℑ(v′), state a p-adic number cor-
respondence, and then take a valuation of it. Universal
coefficient theorems [33],
0 → Ext(Hq−1(X, A), G) → Hq(X, A; G)
→ Hom(Hq(X, A), G) → 0,
(32)
8
could be described as
0 → µl → E[l] → Z/lZ → 0,
(33)
making an exact sequence, with ℜ(s) value in the mid-
dle level between populational ℜ(v) value and its fractal
ℜ(v′′) value. E[l] → Z/lZ is an injection and Z/lZ → 0
an epimorphism. The image of the former is the kernel
of the latter. Homology backwards is a homomorphism
of the cohomology, and the exact sequence splits. These
are abelian groups and E[l] ∼= µl ⊕Z/lZ; Z/lZ ∼= E[l]⊕0.
A real level is constituted by a direct sum of a potential
level below and its own potential. A potential level is
constituted by a direct sum of a real level below and a
real level above. E[l]/µl ∼= Z/lZ; Z/lZ/E[l] ∼= 0 are Ga-
lois actions and a representation of an ´etale topology ℓ is
obtained, concomitantly with information of interactions
among different levels of hierarchies. Species should ap-
pear two layers above the population layer. [3] reports
results where the point mutation rate is on the order of
10−8 and speciation is on the order of 10−25, roughly
above a square of 10−8 over 10−8. This calculation could
be modeled by a simple critical phenomenon of dendro-
gram percolation.
In this model, approaching 1/2 − 0
probability of mutation maintenance leads to divergence
in cluster size. Regarding non-trivial ζ(w) = 0 as a seed
for speciation, a ∼ 108 population is on the same order as
a branch for being identical to ancestors or different from
them at each genome base pair. A dendrogram can be
regarded as a phylogenetic tree for dividing cells, which
is common to both asexually propagating organisms and
a constituent of sexually reproducing organisms at the
level of cell division of germ line cells, strictly correlated
to mutation during cell cycle processes. These facts ex-
hibit ℓ and Galois actions can adequately describe inter-
hierarchical interactions.
The
logic
above
would
suggest
application
of
Grothendieck groups. Let the situation be a Noetherian
ring, i.e., B is the ring. Let F(B) be the set of all isomor-
phisms of B-modules. Let CB be the free abelian group
generated by F(B). The short exact sequence above is
associated with (µl) − (E[l]) + (Z/lZ) of CB (() is an
isomorphism). Let DB be the subgroup of CB. The quo-
tient group CB/DB is a Grothendieck group of B related
to potential of s, w layers, denoted by K(B). If E[l] is
a finitely generated B-module, γ(E[l]) would be the im-
age of (E[l]) in K(B). There exists a unique homomor-
phism λ0 : K(B) → G such that λ(E[l]) = λ0(γ(E[l]))
for all E[l] when G is an abelian group of the B-module.
This representation corresponds to the Stone-von Neu-
mann theorem in this restricted situation.
B is gen-
erated by γ(B/p) when p corresponds to species in a
biological sense. If B is a principal ideal domain con-
stituting a single niche without cooperation of distin-
guished niches, K(B) ∼= Z, and this is suitable when
considering biological numbers for individuals. Consid-
ering different E[l], Ml, and Nl, and the set of all iso-
morphisms of a flat B-module F1(B), γ1(Ml) · γ1(Nl) =
γ1(Ml ⊗ Nl); γ1(Ml) · γ(Nl) = γ(Ml ⊗ Nl); K1(A) ∼= Z
with tensor products.
Furthermore, if B is regular,
K1(B) → K(B) is an isomorphism.
The sum of in-
teractions for different niches (not interacting between
distinguished niches) is thus calculable as integers by a
Grothendieck group. If the calculation does not lead to
integers, the situation involves interactions among dis-
tinguished niches. Algebraic expansion of this ring thus
introduces entirely different niches to the original ring.
If a ∈ K, f(x) = xl − x − a, α ∈ ¯K, f(α) = 0, α /∈ K(α ∈
∂K), f(x) is irreducible on K, L = K(α) is a Galois ex-
tension, and Gal(L/K) ∼= Z/lZ. α is from the hierarchy
above based on a new ideal.
To unify the sections introducing Galois Hi and the
preceding sections regarding the time arrow, consider
X, Y , which are eigen and smooth connected algebraic
curves on an algebraic closed field.
Hi(X¯k, Qℓ)
pr∗
1
−−→ Hi(X¯k ׯk Y¯k, Qℓ)
∪cl(γ)
−−−−→ Hi+2d(X¯k ׯk Y¯k, Qℓ(d))
pr2∗
−−−→ Hi(Y¯k, Qℓ),
(34)
when γ is an algebraic correspondence from Y to X. If
we assume X and Y correspond to different time points,
the above diagram,
γ∗ : Hi(X¯k, Qℓ) → Hi(Y¯k, Qℓ)
(35)
describes the time development of the system. To dissect
the contributions of each component on the time devel-
oping system, let κm be an m-dimensional expansion of
κ, which is a finite field of a residue field of an integer
ring OK on K. When the eigen smooth scheme Y is on
κ,
2d
∑
i=0
(−1)iTr(Frobm
v ; Hi(Y¯k, Qℓ)) = ♯Y (κm)
(36)
[34] [35].
When Y is finite, a congruent zeta function is
Z(Y, T) = exp(
∞
∑
n=1
♯Y (κn)
n
T n).
(37)
Setting
Pi(Y, T) = det(1 − FrobvT; Hi(Y¯k, Qℓ))
(38)
results in
Z(Y, T) =
2 dim Y
∏
i=0
Pi(Y, T)(−1)i+1.
(39)
To separate each contribution of Hi, consider Weil con-
jectures [36] [37], and Pi(Y, T) and Pj(Y, T) are dis-
joint when i ̸= j. Pi(Y, T) and Tr(Frobm
v ; Hi(Y¯k, Qℓ))
are thus calculable and this deciphers each contribution
of Pi(Y, T)s. Examples of the calculation are provided
in Tables II & III. Generally, large positive zeta values
represent highly adapted situations, whereas large neg-
ative zeta values represent highly disadapted situations
9
TABLE II. Calculations for Washidu East quadrat
Z (congruent) P. pallidum D. purpureum P. violaceum
May
-
-
-
June
0.009378
151.1
9.272
July
-
-
-
August
-
-
-
September
114.7
30.89
-
October
334.6
-540.4
-
November
0.02561
-
-54.13
December
-
-
-
January
-
-
-
P0
P. pallidum D. purpureum P. violaceum
May
-
-
-
June
-1.288
-0.06806
-0.1520
July
-
-
-
August
-
-
-
September -1.163
-0.7248
-
October
-0.8250
0.02954
-
November -1.002
-
0.1790
December -
-
-
January
-
-
-
P1
P. pallidum D. purpureum P. violaceum
May
-
-
-
June
0.7635
-10.29
-1.253
July
-
-
-
August
-
-
-
September -133.4
-22.39
-
October
-276.1
-15.96
-
November 0.7480
-
-9.689
December -
-
-
January
-
-
-
P2
P. pallidum D. purpureum P. violaceum
May
-
-
-
June
-63.23
1.000
0.8886
July
-
-
-
August
-
-
-
September 1.000
1.000
-
October
1.000
0.9999
-
November -29.13
-
1.000
December -
-
-
January
-
-
-
P. pallidum: Polysphondylium pallidum; D. purpureum:
Dictyostelium purpureum; P. violaceum: Polysphondylium
violaceum. - are undefinable.
and zero values are neutral situations.
P0, P1, P2 cor-
respond to ℜ(v), ℜ(v′), ℜ(v′′).
For ℜ(v), ℜ(v′′), values
close to zero represent large contributions, and for ℜ(v′),
large values represent large contributions. The inverses
of ℜ(v), ℜ(v′′) scale for ℜ(v′). The important point here
is that by utilizing a congruent zeta function, we can
visualize a contribution from each hierarchy.
From these theorems, we can deduce that P2 is a pen-
cil on elliptic curves with a section of order two and an
additional multisection. Setting ζ = e2πi/3 = (eπi/3)2 on
the initial condition of P2 at the point xa = 0,
t = ζ + 1, X(ζ + 1) =
1
1 − ζ , X′(t) = 1
3.
(40)
In the PzDom model [1], 1/ℑ(s−1) ≈ eπi/3 for predicting
TABLE III. Calculations for Washidu West quadrat.
Z (congruent) P. pallidum D. purpureum P. violaceum
May
-
-
-
June
-
-
-
July
8.135
0.002196
97.00
August
123.7
29.31
-
September
26.54
-106.1
0.0001892
October
99.51
-
26.36
November
-
-
-
December
-
-
-
January
-
-
-
P0
P. pallidum D. purpureum P. violaceum
May
-
-
-
June
-
-
-
July
-0.1936
-2.208
-0.1174
August
-1.306
-0.9804
-
September -0.6856
0.08601
-8.157
October
-1.141
-
-0.7729
November -
-
-
December -
-
-
January
-
-
-
P1
P. pallidum D. purpureum P. violaceum
May
-
-
-
June
-
-
-
July
-1.483
0.9810
-11.39
August
-161.6
-28.74
-
September -18.19
-9.126
1.000
October
-113.6
-
-20.37
November -
-
-
December -
-
-
January
-
-
-
P2
P. pallidum D. purpureum P. violaceum
May
-
-
-
June
-
-
-
July
0.9417
-202.2
1.000
August
1.000
1.000
-
September 1.000
1.000
-647.9
October
1.000
-
1.000
November -
-
-
December -
-
-
January
-
-
-
P. pallidum: Polysphondylium pallidum; D. purpureum:
Dictyostelium purpureum; P. violaceum: Polysphondylium
violaceum. - are undefinable.
the future and t is an addition of 1 to interactive (eπi/3)2
if ℜ(s − 1) is neglectable. When in close proximity to
trivial zero points of Riemann ζ, t ∼ 1 and X(t) ∼ 1.
X′(t) = 1
3 thus represents a (2 + 1)-dimensional system.
System dimensions are thus reduced to 2+1. For re-
producing the kernels, let q be in (Q∞)Γ(H∗). Then,
q(w)dw2 = 12
π (
∫
H
q(¯z)ℑ(z)2
(z − w)4 |dz|2)dw2,
(41)
where w = α/β and z := (αζ + ¯α)/((βζ + ¯β). The term
in parentheses is the reproduced kernel (Prop. 5.4.9 of
[10]).
Now consider q difference Painlev´e VI with ˆgl3 hierar-
chy. q could be equal to −s, and y(x + 1) = 1−qx
1−q y(x) =
10
(∑x−1
i=0 qi)y(x) can be converted from q to −s, when
x → ∞.
Setting |q| > 1, t as an independent variable, and f, g
as dependent variables,
T(g) = (f − ta1)(f − ta2)b3b4
g(f − a3)(f − a4)
, T −1(f)
= (g − tb1)(g − tb2)a3a4
f(g − b3)(g − b4)
,
(42)
where
f = −A12
0
A12
1
, g =
(A12
0 + x1A12
1 )(A12
0 + x1qα1+1A12
1 )
q(A11
0 (A12
1 )2 − A11
1 A12
0 A12
1 + qβ2+1(A12
0 )2).
(43)
A12
0 = qα1+α2+2x1x2ω13 ¯w32,
(44)
A12
1 = qα1+1x1ω11 ¯w12 + qα2+1x2ω12 ¯w22,
(45)
A11
0 = qα1+α2+2x1x2(1 + ω13 ¯w31),
(46)
A11
1 = −qα1+1x1(1 + ω12 ¯w21 + ω13 ¯w31)
−qα2+1x2(1 + ω11 ¯w11 + ω13 ¯w31)
(47)
and considering q
=
−b ln D of the PzDom model
[1], local time development can be easily calculated.
(a1, a2, a3, a4); (b1, b2, b3, b4) have 4 parameters interact-
ing with each other in this soliton equation of similarity
reduction [38] [23]. In other words, we are treating a di-
rect sum of two Virasoro algebras, or a Majorana fermion
and a super-Virasoro algebra [25].
III.6.
Further consideration of 1+1 dynamics
There is another way of considering system dynam-
ics with q, starting from a Young tableau. Let S be a
finite or countable set, for example, as the measures of
species density as SpecZ. For ℜ(s) ≤ 1/2, let an absolute
value of an absolute zeta function ζK = ζGm/F1(x, y) =
|
s(x,y)
(s−1)(o,y)|; x, y ∈ S where Gm = GL(1). For ℜ(s) > 1/2,
and let an absolute value of an inverse of an absolute
zeta function ζK =
1
ζGm/F1(x,y) = | (s−1)(o,y)
s(x,y)
|; x, y ∈ S.
ζK becomes a Martin kernel.
Let a distance function
Dδ(x, y) = ∑
z∈S Cz(|ζK(z, x) − ζK(z, y)| + |δzx − δzy|),
where δ is the delta function.
For a distance space
(S, Dδ), a topology of S determined by Dδ is a discrete
topology and (S, Dδ) is totally bounded. A completion
of (S, Dδ) will be set as ˆS.
Let a Martin boundary
∂S = ˆS\S be a (d − 1)-dimensional species density not
restricted to a random walk or transition probability. Sd
represents all possibilities of Sd−1 with a time dimen-
sion. Furthermore, a set of Sd−1 can be expressed by a
Young tableau in a Frobenius coordinate system. Taking
a Maya diagram of the tableau distributes the data to
a single dimension. Therefore, the 3-dimensional system
is in fact represented as a 1-dimensional system, a set of
F1 = Fq. In this context, a set of the individual numbers
of species is over Z and a time X is a flat algebra Λ-space
over Z. A Λ-structure on X is ψp : X → X, where ψ is
X ×SpecZ SpecFpc. In other words, Λ = Z[Gal(Z/Fpc)].
pc = 1 when there is no hierarchy/period in our anal-
ysis and, for example, pc = 2 in protein or species
data sets described above. Therefore, the hierarchy ex-
tends from F1 to F2.
Mn/F1 = HomGm/F1(An, An) =
ζK; GLn/F1 = AutGm/F1(An) = Sn and thus s ∈ Gm and
s − 1 ∈ F1 when ℜ(s) ≤ 1/2 and s − 1 ∈ Gm and s ∈ F1
when ℜ(s) > 1/2. q ∈ Gm and Spec(q) is Spec(s) or
Spec(s − 1). Since D = es/b is calculable in [1] with a
root of time t, temperature bt at time point t2 and tem-
perature bt−1 at time point (t−1)2 when time is properly
scaled, the dynamics of q can be calculated by this basal
information. See [39] for further details in this respect as
relates to Grothendieck’s Riemann Roch theorem. This
is another explanation as to why a 1-dimensional system
with a certain topology leads to 3 + 1 dynamics.
III.7.
℘ as evaluations for interactions
Take Wallis’ formula:
lim
n→∞
1
√n ·
2 · 4 · • • • · (2n)
1 · 3 · • • • · (2n − 1) = √π.
(48)
The upper product of even numbers could be a product
of bosonic multiplications, and the lower product of odd
numbers could be that of fermionic multiplications. The
square of them divided by n as an average number of
actions would result in π.
π is thus the number ratio
of boson multiplications and fermion multiplications. In
other words, an area of a circle corresponds to boson ac-
tions and the square of the radius corresponds to fermion
actions. Globally there are ∼ 3 times more bosonic ac-
tions than fermionic actions. For further expansion for
the bosonic even −w (without w = 0) with µ(n) = 1
[1], Weierstraß ℘(1/n) = ∑negative even̸=0
w=−2
(1/n)w and a
((w/2+1)×n)(n×1) matrix would calculate a set of patch
quality Pw of bosons involving a future status of w = −2.
Similarly, even −s with µ(n) = −1 [1], −℘(1/n) =
− ∑negative even̸=0
s=−2
(1/n)s, and a ((s/2 + 1) × n)(n × 1)
matrix would calculate a set of patch quality −Ps of
fermions involving a future status of s = −2. Regard-
ing w = s−1, P(w) = Pw −Ps=w+1 = ζ(w)+n+n2 and
the Riemann ζ function can be related to patch quality.
Population bursts with these even w (odd s) could be
calculated by Pw → +∞ with negative even w (negative
odd s), or in lower extent of bursting, Ps → ∓∞ with
w → 1 ∓ 0(s → 2 ∓ 0). Since P(0) ̸= 0 and P(0) → +∞,
considering P(w) = ℘(1/n)+℘(1/n)/n and ak, bk as zero
11
TABLE IV. Weierstraß ζ values.
Weierstraß ζ WE P. pallidum
WE D. purpureum
WE P. violaceum
May
June
2.290e15 - 5.081e15*I
5.648e51 + 1.513e52*I
3.036e32 + 1.2783e32*I
July
August
September
-9.284e28 - 2.716e28*I -1.501e23 + 3.448e23*I
October
-3.307e36 - 2.666e37*I -1.220e35 - 2.047e35*I
November
3.579e14 - 1.003e15*I
2.065e59 + 9.395e59*I
December
January
Weierstraß ζ WW P. pallidum
WW D. purpureum
WW P. violaceum
May
June
July
2.329e35 + 1.735e35*I 7.052e8 - 2.352e10*I
4.950e53 + 1.630e54*I
August
-4.121e28 - 1.547e28*I -1.075e22 + 3.286e22*I
September
1.493e39 + 1.008e39*I 6.076e48 + 1.023e49*I
1.220 + 0.02924*I
October
-4.379e28 - 1.562e28*I
-1.440e22 + 4.328e22*I
November
December
January
WE: Washidu East quadrat; WW: Washidu West quadrat;
P. pallidum: Polysphondylium pallidum; D. purpureum:
Dictyostelium purpureum; P. violaceum: Polysphondylium
violaceum. Weierstraß ζ are calculated from ℘ on an elliptic
curve [0, 1], expanded to 30-th order. Constants of
integration were neglected for ζ′ = −℘.
points and poles of the function,
fP (1/n) = CP
∞
∏
k=1
℘(1/n) − ℘(ak)
℘(1/n) − ℘(bk)
×
∞
∏
k=1
℘(1/n)/n − ℘(ak)/n
℘(1/n)/n − ℘(bk)/n = 0
(49)
because the constant CP = 0 when w = 0 [40]. Thus
w = 0(s = 1) means every singularity can be considered
as a zero ideal adopting fP . w → 0 means a general limit
of limw→0 ln s
w = 1. We can regard a logarithm of s as a
fitness when the fitness is sufficiently small. A fixed point
of the observer at s = 1 implies everything combined
to the zero ideals.
If we regard Weierstraß ζ(z; Λ) =
1
z + ∑
w∈Λ∗(
1
z−w + 1
w +
z
w2 ) (not Riemann zeta) as a
distribution function, an additive operation for fractal
dimensions s1, s2 results in
ζ(s1 + s2) = ζ(s1) + ζ(s2) + 1
2
℘′(s1) − ℘′(s2)
℘(s1) − ℘(s2) .
(50)
This means the third term on the right is a contribution
of different fractal hierarchies, besides a direct sum of
distribution functions. Tables IV to VII— present val-
ues for the Weierstraß zeta function, Weierstraß ℘, ℘′,
and interaction terms.
Note that at Washidu West in
September, Pv-Dp-Pp interacted strongly in that order.
In October, there is also a strong interaction of Pv-Pp.
Compared with Washidu West, Washidu East exhibited
weaker interaction and was dominated by Pp.
For further clarification, regarding ℘ as an elliptic func-
tion,
℘′2 = 4℘3 − g2℘ − g3
(51)
is a normal form without multiple root. Rationals exist,
F(℘(u)), G(℘(u)) as Legendre canonical forms of elliptic
TABLE V. ℘ values.
℘
WE P. pallidum
WE D. purpureum
WE P. violaceum
May
June
1.709e16 + 9.720e15*I -2.966e51 + 1.066e51*I -1.304e32 + 2.778e32*I
July
August
September 5.052e28 - 1.081e29*I
-5.829e23 - 4.100e23*I
October
1.694e37 - 1.838e35*I
1.622e35 - 7.066e34*I
November 3.600e15 + 1.691e15*I
-9.903e58 + 2.100e58*I
December
January
℘
WW P. pallidum
WW D. purpureum
WW P. violaceum
May
June
July
-1.334e35 + 1.676e35*I 1.259e11 + 1.727e10*I
-2.717e53 + 7.949e52*I
August
2.719e28 - 4.871e28*I
-6.167e22 - 3.546e22*I
September -5.774e38 + 7.958e38*I -1.837e38 + 7.905e37*I -0.8200 + 4.042*I
October
6.491e22 - 1.137e22*I
3.434e17 - 1.790e18*I
November
December
January
WE: Washidu East quadrat; WW: Washidu West quadrat;
P. pallidum: Polysphondylium pallidum; D. purpureum:
Dictyostelium purpureum; P. violaceum: Polysphondylium
violaceum. ℘ were calculated from an elliptic curve [0, 1],
expanded to 30-th order.
TABLE VI. ℘′ values I.
℘′
WE P. pallidum
WE D. purpureum
WE P. violaceum
May
June
3.841e16 - 5.488e16*I
1.938e50 + 5.613e50*I
2.450e32 + 1.2734e32*I
July
August
September -1.181e29 - 7.904e28*I -9.494e23 + 8.939e23*I
October
1.048e36 - 1.027*I
-3.553e34 - 1.217e35*I
November 7.322e15 - 1.234e16*I
2.058e57 + 1.008e58*I
December
January
WE: Washidu East quadrat; WW: Washidu West quadrat;
P. pallidum: Polysphondylium pallidum; D. purpureum:
Dictyostelium purpureum; P. violaceum: Polysphondylium
violaceum. ℘′ were calculated from an elliptic curve [0, 1],
expanded to 30-th order, and differentiated.
TABLE VII. ℘′ values II.
℘′
WW P. pallidum
WW D. purpureum
WW P. violaceum
May
June
July
1.162e35 + 9.874e34*I 1.594e11 - 6.432e11*I
1.231e52 + 4.372e52*I
August
-5.395e28 - 4.181e28*I -9.400e22 + 1.055e23*I
September 4.088e38 + 3.182e38*I 3.442e47 + 6.318e47*I
14.97 - 8.365*I
October
-5.744e28 - 4.312e28*I
-1.223e23 + 1.360e23*I
November
December
January
WE: Washidu East quadrat; WW: Washidu West quadrat;
P. pallidum: Polysphondylium pallidum; D. purpureum:
Dictyostelium purpureum; P. violaceum: Polysphondylium
violaceum. ℘′ were calculated from an elliptic curve [0, 1],
expanded to 30-th order, and differentiated.
TABLE VIII. Hetero-interaction terms.
hetero-interaction WE
WW
Pp-Dp (June)
0.001160 - 0.09419*I
Pp-Dp (Jul) 0.01142 - 0.3558*I
Pp-Pv (June)
0.01818 - 0.4494*I
Pp-Pv (Jul)
0.0008154 - 0.08021*I
Dp-Pv (June)
0.001160 - 0.09419*I
Dp-Pv (Jul) 0.0008154 - 0.08021*I
September
0.09055 - 0.5885*I
August
0.09149 - 0.6049*I
October
0.03433 - 0.3021*I
Pp-Dp (Sep) -2.372e8 + 3.704e8*I
November
0.0003791 - 0.05081*I Pp-Pv (Sep) 0.008871 - 0.2633*I
Dp-Pv (Sep) -1.659e8 - 1.791e9*I
October
-3.728e5 - 3.975e5*I
WE: Washidu East quadrat; WW: Washidu West quadrat;
P. pallidum, Pp: Polysphondylium pallidum; D. purpureum,
Dp: Dictyostelium purpureum; P. violaceum, Pv:
Polysphondylium violaceum.
12
TABLE IX. F values and contributions.
F
WE
major
WW
major
Pp-Dp (June) 0.7693+8.182*I Pp
Pp-Dp (Jul) 0.5752+5.331*I Dp
Pp-Pv (June) 0.7693+8.182*I Pp
Pp-Pv (Jul)
1.262+39.31*I
Pp
Dp-Pv (June) 1.258+31.10*I
Pv
Dp-Pv (Jul) 0.5752+5.331*I Dp
September
3.078+14.99*I
Pp
August
2.879+13.80*I
Dp
October
4.907+38.67*I
Dp
Pp-Dp (Sep) 1.790+53.12*I
Pp
November
0.7481+7.726*I Pp
Pp-Pv (Sep) 0.3186+2.028*I Pv
Dp-Pv (Sep) 0.3186+2.028*I Pv
October
2.905+13.93*I
Pv
WE: Washidu East quadrat; WW: Washidu West quadrat;
Pp: Polysphondylium pallidum; Dp: Dictyostelium
purpureum; Pv: Polysphondylium violaceum. Major: species
that had a major impact on dynamics.
TABLE X. g2, g3 values I.
Normal form
g2
g3
int. const. synch.
Pp-Dp (June) 3.066e103-2.529e103*I -7.698e119+1.343e119*I -
+
anti
Pp-Pv (June) -2.408e65-2.899e65*I
1.298e81+7.295e81*I
+
-
anti
Dp-Pv (June) 3.066e103-2.529e103
-3.028e135-1.182e136*I
-
+
September
-3.651e58-4.368e58*I
-3.373e81-4.044e82*I
+
+
for
October
1.159e75-2.991e73*I
-1.859e110+8.674e109*I -
+
anti
November
3.747e118-1.663e118*I -1.630e134-3.473e132*I
-
+
integrals, such that any elliptic function f(u) = F(℘) +
G(℘)℘′. Thus a particular state during time procedure ℘′
can be related to any elliptic function form by a particular
pair of Legendre canonical forms. Utilizing Weierstraß ℘
is thus closely related to abstraction of interaction of the
states, with a cube of ℘ itself.
Setting Ω as a period
of f(u), the canonical form K(Ω) ∼= C[x, y]/(y2 − 4x3 +
g2x + g3), where C[x, y] is an integral domain. The ideal
thus characterizes the observation phenomena related to
F, G.
To develop the evaluation, s can be regarded as the el-
liptic function f(u) via p, l double periodicity, and a lin-
ear plot of f(u) against ℘′ shows F, G values. Basically,
due to empirically massive values for ℘′, G ∼ 0 and F
are almost identical to either of the s values selected for
calculating the interaction. By this method, one can eval-
uate which of the interacting partners plays a major role
in the interaction. The results are shown in Table IX; in
WE, the climax species Pp dominated, while in WW, pi-
oneering species Dp and Pv had significant roles [3]. Note
that F, G are solutions for corresponding hypergeometric
differential equations. Thus g2, g3 become apparent dur-
ing the time development process. ω can be calculated by
g2 = 60 ∑
ω∈Λ′
1
ω4 , g3 = 140 ∑
ω∈Λ′
1
ω6 . Riemann’s theta
relations showed how a (3 + 1)-dimensional system could
be rearranged to a 2 + 2 system. Table X & XI shows
calculated values for g2, g3 in normal form of the elliptic
curves.
IV.
DISCUSSION
Here we move to some more miscellaneous parts asso-
ciated with eliminating fluctuations. Regarding the uti-
lization of hyperbolic geometry (logarithmic-adic space)
and blowing up for resolution of singularity, see our ear-
lier work [5].
From generalized function theories, the
idea of cohomology naturally emerges and if we set op-
erator III in terms of cohomology, the Hp = 0(p ≥ 1)
TABLE XI. g2, g3 values II.
Normal form g2
g3
int. const. synch.
Pp-Dp (Jul) -4.118e70-1.788e71*I
2.322e82+2.096e81*I
+
-
anti
Pp-Pv (Jul)
-1.728e107+2.700e107*I -6.829e142+7.054e141*I +
+
for
Dp-Pv (Jul) 1.691e198-1.728e107*I
-3.697e118+1.709e118*I -
+
anti
August
-1.060e58-6.532e57*I
-2.706e79-8.852e80*I
+
+
for
Pp-Dp (Sep) -9.168e77-4.559e78*I
-5.398e116-7.349e116*I
+
+
for
Pp-Pv (Sep) -1.199e78-3.676e78*I
-1.584e79+1.834e78*I
+
+
Dp-Pv (Sep) 1.099e77-1.162e77*I
-3.794e77-5.397e77*I
-
+
October
1.634e46-5.905e45*I
4.963e63+3.128e64*I
-
-
int.: positive or negative effect of an interaction term on ℘′
dynamics; const.: positive or negative effect of a constant on
℘′ dynamics; synch.: coupling between g2 and g3 against the
dynamics.
(p are primes and 1) cohomology and the Kawamata-
Viehweg vanishing theorem are fulfilled.
This clearly
demonstrates that investment in adaptation in the higher
order hierarchies diminishes chaotic behavior in the hier-
archies. This is because our complex manifold is a Stein
manifold (s is a Schwartz distribution). Furthermore, an
empirical process is already introduced as “Paddelbewe-
gung” in [1], inspired by Hermann Weyl’s work. Other
possible developments for this work include utilizing a
Riemann scheme and hypergeometric differential equa-
tions or Painlev´e VI equations for the hierarchical time-
developing model. Consideration of an array of model
types would plausibly allow exploration in relation to
Galois theory and ´etale cohomology to interpret the hier-
archical structures of natural systems, especially in bio-
logical contexts. This thus represents fruitful terrain for
future research.
Finally, adopting the Atiyah-Singer index theorem, a
twisted (fractal) property, Euler number of
∫
B e(TB) is
obviously equal to its topological Euler characteristic,
χ(B) = ∑(−1)ll. Hence, the analytical index of Euler
class (Poincar´e dual) should be the same. For evaluation
of agreement, the Chern class should be (−1)ll. On the
other hand, analytically, the Hirzebruch signature (char-
acteristic from species) of B is (−1)n ∫
B
∏n
i=1
pi
tanh pi ,
where
pi
tanh pi = ∑
k≥0
22kB2k
(2k)! p2k
i . Topologically, this is
equivalent to the L genus.
We are thus able to extend the methodology for the
“small s” metric to characterize dynamical system hier-
archy (adaptation and contributions) and interactions,
using only abundance data along time development.
ACKNOWLEDGMENTS
This research was primarily funded by the Center of
Innovation (COI) Program of Japan Science and Tech-
nology Agency. Additional funding was provided by Ky-
oto University. I extend my gratitude to all the reviewers
and colleagues who provided useful information and in-
sights which helped to materially improve this work.
13
[1] S.
Adachi,
Exploring
group
theory
and
topol-
ogy
for
analyzing
the
structure
of
biological
hi-
erarchies,
arXiv:1603.00959v8
[q-bio.PE]
(2019).
https://arxiv.org/abs/1603.00959.
[2] H. A. Bethe, Zur Theorie der Metalle, I. Eigenwerte und
Eigenfunktionen der linearen Atomkette, Z. Phys. 71,
205-226 (1931). https://doi.org/10.1007/BF01341708.
[3] S. Adachi, Eastern Japanese Dictyostelia species adapt
while populations exhibit neutrality, Evol. Biol. 42, 212-
222 (2015). https://doi.org/10.1007/s11692-015-9312-0.
[4] L. Medlin, H.J. Elwood, S. Stickel, and M.L. Sogin,
The characterization of enzymatically amplified eukary-
otic 16S-like rRNA-coding regions, Gene 71, 491-499
(1988) . https://doi.org/10.1016/0378-1119(88)90066-2.
[5] S. Adachi,
Rigid geometry solves “curse of dimen-
sionality” effects in clustering methods:
An applica-
tion to omics data, PLOS ONE 12, e0179180 (2017).
https://doi.org/10.1371/journal.pone.0179180.
[6] M. Rapoport, and Th. Zink, Period Spaces for p-divisible
Groups(Princeton University Press, Princeton, 1996).
[7] Y. Varshavsky, p-adic uniformization of unitary Shimura
varieties, Publ. Math. IH´ES 87, 57-119 (1998).
[8] G.E. Andrews, R.J. Baxter, and P.J. Forrester, Eight-
vertex SOS model and generalized Rogers-Ramanujan-
type identities,
J. Stat. Phys. 35,
193-266 (1984).
https://doi.org/10.1007/BF01014383.
[9] R.J. Baxter,
Eight-vertex model in lattice statistics
and one-dimensional anisotropic Heisenberg model, Ann.
Phys.
76,1-71
(1973).
https://doi.org/10.1016/0003-
4916(73)90440-5.
[10] J.H. Hubbard, Teichm¨uller Theory and Applications to
Geometry, Topology and Dynamics, Vol. 1, Teichm¨uller
Theory(Matrix Editions, Ithaca, 2006).
[11] J. H. Silverman, The Arithmetic of Elliptic Curves, Grad-
uate Texts in Mathematics, Vol. 106(Springer-Verlag,
New York, 1986).
[12] M.
H.
Stone,
Linear
transformations
in
Hilbert
space:
III. Operational methods and group theory,
Proc.
Natl.
Acad.
Sci.
USA
16,
172-175
(1930).
https://doi.org/10.1073/pnas.16.2.172.
[13] M.
H.
Stone,
On
one-parameter
unitary
groups
in Hilbert space,
Ann.
Math.
33,
643-648
(1932).
https://doi.org/10.2307/1968538.
[14] J. von Neumann, Die Eindeutigkeit der Schr¨odingerschen
Operatoren,
Math.
Ann.
104,
570-578
(1931).
https://doi.org/10.1007/BF01457956.
[15] J. von Neumann, Ueber Einen Satz von Herrn M.
H. Stone,
Ann. Math. Ser. 2 33,
567-573 (1932).
https://doi.org/10.2307/1968535.
[16] J.G. Cramer, The transactional interpretation of quan-
tum mechanics, Rev. Mod. Phys. 58, 647-687 (1986).
https://doi.org/10.1103/RevModPhys.58.647.
[17] J. Guckenheimer, and P. Holmes, Nonlinear Oscilla-
tions, Dynamical Systems, and Bifurcations of Vec-
tor Fields(Springer-Verlag, New York-Berlin-Heidelberg-
Tokyo, 1983).
[18] L. Schwartz, Th´eorie des Distributions(Hermann & Cie,
Paris, 1966).
[19] M.
Eichler,
Quatern¨are
quadratische
Formen
und
die Riemannsche Vermutung f¨ur die Kongruenzzeta-
funktion,
Arch.
Math.
(Basel)
5,
355-366
(1954).
https://doi.org/10.1007/BF01898377.
[20] Y. I. Manin, Frobenius Manifolds, Quantum Cohomol-
ogy, and Moduli Spaces(AMS Colloquium Publications,
Providence, 1999).
[21] H.
Kawakami,
A.
Nakamura,
and
H.
Sakai,
De-
generation
scheme
of
4-dimensional
Painlev´e-type
equations,
arXiv:1209.3836v3
[math.CA]
(2016).
https://arxiv.org/abs/1209.3836.
[22] N.
Iorgov,
O.
Lisovyy,
and
J.
Teschner,
Isomon-
odromic
tau-Functions
from
Liouville
conformal
blocks,
Commun.
Math.
Phys.
336,
671
(2015).
https://doi.org/10.1007/s00220-014-2245-0.
[23] S. Kakei, and T. Kikuchi, The sixth Painlev´e equa-
tion as similarity reduction of ˆgl3 generalized Drinfel’d-
Sokolov hierarchy, Lett. Math. Phys. 79, 221-234 (2007).
https://doi.org/10.1007/s11005-007-0144-4.
[24] O. Gamayun, N. Iorgov, and O. Lisovyy, Conformal
field theory of Painlev´e VI, J. High Energy Phys. 10,
38 (2012). https://doi.org/10.1007/JHEP10(2012)038.
[25] M.A.
Bershtein,
and
A.
I.
Shchechkin,
Bilin-
ear
equations
on
Painlev´e
τ
functions
from
CFT,
Commun.
Math.
Phys.
339,
1021-1061
(2015).
https://doi.org/10.1007/s00220-015-2427-4.
[26] P. Gavrylenko, and O. Lisovyy, Fredholm determinant
and Nekrasov sum representations of isomonodromic tau
functions, Commun. Math. Phys. 363, 1-58 (2016).
https://doi.org/10.1007/s00220-018-3224-7.
[27] O. Lisovyy, and Y. Tykhyy, Algebraic solutions of sixth
Painlev´e equation, J. Geometry Phys. 85, 124-163 (2014).
https://doi.org/10.1016/j.geomphys.2014.05.010.
[28] E.
Meinrenken,
Clifford
Algebras
and
Lie
Theory(Springer-Verlag, Berlin-Heidelberg, 2013).
[29] M. Sato, Theory of hyperfunctions, I., J. Fac. Sci. Univ.
Tokyo, Sect., I. 8, 139-193 (1959).
[30] M. Sato, Theory of hyperfunctions, II., J. Fac. Sci. Univ.
Tokyo, Sect., I. 8, 387-437 (1960).
[31] M. Morimoto, Sur la d´ecomposition du faisceau des ger-
mes de singularit´es, d’hyperfonctions, J. Fac. Sci. Univ.
Tokyo Sect, IA 17, 215-239 (1970).
[32] M. Sato, Hyperfunctions and partial differential equa-
tions, in: Proc. Intern. Conf. on Functional Analysis and
Related Topics (Todai Shuppankai, Tokyo, 1969), pp. 91-
94.
[33] R. Bott, R., and L.W. Tu, Differential Forms in Algebraic
Topology(Springer-Verlag, New York, 1982).
[34] P. Deligne, Cohomologie ´Etale, Lecture Notes in Mathe-
matics 569(Springer-Verlag, Berlin, 1977).
[35] M. Rapoport, and Th. Zink, ¨Uber die lokale Zetafunk-
tion von Shimuravariet¨aten. Monodromiefiltration und
verschwindende Zyklen in ungleicher Charakteristik, In-
vent. Math. 68, 21-101 (1982).
[36] P.
Deligne,
La
Conjecture
de
Weil.
I,
Publ.
Math.
IH´ES
43,
273-307
(1974).
https://doi.org/10.1007/BF02684373.
[37] P.
Deligne,
La
Conjecture
de
Weil.
II,
Publ.
Math.
IH´ES
52,
137-252
(1980).
https://doi.org/10.1007/BF02684780.
[38] M. Jimbo,
and H. Sakai,
A q-analog of the sixth
Painlev´e equation, Lett. Math. Phys. 38, 145-154 (1996).
https://doi.org/10.1007/BF00398316.
[39] J.
Borger,
Lambda-rings
and
the
field
with
one
element,
arXiv:0906.3146v1
[math.NT]
(2009).
https://arxiv.org/abs/0906.3146.
14
[40] L.V.
Ahlfors,
Complex
Analysis(McGraw-Hill
Book
Company, New York, 1979).
See Supplemental Material at [URL will be inserted by
publisher] for LC/MS data. All files related to a pub-
lished paper are stored as a single deposit and assigned
a Supplemental Material URL. This URL appears in the
article’s reference list.
| 2019 | Induction of hierarchy and time through one-dimensional probability space with certain topologies | 10.1101/780882 | null | creative-commons |
1
Coevolutionary Analysis and Perturbation-Based
Network Modeling of the SARS-CoV-2 Spike Protein
Complexes with Antibodies: Binding-Induced Control
of Dynamics, Allosteric Interactions and Signaling
Gennady M. Verkhivker,1,2 ‡ Luisa Di Paola3
1Keck Center for Science and Engineering, Schmid College of Science and Technology,
Chapman University, One University Drive, Orange, CA 92866, USA
2 Depatment of Biomedical and Pharmaceutical Sciences, Chapman University School of
Pharmacy, Irvine, CA 92618, USA
3Unit of Chemical-Physics Fundamentals in Chemical Engineering, Department of Engineering,
Università Campus Bio-Medico di Roma, via Álvaro del Portillo 21, 00128 Rome, Italy
‡corresponding author
E-mail: verkhivk@chapman.edu
2
Abstract
The structural and biochemical studies of the SARS-CoV-2 spike glycoproteins and complexes
with highly potent antibodies have revealed multiple conformation-dependent epitopes
highlighting the link between conformational plasticity of spike proteins and capacity for eliciting
specific binding and broad neutralization responses. In this study, we used coevolutionary
analysis, molecular simulations, and perturbation-based hierarchical network modeling of the
SARS-CoV-2 S complexes with H014, S309, S2M11 and S2E12 antibodies targeting distinct
epitopes to explore molecular mechanisms underlying binding-induced modulation of dynamics,
stability and allosteric signaling in the spike protein trimers. The results of this study revealed
key regulatory centers that can govern allosteric interactions and communications in the SARS-
CoV-2 spike proteins. Through coevolutionary analysis of the SARS-CoV-2 spike proteins, we
identified highly coevolving hotspots and functional clusters forming coevolutionary networks.
The results revealed significant coevolutionary couplings between functional regions separated
by the medium-range distances which may help to facilitate a functional cross-talk between
distant allosteric regions in the SARS-CoV-2 spike complexes with antibodies. We also
discovered a potential mechanism by which antibody-specific targeting of coevolutionary
centers can allow for efficient modulation of allosteric interactions and signal propagation
between remote functional regions. Using a hierarchical network modeling and perturbation-
response scanning analysis, we demonstrated that binding of antibodies could leverage direct
contacts with coevolutionary hotspots to allosterically restore and enhance couplings between
spatially separated functional regions, thereby protecting the spike apparatus from membrane
fusion. The results of this study also suggested that antibody binding can induce a switch from a
moderately cooperative population-shift mechanism, governing structural changes of the ligand-
3
free SARS-CoV-2 spike protein, to antibody-induced highly cooperative mechanism that can
better withstand mutations in the functional regions without significant deleterious consequences
for protein function. This study provides a novel insight into allosteric regulatory mechanisms of
SARS-CoV-2 S proteins, showing that antibodies can modulate allosteric interactions and
signaling of spike proteins, providing a plausible strategy for therapeutic intervention by targeting
specific hotspots of allosteric interactions in the SARS-CoV-2 proteins.
4
Introduction
The coronavirus disease 2019 (COVID-19) pandemic associated with the severe acute respiratory
syndrome (SARS)1-5 has been at the focal point of biomedical research. SARS-CoV-2 infection
is transmitted when the viral spike (S) glycoprotein binds to the host cell receptor leading to the
entry of S protein into host cells and membrane fusion.6-8 The full-length SARS-CoV-2 S protein
consists of two main domains, amino (N)-terminal S1 subunit and carboxyl (C)-terminal S2
subunit. The subunit S1 is involved in the interactions with the host receptor and includes an N-
terminal domain (NTD), the receptor-binding domain (RBD), and two structurally conserved
subdomains (SD1 and SD2). Structural and biochemical studies have shown that the mechanism
of virus infection may involve spontaneous conformational transformations of the SARS-CoV-
2 S protein between a spectrum of closed and receptor-accessible open forms, where RBD
continuously switches between “down” and “up” positions where the latter can promote
binding with the host receptor ACE2.9-11 The crystal structures of the S-RBD in the complexes
with human ACE2 enzyme revealed structurally conserved binding mode shared by the SARS-
CoV and SARS-CoV-2 proteins in which an extensive interaction network is formed by the
receptor binding motif (RBM) of the RBD region.12-16 These studies established that binding of
the SARS-CoV-RBD to the ACE2 receptor can be a critical initial step for virus entry into target
cells. The rapidly growing body of cryo-EM structures of the SARS-CoV-2 S proteins detailed
distinct conformational arrangements of S protein trimers in the prefusion form that are
manifested by a dynamic equilibrium between the closed (“RBD-down”) and the receptor-
accessible open (“RBD-up”) form required for the S protein fusion to the viral membrane.17-26
The cryo-EM characterization of the SARS-CoV-2 S trimer demonstrated that S protein may
populate a spectrum of closed states by fluctuating between structurally rigid locked-closed form
5
and more dynamic, closed states preceding a transition to the fully open S conformation.26
Structural and biophysical studies employed protein engineering to generate prefusion-stabilized
SARS-CoV-2 S variants by introducing disulfide bonds and proline mutations to modulate
stability of the S2 subunit and the inter-subunit boundaries, and consequently prevent refolding
changes that accompany acquisition of the postfusion state.27 By combining targeted
mutagenesis and cryo-EM structure determination, recent biophysical investigations demonstrated
that modifications in the contact regions between the RBD and S2 domains via S383C/D985C
double mutation can lead to the thermodynamic prevalence of the closed-down conformation,
while the quadruple mutant (A570L/T572I/F855Y/N856I) perturbing the inter-protomer contacts
can shift the equilibrium towards the open form with the enhanced binding propensities for the
ACE2 host receptor.28 Protein engineering and cryo-EM studies of a prefusion-stabilized SARS-
CoV-2 S ectodomain trimer using the inter-protomer disulfide bonds (S383C/D985C,
G413C/P987C, T385C/T415C) between RBD and S2 regions can lock the trimer in the closed
state and enhance the SARS-CoV-2 S resistance to proteolysis.29 Targeted design of thermostable
SARS-CoV-2 spike trimers further specified how disulfide-bonded S-protein trimer variants
imposing stabilization in the strategically located inter-protomer positions (S383-D385) and
(G413-V987) can promote dramatic thermodynamic shifts towards the prefusion closed states
with only ~20% of the population corresponding to the open state.30 Recent biochemical studies
of the SARS-CoV-2 S mutants with the enhanced infectivity profile39-41 discovered that a highly
active D614G mutation can exert its dramatic functional effect on virus infectivity by radically
shifting the population of the SARS-CoV-2 S trimer towards open states.31 The cryo-EM and
sophisticated tomography tools determined the high-resolution structure and distribution of S
trimers in situ on the virion surface.32 These studies confirmed a general mechanism of population
6
shifts between different functional states of the SARS-CoV-2 S trimers, suggesting that RBD
epitopes can become stochastically exposed to the interactions with the host receptor ACE2.
Biophysical analysis of SARS-CoV-2 S trimer on virus particles revealed four distinct
conformational states for the S protein and a sequence of conformational transitions through an
obligatory intermediate in which all three RBD domains in the closed conformations are oriented
towards the viral particle membrane.33 Cryo-EM structural studies also mapped a mechanism of
conformational events associated with ACE2 binding, showing that the compact closed form of
the SARS-CoV-2 S protein becomes weakened after furin cleavage between the S1 and S2
domains, leading to the increased population of partially open states and followed by ACE2
recognition that can accelerate transformation to a fully open and ACE2-bound form priming the
protein for fusion activation.34
The early biochemical studies of SARS S proteins with antibodies (Abs) suggested that RBD
regions of S proteins contain multiple conformation-dependent epitopes capable of inducing
potent neutralizing Ab responses, thus revealing the link between conformational heterogeneity
of S proteins and capacity for eliciting binding with highly potent neutralizing Abs.35
Subsequently, it was shown that major neutralizing epitopes of SARS-CoV may have been
preserved during cross-species transmission, and that RBD-targeted Abs have a potential for
broad protection against both human and animal SARS-CoV variants.36 The SARS-CoV-2 S
protein–targeting monoclonal antibodies (mAbs) with potent neutralizing activity are of
paramount importance and are actively pursued as therapeutic interventions for COVID-19
virus.37-40 The rapidly growing structural studies of SARS-CoV and SARS-CoV-2–neutralizing
Abs targeting the RBD have suggested potential mechanisms underlying inhibition of the
association between the S protein and ACE2 host receptor. The early structure of SARS-CoV-
7
RBD complex with a neutralizing Ab 80R showed that the epitope on the S1 RBD overlapped
closely with the ACE2-binding site, suggesting that a direct interference mechanism may be
responsible for the neutralizing activity.41 However, several SARS-CoV–specific neutralizing
Abs such as m396, 80R, and F26G19 that block the RBM motif in the open S conformation
did not exhibit a strong neutralizing activity against SARS-CoV-2 protein. The crystal structure of
a neutralizing Ab CR3022 in the complex with the SARS-CoV-2 S-RBD revealed binding to a
highly conserved epitope that is located away from the ACE2-binding site but could only be
accessed when two RBDs adopt the “up” conformation.42 Subsequent structural and surface
plasmon resonance studies confirmed that CR3022 binds the RBD of SARS-CoV-2 displaying
strong neutralization by allosterically perturbing the interactions between the RBD regions and
ACE2 receptor.43 The proposed neutralization mechanism of SARS-CoV-2 through
destabilization of the prefusion S conformation can provide a resistance mechanism to virus
escape which can be contrasted with Abs directly competing for the ACE2-binding site and
often susceptible to immune evasion. Potent neutralizing Abs from COVID-19 patients examined
through electron microscopy studies confirmed that the SARS-CoV-2 S protein features multiple
distinct antigenic sites, including RBD-based and non-RBD epitopes.44 These studies also
suggested that some Abs may function by allosterically interfering with the host receptor binding
and causing conformational changes in the S protein that can obstruct other epitopes and block
virus infection without directly interfering with ACE2 recognition. Cryo–EM characterization of
the SARS-CoV-2 S trimer in complex with the H014 Fab fragment revealed a new conformational
epitope that is accessible only when the RBD is in the up conformation.45 Biochemical and
virological studies demonstrated that H014 prevents attachment of SARS-CoV-2 to the host cell
receptors and can exhibit broad cross-neutralization activities by leveraging conserved nature of
8
the RBD epitope and a partial overlap with ACE2-binding region. The recently reported mAb
S309 potently neutralizes both SARS-CoV-2 and SARS-CoV through binding to a conserved
RBD epitope which is distinct from the RBM region and accessible in both open and closed
states, so that there is no completion between S309 and ACE2 for binding to the SARS-CoV-2
S protein.46 Two ultra-potent Abs S2M11 and S2E12 targeting the overlapping RBD epitopes
were recently reported, revealing Ab-specific modulation of protein responses and adaptation of
different functional states for the S trimer.47 Cryo-EM structures showed that S2M11 can
recognize and stabilize S protein in the closed conformation by binding to a quaternary epitope
spanning two RBDs of the adjacent protomers in the S trimer, while S2E12 binds to a tertiary
epitope contained within one S protomer and shifts the conformational equilibrium towards a fully
open S trimer conformation.47 The mAbs isolated from 10 convalescent COVID-19 patients
showed neutralizing activities against authentic SARS-CoV-2, with the mAb 4A8 displaying
high potency by binding to the NTD of the S protein conformation with one RBD in “up”
conformation and the other two RBDs in “down” conformation.48 Interestingly, none of the
isolated mAbs recognize the RBD and inhibit binding of SARS-CoV-2 S protein to ACE2,
suggesting that allosterically regulated mechanisms may underlie the functional effects and
experimentally observed efficient cross-neutralization.48 Moreover, it was proposed that
combining NTD-targeting 4A8 with RBD-targeting Abs may help in the design of “cocktail”
therapeutics to combat the escaping mutations of the virus.
The most recent investigation reported discovery of an ultra-potent synthetic nanobody Nb6 that
neutralizes SARS-CoV-2 by stabilizing the fully inactive down S conformation preventing
binding with ACE2 receptor.49 Affinity maturation and structure-guided design produced a
trivalent nanobody, mNb6-tri that simultaneously binds to all three RBDs, yielding the
9
remarkably high affinity for S protein and completely blocking the S-ACE2 interactions by
occupying the binding site and locking spike protein in the inactive, receptor-inaccessible state. In
general, Abs tend to bind to the most easily accessible regions of the virus, where viruses can
tolerate mutations and thereby escape immune challenge. The emerging body of recent studies
suggested that properly designed cocktails of Abs can provide a broad and efficient cross-
neutralization effects through synergistic targeting of conserved and more variable SARS-CoV-
2 RBD epitopes, thereby offering a robust strategy to combat virus resistance.45-49
Computational modeling and molecular dynamics (MD) simulations have been instrumental in
predicting conformational and energetic mechanisms of SARS-CoV-2 functions.50-55
Microsecond, all-atom MD simulations of the full-length SARS-CoV-2 S glycoprotein embedded
in the viral membrane, with a complete glycosylation profile were recently reported, providing the
unprecedented level of details about open and closed structures.51 MD simulations of the SARS-
CoV-2 spike glycoprotein identified differences in flexibility of functional regions that may be
important for modulating the equilibrium changes and binding to ACE2 host receptor.52
Computational studies examined SARS-CoV-2 S trimer interactions with ACE2 enzyme using
the recent crystal structures53-62 providing insights into the key determinants of the binding
affinity and selectivity. A comprehensive study employed MD simulations to reveal a balance of
hydrophobic interactions and elaborate hydrogen-bonding network in the SARS-CoV-2-RBD
interface.59 Molecular mechanisms of the SARS-CoV-2 binding with ACE2 enzyme were
analyzed in our recent study using coevolution and conformational dynamics.62 Using protein
contact networks and perturbation response scanning based on elastic network models, we
recently discovered existence of allosteric sites on the SARS-CoV- 2 spike protein.63 By using
molecular simulations and network modeling we recently presented the first evidence that the
10
SARS-CoV-2 spike protein can function as an allosteric regulatory engine that fluctuates between
dynamically distinct functional states.64
In this study, we used a battery of computational approaches to explore and simulate molecular
mechanisms underlying responses of the SARS-CoV-2 S proteins to binding of a panel of Abs
(H014, S309, S2M11 and S2E12) that target distinct epitopes in the RBD regions. Using
coevolutionary analysis, molecular simulations, and perturbation-based hierarchical network
modeling of the SARS-CoV-2 S complexes with these Abs, we examined binding-induced
modulation of dynamics, stability and allosteric interactions in the S protein trimers. The results
of this study revealed structural topography of coevolutionary couplings and network
connectivity that may determine mechanisms of allosteric signaling in the SARS-CoV-2 S
proteins. We show that specific Ab targeting of conserved centers and coevolutionary hotspots
in the S protein that are distinct from RBM region can allow not only for productive binding,
but also for efficient Ab-induced modulation of long-range interactions between the S1 and S2
units of the SARS-CoV-2 S protein. Using perturbation-based network modeling, we find that
targeted binding of the Abs could leverage direct contacts with coevolutionary hotspots to
effectively restore allosteric potential of the S1 regions in the open states, thereby strengthening
the allosteric interaction network and protecting the S protein machinery from dissociation of S1
subunit required for membrane fusion. The results of this study provide a novel insight into
allosteric regulatory mechanisms of SARS-CoV-2 S proteins showing that the examined Abs can
uniquely modulate signal communication providing a plausible strategy for therapeutic
intervention by targeting specific hotspots of allosteric interactions in the SARS-CoV-2 proteins.
11
Materials and Methods
Sequence Conservation and Coevolutionary Analyses
Multiple sequence alignment (MSA) was obtained using the MAFFT approach65 and homologues
were obtained from UNIREF90.66,67 We employed Kullback-Leibler (KL) sequence
conservation score KLConsScore using MSA profiles generated by hidden Markov models in
Pfam database for the SARS-CoV S glycoproteins.68,69 Three Pfam domains were utilized
corresponding to S1, the NTD (bCoV_S1_N, Betacoronavirus-like spike glycoprotein S1, N-
terminal, Pfam:PF16451, Uniprot SPIKE_CVHSA, pdb id 6CS0, residues 33-324), the RBD
(bCoV_S1_RBD, Betacoronavirus spike glycoprotein S1, receptor binding, Pfam:PF09408,
Uniprot SPIKE_CVHSA, pdb id 6CS0, residues 335-512) and the new C-terminal domain, CTD
(CoV_S1_C Coronavirus spike glycoprotein S1, C-terminal. Pfam:PF19209, Uniprot
SPIKE_CVHSA, pdb id 6CS0, residues 522-580). S2 is described in the family Pfam:PF01601
(Uniprot SPIKE_CVHSA, pdb id 6CS0, residues 622-1120) which contains an additional S2′
cleavage site, a fusion peptide, internal fusion peptide, heptad repeat 1/2 domains, and the
transmembrane domain. The following Uniprot entries were used for sequence analysis : P59594:
SPIKE_SARS (previously SPIKE_CVHSA) (pdb id 6CS0) and P0DTC2: SPIKE_SARS2 (pdb id
6VXX, 6VYB).
The KL conservation is calculated according to the following formula:
1
( )
ln
( )
N
i
i
P i
KLConsScore
Q i
=
= ∑
(1)
Here,
( )
P i is the frequency of amino acid i in that position and
( )
Q i is the background frequency
of the amino acid in nature calculated using an amino acids background frequency distribution
obtained from the UniProt database.70 To evaluate coevolutionary couplings in the SARS-CoV-
12
2 S glycoproteins, we used MISTIC approach71-73 in which computation of residue covariation
was done with three different direct coupling analysis (DCA) methods: mean field DCA
(mfDCA),74-76 pseudo-likelihood maximization DCA (plmDCA)77,78 and multivariate Gaussian
modeling DCA (gaussianDCA).79,80 For each residue, we computed cumulative covariation score
(CScore) parameter, that evaluates to what degree a given position participates in the
coevolutionary network. CScore is a derived score per position that characterizes the extent of
coevolutionary couplings shared by a given residue. This score is calculated as the sum of
covariation scores above a certain threshold (typically top 5% of the covariation scores) for every
position pair where the particular position appears.
Coarse-Grained Molecular Simulations
Coarse-grained (CG) models are computationally effective approaches for simulations of large
systems over long timescales. In this study, CG-CABS model81-85 was used for simulations of
the cryo-EM structures of the SARS-CoV-2 S complexes with H014, S309, S2M11, and S2E12
Abs. In this model, the amino acid residues are represented by Cα, Cβ, the center of mass of
side chains and another pseudoatom placed in the center of the Cα-Cα pseudo-bond.81-83 We
employed CABS-flex approach that efficiently combines a high-resolution coarse-grained
model and efficient search protocol capable of accurately reproducing all-atom MD simulation
trajectories and dynamic profiles of large biomolecules on a long time scale.81-85 The sampling
scheme of the CABS model used in our study is based on Monte Carlo replica-exchange
dynamics and involves a sequence of local moves of individual amino acids in the protein
structure as well as moves of small fragments.81-83 CABS-flex standalone package dynamics
implemented as a Python 2.7 object-oriented package was used for fast simulations of protein
structures.85 A total of 1,000 independent CG-CABS simulations were performed for each of the
13
studied systems. In each simulation, the total number of cycles was set to 10,000 and the number
of cycles between trajectory frames was 100. The cryo-EM structures of the SARS-CoV-2 S
trimer complexes with a panel of Abs including H014, S309, S2M11, and S2E12 were used in
CG-CABS simulations (Figures 1,2). These structures included the partially open and fully open
forms of the SARS-CoV-2 S trimer in the complex with H014 (Figure 1A,B), the partially
closed and fully closed S trimer forms bound with S309 (Figure 1C,D), the fully closed S trimer
form complexed with S2M11 (Figure 1E), and the fully open S trimer form in the complex with
S2E12 (Figure 1F). All structures were obtained from the Protein Data Bank.86,87 During
structure preparation stage, protein residues in the crystal structures were inspected for missing
residues and protons. Hydrogen atoms and missing residues were initially added and assigned
according to the WHATIF program web interface.88,89 The structures were further pre-processed
through the Protein Preparation Wizard (Schrödinger, LLC, New York, NY) and included the
check of bond order, assignment and adjustment of ionization states, formation of disulphide
bonds, removal of crystallographic water molecules and co-factors, capping of the termini,
assignment of partial charges, and addition of possible missing atoms and side chains that were
not assigned in the initial processing with the WHATIF program. The missing loops in the cryo-
EM structures were also reconstructed using template-based loop prediction approaches
ModLoop90 and ArchPRED91 The conformational ensembles were also subjected to all-atom
reconstruction using PULCHRA method92 and CG2AA tool93 to produce atomistic models of
simulation trajectories. The protein structures were then optimized using atomic-level energy
minimization with a composite physics and knowledge-based force fields as implemented in the
3Drefine method.94
14
Figure 1. The cryo-EM structures of the SARS-CoV-2 S protein trimer complexes with a panel
of Abs used in this study. (A) The cryo-EM structure of the SARS-CoV-2 S protein trimer with
two RBDs in the open state complexed with two H014 Fab (pdb id 7CAI).45 (B) The cryo-EM
structure of the SARS-CoV-2 S protein trimer with three RBD in the open state complexed
with three H014 Fab (pdb id 7CAK).45 (C) The cryo-EM structure of the SARS-CoV-2 S protein
trimer with two RBDs in the closed form and one RBD in the open state bound with the two S309
neutralizing Fab fragments (pdb id 6WPT).46 (D) The cryo-EM structure of the SARS-CoV-2 S
protein trimer with all three RBDs in the closed form bound with the three S309 neutralizing Fab
fragments (pdb id 6WPS).46 (E) The cryo-EM structure of the SARS-CoV-2 S protein trimer
15
with all three RBDs in the closed-down form bound with the three S2M11 neutralizing Fab
fragments (pdb id 7K43).47 (F) The cryo-EM structure of the SARS-CoV-2 S protein trimer
with all three RBDs in the open-up form bound with the three S2ME12 neutralizing Fab fragments
(pdb id 7K4N).47 The SARS-CoV-2 S proteins are shown in surface representation, with protomer
A in green, protomer B in cyan, and protomer C in magenta. The Ab structures are shown in
ribbons and colored in maroon. All structures are annotated and open/closed (up/down)
conformations of S protomers are indicated.
16
Figure 2. The binding epitopes of the SARS-CoV-2 S protein trimer complexes with a panel
of Abs used in this study. The top view highlighting the binding epitopes is shown for the cryo-
EM structure of the SARS-CoV-2 S protein trimer with H014 (A,B), S309 (C,D), S2M11 (E), an
S2E12 (F). Note, S2M11 recognizes a quaternary epitope comprising distinct regions of two
neighboring RBDs within an S trimer (E). The SARS-CoV-2 S proteins are shown in surface
representation, with protomer A in green, protomer B in cyan, and protomer C in magenta. The
Ab structures are shown in ribbons and colored in maroon. All structures are annotated and
open/closed (up/down) conformations of S protomers are indicated.
Protein Stability and Mutational Scanning Analysis
To compute protein stability changes in the SARS-CoV-2 trimer mutants, we conducted a
systematic alanine scanning of protein residues in the SARS-CoV-2 trimer mutants. BeAtMuSiC
approach was employed that is based on statistical potentials describing the pairwise inter-residue
distances, backbone torsion angles and solvent accessibilities, and considers the effect of the
mutation on the strength of the interactions at the interface and on the overall stability of the
complex.95-97 The binding free energy of protein-protein complex can be expressed as the
difference in the folding free energy of the complex and folding free energies of the two protein
binding partners:
com
A
B
bind
G
G
G
G
∆
=
−
−
(2)
The change of the binding energy due to a mutation was calculated then as the following:
mut
wt
bind
bind
bind
G
G
G
∆∆
= ∆
− ∆
(3)
17
We leveraged rapid calculations based on statistical potentials to compute the ensemble-averaged
binding free energy changes using equilibrium samples from MD trajectories. The binding free
energy changes were computed by averaging the results over 1,000 equilibrium samples for each
of the studied systems.
Protein Contact Networks and Network Clustering
The protein contact network is a network whose nodes are the protein residues and links are active
contacts between residues in the protein structure. The protein contact network is an undirected,
unweighted graph; it is built on the basis of the distance matrix d, whose generic element dij records
the Euclidean distance between the ith and the jth residue (measured between the corresponding α
carbons). A detailed description of network construction and significance of network descriptors
is presented in our previous studies.98-100 The active network links are defined using a range of
contacts between 4 Å and 8 Å. The description of the network is given by the following adjacency
matrix :
������������������������������������ = �1
������������������������ 4 Å < ������������������������������������ < 8 Å
0
������������������������ℎ������������������������������������������������������������������������
(4)
where ������������������������������������ is the distance between the residues. The node degree describes the number of links
each residue has with other residues, defined as:
������������������������ = ∑ ������������������������������������
������������
(5)
We previously demonstrated that spectral network clustering targets functional modules in
proteins.98 The network clustering is based on the spectral decomposition of the network Laplacian
������������ defined as:
18
������������ = ������������ − ������������
(6)
where ������������ is the degree matrix, a diagonal matrix whose diagonal is the degree vector, and ������������ is the
adjacency matrix. We used the eigenvector corresponding to the second minor eigenvalue ������������������������ of
the Laplacian (Fiedler’s vector) to assign nodes (residues) to different clusters. We introduced a
novel feature in the hierarchical binary algorithm to compute any number of clusters ������������������������������������������������������������ (no
more only powers of two): we parted the whole range of values of ������������������������ into ������������������������������������������������������������ parts, of length
������������������������������������������������������������; so residues are assigned to the first cluster if the corresponding component falls between
min (������������������������) and min(������������������������) + ������������������������������������������������������������ the generic i-th cluster, thus, is that made of residues corresponding
to ������������������������ components comprised in the range [min(������������������������) + (������������ − 1) ⋅ ������������������������������������������������������������, min(������������������������) + ������������ ⋅ ������������������������������������������������������������].
Once the network is divided into a given number of clusters (powers of two), we define the
participation coefficient, defined as:
������������������������ = 1 − �
������������������������
�������������������������������������
2
(7)
where ������������������������ is the overall node degree, while ������������������������������������ is the node degree including only links with nodes
(residues) that belong to their own cluster. The participation coefficient ������������ describes the propensity
of residue nodes to participate into inter-cluster communication. We designate as highly active
communication residues the nodes with P>0.75, based on our previous studies showing that such
residues typically correspond to important regulatory centers of signal transmission between
protein domains.98 The proposed methodology of network clustering was implemented as
Cytoscape plugin.101
In the framework of hierarchical network modeling approach, we also employed a graph-based
representation of protein structures102-104 with residues as network nodes and the inter-residue
edges as residue interactions to construct the residue interaction networks using dynamic
correlations104 and coevolutionary residue couplings 105 as detailed in our previous studies.105-107
19
The ensemble of shortest paths is determined from matrix of communication distances by the
Floyd-Warshall algorithm.108 Network graph calculations were performed using the python
package NetworkX.109 Using the constructed protein structure networks, we computed the
residue-based betweenness parameter. The short path betweenness centrality of residue i is
defined to be the sum of the fraction of shortest paths between all pairs of residues that pass through
residuei :
( )
( )
N
jk
b
i
j k
jk
g
i
C n
g
<
=∑
(8)
where
jk
g
denotes the number of shortest geodesics paths connecting j and k , and
( )
jk
g
i is the
number of shortest paths between residues j and k passing through the node
in .
Perturbation Response Scanning
Perturbation Response Scanning (PRS) approach110,111 was used to estimate residue response to
external forces applied systematically to each residue in the protein system. This approach has
successfully identified hotspot residues driving allosteric mechanisms in single protein domains
and large multi-domain assemblies.112-117 The implementation of this approach follows the
protocol originally proposed by Bahar and colleagues112,113 and was described in details in our
previous studies.64 In brief, through monitoring the response to forces on the protein residues,
the PRS approach can quantify allosteric couplings and determine the protein response in
functional movements. In this approach, it 3N × 3N Hessian matrix ������������ whose elements represent
second derivatives of the potential at the local minimum connect the perturbation forces to the
residue displacements. The 3N-dimensional vector ������������������������ of node displacements in response to 3N-
20
dimensional perturbation force follows Hooke’s law ������������ = ������������ ∗ ������������������������. A perturbation force is applied
to one residue at a time, and the response of the protein system is measured by the displacement
vector ∆������������(������������) = ������������−������������������������(������������) that is then translated into N×N PRS matrix. The second derivatives
matrix ������������ is obtained from simulation trajectories for each protein structure, with residues
represented by ������������������������ atoms and the deviation of each residue from an average structure was calculated
by ∆������������������������(������������) = ������������������������(������������) − 〈������������������������(������������)〉, and corresponding covariance matrix C was then calculated
by ∆������������∆������������������������. We sequentially perturbed each residue in the SARS-CoV-2 spike structures by
applying a total of 250 random forces to each residue to mimic a sphere of randomly selected
directions.64 The displacement changes, ∆������������������������ is a 3N-dimensional vector describing the linear
response of the protein and deformation of all the residues. Using the residue displacements upon
multiple external force perturbations, we compute the magnitude of the response of residue k as
2
)
(i
k
ΔR
averaged over multiple perturbation forces F(i), yielding the ikth element of the N×N
PRS matrix.112,113 The average effect of the perturbed effector site ������������ on all other residues is
computed by averaging over all sensors (receivers) residues ������������ and can be expressed
as〈(∆������������������������)2〉������������������������������������������������������������������������������������������������. The effector profile determines the global influence of a given residue node
on the perturbations in other protein residues and can be used as proxy for detecting allosteric
regulatory hotspots in the interaction networks. In turn, the j th column of the PRS matrix
describes the sensitivity profile of sensor residue j in response to perturbations of all residues and
its average is denoted as 〈(∆������������������������)2〉������������������������������������������������������������������������. The sensor profile measures the ability of residue j to
serve as a receiver (or transmitter) of dynamic changes in the system.
21
Results and Discussion
Sequence Analysis Links Evolutionary Patterns in SARS-CoV S Proteins with Antibody
Binding Preferences
To determine the evolutionary patterns in the SARS-CoV S proteins and characterize the extent
of conservation and variability of the S1 and S2 subunits, we utilized KL sequence conservation
score.71-73 Consistent with previous studies118-120 we found that S1 RBD is less conserved than
domains in the S2 subunit (Figure 3A,B). The S2 subunit contains an N-terminal hydrophobic
fusion peptide (FP), fusion peptide proximal region (FPPR), heptad repeat 1 (HR1), central helix
region (CH), connector domain (CD), heptad repeat 2 (HR2), transmembrane domain (TD) and
cytoplasmic tail (CT). The S1 domains are situated above the S2 subunit, covering and protecting
the fusion apparatus. The results confirmed the higher conservation of the S2 subunit particularly
highlighting conservation of the HR1 (residues 910-985), CH (residues 986-1035), CD (residues
1068-1163), HR2 (residues 1163-1211) , and TD regions (residues 1211-1234) (Figure 3). The
S2 subunit regions are highly conserved in the SARS-CoV and the SARS-CoV-2 variants while
the S1 subunit was more diverse in the NTD and RBD regions.
Among most conserved residues in the S2 subunit are clusters of conserved cysteine residues
forming disulfide bridges that are crucial for stabilization of both pre-fusion and post-fusion
SARS-CoV-2 spike protein conformation. The S proteins can contain up to 40 cysteine residues,
36 of which are conserved in the S proteins of various SARS-coronaviruses.121 The conserved
cysteine cluster in the TD region 1220-CCMTSCCSC-1228 displayed high conservation scores,
with C1121, M1222, and C1225 featuring top 1% of conservation scores (Figure 3A). Indeed,
mutagenesis of the cysteine cluster I (1220-CCMTS-24), located immediately proximal to the TD
showed the 55% reduction in S-mediated cell fusion as compared to the wild-type S protein.122
22
The proximal cysteine cluster 1225-CCSC-1228 is similarly important as alanine mutations in this
cluster resulted in the 60% reduction of S-mediated cell fusion.122 At the same time, the nearest
cysteine-rich cluster 1230-CSCGSCCK-1237 featured only one highly conserved C1235.
According to the experimental data, mutations in this region caused only a moderate 15% reduction
in cell fusion122, indicating that functional role of these clusters may be closely linked with the
conservation level of cysteine residues.
As cysteine residues play critical roles in structural stabilization of proteins, we focused our
attention on these residues and their locations in the sequences. Interestingly, the most conserved
S2 positions included cysteine residues C720, C725, C731, C742, C822, C822, C833,
C1014,C1025, and C1064 (Figure 3A). A conserved region flanked by C822 and C833 is known
to be important for interactions with components of the SARS-CoV S trimer to control the
activation of membrane fusion.123 In addition, other conserved residues included Y819, I800,
L803, D830, L831, Y855, H1030, P1039, H1046 (Figure 3A,C). This is consistent with the
experimental mutagenesis study based on cell-cell fusion and pseudovirion infectivity assay
showing a critical role of the core-conserved residues C822, D830, L831, and C833 residues.123
Some of these residues are located C-terminal to the SARS-CoV S2 cleavage site at R797 forming
a highly conserved region 798-SFIEDLLFNKVTLADAGF-815 that plays an important role for
membrane fusion.124 Among highly conserved S protein regions are also six clusters of cysteine
residues in the S2 subunit forming disulfide bridges crucial for stabilization of both pre-fusion and
post-fusion SARS-CoV-2 spike protein conformations125-127 (Figures 3,4).
23
Figure 3. Sequence conservation analysis of the SARS-CoV-2 S glycoprotein. (Top panel) A
schematic representation of domain organization and residue range for the full-length SARS-
CoV-2 spike (S) protein. The subunits S1 and S2 include NTD RBD, C-terminal domain 1(CTD1),
C-terminal domain 2 (CTD2), S1/S2 cleavage site (S1/S2), S2’ cleavage site (S2’), fusion peptide
(FP), fusion peptide proximal region (FPPR), heptad repeat 1 (HR1), central helix region (CH),
connector domain (CD), heptad repeat 2 (HR2), transmembrane domain (TM), and cytoplasmic
tail (CT).
(Panel A) The KL conservation score for SARS-CoV-RBD S protein. High KL scores indicate
highly conserved sites and low scores correspond to more variable positions. Three Pfam domains
24
were utilized corresponding to S1, the NTD (bCoV_S1_N, Betacoronavirus-like spike
glycoprotein S1, N-terminal, Pfam:PF16451, Uniprot SPIKE_CVHSA, pdb id 6CS0, residues 33-
324), the RBD (bCoV_S1_RBD, Betacoronavirus spike glycoprotein S1, receptor binding,
Pfam:PF09408, Uniprot SPIKE_CVHSA, pdb id 6CS0, residues 335-512) and the new C-terminal
domain, CTD (CoV_S1_C Coronavirus spike glycoprotein S1, C-terminal. Pfam:PF19209,
Uniprot SPIKE_CVHSA, pdb id 6CS0, residues 522-580). S2 is described in the family
Pfam:PF01601 (Uniprot SPIKE_CVHSA, pdb id 6CS0, residues 622-1120). The KL scores for
the S1-NTD residues are shown in green bars, for the S1-RBD regions in the red bars, and for S2
residues in blue bars. The KL conservation scores for the epitope residues of all studied Abs are
shown in filled maroon-colored circles.
(Panel B) A close-up view of KL conservation scores for RBD regions of the SARS-CoV-2 S
protein (Pfam:PF09408, Uniprot P0DTC2: SPIKE_SARS2 (pdb id 6VXX, 6VYB residue
numbering) is shown in red bars. The KL scores are highlighted for the binding epitope residues
of H014 (filled maroon-colored circles), S309 (filled blue circles), and S2M11/S2E12 (filled green
circles).
(Panel C) The structural organization of the SARS-CoV-2 S protein major domains is shown for
a single protomer. The subunits S1 regions are annotated as follows : NTD (residues 14-306) in
light blue; RBD (residues 331-528) in yellow; CTD1 (residues 528-591) in orange; CTD2
(residues 592-686) in wheat color ; upstream helix (UH) (residues 736-781) in red; HR1 (residues
910-985) in pink; CH (residues 986-1035) in hot pink; antiparallel core β-sheet (residues 711-
736, 1045-1076) (in blue).
25
Figure 4. Sequence and structural conservation of cysteine clusters in the SARS-CoV-2 spike
prefusion and postfusion states. (A) The cryo-EM structure of the SARS-CoV-2 S protein in the
prefusion form is shown in ribbons. The protomer A is in green, protomer B in red, and protomer
C in blue colors. The positions of the conserved cysteine clusters are shown in yellow spheres. (B)
Structural arrangement of the conserved cluster formed by C720, C725, C731, and C742 in the
UH region. The UH fragment is shown in red ribbons and conserved cysteine sites are shown in
yellow spheres and annotated. (C) Structural organization of the conserved cysteine cluster in the
β-hairpin region formed by C1014 and C1025, C1064 and C1108. The protein fragment is shown
in red ribbons and conserved cysteines are in yellow spheres and annotated. (D) The cryo-EM
structure of the SARS-CoV-2 S protein in the postfusion form. The protomer A is in green,
26
protomer B in red, and protomer C in blue colors. The conserved cysteine clusters are in yellow
spheres. (E) Structural arrangement of conserved cluster formed by C720, C725, C731, and C742
in the postfusion state. (F) Structural organization of a conserved cysteine cluster (C1014 and
C1025, C1064 and C1108) in the post-fusion state. The conserved cysteine sites are shown in
yellow spheres and annotated.
Structural analysis demonstrated that the core elements of S2 regions anchored by several cysteine
clusters are highly preserved in the SARS-CoV-2 spike prefusion and postfusion states (Figure
4A,D) despite massive conformational changes of the SARS-CoV-2 S2 machinery.125,126 Some
of these regions include cysteine clusters formed by C720, C725, C731, and C742 in the upstream
helix (UH) regions. (Figure 4B,E). Another conserved cysteine cluster of disulfide bonds is
formed in the β-hairpin region (residues 1045-1076) located downstream of the CH region by
residues C1014 and C1025 (C1032 and C1043 respectively in SPIKE_SARS2 sequence
numbering) as well as residues C1064 and C1108 (C1082 and C1126 in SPIKE_SARS2 sequence
numbering) (Figure 4C,F). This conserved segment of S2 subunit is a part of the antiparallel core
β-sheet assembled from an N-terminal β-strand (β46) and a C-terminal β-hairpin (β49–β50). The
top conserved RBD positions included C336, R355, C361, F374, F377, C379, L387, C391, D398,
G413, N422, Y423, L425, F429, C432, and W436 (Figure 3B). The RBD region includes eight
conserved cysteine residues, six of which form three disulfide linkages (C336–C361, C379–
C432 and C391–C525), which stabilize the β-sheet RBD structure in the SARS-CoV-2 S protein.
The crystal structures of S-proteins highlighted that two of these disulfide bonds are potentially
redox-active, facilitating the primal interaction between the receptor and the spike protein.121
27
We particularly focused on conservation patterns of the SARS-CoV-2 RBD residues forming
binding epitopes for H014, S309, S2E12 and S2M11 Abs. The H014 epitope is fairly large and
broadly distributed across RBD regions (residues 368 to 386, 405 to 408 and 411 to 413, 439,
and 503) forming a cavity on one side of the RBD (Tables S1 and S2, Supporting Information).
Although most of contacts are formed with moderately conserved residues, H014 makes favorable
interactions with two most highly conserved F377 and C379 positions in the RBD region (Figure
3B). Additionally, H014 makes strong interaction contacts with several other conserved RBD
positions including Y380, S383, P384, K386, G413, and W436 (Figure 3B). Of particular
importance are H014 contacts with S383 and G413 residues that are located at the inter-protomer
boundaries (S383-D385) and (G413-V987) and could function as regulatory switches of S protein
equilibrium.30 It is worth mentioning that disulfide-bonded S-protein trimer variants S383C
/D985C at the RBD to S2 boundaries can lead to a predominant population of the prefusion closed
states.30 Notably, the conformational epitope for H014 is only accessible when the RBD in an
open conformation. According to our analysis, H014 interactions with conserved positions F377,
C379, S383 located away from the RBM region could be important for binding and modulation
of the enhanced cross-neutralization activities.
S309 engages an epitope distinct from the RBM making contacts with two most conserved RBD
positions C336 and C361 (Figure 3B, Tables S3 and S4, Supporting Information). These residues
form one of the disulfide linkages Cys336–Cys361 that stabilize the β-sheet RBD structure. S309
forms particularly strong contacts with neighboring residues L335 and P337 that displayed more
moderate conservation. Our analysis indicated that the KL evolutionary score of the S309
contact RBD positons is considerably higher than average, with several interacting residues such
28
as T333, C336, V341, R355, C361 displaying particularly strong conservation (Figure 3B). We
suggest that binding to these highly conserved and structurally rigid positions located away from
the ACE2-binding interface may contribute to a broad neutralizing activity of S309. Indeed,
viruses may have evolved to maintain the sensitive regions of their structure inaccessible to the
immune system. As a result, Abs tend to bind to the most easily accessible regions of the virus,
where viruses can tolerate mutations. By targeting evolutionary conserved RBD epitopes H014
and S309 Abs can potentially better combat virus resistance. Based on this analysis, we
hypothesized that binding of H014 and S309 to conserved epitopes of S proteins that are distinct
from ACE2-binding site may involve allosterically regulated mechanism in which Abs induce
long-range alterations in the interaction networks and allosteric communications.
S2M11 recognizes a quaternary epitope through electrostatic interactions and shape
complementarity, comprising distinct regions of two neighboring RBDs within an S trimer.47 This
Ab is believed to induce inhibition of membrane fusion through conformational trapping of
SARS-CoV-2 S trimer in the closed state.47 S2M11 forms contacts with highly conserved sites
F374 and W436 on one RBD and makes interactions with F486 on the RBM motif of the other
RBD (Table S5, Supporting Information, and Figure 3B). We also characterized the inter-
molecular contacts formed by S2E12 (Table S6, Supporting Information) that targets the RBM
motif and can block the attachment to the ACE receptor. According to our analysis, the contact
positions targeted by these Abs are only moderately conserved as the epitope overlaps with a more
variable RBD region (Figure 3A,B).
29
Coevolutionary Analysis of the SARS-CoV-2 Proteins Reveals Regulatory Centers and
Functional Role of the Epitope Regions in the Network of Evolutionarily Coupled Residues
Coevolutionary dependencies of protein residues can mediate protein recognition and are often
spatially close to each other, forming clusters of interacting residues that are located near
functionally important sites.129-132 The coevolving residues may be clustered in mobile regions
and form interaction networks of evolutionarily coupled residues that facilitate protein
conformational changes.133 Using MISTIC approach71-73 we determined coevolutionary
dependencies between S protein residues using plmDCA (Figure 5) and gaussianDCA models
(Figure S1, Supporting Information). The network of coevolutionary couplings in the SARS-
CoV-2 S structures was then constructed in which the nodes represented protein residues and
links corresponded to coevolutionary dependencies between these residues. To identify critical
nodes of this coevolutionary network that may coordinate and transmit coevolutionary signals, we
computed plmDCA-based cScore profiles (Figure 5A) that measure the global influence of a
given position in a coevolutionary network. This score is calculated as the sum of covariation
scores above a certain threshold ( top 5% of the covariation scores) for every position pair where
the particular position appears. Using this approach, we quantified coevolutionary relationships
between residues, identified coevolutionary couplings for functionally important regions and also
mapped high CScore positions onto the binding epitopes for studied SARS-CoV-2 complexes
(Figure 5B).
We focused our analysis on the computed distribution of plmDCA-based CScore profiles (Figure
5A) that revealed several important trends. First, the results revealed an appreciable density of
coevolving centers in the S1 subunit, primarily in the RBD and especially CTD1 regions. This
pattern can be further illustrated by a circular representation of the pairwise coevolutionary scores
30
(Figure S2, Supporting Information) showing the greater concentration of coevolutionary links
anchored by the CTD1 regions (residues 528-591). The distribution of CScores pointed to the
significantly higher density of coevolutionary couplings in the tightly packed S2 subunit (Figure
5A). The residues with significant CScore values are distributed across various S2 regions,
including UH (residues 736-781), CH (residues 986-1035) , HR1 region (residues 910-985), HR2
(residues 1163-1211) and β-hairpin (BH) region (residues 1035-1071). A dense network of
coevolutionary coupled residues in the S2 regions can be therefore detected as evident from a
graphical annotation of the pairwise coevolutionary scores (Figure S3, Supporting Information).
Interestingly, the distribution of CScore values for the epitope residues showed that many contact
positions are aligned with highly coevolving residues (Figure 5B). Among high cScore sites that
establish intermolecular contacts with Abs are C336, R355, C361, F374, F377, C379, L387,
C391, D398, G413, N422, Y423, L425, F429, C432, W436 (Figure 5B). Some of these residues
are highly conserved (C336, C361, C379) while other sites exhibited moderate to high
conservation level (L387, W436).
Structural analysis of coevolutionary hotspots corresponding to the local maxima of the
distribution revealed presence of clusters situated in functional regions (Figure 6). We observed
that coevolutionary centers can be localized in the key regions of the SARS-CoV-2 S protein,
occupying the proximity of the SA1/S2 cleavage site, the HR1 and CH regions of S2 subunit as
well as RBD and CTD1 regions in the S1 domain (Figure 6). Of particular interest are several
coevolutionary hotspots located near a well-recognized cleavage site at the S1/S2 boundary. The
furin cleavage site emerges as a disordered loop (residues 655-GICASYHTVSLLRST-669 in the
SPIKE_SARS
sequence
or
residues
669-GICASYQTQT-NSPRRARSVA-688
in
SPIKE_SARS2 sequence).
31
Figure 5. Coevolutionary profiles of the SARS CoV-2 S proteins. (A) The plmDCA-based
coevolutionary Cscore profile for the SARS-CoV-2 S proteins (P0DTC2: SPIKE_SARS2
sequence numbering). The Cscore values are shown for the S1-NTD residues in green bars
(Pfam:PF16451), for the RBD in red bars (Pfam:PF09408) and for S2 regions in blue bars
(Pfam:PF01601). (B) A close-up of the CSscore profile for the RBD regions is shown in red
bars. The CScores for the binding epitope residues of H014, S309, S2M11, and S2E12 are shown
in filled green circles. (C) The distribution of the inter-residue contacts in the S1-RBD regions
(red bars) and S2 regions (blue bars). The highly coevolving centers in the RBD regions are in
maroon-colored filled circles and the high CSscore residues in S2 regions are in orange-colored
32
filled circles. (D) The distance probability distribution of directly coupled residue pairs in the
studied SARS-CoV-2 S complexes is shown in red filled bars.
Our analysis showed that residues immediately C-terminal to the S1/S2 cleavage site such as
S670, K672, S673, Y676, M677, S678, S681 featured high cScore values and formed a cluster
of coevolutionary centers in the S2 subunit (Figures 5,6). The multi-basic S1/S2 site in SARS-
CoV-2 harbors multiple arginine residues and is involved in proteolytic cleavage of the S protein
which is critical for viral entry into cells.134 A particular relevance of this site stems from the fact
that sequence of the S1/S2 site enables cleavage by furin in SARS-CoV-2 but not in SARS-CoV
or MERS-CoV viruses.135 The experimental data also showed that SARS-CoV S-mediated virus
entry is based on sequential proteolytic cleavage at two distinct sites, with cleavage at the S1/S2
boundary (R667) promoting subsequent cleavage at the S2′ position (R797) triggering membrane
fusion.136 Interestingly, our results indicated a high coevolutionary signal for R667 at the S1/S2
boundary but only a moderate Cscore for the conserved R797 position (Figure 5). These findings
are generally consistent with coevolutionary patterns found in disordered protein regions showing
that disordered residues whose function requires specific recognition and disorder-order transition
upon binding can exhibit a high degree of coevolutionary signal.137
Although many coevolving centers in the S2 subunit are located inside the protein core and
generally stable, these regions are involved in gigantic conformational rearrangements to the post-
fusion state that require a nontrivial cooperation between these regions to dramatically rearrange
the interaction network. Several important clusters of highly coevolving centers are localized in
the HR1 regions (N925, A930, K947, N953, L959, F970, V976, L977, and L984) and CH regions
( P987, E988, I993, D994, R995, I997, L1004, Y1007, T1027, and L1039) (Figures 5,6).
33
Figure 6. Structural analysis of coevolutionary hotspots in the SARS-CoV-2 S proteins.
(A) Structural map of high CSscore residues shown in red spheres is projected on the cryo-EM
structure of the SARS-CoV-2 S protein. (B) A close-up of coevolutionary centers mapped onto a
single protomer of the S protein. The protomer is shown in cyan ribbons and high CSscore
positions are depicted in red spheres. The map shows localization of coevolutionary hotspots in
the key regions of the SARS-CoV-2 S protein, occupying the proximity of the SA1/S2 cleavage
site, the HR1 and CH regions of S2 subunit as well as RBD and CTD1 regions in the S1 domain.
34
An interesting cluster of coevolving centers is formed by residues from different S regions
surrounding the C-terminus of the central HR1-CH helices (Figure 6). This cluster included HR1
residues Q920, N925, F927, β-hairpin (BH) motif residue F1052, F898 (from connecting region
841-911) and several other hydrophobic positions F800 and F802 from the region upstream of the
fusion peptide FP (816-SFIEDLLFNKVTLADAGF-833) (Figure 6B). These results showed that
a number of the interface core and inter-protomer centers in the S2 subunit featured a significant
coevolutionary signal. Consistent with the coevolutionary studies of protein complexes,138 we
found that coevolutionary signal can be significant for the S2 positions involved in multiple
interactions at critical junctures of UH, HR1 and CH regions. Hence, the increased structural and
functional constraints for sites involved in significant number of inter-residue contacts can often
imply the higher coevolution values. We found that both conservation and coevolutionary signals
can increase for the S2 core residues involved in the inter-protomer interfacial contacts. However,
the S2 core residues with the strongest coevolutionary signal and highest Cscore values could
feature different level of conservation. In particular, the strongest coevolutionary centers in the
S2 regions included fairly moderately conserved residues F1089, R983, N925 and L984 as well
as strongly conserved A893 , V915, L916, and A1190 positions. In general, our results indicated
that S2 core residues subjected to more structural constraints and inter-residue contacts can exhibit
the higher residue conservation and coevolution values.
To further probe the notion that the interface core residues can exhibit both the higher level of
conservation and coevolution, we computed the average number of the inter-residue contacts for
each S protein residue and aligned this distribution with the top Cscore positions in the S1/RBD
and S2 regions (Figure 5C). The results indicated that coevolutionary centers tend to have a fairly
significant number of the interacting contacts and can be involved in multiple interactions. In
35
particular, coevolutionary hotspots in the RBD regions were often aligned with the peaks of the
contact distribution, supporting the notion that the level of coevolution may be greater in residues
involved in multiple interactions.138 It is worth noting, however, that sites with the largest number
of inter-residue contacts may not necessarily correspond to the most conserved positions or
residues with the highest CScore value. In fact, even though coevolutionary hotspots in the S2
subunit have a significant number of the inter-residue contacts, the distribution peaks
corresponded to residues F718, V729, I742, F782 with moderate levels of conservation and
coevolution (Figure 5C).
We also computed the distance probability distribution of coevolving directly coupled residue
pairs in the studied SARS-CoV-2 S structures (Figure 5D). The profile showed several local
maxima at 3.2 Å, 4.7 Å and a much broader area with a shallow peak near 7 Å - 8Å (Figure
5D). It is evident that the first two peaks reflect physical interactions between residues including
hydrogen bonding and hydrophobic residue pairs. Hence, direct coevolutionary residue couplings
in the SARS-CoV-2 S structures are dominated by spatially proximal residue pairs, that is
consistent with large-scale investigations of direct coevolutionary couplings in proteins
suggesting that coevolutionary signals are stronger for locally interacting residues than for residues
involved in long-range interactions in allosteric networks.139 Nonetheless, the distribution also
highlighted another intermediate range of coevolutionary couplings at 7 Å - 8Å that is beyond
direct inter-residue physical contacts and may reflect strong couplings between spatially proximal
functional regions (Figure 5D). This third distribution peak can correspond to coevolutionary
couplings anchored by CTD1 regions (residues 529-591) in the S1 subunit that is believed to
function as allosteric connector between RBD and FPPR regions by communicating signal from
and to the fusion peptide.26
36
Figure 7. Structural maps of coevolutionary centers in the epitope regions of the SARS-CoV-2
complexes with Abs. (A) Structural map of coevolutionary centers in the S complex with H014
(pdb id 7CAI/7CAK) projected onto a single “up” protomer shown in green ribbons. The
coevolutionary centers are in spheres and high CSscore hotspots from the binding epitope are in
red spheres. A close-up of the H014 binding epitope with the coevolving centers involved in direct
contacts with H014 in red spheres and annotated. (B) Structural map and close-up of
coevolutionary centers in the S complex with S309 (pdb id 6WPT). The coevolving centers
involved in direct contacts with S309 in red spheres and annotated. (C,D) Structural map and
close-up of coevolutionary centers in the S complex with S2M11 (pdb id 7K43) and S2E12 (pdb
37
id 7K4N). The coevolving centers involved in direct contacts with S2M11 and S2E12 are in red
spheres and annotated.
Hence, our results revealed the presence of a significant coevolutionary signal between functional
regions separated by the medium-range distances which may help to facilitate a long-range cross-
talk between distant allosteric regions in the S1 and S2 subunits.
Structural mapping of coevolutionary centers highlighted global connectivity of the
coevolutionary network spanning from the epitope binding site towards the CTD1 region and
regions in the S2 subunit (Figure 7). Collectively, these clusters could form modules of a
coevolutionary network that may allow for efficient allosteric interactions and communications
in the SARS-CoV-2 S proteins. In the SARS-CoV-2 complexes with H014, the contact interface
is fairly large and involves a significant stretch of the RBD residues. Several high CScore
residues W436, G413, F374, F377, and C379 are involved in the interactions with H014. In
particular, multiple favorable interactions are formed by F377 with Y105, T58, S59, D60, Y50
of H014 and by C379 with N55, T58, G56,G57 positions of H014 (Figure 7A). S309 binding
involves interactions with the highly conserved C336 and C361 positions that also correspond to
coevolutionary hotspots and could anchor a network of evolutionary coupled residues in the S
protein (Figure 7B). S2M11 interacts with the evolutionary conserved RBD sites F374 and W436
that also displayed high CScore values (Figure 7C). A smaller patch of coevolutionary centers is
involved in contacts with S2E12 that connects the epitope with the S2 subunit via CTD1 region
that serves as a mediating hub of coevolutionary clusters in the S1 (Figure 7D). Collectively, these
clusters could form modules of a coevolutionary network that may allow for efficient allosteric
interactions and communications in the SARS-CoV-2 S proteins.
38
Conformational Dynamics and Mutational Scanning Reveal Modulation of Protein
Stability and Binding Energy Hotspots of the SARS-CoV-2 Spike Complexes
We employed multiple CABS-CG simulations followed by atomistic reconstruction and
refinement to provide a detailed comparative analysis of dynamic landscapes that are
characteristic of the SARS-CoV-2 S trimer complexes with H014, S309, S2M11, and S2E12 Abs.
Using these simulations, we examined how Ab binding could affect the global dynamic profiles
of the closed, partially open, an open states revealing the important regions of flexibility (Figure
8). The analysis of the inter-residue contact maps (Figure S4, Supporting Information)140 and
inter-residue distance maps (Figure S5, Supporting Information)141 in the SARS-CoV-2 S
complexes with Abs indicated that the density of the interaction contacts is significantly greater in
the densely packed S2 domains. The overall packing density of the closed S protein
conformations complexed with S309 and S2M11 is also markedly higher as compared to the
partially open and open states. Molecular simulations of the SARS-CoV-2 S complexes provided
a quantitative picture of the differences in flexibility of the S protein states and the effect of Ab
binding on modulation protein stability. A comparative analysis of the dynamics profiles showed
that H014 binding can induce the significant dynamic changes by considerably reducing thermal
fluctuations in the S1 regions of the Ab-interacting open protomers as compared to the unbound
trimer form (Figure 8A,B). We observed small thermal fluctuations with RMSF < 1.0 Å for the
S1 epitope positions (residues 368 to 386, 405 to 408 and 411 to 413, 439, and 503) that were
considerably rigidified in both H014 complexes (Figure 8A,B). These findings are consistent with
the experimental structural data suggesting that Ab-induced structural changes could trigger
stabilization changes in both the RBD and NTD regions.45
39
Figure 8. CABS-GG conformational dynamics of the SARS-CoV-2 S complexes. A schematic
representation of domain organization and residue range for the full-length SARS-CoV-2 S
protein is shown above conformational dynamics profiles. (A,B) The root mean square
fluctuations (RMSF) profiles from simulations of the cryo-EM structures of the SARS-CoV-2 S
trimer with H014. (C,D) The RMSF profiles from simulations of the cryo-EM structures of the
SARS-CoV-2 S trimer with S309. (E) The RMSF profiles from simulations of the cryo-EM
structures of the SARS-CoV-2 S trimer with all three RBDs in the closed form bound with
S2M11. (F) The RMSF profiles from simulations of the cryo-EM structure of the SARS-CoV-
2 S protein trimer with all three RBDs in the open-up form bound with S2E12. The profiles
40
for protomer chains A,B and C are shown in green, red and blue bars respectively. The RMSF
profiles for the unbound forms of S protein trimer are shown in light grey bars.
Conformational dynamics of the SARS-CoV-2 S protein complex with S309 showed only minor
changes in the flexibility upon binding, particularly in the complex with S309 bound to 2 closed
protomers (Figure 8C). In this case, the unbound open protomer displayed an appreciable
flexibility, while the NTD regions of S309-bound closed protomers also showed some degree of
mobility. In the S309 complex with 3 Abs bound to closed protomers, we found that stability of
the closed S protein is protected, with only minor changes in the local dynamics between unbound
and S309-bound S forms (Figure 8D). S2M11 functions by locking down the SARS-CoV-2 S
trimer in the closed state through binding to a quaternary epitope. Conformational dynamics profile
of the S protein complex with S2M11 in the closed form reflected this mechanism by featuring an
extremely stable SARS-CoV-2 S conformation in which both S1 and S2 regions were virtually
immobilized and displayed only very minor thermal fluctuations (Figure 8E). Interestingly,
according to our analysis, this is the most stable bound form of the SARS-CoV-2 S protein among
studied systems, suggesting that ultra-potent neutralization effect may be partly determined by the
extreme thermodynamic stabilization of the closed-down S protein form. The mechanism of
S2E12 neutralization of SARS-CoV-2 S protein is based on direct targeting of the RBM regions
and interfering with ACE2 binding. A relatively small binding epitope in the S2E12 complex with
the fully open form of S protein produced the dynamics profile where NTD and RBD regions
showed an appreciable degree of mobility, while the S2 regions were mostly immobilized (Figure
8F). The important finding of this analysis was that H014, S309 and S2M11 Abs can exert
modulation of the conformational dynamics leading to a significant stabilization of both S1 and
41
S2 regions in the open protein forms, which may effectively counteract the intrinsic flexibility of
the receptor-accessible, open S conformations and thus induce potent neutralization effects.
To establish connection between dynamics and energetics of the SARS-CoV-2 binding, we
employed the conformational ensembles generated in simulations and performed a systematic
alanine scanning of the protein residues (Figure 9). The results revealed a wide range of important
binding hotspots in the S protein complexes with H014 (Figure 9A,B). This is consistent with the
dynamics profile showing a broad stabilization of the RBD regions, including the epitope residues
and RBM positions. In particular, alanine scanning showed a significant contribution of
conserved RBD residues F374, F377, K378, C379, Y380, P384, T385 as well as N437, V503,
Y508 (Figure 9B). Among binding energy hotspots we detected some of the highly conserved
positions and several coevolutionary centers such as F374, F377, and C379 residues. We argue
that through interactions with major coevolutionary centers in the conserved RBD epitope,
H014 may exert its long-range effect by propagating binding signal through clusters of proximal
coevolutionary pairs in the RBD and CTD1 regions. The noticeably fewer number of binding
hotspots were seen in the S309 complexes with partially closed (2-down) and fully closed forms
of the S protein (Figure 9C,D). The determined binding hotspots L335, P337, T345, and L441 are
characterized by only moderate conservation and CScore values. S309 also makes weaker
contacts with conserved and coevolutionary important RBD centers C336 and C361. However,
the binding free energy changes caused by alanine mutations in these positions are fairly moderate
(~0.7 - 0.8 kcal/mol). The binding energy hotspots in the S2M11 complex with S protein occupy
two different regions, where one group includes conserved RBD sites F374 and W436 that also
displayed high CScore values (Figure 9E). Another group of binding energy hotspot positions
42
includes moderately conserved residues Y449, F456, F484, F486, Y489 that form a critical patch
of the RBM binding interface with the host receptor.
Figure 9. Alanine scanning of the binding epitope residues in the SARS-CoV-2 S complexes
with a panel of Abs. The binding free energy changes upon alanine mutations for the epitope
residues in the SARS-CoV-2 S complex with H014 - two RBDs in the open state, pdb id 7CAI
(panel A), SARS-CoV-2 S complex with H014 - three RBD in the open state, pdb id 7CAK (panel
B), SARS-CoV-2 S complex with S309 - two RBDs in the closed form, pdb id 6WPT (panel C),
SARS-CoV-2 S complex with S3090 - three RBDs in the closed form, pdb id 6WPS (panel D),
SARS-CoV-2 S complex with S2M11 - three RBDs in the closed form, pdb id 7K43 (panel E),
SARS-CoV-2 S complex with S2E12 - three RBDs in the closed form, pdb id 7K4N (panel F).
43
The computed binding free energy changes values are shown in bars. The binding interface
residues are determined for each complex based on the average interaction contacts that persist
during simulation of a given complex.
We previously showed that a conserved segment 486-FNCYFPL-492 in the RBD region emerged
as a central binding energy hotspot in the SARS-CoV-2 complex with ACE2 receptor.62 Hence,
through binding to two adjacent protomers S2M11 can simultaneously block interface with ACE2
and influence long-range couplings using a network of coevolutionary coupled residues. In
general, the results indicated that the interactions S014 and S2M11 Abs can lead to stabilization
of the S conformations and emergence of multiple binding energy hotspots. By targeting these
centers H014 and S309 can exert their neutralizing effect by achieving a strong binding affinity
with the SARS-CoV-2 S protein and also by strengthening long-range couplings of S1 an S2
regions (Figure 9).
Hierarchical Network Modeling Reveals Mediating Centers of Allosteric Interactions in the
SARS-CoV-2 Spike Complexes
We applied a hierarchical-based network modeling approach in which the residue interactions
and network couplings are described with the increasing level of atomistic details and complexity.
First, a protein contact network was implemented to highlight the topological role of residues in
protein structure activity and identify residues mostly responsible for signal transmission
throughout the protein structure. In this simplified model, the protein residues correspond to
network nodes and inter-residue contacts are considered as active links based on distance criteria
as described in our previous studies.98-100 Based on the hierarchical clustering algorithm, we
computed the average participation coefficient ������������ values that measure the contribution of residue
44
nodes in communication between different clusters (functional domains). To focus analysis on
several prominent cases, we reported the communicating residues in the SARS-CoV-2
structures bound with H014 (Tables S7,S8, Supporting Information) and S309 (Tables S9,S10,
Supporting Information). The results indicated that the majority of the inter-cluster communcating
sites are localized in the RBD and especially CTD1 regions for SARS-CoV-2 S complexes with
H014 (Tables S7,S8, Supporting Information). In this case, by attentuating mobility of the
interacting RBD regions H014 binding may activate allosteric interaction networks and
communications between S1 and S2 regions with CTD1 residues acting as global mediating
centers of long-range interactions. The distribution of communicating positions in the SARS-CoV-
2 S complexes with S309 (Tables S9,S10, Supporting Information) revealed an appreciably larger
number of potential mediatig centers with significant communication propensities. Moreover,
these positions corresponded to different regions, including a significant number of mediating hubs
in the UH, CH and HR1 regions of S2 subunit as well as residues in the CTD1 regions of S1.
These preliminary findings suggested that allosteric interaction networks in the SARS-CoV-2 S
complexes with S309 could be broadly distributed, which can arguably reflect strengthening of
allosteric couplings between S1 and S2 subunits as S309 locks the down-regulated form of the S
protein. In the framework of the hierarchical approach, we also explored a more detailed model
of the residue interaction networks by using a graph-based representation with residues as network
nodes and the inter-residue edges defined by both dynamic correlations104 and coevolutionary
residue couplings 105 as detailed in our previous studies.105-107 Using the results of simulations,
the ensemble-averaged distributions of the betweenness centrality were computed for the SARS-
CoV-2 S complexes with Abs (Figure 10). We found that the high centrality residues can be
assembled in tight interaction clusters localized in the key functional regions of the S protein. In
45
the SARS-CoV-2 S protein complexes with H014, the centrality profiles featured strong and dense
peaks in the RBD and CTD1 regions of S1 as well as another peak in the CH region of S2 (residues
986-1035) (Figure 10A,B). The centrality peaks also aligned well with the hinge centers of S1
(residues 315-320, 569-572), indicating that these dynamically important control points could
also mediate communication in the residue interaction networks. The network centrality analysis
also revealed clusters of distribution peaks in the SARS-CoV-2 S complexes featuring the fully
closed conformation (Figure 10D,E). In these structures S309 and S2M11 induce a strong
stabilization effect and lock the S protein in the closed state. According to our results, these
structurally stable states can also feature a broadly distributed allosteric network mediated by
functional sits in both S1 and S2 subunits, primarily CTD1 regions (residues 529-591) UH
(residues 736-781), CH (residues 986-1035), and β-hairpin (BH) region (residues 1035-1071).
The dominant clusters of centrality peaks located in the RBD and CTD1 regions of S1 and CH
regions of S2 can be seen in the S complex with S2E12 (Figure 10F). This showed that S2E12
binding may activate the increased mediating capacity of CTD1 regions and strengthen allosteric
interactions between S1 and S2 regions.
Structural mapping of high centrality sites highlighted differences between network organizations
in the SARS-CoV-2 complexes (Figure 11). In the complexes with H014 the high centrality sites
are concentrated near CTD1 regions that could strengthen couplings at the inter-domain
boundaries between S1 and S2 (Figure 11A,B). We argue that H014 binding may increase the
allosteric potential of the RBD and CTD1 regions and activate communication between the RBD
and S2 via CTD1 regions. Of particular interest is a dense network of mediating centers in the
complexes with S309 and S2M11 (Figure 11C-E), showing that these Abs may facilitate a broad
allosteric interaction network between S1 and S2 functional regions.
46
Figure 10. The residue-based betwenness centrality profiles in the SARS-CoV-2 S complexes
with a panel of Abs. The centrality values are computed by averaging the results over 1,000
representative samples from CABS-CG simulations and atomistic reconstruction of trajectories.
(A) The centrality profile is shown for the SARS-CoV-2 S complex with H014 - two RBDs in
the open state ( A), SARS-CoV-2 S complex with H014 - three RBD in the open state (B), SARS-
CoV-2 S complex with S309 - two RBDs in the closed form (C), SARS-CoV-2 S complex with
S309 - three RBDs in the closed form (D), SARS-CoV-2 S complex with S2M11 - three RBDs in
the closed form (E), SARS-CoV-2 S complex with S2E12 - three RBDs in the closed form ( F).
The profiles for protomer chains A,B and C are shown in green, red and blue bars respectively.
47
Figure 11. Structural maps of high centrality clusters in the SARS-CoV-2 S complexes. (A)
Structural map for the SARS-CoV-2 S complex with H014 - two RBDs in the open state ( A),
SARS-CoV-2 S complex with H014 - three RBD in the open state (B), SARS-CoV-2 S complex
with S309 - two RBDs in the closed form (C), SARS-CoV-2 S complex with S3090 - three
RBDs in the closed form (D), SARS-CoV-2 S complex with S2M11 - three RBDs in the closed
form (E), SARS-CoV-2 S complex with S2E12 - three RBDs in the closed form ( F). The
protomer A is shown in green ribbons, protomer B in cyan ribbons, and protomer C in blue ribbons.
The bound Abs are depicted in dark pink-colored ribbons. The high centrality residue clusters are
shown in red spheres.
48
Perturbation Response Scanning Identifies Regulatory Hotspots of Allosteric Interactions in
Different Conformational States of the SARS-CoV-2 Spike Trimer
Using the PRS method112,113 we quantified the allosteric effect of each residue in the SARS-CoV-
2 complexes in response to external perturbations. PRS analysis produced the residue-based
effector response profiles in different functional states of the unbound SARS-CoV-2 S trimer
(Figure 12) and SARS-CoV-2 S complexes with H014, S309, S2M11, and S2E12 Abs (Figure 13).
The effector profiles estimate the propensities of a given residue to influence dynamic changes
in other residues and are applied to identify regulatory hotspots of allosteric interactions as the
local maxima along the profile. The central hypothesis tested in the PRS analysis is that Ab
binding can incur measurable and functionally relevant changes by modulating the effector
profiles of the unbound SARS-CoV-2 S protein timer. Moreover, we conjectured that binding can
differentially affect the effector response profiles and allosteric interaction networks in distinct
functional forms of the SARS-CoV-2 S protein. By systematically comparing the PRS profiles in
the unbound and bound S protein forms, we determined the distribution of regulatory allosteric
centers and clarified the role of specific functional regions in controlling allosteric
conformational changes.
To establish the baseline for comparison of allosteric profiles in the SARS-CoV-2 complexes we
first computed PRS effector profiles for the unbound S protein in the closed (3 RBDs-down),
partially open (2RBDs-down, 1RBD-up) and open states (1RBD-down, 2 RBDs-up) (Figure 12).
In this analysis of the unbound forms of the SARS-CoV-2 S trimer, we used the cryo-EM structure
of the full-length quadruple mutant (A507L T572I F855Y N856I) in the all-down closed
prefusion conformation (pdb id 6X2C), the partially open (1-up) conformation (pdb id 6X2A),
and the open (2-up) conformation (pdb id 6X2B).28 These structures revealed thermodynamic
49
and dynamic differences between different states in which dynamic switch centers are responsible
for modulation of allosteric changes between the closed and open S states.28 By using these
structures as reference states for the PRS analysis, we also examined how Ab binding can alter the
localization and effect of these regulatory switch points on allosteric interactions.
The results showed that the effector profile in the unbound S structures can remain largely
conserved, displaying the largest peaks in the functional regions of the S2 subunit (Figure 12A-
C). The major effector peaks corresponded to residues 756-758 in the UH region, HR1 region
(residues 910-985), residues 886-890 and BH region (residues 1035-1071). In the closed and
partially open forms, we also observed a secondary peak corresponding to CTD1 regions (residues
529-591) in the S1 subunit. Only a minor effector density was seen in the RBD regions (Figure
12A-C). At the same time, these regions can serve as primary sensors of allosteric signaling in
the S trimers for the partially open and open states (Figure S6, Supporting Information). These
results are consistent with our latest studies of the SARS-CoV-2 S structures demonstrating that
the broadly distributed effector density can be seen only in the fully locked closed state, while
allosteric couplings in more dynamic closed and open forms may be largely governed by the
regulatory centers in core of the S2 subunit.64 The reduced density of the effector centers in the
RBD regions indicated that the allosteric signaling in the dynamic closed and open forms may
be primarily one-directional, in which allosteric centers in the S2 core regions could dictate
allosteric changes in the RBD regions. Based on this preliminary evidence, we suggested that
efficient allosteric signaling between S1 and S2 subunits and broad allosteric networks may be
salient features of the thermodynamically locked closed S form.
50
Figure 12. The PRS effector profiles in the closed, partially open and open states of the SARS-
CoV-2 spike trimers. (A) The PRS effector profile is shown for the ligand-free SARS-CoV-2 S
trimer in the partially open (1 RBD up) conformation (pdb id 6X2A). (B) The PRS effector
profile for the ligand-free SARS-CoV-2 S trimer in the open (2 RBDs up) conformation (pdb id
6X2B). (C) The PRS effector profile for the ligand-free SARS-CoV-2 S trimer in the closed (3
RBDs-down) prefusion conformation (pdb id 6X2C). The profiles are shown for protomer in green
lines, protomer B in red lines, and protomer C in blue lines. Structural maps of the PRS effector
profiles are shown for the partially open state of the SARS-CoV-2 S prefusion trimer (D), open
state (E), and closed state (F). The color gradient from blue to red indicates the increasing effector
propensities.
51
We conjectured that Ab binding at different RBD epitopes may affect not only local interactions
and stability near the binding sites but also have long-range effect by modulating allosteric effector
potential of SARS-CoV-2 S regions and altering allosteric interaction networks. This can allow
for highly cooperative motions in which many spatially distributed effector residues are in the
allosteric network that links distant S1 and S2 functional regions. In this model, allostery requires
an effector ligand to stabilize the interactions in the closed S state over those in the open S states.
The central result of this analysis is that the studied neutralizing Abs could effectively restore
allosteric potential of the RBD and CTD1 regions in the closed and open states without
compromising the effector potential of S2 regions, thereby introducing ligand-induced
cooperativity and strengthening the broad allosteric interaction network (Figure 13). Indeed,
the results showed that H014 binding induced significant changes in the effector profile of the
RBD and CTD1 regions for “up” protomers while preserving and enhancing the effector capacity
of the CH and CD residues (Figure 13A,B). In the partially open form of the SARS-CoV-2
complex, the effector peak center corresponded to RBD residues S383, F377, K378, C379, Y380,
G381, V382, S383, P384 involved in direct productive interactions with H014 (Figure 13A).
Several of these effector centers (F377, C379, and S383) also corresponded to conserved
positions implicated in mediating coevolutionary couplings. In addition, we noticed a significant
increase of the effector potential for the CTD1 regions in the open protomers. Similar findings
were observed in the open form ( 3RBDs - up) of the SARS-CoV-2 S complex with H014 where
the allosteric potential of RBD regions was markedly enhanced (Figure 13B).
Structural maps of the effector profiles further illustrated these findings by showing the increased
effector density penetrating into the RBDs of protomers that interact with H014 (Figure 13A,B).
52
These results suggested that H014 binding could modulate the effector propensities of the RBD
residues and restore the allosteric potential of the S1 regions, leading to strengthening of the
network of allosteric interactions in the complex. The effector profiles also indicated the increased
density and clustering of effector peaks distributed across RBD, CTD1, UH and CH regions
(Figures 13A,B).
The effector profiles of SARS-CoV-2 complexes with S309 showed the significantly increased
effector potential of the RBD and CTD1 regions, which is manifested in the emergence of
dominant and broad peaks in these S1 regions (Figure 13C,D). Among emerging effector peaks
in the RBD regions are T333, N334, C361, V362, A363, L390 and C391. These residues
correspond to highly conserved and structurally stable positions involved in stabilizing disulfide
linkages, C336–C361 and C391–C525 that anchor the β-sheet structure in the RBD regions. The
S309-induced modulation of allosteric effector propensities could be also amplified by the fact
that this Ab appears to thermodynamically stabilize partially closed and closed forms of the S
protein where two or all three RBDs assume the down-regulated conformation (Figure 13C,D).
Structural maps highlighted a significant expansion of the effector density towards S1 regions
and boundaries between S1 and S2 subunits (Figure 13C,D). By strengthening allosteric couplings
between S1 and S2 subunits, S309 could arguably lock the down-regulated form to ensure S1-
based protection of the fusion machinery.
53
Figure 13. The PRS effector profiles for the SARS-CoV-2 S complexes with Abs. (A) The
PRS effector distributions and structural maps of the effector profile are shown for the SARS-
CoV-2 S complex with H014 - two RBDs in the open state (A), SARS-CoV-2 S complex with
H014 - three RBD in the open state (B), SARS-CoV-2 S complex with S309 - two RBDs in the
closed form (C), SARS-CoV-2 S complex with S3090 - three RBDs in the closed form (D),
SARS-CoV-2 S complex with S2M11 - three RBDs in the closed form (E), SARS-CoV-2 S
complex with S2E12 - three RBDs in the closed form (F). The profiles are shown for protomer
in green bars, protomer B in red bars, and protomer C in blue bars. Structural maps of the PRS
effector profiles are shown with the color gradient from blue to red indicating the increasing
effector propensities.
54
S2M11 binds a quaternary epitope comprising distinct regions of two neighboring RBDs within
an S trimer and induced stabilization of SARS-CoV-2 S in the closed conformational state. We
found that S2M11 binding can promote the increased effector potential of conserved and
structurally stable residues that are not directly involved in binding contacts (Figure 13E). Indeed,
S2M11 can induce the increased effector potential in the RBD regions, particularly residues F374,
F377, K378 C379, Y380 (Figure 13E). S2E12 recognizes an RBD epitope overlapping with the
RBM that is partially buried at the interface between protomers in the closed S trimer and therefore
S2E12 can only interact with open RBDs. According to our results, S2E12 binding can cause a
similar redistribution of allosteric effector potential in the S1 regions and activate effector capacity
of conserved stretch of residues in the β-sheet of the RBD regions (Figure 13F). Strikingly,
S2E12-induced modulation of the effector propensities in the fully open S conformation can
effectively restore allosteric potential for the RBD and CTD1 S1 regions, while strengthening
peaks in the UH and CH regions of S2 subunit. As a result, as highlighted by structural projection
of the effector profiles, the potential effector centers become spatially distributed across the S
trimer structure (Figure 13F).
The PRS sensor profiles describe the propensity of flexible residues to serve as carriers of large
allosteric conformational changes. We found that Ab binding can induce changes in the shape and
peaks of sensor profiles in the partially open and closed S forms (Figure S7, Supporting
Information). The sensor peaks in the RBD regions that are prevalent in the unbound S protein
become partially suppressed, featuring now a broader distribution of small peaks localized across
S1 and S2 regions. This suggested that the sensor propensities of the RBD regions can be
significantly reduced in the SARS-CoV-2 S complexes, which reflects conformational and
dynamic constraints imposed by Abs on large structural changes.
55
The results of the PRS analysis also suggested that allosteric mechanisms underlying Ab binding
to S proteins may bear signs of ligand-induced cooperativity in which the effector can shift the
distribution of local interactions and energies for many residues.142 Based on our results, we argue
that Abs can induce a switch from a moderately cooperative population-shift mechanism of the
unbound S protein to a highly cooperative ligand-induced allosteric mechanism. While the
allosteric interaction network of the unbound S protein in the population-shift mechanism tends
to be more dispersed and smaller, binding can induce a large and dense allosteric network that
efficiently couples local changes in the distant S1 and S2 regions.
In this context, it is worth noting that cooperative allosteric mechanisms with a broad allosteric
network tend to better withstand mutations in the functional regions without significant deleterious
consequences for protein function.142 Accordingly, it may be suggested that the ligand-induced
cooperative allosteric effect produced by Ab binding may enhance resistance against mutations
so that mutational changes would not easily alter conformational preferences and expose the RBD
regions to interactions with the host receptor.143-147 In some contrast, a less cooperative population-
shift mechanism in the unbound S protein may be more susceptible and vulnerable to mutations
of residues in the communication network, which may allow individual mutations at the regulatory
switch centers to alter conformational equilibrium and potentially increase population of the
receptor-accessible open S conformations.
56
Conclusions
This study examined molecular mechanisms underlying SARS-CoV-2 S protein binding with a
panel of highly potent Abs through the lens of coevolutionary relationships and ligand-induced
modulation of allosteric interaction networks. The results revealed key functional regions and
regulatory centers that govern coevolutionary couplings and allosteric interactions in the SARS-
CoV-2S protein complexes. We found that Ab-specific targeting of coevolutionary hotspots
in the S protein can allow for efficient modulation of long-range interactions between S1 and S2
units by propagating signal through clusters of spatially proximal coevolutionary coupled residues.
The results revealed strong coevolutionary signal between functional regions separated by the
medium-range distances which may help to facilitate a long-range cross-talk between distant
allosteric regions. Conformational dynamics and binding energetics analyses showed that
binding of Abs can lead to significant stabilization of both S1 and S2 regions which may be
relevant in rationalization of potent neutralization effects. The PRS analysis of the unbound and
bound SARS-CoV-2 S proteins showed that Abs can promote formation of highly cooperative
and broad allosteric networks that restore and enhance couplings between S1 and S2 regions,
thereby inhibiting dissociation of S1 subunit from the spike apparatus required for membrane
fusion. By systematically comparing the PRS profiles, we clarified the role of specific functional
regions in regulating allosteric interactions. The results of this study provide a novel insight
into allosteric regulatory mechanisms of SARS-CoV-2 S proteins showing that Abs can uniquely
modulate signal communication providing a plausible strategy for therapeutic intervention by
targeting specific hotspots of allosteric interactions in the SARS-CoV-2 proteins.
57
AUTHOR INFORMATION
* Corresponding Author
Phone: 714-516-4586; Fax: 714-532-6048; E-mail: verkhivk@chapman.edu
The authors declare no competing financial interest.
Acknowledgment This work was partly supported by institutional funding from Chapman
University. The author acknowledges support by the Kay Family Foundation Grant A20-0032.
ABBREVIATIONS
SARS, Severe Acute Respiratory Syndrome; RBD, Receptor Binding Domain; ACE2,
Angiotensin-Converting Enzyme 2 (ACE2); NTD, N-terminal domain; RBD, receptor-binding
domain; CTD1, C-terminal domain 1; CTD2, C-terminal domain 2; FP, fusion peptide; FPPR,
fusion peptide proximal region; HR1, heptad repeat 1; CH, central helix region; CD, connector
domain; HR2, heptad repeat 2; TM, transmembrane anchor; CT, cytoplasmic tail.
58
References
(1) Li, Q.; Guan, X.; Wu, P.; Wang, X.; Zhou, L.; Tong, Y.; Ren, R.; Leung, K. S. M.; Lau, E. H.
Y.; Wong, J. Y.; Xing, X.; Xiang, N.; Wu, Y.; Li, C.; Chen, Q.; Li, D.; Liu, T.; Zhao, J.; Liu, M.;
Tu, W.; Chen, C.; Jin, L.; Yang, R.; Wang, Q.; Zhou, S.; Wang, R.; Liu, H.; Luo, Y.; Liu, Y.; Shao,
G.; Li, H.; Tao, Z.; Yang, Y.; Deng, Z.; Liu, B.; Ma, Z.; Zhang, Y.; Shi, G.; Lam, T. T. Y.; Wu, J.
T.; Gao, G. F.; Cowling, B. J.; Yang, B.; Leung, G. M.; Feng, Z. Early transmission dynamics in
Wuhan, China, of novel coronavirus-infected pneumonia. N. Engl. J. Med. 2020, 382, 1199-1207.
(2)Wang, C.; Horby, P. W.; Hayden, F. G.; Gao, G. F. A novel coronavirus outbreak of global
health concern. Lancet 2020, 395, 470-473.
(3)Yi, Y.; Lagniton, P. N. P.; Ye, S.; Li, E.; Xu, R. H. COVID-19: what has been learned and to
be learned about the novel coronavirus disease. Int. J. Biol. Sci. 2020, 16, 1753-1766.
(4) Wu, A.; Peng, Y.; Huang, B.; Ding, X.; Wang, X.; Niu, P.; Meng, J.; Zhu, Z.; Zhang, Z.; Wang,
J.; Sheng, J.; Quan, L.; Xia, Z.; Tan, W.; Cheng, G.; Jiang, T. Genome composition and divergence
of the novel coronavirus (2019-nCoV) originating in China. Cell Host Microbe 2020, 27, 325-328.
(5) Tai, W.; He, L.; Zhang, X.; Pu, J.; Voronin, D.; Jiang, S.; Zhou, Y.; Du, L. Characterization
of the receptor-binding domain (RBD) of 2019 novel coronavirus: implication for development of
RBD protein as a viral attachment inhibitor and vaccine. Cell. Mol. Immunol. 2020, 17, 613-620.
(6) Hoffmann, M.; Kleine-Weber, H.; Schroeder, S.; Krüger, N.; Herrler, T.; Erichsen, S.;
Schiergens, T. S.; Herrler, G.; Wu, N. H.; Nitsche, A.; Müller, M. A.; Drosten, C.; Pöhlmann, S.
SARS-CoV-2 cell entry depends on ACE2 and TMPRSS2 and is blocked by a clinically proven
protease inhibitor. Cell 2020, 181, 271-280.e8.
(7) Lu, R.; Zhao, X.; Li, J.; Niu, P.; Yang, B.; Wu, H.; Wang, W.; Song, H.; Huang, B.; Zhu, N.;
Bi, Y.; Ma, X.; Zhan, F.; Wang, L.; Hu, T.; Zhou, H.; Hu, Z.; Zhou, W.; Zhao, L.; Chen, J.; Meng,
59
Y.; Wang, J.; Lin, Y.; Yuan, J.; Xie, Z.; Ma, J.; Liu, W. J.; Wang, D.; Xu, W.; Holmes, E. C.; Gao,
G. F.; Wu, G.; Chen, W.; Shi, W.; Tan, W. Genomic characterisation and epidemiology of 2019
novel coronavirus: implications for virus origins and receptor binding. Lancet 2020, 395, 565-574.
(8) Duan, L.; Zheng, Q.; Zhang, H.; Niu, Y.; Lou, Y.; Wang, H. The SARS-CoV-2 spike
glycoprotein biosynthesis, structure, function, and antigenicity: Implications for the design of
spike-based vaccine immunogens. Front. Immunol. 2020, 11, 576622.
(9) Wang, Q.; Zhang, Y.; Wu, L.; Niu, S.; Song, C.; Zhang, Z.; Lu, G.; Qiao, C.; Hu, Y.; Yuen, K.
Y.; Zhou, H.; Yan, J.; Qi, J. Structural and functional basis of SARS-CoV-2 entry by using human
ACE2. Cell 2020, 181, 894-904.e9.
(10) Wan, Y.; Shang, J.; Graham, R.; Baric, R. S.; Li, F. Receptor recognition by the novel
coronavirus from Wuhan: An analysis based on decade-long structural studies of SARS
coronavirus. J. Virol. 2020, 94, e00127-20.
(11) Shang, J.; Wan, Y.; Luo, C.; Ye, G.; Geng, Q.; Auerbach, A.; Li, F. Cell entry mechanisms
of SARS-CoV-2. Proc. Natl. Acad. Sci. U. S. A. 2020, 117, 11727-11734.
(12) Wang, Q.; Zhang, Y.; Wu, L.; Niu, S.; Song, C.; Zhang, Z.; Lu, G.; Qiao, C.; Hu, Y.; Yuen,
K. Y.; Zhou, H.; Yan, J.; Qi, J. Structural and Functional Basis of SARS-CoV-2 Entry by Using
Human ACE2. Cell 2020, 181, 894-904.e9.
(13) Wan, Y.; Shang, J.; Graham, R.; Baric, R. S.; Li, F. Receptor Recognition by the Novel
Coronavirus from Wuhan: an Analysis Based on Decade-Long Structural Studies of SARS
Coronavirus. J. Virol. 2020, 94, e00127-20.
(14) Li, F.; Li, W.; Farzan, M.; Harrison, S. C. Structure of SARS coronavirus spike receptor-
binding domain complexed with receptor. Science 2005, 309, 1864-1868.
60
(15) Lan, J.; Ge, J.; Yu, J.; Shan, S.; Zhou, H.; Fan, S.; Zhang, Q.; Shi, X.; Wang, Q.; Zhang, L.;
Wang, X. Structure of the SARS-CoV-2 spike receptor-binding domain bound to the ACE2
receptor. Nature 2020, 581, 215-220.
(16) Shang, J.; Ye, G.; Shi, K.; Wan, Y.; Luo, C.; Aihara, H.; Geng, Q.; Auerbach, A.; Li, F.
Structural basis of receptor recognition by SARS-CoV-2. Nature 2020, 581, 221-224.
(17) Kirchdoerfer, R. N.; Cottrell, C. A.; Wang, N.; Pallesen, J.; Yassine, H. M.; Turner, H. L.;
Corbett, K. S.; Graham, B. S.; McLellan, J. S.; Ward, A. B. Pre-fusion structure of a human
coronavirus spike protein. Nature 2016, 531, 118-121.
(18) Walls, A. C.; Tortorici, M. A.; Bosch, B. J.; Frenz, B.; Rottier, P. J. M.; DiMaio, F.; Rey,
F. A.; Veesler, D. Cryo-electron microscopy structure of a coronavirus spike glycoprotein trimer.
Nature 2016, 531, 114-117.
(19) Gui, M.; Song, W.; Zhou, H.; Xu, J.; Chen, S.; Xiang, Y.; Wang, X. Cryo-electron microscopy
structures of the SARS-CoV spike glycoprotein reveal a prerequisite conformational state for
receptor binding. Cell Res. 2017, 27, 119-129.
(20) Walls, A. C.; Xiong, X.; Park, Y. J.; Tortorici, M. A.; Snijder, J.; Quispe, J.; Cameroni, E.;
Gopal, R.; Dai, M.; Lanzavecchia, A.; Zambon, M.; Rey, F. A.; Corti, D.; Veesler, D. Unexpected
receptor functional mimicry elucidates activation of coronavirus fusion. Cell 2019, 176, 1026-
1039.e15.
(21) Yuan, Y.; Cao, D.; Zhang, Y.; Ma, J.; Qi, J.; Wang, Q.; Lu, G.; Wu, Y.; Yan, J.; Shi, Y.;
Zhang, X.; Gao, G. F. Cryo-EM structures of MERS-CoV and SARS-CoV spike glycoproteins
reveal the dynamic receptor binding domains. Nat. Commun. 2017, 8, 15092.
(22) Song, W.; Gui, M.; Wang, X.; Xiang, Y. Cryo-EM structure of the SARS coronavirus spike
glycoprotein in complex with its host cell receptor ACE2. PLoS Pathog. 2018, 14, e1007236.
61
(23) Kirchdoerfer, R. N.; Wang, N.; Pallesen, J.; Wrapp, D.; Turner, H. L.; Cottrell, C. A.;
Corbett, K. S.; Graham, B. S.; McLellan, J. S.; Ward, A. B. Stabilized coronavirus spikes are
resistant to conformational changes induced by receptor recognition or proteolysis. Sci. Rep. 2018,
8, 15701.
(24) Walls, A. C.; Park, Y. J.; Tortorici, M. A.; Wall, A.; McGuire, A. T.; Veesler, D. Structure,
Function, and Antigenicity of the SARS-CoV-2 Spike Glycoprotein. Cell 2020, 181, 281-292.e6.
(25) Wrapp, D.; Wang, N.; Corbett, K. S.; Goldsmith, J. A.; Hsieh, C. L.; Abiona, O.; Graham,
B. S.; McLellan, J. S. Cryo-EM structure of the 2019-nCoV spike in the prefusion conformation.
Science 2020, 367, 1260-1263.
(26) Cai, Y.; Zhang, J.; Xiao, T.; Peng, H.; Sterling, S. M.; Walsh, R. M., Jr.; Rawson, S.;
Rits-Volloch, S.; Chen, B. Distinct conformational states of SARS-CoV-2 spike protein. Science
2020, 369, 1586-1592.
(27) Hsieh, C. L.; Goldsmith, J. A.; Schaub, J. M.; DiVenere, A. M.; Kuo, H. C.; Javanmardi,
K.; Le, K. C.; Wrapp, D.; Lee, A. G.; Liu, Y.; Chou, C. W.; Byrne, P. O.; Hjorth, C. K.;
Johnson, N. V.; Ludes-Meyers, J.; Nguyen, A. W.; Park, J.; Wang, N.; Amengor, D.; Lavinder,
J. J.; Ippolito, G. C.; Maynard, J. A.; Finkelstein, I. J.; McLellan, J. S. Structure-based design of
prefusion-stabilized SARS-CoV-2 spikes. Science 2020, 369, 1501-1505.
(28) Henderson, R.; Edwards, R. J.; Mansouri, K.; Janowska, K.; Stalls, V.; Gobeil, S. M. C.;
Kopp, M.; Li, D.; Parks, R.; Hsu, A. L.; Borgnia, M. J.; Haynes, B. F.; Acharya, P. Controlling
the SARS-CoV-2 spike glycoprotein conformation. Nat. Struct. Mol. Biol. 2020,
27, 925-933.
62
(29) McCallum, M.; Walls, A. C.; Bowen, J. E.; Corti, D.; Veesler, D. Structure-guided covalent
stabilization of coronavirus spike glycoprotein trimers in the closed conformation. Nat. Struct.
Mol. Biol. 2020, 27, 942-949.
(30) Xiong, X.; Qu, K.; Ciazynska, K. A.; Hosmillo, M.; Carter, A. P.; Ebrahimi, S.; Ke, Z.;
Scheres, S. H. W.; Bergamaschi, L.; Grice, G. L.; Zhang, Y.; Nathan, J. A.; Baker, S.; James,
L. C.; Baxendale, H. E.; Goodfellow, I.; Doffinger, R.; Briggs, J. A. G. A thermostable, closed
SARS-CoV-2 spike protein trimer. Nat. Struct. Mol. Biol. 2020, 27, 934-941.
(31) Yurkovetskiy, L.; Wang, X.; Pascal, K. E.; Tomkins-Tinch, C.; Nyalile, T. P.; Wang, Y.;
Baum, A.; Diehl, W. E.; Dauphin, A.; Carbone, C.; Veinotte, K.; Egri, S. B.; Schaffner, S. F.;
Lemieux, J. E.; Munro, J. B.; Rafique, A.; Barve, A.; Sabeti, P. C.; Kyratsous, C. A.; Dudkina,
N. V.; Shen, K.; Luban, J. Structural and functional analysis of the D614G SARS-CoV-2 spike
protein variant. Cell 2020, 183, 739-751.e8.
(32) Turoňová, B.; Sikora, M.; Schürmann, C.; Hagen, W. J. H.; Welsch, S.; Blanc, F. E. C.;
von Bülow, S.; Gecht, M.; Bagola, K.; Hörner, C.; van Zandbergen, G.; Landry, J.; de Azevedo,
N. T. D.; Mosalaganti, S.; Schwarz, A.; Covino, R.; Mühlebach, M. D.; Hummer, G.; Krijnse
Locker, J.; Beck, M. In situ structural analysis of SARS-CoV-2 spike reveals flexibility mediated
by three hinges. Science 2020, 370, 203-208.
(33) Lu, M.; Uchil, P. D.; Li, W.; Zheng, D.; Terry, D. S.; Gorman, J.; Shi, W.; Zhang, B.;
Zhou, T.; Ding, S.; Gasser, R.; Prevost, J.; Beaudoin-Bussieres, G.; Anand, S. P.; Laumaea,
A.; Grover, J. R.; Lihong, L.; Ho, D. D.; Mascola, J.; Finzi, A.; Kwong, P. D.; Blanchard, S.
C.; Mothes, W. Real-time conformational dynamics of SARS-CoV-2 spikes on virus particles.
bioRxiv 2020, doi: 10.1101/2020.09.10.286948.
63
(34) Benton, D. J.; Wrobel, A. G.; Xu, P.; Roustan, C.; Martin, S. R.; Rosenthal, P. B.; Skehel,
J. J.; Gamblin, S. J. Receptor binding and priming of the spike protein of SARS-CoV-2 for
membrane fusion. Nature 2020, doi: 10.1038/s41586-020-2772-0.
(35) He, Y.; Lu, H.; Siddiqui, P.; Zhou, Y.; Jiang, S. Receptor-binding domain of severe acute
respiratory syndrome coronavirus spike protein contains multiple conformation-dependent
epitopes that induce highly potent neutralizing antibodies. J. Immunol. 2005, 174, 4908-4915.
(36) He, Y.; Li, J.; Li, W.; Lustigman, S.; Farzan, M.; Jiang, S. Cross-neutralization of human
and palm civet severe acute respiratory syndrome coronaviruses by antibodies targeting the
receptor-binding domain of spike protein. J. Immunol. 2006, 176, 6085-92.
(37) Chen, X.; Li, R.; Pan, Z.; Qian, C.; Yang, Y.; You, R.; Zhao, J.; Liu, P.; Gao, L.; Li, Z.;
Huang, Q.; Xu, L.; Tang, J.; Tian, Q.; Yao, W.; Hu, L.; Yan, X.; Zhou, X.; Wu, Y.; Deng,
K.; Zhang, Z.; Qian, Z.; Chen, Y.; Ye, L. Human monoclonal antibodies block the binding of
SARS-CoV-2 spike protein to angiotensin converting enzyme 2 receptor. Cell Mol. Immunol.
2020, 17, 647-649.
(38) Wrapp, D.; De Vlieger, D.; Corbett, K. S.; Torres, G. M.; Wang, N.; Van Breedam, W.;
Roose, K.; van Schie, L.; Hoffmann, M.; Pöhlmann, S.; Graham, B. S.; Callewaert, N.;
Schepens, B.; Saelens, X.; McLellan, J. S. Structural Basis for Potent Neutralization of
Betacoronaviruses by Single-Domain Camelid Antibodies. Cell 2020, 181, 1004-1015.e15.
(39) Yu, F.; Xiang, R.; Deng, X.; Wang, L.; Yu, Z.; Tian, S.; Liang, R.; Li, Y.; Ying, T.;
Jiang, S. Receptor-binding domain-specific human neutralizing monoclonal antibodies against
SARS-CoV and SARS-CoV-2. Signal Transduct. Target. Ther. 2020, 5, 212.
64
(40) Wang, C.; Li, W.; Drabek, D.; Okba, N. M. A.; van Haperen, R.; Osterhaus, A.; van
Kuppeveld, F. J. M.; Haagmans, B. L.; Grosveld, F.; Bosch, B. J. A human monoclonal antibody
blocking SARS-CoV-2 infection. Nat. Commun. 2020, 11, 2251.
(41) Hwang, W. C.; Lin, Y.; Santelli, E.; Sui, J.; Jaroszewski, L.; Stec, B.; Farzan, M.;
Marasco, W. A.; Liddington, R. C. Structural basis of neutralization by a human anti-severe acute
respiratory syndrome spike protein antibody, 80R. J. Biol. Chem. 2006, 281, 34610-34616.
(42) Yuan, M.; Wu, N. C.; Zhu, X.; Lee, C. D.; So, R. T. Y.; Lv, H.; Mok, C. K. P.; Wilson, I.
A. A highly conserved cryptic epitope in the receptor binding domains of SARS-CoV-2 and
SARS-CoV. Science 2020, 368, 630-633.
(43) Huo, J.; Zhao, Y.; Ren, J.; Zhou, D.; Duyvesteyn, H. M. E.; Ginn, H. M.; Carrique, L.;
Malinauskas, T.; Ruza, R. R.; Shah, P. N. M.; Tan, T. K.; Rijal, P.; Coombes, N.; Bewley, K.
R.; Tree, J. A.; Radecke, J.; Paterson, N. G.; Supasa, P.; Mongkolsapaya, J.; Screaton, G. R.;
Carroll, M.; Townsend, A.; Fry, E. E.; Owens, R. J.; Stuart, D. I. Neutralization of SARS-CoV-
2 by destruction of the prefusion spike. Cell Host Microbe 2020, 28, 445-454.e6.
(44) Brouwer, P. J. M.; Caniels, T. G.; van der Straten, K.; Snitselaar, J. L.; Aldon, Y.; Bangaru,
S.; Torres, J. L.; Okba, N. M. A.; Claireaux, M.; Kerster, G.; Bentlage, A. E. H.; van Haaren,
M. M.; Guerra, D.; Burger, J. A.; Schermer, E. E.; Verheul, K. D.; van der Velde, N.; van der
Kooi, A.; van Schooten, J.; van Breemen, M. J.; Bijl, T. P. L.; Sliepen, K.; Aartse, A.; Derking,
R.; Bontjer, I.; Kootstra, N. A.; Wiersinga, W. J.; Vidarsson, G.; Haagmans, B. L.; Ward, A.
B.; de Bree, G. J.; Sanders, R. W.; van Gils, M. J. Potent neutralizing antibodies from COVID-
19 patients define multiple targets of vulnerability. Science 2020, 369, 643-650.
(45) Lv, Z.; Deng, Y. Q.; Ye, Q.; Cao, L.; Sun, C. Y.; Fan, C.; Huang, W.; Sun, S.; Sun, Y.;
Zhu, L.; Chen, Q.; Wang, N.; Nie, J.; Cui, Z.; Zhu, D.; Shaw, N.; Li, X. F.; Li, Q.; Xie, L.;
65
Wang, Y.; Rao, Z.; Qin, C. F.; Wang, X. Structural basis for neutralization of SARS-CoV-2 and
SARS-CoV by a potent therapeutic antibody. Science 2020, 369, 1505-1509.
(46) Pinto, D.; Park, Y. J.; Beltramello, M.; Walls, A. C.; Tortorici, M. A.; Bianchi, S.; Jaconi,
S.; Culap, K.; Zatta, F.; De Marco, A.; Peter, A.; Guarino, B.; Spreafico, R.; Cameroni, E.;
Case, J. B.; Chen, R. E.; Havenar-Daughton, C.; Snell, G.; Telenti, A.; Virgin, H. W.;
Lanzavecchia, A.; Diamond, M. S.; Fink, K.; Veesler, D.; Corti, D. Cross-neutralization of
SARS-CoV-2 by a human monoclonal SARS-CoV antibody. Nature 2020, 583, 290-295.
(47) Tortorici, M. A.; Beltramello, M.; Lempp, F. A.; Pinto, D.; Dang, H. V.; Rosen, L. E.;
McCallum, M.; Bowen, J.; Minola, A.; Jaconi, S.; Zatta, F.; De Marco, A.; Guarino, B.;
Bianchi, S.; Lauron, E. J.; Tucker, H.; Zhou, J.; Peter, A.; Havenar-Daughton, C.;
Wojcechowskyj, J. A.; Case, J. B.; Chen, R. E.; Kaiser, H.; Montiel-Ruiz, M.; Meury, M.;
Czudnochowski, N.; Spreafico, R.; Dillen, J.; Ng, C.; Sprugasci, N.; Culap, K.; Benigni, F.;
Abdelnabi, R.; Foo, S. C.; Schmid, M. A.; Cameroni, E.; Riva, A.; Gabrieli, A.; Galli, M.;
Pizzuto, M. S.; Neyts, J.; Diamond, M. S.; Virgin, H. W.; Snell, G.; Corti, D.; Fink, K.; Veesler,
D. Ultrapotent human antibodies protect against SARS-CoV-2 challenge via multiple
mechanisms. Science 2020, 370, 950-957.
(48) Chi, X.; Yan, R.; Zhang, J.; Zhang, G.; Zhang, Y.; Hao, M.; Zhang, Z.; Fan, P.; Dong,
Y.; Yang, Y.; Chen, Z.; Guo, Y.; Li, Y.; Song, X.; Chen, Y.; Xia, L.; Fu, L.; Hou, L.; Xu,
J.; Yu, C.; Li, J.; Zhou, Q.; Chen, W., A neutralizing human antibody binds to the N-terminal
domain of the Spike protein of SARS-CoV-2. Science 2020, 369, 650-655.
(49) Schoof, M.; Faust, B.; Saunders, R. A.; Sangwan, S.; Rezelj, V.; Hoppe, N.; Boone, M.;
Billesbølle, C. B.; Puchades, C.; Azumaya, C. M.; Kratochvil, H. T.; Zimanyi, M.; Deshpande,
I.; Liang, J.; Dickinson, S.; Nguyen, H. C.; Chio, C. M.; Merz, G. E.; Thompson, M. C.;
66
Diwanji, D.; Schaefer, K.; Anand, A. A.; Dobzinski, N.; Zha, B. S.; Simoneau, C. R.; Leon,
K.; White, K. M.; Chio, U. S.; Gupta, M.; Jin, M.; Li, F.; Liu, Y.; Zhang, K.; Bulkley, D.;
Sun, M.; Smith, A. M.; Rizo, A. N.; Moss, F.; Brilot, A. F.; Pourmal, S.; Trenker, R.; Pospiech,
T.; Gupta, S.; Barsi-Rhyne, B.; Belyy, V.; Barile-Hill, A. W.; Nock, S.; Krogan, N. J.; Ralston,
C. Y.; Swaney, D. L.; García-Sastre, A.; Ott, M.; Vignuzzi, M.; Walter, P.; Manglik, A. An
ultrapotent synthetic nanobody neutralizes SARS-CoV-2 by stabilizing inactive Spike. Science
2020, 370, 1473-1479.
(50) Woo, H.; Park, S. J.; Choi, Y. K.; Park, T.; Tanveer, M.; Cao, Y.; Kern, N. R.; Lee, J.;
Yeom, M. S.; Croll, T. I.; Seok, C.; Im, W. Developing a fully glycosylated full-length SARS-
CoV-2 spike protein model in a viral membrane. J. Phys. Chem. B 2020, 124, 7128-7137.
(51) Casalino, L.; Gaieb, Z.; Goldsmith, J.A.; Hjorth, C.K.; Dommer, A. C.; Harbison, A. M.;
Fogarty, C. A.; Barros, E. P.; Taylor, B. C.; McLellan, J.S.; Fadda, E.; Amaro, R. E. Beyond
shielding: The roles of glycans in the SARS-CoV-2 spike potein. ACS Cent. Sci. 2020, 6, 10,
1722-1734.
(52) Kalathiya, U.; Padariya, M.; Mayordomo, M.; Lisowska, M.; Nicholson, J.; Singh, A.;
Baginski, M.; Fahraeus, R.; Carragher, N.; Ball, K.; Haas, J.; Daniels, A.; Hupp, T. R.; Alfaro,
J. A. Highly conserved homotrimer cavity formed by the SARS-CoV-2 spike glycoprotein: A
novel binding site. J. Clin. Med. 2020, 9, 1473.
(53) Brielle, E. S.; Schneidman-Duhovny, D.; Linial, M. The SARS-CoV-2 exerts a distinctive
strategy for interacting with the ACE2 human receptor. Viruses 2020, 12,497.
(54) Spinello, A.; Saltalamacchia, A.; Magistrato, A. Is the Rigidity of SARS-CoV-2 spike
receptor-binding motif the hallmark for Its enhanced infectivity? Insights from all-atom
simulations. J. Phys. Chem. Lett. 2020, 11, 4785-4790.
67
(55) Chen, Y.; Guo, Y.; Pan, Y.; Zhao, Z. J. Structure analysis of the receptor binding of 2019-
nCoV. Biochem. Biophys. Res. Commun. 2020, 525, 135-40.
(56) Othman, H.; Bouslama, Z.; Brandenburg, J. T.; da Rocha, J.; Hamdi, Y.; Ghedira, K.; Srairi-
Abid, N.; Hazelhurst, S. Interaction of the spike protein RBD from SARS-CoV-2 with ACE2:
Similarity with SARS-CoV, hot-spot analysis and effect of the receptor polymorphism. Biochem.
Biophys. Res. Commun. 2020, 527, 702-708.
(57) Ghorbani, M.; Brooks, B. R.; Klauda, J. B. Critical Sequence Hot-spots for Binding of nCOV-
2019
to
ACE2
as
Evaluated
by
Molecular
Simulations.
bioRxiv
2020,
doi:
10.1101/2020.06.27.175448.
(58) Veeramachaneni, G. K.; Thunuguntla, V.; Bobbillapati, J.; Bondili, J. S. Structural and
simulation analysis of hotspot residues interactions of SARS-CoV 2 with human ACE2 receptor.
J. Biomol. Struct. Dyn. 2020, 1-11.
(59) Wang, Y.; Liu, M.; Gao, J. Enhanced receptor binding of SARS-CoV-2 through networks of
hydrogen-bonding and hydrophobic interactions. Proc. Natl. Acad. Sci. U. S. A. 2020, 117, 13967-
13974.
(60) Laurini, E.; Marson, D.; Aulic, S.; Fermeglia, M.; Pricl, S. Computational alanine scanning
and structural analysis of the SARS-CoV-2 Spike protein/angiotensin-converting enzyme 2
complex. ACS Nano 2020, 14, 11821-11830.
(61) Ali, A.; Vijayan, R. Dynamics of the ACE2-SARS-CoV-2/SARS-CoV spike protein interface
reveal unique mechanisms. Sci. Rep. 2020, 10, 14214.
(62) Verkhivker, G.M. Coevolution, dynamics and allostery conspire in shaping cooperative
binding and signal transmission of the SARS-CoV-2 spike protein with human angiotensin-
converting enzyme 2. Int. J. Mol. Sci. 2020, 21, 8268.
68
(63) Di Paola, L.; Hadi-Alijanvand, H.; Song, X.; Hu, G.; Giuliani, A. The discovery of a putative
allosteric site in the SARS-CoV-2 spike protein using an integrated structural/dynamic approach.
J. Proteome Res. 2020, 19, 4576-4586.
(64) Verkhivker, G.M. Molecular simulations and network modeling reveal an allosteric signaling
in the SARS-CoV-2 spike proteins. J. Proteome Res. 2020, 19, 4587-4608.
(65) Rozewicki, J.; Li, S.; Amada, K. M.; Standley, D. M.; Katoh, K. MAFFT-DASH: integrated
protein sequence and structural alignment. Nucleic Acids Res. 2019, 47, W5-W10.
(66) Suzek, B. E.; Huang, H.; McGarvey, P.; Mazumder, R.; Wu, C. H. UniRef: comprehensive
and non-redundant UniProt reference clusters. Bioinformatics 2007, 23, 1282-1288.
(67) Suzek, B. E.; Wang, Y.; Huang, H.; McGarvey, P. B.; Wu, C. H. UniRef clusters: a
comprehensive and scalable alternative for improving sequence similarity searches.
Bioinformatics 2015, 31, 926-932.
(68) El-Gebali, S.; Mistry, J.; Bateman, A.; Eddy, S. R.; Luciani, A.; Potter, S. C.; Qureshi, M.;
Richardson, L. J.; Salazar, G. A.; Smart, A.; Sonnhammer, E. L. L.; Hirsh, L.; Paladin, L.;
Piovesan, D.; Tosatto, S. C. E.; Finn, R. D. The Pfam protein families database in 2019. Nucleic
Acids Res. 2019, 47, D427-D432.
(69) Mistry, J.; Chuguransky, S.; Williams, L.; Qureshi, M.; Salazar, G. A.; Sonnhammer, E.
L. L.; Tosatto, S. C. E.; Paladin, L.; Raj, S.; Richardson, L. J.; Finn, R. D.; Bateman, A., Pfam:
The protein families database in 2021. Nucleic Acids Res 2020. doi: 10.1093/nar/gkaa913.
(70) Wu, C. H.; Apweiler, R.; Bairoch, A.; Natale, D. A.; Barker, W. C.; Boeckmann, B.; Ferro,
S.; Gasteiger, E.; Huang, H.; Lopez, R.; Magrane, M.; Martin, M. J.; Mazumder, R.; O'Donovan,
C.; Redaschi, N.; Suzek, B. The Universal Protein Resource (UniProt): an expanding universe of
protein information. Nucleic Acids Res. 2006, 34, D187-D191.
69
(71) Marino Buslje, C.; Teppa, E.; Di Domenico, T.; Delfino, J. M.; Nielsen, M. Networks of high
mutual information define the structural proximity of catalytic sites: implications for catalytic
residue identification. PLoS Comput. Biol. 2010, 6, e1000978.
(72) Simonetti, F. L.; Teppa, E.; Chernomoretz, A.; Nielsen, M.; Marino Buslje, C. MISTIC:
Mutual information server to infer coevolution. Nucleic Acids Res 2013, 41, W8-W14.
(73) Colell, E. A.; Iserte, J. A.; Simonetti, F. L.; Marino-Buslje, C., MISTIC2: comprehensive
server to study coevolution in protein families. Nucleic Acids Res 2018, 46, W323-W328.
(74) Morcos, F.; Pagnani, A.; Lunt, B.; Bertolino, A.; Marks, D. S.; Sander, C.; Zecchina, R.;
Onuchic, J. N.; Hwa, T.; Weigt, M. Direct-coupling analysis of residue coevolution captures
native contacts across many protein families. Proc. Natl. Acad. Sci. U. S. A. 2011, 108, E1293-
E1301.
(75) Morcos, F.; Hwa, T.; Onuchic, J. N.; Weigt, M. Direct coupling analysis for protein contact
prediction. Methods Mol. Biol. 2014, 1137, 55-70.
(76) Kaján, L.; Hopf, T. A.; Kalaš, M.; Marks, D. S.; Rost, B. FreeContact: fast and free software
for protein contact prediction from residue co-evolution. BMC Bioinformatics 2014, 15, 85.
(77) Seemayer, S.; Gruber, M.; Söding, J. CCMpred--fast and precise prediction of protein
residue-residue contacts from correlated mutations. Bioinformatics 2014, 30 , 3128-3130.
(78) Ekeberg, M.; Lövkvist, C.; Lan, Y.; Weigt, M.; Aurell, E. Improved contact prediction in
proteins: using pseudolikelihoods to infer Potts models. Phys. Rev. E. Stat. Nonlin. Soft Matter
Phys. 2013, 87, 012707.
(79) Baldassi, C.; Zamparo, M.; Feinauer, C.; Procaccini, A.; Zecchina, R.; Weigt, M.; Pagnani,
A. Fast and accurate multivariate Gaussian modeling of protein families: predicting residue
contacts and protein-interaction partners. PLoS One 2014, 9, e92721.
70
(80) Gueudre, T.; Baldassi, C.; Zamparo, M.; Weigt, M.; Pagnani, A., Simultaneous identification
of specifically interacting paralogs and interprotein contacts by direct coupling analysis. Proc.
Natl. Acad. Sci. U. S. A. 2016, 113, 12186-12191.
(81) Kolinski, A. Protein modeling and structure prediction with a reduced representation. Acta
Biochim. Pol. 2004, 51, 349-371.
(82) Kmiecik, S.; Gront, D.; Kolinski, M.; Wieteska, L.; Dawid, A.E.; Kolinski, A. Coarse-grained
protein models and their applications. Chem. Rev. 2016, 116, 7898-7936.
(83) Kmiecik, S.; Kouza, M.; Badaczewska-Dawid, A.E.; Kloczkowski, A.; Kolinski, A. Modeling
of protein structural flexibility and large-scale dynamics: Coarse-grained simulations and elastic
network models. Int. J. Mol. Sci. 2018, 19, e3496.
(84) Ciemny, M.P.; Badaczewska-Dawid, A.E.; Pikuzinska, M.; Kolinski, A.; Kmiecik, S.
Modeling of disordered protein structures using monte carlo simulations and knowledge-based
statistical force fields. Int. J. Mol. Sci. 2019, 20, e606.
(85) Kurcinski, M.; Oleniecki, T.; Ciemny, M.P.; Kuriata, A.; Kolinski, A.; Kmiecik, S. CABS-
flex standalone: A simulation environment for fast modeling of protein flexibility. Bioinformatics
2019, 35, 694-695.
(86) Berman, H.M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T.N.; Weissig, H.; Shindyalov,
I.N.; Bourne, P.E. The Protein Data Bank. Nucleic Acids Res. 2000, 28, 235-242.
(87) Rose, P. W.; Prlic, A.; Altunkaya, A.; Bi, C.; Bradley, A. R.; Christie, C. H.; Costanzo, L. D.;
Duarte, J. M.; Dutta, S.; Feng, Z.; Green, R. K.; Goodsell, D. S.; Hudson, B.; Kalro, T.; Lowe, R.;
Peisach, E.; Randle, C.; Rose, A. S.; Shao, C.; Tao, Y. P.; Valasatava, Y.; Voigt, M.; Westbrook,
J. D.; Woo, J.; Yang, H.; Young, J. Y.; Zardecki, C.; Berman, H. M.; Burley, S. K. The RCSB
71
protein data bank: integrative view of protein, gene and 3D structural information. Nucleic Acids
Res. 2017, 45, D271-D281.
(88) Hooft, R. W.; Sander, C.; Vriend, G. Positioning hydrogen atoms by optimizing hydrogen-
bond networks in protein structures. Proteins 1996, 26, 363-376.
(89) Hekkelman, M. L.; Te Beek, T. A.; Pettifer, S. R.; Thorne, D.; Attwood, T. K.; Vriend, G.
WIWS: A protein structure bioinformatics web service collection. Nucleic Acids Res. 2010, 38,
W719-W723.
(90) Fiser, A.; Sali, A. ModLoop: Automated modeling of loops in protein structures.
Bioinformatics 2003, 19, 2500-25001.
(91) Fernandez-Fuentes, N.; Zhai, J.; Fiser, A. ArchPRED: A template based loop structure
prediction server. Nucleic Acids Res. 2006, 34, W173-W176.
(92) Rotkiewicz, P.; Skolnick, J. Fast procedure for reconstruction of full-atom protein models
from reduced representations. J. Comput. Chem. 2008, 29, 1460-1465.
(93) Lombardi, L. E.; Marti, M. A.; Capece, L. CG2AA: backmapping protein coarse-grained
structures. Bioinformatics 2016, 32, 1235-1237.
(94) Bhattacharya, D.; Nowotny, J.; Cao, R.; Cheng, J. 3Drefine: an interactive web server for
efficient protein structure refinement. Nucleic Acids Res. 2016, 44, W406-W409.
(95) Dehouck, Y.; Kwasigroch, J. M.; Rooman, M.; Gilis, D. BeAtMuSiC: Prediction of changes
in protein-protein binding affinity on mutations. Nucleic Acids Res. 2013, 41, W333-W339.
(96) Dehouck, Y.; Gilis, D.; Rooman, M. A new generation of statistical potentials for proteins.
Biophys. J. 2006, 90, 4010-4017.
72
(97) Dehouck, Y.; Grosfils, A.; Folch, B.; Gilis, D.; Bogaerts, P.; Rooman, M. Fast and accurate
predictions of protein stability changes upon mutations using statistical potentials and neural
networks: PoPMuSiC-2.0. Bioinformatics 2009, 25, 2537-2543.
(98) Tasdighian, S.; Di Paola, L.; De Ruvo, M.; Paci, P.; Santoni, D.; Palumbo, P.; Mei, G.; Di
Venere, A.; Giuliani, A. Modules identification in protein Structures: The topological and
geometrical solutions. J. Chem. Inf. Model. 2014, 54, 159-168.
(99) De Ruvo, M.; Giuliani, A.; Paci, P.; Santoni, D.; Di Paola, L. Shedding light on protein-ligand
binding by graph theory: The topological nature of allostery. Biophys. Chem. 2012, 165-166, 21–
29.
(100) Di Paola, L.; Paci, P.; Santoni, D.; De Ruvo, M.; Giuliani, A. Proteins as sponges: A
statistical journey along protein structure organization principles. J. Chem. Inf. Model. 2012, 52,
474-482.
(101) Cumbo, F.; Paci, P.; Santoni, D.; Di Paola, L.; Giuliani, A. GIANT: A Cytoscape plugin for
modular tetworks. PLoS One 2014, 9, e105001.
(102) Brinda, K.V.; Vishveshwara, S. A network representation of protein structures: Implications
for protein stability. Biophys. J. 2005, 89, 4159-4170.
(103) Vijayabaskar, M.S.; Vishveshwara, S. Interaction energy based protein structure networks.
Biophys. J. 2010, 99, 3704-3715.
(104) Sethi, A.; Eargle, J.; Black, A.A.; Luthey-Schulten, Z. Dynamical networks in tRNA:protein
complexes. Proc. Natl. Acad. Sci. U.S.A. 2009, 106, 6620-6625.
(105) Stetz, G.; Verkhivker, G. M. Computational analysis of residue interaction networks and
coevolutionary relationships in the Hsp70 chaperones: A community-hopping model of allosteric
regulation and communication. Plos Comput. Biol. 2017, 13, e1005299.
73
(106) Czemeres, J.; Buse, K.; Verkhivker, G. M. Atomistic simulations and network-based
modeling of the Hsp90-Cdc37 chaperone binding with Cdk4 client protein: A mechanism of
chaperoning kinase clients by exploiting weak spots of intrinsically dynamic kinase domains. PLoS
One 2017, 12, e0190267.
(107) Stetz, G.; Verkhivker, G. M. Dancing through life: Molecular dynamics simulations and
network-centric modeling of allosteric mechanisms in Hsp70 and Hsp110 chaperone proteins.
PLoS One 2015, 10, e0143752.61.
(108) Floyd, R.W. Algorithm 97: Shortest path. Commun. A.C.M. 1962, 5, 345.
(109) Hagberg, A.A.; Schult, D.A.; Swart, P.J. Exploring network structure, dynamics, and
function using NetworkX, in : G. Varoquaux, T. Vaught, J. Millman (Eds.), Proceedings of the 7th
Python in Science Conference (SciPy2008), Pasadena, 2008, pp. 11-15.
(110) Atilgan, C.; Atilgan, A. R., Perturbation-response scanning reveals ligand entry-exit
mechanisms of ferric binding protein. PLoS Comput. Biol. 2009, 5, e1000544.
(111) Atilgan, C.; Gerek, Z. N.; Ozkan, S. B.; Atilgan, A. R., Manipulation of conformational
change in proteins by single-residue perturbations. Biophys. J. 2010, 99, 933-943.
(112) General, I. J.; Liu, Y.; Blackburn, M. E.; Mao, W.; Gierasch, L. M.; Bahar, I., ATPase
subdomain IA is a mediator of interdomain allostery in Hsp70 molecular chaperones. PLoS
Comput. Biol. 2014, 10, e1003624.
(113) Dutta, A.; Krieger, J.; Lee, J. Y.; Garcia-Nafria, J.; Greger, I. H.; Bahar, I., Cooperative
Dynamics of Intact AMPA and NMDA Glutamate Receptors: Similarities and Subfamily-Specific
Differences. Structure 2015, 23, 1692-1704.
(114) Penkler, D.; Sensoy, O.; Atilgan, C.; Tastan Bishop, O., Perturbation-Response Scanning
Reveals Key Residues for Allosteric Control in Hsp70. J. Chem. Inf. Model. 2017, 57, 1359-1374.
74
(115) Penkler, D. L.; Atilgan, C.; Tastan Bishop, O., Allosteric Modulation of Human Hsp90alpha
Conformational Dynamics. J. Chem. Inf. Model. 2018, 58, 383-404.
(116) Stetz, G.; Tse, A.; Verkhivker, G. M., Dissecting Structure-Encoded Determinants of
Allosteric Cross-Talk between Post-Translational Modification Sites in the Hsp90 Chaperones.
Sci. Rep. 2018, 8, 19.
(117) Jalalypour, F.; Sensoy, O.; Atilgan, C., Perturb-Scan-Pull: A Novel Method Facilitating
Conformational Transitions in Proteins. J. Chem. Theory Comput. 2020, 16, 3825-3841.
(118) Jaimes, J. A.; André, N. M.; Chappie, J. S.; Millet, J. K.; Whittaker, G. R. Phylogenetic
Analysis and Structural Modeling of SARS-CoV-2 Spike Protein Reveals an Evolutionary Distinct
and Proteolytically Sensitive Activation Loop. J. Mol. Biol. 2020, 432, 3309-3325.
(119) Shah, M.; Ahmad, B.; Choi, S.; Woo, H.G. Sequence variation of SARS-CoV-2 spike protein
may facilitate stronger interaction with ACE2 promoting high infectivity. Research Square 2020.
DOI: 10.21203/rs.3.rs-16932/v1.
(120) Armijos‐Jaramillo, V.; Yeager, J.; Muslin, C.; Perez‐Castillo, Y. SARS‐CoV‐2, an
evolutionary perspective of interaction with human ACE2 reveals undiscovered amino acids
necessary for complex stability. Evol. Appl. 2020, 00: 1-11.
(121) Singh, J.; Dhindsa, R. S.; Misra, V.; Singh, B. SARS-CoV2 infectivity is potentially
modulated by host redox status. Comput. Struct. Biotechnol. J. 2020, 18, 3705-3711.
(122) Petit, C. M.; Chouljenko, V. N.; Iyer, A.; Colgrove, R.; Farzan, M.; Knipe, D. M.;
Kousoulas, K. G. Palmitoylation of the cysteine-rich endodomain of the SARS-coronavirus spike
glycoprotein is important for spike-mediated cell fusion. Virology 2007, 360, 264-274.
75
(123) Madu, I. G.; Belouzard, S.; Whittaker, G. R., SARS-coronavirus spike S2 domain flanked
by cysteine residues C822 and C833 is important for activation of membrane fusion. Virology
2009, 393, 265-71.
(124) Madu, I. G.; Roth, S. L.; Belouzard, S.; Whittaker, G. R., Characterization of a highly
conserved domain within the severe acute respiratory syndrome coronavirus spike protein S2
domain with characteristics of a viral fusion peptide. J. Virol. 2009, 83, 7411-7421.
(125) Walls, A. C.; Tortorici, M. A.; Snijder, J.; Xiong, X.; Bosch, B. J.; Rey, F. A.; Veesler,
D. Tectonic conformational changes of a coronavirus spike glycoprotein promote membrane
fusion. Proc. Natl. Acad. Sci. U. S. A. 2017, 114, 11157-11162.
(126) Fan, X.; Cao, D.; Kong, L.; Zhang, X., Cryo-EM analysis of the post-fusion structure of
the SARS-CoV spike glycoprotein. Nat. Commun. 2020, 11, 3618.
(127) Mercurio, I.; Tragni, V.; Busto, F.; De Grassi, A.; Pierri, C. L. Protein structure analysis
of the interactions between SARS-CoV-2 spike protein and the human ACE2 receptor: from
conformational changes to novel neutralizing antibodies. Cell. Mol. Life Sci. 2020, 1-22.
(128) Hati, S.; Bhattacharyya, S. Impact of thiol-disulfide balance on the binding of Covid-19
spike protein with angiotensin-converting enzyme 2 receptor. ACS Omega 2020, 5, 16292-16298.
(129) Lee, B. C.; Park, K.; Kim, D. Analysis of the residue-residue coevolution network and the
functionally important residues in proteins. Proteins 2008, 72, 863-872.
(130) Chakrabarti, S.; Panchenko, A. R. Coevolution in defining the functional specificity.
Proteins 2009, 75, 231-240.
(131) Chakrabarti, S.; Panchenko, A. R. Structural and functional roles of coevolved sites in
proteins. PLoS One 2010, 5, e8591.
76
(132) Teppa, E.; Wilkins, A. D.; Nielsen, M.; Buslje, C. M. Disentangling evolutionary signals:
conservation, specificity determining positions and coevolution. Implication for catalytic residue
prediction. BMC Bioinformatics 2012, 13, 235.
(133) Jeon, J.; Nam, H. J.; Choi, Y. S.; Yang, J. S.; Hwang, J.; Kim, S. Molecular evolution of
protein conformational changes revealed by a network of evolutionarily coupled residues. Mol.
Biol. Evol. 2011, 28, 2675-85.
(134) Hoffmann, M.; Kleine-Weber, H.; Pöhlmann, S. A multibasic cleavage site in the spike
Pprotein of SARS-CoV-2 is essential for infection of human lung cells. Mol. Cell 2020, 78 779-
784.e5.
(135) Örd, M.; Faustova, I.; Loog, M., The sequence at Spike S1/S2 site enables cleavage by furin
and phospho-regulation in SARS-CoV2 but not in SARS-CoV1 or MERS-Co. Sci. Rep. 2020, 10,
16944.
(136) Belouzard, S.; Chu, V. C.; Whittaker, G. R., Activation of the SARS coronavirus spike
protein via sequential proteolytic cleavage at two distinct sites. Proc. Natl. Acad. Sci. U.S.A. 2009,
106 , 5871-5876.
(137) Jeong, C. S.; Kim, D. Coevolved residues and the functional association for intrinsically
disordered proteins. Pac. Symp. Biocomput. 2012, 140-51.
(138) Teppa, E.; Zea, D. J.; Marino-Buslje, C. Protein-protein interactions leave evolutionary
footprints: High molecular coevolution at the core of interfaces. Protein Sci. 2017, 26, 2438-2444.
(139) Anishchenko, I.; Ovchinnikov, S.; Kamisetty, H.; Baker, D. Origins of coevolution between
residues distant in protein 3D structures. Proc. Natl. Acad. Sci. U.S.A. 2017, 114, 9122-9127.
(140) Pietal, M. J.; Tuszynska, I.; Bujnicki, J. M., PROTMAP2D: visualization, comparison and
analysis of 2D maps of protein structure. Bioinformatics 2007, 23, 1429-1430.
77
(141) Chen, J. E.; Huang, C. C.; Ferrin, T. E. RRDistMaps: a UCSF Chimera tool for viewing
and comparing protein distance maps. Bioinformatics 2015, 31, 1484-1486.
(142) Weinkam, P.; Pons, J.; Sali, A. Structure-based model of allostery predicts coupling between
distant sites. Proc. Natl. Acad. Sci. U.S.A. 2012, 109, 4875-4880.
(143) Li, Q.; Wu, J.; Nie, J.; Zhang, L.; Hao, H.; Liu, S.; Zhao, C.; Zhang, Q.; Liu, H.; Nie,
L.; Qin, H.; Wang, M.; Lu, Q.; Li, X.; Sun, Q.; Liu, J.; Huang, W.; Wang, Y. The impact of
mutations in SARS-CoV-2 spike on viral infectivity and antigenicity. Cell 2020, 182, 1284-
1294.e9.
(144) Korber, B.; Fischer, W. M.; Gnanakaran, S.; Yoon, H.; Theiler, J.; Abfalterer, W.;
Hengartner, N.; Giorgi, E. E.; Bhattacharya, T.; Foley, B.; Hastie, K. M.; Parker, M. D.;
Partridge, D. G.; Evans, C. M.; Freeman, T. M.; de Silva, T. I.; McDanal, C.; Perez, L. G.;
Tang, H.; Moon-Walker, A.; Whelan, S. P.; LaBranche, C. C.; Saphire, E. O.; Montefiori, D. C.
Tracking changes in SARS-CoV-2 Spike: Evidence that D614G increases infectivity of the
COVID-19 virus. Cell 2020, 182, 812-827.e19.
(145) Laha, S.; Chakraborty, J.; Das, S.; Manna, S. K.; Biswas, S.; Chatterjee, R.
Characterizations of SARS-CoV-2 mutational profile, spike protein stability and viral
transmission. Infect. Genet. Evol. 2020, 85, 104445.
(146) Mansbach, R. A.; Chakraborty, S.; Nguyen, K.; Montefiori, D.; Korber, B.; Gnanakaran,
S. The SARS-CoV-2 spike variant D614G favors an open conformational state. bioRxiv 2020,
doi: 10.1101/2020.07.26.219741. 2007.2026.219741.
(147) Gobeil, S.; Janowska, K.; McDowell, S.; Mansouri, K.; Parks, R.; Manne, K.; Stalls,
V.; Kopp, M.; Henderson, R.; Edwards, R. J.; Haynes, B. F.; Acharya, P. D614G mutation
78
alters SARS-CoV-2 spike conformational dynamics and protease cleavage susceptibility at the
S1/S2 junction. bioRxiv 2020, doi: 10.1101/2020.10.11.335299.
TAN, SUIS (SUP) = ONT a SIUY (e-COWN, Tp)
ee o_o ee
6WPS, 38309 ( 3-down) 7K43, 3S2M11 (3-down)
VA
FORM, € 1014 (é-up, 1-Cown)
up, 1-<down) TCAK, 3014 (3-up) SWPT 2 S309 (2-down, 1-up)
bp
4
Ho14
B c
~<down) 7K43, 3S2M11 (3-down) TK4N 3S2E12 (3-up)
686 828 8!
=
a
e
bd
8
&
2
s
$1/S2 | $2’
a
=
.
:
a re
;| Soca.
g -
i, NTE
8 Lg
. 7
3 e z SFA
3
3
E
é
=x
SESRERRSSISSSERETES
Residue Number
1
f
We Evolutionary Conservation Score >
st
i
. al
a tH:
Hl
a
ae
BLASER STSSTSIESSSSGSTSSES
Ghaaicheae Mosnal
528 591 686 828 853 910 985 1035 1068 11631211 1234 12
sisz| sz’ $2
< | i EE
SEsgegeeesgsges22528822223
Residue Number
: BSE ee
‘$30 oBtope contact posimons.
3 Ehaszov: Nase epiage contact postions
— :
W KL Evolutionary Conservation Score #
2 +
.
c725
C1014
C102:
C1108
C1064
C742
C1014
C102:
C1014
C10:
}3992292929922953
sidue Number
EES I te crate
SRRSSRERERRE
ewes
one
D’
RESERTR ER EE
#2332
Residue Number
* Comuojenpeewmue oo = 3s 5 «= & “¢
‘smog enprean asl & > = a=
Aysueg e2uNsig senpjsey Guisjoneo3
a
3
E
z
3
3
$
«
EgSESES
g
Ta
cluster
HR1/CH/BH
cluster
HR1/CH
114 306 331 528501686828 BSS (910 98
| s1 I $2
j — j BEEBE tare ii
le thi '
| Sie | “sane | “ny
‘ail nn
331 528 SOT 686 BZB_ «BSS 910 985 1035 1068 1163 1211 1234 1273
| s2 |
* ase
i ~~ i
i
i 1 oe
ul li, J al | ul ld eal
|-——= = f= — _
ahaa, thy i ‘| rt | | i
(}BRISRESSEG Sieere =
OWT 25509 (<-Gown, T-up)
TK4N 3 S2E12 (3-up)
ee
TCAI, 2 H014 (2-up, 1-down)
£ rng 8
Sow TOMI, OOOO OT POTS 6 See eee
ini Book . RSS See Smglze
i
Hs
corre
-Ararage Response of Site to Other Res
SOPRA TTENTT TTT TES “SECC EEE:
Number (Eftectovinfluescer Ste)" Retsdue Number (Effector Influencer
RIE EEES Si Seat
2 RSS RS SSRe See -
" “
i.
La
Resioon Number (Efctorintvence te
| REESE IEE SI Sa . 4
‘aaanrememeneen D
| BRESiah pees sama ses gS q
% |.
a
a.
ri 2
I
. |
i"
j
i
d
UAL
are | tp tt Sammgn aann or euoen earn
pa i FF
Resins Number (Eflectortintwencer Site
| REESE IEE SI Sa ¢ -
Zz:
I
E
eiah pepminn inees st Get ts
sal
| 2021 | Coevolutionary Analysis and Perturbation-Based Network Modeling of the SARS-CoV-2 Spike Protein Complexes with Antibodies: Binding-Induced Control of Dynamics, Allosteric Interactions and Signaling | 10.1101/2021.01.19.427320 | [
"Verkhivker Gennady M.",
"Di Paola Luisa"
] | null |
1
Optogenetic actuator/ERK biosensor circuits identify MAPK network
1
nodes that shape ERK dynamics
2
3
4
5
Coralie Dessauges1, Jan Mikelson2, Maciej Dobrzyński1, Marc-Antoine Jacques1,
6
Agne Frismantiene1, Paolo Armando Gagliardi1, Mustafa Khammash2, Olivier Pertz1
7
8
1Institute of Cell Biology, University of Bern, Baltzerstrasse 4, 3012 Bern, Switzerland
9
2Department of Biosystems Science and Engineering, ETH Zurich, Mattenstrasse 26,
10
4058 Basel, Switzerland
11
12
13
Key words
14
15
ERK dynamics, MAPK network, signaling robustness, optogenetics, single cell biology
16
17
Abstract
18
19
Combining single-cell measurements of ERK activity dynamics with perturbations provides
20
insights into the MAPK network topology. We built circuits consisting of an optogenetic
21
actuator to activate MAPK signaling and an ERK biosensor to measure single-cell ERK
22
dynamics. This allowed us to conduct RNAi screens to investigate the role of 50 MAPK
23
proteins in ERK dynamics. We found that the MAPK network is robust against most node
24
perturbations. We observed that the ERK-RAF and the ERK-RSK2-SOS negative feedbacks
25
operate simultaneously to regulate ERK dynamics. Bypassing the RSK2-mediated feedback,
26
either by direct optogenetic activation of RAS, or by RSK2 perturbation, sensitized ERK
27
dynamics to further perturbations. Similarly, targeting this feedback in a human ErbB2-
28
dependent oncogenic signaling model increased the efficiency of a MEK inhibitor. The RSK2-
29
mediated feedback is thus important for the ability of the MAPK network to produce consistent
30
ERK outputs and its perturbation can enhance the efficiency of MAPK inhibitors.
31
32
2
Introduction
33
34
The extracellular signal-regulated kinase (ERK) is part of the mitogen-activated protein
35
kinase (MAPK) signaling network and regulates a large variety of fate decisions. While
36
ERK can be activated by several extracellular inputs, ERK signaling has mostly been
37
studied in the context of receptor tyrosine kinases (RTKs). Upon binding of their
38
cognate growth factors (GFs), RTKs activate a complex signaling cascade with the
39
following hierarchy: (1) recruitment of adaptor molecules such as GRB2 (Schlessinger
40
2000), (2) activation of RAS GTPases through Guanine nucleotide exchange factors
41
(GEFs) and GTPase activating proteins (GAPs) (Cherfils and Zeghouf 2013), (3)
42
triggering of a tripartite RAF, MEK, ERK kinase cascade that is further regulated by a
43
variety of binding proteins (Lavoie et al. 2020), (4) ERK-mediated phosphorylation of
44
a large number of substrates. Due to its central role in fate decisions, MAPK network
45
dysregulation is causative for a large number of diseases including cancer (Rauen
46
2013; Samatar and Poulikakos 2014).
47
As for other signaling pathways (Purvis and Lahav 2013), temporal patterns of ERK
48
activity, hereafter referred to as ERK dynamics, rather than steady states control fate
49
decisions (Santos et al. 2007; Avraham and Yarden 2011; Ryu et al. 2015; Albeck et
50
al. 2013). These specific ERK dynamics have been shown to arise from feedbacks in
51
the MAPK network. For example, a negative feedback (NFB) from ERK to RAF can
52
produce adaptive or oscillatory ERK dynamics (Santos et al. 2007; Kholodenko et al.
53
2010; Avraham and Yarden 2011). The ERK-RAF NFB was also shown to buffer
54
against MAPK node perturbations (Sturm et al. 2010; Fritsche-Guenther et al. 2011).
55
This property might allow cells to produce consistent ERK outputs despite
56
heterogeneous node expression (Blüthgen and Legewie 2013). In this work, we
57
specifically refer to the ability of the MAPK network to produce consistent ERK
58
dynamics in presence of node perturbations as signaling robustness. While several
59
NFBs have been mapped experimentally in the MAPK network (Lake et al. 2016), their
60
contribution to this signaling robustness and shaping ERK dynamics remains largely
61
unknown.
62
Single-cell biosensor imaging has provided new insights into MAPK signaling that
63
were not accessible with biochemical, population-averaged measurements. This
64
showed that the MAPK network can produce a wide variety of ERK dynamics such as
65
transient (Ryu et al. 2015), pulsatile (Albeck et al. 2013), oscillatory (Shankaran et al.
66
2009) and sustained dynamics (Ryu et al. 2015; Blum et al. 2019). Mathematical
67
modeling has provided insights into the network’s structures that decode different
68
signaling inputs into specific ERK dynamics (Santos et al. 2007; Shankaran et al.
69
2009; Nakakuki et al. 2010; Ryu et al. 2015). Combined modeling/experimental
70
approaches helped to shed light on various subparts of the MAPK network, including
71
the epidermal growth factor receptor (EGFR) module (Koseska and Bastiaens 2020),
72
the RAS module (Schmick et al. 2015; Erickson et al. 2019), and the tripartite
73
RAF/MEK/ERK cascade (Ferrell and Bhatt 1997; Kholodenko 2000; Orton et al. 2005;
74
Santos et al. 2007; Ryu et al. 2015; Kochańczyk et al. 2017; Arkun and Yasemi 2018).
75
However, the low experimental throughput to measure ERK dynamics, or other MAPK
76
3
network nodes, has precluded a global understanding of the specific functions of the
77
nodes present in the network.
78
Here, we built multiple genetic circuits consisting of optogenetic actuators together
79
with an ERK biosensor to simultaneously activate ERK from different nodes in the
80
MAPK network and report single-cell ERK dynamics. These circuits allowed us to
81
investigate the role of 50 MAPK signaling nodes in ERK dynamics regulations with
82
RNA interference (RNAi). We observed that most perturbations of individual nodes
83
resulted in mild ERK dynamics phenotypes despite targeting major MAPK signaling
84
nodes. Further, the ERK dynamics induced by various perturbations suggest that two
85
NFBs (ERK-RAF and ERK-RSK2-SOS) act simultaneously to regulate ERK dynamics.
86
Targeting the RSK2-mediated NFB increased the efficiency of additional MAPK
87
network perturbations both in our optogenetic systems and in an ErbB2-driven
88
oncogenic ERK signaling model. This suggests that the RSK2-mediated feedback
89
plays a role in MAPK signaling robustness and can be targeted for potent inhibition of
90
oncogenic ERK signaling.
91
4
Results
92
An optogenetic actuator-biosensor genetic circuit to study input-dependent
93
ERK dynamics
94
In order to measure ERK dynamics in response to dynamic RTK input, we built a
95
genetically-encoded circuit made of an optogenetic RTK actuator and an ERK
96
biosensor (Figure 1A). We chose optoFGFR, which consists of a myristoylated
97
intracellular domain of the fibroblast growth factor receptor 1 (FGFR1) fused to a CRY2
98
domain and tagged with mCitrine (Kim et al. 2014). Upon stimulation with blue light,
99
optoFGFR dimerizes and trans-autophosphorylates, leading to the activation of the
100
MAPK/ERK, phosphoinositide 3-kinase (PI3K)/AKT, and phospholipase C (PLC)/Ca2+
101
pathways. As ERK biosensor, we used ERK-KTR-mRuby2 that is spectrally
102
compatible with optoFGFR. ERK-KTR reversibly translocates from the nucleus to the
103
cytosol upon ERK activation (Regot et al. 2014). We used a nuclear Histone 2B (H2B)-
104
miRFP703 marker to identify and track single cells. After stably inserting these
105
constructs into murine NIH3T3 fibroblasts, we used automated time-lapse microscopy
106
to stimulate selected fields of view with defined blue light input patterns to activate
107
optoFGFR. The corresponding ERK-KTR/H2B signals were recorded with a 1-minute
108
temporal resolution. We observed that a 100 ms light pulse leads to reversible ERK-
109
KTR translocation from the nucleus to the cytosol, indicative of transient ERK
110
activation (Figure 1B, Appendix Movie S1). At the end of each experiment, we imaged
111
the mCitrine signal to evaluate optoFGFR expression levels. We built a computer
112
vision pipeline to automatically track each nucleus, compute ERK activity as the
113
cytosolic/nuclear ratio of the ERK-KTR signals and correlate single-cell ERK
114
responses with optoFGFR levels (Figure 1C). We then use this pipeline to evaluate
115
the sensitivity and specificity of our system with dose response experiments using the
116
FGFR inhibitor SU5402, the RAF inhibitor RAF709, the MEK inhibitor U0126 and the
117
ERK inhibitor SCH772984 (Appendix Figure S1A).
118
119
To evaluate light-dependent optoFGFR activation dynamics, we engineered a
120
mScarlet-tagged optoFGFR that is spectrally orthogonal to CRY2 absorption
121
(Appendix Figure S1B). Total internal reflection (TIRF) microscopy visualized the
122
formation of optoFGFR clusters in response to blue light-mediated dimerization in the
123
plasma membrane (Appendix Figure S1B, blue arrows, Appendix Movie S2).
124
Consistently with CRY2’s dissociation half-life (Duan et al. 2017), these optoFGFR
125
clusters appeared within 20 seconds after a blue light pulse and disappeared after ~
126
5 minutes (Appendix Figure S1C). We assume that optoFGFR is active in its clustered
127
form in which transphosphorylation occurs and inactive in its monomeric form due to
128
tonic cytosolic phosphatase activity (Lemmon et al. 2016). As documented previously
129
(Kim et al. 2014), light stimulation also triggered optoFGFR endocytosis (Appendix
130
Figure S1B, red arrows).
131
132
5
Directly following light stimulation, we systematically observed a short ERK
133
inactivation period, that we refer to as “dip”, lasting 2-3 minutes before activation of a
134
strong ERK activity (Appendix Figure S1D, green rectangle). This light-induced ERK
135
dip was insensitive to SCH772984-mediated ERK inhibition but could be suppressed
136
by Cyclosporin A-mediated calcineurin inhibition. Calcineurin is a Ca2+-dependent
137
phosphatase that dephosphorylates Ser383 in Elk1 (Sugimoto et al. 1997). As ERK-
138
KTR contains an Elk-1 docking domain phosphorylated by ERK (Regot et al. 2014),
139
we hypothesized that it could be negatively affected by optoFGFR-evoked Ca2+ input
140
(Kim et al. 2014) (Appendix Figure S1E). Consistently, Ionomycin-evoked increase in
141
cytosolic Ca2+ induced a dip in absence of light stimulation (Appendix Figure S1F).
142
143
144
Figure 1: An optogenetic actuator-biosensor genetic circuit to study input-dependent ERK
145
dynamics. (A) Schematic representation of the optoFGFR system consisting of the optogenetic FGF
146
receptor (optoFGFR) tagged with mCitrine, the ERK biosensor (ERK-KTR) tagged with mRuby2 and a
147
nuclear marker (H2B) tagged with miRFP703. (B) Time lapse micrographs of ERK-KTR dynamics in
148
response to a 470 nm light pulse. Using a 20x air objective, ERK-KTR and H2B channels were acquired
149
every 1 minute and the optoFGFR channel was acquired once at the end of the experiment. Scale bar:
150
50 μm. (C) Image analysis pipeline developed to quantify single-cell ERK dynamics. Nuclear and
151
cytosolic ERK-KTR signals were segmented based on the H2B nuclear mask. Single-cell ERK activity
152
was then calculated as the cytosolic/nuclear ERK-KTR ratio. Single-cell optoFGFR intensity was
153
measured under the cytosolic ERK-KTR mask and used as a proxy for single-cell optoFGFR
154
expression.
155
Different optoFGFR inputs trigger transient, oscillatory and sustained ERK
156
dynamics
157
Next, we characterized optoFGFR-triggered ERK dynamics in response to a single
158
light pulse of different intensities and durations (Figure 2A). As ERK dynamics
159
A
C
B
FGFRcyto
FGFRcyto
SOS
MEK
GRB2
FRS2
RAF
RAS
ERK
P
ERK-KTR
ERK-KTR
ERK inactive
ERK active
H2B
-5 min
+5 min
+10 min
+15 min
+30 min
ERK-KTR
mRuby2
H2B
miRFP703
optoFGFR
mCitrine
Post time-lapse
t
ERK-KTR
segmentation
- Single cell tracking
- Full-length trajectories selection
- Receptor intensity matching
ERK-KTR
optoFGFR
H2B
optoFGFR
segmentation
0
10
20
30
40
0.5
1.0
ERK activity
Time [min]
Single cell
Average
Light input
Nuc
Cyto
Nuc
Cyto
Intensity (log10)
Density
optoFGFR distribution
P
P
P
P
P
6
depended on light power density, as well as pulse duration, we defined the light dose
160
(D, mJ/cm2) as their product to quantify the total energy received per illuminated area.
161
To characterize ERK dynamics, we extracted the amplitude at the maximum of the
162
peak (maxPeak), and the full width at half maximum (FWHM) of the ERK trajectories
163
(Figure 2B). With increasing light doses, ERK peaks increased both in duration and
164
amplitude, until the latter reached saturation. Based on these observations, we
165
selected 180 mW/cm2 and 100 ms (D = 18 mJ/cm2) as the minimal light input to
166
generate an ERK transient of maximal amplitude. Using this light dose, we then
167
investigated ERK dynamics in response to multiple light pulses delivered at different
168
intervals (Figure 2C). All stimulation regimes led to identical maximal ERK amplitude
169
(Figure EV1A) and adaptation kinetics when optoFGFR input ceased (Figure EV1B).
170
Repeated light inputs applied at 10- or 20-minute intervals evoked population-
171
synchronous ERK transients. In contrast, repeated light inputs applied at higher
172
frequencies (2-minute intervals) led to sustained ERK dynamics. Given CRY2’s 5-
173
minute dissociation half-life (Appendix Figure S1B-C) (Duan et al. 2017), this suggests
174
that light pulses delivered at a 2-minute interval reactivate optoFGFR faster than it
175
deactivates, leading to sustained optoFGFR activity. Hierarchical clustering of ERK
176
responses to sustained optoFGFR input highlighted the presence of sustained and
177
oscillatory single-cell ERK dynamics (Figure 2D). Classification of ERK trajectories
178
based on optoFGFR expression revealed that sustained/oscillatory ERK dynamics
179
correlated with high/low optoFGFR levels (Figure 2E, Appendix Movie S3). Oscillatory
180
ERK dynamics were also observed in optoFGFR high expressing cells in response to
181
low light input (Figure 2F). Thus, sustained optoFGFR input can trigger sustained or
182
oscillatory ERK dynamics depending on the input strength, a combination of light
183
energy and optoFGFR expression.
184
7
185
Figure 2: Different optoFGFR inputs trigger transient, oscillatory and sustained ERK dynamics.
186
(A) ERK responses to increasing light power densities and pulse durations of 470 nm transient light
187
input. The light dose “D” is calculated as the product of the power density and pulse duration. (B)
188
Quantification of the maxPeak (maximal ERK amplitude of the trajectory) and the FWHM (full width at
189
half maximum) of single-cell ERK responses shown in (A) (Nmin = 40 cells per condition). (C) ERK
190
responses to 470 nm light pulses delivered every 20, 10, 5 and 2 minutes respectively (D = 18 mJ/cm2).
191
(D) Hierarchical clustering (Euclidean distance and Ward D2 linkage) of trajectories from the 2-minute
192
interval stimulation shown in (C) (referred to as “sustained”) (N = 60 cells). The number of clusters was
193
empirically defined to resolve the different ERK dynamics. The average ERK responses per cluster are
194
displayed on the right. (E) Separation of the trajectories shown in (D) in low and high optoFGFR cells,
195
based on the log10 intensity of optoFGFR-mCitrine. (F) ERK responses to increasing doses of
196
sustained optoFGFR input. Single-cell ERK trajectories were divided in low (top panel) and high (bottom
197
panel) optoFGFR expression.
198
0
10
20
30
40 0
10
20
30
40 0
10
20
30
40
0
10
20
30
40
0.0
0.5
1.0
Time [min]
0 mW/cm2
5 mW/cm2
20 mW/cm2
180 mW/cm2
50 ms
100 ms
1 s
0.0
0.5
1.0
0.0
0.5
1.0
ERK activity
A
B
C
Time
ERK activity
FWHM
maxPeak
Feature extraction
2 min interval
10 min interval
20 min interval
5 min interval
E
D
ERK trajectories
2 min interval
Dose (D) = 0 mJ/cm2
D = 0
D = 0
D = 0.25
D = 0.5
D = 5
D = 1
D = 2
D = 20
D = 9
D = 18
D = 180
Time [min]
25
50
75
5
25
50
75
5
ERK activity
25
50
75
5
0.25
0.50
0.75
1.00
25
50
75
5
F
0 mJ/cm2
5 mJ/cm2
18 mJ/cm2
2.5 mJ/cm2
0.0
0.5
1.5
����
����
����
Intensity (log10)
Density
1.0
optoFGFR distribution
Light power density
Light dose
Pulse duration
Clusters
1
3
2
4
4
3
2
1
Low
High
25
50
75
5
0.5
1.0
Average
ERK activity
0.25
0.50
0.75
1.00
ERK activity
25
50
75
5
0.25
0.50
0.75
1.00
Time [min]
Low
optoFGFR
1.00
0.25
0.50
0.75
1.00
ERK activity
0.25
0.50
0.75
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
Time [min]
Time [min]
Time [min]
0.3
1.2
ERK
activity
FWHM [min]
maxPeak
5 mW/cm2
20 mW/cm2
180 mW/cm2
Power density
50ms
100ms
1s
12
9
6
NA
0.8
1.2
50ms
100ms
1s
0.4
0 mW/cm2
Pulse duration
Oscillatory
Sustained
NA
NA
NA
Average
Single cell
High
optoFGFR
Average
Single cell
Low
optoFGFR
Average
Single cell
High
optoFGFR
Average
Single cell
8
ERK dynamics evoked by optoFGFR versus endogenous RTKs highlight
199
different MAPK regulatory mechanisms
200
Because of the absence of an ectodomain, optoFGFR must be considered as a
201
prototypic RTK that lacks some regulatory mechanisms inherent to the native FGFR.
202
To evaluate if optoFGFR is relevant for studying the MAPK network, we compared
203
ERK dynamics evoked by optoFGFR inputs versus stimulation of the endogenous
204
FGFR or EGFR using increasing concentrations of basic FGF (bFGF) and EGF. All
205
bFGF concentrations led to an ERK peak similar in amplitude to sustained optoFGFR
206
input (Figure 3A, EV1C, compared to EV1A). However, FGFR inputs led to different
207
ERK dynamics than optoFGFR: 1 ng/ml bFGF led to damped ERK oscillations
208
followed by steady state sustained ERK activity, while 10 and 100 ng/ml bFGF
209
concentrations led to a first ERK peak followed by a strong adaptation. The biphasic
210
behavior induced by increasing bFGF concentrations was previously documented to
211
emerge from the competition of bFGF for FGFR and heparan sulfate proteoglycan co-
212
receptors (Kanodia et al. 2014; Blum et al. 2019). It is thus not surprising that
213
optoFGFR, that lacks these extracellular interactions, produced different ERK
214
dynamics than FGFR. All EGF concentrations led to an ERK peak similar in amplitude
215
to optoFGFR and FGFR inputs (Figure 3B, EV1D). As for bFGF, 1 ng/ml EGF
216
concentration evoked damped oscillatory ERK dynamics that decreased at higher
217
EGF concentrations. However, EGFR inputs led to strong ERK adaptation, not
218
observed in response to optoFGFR inputs, suggesting the existence of different
219
regulatory mechanisms.
220
Both oscillatory and transient ERK dynamics can be explained by the presence of NFB
221
(Kholodenko et al. 2010). Thus, we wondered if the different ERK dynamics induced
222
by optoFGFR or EGFR input emerge from differences in downstream NFBs. We
223
reasoned that if EGFR induces different NFBs than optoFGFR, pre-stimulating cells
224
with EGF should activate these feedbacks, and affect subsequent optoFGFR-evoked
225
ERK dynamics. To test this, we pre-stimulated cells with sustained EGFR input,
226
subsequently applied sustained optoFGFR input, and evaluated ERK dynamics
227
(Figure 3C). Pre-stimulation with 100 ng/ml EGF led to the characteristic adaptive ERK
228
transient. Subsequent application of optoFGFR input yielded sustained ERK
229
responses similar in amplitude and duration to non-pre-stimulated cells. However,
230
EGF pre-stimulation led to a reduction of synchronous optoFGFR-evoked ERK
231
oscillations in low optoFGFR expressing cells.
232
To provide intuition about the MAPK network circuitries leading to different ERK
233
dynamics in response to optoFGFR and EGFR inputs, as well as the origin of the
234
oscillatory behavior, we built a mathematical model consisting of the RAS GTPase
235
and the three-tiered RAF/MEK/ERK network (Figure 3D, Appendix Table S1). We
236
used ordinary differential equations with Michaelis-Menten kinetics (see Material and
237
methods, Appendix Table S2 and S3). To account for the oscillatory ERK dynamics in
238
response to EGFR and optoFGFR inputs, we included the well-documented ERK-RAF
239
NFB (Kholodenko et al. 2010; Santos et al. 2007; Fritsche-Guenther et al. 2011; Blum
240
et al. 2019). We also included a receptor level inactivation process for EGFR, but not
241
for optoFGFR, to account for EGF-dependent regulatory mechanisms. We used a
242
9
Bayesian inference approach (Mikelson and Khammash 2020) to infer the model
243
parameters from averaged ERK trajectories in response to sustained low optoFGFR
244
input with or without sustained EGFR pre-stimulation (Figure 3E). After identification
245
of parameters that allowed the model to capture the training dataset (Figure 3F), we
246
simulated ERK dynamics evoked by low EGFR input (adaptative, oscillatory ERK
247
dynamics), high EGFR input (adaptative ERK dynamics without oscillation) and
248
sustained high optoFGFR input (sustained ERK dynamics) (Figure 3G). We observed
249
that our model with a NFB and EGFR inactivation was able to predict ERK dynamics
250
evoked by different EGFR and optoFGFR input strengths, while two simpler models
251
(one with only the EGFR inactivation reaction, but no NFB (Figure EV1E-G) and one
252
with only the NFB, but no EGFR inactivation (Figure EV1H-J) were not able to
253
reproduce experimentally observed ERK dynamics.
254
This suggested that oscillatory optoFGFR-evoked ERK dynamics emerge from a NFB
255
also present downstream of endogenous EGFR, while additional regulatory
256
mechanisms seem to be required for the strong ERK transient adaptation following
257
EGFR input. These mechanisms might consist of receptor-level regulations such as
258
endocytosis, which was recently shown to be an important regulator of the transient
259
adaptive EGF-triggered ERK dynamics in different cell systems (Kiyatkin et al. 2020;
260
Gerosa et al. 2020). While optoFGFR also gets endocytosed (Appendix Figure S1B,
261
(Kim et al. 2014)), it most likely is insensitive to inactivation by endosome acidification
262
since it lacks an ectodomain (Huotari and Helenius 2011). Additionally, light-mediated
263
optoFGFR dimerization might occur both at the plasma and endo-membranes,
264
allowing for reactivation of endocytosed optoFGFR. The hypothesis that a receptor
265
level mechanism is important for strong adaptation was further supported by inhibition
266
of optoFGFR with the FGFR kinase inhibitor (SU5402), which shifted ERK dynamics
267
from sustained to transient in a dose response-dependent manner (Figure EV1K).
268
Thus, these results suggest that optoFGFR lacks receptor-dependent regulatory
269
mechanisms but allows us to investigate the intracellular MAPK feedback structure
270
shaping ERK dynamics. In our model, we used the well-established ERK-RAF NFB.
271
However, several NFBs have been mapped in the MAPK signaling cascade, whose
272
role in shaping ERK dynamics is still unknown and which could also be responsible
273
for the observed oscillatory ERK dynamics.
274
10
275
Figure 3: ERK dynamics evoked by optoFGFR versus endogenous RTKs highlight different
276
MAPK regulatory mechanisms. (A-B) Single-cell ERK trajectories under increasing concentrations of
277
sustained (A) bFGF or (B) EGF input added at t = 5 minutes. (C) ERK responses of cells stimulated
278
with sustained optoFGFR input (D = 18 mJ/cm2) at t = 24 minutes without or with 100 ng/ml EGF
279
sustained pre-stimulation at t = 5 minutes. Average ERK responses for optoFGFR high and low
280
expression levels are shown (N = 20 cells for low and high optoFGFR, randomly selected out of at least
281
80 cells). (D) Mathematical model topology consisting of the RAS GTPase, the MAPK three-tiered (RAF,
282
MEK, ERK) network and the ERK-KTR reporter. EGFR and optoFGFR inputs both activate the
283
RAS/RAF/MEK/ERK cascade and the ERK-RAF NFB. EGFR activity is under receptor-dependent
284
regulations. (E) Training dataset consisting of the average ERK responses evoked by sustained low
285
optoFGFR input with or without pre-stimulation with 100 ng/ml sustained EGF. (F) Simulation of ERK
286
responses from the training dataset, including the maximum a posteriori (MAP) estimate, the posterior
287
envelope indicating the predictive density of our estimation, as well as an example trajectory. (G)
288
Predictions of the model for ERK responses evoked by 1 ng/ml EGF, 100 ng/ml EGF and sustained
289
high optoFGFR inputs. Note that for low EGFR input (1 ng/ml), the model predicts both adaptive and
290
oscillatory ERK responses.
291
292
293
A
B
ERK activity
Time [min]
0.25
0.50
0.75
1.00
40
10
20
0
30
40
10
20
0
30
40
10
20
0
30
ERK activity
Time [min]
0.25
0.50
0.75
1.00
40
10
20
0
30
40
10
20
0
30
40
10
20
0
30
10 ng/ml
1 ng/ml
100 ng/ml
100 ng/ml EGF
18 mJ/cm2 light
Time [min]
0.25
0.50
0.75
ERK activity
40
20
0
60
1.00
Time [min]
40
20
0
60
Average low
optoFGFR
Average high
optoFGFR
Time
Light
EGF
C
Light
Light
EGF
bFGF
10 ng/ml
1 ng/ml
100 ng/ml
EGF
Single cell
D
MEK
RAF
RAS
ERK
EGFR
optoFGFR
Receptor
inactivation
ERK-KTR
Training
Predictions
EGFR
Data
NFB
Maximum a posteriori
estimate
Example trajectory
Posterior envelope
40
10
20
0
30
50
ERK activity
0.4
0.6
0.8
40
10
20
0
30
50
0.4
0.6
0.8
40
10
20
0
30
50
0.4
0.6
0.8
Time [min]
40
10
20
0
30
50
60
70
0.4
0.6
0.8
ERK activity
40
10
20
0
30
50
0.4
0.6
0.8
High optoFGFR expression
Low optoFGFR expression
1 ng/ml EGF
100 ng/ml EGF
100 ng/ml EGF +
Low optoFGFR expression
40
10
20
0
30
50
60
70
0.4
0.6
0.8
ERK activity
40
10
20
0
30
50
0.4
0.6
0.8
EGFR input
optoFGFR input
Training dataset
E
F
G
P
P
P
P
11
RNA interference screen reveals that ERK dynamics remain unaffected in
294
response to perturbation of most MAPK signaling nodes
295
We then explored the network circuitry that shapes optoFGFR-evoked ERK dynamics
296
with an RNA interference (RNAi) screen targeting 50 MAPK signaling nodes. We
297
focused our screen on sustained optoFGFR input which captured the largest amount
298
of information about ERK dynamics when compared to other stimulation schemes: it
299
led to sustained and oscillatory ERK dynamics (Figure 2E,F) while recapitulating the
300
rapid increase of ERK activity and adaptation observed with transient input (Figure
301
EV1A,B). We used a bioinformatic approach to select 50 known interactors of the
302
tripartite RAF/MEK/ERK cascade downstream of the FGFR receptor that were
303
detected in a NIH3T3 proteome (Schwanhäusser et al. 2011) (Figure 4A, Appendix
304
Table S4). We used the siPOOL technology to specifically knockdown (KD) these 50
305
MAPK signaling nodes while limiting off-target effects (Hannus et al. 2014). We first
306
validated KD efficiency by quantifying transcript levels with different siPOOL
307
concentrations targeting the ERK and MEK isoforms (Figure EV2A) and observed
308
strong KD with 10 nM siRNA concentration. We then evaluated the effect of ERK1 or
309
ERK2 KD on ERK dynamics. We observed only subtle phenotypes compared to the
310
non-targeting siRNA (CTRL) used as negative control (Figure 4B), even though
311
efficient KD was observed at protein level (Figure 4C). However, combined
312
ERK1/ERK2 KD strongly suppressed ERK dynamics indicating that the latter is not
313
affected by the perturbation of individual ERK isoforms as previously reported
314
(Fritsche-Guenther et al. 2011; Ornitz and Itoh 2015). Due to its strong phenotype, we
315
used ERK1/ERK2 KD as positive control throughout our screen.
316
We performed three replicates of the screen targeting the 50 nodes. Despite efficient
317
KD quantified for different nodes (Figure EV2B), visual inspection of ERK trajectories
318
only revealed subtle ERK dynamics phenotypes for a limited number of node
319
perturbations (Figure EV2C,D). We used a feature-based approach to evaluate the
320
effect of each perturbation on ERK dynamics. We focused our analysis on ERK
321
responses evoked by high optoFGFR input to limit the single-cell heterogeneity due to
322
optoFGFR expression variability. We quantified the average ERK activity before
323
stimulation (baseline), the maximal ERK amplitude during stimulation (maxPeak), and
324
the ERK amplitude at a fixed time point after response adaptation in the negative
325
control (ERKpostStim). To evaluate these phenotypes, we z-scored the features
326
associated to each perturbation to those of the negative control (Figure 4D, see
327
Material and method for details). While many phenotypes were statistically significant,
328
most of them remained mild as observed by visually inspection of the feature
329
distributions (Figure EV3A). Apart from ERK1+2 KD, only GRB2, PTK2 and ERK2 led
330
to a reduction of ERK amplitude (maxPeak). KD of negative regulators such as
331
SPROUTY 2,3 and 4, or phosphatases such as PP2A and several dual-specificity
332
phosphatases (DUSPs) led to increased ERK amplitude. Increased basal ERK activity
333
was observed for RKIP, PP2A, DUSP4 and DUSP6 KDs, indicating a function in
334
regulating basal ERK levels. Prolonged ERK activity (ERKpostStim) was observed in
335
12
response to KD of RKIP, PP2A, ERK2, DUSP1,2,3,4,6 and strikingly for RSK2 KD
336
(Figure EV3B), suggesting a role of these nodes in ERK adaptation.
337
Because both visual inspection of trajectories, as well as our feature-based approach
338
might miss more subtle ERK dynamics phenotypes, we used CODEX (Jacques et al.
339
2021), a data-driven approach to identify patterns in single-cell time-series based on
340
convolutional neural networks (CNNs) (Figure EV3C). We trained a CNN to classify
341
ERK trajectories that originate from different siRNA perturbations and selected the ten
342
perturbations for which the CNN classification accuracy was the highest (Appendix
343
Table S4, “CODEX accuracy”, see Material and methods for details). Projection of the
344
CNN features in a t-distributed stochastic neighbor embedding (t-SNE) space revealed
345
different clusters of ERK trajectories (Figure EV3D). Comparison of the ten trajectories
346
with the highest classification confidence identified by CODEX to randomly selected
347
ERK trajectories for low or high optoFGFR expression highlighted ERK phenotypes
348
not accessible to visual inspection and the feature-based approach (Figure 4E).
349
CODEX identified some of the perturbations that affect ERK amplitude, baseline or
350
adaptation observed with the feature-based approach. However, it also highlighted
351
perturbations affecting oscillatory ERK dynamics. PP2A KD led to sustained oscillatory
352
behavior. PLCG1 KD resulted in a first peak followed by damped oscillations, and
353
absence of the dip. As phospholipase C mediates Ca2+ signaling in response to FGFR
354
activation (Ornitz and Itoh 2015), this further validates the role of Ca2+ signaling in
355
formation of the dip (Appendix Figure S1D-F). RAPGEF1 KD led to oscillatory ERK
356
responses of different amplitudes. RSK2, ERK2 and CRAF KD displayed reduced
357
oscillatory ERK behavior.
358
To validate the latter oscillatory ERK dynamics phenotypes, we evaluated the
359
proportion of oscillatory trajectories (trajectories with at least 3 peaks) for each
360
perturbation, both for high and low optoFGFR input (Figure 4F). This confirmed that
361
RSK2, CRAF and ERK2 KD led to decreased oscillatory ERK dynamics. We also
362
observed that these perturbations reduced ERK oscillations in cells stimulated with 1
363
ng/ml EGF (Figure EV3E-G), suggesting a role of these nodes in the regulation of ERK
364
oscillations in the context of a native RTK.
365
ERK2 and CRAF isoforms are implicated in the well-established ERK-RAF NFB,
366
known to regulate ERK dynamics (Santos et al. 2007; Ryu et al. 2015; Blum et al.
367
2019), and to enable consistent ERK dynamics under MEK or ERK perturbations
368
(Fritsche-Guenther et al. 2011; Sturm et al. 2010). RSK2 encodes the p90 ribosomal
369
S6 kinase 2 protein, an ERK substrate regulating survival and proliferation (Cargnello
370
and Roux 2011; Yoo et al. 2015). RSK2 is also known to be involved in an ERK-
371
induced NFB targeting SOS (Douville and Downward 1997; Saha et al. 2012; Lake et
372
al. 2016), whose significance in the regulation of ERK dynamics has been less well
373
studied. In addition to dampening ERK oscillations, RSK2 KD also led to slower ERK
374
adaptation when optoFGFR input ceased (Figure 4D, EV3A,B), suggesting an
375
important role of this NFB in ERK dynamics regulation. Our results suggest that the
376
ERK-RAF and ERK-RSK2-SOS NFBs simultaneously operate within the MAPK
377
network to generate ERK oscillations and raise the question whether both NFBs
378
contribute to the strong MAPK signaling robustness observed in our screen.
379
13
380
Figure 4: RNA interference screen reveals that ERK dynamics remain unaffected in response to
381
perturbation of most MAPK signaling nodes. (A) RNAi perturbation targets referred to by their
382
protein names. Nodes were spatially grouped based on the hierarchy of interactions within the MAPK
383
network and color-coded for their function. (B) ERK responses to sustained optoFGFR input (D = 18
384
mJ/cm2) in cells transfected with 10 nM siRNA against ERK1, ERK2 or a 5 nM combination of each
385
(ERK1+2). A non-targeting siRNA (CTRL) was used as control (N = 15 cells from low and high
386
optoFGFR levels). (C) Western blot analysis of cells transfected with 10 nM siRNA against ERK1, ERK2
387
or a 5 nM combination of each (ERK1+2). (D) Z-Score evaluation of the baseline, maxPeak and
388
SPRY1
FGFR1
FRS2
GRB2
SHC1
SPRY4
SPRY2
PTPN6
GAB1
PTPN11
SPRY3
DUSP26
DUSP22
DUSP10
DUSP16
DUSP4
DUSP8
DUSP1
DUSP3
DUSP6
DUSP2
DUSP9
SOS2
SOS1
RAPGEF3
NF1
RAPGEF1
RASA1
RASGRP1
PP2A
MEK2
CNKSR1
ERK1
KSR1
NCK1
SRC
NCK2
PTK2
CRKL
SHIP
PLCG1
RRAS
NRAS
RKIP
YWHAZ
HRAS
RAP1A
RAP1B
KRAS
YWHAG
CRAF
RSK2
PEA15
ERK2
MEK1
ARAF
BRAF
DUSP14
A
B
GAPDH
1.00
0.25
0.50
0.75
ERK2
ERK1+2
CTRL
ERK1
20
40
0
10
30
1.00
0.25
0.50
0.75
20
40
0
10
30
ERK2
ERK1+2
CTRL
ERK1
ERK1
ERK2
Time [min]
ERK activity
C
Receptor
proximal layer
Membrane
MAPK
cascade
Downstream
D
F
Receptor
GEFs, GAPs, GTPases
Kinases
Phosphatases
Adaptors, scaffolds, antagonists
44 kDa
42 kDa
maxPeak
Time
ERK activity
baseline
ERKpostStim
optoFGFR expression
high
low
Proportion of
oscillating cells
0.2
0.0
0.2
0.4
PP2A
PEA15
RAPGEF1
DUSP8
KRAS
DUSP14
NCK2
SRC
DUSP9
PLCG1
YWHAZ
DUSP10
SPRY2
RAP1B
RAP1A
DUSP26
YWHAG
RAPGEF3
CTRL
RRAS
DUSP22
ARAF
SHC1
DUSP3
NCK1
PTK2
DUSP2
SOS2
PTPN11
CRKL
RASGRP1
NF1
RASA1
CNKSR1
SOS1
RKIP
SPRY4
GAB1
DUSP1
NRAS
DUSP16
MEK1
KSR1
DUSP4
HRAS
BRAF
ERK1
DUSP6
FRS2
SPRY3
MEK2
SPRY1
PTPN6
SHIP
GRB2
RSK2
CRAF
ERK2
ERK1+2
0.4
*** ****
****
****
***
***
****
****
*
*
**
**
*
**
*
*
*
*
Time
ERK activity
Peak 1 Peak 2 Peak 3
Oscillating cells
CTRL
CTRL
ERK activity
GRB2
RAPGEF1
FRS2
CRAF
DUSP6
PLCG1
PP2A
ERK2
ERK1+2
RSK2
ERK activity
Time [min]
1.00
0.25
0.50
0.75
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
1.00
0.25
0.50
0.75
1.00
0.25
0.50
0.75
1.00
0.25
0.50
0.75
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
1.00
0.25
0.50
0.75
1.00
0.25
0.50
0.75
CODEX
Low
optoFGFR
High
optoFGFR
CODEX
Low
optoFGFR
High
optoFGFR
E
Features
37 kDa
DUSP26
DUSP22
DUSP16
DUSP14
DUSP10
DUSP9
DUSP8
DUSP6
DUSP4
DUSP3
DUSP2
DUSP1
PEA15
RSK2
ERK2
ERK1
MEK2
MEK1
CRAF
BRAF
ARAF
PP2A
KSR1
CNKSR1
YWHAG
YWHAZ
RKIP
RRAS
NRAS
KRAS
HRAS
RAP1B
RAP1A
NF1
RASA1
RASGRP1
RAPGEF3
RAPGEF1
SOS2
SOS1
NCK2
NCK1
SRC
PLCG1
SHIP
PTPN11
PTPN6
SPRY4
SPRY3
SPRY2
SPRY1
CRKL
PTK2
SHC1
GAB1
GRB2
FRS2
ERK1+2
CTRL
baseline
maxPeak
ERKpostStim
2
-2
zScore to CTRL
NS
14
ERKpostStim of single-cell ERK responses under sustained high optoFGFR input (D = 18 mJ/cm2). The
389
z-score was calculated by comparing each RNAi perturbation to the CTRL KD (Nmin = 126 cells per
390
treatment, from 3 technical replicates). Non-significant (NS) results are in grey (see Figure EV3A for
391
statistical results). (E) Single-cell ERK trajectories (sustained optoFGFR input, D = 18 mJ/cm2) for the
392
RNAi perturbations classified with the highest accuracy by CODEX. Top lines show single-cell ERK
393
trajectories for which CODEX had the highest classification confidence in the validation set (N = 10).
394
Bottom lines show single-cell ERK trajectories for low and high optoFGFR cells (N = 30 for each
395
condition, randomly selected out of at least 212 cells per perturbation from 3 technical replicates). For
396
easier visualization, the CTRL condition is shown twice. (F) Proportion of oscillating cells (trajectories
397
with at least 3 peaks) per RNAi perturbation for low and high optoFGFR expression (sustained
398
optoFGFR input, D = 18 mJ/cm2, Nmin = 61 cells for low and 126 for high optoFGFR per perturbation
399
from 3 technical replicates). Perturbations were ordered based on the proportion of oscillating cells with
400
low optoFGFR expression. Statistical analysis was done using a pairwise t-test, comparing each
401
perturbation against the CTRL for each receptor level independently (*<0.05, **<0.005, ***<0.0005,
402
****<0.00005, FDR p-value correction method).
403
Direct optogenetic activation of RAS highlights different ERK dynamics
404
phenotypes than optoFGFR input
405
To further explore the role of MAPK feedbacks in MAPK signaling robustness, we used
406
optoSOS (Johnson et al. 2017), an optogenetic actuator that activates RAS, and thus
407
bypasses the RSK2-mediated NFB regulation (Figure 5A). OptoSOS consists of a
408
membrane anchored light-activatable iLID domain, and an mCitrine-tagged SspB
409
domain fused to SOS’s catalytic GEF domain. It was stably integrated into cells
410
expressing ERK-KTR and H2B. Because iLID displays faster dissociation rates than
411
CRY2 (t1/2= 30 seconds for iLID versus ~ 5 minutes for CRY2 (Duan et al. 2017;
412
Benedetti et al. 2018)), optoSOS required repeated light pulses to prolong its
413
membrane recruitment and produce a robust ERK response (Figure 5B). Five
414
consecutive 100 ms light pulses at 6 W/cm2 (D = 0.6 J/cm2) applied at 20-second
415
intervals, provided the minimal light input to induce a saturated ERK amplitude (Figure
416
EV4A). Application of this light input at 2-minute intervals evoked sustained ERK
417
dynamics with small fluctuations at the same frequency as the light input pattern,
418
reflecting the fast optoSOS reversion to the dark state (Figure 5C). OptoSOS did not
419
induce ERK oscillations (Figure EV4B), even in cells with low optoSOS expression or
420
at lower light doses (Figure 5D). However, ERK amplitudes correlated with optoSOS
421
expression level, low optoSOS levels led to low ERK amplitudes, while high actuator
422
expression levels resulted in high ERK amplitudes. Using the minimal light input to
423
trigger saturating ERK amplitude, both optoSOS and optoFGFR led to steep ERK
424
activation and fast adaptation when light stimulation ceased (compare Figures 2C and
425
5C), as well as similar ERK amplitudes in cells expressing high actuator levels (Figure
426
5E). However, high optoSOS expression levels moderately increased ERK activity
427
baseline levels in comparison to optoFGFR (Figure EV4C), suggesting that this
428
system is leaky to some extent.
429
Using this specific light input, we performed siRNA screens targeting MAPK signaling
430
nodes downstream of optoSOS in triplicates (Figure EV4D,E). We extracted the
431
baseline, maxPeak, ERKpostStim features from optoSOS high expressing cells
432
(Figure EV4F) and z-scored feature values to the negative control (Figure 5F). We
433
15
observed more prominent ERK amplitude phenotypes in response to optoSOS input
434
than to optoFGFR input. Some of these phenotypes are shown in Figure 5G. Most
435
prominently, CRAF, ERK2, DUSP4 KD led to a stronger reduction in ERK amplitude
436
than observed with optoFGFR input. RSK2 KD also reduced ERK amplitude,
437
suggesting that it also regulates nodes downstream of RAS. However, RSK2 KD did
438
not decrease ERK adaptation following optoSOS input removal (Figure EV4G),
439
suggesting that it is not involved in NFB regulation in this system. PP2A KD did not
440
induce increased ERK amplitude or baseline as observed in the optoFGFR system.
441
As for optoFGFR input, DUSP6 KD increased basal ERK activity and decreased
442
adaptation (Figure EV4G). DUSP22 KD led to increased amplitude, without affecting
443
ERK baseline and adaptation. NF1 KD, which encodes a RAS-specific GAP, led to
444
increased ERK baseline and slower adaptation (Figure EV4G), without affecting ERK
445
amplitude. The NF1 baseline phenotype, that was not observed in the optoFGFR
446
system, might emerge from the optoSOS-mediated low levels of RAS activation due
447
to the optoSOS system’s leakiness (Figure EV4C), that can then be amplified by loss
448
of NF1’s RAS GAP activity. The finding that perturbation of specific nodes (e.g. ERK2
449
and CRAF) leads to more penetrant phenotypes in response to optoSOS versus
450
optoFGFR input suggested that the RAS/RAF/MEK/ERK part of the network is more
451
sensitive to perturbations than optoFGFR-triggered network, suggesting that the
452
RSK2 NFB that operates above RAS contributes to MAPK signaling robustness.
453
16
454
Figure 5: Direct optogenetic activation of RAS highlights different ERK dynamics phenotypes
455
than optoFGFR input. (A) Schematic representation of ERK signaling induced by optoSOS versus
456
optoFGFR input. (B) ERK dose responses under transient optoSOS input consisting of different
457
numbers of repeated 470 nm pulses (1x, 2x, 3x, 4x and 5x pulses applied at 20-second intervals, D =
458
0.6 J/cm2). Repeated pulses are depicted as a single stimulation (blue bar). (C) ERK responses to
459
optoSOS inputs consisting of 5 repeated 470 nm light pulses delivered every 20, 10, 5 and 2 minutes
460
respectively (D = 0.6 J/cm2). (D) ERK responses to increasing light doses of sustained optoSOS input
461
consisting of 2-minute interval input, each input made of 5 repeated light pulses. Cells were divided in
462
low and high optoSOS expression levels based on the log10 intensity of the optoSOS-mCitrine. (E)
463
Quantification of the maxPeak of single-cell ERK responses under sustained optoFGFR (Figure 2F, D
464
= 18 mJ/cm2) and optoSOS (Figure 5D, D = 0.6 J/cm2) input for low or high expression of each
465
optogenetic system (N = 40 cells per condition). Statistical analysis was done using a Wilcoxon test,
466
comparing each condition to each other (Nmin = 48 cells per condition, NS: non-significant, *<0.05,
467
**<0.005, ***<0.0005, ****<0.00005, FDR p-value correction method). (F) Z-Score evaluation of the
468
A
B
D
E
maxPeak
Time
ERK activity
0.5
1.0
optoFGFR optoSOS
1.5
NS
****
*
NS
low
high
Expression
ERK activity at maxPeak
0 J/cm2
0.05 J/cm2
0.6 J/cm2
0.01 J/cm2
ERK activity
1.00
0.25
0.50
0.75
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
1.00
0.25
0.50
0.75
Time [min]
Light dose
0 x
0.25
0.50
0.75
1.00
Time [min]
ERK activity
0
10
20
30
40
C
2 min interval
10 min interval
20 min interval
25
50
75
5
25
50
75
5
ERK activity
0.25
0.50
0.75
1.00
25
50
75
5
25
50
75
5
5 min interval
Time [min]
0
10
20
30
40
0
10
20
30
40
0
10
20
30
40
Number of pulses [0.6 J/cm2]
1 x
3 x
5 x
SOScat
RSK2
FGFRcyto
FGFRcyto
SOS
MEK
GRB2
FRS2
RAF
RAS
ERK
P
P
P
P
P
Low
optoSOS
High
optoSOS
F
G
maxPeak
Time
ERK activity
baseline
ERKpostStim
Features
CTRL
NF1
ERK2
DUSP6
CRAF
RSK2
ERK activity
Time [min]
1.00
0.25
0.50
0.75
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
20
40
0
10
30
1.00
0.25
0.50
0.75
Low
optoSOS
High
optoSOS
20
40
0
10
30
DUSP22
DUSP26
DUSP22
DUSP16
DUSP14
DUSP10
DUSP9
DUSP8
DUSP6
DUSP4
DUSP3
DUSP2
DUSP1
PEA15
RSK2
ERK2
ERK1
MEK2
MEK1
CRAF
BRAF
ARAF
PP2A
KSR1
CNKSR1
YWHAG
YWHAZ
RKIP
RRAS
NRAS
KRAS
HRAS
RAP1B
RAP1A
NF1
RASA1
ERK1+2
CTRL
baseline
maxPeak
ERKpostStim
2
-2
zScore to CTRL
NS
17
baseline, maxPeak and ERKpostStim of single-cell ERK responses under sustained high optoSOS
469
input (D = 0.6 J/cm2). The z-score was calculated by comparing each RNAi perturbation to the CTRL
470
KD (Nmin = 33 cells per treatment, from 3 technical replicates). Non-significant (NS) results are in grey
471
(see Figure EV4F for statistical results). (G) Single-cell ERK trajectories for low and high optoSOS cells
472
for selected RNAi perturbations (N = 40 randomly selected out of at least 193 trajectories from 3
473
technical replicates).
474
Perturbation of the RSK2-mediated NFB increases the efficiency of RAF, MEK
475
and ERK targeting drugs
476
To further investigate the role of the RSK2-mediated NFB in MAPK signaling
477
robustness, we performed dose response experiments using different MAPK inhibitors
478
and compared ERK amplitudes evoked by optoFGFR (RSK2-feedback dependent)
479
versus optoSOS (RSK2-feedback independent) input, as well as optoFGFR input in
480
absence/presence of RSK2 perturbation. We used drugs targeting B/CRAF (RAF709),
481
MEK (U0126) and ERK (SCH772984). We evaluated the inhibition efficiency by
482
measuring ERK amplitude at a fixed time point, focusing on ERK responses evoked
483
by high optoFGFR or optoSOS inputs to limit the single-cell heterogeneity due to
484
expression variability of the optogenetic actuator. All inhibitors led to a stronger
485
reduction of ERK amplitude and EC50 in response to optoSOS versus optoFGFR input
486
(Figure 6A-C, EV5A, Appendix Table S5). Visual evaluation of ERK amplitude
487
distributions (Figure 6B) and quantification of their standard deviations (Figure 6D)
488
revealed more compact ERK amplitude distributions in presence of increasing drug
489
concentrations in response to optoSOS versus optoFGFR input. This suggests a more
490
homogeneous drug inhibition in the cell population in response to optoSOS input. We
491
then performed the identical experiments in CTRL or RSK2 KD cells in response to
492
optoFGFR input (Figure 6E-H, EV5B, Appendix Table S6). RSK2 KD led to increased
493
inhibition of ERK amplitudes, decreased EC50, and more compact ERK amplitude
494
distributions in response to increasing drug concentration than in CTRL KD cells.
495
Similar results were observed when the RSK2-mediated feedback was inhibited using
496
the RSK inhibitor SL0101 (Smith et al. 2005) (Figure EV5C-F, Appendix Table S7).
497
Thus, inhibition of the RSK2-mediated NFB sensitizes ERK responses to RAF, MEK
498
or ERK drug perturbations. Note that drug mediated ERK amplitude inhibition was
499
stronger in response to optoSOS input than to optoFGFR input with RSK2 KD or RSK
500
inhibition, suggesting that additional mechanisms to the RSK2-mediated feedback
501
contribute to MAPK signaling robustness. However, our results suggest that
502
perturbation of the RSK2-mediated feedback can be exploited to enhance the
503
efficiency of MAPK-targeting drugs, reducing ERK amplitudes more homogeneously
504
across the cell population.
505
18
506
Figure 6: Perturbation of the RSK2-mediated NFB increases the efficiency of RAS, MEK and ERK
507
targeting drugs. (A) Schematic representation of the optoFGFR (RSK2-mediated feedback
508
dependent) and optoSOS (RSK2-mediated feedback independent) systems targeted with the B/CRAF
509
(RAF709), the MEK (U0126) or the ERK (SCH772984) inhibitor. (B) Single-cell ERK amplitudes from
510
sustained high optoFGFR input (D = 18 mJ/cm2) or optoSOS input (D = 0.6 J/cm2) under different
511
concentrations of the MAPK inhibitors, extracted at a fixed time point (tfixed optoFGFR = 15 minutes, tfixed
512
optoSOS = 10 minutes, N = 200 cells with high optoFGFR or optoSOS expression per condition randomly
513
selected from 3 technical replicates). (C) A Hill function was fit to the normalized mean ERK activity as
514
shown in (B) (Nmin = 200 cells per condition). Shaded area indicates the 95% CI and dashed lines the
515
EC50. (D) Normalized standard deviation of ERK amplitudes shown in (B) (Nmin = 200 cells per
516
condition). (E) Schematic representation of the optoFGFR system treated with CTRL KD (RSK2-
517
mediated feedback dependent) or RSK2 KD (RSK2-mediated feedback independent) targeted with the
518
B/CRAF (RAF709), the MEK (U0126) or the ERK (SCH772984) inhibitor. (F) Single-cell ERK
519
amplitudes from sustained high optoFGFR input (D = 18 mJ/cm2) under different concentrations of the
520
MAPK inhibitors, extracted at a fixed time point (tfixed optoFGFR = 15 minutes, N = 70 cells with high
521
optoFGFR expression per condition (apart from RSK2 KD + 0 μM U0126 (32 cells), randomly selected
522
from 2 technical replicates for RSK2 KD and 1 replicate for CTRL KD). (G) A Hill function was fit to the
523
normalized mean ERK activity as shown in (F) (Nmin = 32 cells per perturbation). Shaded area indicates
524
the 95% CI and dashed lines the EC50. (H) Normalized standard deviation of ERK amplitudes shown in
525
(F) (Nmin = 32 cells per perturbation).
526
B
optoFGFR
optoSOS
A
RSK2
SOS
MEK
GRB2
FRS2
RAF
RAS
ERK
P
P
P
SCH772984
RAF709
U0126
optoSOS
input
optoFGFR
input
RAF709
U0126
SCH772984
EC50 ��������
EC50 ��������
EC50 =��������
EC50���������
EC50 =�������
EC50���������
C
EC50 ��������
EC50 ��������
EC50 =��������
EC50���������
EC50 =�������
EC50���������
D
H
Concentration [�M]
optoFGFR
optoSOS
Rescaled
ERK activity at tfixed
0.0
0.5
1.0
0
2
5 10 20 50
1
0
2
5 10 20 50
1
10
0
2
1
5
Concentration [�M] (log10)
optoFGFR
CTRL KD
RSK2 KD
Rescaled
ERK activity at tfixed
Concentration [�M] (log10)
0.0
0.5
1.0
0
2
5 10 20 50
1
0
2
5 10 20 50
1
10
0
2
1
5
E
RSK2
SOS
MEK
GRB2
FRS2
RAF
RAS
ERK
P
P
P
SCH772984
RSK2 KD
RAF709
U0126
optoFGFR
input
F
G
optoFGFR
optoSOS
0
1
Normalized
SD
Concentration [�M]
0
2
5
10
20
50
0
2
5
10
20
50
0
0.5
1
2
5
10
CTRL KD
RSK2 KD
Concentration [�M]
0
2
5
10
20
50
0
2
5
10
20
50
0
0.5
1
2
5
10
0
1
Normalized
SD
ERK activity at tfixed
CTRL KD
RSK2 KD
optoFGFR
0.5
1.0
0
2
5
10
20
50
0
2
5
10
20
50
0
0.5
1
2
5
10
Concentration [�M]
RAF709
U0126
SCH772984
ERK activity at tfixed
0.5
1.0
0
2
5
10
20
50
0
2
5
10
20
50
0
0.5
1
2
5
10
19
Targeting the RSK2-mediated feedback in an ErbB2 oncogenic signaling model
527
increases MEK inhibition efficiency
528
The results above suggested an important role of the RSK2-mediated feedback in
529
MAPK signaling robustness against node perturbation in response to optogenetic
530
inputs in NIH3T3 cells. To test if this feedback also contributes to MAPK signaling
531
robustness in a disease-relevant system, we evaluated its function in MCF10A cells,
532
a breast epithelium model, using either wild-type (WT) or overexpressing ErbB2
533
(referred to as ErbB2over) recapitulating the ErbB2 amplification observed in 20% of all
534
breast cancers (Arteaga and Engelman 2014; Yarden and Pines 2012). We chose this
535
specific model system because ErbB2 amplification leads to constitutive RTK input on
536
the MAPK network, while retaining an intact downstream feedback structure (Figure
537
7A). This contrasts with other cancer model systems in which additional mutations
538
might lead to RAS or RAF overactivation, and thus disrupt the feedback architecture.
539
Further, previous work has highlighted the role of NFBs in ERK pulse formation in
540
MCF10A cells (Kochańczyk et al. 2017), suggesting that EGFR and ErbB2 trigger a
541
MAPK network with similar feedback circuitry as optoFGFR.
542
As described before (Albeck et al. 2013), WT cells displayed asynchronous low
543
frequency ERK pulses in the absence of EGF, and high frequency ERK pulses in
544
presence of EGF (Figure 7B). In marked contrast, ErbB2over cells displayed high
545
frequency ERK pulses, even in the absence of EGF (Figure 7C). To investigate the
546
role of the RSK2-mediated feedback in MAPK signaling robustness, we performed a
547
U0126 dose response in EGF-stimulated MCF10A WT cells and found that 3 µM
548
U0126 decreased ERK amplitude without fully suppressing the response (Figure
549
EV5G,H). As observed in response to optogenetic inputs, RSK inhibition with 50 µM
550
SL0101 led to a mild reduction in ERK amplitude. However, in combination with 3 µM
551
U0126, ERK amplitude was decreased to the level of unstimulated cells. Similar
552
results were observed in ErbB2over cells (Figure EV5I), suggesting that RSK2
553
perturbation increases the sensitivity of ERK responses to MEK inhibition.
554
As averaging ERK dynamics can hide asynchronous single-cell signaling activity, we
555
further investigated the effect of these perturbations on single-cell trajectories using
556
CODEX (Jacques et al. 2021) (see Material and methods for details). For WT cells, a
557
tSNE projection of the CNN features built from single-cell ERK trajectories hinted that
558
the CNN was able to construct features separating the treatments into well-defined
559
clusters (Figure 7D, EV5J). Clustering of the CNN features confirmed the existence of
560
discrete ERK dynamics clusters (Figure 7E) whose composition correlated with the
561
treatments (Figure 7F). To characterize the dynamics captured by each cluster, we
562
extracted the medoid trajectory and its 4 closest neighbors from each cluster (Figure
563
7G). This revealed that non-stimulated cells mostly display low frequency ERK activity
564
pulses (cluster 4) or absence of pulses (cluster 5). Cells stimulated with EGF without
565
inhibitor displayed ERK pulses of high amplitude (cluster 1). SL0101-treated cells
566
displayed a sustained ERK activation at low amplitude (cluster 3). U0126-treated cells
567
still displayed prominent ERK pulses but at a lower amplitude than EGF-treated cells
568
in absence of drug (cluster 2). Finally, in cells treated with both U0126 and SL0101,
569
20
almost no ERK activity was observed (cluster 5). For ErbB2over cells, we observed that
570
the CNN features were forming a more continuous space with less distinct clusters
571
(Figure 7H,I, EV5K). A heterogeneous mix of ERK trajectory clusters was observed
572
for the different treatments (Figure 7J,K). Untreated cells mostly displayed high
573
frequency ERK pulses that were either sharp (cluster 3) or wider (cluster 2). SL0101-
574
treated cells were almost equally shared between cluster 1 (relatively flat high
575
amplitude ERK trajectories), cluster 2, cluster 4 (low amplitude ERK pulses) and
576
cluster 5 (low baseline activity). U0126 led to a less heterogeneous mix mostly
577
consisting of ERK trajectories from cluster 4 and 5. The U0126/SL0101 combination
578
shifted most cells to cluster 5, indicating an efficient inhibition of ERK activity at a
579
suboptimal U0126 concentration. Thus, RSK inhibition also sensitizes the MAPK
580
network to U0126-mediated MEK inhibition both in MCF10A WT and ErbB2over cells.
581
582
583
A
H
K
yTSNE
P
P P
P
ErbB2
ErbB2over
RSK2
SOS
MEK
GRB2
RAF
RAS
ERK
P
P
P
SL0101
U0126
no inhibitor
xTSNE
yTSNE
D
G
WT
P
P
EGFR
EGF
B
+ 10 ng/ml EGF
no EGF
no EGF
ErbB2over
WT
ERK activity
Time [hours]
0
5
10
15
1.0
0
5
10
15
0
5
10
15
1.5
0.5
C
Time [hours]
50 �M SL0101
3 �M U0126
3 �M U0126
+ 50 �M SL0101
no inhibitor
xTSNE
yTSNE
xTSNE
Treatment
50 �M SL0101
3 �M U0126
3 �M U0126
+ 50 �M SL0101
no inhibitor
Treatment
CNN
clusters
E
F
I
J
yTSNE
xTSNE
0.0
0.5
1.0
50 �M SL0101
3 �M U0126
3 �M U0126
+ 50 �M SL0101
no inhibitor
CNN clusters proportion
no inhibitor
0.0
0.5
1.0
50 �M SL0101
3 �M U0126
3 �M U0126
+ 50 �M SL0101
no inhibitor
CNN clusters proportion
Cluster 1
Cluster 2
Cluster 3
Cluster 4
Cluster 5
Time [hours]
ERK activity
0
5
10
15
0
5
10
15
0
5
10
15
0
5
10
15
0.5
1.0
1.5
2.0
0
5
10
15
Cluster 1
Cluster 2
Cluster 3
Cluster 4
Cluster 5
Time [hours]
ERK activity
0
5
10
15
0
5
10
15
0
5
10
15
0
5
10
15
0.5
1.0
1.5
2.0
0
5
10
15
1
2
3
4
5
+ EGF
- EGF
CNN
clusters
1
2
3
4
5
- EGF
+ EGF
- EGF
- EGF
21
Figure 7: Targeting the RSK2-mediated feedback in an ErbB2 oncogenic signaling model
584
increases MEK inhibition efficiency. (A) Schematic representation of MAPK signaling in response to
585
EGFR input in MCF10A WT cells or oncogenic ErbB2 input in ErbB2 overexpressing (ErbB2over) cells.
586
(B-C) Single-cell ERK responses in MCF10A WT cells without or with stimulation with 10 ng/ml EGF at
587
t = 30 minutes (B) and in unstimulated MCF10A ErbB2over cells (C). (D) tSNE projection of CODEX’s
588
CNN features from ERK trajectories of MCF10A WT cells without EGF stimulation, or with 10 ng/ml
589
EGF stimulation added at t = 30 min in absence of perturbation, with 50 μM SL0101, 3 μM U0126 or a
590
combination of both. (E) t-SNE projection of CODEX’s CNN features shown in (D) colored by the CNN
591
feature clusters. Black diamonds indicate the position of the medoid and its 4 closest neighbor
592
trajectories for each cluster. (F) Distribution of the trajectories in the CNN features clusters per
593
treatment. Colors are as shown in (E). (G) Medoid trajectories and their 4 closest neighbors per cluster
594
highlighted in (E) (black diamonds). (H) tSNE projection of CODEX’s CNN features from ERK
595
trajectories of non-stimulated ErbB2 overexpressing cells without perturbation, with 50 μM SL0101, 3
596
μM U0126 or a combination of both. (I) t-SNE projection of CODEX’s CNN features shown in (H) colored
597
by the CNN feature clusters. Black diamonds indicate the position of the medoid and its 4 closest
598
neighbor trajectories for each cluster. (J) Distribution of the trajectories in the CNN features clusters per
599
treatment. Colors are as shown in (I). (K) Medoid trajectories and their 4 closest neighbors per cluster
600
highlighted in (I) (black diamonds).
601
22
Discussion
602
Optogenetic actuator-biosensor circuits allow for feedback structure mapping
603
in the MAPK network
604
ERK dynamics is crucial for fate decisions. Yet, the topology of the network enabling
605
the cells to sense different inputs and convert this information into finely tuned ERK
606
dynamics remains poorly understood. We developed genetic circuits consisting of
607
optogenetic actuators and an ERK biosensor (Figure 1A, 5A) that allow for a large-
608
scale interrogation of single-cell ERK dynamics and investigated the effects of 50 RNAi
609
perturbations targeting components of the MAPK signaling network (Figure 4A). In our
610
optoFGFR screen, we only observed a small number of penetrant ERK dynamics
611
phenotypes (Figure 4D-F), implying that the MAPK network can buffer against
612
perturbations of most of its components. We cannot exclude that in some cases, even
613
on the relatively short 72 hours timescale of the RNAi experiment, compensation by
614
upregulation of specific nodes might occur. However, our data suggest that the MAPK
615
network topology allows for MAPK signaling robustness – the production of consistent
616
ERK outputs in presence of node perturbation. This might emerge from isoform
617
redundancy for multiple nodes in the network, as observed for single or combined ERK
618
isoforms perturbation (Figure 4B), but also for individual perturbation of RAS, RAF,
619
MEK isoforms. Another mechanism might involve NFBs that have been shown to
620
decrease the network sensitivity to node perturbation (Sturm et al. 2010; Fritsche-
621
Guenther et al. 2011). Our screen suggested that RSK2, that mediates a NFB from
622
ERK to SOS (Douville and Downward 1997; Saha et al. 2012), both regulates ERK
623
dynamics (Figure 4D-F) and plays a role in MAPK signaling robustness (Figure 6E-H).
624
In addition, our data suggest that the well-studied ERK-RAF NFB, which has been
625
shown to buffer against MAPK node perturbations (Sturm et al. 2010; Fritsche-
626
Guenther et al. 2011), also regulates ERK dynamics (Figure 4F). We therefore
627
speculate that both feedbacks operate simultaneously in the MAPK network, and act
628
at multiple levels within the cascade to warrant MAPK signaling robustness.
629
Consistently with this hypothesis, we observed that the optoSOS-triggered network,
630
which is not under the RSK2 NFB regulation, shows an increased sensitivity in ERK
631
amplitude to perturbation of some nodes (Figure 5F,G). Indeed, ERK2 and CRAF
632
perturbations, which led to loss of ERK oscillations, had relatively mild amplitude
633
phenotypes in response to optoFGFR input, while both perturbations led to strong ERK
634
amplitude phenotypes in response to optoSOS input. Because these phenotypes were
635
not observed with other ERK and RAF isoforms, we propose that ERK2 and CRAF
636
are the isoforms involved in the classic ERK-RAF NFB. Additional feedbacks have
637
been reported within the MAPK network (Langlois et al. 1995; Lake et al. 2016;
638
Kochańczyk et al. 2017), and even if they have not been highlighted in our screen,
639
they might also regulate ERK dynamics.
640
While providing the experimental throughput to perturb and analyze ERK dynamics at
641
scale, optoFGFR, that lacks an ectodomain, evoked different ERK dynamics than
642
endogenous RTKs such as FGFR and EGFR (Figure 3A,B compared to Figure 2F).
643
23
These different ERK dynamics emerge likely because of receptor-level interactions
644
that involve competition of bFGF for FGFR and heparan sulfate proteoglycan co-
645
receptors (Kanodia et al. 2014; Blum et al. 2019) in the case of FGFR, or receptor
646
endocytosis in the case of EGFR (Kiyatkin et al. 2020; Gerosa et al. 2020). Our
647
combined modeling and experimental approach suggested that optoFGFR and EGFR
648
share similar downstream MAPK network circuitries and NFBs (Figure 3C-G).
649
OptoFGFR therefore provides a simplified system that allowed us to focus on
650
intracellular feedback structures, without confounding receptor level regulations. Our
651
Bayesian inference modeling approach, that is parameter agnostic, could provide
652
simple intuitions about the receptor-level and negative feedback structures that shape
653
ERK dynamics in response to optoFGFR and EGFR inputs. However, even if we had
654
access to many ERK dynamics phenotypes, our modeling approach did not allow us
655
to explore more sophisticated MAPK network topologies such as the presence of two
656
NFBs or multiple node isoforms. We interpreted our data using some of the feedback
657
structures that have been previously experimentally documented and modelled but
658
cannot formally exclude that the observed ERK dynamics emerge from different
659
network structures. In the future, information about the different nodes and their
660
dynamics might allow to further constrain the model topology and parameter space,
661
and hopefully address this limitation.
662
Additional novel insights into regulation of ERK dynamics
663
Our optoFGFR and optoSOS screens provided new system-wide insights into the
664
regulation of the MAPK network. Strikingly, the same perturbations induced different
665
ERK dynamics phenotypes in the optoFGFR and optoSOS screens. This might occur
666
because some regulators target the MAPK network at multiple levels, differently
667
affecting ERK responses triggered with optoFGFR or optoSOS inputs. Additionally, as
668
the two optogenetic systems are under the regulation of one versus two
669
simultaneously occurring NFBs, they might have different sensitivities to perturbations,
670
as discussed above.
671
With respect to the optoFGFR system, GRB2 KD led to a reduction of ERK amplitude
672
(Figure 4D,E). GRB2 acts as the RTK-proximal adaptor to activate SOS (Chardin et
673
al. 1993; Belov and Mohammadi 2012). As GRB2 operates at the start of the cascade,
674
outside of most NFBs, heterogeneity in its expression levels might be less easily
675
buffered out. PLCG1 KD increased damped oscillatory behavior (Figure 4E,F).
676
Phospholipase Cɣ1 activates calcium signaling, which has itself been shown to
677
regulate RAS/MAPK signaling in a calcium spike frequency-dependent manner
678
(Kupzig et al. 2005; Cullen and Lockyer 2002). Further investigation will be required
679
to understand the significance of this crosstalk. RKIP KD resulted in higher ERK
680
baseline and slower ERK adaptation post stimulation, without affecting ERK amplitude
681
(Figure 4D). RKIP (RAF kinase inhibitor protein) prevents MEK phosphorylation by
682
CRAF (Yeung et al. 2000), suggesting that RKIP-dependent regulation is specifically
683
involved in keeping basal ERK activity low. With respect to phosphatases, none of
684
their perturbations led to a strong phenotype such as sustained ERK dynamics post
685
24
stimulation for example. The strongest phenotype was observed for PP2A KD that led
686
to increased ERK amplitude, baseline, and slower adaptation (Figure 4D, EV3A). This
687
might occur because the protein phosphatase 2A is an ubiquitous phosphatase that
688
acts at multiple levels by dephosphorylating SHC1, MEK1, MEK2, ERK1 and ERK2,
689
as well as a large number of other proteins (Junttila et al. 2008; Saraf et al. 2010). The
690
observation that in optoFGFR-low PP2A KD cells, ERK dynamics displayed increased
691
amplitude but still oscillated rather than exhibiting sustained behavior, suggests that
692
NFBs might buffer against the loss of phosphatase regulation to some extent.
693
Perturbation of the nuclear DUSPs, DUSP1,2,4, the atypical DUSP3 and most strongly
694
the cytosolic DUSP6 (Patterson et al. 2009) led to higher ERK baseline, reduced
695
adaptation, with only limited effects on amplitude (Figure 4D, EV3A). Consistently,
696
DUSP6 has previously been proposed to pre-emptively dephosphorylate MAPKs to
697
maintain low ERK activity baseline levels at resting state (Huang and Tan 2012). Our
698
results indicate that perturbation of single DUSPs might not be compensated by the
699
others, suggesting that individual DUSPs might regulate specific substrates within the
700
MAPK network. Except for DUSP6, KD of the different DUSPs did not significantly
701
affect oscillatory ERK behavior in optoFGFR-low cells (Figure 4F), suggesting that
702
they are not involved in the MAPK feedback circuitry that operates on timescales of
703
minutes.
704
The optoSOS screen revealed stronger ERK amplitude phenotypes, especially for
705
ERK2 and CRAF KD (Figure 5F versus 4D). Unlike for optoFGFR input, RSK2 KD did
706
not result in slower ERK adaptation, suggesting that ERK responses triggered by the
707
optoSOS input are not regulated by the RSK2-mediated NFB. However, RSK2 KD led
708
to a reduction of ERK amplitude, also observed to a lesser extent in response to
709
optoFGFR input, suggesting a role of RSK2 in ERK amplitude regulation downstream
710
of RAS. With respect to phosphatases, PP2A KD led to decreased amplitude, a
711
different phenotype than in response to optoFGFR input. This might occur because of
712
the broad specificity PP2A phosphatase, which might lead to different phospho-
713
proteomes in response to optoSOS versus optoFGFR input. Similar phenomena might
714
apply for most of the DUSPs.
715
The RSK2-mediated feedback can be targeted to potently inhibit oncogenic
716
ErbB2 signaling
717
Our data suggest that the RSK2-mediated NFB is important for MAPK signaling
718
robustness downstream of our prototypic optoFGFR RTK (Figure 6). We found that
719
the RSK2-mediated NFB likely also operates downstream of EGFR and oncogenic
720
ErbB2 signaling in MCF10A cells (Figure 7). In response to EGF stimulation, or ErbB2
721
overexpression, a subset of RSK-inhibited cells displayed wider ERK pulses,
722
suggesting that the RSK2 NFB is also involved in ERK adaptation in this system
723
(Figure 7G cluster 3, Figure 7K cluster 1 and 2). Further, RSK inhibition led to a high
724
heterogeneity of ERK dynamics within the cell population especially visible in the case
725
of ErbB2 overexpressing cells (Figure 7J), which might result from the reduced ability
726
of the MAPK network to cope with nodes expression noise in absence of the RSK2
727
25
NFB. In EGF-treated cells, combination of RSK and suboptimal MEK inhibition led to
728
strong and homogeneous ERK inhibition (Figure 7E-G, cluster 5). In the ErbB2
729
overexpressing cells, combined RSK/MEK inhibition shifted most of the cell population
730
to flat, low amplitude ERK dynamics, enabling to further inhibit a large number of cells
731
when compared to suboptimal MEK inhibition only (Figure 7I-K, cluster 5). These
732
results suggest that pharmacological inhibition of the RSK2-mediated NFB can be
733
used to reduce MAPK signaling robustness, sensitizing the network to MEK
734
perturbation. Such non-trivial drug combinations might allow for homogeneous
735
inhibition of ERK dynamics in most of the cells in a population. This homogeneous
736
inhibition might mitigate the emergence of drug-tolerant persister cells from cell
737
subpopulations that display residual ERK activity in response to inhibition of a single
738
node. Our results imply that efficient pharmacological inhibition of the MAPK network
739
requires precise understanding of its topology. The RSK2 NFB is an example of a
740
druggable node that can be exploited to target MAPK signaling robustness.
741
742
Our scalable experimental pipeline provides new insight into the MAPK network wiring
743
that produces ERK dynamics. However, our perturbation approach only highlighted
744
very subtle ERK dynamics phenotypes, precluding a complete understanding of the
745
MAPK network. We envision that this will require more precise knowledge about the
746
dynamics of MAPK network nodes and their interactions in response to defined inputs
747
and perturbations. Such data can now be produced at scale using optogenetic
748
actuator/biosensor circuits as those we describe in this work. This information might
749
allow for faithful parametrization of more complex models. With the increasing amount
750
of optogenetic actuators and biosensors available, similar genetic circuits could also
751
be designed to study the dynamics of other signaling pathways at scale.
752
26
Materials and methods
753
754
Cell culture and reagents
755
NIH3T3 cells were cultured in DMEM high glucose medium with 5% fetal bovine
756
serum, 4 mM L-glutamine, 200 U/ml penicillin and 200𝜇g/ml streptomycin at 37°C with
757
5% CO2. All imaging experiments with NIH3T3 were done in starving medium
758
consisting of DMEM high glucose supplemented with 0.5% BSA (Sigma), 200 U/ml
759
penicillin, 200 μg/ml streptomycin and 4 mM L-Glutamine. MCF10A human mammary
760
cells were cultured in DMEM:F12 supplemented with 5% horse serum, 20 ng/ml
761
recombinant human EGF (Peprotech), 10 μg/ml insulin (Sigma), 0.5 μg/ml
762
hydrocortisone (Sigma), 200 U/ml penicillin and 200 μg/ml streptomycin. All imaging
763
experiments with MCF10A were done in starving medium consisting in DMEM:F12
764
supplemented with 0.3% BSA, 0.5 μg/ml hydrocortisone, 200 U/ml penicillin and 200
765
μg/ml streptomycin. For growth factor stimulations, we used human EGF (AF-100,
766
Peprotech) and human basic FGF (F0291, Sigma). Chemical perturbations were done
767
with SU-5402 (SML0443, Sigma), RAF709 (HY-100510, Lucerna Chem), U0126
768
(S1102, Selleck chemicals, Lubio), SCH772984 (HY-50846, Lucerna-Chem), SL0101
769
(559285, Sigma), Cyclosporine A (10-1119, Lucerna-chem) and Ionomycin (sc-3592,
770
Santa Cruz). Selection of the cells post transfection was done using Puromycin
771
(P7255, Sigma), Blasticidin S HCI (5502, Tocris) and Hygromycin B (sc-29067, Lab
772
Force).
773
774
Plasmids and stable cell line generation
775
The optoFGFR construct was a gift from Won Do Heo (Addgene plasmid # 59776)
776
(Kim et al. 2014). It consists of the myristoylated FGFR1 cytoplasmic region fused with
777
the PHR domain of the cryptochrome2 and tagged with mCitrine. It was cloned in a
778
lentiviral backbone for stable cell line generation. A modified version of the optoFGFR
779
tagged with the red fluorophore mScarlet (Bindels et al. 2017) was cloned in a
780
PiggyBac plasmid pPBbSr2-MCS (blasticidin resistance), a gift from Kazuhiro Aoki.
781
The optoSOS construct is a modified version of the tRFP-SSPB-SOScat-P2A-iLID-
782
CAAX (Addgene plasmid #86439) (Johnson et al. 2017), in which we replaced the
783
tRFP by mCitrine. The construct was cloned in the pPB3.0.Puro, an improved
784
PiggyBac plasmid generated in our lab with puromycin resistance. The ERK-KTR-
785
mRuby2 and ERK-KTR-mTurquoise2 reporters were generated by fusing the ERK
786
Kinase Translocation Reporter (ERK-KTR) (Regot et al. 2014) with mRuby2 (Lam et
787
al. 2012) or mTurquoise2 (Goedhart et al. 2012). The nuclear marker H2B-miRFP703
788
is a fusion of the human H2B clustered histone 11 (H2BC11) with the monomeric near-
789
infrared fluorescent protein miRFP703 (Shcherbakova et al. 2016) (Addgene plasmid
790
#80001). ERK-KTR-mRuby2, ERK-KTR-mTurquoise2 and H2B-miRFP703 were
791
cloned in the PiggyBac plasmids pPB3.0.Hygro, pSB-HPB (gift of David Hacker,
792
EPFL, (Balasubramanian et al. 2016)) and pPB3.0.Blast, respectively. All constructs
793
in PiggyBac plasmids were co-transfected with the helper plasmid expressing the
794
transposase (Yusa et al. 2011) for stable insertion using the jetPEI (Polyplus)
795
transfection reagent for NIH3T3 cells or FuGene (Promega) transfection reagent for
796
27
MCF10A cells. After antibiotic selection, NIH3T3 cells were FACS-sorted to generate
797
stable cell lines homogeneously expressing the biosensors. In the case of MCF10A
798
cells, clones with uniform biosensor expression were isolated. To generate ErbB2
799
overexpressing MCF10A cells, lentiviral transduction using a pHAGE-ERBB2
800
construct (a gift from Gordon Mills & Kenneth Scott, Addgene plasmid #116734, (Ng
801
et al. 2018)) was performed in the presence of 8 μg/ml polybrene (TR1003, Sigma) in
802
cells already expressing H2B-miRFP703 and ERK-KTR-mTurquoise2. Cells were
803
further selected with 5 μg/ml puromycin.
804
805
Live imaging of ERK dynamics
806
NIH3T3 cells were seeded in 96 well 1.5 glass bottom plates (Cellvis) coated with 10
807
μg/ml Fibronectin (Huber lab) using 1.5 x 103 cells/well and incubated for 24 hours.
808
MCF10A cells were seeded in 24-well 1.5 glass bottom plates (Cellvis) coated with 5
809
μg/ml Fibronectin (Huber lab) at 1 x 105 cells/well and incubated for 48 hours. NIH3T3
810
cells were washed with PBS and incubated in starving medium for 4 hours in the dark
811
before starting the experiment. MCF10A cells were starved for 7 hours before starting
812
the experiments. In experiments involving drug perturbations, cells were incubated for
813
2 hours (or 1 hour in MCF10A experiments) with the inhibitor(s). Imaging was
814
performed with an epifluorescence Eclipse Ti inverted fluorescence microscope
815
(Nikon) using a Plan Apo air 20x (NA 0.8) objective. Nikon Perfect Focus System
816
(PFS) was used to keep cells in focus throughout the experiment. Illumination was
817
done with a SPECTRA X light engine (Lumencor) with the following filters (Chroma):
818
mTurquoise2: 440 nm LED, 470lp, 69308 CFP/YFP/mCherry-ET, CFP 458-482;
819
mCitrine: 508 nm LED, ET500/20x, 69308bs, ET535/30m; mRuby2 and mCherry: 555
820
nm LED, ET575/25x, 69008bs, 59022m, miRFP703: 640 nm LED, ET640/30x,
821
89100bs Sedat Quad, 84101m Quad. Images were acquired with an Andor Zyla 4.2
822
plus camera at a 16-bit depth. Image acquisition and optogenetic stimulation were
823
controlled with the NIS-Element JOBS module. For NIH3T3 experiments, ERK-KTR-
824
mRuby2 and H2B-miRFP703 were acquired at 1-minute interval and 470 nm light
825
inputs were delivered at specific frequencies and intensities (see below). MCF10A
826
image acquisition was performed at 5-minute time resolution. Growth factor
827
stimulations were done by manually pipetting EGF and bFGF during the experiment.
828
We used mCitrine intensity to quantify the expression level of the optogenetic
829
constructs. However, as mCitrine excitation leads to optoFGFR or optoSOS activation,
830
we acquired one frame with the ERK-KTR-mRuby2, the H2B-miRFP703 and the
831
mCitrine-tagged optoFGFR or optoSOS only at the end of each NIH3T3 experiments.
832
All experiments were carried on at 37°C with 5% CO2.
833
834
Optogenetic stimulation
835
Light stimulations were delivered with a 470 nm LED light source that was hardware-
836
triggered by the camera to generate light pulses of reproducible duration. Light
837
stimulations of defined intensity and duration were programmed to be automatically
838
delivered at specific timepoints. To define the dose of light received by the cells, we
839
measured the 470 nm light intensity at the focal plane using an optical power meter
840
28
(X-Cite Power Meter, Lumen Dynamics Group) and converted this value to a power
841
density as
842
843
𝐿𝑖𝑔ℎ𝑡 𝑝𝑜𝑤𝑒𝑟 𝑑𝑒𝑛𝑠𝑖𝑡𝑦 = 𝐿𝑖𝑔ℎ𝑡 𝑖𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦 ×
1
𝜋 × 5
𝐹𝑁
2 × 𝑀𝑎𝑔𝑛𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛=
! >𝑚𝑊
𝑐𝑚!A
844
845
with FN = 18 mm. The obtained value was then multiplied by the duration of the pulse
846
to obtain the dose of light received by the cells for each light pulse.
847
848
𝐿𝑖𝑔ℎ𝑡 𝑑𝑜𝑠𝑒 (𝐷) = 𝐿𝑖𝑔ℎ𝑡 𝑝𝑜𝑤𝑒𝑟 𝑑𝑒𝑛𝑠𝑖𝑡𝑦 × 𝑃𝑢𝑙𝑠𝑒 𝑑𝑢𝑟𝑎𝑡𝑖𝑜𝑛 = >𝑚𝑊 × 𝑠
𝑐𝑚!
A = > 𝑚𝐽
𝑐𝑚!A
849
850
For stimulation of the optoFGFR cells, the 470 nm LED intensity was limited to a low
851
dose by combining a ZET470/10x filter and a ND filter 5% (Chroma). Transient
852
stimulations were done with a single pulse, while sustained stimulations were done
853
with single pulses delivered every 2 minutes. For stimulation of the optoSOS cells, we
854
used the 470 nm LED with a ET470/24x filter (no ND filter). Transient stimulations
855
were done with 5 pulses repeated at 20-second intervals, while sustained stimulations
856
were done using 5 pulses repeated at 20-second intervals, delivered every 2 minutes.
857
858
Figures
System
Power density
Pulse duration
Dose
Stimulation
pattern
1B,C, Appendix
S1A, Appendix S1D
optoFGFR 180 mW/cm2
1 x 100 ms
18 mJ/cm2
transient
Appendix S1B,C
optoFGFR
(mScarlet)
> 180 mW/cm2
1 x 100 ms
> 18 mJ/cm2
transient
2A,B
optoFGFR variable
variable
variable
transient
2C, EV1A,B
optoFGFR 180 mW/cm2
1 x 100 ms
18 mJ/cm2
variable
2D,E, 3C,E, EV1K,
4B,D-F, EV2C,D,
EV3A-D, 5E, EV4C,
6B-D,F-H,
EV5A,B,D-F
optoFGFR 180 mW/cm2
1 x 100 ms
18 mJ/cm2
sustained
2F
optoFGFR variable
1 x 100 ms
variable
sustained
5B
optoSOS
6 W/cm2
variable x 100 ms
(20 sec interval)
0.6 J/cm2
transient
EV4A
optoSOS
variable
variable x 100ms
(20-sec interval)
variable
transient
5C
optoSOS
6 W/cm2
5 x 100 ms (20-
0.6 J/cm2
variable
29
sec interval)
5D
optoSOS
variable
5 x 100 ms (20-
sec interval)
variable
sustained
5E-G, EV4B-G, 6B-
D, EV5A
optoSOS
6 W/cm2
5 x 100 ms (20-
sec interval)
0.6 J/cm2
sustained
859
TIRF imaging of optoFGFR dynamics
860
Cells were seeded at a density of 1 x 103 per well in 96 well 1.5 glass bottom plates
861
(Cellvis) coated with 10 μg/ml Fibronectin (Huber lab) and incubated for 24 hours at
862
37°C with 5% CO2. Before imaging, cells were washed with PBS and incubated in
863
starving medium for 4 hours in the dark. Imaging was performed with an
864
epifluorescence Eclipse Ti inverted fluorescence microscope (Nikon) using a CFI
865
Apochromat TIRF 100x oil (NA 1.49). Images were acquired with an Andor Zyla 4.2
866
plus camera at a 16-bit depth. TIRF images were acquired with a 561 nm laser using
867
a ET575/25 filter in front of the ZT488/561rpc (Chroma) to prevent nonspecific
868
activation of the CRY2. MetaMorph software (Universal Imaging) was used for
869
acquisition. TIRF images of the optoFGFR-mScarlet were acquired at a 20-second
870
interval. Optogenetic stimulation was done using a 470 nm LED (SPECTRA X,
871
Lumencor) (Appendix Figure S1B). All experiments were carried on at 37°C with 5%
872
CO2.
873
874
Image processing pipeline
875
Nuclear segmentation was done in CellProfiler 3.0 (McQuin et al. 2018) using a
876
threshold-based approach of the H2B channel. In the case of MCF10A cells, nuclear
877
segmentation was preceded by prediction of nuclear probability using a random forest
878
classifier based on different pixel features available in Ilastik software (Berg et al.
879
2019). To measure the ERK-KTR fluorescence in the cytosol, the nuclear mask was
880
first expanded by 2 pixels to exclude the blurred edges of the nucleus. The new mask
881
was then further expanded by 4 pixels in a threshold-based manner to obtain a “ring”
882
area corresponding to the cytoplasmic ERK-KTR. ERK activity was obtained by
883
calculating the ratio between the average cytosolic pixel intensity and the average
884
nuclear pixel intensity. Single-cell tracking was done on nuclear centroids with
885
MATLAB using μ-track 2.2.1 (Jaqaman et al. 2008). The final images containing the
886
ERK-KTR-mRuby2, H2B-miRFP703 and the optoFGFR-mCitrine (or optoSOS-
887
mCitrine) channels were processed using the same CellProfiler settings as the time
888
lapse images. Intensity of the mCitrine was extracted under the ERK-KTR cytoplasmic
889
mask and used to classify cells into low or high expressors in a threshold-based
890
manner. For optoFGFR-evoked ERK responses, the threshold was defined empirically
891
to separate oscillatory and non-oscillatory ERK responses (low < -1.75 (log10 mCitrine
892
intensity) < high). For optoSOS-evoked ERK responses, the threshold was defined
893
empirically to separate cells with low or high ERK response amplitudes (low < -1.25
894
30
(log10 mCitrine intensity) < high). The same thresholds were kept across experiments
895
to compare low and high expressors.
896
The optoFGFR-mScarlet dimers/oligomers were segmented using the pixel
897
classification module from Ilastik (Berg et al. 2019). OptoFGFR dimers, cell
898
background and trafficking vesicles were manually annotated on images before and
899
after the light stimulation. A probability map of the optoFGFR dimers classification was
900
exported as TIFF for each frame. We then computed the mean of pixel intensities from
901
the binarized mask obtained with Ilastik using Fiji (Appendix Figure S1C).
902
903
Quantification of ERK activity
904
We wrote a set of custom R scripts to automatically calculate the ERK-KTR
905
cytoplasmic/nuclear ratio as a proxy for ERK activity for each single-cell, link single-
906
cell ERK responses with the corresponding optoFGFR/optoSOS intensity value and
907
export the corresponding ERK single-cell trajectories. For NIH3T3 data, outliers in
908
ERK single-cell trajectories were removed using a clustering-based approach
909
(https://github.com/pertzlab/Outlier_app). Trajectories with an ERK activity higher than
910
0.8 or lower than 0.2 before stimulation, above 1.6 during the whole experiment or
911
displaying single time point spiking values were removed. For MCF10A data,
912
trajectories with an ERK activity above 2 or shorter than 90% of the total experiment
913
duration were removed. All the R codes used for further analysis are available as
914
supplementary information (see Data availability section). Hierarchical clustering
915
analysis of single-cell trajectories (Figure 2D, EV3F,G, EV4B) was done using Time
916
Course Inspector (Dobrzyński et al. 2019).
917
918
Modeling
919
The model for the EGF and light stimulated ERK cascade is a kinetic model,
920
representing the EGF receptor, the inter-cellular proteins (RAS, RAF, MEK, ERK) as
921
well as a negative feedback (NFB) from ERK to RAF and the inactivation of the EGF
922
receptor in the form of endocytosis (Figure 3D). We explicitly modelled the ERK-KTR
923
readout through nuclear and cytosolic KTR. The initial fraction of cytosolic KTR is
924
estimated from the data through the parameter 𝑘𝑡𝑟"#"$. The KTR readout 𝑌(𝑡) was taken
925
to be the ratio of cytosolic KTR over nuclear KTR with additive Gaussian noise
926
927
𝑌(𝑡) =
%&'
%&'∗ + 𝜖
928
929
𝜖 ∼ 𝑁𝑜𝑟𝑚𝑎𝑙(0, 𝜎!)
930
931
where the variance of the measurement noise 𝜎! was estimated from the data.
932
Appendix Table S1 shows all modelled species, their notation used for the equation,
933
as well as the initial values. We assume that in the beginning of the experiment, all
934
species are in the inactive form, reflecting the fact that the cells have been starved.
935
The total concentrations of all species have been normalized to 1. The model
936
equations are shown in Appendix Table S2. The phosphorylation events are modelled
937
31
with Michaelis-Menten kinetics. The NFB is modelled through the modeling species
938
𝑁𝐹𝐵 and its “active” version 𝑁𝐹𝐵∗ which affects the dephosphorylation rate of 𝑅𝐴𝐹
939
linearly. The activation, endocytosis, and recycling of the EGF receptor is modelled
940
linearly. The model parameters are described in Appendix Table S3. For the modeling
941
of the two smaller models (without feedback (Figure EV1E) or without endocytosis
942
(Figure EV1H)), we set the corresponding parameters (𝑘#)* and 𝑟!,,) to zero.
943
For the parameter inference, we used a Nested Sampling algorithm as described in
944
(Mikelson and Khammash 2020). The inference was performed on the ETH High-
945
performance Cluster Euler and was done using the parallel implementation on 48
946
cores. The algorithm was run for 24 hours or until the algorithm stopped because the
947
termination criterion 𝛥-./0 (see (Mikelson and Khammash 2020) for details) was −∞.
948
As prior distributions, we chose for all parameters non-informative log-uniform priors
949
between 10-5 and 105, except for 𝑘𝑡𝑟"#"$ for which we chose a uniform prior on the
950
interval [0, 1] and for 𝜎 for which we chose a log-uniform prior between 10-5 and 1.
951
Predictive distributions can be found on Figure 3F,G, EV1F,G,I,J.
952
953
RNAi perturbation screen
954
We used Ingenuity Pathway Analysis (IPA, Qiagen) to select proteins directly
955
interacting with ERK, MEK, RAF, RAS and FGFR, that are known to be expressed in
956
NIH3T3 cells using a proteomics approach (Schwanhäusser et al. 2011; Jensen et al.
957
2009) (Appendix Table S4). We then imported this protein list in STRING (Jensen et
958
al. 2009) to generate an interaction network with a minimum interaction score of 0.4.
959
The final interactome was manually modified to display the protein names to facilitate
960
the readout (Figure 4A). We targeted these selected proteins with RNA interference,
961
using the siPOOL technology (one siPOOL containing a mix of 30 siRNAs targeting
962
the same gene (Hannus et al. 2014), sequences available in the Data availability
963
section). We arranged the siPOOLs in a 96 well plate format (in columns 2-5 and 8-
964
11, one well per siPOOL) with the non-targeting siRNA (CTRL) and the positive control
965
(mix of 5 nM siPOOL against ERK1 and 5 nM siPOOL against ERK2) placed
966
alternately in columns 1, 6, 7 and 12. Cells were reverse transfected using RNAiMAX
967
(Thermofisher, 13778150) following the recommended siPOOL transfection protocol
968
(https://sitoolsbiotech.com/protocols.php).
OptoFGFR-expressing
cells
were
969
transfected with 10 nM of siPOOL in a 96 well 1.5 glass bottom plate (Cellvis) coated
970
with 10 μg/ml Fibronectin (Huber Lab) at 0.3 x 103 cells/well density and incubated for
971
72 hours at 37°C and 5% CO2. For the imaging, the 96 well plate was divided into 15
972
sub-experiments, each sub-experiment consisting of a negative control well, a positive
973
control well and 4 wells with different siPOOLs. We selected 2 FOVs per well and
974
programmed the microscope to run the 15 experiments sequentially, acquiring the
975
ERK-KTR-mRuby2 and the H2B-miRFP703 channels with a 1-minute interval,
976
stimulating the cells with sustained optoFGFR input (2-minute intervals, D = 18
977
mJ/cm2), and acquiring a final frame with ERK-KTR-mRuby2, H2B-miRFP703 and
978
optoFGFR-mCitrine (Figure 4B,D-F, EV2C,D, EV3A-D, 6F-H, EV5B). For the optoSOS
979
system, we limited the perturbation screen to targets acting below RAS (Figure 5F,G,
980
EV4D-G). Stimulations were done with sustained optoSOS input (5 repeated pulses
981
32
at 2-minute intervals, D = 0.6 J/cm2). For EGF experiments, cells were stimulated with
982
1 ng/ml EGF at t = 5 minutes (Figure EV3E-G).
983
984
Real-time qPCR
985
Cells were transfected with different concentrations of siPOOL in a 24 well plate at 5
986
x 103 cells/well density and incubated at 37°C with 5% CO2 for 72 hours before RNA
987
isolation. Reverse transcription was done with the ProtoScript II reverse transcriptase
988
kit (Bioconcept, M0368L). Real-time qPCR reactions were run using the MESA Green
989
pPCR MasterMix Plus for SYBR Green assay (Eurogenetec, RT-SY2X-03+WOU) on
990
the Rotor-Gen Q device (Qiagen). Each sample was tested in triplicate. Expression
991
level of the gene of interest was calculated using the 2-ΔΔCt method with GAPDH
992
expression level as internal control (Figure EV2A). The following primers were used
993
for the RT-qPCR reaction (designed with the Real-time PCR (TaqMan) Primer and
994
Probes Design Tool from GenScript).
995
996
Target
Forward sequences
Reverse sequences
ERK1
5’-GGTTGTTCCCAAATGCTGACT-3’
5’-CAACTTCAATCCTCTTGTGAGGG-3’
ERK2
5’-TCCGCCATGAGAATGTTATAGGC-3’
5’-GGTGGTGTTGATAAGCAGATTGG-3’
MEK1
5’-AAGGTGGGGGAACTGAAGGAT-3’
5’-CGGATTGCGGGTTTGATCTC-3’
MEK2
5’-GTTACCGGCACTCACTATCAA C-3’
5’-CCTCCAGCCGCTTCCTTTG-3’
GAPDH
5'-ACCCAGAAGACTGTGGATGG-3'
5'-TCAGCTCAGGGATGACCTTG-3'
997
Immunoblotting
998
Cells were transfected with 10 nM siPOOL in 6 well plates at 6 x 104 cells/well density
999
and incubated at 37°C with 5% CO2 for 72 hours. Cells were lysed in a buffer
1000
containing 10 mM Tris HCl, 1 mM EDTA and 1% SDS. Protein concentration was
1001
determined with the BCATM protein assay kit (Thermofisher, 23227). Home cast 10%
1002
SDS gels or Novex 4%-20% 10 well Mini Gels (Thermofisher, XP04200) were used
1003
for SDS page. Transfer was done using PVDF membranes and a Trans-Blot SD Semi-
1004
Dry Electrophoretic Transfer Cell (Bio-Rad). Imaging was done with an Odyssey
1005
Fluorescence scanner (Li-COR) (Figure 4C, EV2B). The following primary antibodies
1006
were used: anti-total ERK (M7927, Sigma), anti-MEK1 (ab32091, Abcam), anti-MEK2
1007
(ab32517, Abcam), anti-BRAF (sc-5284, Santa Cruz), anti-CRAF (9422S, Cell
1008
Signaling Technology), anti-SOS1 (610096, Biosciences), anti-GRB2 (PA5-17692,
1009
Invitrogen) and anti-RSK2 (sc-9986, Santa Cruz). Anti-GAPDH (sc-32233, Santa
1010
Cruz) or anti-Actin (A2066, Merck) was used as protein of reference. For the
1011
secondary antibodies, we used the IRDye 680LT donkey anti-mouse IgG (926-68022,
1012
Li-COR), IRDye 800CW goat anti-mouse (926-32210, Li-COR) and IRDye 800CW
1013
donkey anti-rabbit (926-32213, Li-COR). Protein quantification was done with the
1014
Image StudioTM Lite software.
1015
33
1016
Time-series feature extraction
1017
We used custom scripts to extract features of ERK responses to transient optoFGFR
1018
input (Figure 2B, EV1A,B), sustained GF input (Figure EV1C,D) and transient
1019
optoSOS input (EV4A). The maximum peak (maxPeak) is the absolute value of the
1020
highest ERK activity in the trajectory. To estimate the full width at half maximum
1021
(FWHM), we first removed the baseline of the trajectories and increased their sampling
1022
frequency by a factor 30 with spline interpolation. On the resulting trajectory, we
1023
applied a “walk” procedure to quantify the FWHM. In this method, a pointer walks left
1024
and right (i.e. opposite and along the direction of time respectively) from the maximum
1025
point of the trajectory. The pointer stops whenever the half maximum value is crossed.
1026
Both stops define a left and a right border, the time difference between these 2-border
1027
time-points gives the FWHM. To avoid reporting aberrant FWHM values in cases
1028
where a peak cannot be clearly defined, we excluded FWHM calculation for
1029
trajectories where the fold change between the baseline (mean activity before
1030
stimulation) and the maximum value of the trajectory was below a threshold manually
1031
defined. ERKpostStim is the absolute value of ERK activity extracted 9 minutes after
1032
the last stimulation pulse to evaluate ERK adaptation. Statistical analysis (Figure
1033
EV1A,B) was done by comparing all conditions to the 20-minute interval stimulation
1034
patterns with a Wilcoxon test using the FDR p-value correction (NS: non-significant,
1035
*<0.01, **<0.001, ***<0.0001, ****<0.00001).
1036
To evaluate ERK phenotypes under siRNA perturbations in response to sustained
1037
optoFGFR or optoSOS input (Figure 4D, EV3A, 5F, EV4F), we extracted the baseline
1038
(average ERK activity on 5 timepoints before stimulation), the maxPeak (maximum
1039
ERK activity within a 10-minute time window following the start of the stimulation) and
1040
the ERKpostStim (ERK activity at a fixed timepoint post-stimulation (toptoFGFR = 42 min
1041
and toptoSOS = 40 min)) from 3 technical replicates. To avoid heterogeneity due to
1042
differences in optogenetic expression, we focused our analysis on cells with high
1043
optogenetic expression. The obtained baseline, maxPeak and ERKpostStim for each
1044
siRNA perturbation was z-scored to the non-targeting siRNA (CTRL). Non-significant
1045
results were manually set to grey. Statistical analysis was done by comparing each
1046
perturbation to the control with a Wilcoxon test using the FDR p-value correction (NS:
1047
non-significant, *<0.05, **<0.005, ***<0.0005, ****<0.00005).
1048
For the comparison of both optogenetic systems (Figure 5E, EV4C), ERK baseline
1049
was obtained by averaging ERK activity on 5 timepoints before stimulation and ERK
1050
maxPeak was extracted within a 10-minute time window following the start of the
1051
stimulation. Statistical analysis was done by comparing low and high expressing cells
1052
within and across optogenetic systems with a Wilcoxon test using the FDR p-value
1053
correction (NS: non-significant, *<0.05, **<0.005, ***<0.0005, ****<0.00005).
1054
To quantify the efficiency of the three MAPK inhibitors on the reduction of ERK
1055
amplitudes under sustained high optoFGFR or optoSOS input (Figure 6), extraction of
1056
the maxPeak was limited by the fact that several concentrations led to a full
1057
suppression of ERK amplitudes. Therefore, we extracted ERK amplitudes at a fixed
1058
time point following the start of the stimulation (tfixed optoFGFR = 15 minutes, tfixed optoSOS =
1059
34
10 minutes). The obtained ERK amplitudes were then plotted for each concentration
1060
for a fixed number of cells randomly selected (Figure 6B,F, EV5D). To calculate the
1061
EC50 of each drug, we normalized the data by setting the mean ERK responses of the
1062
non-treated condition to 1 and the mean ERK responses of the maximum
1063
concentration to 0. EC50 then was calculated by fitting a Hill function to the mean ERK
1064
activity of each concentration (Figure 6C,G, EV5E, Appendix Table 5-7). The
1065
heterogeneity of ERK amplitude at the fixed time point was evaluated by computing
1066
the normalized standard deviation of the extracted ERK activity per condition (Figure
1067
6D,H, EV5F).
1068
1069
Identification of ERK dynamics phenotypes using CODEX
1070
To investigate ERK dynamics phenotypes to siRNA perturbations, we first trained a
1071
convolutional neural network (CNN) to classify input ERK trajectories into any of the
1072
siRNA-perturbed conditions (Figure EV3C). For this purpose, we used a CNN
1073
architecture composed of 4 1D-convolution layers with 20 kernels of size 5, followed
1074
by a convolution layer with 20 kernels of size 3 and one layer of 10 kernels of size 3.
1075
The responses are then pooled with global average pooling to generate a vector of 10
1076
features that is passed to a (10,63) fully connected layer for classification. Each
1077
convolutional layer is followed by ReLU and batch normalization. The CNN was trained
1078
to minimize the cross-entropy loss, with L2 weight penalty of 1e-3.
1079
To identify siRNA treatments that induced a distinctive phenotype, we selected the 10
1080
conditions for which the CNN classification precision was the highest on the validation
1081
set (Appendix Table S4, “CODEX accuracy”). To these 10 conditions, we also added
1082
the negative control (non-targeting siRNA (CTRL)). We trained a second CNN, with
1083
the same architecture and training parameters, but limited to recognizing the 11
1084
selected treatments to obtain a clear embedding of these hits. With this new model,
1085
we extracted the features used for the classification of the trajectories (i.e. the input
1086
representation after the last convolution layer) and projected them with tSNE (Python’s
1087
sklearn implementation, perplexity of 100, learning rate of 600 and 2500 iterations)
1088
(Figure EV3D). We selected 10 prototype curves for each treatment by taking the
1089
trajectories for which the second CNN’s classification confidence (i.e. the probability
1090
for the actual class of the inputs) were the highest in the validation set (Figure 4E,
1091
“CODEX”).
1092
To visualize the ERK dynamics landscape in MCF10A WT cells and in MCF10A cells
1093
overexpressing ErbB2, we trained one CNN for each cell line. These CNNs were
1094
trained to recognize the drug treatment applied on cells, using single-cell ERK traces
1095
as input. The architecture of the CNNs is the same as described previously. The only
1096
difference lies in the number of outputs in the final fully connected layer, which were
1097
set to the number of drug treatments. Features used for the classification of the
1098
trajectories were then projected with tSNE (Figure 7D,H, EV5J,K).
1099
To identify clusters gathering similar ERK dynamics (Figure 7E,F,I,J), we clustered
1100
trajectories based on their CNN features using a partition around medoids (PAM). This
1101
iterative algorithm is similar to K-means clustering. PAM defines the cluster centers
1102
(i.e. the medoids) as the observed data points which minimize the median distances
1103
35
to all other points in its own cluster. This makes PAM more robust to outliers than K-
1104
means which uses the average coordinates of a cluster to define its center.
1105
Representative trajectories were obtained by taking the medoids of each cluster and
1106
their four closest neighbors (Figure 7G,K). Distances between points were defined
1107
with the Manhattan distance between the scaled CNN features (zero mean and unit
1108
variance). We manually verified that these clusters captured an actual trend by
1109
visualizing trajectories in each cluster with the interactive CODEX application.
1110
1111
Peak detection and classification of oscillatory trajectories
1112
The number of ERK activity peaks was calculated with a custom algorithm that detects
1113
local maxima in time series. First, we applied a short median filter to smoothen the
1114
data with a window width of 3 time points. Then, we ran a long median filter to estimate
1115
the long-term bias with a window width of 15 time points. This bias was then subtracted
1116
from the smoothed time series and we only kept the positive values. If no point in this
1117
processed trajectory was exceeding a manual threshold of 0.075, all variations were
1118
considered as noise and no peak was extracted from the trajectory. The remaining
1119
trajectories were then rescaled to [0,1]. Finally, peaks were detected as points that
1120
exceeded a threshold which was manually set to 0.1. Peaks that were found before
1121
the first stimulation or after the last stimulation were filtered out.
1122
The classification of trajectories into oscillatory and non-oscillatory behaviors was
1123
performed after the peak detection step. Cells were called oscillatory if at least 3 peaks
1124
were detected with the peak detection procedure (Figure 4F). Statistical analysis was
1125
done using a pairwise t-test comparing each perturbation to the control for high and
1126
low levels of optoFGFR independently, with FDR p-value correction (*<0.05, **<0.005,
1127
***<0.0005, ****<0.00005).
1128
1129
Data availability
1130
The datasets used in this study as well as all R codes used for further analysis are
1131
available at https://data.mendeley.com/datasets/st36dd7k23/1. Source code for the
1132
inference algorithm, model files and results are available at
1133
https://github.com/Mijan/LFNS_optoFGFR.
1134
1135
Acknowledgements
1136
1137
This work was supported by SystemsX.ch, Swiss Cancer League and Swiss National
1138
Science Foundation grants to Olivier Pertz, by the H2020-MSCA-IF, project No. 89631
1139
- NOSCAR to Agne Frismantiene and by the European Union’s Horizon 2020 and
1140
innovation program under grant agreement No. 730964 (TRANSVAC project) to
1141
Mustafa Khammash. We thank Won Do Heo for sharing the optoFGFR plasmid,
1142
Kazuhiro Aoki for sharing the pPBbSr2-MCS plasmid, and David Hacker for sharing
1143
the pSB-HPB plasmid. We thank the Microscopy Imaging Center of the University of
1144
Bern for its support.
1145
1146
Authors contribution
1147
36
1148
O.P. and C.D. designed the study. C.D. developed the optogenetic systems and
1149
imaging pipelines. CD performed the experiment and image analysis on NIH3T3. A.F
1150
and P.A.G. performed the experiments and image analysis on MCF10A. M.D.
1151
developed the processing pipelines. C.D processed the data. C.D., M.-A.J., A.F and
1152
P.A.G. analyzed the data. M.-A.J. conducted the CNN analysis. J.M. performed
1153
mathematical modeling. O.P and M.K. supervised the work. O.P., C.D. and J.M. wrote
1154
the paper.
1155
1156
Conflict of interest
1157
1158
The authors declare having no conflict of interest.
1159
37
References
1160
1161
Albeck, J.G., Mills, G.B. and Brugge, J.S. 2013. Frequency-modulated pulses of ERK activity transmit
1162
quantitative proliferation signals. Molecular Cell 49(2), pp. 249–261.
1163
1164
Arkun, Y. and Yasemi, M. 2018. Dynamics and control of the ERK signaling pathway: Sensitivity,
1165
bistability, and oscillations. Plos One 13(4), p. e0195513.
1166
1167
Arteaga, C.L. and Engelman, J.A. 2014. ERBB receptors: from oncogene discovery to basic science
1168
to mechanism-based cancer therapeutics. Cancer Cell 25(3), pp. 282–303.
1169
1170
Avraham, R. and Yarden, Y. 2011. Feedback regulation of EGFR signalling: decision making by early
1171
and delayed loops. Nature Reviews. Molecular Cell Biology 12(2), pp. 104–117.
1172
1173
Balasubramanian, S., Wurm, F.M. and Hacker, D.L. 2016. Multigene expression in stable CHO cell
1174
pools generated with the piggyBac transposon system. Biotechnology Progress 32(5), pp. 1308–
1175
1317.
1176
1177
Belov, A.A. and Mohammadi, M. 2012. Grb2, a double-edged sword of receptor tyrosine kinase
1178
signaling. Science Signaling 5(249), p. pe49.
1179
1180
Benedetti, L., Barentine, A.E.S., Messa, M., Wheeler, H., Bewersdorf, J. and De Camilli, P. 2018.
1181
Light-activated protein interaction with high spatial subcellular confinement. Proceedings of the
1182
National Academy of Sciences of the United States of America 115(10), pp. E2238–E2245.
1183
1184
Berg, S., Kutra, D., Kroeger, T., et al. 2019. ilastik: interactive machine learning for (bio)image
1185
analysis. Nature Methods 16(12), pp. 1226–1232.
1186
1187
Bindels, D.S., Haarbosch, L., van Weeren, L., et al. 2017. mScarlet: a bright monomeric red
1188
fluorescent protein for cellular imaging. Nature Methods 14(1), pp. 53–56.
1189
1190
Blum, Y., Mikelson, J., Dobrzyński, M., et al. 2019. Temporal perturbation of ERK dynamics reveals
1191
network architecture of FGF2/MAPK signaling. Molecular Systems Biology 15(11), p. e8947.
1192
1193
Blüthgen, N. and Legewie, S. 2013. Robustness of signal transduction pathways. Cellular and
1194
Molecular Life Sciences 70(13), pp. 2259–2269.
1195
1196
Cargnello, M. and Roux, P.P. 2011. Activation and function of the MAPKs and their substrates, the
1197
MAPK-activated protein kinases. Microbiology and Molecular Biology Reviews 75(1), pp. 50–83.
1198
1199
Chardin, P., Camonis, J.H., Gale, N.W., et al. 1993. Human Sos1: a guanine nucleotide exchange
1200
factor for Ras that binds to GRB2. Science 260(5112), pp. 1338–1343.
1201
1202
Cherfils, J. and Zeghouf, M. 2013. Regulation of small GTPases by GEFs, GAPs, and GDIs.
1203
Physiological Reviews 93(1), pp. 269–309.
1204
1205
Cullen, P.J. and Lockyer, P.J. 2002. Integration of calcium and Ras signalling. Nature Reviews.
1206
Molecular Cell Biology 3(5), pp. 339–348.
1207
1208
Dobrzyński, M., Jacques, M.-A. and Pertz, O. 2019. Mining single-cell time-series datasets with Time
1209
Course Inspector. Bioinformatics.
1210
1211
38
Douville, E. and Downward, J. 1997. EGF induced SOS phosphorylation in PC12 cells involves P90
1212
RSK-2. Oncogene 15(4), pp. 373–383.
1213
1214
Duan, L., Hope, J., Ong, Q., et al. 2017. Understanding CRY2 interactions for optical control of
1215
intracellular signaling. Nature Communications 8(1), p. 547.
1216
1217
Erickson, K.E., Rukhlenko, O.S., Posner, R.G., Hlavacek, W.S. and Kholodenko, B.N. 2019. New
1218
insights into RAS biology reinvigorate interest in mathematical modeling of RAS signaling. Seminars
1219
in Cancer Biology 54, pp. 162–173.
1220
1221
Ferrell, J.E. and Bhatt, R.R. 1997. Mechanistic studies of the dual phosphorylation of mitogen-
1222
activated protein kinase. The Journal of Biological Chemistry 272(30), pp. 19008–19016.
1223
1224
Fritsche-Guenther, R., Witzel, F., Sieber, A., et al. 2011. Strong negative feedback from Erk to Raf
1225
confers robustness to MAPK signalling. Molecular Systems Biology 7, p. 489.
1226
1227
Gerosa, L., Chidley, C., Fröhlich, F., et al. 2020. Receptor-Driven ERK Pulses Reconfigure MAPK
1228
Signaling and Enable Persistence of Drug-Adapted BRAF-Mutant Melanoma Cells. Cell Systems
1229
11(5), p. 478–494.e9.
1230
1231
Goedhart, J., von Stetten, D., Noirclerc-Savoye, M., et al. 2012. Structure-guided evolution of cyan
1232
fluorescent proteins towards a quantum yield of 93%. Nature Communications 3, p. 751.
1233
1234
Hannus, M., Beitzinger, M., Engelmann, J.C., et al. 2014. siPools: highly complex but accurately
1235
defined siRNA pools eliminate off-target effects. Nucleic Acids Research 42(12), pp. 8049–8061.
1236
1237
Huang, C.-Y. and Tan, T.-H. 2012. DUSPs, to MAP kinases and beyond. Cell & bioscience 2(1), p.
1238
24.
1239
1240
Huotari, J. and Helenius, A. 2011. Endosome maturation. The EMBO Journal 30(17), pp. 3481–3500.
1241
1242
Jacques, M.A., Dobrzyński, M., Gagliardi, P.A., Sznitman, R. and Pertz, O. 2021. CODEX, a neural
1243
network approach to explore signaling dynamics landscapes. Molecular Systems Biology 17(4), p.
1244
e10026.
1245
1246
Jaqaman, K., Loerke, D., Mettlen, M., et al. 2008. Robust single-particle tracking in live-cell time-lapse
1247
sequences. Nature Methods 5(8), pp. 695–702.
1248
1249
Jensen, L.J., Kuhn, M., Stark, M., et al. 2009. STRING 8--a global view on proteins and their
1250
functional interactions in 630 organisms. Nucleic Acids Research 37(Database issue), pp. D412-6.
1251
1252
Johnson, H.E., Goyal, Y., Pannucci, N.L., Schüpbach, T., Shvartsman, S.Y. and Toettcher, J.E. 2017.
1253
The spatiotemporal limits of developmental erk signaling. Developmental Cell 40(2), pp. 185–192.
1254
1255
Junttila, M.R., Li, S.-P. and Westermarck, J. 2008. Phosphatase-mediated crosstalk between MAPK
1256
signaling pathways in the regulation of cell survival. The FASEB Journal 22(4), pp. 954–965.
1257
1258
Kanodia, J., Chai, D., Vollmer, J., et al. 2014. Deciphering the mechanism behind Fibroblast Growth
1259
Factor (FGF) induced biphasic signal-response profiles. Cell Communication and Signaling 12, p. 34.
1260
1261
Kholodenko, B.N. 2000. Negative feedback and ultrasensitivity can bring about oscillations in the
1262
mitogen-activted protein kinase cascades. Eur. J. Biochem 267, pp. 1583–1588.
1263
39
1264
Kholodenko, B.N., Hancock, J.F. and Kolch, W. 2010. Signalling ballet in space and time. Nature
1265
Reviews. Molecular Cell Biology 11(6), pp. 414–426.
1266
1267
Kim, N., Kim, J.M., Lee, M., Kim, C.Y., Chang, K.-Y. and Heo, W.D. 2014. Spatiotemporal control of
1268
fibroblast growth factor receptor signals by blue light. Chemistry & Biology 21(7), pp. 903–912.
1269
1270
Kiyatkin, A., van Alderwerelt van Rosenburgh, I.K., Klein, D.E. and Lemmon, M.A. 2020. Kinetics of
1271
receptor tyrosine kinase activation define ERK signaling dynamics. Science Signaling 13(645).
1272
1273
Kochańczyk, M., Kocieniewski, P., Kozłowska, E., et al. 2017. Relaxation oscillations and hierarchy of
1274
feedbacks in MAPK signaling. Scientific Reports 7, p. 38244.
1275
1276
Koseska, A. and Bastiaens, P.I.H. 2020. Processing temporal growth factor patterns by an epidermal
1277
growth factor receptor network dynamically established in space. Annual Review of Cell and
1278
Developmental Biology 36, pp. 359–383.
1279
1280
Kupzig, S., Walker, S.A. and Cullen, P.J. 2005. The frequencies of calcium oscillations are optimized
1281
for efficient calcium-mediated activation of Ras and the ERK/MAPK cascade. Proceedings of the
1282
National Academy of Sciences of the United States of America 102(21), pp. 7577–7582.
1283
1284
Lake, D., Corrêa, S.A.L. and Müller, J. 2016. Negative feedback regulation of the ERK1/2 MAPK
1285
pathway. Cellular and Molecular Life Sciences 73(23), pp. 4397–4413.
1286
1287
Lam, A.J., St-Pierre, F., Gong, Y., et al. 2012. Improving FRET dynamic range with bright green and
1288
red fluorescent proteins. Nature Methods 9(10), pp. 1005–1012.
1289
1290
Langlois, W.J., Sasaoka, T., Saltiel, A.R. and Olefsky, J.M. 1995. Negative feedback regulation and
1291
desensitization of insulin- and epidermal growth factor-stimulated p21ras activation. The Journal of
1292
Biological Chemistry 270(43), pp. 25320–25323.
1293
1294
Lavoie, H., Gagnon, J. and Therrien, M. 2020. ERK signalling: a master regulator of cell behaviour,
1295
life and fate. Nature Reviews. Molecular Cell Biology 21(10), pp. 607–632.
1296
1297
Lemmon, M.A., Freed, D.M., Schlessinger, J. and Kiyatkin, A. 2016. The dark side of cell signaling:
1298
positive roles for negative regulators. Cell 164(6), pp. 1172–1184.
1299
1300
McQuin, C., Goodman, A., Chernyshev, V., et al. 2018. CellProfiler 3.0: Next-generation image
1301
processing for biology. PLoS Biology 16(7), p. e2005970.
1302
1303
Mikelson, J. and Khammash, M. 2020. Likelihood-free nested sampling for parameter inference of
1304
biochemical reaction networks. PLoS Computational Biology 16(10), p. e1008264.
1305
1306
Nakakuki, T., Birtwistle, M.R., Saeki, Y., et al. 2010. Ligand-specific c-Fos expression emerges from
1307
the spatiotemporal control of ErbB network dynamics. Cell 141(5), pp. 884–896.
1308
1309
Ng, P.K.-S., Li, J., Jeong, K.J., et al. 2018. Systematic functional annotation of somatic mutations in
1310
cancer. Cancer Cell 33(3), p. 450–462.e10.
1311
1312
Ornitz, D.M. and Itoh, N. 2015. The Fibroblast Growth Factor signaling pathway. Wiley
1313
interdisciplinary reviews. Developmental biology 4(3), pp. 215–266.
1314
1315
40
Orton, R.J., Sturm, O.E., Vyshemirsky, V., Calder, M., Gilbert, D.R. and Kolch, W. 2005.
1316
Computational modelling of the receptor-tyrosine-kinase-activated MAPK pathway. The Biochemical
1317
Journal 392(Pt 2), pp. 249–261.
1318
1319
Patterson, K.I., Brummer, T., O’Brien, P.M. and Daly, R.J. 2009. Dual-specificity phosphatases:
1320
critical regulators with diverse cellular targets. The Biochemical Journal 418(3), pp. 475–489.
1321
1322
Purvis, J.E. and Lahav, G. 2013. Encoding and decoding cellular information through signaling
1323
dynamics. Cell 152(5), pp. 945–956.
1324
1325
Rauen, K.A. 2013. The RASopathies. Annual Review of Genomics and Human Genetics 14, pp. 355–
1326
369.
1327
1328
Regot, S., Hughey, J.J., Bajar, B.T., Carrasco, S. and Covert, M.W. 2014. High-sensitivity
1329
measurements of multiple kinase activities in live single cells. Cell 157(7), pp. 1724–1734.
1330
1331
Ryu, H., Chung, M., Dobrzyński, M., et al. 2015. Frequency modulation of ERK activation dynamics
1332
rewires cell fate. Molecular Systems Biology 11(11), p. 838.
1333
1334
Saha, M., Carriere, A., Cheerathodi, M., et al. 2012. RSK phosphorylates SOS1 creating 14-3-3-
1335
docking sites and negatively regulating MAPK activation. The Biochemical Journal 447(1), pp. 159–
1336
166.
1337
1338
Sakaue-Sawano, A., Kurokawa, H., Morimura, T., et al. 2008. Visualizing spatiotemporal dynamics of
1339
multicellular cell-cycle progression. Cell 132(3), pp. 487–498.
1340
1341
Samatar, A.A. and Poulikakos, P.I. 2014. Targeting RAS-ERK signalling in cancer: promises and
1342
challenges. Nature Reviews. Drug Discovery 13(12), pp. 928–942.
1343
1344
Santos, S.D.M., Verveer, P.J. and Bastiaens, P.I.H. 2007. Growth factor-induced MAPK network
1345
topology shapes Erk response determining PC-12 cell fate. Nature Cell Biology 9(3), pp. 324–330.
1346
1347
Saraf, A., Oberg, E.A. and Strack, S. 2010. Molecular determinants for PP2A substrate specificity:
1348
charged residues mediate dephosphorylation of tyrosine hydroxylase by the PP2A/B’ regulatory
1349
subunit. Biochemistry 49(5), pp. 986–995.
1350
1351
Schlessinger, J. 2000. Cell signaling by receptor tyrosine kinases. Cell 103(2), pp. 211–225.
1352
1353
Schmick, M., Kraemer, A. and Bastiaens, P.I.H. 2015. Ras moves to stay in place. Trends in Cell
1354
Biology 25(4), pp. 190–197.
1355
1356
Schwanhäusser, B., Busse, D., Li, N., et al. 2011. Global quantification of mammalian gene
1357
expression control. Nature 473(7347), pp. 337–342.
1358
1359
Shankaran, H., Ippolito, D.L., Chrisler, W.B., et al. 2009. Rapid and sustained nuclear-cytoplasmic
1360
ERK oscillations induced by epidermal growth factor. Molecular Systems Biology 5, p. 332.
1361
1362
Shcherbakova, D.M., Baloban, M., Emelyanov, A.V., Brenowitz, M., Guo, P. and Verkhusha, V.V.
1363
2016. Bright monomeric near-infrared fluorescent proteins as tags and biosensors for multiscale
1364
imaging. Nature Communications 7, p. 12405.
1365
1366
Smith, J.A., Poteet-Smith, C.E., Xu, Y., Errington, T.M., Hecht, S.M. and Lannigan, D.A. 2005.
1367
41
Identification of the first specific inhibitor of p90 ribosomal S6 kinase (RSK) reveals an unexpected
1368
role for RSK in cancer cell proliferation. Cancer Research 65(3), pp. 1027–1034.
1369
1370
Sturm, O.E., Orton, R., Grindlay, J., et al. 2010. The mammalian MAPK/ERK pathway exhibits
1371
properties of a negative feedback amplifier. Science Signaling 3(153), p. ra90.
1372
1373
Sugimoto, T., Stewart, S. and Guan, K.L. 1997. The calcium/calmodulin-dependent protein
1374
phosphatase calcineurin is the major Elk-1 phosphatase. The Journal of Biological Chemistry 272(47),
1375
pp. 29415–29418.
1376
1377
Yarden, Y. and Pines, G. 2012. The ERBB network: at last, cancer therapy meets systems biology.
1378
Nature Reviews. Cancer 12(8), pp. 553–563.
1379
1380
Yeung, K., Janosch, P., McFerran, B., et al. 2000. Mechanism of suppression of the
1381
Raf/MEK/extracellular signal-regulated kinase pathway by the raf kinase inhibitor protein. Molecular
1382
and Cellular Biology 20(9), pp. 3079–3085.
1383
1384
Yoo, S.-M., Cho, S.J. and Cho, Y.-Y. 2015. Molecular targeting of erks/rsk2 signaling axis in cancer
1385
prevention. Journal of cancer prevention 20(3), pp. 165–171.
1386
1387
Yusa, K., Zhou, L., Li, M.A., Bradley, A. and Craig, N.L. 2011. A hyperactive piggyBac transposase for
1388
mammalian applications. Proceedings of the National Academy of Sciences of the United States of
1389
America 108(4), pp. 1531–1536.
1390
1391
| 2022 | Optogenetic actuator/ERK biosensor circuits identify MAPK network nodes that shape ERK dynamics | 10.1101/2021.07.27.453955 | [
"Dessauges Coralie",
"Mikelson Jan",
"Dobrzyński Maciej",
"Jacques Marc-Antoine",
"Frismantiene Agne",
"Gagliardi Paolo Armando",
"Khammash Mustafa",
"Pertz Olivier"
] | creative-commons |
1
Chelator sensing and lipopeptide interplay mediates molecular
1
interspecies interactions between soil bacilli and pseudomonads
2
3
Sofija Andric1*, Thibault Meyer1,Φ *, Augustin Rigolet1, Anthony Argüelles Arias1, Sébastien Steels1,
4
Grégory Hoff1,#, Monica Höfte2, René De Mot3, Andrea McCann4, Edwin De Pauw4 and Marc
5
Ongena1
6
7
1Microbial Processes and Interactions Laboratory, Terra Teaching and Research Center, Gembloux Agro-Bio
8
Tech, University of Liège, Gembloux, Belgium
9
2Laboratory of Phytopathology, Department of Plants and Crops, Faculty of Bioscience engineering, Ghent
10
University, Gent, Belgium
11
3 Centre of Microbial and Plant Genetics, Faculty of Bioscience Engineering, University of Leuven, Heverlee,
12
Belgium
13
4 Mass Spectrometry Laboratory, , MolSys Research Unit, Department of Chemistry, University of Liège,
14
Belgium
15
Φ Current address: UMR Ecologie Microbienne, F-69622, University of Lyon, Université Claude Bernard Lyon
16
1, CNRS, INRAE, VetAgro Sup, Villeurbanne, France
17
# Current address: Ecology and Biodiversity, Department of Biology, Utrecht University, Padualaan 8, 3584
18
CH, Utrecht, The Netherlands
19
20
* Equal contribution
21
22
2
23
Abstract
24
25
Some bacterial species are important members of the rhizosphere microbiome and confer
26
protection to the host plant against pathogens. However, our knowledge of the multitrophic
27
interactions determining the ecological fitness of these biocontrol bacteria in their highly competitive
28
natural niche is still limited. In this work, we investigated the molecular mechanisms underlying
29
interactions between B. velezensis, considered as model plant-associated and beneficial species in
30
the Bacillus genus, and Pseudomonas as a rhizosphere-dwelling competitor. Our data show that B.
31
velezensis boosts its arsenal of specialized antibacterials upon the perception of the secondary
32
siderophore enantio-pyochelin produced by phylogenetically distinct pseudomonads and some
33
other genera. We postulate that B. velezensis has developed some chelator sensing systems to
34
learn about the identity of its surrounding competitors. Illustrating the multifaceted molecular
35
response of Bacillus, surfactin is another crucial component of the secondary metabolome
36
mobilized in interbacteria competition. Its accumulation not only enhances motility but,
37
unexpectedly, the lipopeptide also acts as a chemical trap that reduces the toxicity of other
38
lipopeptides released by Pseudomonas challengers. This in turn favors the persistence of Bacillus
39
populations upon competitive root colonization. Our work thus highlights new ecological roles for
40
bacterial secondary metabolites acting as key drivers of social interactions.
41
42
43
3
44
Soil is one of the richest ecosystems in terms of microbial diversity and abundance1. However, the
45
scarcity of resources makes it one of the most privileged environments for competitive interspecies
46
interactions2,3. A subset of the bulk soil microbes has evolved to dwell in the rhizosphere
47
compartment surrounding roots due to continued nutrient-enriched exudation from the plant.
48
Compared to bulk soil, microbial warfare in the rhizosphere is presumably even more intense as the
49
habitat is spatially restricted and more densely populated3. Besides rivalry for nutrients
50
(exploitative), interference competition is considered a key factor driving microbial interactions and
51
community assembly. This competition can involve signal interference or toxins deployed by
52
contact-dependent delivery systems4,5 but is mainly mediated at distance through the emission of
53
various molecular weapons. The molecular basis of interference interactions and their phenotypic
54
outcomes between diverse soil bacterial species have been amply investigated in the last
55
decade6,7.
56
Bacilli belonging to the B. subtilis complex are ubiquitous members of the rhizosphere
57
microbiome2,8,9. Among these species, B. velezensis has emerged as plant-associated model
58
bacilli, displaying strong potential as biocontrol agent reducing diseases caused by
59
phytopathogens10. B. velezensis distinguishes itself from other species of the B. subtilis group by its
60
richness in biosynthetic gene clusters (BGCs, representing up to 13% of the whole genome)
61
responsible for the synthesis of bioactive secondary metabolites (BSMs)11,12. This chemically-
62
diverse secondary metabolome includes volatiles, terpenes, non-ribosomal (NR) dipeptides, cyclic
63
lipopeptides (CLPs) and polyketides (PKs), but also ribosomally synthesized lantibiotics and larger
64
bacteriocins (RiPP)13,14. BSMs are involved in biocontrol activity via direct inhibition of pathogenic
65
microbes and/or via stimulation of the plant immune system15,16. From an ecological viewpoint,
66
BSMs also contribute to competitiveness in the rhizosphere niche thanks to multiple and
67
4
complementary functions as drivers of developmental traits, as antimicrobials, or as signals
68
initiating cross-talk with the host plant17–19.
69
Mostly guided by practical concerns for use as biocontrol agents, research on BSMs has
70
mainly focused on the characterization of their biological activities. However, the impact of
71
environmental factors that may modulate their expression under natural conditions still remains
72
poorly understood. It includes interactions with other organisms sharing the niche. Some recent
73
reports illustrate how soil bacilli may adapt their behavior upon sensing bacterial competitors but
74
almost exclusively focusing on developmental traits (sporulation, biofilm formation, or motility)7.
75
Unlike other genera such as Streptomyces, it remains largely unknown to what extent bacilli in
76
general and B. velezensis in particular, may modulate the expression of their secondary
77
metabolome upon interaction with other bacteria7,20. In this work, we investigated the molecular
78
outcomes of interspecies interactions in which B. velezensis may engage. We selected
79
Pseudomonas as challenger considering that species of this genus are also highly competitive and
80
commonly encountered in rhizosphere microbiomes8. We performed experiments under nutritional
81
conditions mimicking the oligotrophic rhizosphere environment and used contact-independent
82
settings for pairwise interaction which probably best reflect the real situation in soil. Our data
83
revealed that the two bacteria initiate multifaceted interactions mostly mediated by the non-
84
ribosomally synthesized components of their secondary metabolome. We pointed out unsuspected
85
roles for some of these BSMs in the interaction context. Beyond its role as a metal chelator, the
86
Pseudomonas secondary siderophore enantio-pyochelin (E-PCH) acts as a signal triggering dual
87
production of PKs and RiPP in Bacillus, while specific lipopeptides modulate the inhibitory
88
interaction between the two species. This results in marked phenotypic changes in B. velezensis
89
such as higher antibacterial potential, enhanced motility and protective effect via chemical trapping.
90
We also illustrate the relevance of these outcomes in the context of competitive root colonization.
91
92
5
Results
93
B. velezensis modulates its secondary metabolome and boosts antibacterial activity upon
94
sensing Pseudomonas metabolites
95
We used B. velezensis strain GA1 as a BSM-rich and genetically amenable isolate representative
96
of the species. Genome mining with AntiSMASH 5.021 confirmed the presence of all gene clusters
97
necessary for the biosynthesis of known BSMs typically formed by this bacterium (Supplementary
98
Table 1). Based on the exact mass and absence of the corresponding peaks in deletion mutants,
99
most of the predicted non-ribosomal (NR) secondary metabolites were identified in cell-free crude
100
supernatants via optimized UPLC-MS (Supplementary Fig. 1). It includes the whole set of cyclic
101
lipopeptides (CLPs of the surfactin, fengycin and iturin families) and polyketides (PKs difficidin,
102
macrolactin and bacillaene) with their multiple co-produced structural variants, as well as the
103
siderophore bacillibactin. We verified that all these compounds are readily formed in the so-called
104
exudate-mimicking (EM) medium reflecting the specific content in major carbon sources typically
105
released by roots of Solanaceae (such as tomato) plants22. In addition to these NR products, genes
106
encoding RiPPs such as amylocyclicin and amylolysin are also present in GA1, but these
107
compounds could not be reliably detected in culture broths. The NR dipeptide bacilysin was also
108
predicted but not detected. We selected as the main interaction partner the plant-associated
109
Pseudomonas sp. strain CMR12a based on its biocontrol potential and its production of multiple
110
secondary metabolites23–27. Genome mining confirmed the potential of CMR12a to synthesize a
111
range of BSMs, including antimicrobial phenazines, the siderophores pyoverdine (PVD) (structure
112
confirmation in Supplementary Fig. 2) and E-PCH as well as two structurally distinct CLPs, sessilins
113
and orfamides (Supplementary Table 1). In contrast to Bacillus, the capacity to co-produce two
114
different CLPs is a quite rare trait for non-phytopathogenic pseudomonads and represented an
115
additional criterion for selecting strain CMR12a for this study23,27–29. In the case of CMR12a,
116
according to the exact mass and absence of the corresponding peaks in deletion mutants, all these
117
6
compounds were detected in EM culture broth but most of them are more efficiently produced upon
118
growth in casamino acids medium (CAA) commonly used for Pseudomonas cultivation
119
(Supplementary Fig. 1).
120
Our prime objective was to evaluate the intrinsic potential of B. velezensis to react to the
121
perception of Pseudomonas metabolites in an experimental setting avoiding interferences due to
122
diffusion constraints in a semi-solid matrix or due to the formation of impermeable biofilm
123
structures. The first assays were performed by growing GA1 in agitated liquid EM medium
124
supplemented or not with (sterile) BSM-containing spent medium of CAA-grown CMR12a (CFS,
125
cell-free supernatant). At a low dose (2% (v/v)), the addition of this CFS extract led to a marked
126
increase in the production of some GA1 NR metabolites. Significantly higher amounts were
127
measured for surfactins, bacillaene or its dehydrated variant dihydrobacillaene (2H-bae), difficidin
128
or its oxidized form, and bacillibactin (Fig. 1a, b) but not for other compounds such as fengycins,
129
iturins and macrolactins (Fig. 1a).
130
The boost in BSMs synthesis triggered by Pseudomonas CFS was associated with an
131
increase in the antibacterial activity of the corresponding extracts when tested for growth inhibition
132
of Xanthomonas campestris and Clavibacter michiganensis used respectively as representative of
133
Gram-negative and Gram-positive plant pathogenic bacteria of agronomical importance30 (Fig. 1c).
134
Since most of the BSMs are not commercially available and our attempt to purify PKs failed due to
135
chemical instability, we could not use individual compounds for their specific involvement in
136
bacterial inhibition. As an alternative, we generated and tested a range of GA1 knockout mutants
137
including the Δsfp derivative specifically repressed in 4'-phosphopantetheinyl transferase essential
138
for the proper functioning of the PK and NRP biosynthesis machinery. Full loss of anti-
139
Xanthomonas activity in Δsfp extracts indicated a key role for NR BSMs and ruled out the possible
140
involvement of other chemicals known for their antibacterial activity such as bacilysin or RiPPs (Fig.
141
1d). Loss of function of mutants specifically repressed in the synthesis of individual compounds
142
7
pointed out the key role of (oxy)difficidin and to a lower extent of 2H-bae in Xanthomonas inhibition
143
(Fig. 1d). These two PKs are also responsible for GA1 inhibitory activity toward other important
144
bacterial phytopathogens such as Pectobacterium carotovorum, Agrobacterium tumefaciens and
145
Rhodococcus fasciens but are not involved in the inhibition of plant pathogenic Pseudomonas
146
species for which bacilysin may be the active metabolite (Supplementary Fig. 3). However, as
147
illustrated below, B. velezensis does not display significant toxicity against CMR12a and other non-
148
pathogenic soil Pseudomonas isolates tested here. Stimulation of PKs synthesis upon sensing
149
CMR12a is not specific to GA1 and was also observed in other B. velezensis strains with well-
150
known biocontrol potential such as S499, FZB42 and QST71331–33 (Supplementary Fig. 4).
151
In contrast to Xanthomonas, enhanced antibiotic activity against Clavibacter is not mediated
152
by NR products as shown by the fully conserved activity in the Δsfp mutant (Fig. 1d). Therefore, we
153
suspected from genomic data and literature34 that RiPPs such as amylocyclicin could be involved in
154
inhibition. This hypothesis was supported by the 80% reduction in antibiotic potential observed for
155
the ΔacnA mutant knocked out for the corresponding biosynthesis gene (Fig. 1d). Besides, RT-
156
qPCR data revealed a highly induced expression of acnA gene in GA1 cells upon supplementation
157
with CMR12a CFS (Fig. 1e). However, we were not able to provide evidence for higher
158
accumulation of the mature peptide in the medium. Enhanced expression of the acnA gene in
159
presence of Pseudomonas products was also observed for strain S499 (Supplementary Fig. 5).
160
161
E-PCH acts as a signal sensed by Bacillus to stimulate polyketide production
162
We next wanted to identify the signaling molecules secreted by Pseudomonas that are sensed by
163
Bacillus cells and lead to improved BSMs production. For that purpose, we used 2H-bae as an
164
indicator of the Bacillus response because it represents the most consistent and highly boosted
165
polyketide. We first compared the triggering potential of CFS obtained from knockout mutants of
166
CMR12a specifically lacking the different identified metabolites (Supplementary Fig. 1). Only
167
8
extracts from mutants impaired in the production of siderophores and more specifically E-PCH were
168
significantly affected in PKs-inducing potential (Fig. 2a). Possible involvement of this compound
169
was supported by the drastic reduction in the activity of CFS prepared from CMR12a culture in CAA
170
medium supplemented with Fe3+ where siderophore expression is repressed (Fig. 2b,
171
Supplementary Fig. 6). We also performed bioactivity-guided fractionation and data showed that
172
only extracts containing PVD and/or E-PCH displayed consistent PKs-triggering activity
173
(Supplementary Fig. 7). HPLC-purified compounds were also tested independently at a
174
concentration similar to the one measured in CFS CAA extract revealing a much higher PK-
175
triggering activity for E-PCH compared to the main PVD isoform (Fig. 2b). Dose-dependent assays
176
further indicated that supplementation with PVD, as strong chelator35, caused iron limitation in the
177
medium which is sensed by GA1. It is supported by the marked increase in production of the
178
siderophore bacillibactin in GA1 wild-type (Fig. 2c) and by the reduced growth of the ΔdhbC mutant,
179
repressed in bacillibactin synthesis, upon PVD addition (Fig. 2d, Supplementary Fig. 8). This last
180
result indicates that PVD in its ferric form cannot be taken up by GA1 despite the presence of
181
several transporters for exogenous siderophores in B. velezensis similar to those identified in B.
182
subtilis36,37 based on genome comparison (Supplementary Table 2). Therefore, we assumed that
183
iron stress mediated by PVD only induces a rather limited boost in PKs production. We validated
184
that such response is not due to iron starvation by supplementing GA1 culture with increasing
185
doses of the 2,2’-dipyridyl (DIP) chemical chelator that cannot be taken up by Bacillus cells (Fig.
186
2b). By contrast, the addition of E-PCH with a much lower affinity for iron does not activate
187
bacillibactin synthesis (Fig. 2c) and does not affect ΔdhbC growth at the concentrations used (Fig.
188
2d, Supplementary Fig. 8). We conclude that the activity of this compound referred to as secondary
189
siderophore is not related to iron-stress. If internalized, E-PCH can cause oxidative stress and
190
damage in other bacteria as reported for E. coli38,39. However, the absence of toxicity toward GA1
191
indicates that E-PCH is not taken up by Bacillus cells and thus clearly acts as a signal molecule
192
9
perceived at the cell surface. PKs boost also occurred upon addition of CFS obtained from other
193
Pseudomonas isolates producing pyochelin-type siderophores, such as P. protegens Pf-540.
194
However, PKs stimulation was similarly observed in response to P. tolaasii CH3641 which does not
195
form pyochelin, indicating that other unidentified BSMs may act as triggers in other strains
196
(Supplementary Fig. 9).
197
Enhanced motility as distance-dependent and surfactin-mediated response of Bacillus
198
Surfactin production is stimulated by CMR12a CFS and by pure E-PCH (Fig. 1a and
199
Supplementary Fig. 10). Based on mutant loss-of-function analysis, this multifunctional CLP does
200
not contribute to the antibacterial potential of B. velezensis (Fig. 1d and Supplementary Fig. 3) but
201
is known to be notably involved in developmental processes of multicellular communities such as
202
biofilm formation and motility42. Therefore, we wanted to test a possible impact of Pseudomonas on
203
the motile phenotype of B. velezensis upon co-cultivation on plates. We observed distance-
204
dependent enhanced motility on medium containing high agar concentrations (1.5% m/v) which
205
phenotypically resembles the sliding-type of motility illustrated by typical “van Gogh bundles"42 (Fig.
206
3a). This migration pattern is flagellum independent but depends on multiple factors including the
207
synthesis of surfactin which reduces friction at the cell-substrate interface42. We thus suspected
208
such enhanced motility to derive from an increased formation of the lipopeptide. This was
209
supported by the almost full loss of migration of the ΔsrfaA mutant in these interaction conditions
210
(Fig. 3b). Moreover, spatial mapping via MALDI-FT-ICR MS imaging confirmed a higher
211
accumulation of surfactin ions in the interaction zone and around the Bacillus colony when growing
212
at a short or intermediate distance from the Pseudomonas challenger, compared to the largest
213
distance where the motile phenotype is much less visible (Fig. 3c). These data indicate that Bacillus
214
cells in the microcolony perceive a soluble signal diffusing from the Pseudomonas colony over a
215
limited distance.
216
217
10
Interplay between CLPs drives antagonistic interactions
218
Besides modulating secondary metabolite synthesis, we further observed that confrontation with
219
Pseudomonas may also lead to some antagonistic outcomes. GA1 growth as planktonic cells is
220
slightly inhibited upon supplementation of the medium with 2% v/v CMR12a CFS but this inhibition
221
is much more marked at a higher dose (Supplementary Fig. 11). To identify the Pseudomonas
222
compound retaining such antibiotic activity, we tested the effect of CFS from various CMR12a
223
mutants impaired in the synthesis of lipopeptides and/or phenazines. Even if some contribution of
224
other compounds cannot be ruled out, it revealed that the CLP sessilin is mainly responsible for
225
toxic activity toward GA1 grown in liquid cultures but also when the two bacteria are grown at close
226
proximity on gelified EM medium (Fig. 4a). Nevertheless, we observed that the sessilin-mediated
227
inhibitory effect is markedly reduced by delaying CFS supplementation until 6 h of Bacillus culture
228
instead of adding it at the beginning of incubation (Fig. 4b). This suggested that early secreted
229
Bacillus compounds may counteract the toxic effect of sessilin. We hypothesized that surfactin can
230
play this role as it is the first detectable BSM to accumulate in significant amounts in the medium
231
early in the growth phase. We tested the surfactin-deficient mutant in the same conditions and
232
observed that its growth is still strongly affected indicating that no other GA1 compound may be
233
involved in toxicity alleviation. Chemical complementation with purified surfactins restored growth to
234
a large extent, providing further evidence for a protective role of the surfactin lipopeptide (Fig. 4b).
235
Such sessilin-dependent inhibition also occurred when bacteria were confronted on solid CAA
236
medium (Fig. 4c-I) favoring Pseudomonas BSM production. In these conditions, the formation of a
237
white precipitate in the interaction zone was observed with CMR12a wild-type but not when GA1
238
was confronted with the ΔsesA mutant (Fig. 4c). UPLC-MS analysis of ethanol extracts from this
239
white-line area confirmed the presence of sessilin ions but also revealed an accumulation of
240
surfactin from GA1 in the confrontation zone (Fig. 4d). The involvement of surfactins in precipitate
241
formation was confirmed by the absence of this white-line upon testing the ΔsrfaA mutant of GA1
242
11
(Fig. 4c-I, II). The loss of surfactin production and white-line formation was associated with a higher
243
sensitivity of the Bacillus colony to the sessilin toxin secreted by Pseudomonas. Altogether, these
244
data indicate that surfactin acts as a chemical trap and inactivates sessilin via co-aggregation into
245
insoluble complexes.
246
A similar CLP-dependent antagonistic interaction and white-line formation were observed
247
upon co-cultivation of GA1 with P. tolaasii strain CH36 producing tolaasin (Fig. 4c-II), a CLP
248
structurally very similar to sessilin (only differing by two amino acid residues, Supplementary Fig.
249
12). However, this chemical aggregation is quite specific regarding the type of CLP involved, since
250
it was not visible upon the interaction of GA1 with other Pseudomonas strains forming different CLP
251
structural groups that are not toxic for Bacillus (Fig. 4c-III, see Supplementary Fig. 12 and 13 for
252
identification and structures). Sessilin/tolaasin-dependent toxicity and white-line formation were
253
also observed when other surfactin-producing B. velezensis isolates were confronted with CMR12a
254
and CH36 (Supplementary Fig. 14 and 15, respectively). Although the chemical basis and the
255
stoichiometry of such molecular interaction remain to be determined, it probably follows the same
256
rules as observed for the association between sessilins/tolaasins and other endogenous
257
Pseudomonas CLPs such as WLIP or orfamides28 or between CLPs and other unknown
258
metabolites43,44.
259
260
BSMs-mediated interactions drive competitive root colonization
261
Our in vitro data point out how B. velezensis may modulate its secondary metabolome when
262
confronted with Pseudomonas. To appreciate the relevance of our findings in a more realistic
263
context, we next evaluated whether such BSMs interplay may also occur upon root co-colonization
264
of tomato plantlets and possibly impact Bacillus fitness. When inoculated independently, CMR12a
265
colonized roots more efficiently than GA1 within the first 3 days, most probably due to a higher
266
intrinsic growth rate45. Upon co-inoculation, the CMR12a colonization rate was not affected but GA1
267
12
populations were reduced compared to mono-inoculated plantlets (Fig. 5a). UPLC-MS analysis of
268
methanolic extracts prepared from co-bacterized roots (and surrounding medium) revealed
269
substantial amounts of E-PCH (Supplementary Fig. 16) indicating that the molecule is readily
270
formed under these conditions and could therefore also act as a signal in planta. Probably due to
271
the low populations of GA1, we could not detect Bacillus PKs and RiPPs in these extracts.
272
However, a significantly enhanced expression of gene clusters responsible for the synthesis of
273
bacillaene, difficidin and amylocyclicin was observed in GA1 cells co-inoculated with Pseudomonas
274
compared to single inoculation (Fig. 5b). It indicated that the metabolite response observed in GA1
275
in vitro cultures in EM medium may also occur upon competitive colonization where the bacteria
276
feed exclusively on root exudates.
277
Lipopeptides involved in interference interaction are also readily formed upon single and
278
dual root colonization (Fig. 5d). We hypothesized that the inhibitory effect of sessilin may impact the
279
colonization potential of GA1 in presence of CMR12a which was confirmed by the increase in GA1
280
populations co-inoculated with the ΔsesA mutant (Fig. 5c). Moreover, colonization by the ΔsrfaA
281
mutant is more impacted compared to WT when co-cultivated with CMR12a and a significant gain
282
in root establishment is recovered upon co-colonization with the ΔsesA mutant (Fig. 5d). The
283
sessilin-surfactin interplay thus also occurs in planta. Sessilin would confer a competitive
284
advantage to CMR12a during colonization by inhibiting GA1 development but efficient surfactin
285
production on roots may provide some protection to the Bacillus cells.
286
287
Discussion
288
It has been recently reported that Pseudomonas toxin delivery via Type VI secretion system
289
and antibiotic (2,4-diacetylphloroglucinol) production may impact biofilm formation and sporulation
290
in B. subtilis46,47. However, our current understanding of the molecular basis of interactions between
291
soil bacilli and pseudomonads is still rather limited. Here we show that the model species B.
292
13
velezensis can mobilize a substantial part of its secondary metabolome in response to
293
Pseudomonas competitors. To our knowledge, it is the first evidence for enhanced synthesis of
294
both broad-spectrum polyketides and RiPP in Bacillus upon a perception of other bacteria, in
295
contact-independent in vitro settings and upon competitive root colonization. This correlates with an
296
enhanced antibacterial potential which is of interest for biocontrol but which can also be considered
297
as a defensive strategy to persist in its natural competitive niche. Upon sensing Pseudomonas, B.
298
velezensis calls on its antibiotic arsenal but also recruits its surfactin lipopeptide to improve
299
multicellular mobility. This may be viewed as an escape mechanism enabling Bacillus cells to
300
relocate after detecting harmful challengers. Improved motility of B. subtilis has been already
301
described upon sensing competitors such as Streptomyces venezuelae7,48 but no relationship was
302
established with enhanced production of BSMs potentially involved in the process. We also
303
highlight a new role for surfactin acting as a chemical shield to counteract the toxicity of exogenous
304
CLPs. Intraspecies CLP co-precipitation has been reported23 but our results make sense of this
305
phenomenon in the context of interference interaction between two different genera. In planta, this
306
new function of surfactin contributes to Bacillus competitiveness for root invasion. This has to be
307
added to other previously reported implications of surfactin in B. subtilis interspecies interactions,
308
such as interfering with the growth of closely related species in synergy with cannibalism toxins49,
309
inhibiting the development of Streptomyces aerial hyphae50, or participating in the expansion and
310
motility of the interacting species47. We postulate that such Bacillus metabolite response largely
311
contributes to mount a multi-faceted defensive strategy in order to gain fitness and persistence in
312
its natural competitive niche.
313
Furthermore, we exemplify that PKs stimulation in B. velezensis is mainly mediated by the
314
Pseudomonas secondary siderophore pyochelin, although it cannot be excluded that other
315
secreted products may also play a role. Bacillus perceives pyochelin in a way independent of iron
316
stress and piracy, indicating that beyond its iron-scavenging function, this siderophore may also act
317
14
as infochemical in interspecies cross-talk. In the pairwise system used here, E-PCH signaling
318
superimposes the possible effect of iron limitation in the external medium which may also result in
319
enhanced production of antibacterial metabolites by Bacillus, as occasionally reported51. That said,
320
due to the limitation in bioavailable iron, almost all known rhizobacterial species have adapted to
321
produce their iron-scavenging molecules to compete for this essential element52–54. Siderophore
322
production is thus widely conserved among soil-borne bacteria55. It means that upon recognition of
323
exogenous siderophores, any isolate may somehow identify surrounding competitors. However,
324
some of these siderophores are structurally very variable and almost strain-specific (such as PVDs
325
from fluorescent pseudomonads) while some others are much more widely distributed across
326
species and even genera (enterobactin-like, citrate)52. In both cases, their recognition would not
327
provide proper information about the producer because they are too specific or too general,
328
respectively. Interestingly, the synthesis of E-PCH and its structurally very close enantio form is
329
conserved in several but not all Pseudomonas sp.56–58 as well as in a limited number of species
330
belonging to other genera such as Burkholderia59 and Streptomyces60,61. We therefore hypothesize
331
that Bacillus may have evolved some chelator-sensing systems targeting siderophores that are
332
conserved enough to be detected but restricted to specific microbial phylogenetic groups. With this
333
mechanism, soil bacilli would rely on siderophores as public goods to accurately identify
334
competitors and respond in an appropriate way like remodeling its BSM secretome. This novel
335
concept of chelator sensing represents a new facet of siderophore-mediated social interactions.
336
Whether it is used for other secondary siderophores than E-PCH and if so, whether this adaptative
337
trait can be generalized to other soil-dwelling species deserves to be further investigated given its
338
possible impact on soil bacterial ecology. Beyond the notion of specialized metabolites, we point
339
out unsuspected functions for some bacterial small molecules in the context of interactions between
340
clades that are important members of the plant-associated microbiome.
341
342
15
Methods
343
Bacterial strains and growth conditions
344
Strains and plasmids used in this study are listed in Supplementary Table 3. B. velezensis strains
345
were grown at 30 °C on, half diluted, recomposed exudate solid medium (EM)22 or in liquid EM with
346
shaking (160 rpm). Deletion mutants of B. velezensis were selected on appropriate antibiotics
347
(chloramphenicol at 5 µg/ml, phleomycin at 4 µg/ml, kanamycin at 25 µg/ml) on Lysogeny broth
348
(LB) (10 g l-1 NaCl, 5 g l-1 yeast extract and 10 g l-1 tryptone). Pseudomonas sp. strains were grown
349
on King B (20 g l-1 of bacteriological peptone, 10 g l-1 of glycerol and 1.5 g l-1 of K2HPO4, 1.5 g l-1 of
350
MgSO4.7H2O, pH = 7) and casamino acid (CAA) solid and liquid medium (10 g l-1 casamino acid,
351
0.3 g l-1 K2HPO4, 0.5 g l-1 MgSO4 and pH = 7) with shaking (120 rpm), at 30 °C. The
352
phytopathogenic bacterial strains were grown on LB and EM solid and liquid media and with
353
shaking (150 rpm), at 30 °C.
354
Construction of deletion mutants of B. velezensis GA1
355
All deletion mutants were constructed by marker replacement. Briefly, 1 kb of the upstream region
356
of the targeted gene, an antibiotic marker (chloramphenicol, phleomycin or kanamycin cassette)
357
and downstream region of the targeted gene were PCR amplified with specific primers
358
(Supplementary Table 3). The three DNA fragments were linked by overlap PCR to obtain a DNA
359
fragment containing the antibiotic marker flanked by the two homologous recombination regions.
360
This latter fragment was introduced into B. velezensis GA1 by natural competence induced by
361
nitrogen limitation62. Homologous recombination event was selected by chloramphenicol resistance
362
(phleomycin resistance for double mutants or kanamycin resistance for triple mutants) on LB
363
medium. All gene deletions were confirmed by PCR analysis with the corresponding UpF and DwR
364
specific primers and by the loss of the corresponding BSMs production.
365
16
Transformation of the B. velezensis GA1 strain was performed following the protocol
366
previously described62 with some modifications. One fresh GA1 colony was inoculated into LB liquid
367
medium at 37 °C (160 rpm) until reaching an OD600nm of 1.0. Afterwards, cells were washed one
368
time with peptone water and one time with a modified Spizizen minimal salt liquid medium (MMG)
369
(19 g l-1 K2HPO4 anhydrous; 6 g l-1 KH2PO4; 1 g l-1 Na3 citrate anhydrous; 0.2 g l-1 MgSO4 7H2O; 2 g
370
l-1 Na2SO4; 50 µM FeCl3 (sterilized by filtration at 0.22 µm); 2 µM MnSO4; 8 g l-1 glucose; 2 g l-1 L-
371
glutamic acid; pH 7.0), 1 µg of DNA recombinant fragment was added to the GA1 cells suspension
372
adjusted to an OD600nm of 0.01 into MMG liquid medium. One day after incubation at 37 °C with
373
shaking at 165 rpm, bacteria were spread on LB plates supplemented with the appropriate
374
antibiotic to select positive colonies.
375
Construction of deletion mutants of Pseudomonas sp. CMR12a
376
E-PCH and PVD mutants of Pseudomonas sp. CMR12a were constructed using the I-SceI system
377
and the pEMG suicid vector63,64. Briefly, the upstream and downstream region flanking the pchA
378
(C4K39_5481) or the pvdI (C4K39_6027) genes were PCR amplified (primers listed in the
379
Supplementary Table 3), linked via overlap PCR and inserted into the pEMG vector. The resulting
380
plasmid (Supplementary Table 3) was integrated by conjugation into the Pseudomonas sp.
381
CMR12a chromosome via homologous recombination. Kanamycin (25µg/mL) resistant cells were
382
selected on King B agar plates and transformed by electroporation with the pSW-2 plasmid
383
(harboring I-SceI system). Gentamycin (20µg/ml) resistant colonies on agar plates were transferred
384
to King B medium with and without kanamycin to verify the loss of the antibiotic (kanamycin)
385
resistance. Pseudomonas mutants were identified by PCR with the corresponding UpF and DwR
386
specific primers and via the loss of E-PCH or/and PVD production (see section Secondary
387
metabolites analysis).
388
Pseudomonas sp. cell-free supernatant
389
17
Pseudomonas sp. strains were grown overnight on LB solid medium, at 30 °C. The cell suspension
390
was adjusted to OD600nm 0.05 by resuspension in 100 ml of CAA and when appropriate
391
supplemented with 20 µg/l of FeCl3.6H2O (iron supplementation). Cultures were shaken at 120 rpm
392
at 30 °C for 48 h and then centrifuged at 5000 rpm at room temperature (22 °C) for 20 min. The
393
supernatant was filter-sterilized (0.22 µm pore size filters) and stored at -20 °C until use.
394
Dual interactions
395
B. velezensis strains were grown overnight on LB solid medium, at 30 °C. Cells were resuspended
396
in 2 ml of EM liquid medium to a final OD600nm of 0.1 in which 1, 2, or 4% v/v (depending on the
397
experiment and indicated in the figures legends) of Pseudomonas CFS were added while the
398
control remained un-supplemented. B. velezensis liquid cultures were shaken in an incubator at
399
300 rpm at 30 °C for 24 h. Additionally, 2 ml of the (co-)culture supernatants were sampled at 8 h
400
and 24 h, centrifugated at 5000 rpm at room temperature (approx. 22 °C) for 10 min to extract
401
supernatants and collect the cells. Further, cell-free (co-)culture supernatants were filter-sterilized
402
(0.22µm) and used for analytical analysis of secondary metabolites and antibacterial assays. For
403
some experiments using 2H-bae as a marker, the CFS obtained from the double mutant sessilins
404
and orfamides (ΔsesA-ofaBC) was used instead of CFS from CMR12a wild-type because it yielded
405
a higher response and lower inhibition interferences by CLPs. The remaining cells, after
406
supernatant collection, were stored at -80 °C to avoid RNA degradation, until performing RT-qPCR
407
analysis.
408
Antimicrobial activity assays
409
Antibacterial activity of the B. velezensis supernatant generated after dual interaction with
410
Pseudomonas CFS was tested against X. campestris pv. campestris and C. michiganensis subsp.
411
michiganensis. The activity of co-culture supernatants was quantified in microtiter plates (96-well)
412
18
filled with 250 µl of LB liquid medium, inoculated at OD600nm = 0.1 with X. campestris pv. campestris
413
and C. michiganensis subsp. michiganensis and supplemented with 2% or 6% v/v of the
414
supernatants, respectively. The activity of (co-)culture supernatants was estimated by measuring
415
the pathogen OD600nm every 30min during 24 h with a Spectramax® (Molecular Devices,
416
Wokingham, UK), continuously shaken, at 30 °C. For estimating the activity of co-culture
417
supernatants on a solid medium, 5 µl supernatant was applied to a sterile paper disk (5 mm
418
diameter). After drying, disks were placed on LBA square plates previously inoculated with a
419
confluent layer of X. campestris pv. campestris, C. michiganensis subsp. michiganensis, P.
420
carotovorum, P. fuscovaginae, P. cichorii, A. tumefaciens or R. fascians. LB liquid medium was
421
used as a negative control. Plates were incubated at 25 °C for 48 h. Three repetitions were done
422
and the inhibition zones from the edge of the paper discs to the edge of the zone were measured.
423
Antibacterial activity of the different Pseudomonas strains on the B. velezensis strains growth were
424
tested by adding different % (v/v) of the corresponding Pseudomonas CFS in microtiter plates (96-
425
well) filled with 250 µl of EM liquid medium. B. velezensis OD600nm after 7 h was measured with a
426
Spectramax® (Molecular Devices, Wokingham, UK).
427
RNA isolation and RT-qPCR
428
RNA extraction and DNAse treatment were carried out using the NucleoSpin RNA Kit (Macherey
429
Nagel, Germany), following the Gram + manufacturer’s protocol. RNA quality and quantity were
430
performed with Thermo scientific NanoDrop 2000 UV-vis Spectrophotometer. Primer 3 program
431
available online was used for primer design and primers were synthesized by Eurogentec. The
432
primer efficiency was evaluated and primer pairs showing an efficiency between 90 and 110% in
433
the qPCR analysis were selected. Reverse transcriptase and RT-qPCR reactions were conducted
434
using the Luna® Universal One-Step RT-qPCR Kit (New England Biolabs, Ipswich, MA, United
435
States). The reaction was performed with 50 ng of total RNA in a total volume of 20 µL: 10 µL of
436
19
luna universal reaction mix, 0.8 µL of each primer (10 µM), 5 µL of cDNA (50ng), 1 µL of RT
437
Enzyme MIX, 2.4 µl of Nuclease-free water. The thermal cycling program applied on the ABI
438
StepOne was: 55 °C for 10 min, 95 °C for 1 min, 40 cycles of 95 °C for 10 s and 60 °C for 1 min,
439
followed by a melting curve analysis performed using the default program of the ABI StepOne
440
qPCR machine (Applied Biosystems). The real-time PCR amplification was run on the ABI step-one
441
qPCR instrument (Applied Biosystems) with software version 2.3. The relative gene expression
442
analysis was conducted by using the 2ΔCt method65 with the gyrA gene as a housekeeping gene to
443
normalize mRNA levels between different samples. The target genes in this study were dfnA, baeJ
444
and acnA.
445
Secondary metabolite analysis
446
For detection of BSMs, B. velezensis and Pseudomonas sp. were cultured in EM and CAA as
447
described above. After an incubation period of 24 h for B. velezensis, if not differentially indicated,
448
and 48 h for Pseudomonas sp., supernatants of the bacteria were collected and analyzed by UPLC
449
MS and UPLC qTOF MS/MS. Metabolites were identified using Agilent 1290 Infinity II coupled with
450
DAD detector and Mass detector (Jet Stream ESI-Q-TOF 6530) in both negative and positive mode
451
with the parameter set up as follows: parameters: capillary voltage: 3.5 kV; nebulizer pressure: 35
452
psi; drying gas: 8 l/min; drying gas temperature: 300 °C; flow rate of sheath gas: 11 l/min; sheath
453
gas temperature: 350 °C; fragmentor voltage: 175 V; skimmer voltage: 65 V; octopole RF: 750 V.
454
Accurate mass spectra were recorded in the range of m/z = 40-250. An C18 Acquity UPLC BEH
455
column (2.1 × 50 mm × 1.7 µm; Waters, milford, MA, USA) was used at a flow rate of 0.3 ml/min
456
and a temperature of 40 °C. The injection volume was 20 µl and the diode array detector (DAD)
457
scanned a wavelength spectrum between 190 and 600 nm. A gradient of 0.1% formic acid water
458
(solvent A) and acetonitrile acidified with 0.1% formic acid (solvent B) was used as a mobile phase
459
with a constant flow rate at 0.45 ml/min starting at 10% B and raising to 100% B in 20 min. Solvent
460
20
B was kept at 100% for 2 min before going back to the initial ratio. Secondary metabolite
461
quantification was performed by using UPLC–MS with UPLC (Acquity H-class, Waters) coupled to
462
a single quadrupole mass spectrometer (SQD mass analyzer, Waters) using a C18 column
463
(Acquity UPLC BEH C18 2.1 mm × 50 mm, 1.7 µm). Elution was performed at 40 °C with a
464
constant flow rate of 0.6 ml/min using a gradient of Acetonitrile (solvent B) and water (solvent A)
465
both acidified with 0.1% formic acid as follows: 2 min at 15% B followed by a gradient from 15% to
466
95% during 5 min and maintained at 95% up to 9.5 min before going back to initial conditions at 10
467
min during 2 min before next injection. Compounds were detected in both electrospray positive and
468
negative ion mode by setting SQD parameters as follows: cone voltage: 60V; source temperature
469
130 °C; desolvation temperature 400 °C, and nitrogen flow: 1000 l/h with a mass range from m/z
470
300 to 2048. 3D chromatograms were generated using the open-source software MzMine 266.
471
Bioguided fractionation
472
Pseudomonas CFS were concentrated with a C18 cartridge ‘Chromafix, small’ (Macherey-Nagel,
473
Düren, Germany). The column was conditioned with 10 ml of MeOH followed by 10 ml of milliQ
474
water. Then, 20 ml of supernatant flowed through the column. The metabolites were eluted with 1
475
ml of a solution of increasing acetonitrile/water ratio from 5:95 to 100:0 (v/v). The triggering effect of
476
these fractions on Bacillus 2H-bae production was tested in 48 wells microplate containing 1 ml of
477
EM medium inoculated with B. velezensis GA1 (OD600nm = 0.1) and 4% v/v of aforementioned
478
Pseudomonas fractions, growing for 24 h, with shaking at 300 rpm and 30 °C. Afterward, the
479
production of 2H-bae was quantified compared to controls and crude supernatant.
480
Purification of E-PCH and PVD
481
PVD and E-PCH were purified in two steps. Firstly, Pseudomonas CFS were concentrated with a
482
C18 cartridge (as indicated in section Bioguided fractionation) and eluted with 2 times 2 ml of a
483
21
solution of water and ACN (15 and 30% of ACN (v/v)). Secondly, the fractions were injected on
484
HPLC for purification performed on an Eclipse+ C18 column (L = 150 mm, D = 3.0 mm, Particles
485
diameter 5 µm) (Agilent, Waldbronn, Germany). The volume injected was 100 µl. The UV-Vis
486
absorbance was measured with a VWD Agilent technologies 1100 series (G1314A) detector
487
(Agilent, Waldbronn, Germany). The lamp used was a Deuterium lamp G1314 Var Wavelength Det.
488
(Agilent, Waldbronn, Germany). Two wavelengths were selected: 320 nm, used for the detection of
489
E-PCH, and 380 nm, used for the detection of PVD. The fractions containing the PVD and E-PCH
490
were collected directly at the detector output. Further, the purity of the samples was verified by two
491
detectors, a diode array detector (DAD) 190 to 601 nm (steps: 1 nm) and a Q-TOF (tandem mass
492
spectrometry, quadrupole and Time of flight detector combined) (Agilent, Waldbronn, Germany).
493
Electrospray ionization was performed in positive mode (ESI+) (Dual AJS ESI) (Vcap = 3500 V,
494
Nozzle Voltage = 1000 V), with a mass range from m/z 200 to 1500. Finally, the concentration of
495
PVD and E-PCH were estimated by utilization of Beer-Lambert law formula, A = Ɛlc (A:
496
absorbance; Ɛ: molar attenuation coefficient or absorptivity of the attenuating species; l: optical
497
path length and c: concentration of molecule). l value for E-PCH and PVD is 1 cm while Ɛ is 4000
498
L.mol-1.cm-1 or 16000 L.mol-1.cm-1, respectively67. The absorbance was measured with VWR, V-
499
1200 Spectrophotometer, at 320 nm (pH = 8) for E-PCH and 380nm (pH = 5 ) for PVD67. Further,
500
the absorbance value was used for calculating the final concentration. The fragmentation pattern of
501
Pseudomonas sp. CMR12a PVD was obtained by UPLC MS/MS analysis of m/z = 1288.5913 ion in
502
positive mode with fragmentation energy at 75 V and compared to the one described in P.
503
protegens Pf-540.
504
Confrontation, white line formation and motility test
505
For confrontation assays on agar plates, Bacillus and Pseudomonas strains were grown overnight
506
in EM and CAA liquid mediums, respectively. After bacterial washing in peptone water and
507
22
adjustment of OD600nm to 0.1, 5 µl of bacterial suspension was spotted at 1 mm, 5 mm and 7.5 mm
508
distance onto an EM agar plate. For the white line formation experiments, B. velezensis line was
509
applied with a cotton stick and 5 µl of Pseudomonas sp. cell suspensions were spotted at a 5 mm
510
distance onto CAA agar plates. Plates were incubated at 30 °C and images taken after 24 h.
511
Photographs were captured using CoolPix camera (NiiKKOR 60x WIDE OPTICAL ZOOM EDVR
512
4.3-258 mm 1:33-6.5).
513
MALDI-FT-ICR MS imaging
514
Mass spectrometry images were obtained as recently described68 using a FT-ICR mass
515
spectrometer (SolariX XR 9.4T, (Bruker Daltonics, Bremen, Germany)) mass calibrated from 200
516
m/z to 2,300 m/z to reach a mass accuracy of 0.5 ppm. Region of interest from agar microbial
517
colonies was directly collected from the Petri dish and transferred onto an ITO Glass slide (Bruker,
518
Bremen, Germany), previously covered with double-sided conductive carbon tape. The samples
519
were dried under vacuum and covered with an α-cyano-4-hydroxycinnamic acid (HCCA) matrix
520
solution at 5 mg/mL (70 : 30 acetonitrile : water v/v). In total, 60 layers of HCCA matrix were
521
sprayed using the SunCollect instrument (SunChrom, Friedrichsdorf, Germany). FlexImaging 5.0
522
(Bruker Daltonics, Bremen, Germany) software was used for MALDI-FT-ICR MS imaging
523
acquisition, with a pixel step size for the surface raster set to 100 µm.
524
In planta competition
525
For in planta studies, tomato seeds (Solanum lycopersicum var. Moneymaker) were sterilized in
526
75% ethanol with shaking for 2 min. Subsequently, ethanol was removed and seeds were added to
527
the 50 ml sterilization solution (8.5 ml of 15% bleach, 0.01 g of Tween 80 and 41.5 ml of sterile
528
ultra-pure water) and shaken for 10 min. Seeds were thereafter washed five times with water to
529
eliminate stock solution residues. Further, seeds were placed on square Petri dishes (5
530
23
seeds/plate) containing Hoagland solid medium (14 g/l agar, 5 ml stock 1 (EDTA 5,20 mg/l;
531
FeSO4x7H2O 3,90 mg/l; H3B03 1,40 mg/l; MgSO4x7H2O 513 mg/l; MnCl2x4H2O 0,90 mg/l,
532
ZnSO4x7H2O 0,10 mg/l; CuSO4x5H2O 0,05 mg/l; 1 ml in 50 ml stock 1, NaMo04x2H2O 0,02 mg/l 1
533
ml in 50 ml stock 1), 5 ml stock 2 (KH2PO4 170 mg/l; 5 ml stock 3: KN03 316 mg/l, Ca(NO3)2 4H2O
534
825 mg/l), pH = 6,5) and placed in the dark for three days. Afterwards, 10 seeds were inoculated
535
with 2 µl of overnight culture (OD600 = 0.1) of the appropriate strains (control) or with a mix of
536
Bacillus and Pseudomonas cells (95:5 ration) (interaction) and grown at 22 °C under a 16/8 h
537
night/day cycle with constant light for three days. After the incubation period, to determine bacterial
538
colonization levels, bacteria from roots of six plants per condition were detached from roots by
539
vortexing for 1 min in peptone water solution supplemented with 0.1% of Tween. Serial dilutions
540
were prepared and 200 µl of each were plated onto LB medium using plating beads. After 24 h of
541
incubation at 30 °C for Pseudomonas and at 42 °C for Bacillus, colonies were counted.
542
Colonization results (six plants per strain) were log-transformed and statistically analyzed. Three
543
independent assays were performed with six plants each for in planta competition assays. To
544
measure bacterial BSMs production in planta, a rectangle part (1 x 2.5 cm) of medium close to the
545
tomato roots was sampled. BSMs were extracted for 15 min, with 1.5 ml of acetonitrile (85%). After
546
centrifugation for 5 min at 4000 rpm, the supernatant was recovered for UPLC-MS analysis as
547
previously described.
548
Statistical analysis
549
Statistical analyses were performed using GraphPad PRISM software with Student paired T-test or
550
Mann-Whitney test. For multiple comparisons, one-way ANOVA and Tukey tests were used in
551
RStudio 1.1.423 statistical software environment (R language version 4.03)69.
552
553
24
Figure legends
554
Figure 1. Stimulation of BSMs production by B. velezensis GA1 and enhanced anti-bacterial
555
activities in response to Pseudomonas sp. CMR12a secreted metabolites. a. UPLC-MS
556
extracted ions chromatograms (EIC) illustrating the relative abundance of ions corresponding to
557
non-ribosomal metabolites produced by B. velezensis GA1 in CFS-supplemented (2% v/v) EM
558
medium (blue) compared to un-supplemented cultures used as control (red). Red-coloured parts in
559
the representation of lipopeptides and macrolactin illustrate the variable structural traits explaining
560
the occurrence of naturally co-produced variants (multiple peaks) b. Fold increase in GA1 BSM
561
production upon addition of CMR12a CFS (2% v/v) compared to un-supplemented cultures (fold
562
change = 1, red line). Data were calculated based on the relative quantification of the compounds
563
by UPLC-MS (peak area) in both conditions. Mean values were calculated from data obtained in
564
three cultures (repeats) from two independent experiments (n = 6). Statistical significance was
565
calculated using Mann–Whitney test where ‘’****’’ represents significant difference at P<0.0001. c.
566
Enhanced Anti-Xanthomonas campestris (I and II) and anti-Clavibacter michiganensis (III and IV)
567
activities of GA1 extracts (cell-free culture supernatant) after growth in CMR12a CFS-
568
supplemented medium (GA1+CFS) compared to control (GA1). It was assessed both on plates by
569
the increase in inhibition zone around paper disc soaked with 5 µl the GA1 extracts (I and III) and in
570
liquid cultures of the pathogens by reduction of growth upon addition of 4% (v/v) of GA1 extracts (II
571
and IV). Data are from one representative experiment and similar results were obtained in two
572
independent replicates. d. Antibacterial activities of extracts from GA1 WT and mutants impaired in
573
production of specific BSMs. Metabolites not produced by the different mutants are illustrated with
574
red boxes in the table below. All values represent means with error bars indicating SD calculated
575
on data from three cultures (repeats) in two independent experiments (n = 6). Letters a to d indicate
576
statistically significant differences according to one-way analysis of variance (ANOVA) and Tukey’s
577
HSD test (Honestly significantly different, α = 0.05). e. Differential expression of the acnA gene
578
25
encoding the amylocylicin precursor, upon supplementation with CMR12a CFS compared to GA1
579
un-supplemented culture. Mean and SD values, n = 6, “**” indicates statistical significance
580
according to Mann–Whitney test, P<0.01.
581
582
Figure 2: E-PCH as main Pseudomonas trigger of anti-bacterial activity boosted in B.
583
velezensis GA1. a, Effect of GA1 culture supplementation with CFS (2% v/v) from CMR12a WT
584
and various mutants on dihydrobacillaene (2H-bae) production. Metabolites specifically repressed
585
in the different CMR12a mutants are illustrated by red boxes. Fold changes were calculated based
586
on relative quantification of the compounds by UPLC-MS (peak area) in treated cultures compared
587
to un-supplemented controls (fold change = 1, red line). Data are means and SE calculated from
588
three replicate cultures in two (n = 6) or three (n = 9) independent experiments and different letters
589
indicate statistically significant differences (ANOVA and Tukey’s test, α = 0.05). b, Differential
590
production of 2H-bae after addition of 0.35 µM pure PVD, 1.4 µM pure E-PCH, 4% v/v
591
Pseudomonas sp. CMR12a CFS (CFS CAA), CMR12a CFS from iron supplemented culture (CFS
592
CAA+Fe) and different concentration of the iron-chelating agent 2,2'-dipyridyl (DIP). Data are
593
expressed and were statistically treated as described in a with n = 6 in all treatments. c, Dose-
594
dependent effect of pure PVD and E-PCH on bacillibactin and 2H-bae production. GA1 cultures
595
were supplemented with the indicated concentrations of HPLC-purified CMR12a siderophores.
596
Experiments were replicated and data statistically processed as described in b. d, Impact of the
597
addition of pure PVD and E-PCH on the growth of GA1 WT and its ΔdhbC mutant repressed in
598
bacillibactin synthesis. Pseudomonas siderophores were added at a final concentration similar to
599
the one obtained by adding CMR12a CFS at 4% v/v. Means and SD are from three replicates. See
600
Supplementary Figure 8 for detailed data and statistical significance.
601
602
26
Figure 3: Distance- and surfactin-dependent enhanced motility of B. velezensis GA1 in
603
interaction with Pseudomonas CMR12a. a, GA1 motility phenotype on EM gelified medium when
604
cultured alone (left panel) or in confrontation with CMR12a at a short distance (1 cm) (right panel).
605
b, Motility pattern of GA1 or his ΔsrfaA surfactin deficient mutant in confrontation with CMR12a at a
606
short distance. c, MALDI FT-ICR MSI (Mass spectrometry Imaging) heatmaps showing spatial
607
localization and relative abundance of ions ([M+Na]+) corresponding to the C14 surfactin homolog
608
(most abundant) when B. velezensis GA1 is in confrontation with CMR12a at increasing distances
609
(one biological replicate).
610
611
Figure 4: Surfactin attenuates sessilin-mediated toxicity via white-line formation. a, I.
612
Polarized inhibition of GA1 micro-colony development upon co-cultivation at close contact with
613
CMR12a colonies on EM plates. II. Inhibition of GA1 cell growth in EM liquid culture supplemented
614
with 6% v/v of CFS prepared from CMR12a wild-type or mutants repressed in the synthesis of
615
orfamides and phenazines (ΔofaBC-phz), sessilins (ΔsesA), sessilins and orfamides (ΔsesA-
616
ofaBC), sessilins and phenazines (ΔsesA-phz), or of all compounds (ΔsesA-ofaBC-phz). Data show
617
mean and SD calculated from two independent experiments each with three culture replicates (n =
618
6) and different letters indicate statistically significant differences (ANOVA and Tukey’s test, α =
619
0.05). b, Growth inhibition of GA1 WT and ΔsrfaA mutant upon delayed supplementation (added 6
620
h after incubation start) with CFS from CMR12a WT alone or together with pure surfactin as
621
chemical complementation) and with CFS from the sessilin mutant (ΔsesA). Un-supplemented
622
cultures of GA1 were used as control. Experiments were replicated and data statistically processed
623
as described in a. c, White line formation and/or Bacillus inhibition observed upon confrontation of
624
GA1 WT or the surfactin mutant ΔsrfaA with (I) CMR12a or its ΔsesA derivative, (II) P. tolaasii
625
CH36 or its tolaasin defective mutant ΔtolA and (III) other Pseudomonas CLP producers WCU-84,
626
27
SS101, BW11M1, RW10S2. CLPs produced by the individual Pseudomonas strains are mentioned
627
in the chart below. d, 3D representation of UPLC-MS analysis of CLPs that are present in the
628
white-line zone between GA1 and CMR12a (I). It shows the specific accumulation of sessilin and
629
surfactin molecular ions (one biological replicate).
630
631
Figure 5: Competitive colonization assays support the roles of BSMs in Bacillus-
632
Pseudomonas interaction in planta. a, GA1 and CMR12a cell populations as recovered from
633
roots at 3 days post-inoculation (dpi) of tomato plantlets when inoculated alone (GA1, CMR12a) or
634
co-inoculated (co-inoculation). Box plots were generated based on data from three independent
635
assays each involving at least 4 plants per treatment (n=16). The whiskers extend to the minimum
636
and maximum values, and the midline indicates the median. Statistical differences between the
637
treatments were calculated using Mann–Whitney test and ‘’****’’ and ‘’***’’ represent significant
638
differences at P<0.0001 and P<0.001, respectively. b, In planta (3 dpi on tomato roots) relative
639
expression of the dfnA, baeJ and acnA genes responsible for the synthesis of respectively
640
(oxy)difficidin, 2H-bae and amylocyclicin. Graphs show the mean and SD calculated from three
641
biological replicates (n = 3) each involving six plants. Fold change = 1 as red line corresponds to
642
gene expressions in GA1 inoculated alone on roots used as control conditions. Statistical
643
comparison between data in co-colonization setting and control conditions was performed based on
644
T-test (*, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001). c. UPLC-MS EIC illustrating relative in
645
planta production of sessilins and surfactins by monocultures of B. velezensis GA1 (GA1) and co-
646
cultures of wild-types (GA1+CMR12a) and B. velezensis GA1 and Pseudomonas sp. CMR12a
647
impaired in sessilins production (GA1+ΔsesA). d, Cell populations recovered at 3 dpi for GA1 WT
648
(GA1) or the surfactins impaired mutant (ΔsrfaA) co-inoculated with CMR12a WT (CMR12a) or its
649
sessilins KO mutant (ΔsesA). See a for replicates and statistics (**, P<0.01).
650
28
651
References
652
1.
Nayfach, S. et al. A genomic catalog of Earth’s microbiomes. Nat. Biotechnol. (2020)
653
doi:10.1038/s41587-020-0718-6.
654
2.
Fierer, N. Embracing the unknown: disentangling the complexities of the soil microbiome.
655
Nat. Rev. Microbiol. 15, 579–590 (2017).
656
3.
Schmidt, R., Ulanova, D., Wick, L. Y., Bode, H. B. & Garbeva, P. Microbe-driven chemical
657
ecology: past, present and future. ISME Journal.13, 2656–2663 (2019).
658
4.
Tyc, O., Song, C., Dickschat, J. S., Vos, M. & Garbeva, P. The ecological role of volatile and
659
soluble secondary metabolites produced by soil bacteria. Trends Microbiol. 25, 280–292
660
(2017).
661
5.
Bernal, P., Llamas, M. A. & Filloux, A. Type VI secretion systems in plant-associated
662
bacteria. Environ. Microbiol. 20, 1–15 (2018).
663
6.
Abreu, N. A. & Taga, M. E. Decoding molecular interactions in microbial communitiesa.
664
FEMS Microbiol. Rev. 40, 648–663 (2016).
665
7.
Andrić, S., Meyer, T. & Ongena, M. Bacillus responses to plant-sssociated fungal and
666
bacterial communities. Front. Microbiol. 11, (2020).
667
8.
Mendes, R., Garbeva, P. & Raaijmakers, J. M. The rhizosphere microbiome: significance of
668
plant beneficial, plant pathogenic, and human pathogenic microorganisms. FEMS Microbiol.
669
Rev. 37, 634–663 (2013).
670
9.
Müller, D. B., Vogel, C., Bai, Y. & Vorholt, J. A. The plant microbiota: systems-level insights
671
and perspectives. Annu. Rev. Genet. 50, 211–234 (2016).
672
10.
Penha, R. O., Vandenberghe, L. P. S., Faulds, C., Soccol, V. T. & Soccol, C. R. Bacillus
673
lipopeptides as powerful pest control agents for a more sustainable and healthy agriculture:
674
29
recent studies and innovations. Planta 251, 1–15 (2020).
675
11.
Grubbs, K. J. et al. Large-Scale bioinformatics analysis of Bacillus genomes uncovers
676
conserved roles of natural products in bacterial physiology. mSystems 2, 1–18 (2017).
677
12.
Harwood, C. R., Mouillon, J.-M. M., Pohl, S. & Arnau, J. Secondary metabolite production
678
and the safety of industrially important members of the Bacillus subtilis group. FEMS
679
Microbiol. Rev. 42, 721–738 (2018).
680
13.
Ye, M. et al. Characteristics and application of a novel species of Bacillus : Bacillus
681
velezensis. ACS Chem. Biol. 13, 500–505 (2018).
682
14.
Rabbee, M. et al. Bacillus velezensis: A valuable member of bioactive molecules within plant
683
microbiomes. Molecules 24, 1–13 (2019).
684
15.
Pieterse, C. M. J. et al. Induced systemic resistance by beneficial microbes. Annu. Rev.
685
Phytopathol. 52, 347–375 (2014).
686
16.
Köhl, J., Kolnaar, R. & Ravensberg, W. J. Mode of action of microbial biological control
687
agents against plant diseases: relevance beyond efficacy. Front. Plant Sci. 10, 845 (2019).
688
17.
Traxler, M. F. & Kolter, R. Natural products in soil microbe interactions and evolution. Nat.
689
Prod. Rep. 32, 956–970 (2015).
690
18.
Raaijmakers, J. M. & Mazzola, M. Diversity and natural functions of antibiotics produced by
691
beneficial and plant pathogenic bacteria. Annu. Rev. Phytopathol. 50, 403–424 (2012).
692
19.
Li, Y. & Rebuffat, S. The manifold roles of microbial ribosomal peptide-based natural
693
products in physiology and ecology. J. Biol. Chem. 295, 34–54 (2020).
694
20.
Raaijmakers, J. M., de Bruijn, I., Nybroe, O. & Ongena, M. Natural functions of lipopeptides
695
from Bacillus and Pseudomonas: More than surfactants and antibiotics. FEMS Microbiol.
696
Rev. 34, 1037–1062 (2010).
697
21.
Blin, K. et al. antiSMASH 5.0: updates to the secondary metabolite genome mining pipeline.
698
Nucleic Acids Res. 47, W81–W87 (2019).
699
30
22.
Nihorimbere, V. et al. Impact of rhizosphere factors on cyclic lipopeptide signature from the
700
plant beneficial strain Bacillus amyloliquefaciens S499. FEMS Microbiol. Ecol. 79, 176–191
701
(2012).
702
23.
D’aes, J. et al. To settle or to move? The interplay between two classes of cyclic lipopeptides
703
in the biocontrol strain Pseudomonas CMR12a. Environ. Microbiol. 16, 2282–2300 (2014).
704
24.
Hua, G. K. H. & Höfte, M. The involvement of phenazines and cyclic lipopeptide sessilin in
705
biocontrol of Rhizoctonia root rot on bean (Phaseolus vulgaris) by Pseudomonas sp.
706
CMR12a is influenced by substrate composition. Plant Soil 388, 243–253 (2015).
707
25.
Perneel, M. et al. Characterization of CMR5c and CMR12a, novel fluorescent Pseudomonas
708
strains from the cocoyam rhizosphere with biocontrol activity. J. Appl. Microbiol. 103, 1007–
709
1020 (2007).
710
26.
Olorunleke, F. E. et al. Coregulation of the cyclic lipopeptides orfamide and sessilin in the
711
biocontrol strain Pseudomonas sp. CMR12a. Microbiologyopen 6, 1–12 (2017).
712
27.
Olorunleke, F. E., Hua, G. K. H., Kieu, N. P., Ma, Z. & Höfte, M. Interplay between
713
orfamides, sessilins and phenazines in the control of Rhizoctonia diseases by Pseudomonas
714
sp. CMR12a. Environ. Microbiol. Rep. 7, 774–781 (2015).
715
28.
Geudens, N. & Martins, J. C. Cyclic lipodepsipeptides from Pseudomonas spp. - Biological
716
Swiss-Army knives. Front. Microbiol. 9, 1–18 (2018).
717
29.
Omoboye, O. O. et al. Pseudomonas cyclic lipopeptides suppress the rice blast fungus
718
Magnaporthe oryzae by induced resistance and direct antagonism. Front. Plant Sci. 10, 901
719
(2019).
720
30.
Mansfield, J. et al. Top 10 plant pathogenic bacteria in molecular plant pathology. Mol. Plant
721
Pathol. 13, 614–629 (2012).
722
31.
Molinatto, G. et al. Complete genome sequence of Bacillus amyloliquefaciens subsp.
723
plantarum S499, a rhizobacterium that triggers plant defences and inhibits fungal
724
31
phytopathogens. J. Biotechnol. 238, 56–59 (2016).
725
32.
Fan, B. et al. Bacillus velezensis FZB42 in 2018: The gram-positive model strain for plant
726
growth promotion and biocontrol. Front. Microbiol. 9, 3389 (2018).
727
33.
Pandin, C. et al. Complete genome sequence of Bacillus velezensis QST713: A biocontrol
728
agent that protects Agaricus bisporus crops against the green mould disease. J. Biotechnol.
729
278, 10–19 (2018).
730
34.
Scholz, R. et al. Amylocyclicin, a novel circular bacteriocin produced by Bacillus
731
amyloliquefaciens FZB42. J. Bacteriol. 196, 1842–1852 (2014).
732
35.
Youard, Z. A., Mislin, G. L. A., Majcherczyk, P. A., Schalk, I. J. & Reimmann, C.
733
Pseudomonas fluorescens CHA0 produces enantio-pyochelin, the optical antipode of the
734
Pseudomonas aeruginosa siderophore pyochelin. J. Biol. Chem. 282, 35546–35553 (2007).
735
36.
Grandchamp, G. M. et al. Pirated siderophores promote sporulation in Bacillus subtilis. Appl.
736
Environ. Microbiol. 83, 1–17 (2017).
737
37.
Miethke, M. et al. Ferri-bacillibactin uptake and hydrolysis in Bacillus subtilis. Mol. Microbiol.
738
61, 1413–1427 (2006).
739
38.
Adler, C. et al. Catecholate siderophores protect bacteria from pyochelin toxicity. PLoS One
740
7, (2012).
741
39.
Trottmann, F., Franke, J., Ishida, K., García-Altares, M. & Hertweck, C. A pair of bacterial
742
siderophores releases and traps an intercellular signal molecule: An unusual case of natural
743
nitrone bioconjugation. Angew. Chemie 131, 206–210 (2019).
744
40.
Hartney, S. L. et al. Ferric-pyoverdine recognition by Fpv outer membrane proteins of
745
Pseudomonas protegens Pf-5. J. Bacteriol. 195, 765–776 (2013).
746
41.
Rokni-Zadeh, H. et al. Genetic and functional characterization of cyclic lipopeptide white-line-
747
inducing principle (WLIP) production by rice rhizosphere isolate Pseudomonas putida
748
RW10S2. Appl. Environ. Microbiol. 78, 4826–4834 (2012).
749
32
42.
van Gestel, J., Vlamakis, H. & Kolter, R. From cell differentiation to cell collectives: Bacillus
750
subtilis uses division of labor to migrate. PLoS Biol. 13, 1–29 (2015).
751
43.
Munsch, P. & Alatossava, T. The white-line-in-agar test is not specific for the two cultivated
752
mushroom associated pseudomonads, Pseudomonas tolaasii and Pseudomonas ‘reactans’.
753
Microbiol. Res. 157, 7–11 (2002).
754
44.
Salari, F. et al. Draft Genome Sequence of Pseudomonas aeruginosa strain LMG 1272, an
755
atypical white line reaction producer. Microbiol. Resour. Announc. 9, (2020).
756
45.
Xu, Z. et al. Antibiotic bacillomycin D affects iron acquisition and biofilm formation in Bacillus
757
velezensis through a Btr-mediated FeuABC-dependent pathway. Cell Rep. 29, 1192-1202.e5
758
(2019).
759
46.
Powers, M. J., Sanabria-Valentín, E., Bowers, A. A. & Shank, E. A. Inhibition of cell
760
differentiation in Bacillus subtilis by Pseudomonas protegens. J. Bacteriol. 197, 2129–2138
761
(2015).
762
47.
Molina-Santiago, C. et al. The extracellular matrix protects Bacillus subtilis colonies from
763
Pseudomonas invasion and modulates plant co-colonization. Nat. Commun. 10, (2019).
764
48.
Liu, Y., Kyle, S. & Straight, P. D. Antibiotic stimulation of a Bacillus subtilis migratory
765
response. mSphere 3, 1–13 (2018).
766
49.
Rosenberg, G. et al. Not so simple, not so subtle: The interspecies competition between
767
Bacillus simplex and Bacillus subtilis and its impact on the evolution of biofilms. npj Biofilms
768
Microbiomes 2, 15027 (2016).
769
50.
Straight, P. D., Willey, J. M. & Kolter, R. Interactions between Streptomyces coelicolor and
770
Bacillus subtilis: Role of surfactants in raising aerial structures. J. Bacteriol. 188, 4918–4925
771
(2006).
772
51.
Lee, N. et al. Iron competition triggers antibiotic biosynthesis in Streptomyces coelicolor
773
during coculture with Myxococcus xanthus. ISME J. 14, 1111–1124 (2020)
774
33
52.
Kramer, J., Özkaya, Ö. & Kümmerli, R. Bacterial siderophores in community and host
775
interactions. Nat. Rev. Microbiol. 18, 152–163 (2020).
776
53.
Niehus, R., Picot, A., Oliveira, N. M., Mitri, S. & Foster, K. R. The evolution of siderophore
777
production as a competitive trait. Evolution (N. Y). 71, 1443–1455 (2017).
778
54.
Hibbing, M. E., Fuqua, C., Parsek, M. R. & Peterson, S. B. Bacterial competition: Surviving
779
and thriving in the microbial jungle. Nature Reviews Microbiology 8 15–25 (2010).
780
55.
Gu, S. et al. Competition for iron drives phytopathogen control by natural rhizosphere
781
microbiomes. Nat. Microbiol. 5, 1002–1010 (2020).
782
56.
Youard, Z. A., Wenner, N. & Reimmann, C. Iron acquisition with the natural siderophore
783
enantiomers pyochelin and enantio-pyochelin in Pseudomonas species. BioMetals 24, 513–
784
522 (2011).
785
57.
Cornelis, P. Iron uptake and metabolism in pseudomonads. Appl. Microbiol. Biotechnol. 86,
786
1637–1645 (2010).
787
58.
Ronnebaum, T. A. & Lamb, A. L. Nonribosomal peptides for iron acquisition: pyochelin
788
biosynthesis as a case study. Current Opinion in Structural Biology 53, 1–11 (2018).
789
59.
Esmaeel, Q. et al. Burkholderia genome mining for nonribosomal peptide synthetases
790
reveals
a
great
potential
for
novel
siderophores
and
lipopeptides
synthesis.
791
Microbiologyopen 5, 512–526 (2016).
792
60.
Seipke, R. F. et al. The plant pathogen Streptomyces scabies 87-22 has a functional
793
pyochelin biosynthetic pathway that is regulated by TetR- and AfsR-family proteins.
794
Microbiology 157, 2681–2693 (2011).
795
61.
Inahashi, Y. et al. Watasemycin biosynthesis in Streptomyces venezuelae: thiazoline C-
796
methylation by a type B radical-SAM methylase homologue. Chem. Sci. 8, 2823–2831
797
(2017).
798
62.
Jarmer, H., Berka, R., Knudsen, S. & Saxild, H. H. Transcriptome analysis documents
799
34
induced competence of Bacillus subtilis during nitrogen limiting conditions. FEMS Microbiol.
800
Lett. 206, 197–200 (2002).
801
63.
Martínez-García, E. & de Lorenzo, V. Engineering multiple genomic deletions in Gram-
802
negative bacteria: analysis of the multi-resistant antibiotic profile of Pseudomonas putida
803
KT2440. Environ. Microbiol. 13, 2702–2716 (2011).
804
64.
Vacheron, J. et al. T6SS contributes to gut microbiome invasion and killing of an herbivorous
805
pest insect by plant-beneficial Pseudomonas protegens. ISME J. 13, 1318–1329 (2019).
806
65.
Livak, K. J. & Schmittgen, T. D. Analysis of relative gene expression data using real-time
807
quantitative PCR and the 2-ΔΔCT method. Methods 25, 402–408 (2001).
808
66.
Pluskal, T., Castillo, S., Villar-Briones, A. & Orešič, M. MZmine 2: Modular framework for
809
processing, visualizing, and analyzing mass spectrometry-based molecular profile data.
810
BMC Bioinformatics 11, (2010).
811
67.
Hoegy, F., Mislin, G. L. A. & Schalk, I. J. Pyoverdine and pyochelin measurements. Methods
812
Mol. Biol. 1149, 293–301 (2014).
813
68.
Kune, C. et al. Rapid visualization of chemically related compounds using kendrick mass
814
defect as a filter in mass spectrometry imaging. Anal. Chem. 91, 13112–13118 (2019).
815
69.
R Core Team (2020). R: A language and environment for statistical computing. R Foundation
816
for Statistical Computing, Vienna, Austria. https://www.r-project.org/ (2020).
817
818
Acknowledgments
819
We gratefully acknowledge Sébastien Rigali, Alexandre Jousset and Loïc Ongena for critically
820
reading the manuscript. We thank C. Keel for the kind gift of strains and J. Vacheron for the very
821
helpful indications on Pseudomonas mutagenesis. This work was supported by the EU Interreg V
822
France-Wallonie-Vlaanderen portfolio SmartBiocontrol (Bioprotect and Bioscreen projects, avec le
823
35
soutien du Fonds européen de développement régional - Met steun van het Europees Fonds voor
824
Regionale Ontwikkeling), by the European Union Horizon 2020 research and innovation program
825
under grant agreement No. 731077 and by the EOS project ID 30650620 from the FWO/F.R.S.-
826
FNRS. The MALDI FT-ICR SolariX XR was funded by FEDER BIOMED HUB Technology Support
827
(number 2.2.1/996). AA is recipient of a F.R.I.A. fellowship (Formation à la Recherche dans
828
l’Industrie et l’Agriculture) and MO is senior research associate at the F.R.S.-F.N.R.S.
829
830
Author contributions
831
SA, TM, AR and AA performed most of the co-culture and in planta experiments. SA and TM
832
performed most of molecular biology experiments with help of GH and SS for mutant generation
833
and of SS for transcriptomics. TM and GH did genome mining. AR and AA were involved in all
834
aspects of metabolomics using UPLC-MS. Data analysis was done by SA, TM, AA and AR. AM, AA
835
and EDP performed the MALDI FT-ICR experiments and analyzed the data. MH and RDM provided
836
Pseudomonas strains/mutants and also supported the study by providing intellectual input. SA, TM
837
and MO mainly wrote the manuscript. All of the authors commented on the manuscript and
838
contributed to the final form. MO supervised the study.
839
840
Competing interests
841
The authors declare no competing interests.
842
843
2 H -b a e
b a c illa e n e
o x y d iffic id in
d iffic id in
s u rfa c tin s
b a c illib a c tin
0
5
1 0
1 5
4 0
6 0
F o ld c h a n g e in B S M p ro d u c tio n
****
****
****
****
****
****
G A 1 w t
Ds fp
Ds rfa A
Ditu A
Dfe n A
Dm ln A
Dd fn A -Db a e J
Dd fn A
Dd fn M
Db a e J
Db a e S
0
2 0
4 0
6 0
8 0
1 0 0
G ro w th in h ib itio n (% )
X a n th o m o n a s
a
a
a
a
a b
b
c
d
e
e
e
0
5
1 0
1 5
2 0
2 5
0 .0
0 .5
1 .0
1 .5
C e ll d e n sity (O D 6 0 0)
C u ltu re tim e (h o u rs)
0
5
1 0
1 5
2 0
2 5
0 .0
0 .5
1 .0
1 .5
C u ltu re tim e (h o u rs)
C e ll d e n sity (O D 6 0 0)
u n su p p le m e n te d
G A 1
G A 1 + C F S
C T R L
C F S
0
2
4
6
8
1 0
R elative a c n A e x p re ss io n
**
G A 1 w t
Ds fp
Ds fp -Da c n A
Da c n A
0
2 0
4 0
6 0
8 0
1 0 0
G ro w th in h ib itio n (% )
C la v ib a c te r
a
b
c
c
G A 1
G A 1
G A 1 + C F S
G A 1 + C F S
X a n th o m o n a s
C la v ib a c te r
surfactins
iturins
fengycins
macrolactins
difficidin
oxydifficidin
bacillaene
dihydrobacillaene
amylocyclicin
a
b
c
d
e
I
III
II
lV
Figure 1
C M R 1 2 a
Ds e s A
Do fa B C
Ds e s A -p h z
Do fa B C -p h z
Dp v d I
Dp c h A
Dp v d I-p c h A
0
2
4
6
8
1 0
1 2
F o ld c h a n g e in 2 H -b a e p ro d u c tio n
a
a
a
a
a
a b
c
c d
F o ld c h a n g e in 2 H -b a e p ro d u c tio n
C F S
C F S + F e
P V D
E -P C H
D IP 2 5 µ M
D IP 5 0 µ M
D IP 2 0 0 µ M
0
1
2
3
4
5
a
a
b
b
b
b
b
0
2
4
6
F o ld c h a n g e in 2 H -b a e p ro d u c tio n
0 .1 7 5
0 .3 5
0 .7
0 .7
1 .4
2 .8
P V D (mM)
E -P C H (mM)
c
c
c
b
a
a
2 H -b a e
0
2
4
6
8
F o ld c h a n g e in b a c illib a c tin p ro d u c tio n
0 .1 7 5
0 .3 5
0 .7
0 .7
1 .4
2 .8
P V D (mM)
E -P C H (mM)
a
b
c
c
c
b c
B a c illib a c tin
0
2
4
6
8
1 0
1 2
0 .0
0 .2
0 .4
0 .6
0 .8
1 .0
In c u b a tio n tim e (h )
G ro w th O D 6 0 0
G A 1 W T
0
2
4
6
8
1 0
1 2
0 .0
0 .2
0 .4
0 .6
0 .8
1 .0
In c u b a tio n tim e (h )
G ro w th O D 6 0 0
B a cillib a ctin m u ta n t
+ p vd
+ p c h
C o n tro l
sessilin
orfamide
phenazine
PVD
E-PCH
a
b
c
d
Figure 2
CMR12a
GA1
CMR12a
GA1
1 cm
GA1
GA1 ΔsrfaA
CMR12a
CMR12a
1 cm
m/z=1044.6
1 cm
100%
0%
a
b
c
Figure 3
e
d
GA1
CMR12a
ΔofaBC-phz
ΔofaBC
ΔsesA
ΔsesA-ofaBC-phz
GA1
GA1
GA1
GA1
e
d
ΔofaBC-phz
ΔsesA
ΔsesA-ofaBC-phz
GA1
GA1
C o n tro l
C M R 1 2 a
Ds e s A
C o n tro l
C M R 1 2 a
C M R 1 2 a + s u rfa c tin
Ds e s A
0 .0
0 .2
0 .4
0 .6
0 .8
1 .0
1 .2
B a c illu s g ro w th (O D 6 0 0)
a
a
a
a
c
a b
b
G A 1
Ds rfa A
B a c illu s g ro w th (O D 6 0 0)
C o n tro l
C F S
Do fa B C -p h z
Ds e s A
Ds e s A -o fa B C
Ds e s A -p h z
Ds e s A -o fa B C -p h z
0 .0
0 .2
0 .4
0 .6
0 .8
1 .0
a
b
b
b
b
c
d
GA1
CMR12a
CH36
GA1
ΔsesA
ΔsrfaA
ΔsrfaA
CH36
ΔtolA
ΔsrfaA
ΔsrfaA
CMR12a
GA1
ΔtolA
ΔsesA
GA1
BW11M1
RW10S2
SS101
WCU - 64
GA1
GA1
GA1
GA1
1cm
CMR12a
WCU-64
BW11M1
SS101
RW10S2
tolaasins
orfamides
massetolides
xantholysins
putisolvins
CH36
ΔsesA
ΔtolA
Presence
Absence
sessilins
WLIP
a
b
c
d
II
I
I
II
III
Figure 4
Figure 5
G A 1
C M R 1 2 a
G A 1 (c o -in o c u la tio n )
C M R 1 2 a (c o -in o c u la tio n )
1 0 0
1 0 5
1 0 1 0
B a c te ria l p o p u la tio n (C F U /g o f ro o ts )
***
****
****
d fn A
b a e J
a c n A
0
2
4
6
8
1 0
R elative g e n e e xp re ss io n
*
***
****
G A 1 /C M R 1 2 A
G A 1 /Ds e s A
Ds rfa A /C M R 1 2 a
Ds rfa A /Ds e s A
1 0 0
1 0 5
1 0 1 0
B a c illu s p o p u la tio n (C F U /g o f ro o ts )
**
**
a
b
c
d
Intensity
1.50e4
1.00e4
0.50e4
1.50e4
6.00e2
4.00e2
Intensity
1.50e4
6.00e2
4.00e2
Intensity
GA1
GA1 + CMR12a
GA1 + ΔsesA
C15 Surfactin
C15 Surfactin
C15 Surfactin
Sessilin A
9
9.5
10
16
16.5
17
Time (min)
9
9.5
10
16
16.5
17
Time (min)
9
9.5
10
16
16.5
17
Time (min)
Intensity
9
9.5
10
16
16.5
17
Time (min)
9
9.5
10
16
16.5
17
Time (min)
| 2021 | Chelator sensing and lipopeptide interplay mediates molecular interspecies interactions between soil bacilli and pseudomonads | 10.1101/2021.02.22.432387 | [
"Andric Sofija",
"Meyer Thibault",
"Rigolet Augustin",
"Arias Anthony Argüelles",
"Steels Sébastien",
"Hoff Grégory",
"Höfte Monica",
"De Mot René",
"McCann Andrea",
"De Pauw Edwin",
"Ongena Marc"
] | creative-commons |
1
Trait-similarity and trait-hierarchy jointly determine co-occurrences of
1
resident and invasive ant species
2
3
Mark K. L. Wong*†1, Toby P. N. Tsang*2, Owen T. Lewis1 and Benoit Guénard2
4
5
*Mark K. L. Wong and Toby P. N. Tsang are joint first authors.
6
†Correspondence e-mail: mark.wong@zoo.ox.ac.uk ; Tel: +44 (0) 1865 271234
7
1Department of Zoology, University of Oxford, South Parks Road, Oxford OX1 3PS, United
8
Kingdom
9
2School of Biological Sciences, The University of Hong Kong, Pok Fu Lam Road, Hong Kong
10
SAR, China
11
Other authors’ e-mail addresses, in order presented above: paknok@connect.hku.hk,
12
owen.lewis@zoo.ox.ac.uk, bguenard@hku.hk
13
14
Running title: Trait differences and species co-occurrence
15
16
Keywords: coexistence, community assembly, competition, competitive exclusion, functional
17
trait, invasion, limiting similarity, niche
18
19
Authorship
20
MKLW and TPNT conceived the study with inputs from BG and OTL. MKLW conducted
21
fieldwork, collected trait data and built probability density functions. TPNT performed the
22
statistical analyses. MKLW and TPNT wrote the first draft of the manuscript. All authors
23
contributed significantly to revisions.
24
25
Data accessibility
26
The authors confirm that, should the manuscript be accepted, the data supporting the results
27
will be archived in the Dryad Digital Repository, and the data DOI will be included at the end
28
of the article.
29
2
Abstract
30
Interspecific competition, a dominant process structuring ecological communities, acts on
31
species’ phenotypic differences. Species with similar traits should compete intensely (trait-
32
similarity), while those with traits that confer competitive ability should outcompete others
33
(trait-hierarchy). Either or both of these mechanisms may drive competitive exclusion within
34
a community, but their relative importance and interacting effects are rarely studied. We show
35
empirically that spatial associations (pairwise co-occurrences) between an invasive ant
36
Solenopsis invicta and 28 other ant species across a relatively homogenous landscape are
37
explained largely by an interaction of trait-similarity and trait-hierarchy in one morphological
38
trait. We find that increasing trait-hierarchy leads to more negative associations; however these
39
effects are counteracted when species are sufficiently dissimilar (by 37-95%) in their trait
40
ranges. We also show that a model of species co-occurrences integrating trait-similarity and
41
trait-hierarchy consolidates predictions of different theoretical assembly rules. This highlights
42
the explanatory potential of the trait-based co-occurrence approach.
43
3
INTRODUCTION
44
There is perhaps no ecological process that is at once as familiar and as enigmatic as
45
interspecific competition, which can strongly mould the structure of ecological communities
46
(Hutchinson, 1959). While patterns in biodiversity consistent with competitive interactions
47
have been widely documented (Schoener, 1974; Calatayud et al., 2020), precisely how
48
phenotypic differences between species determine the nature of competitive exclusion has
49
remained highly contested.
50
51
Under classical niche theory (MacArthur & Levins, 1967), competitive exclusion leads to co-
52
occurring species having dissimilar niches because species with similar niches compete more
53
intensely. One proxy for the niche dissimilarity between two species is a non-directional or
54
‘absolute’ measure of their dissimilarity in trait space (Fig. 1A) (Carmona et al., 2019a).
55
Accordingly, the competition trait-similarity hypothesis predicts that the likelihood of co-
56
occurrence will always decrease with increasing overlap in trait space, such that co-occurring
57
species display ‘overdispersion’: high absolute dissimilarity in trait space (Fig. 1A).
58
59
In contrast, modern coexistence theory emphasizes that species’ niche dissimilarities are not
60
the only factors determining competitive outcomes (Chesson, 2000). For instance, species can
61
be organized along a competitive hierarchy where differences in competitive ability drive the
62
exclusion of weaker competitors (Kunstler et al., 2012). Directional measures of trait
63
differences, such as the ‘hierarchical difference’ in species’ mean trait values, provide a proxy
64
for differences in competitive ability (Fig. 1B) (Kunstler et al., 2012). Contrary to the
65
competition trait-similarity hypothesis, the competition trait-hierarchy hypothesis predicts that
66
the likelihood of co-occurrence will decrease with increasing hierarchical difference (and
67
dissimilarity), while decreasing hierarchical difference promotes ‘clustering’: the co-
68
occurrence of similar species (Fig. 1B).
69
70
Despite a lasting focus on classical niche theory, empirical support for the competition trait-
71
similarity hypothesis has been mixed (Mayfield & Levine, 2010). Some communities
72
structured by competition show trait clustering consistent with the competition trait-hierarchy
73
hypothesis (Herben & Goldberg, 2014). However, recent studies show that the outcomes of
74
competition between plant species can be predicted by hierarchical differences in traits
75
governing resource acquisition (e.g., leaf area for light interception, Kraft, Godoy & Levine,
76
2015; Kunstler et al., 2016; Perez-Ramos et al., 2019). The majority of trait-based studies
77
applying the framework of modern coexistence theory have focused on plants and microbes,
78
4
and have used experimentally-assembled communities (Grainger et al., 2019), which may not
79
represent adequately the dynamics of natural communities (Carpenter, 1996). Most
80
observational studies investigating the role of competition in structuring communities,
81
however, measure only trait dissimilarities and test for overdispersion (Mittelbach & McGill,
82
2019). In this regard, the potential for species’ trait differences to reflect competitive ability
83
differences may be underestimated.
84
85
Inferences of assembly processes from patterns in community structure are ubiquitous in the
86
literature (Mittelbach & McGill, 2019). However, an inherent and questionable assumption of
87
this approach is that all species within a community are subject to the same ‘dominant’
88
assembly process (Siepielski & McPeek, 2010). Rather than assuming that competition acts
89
uniformly across all species at the community level, it can be informative to investigate
90
whether and how competitive exclusion occurs for individual pairs of species. At this finer
91
scale, competitive outcomes should be driven by an interaction between trait-similarity and
92
trait-hierarchy (Chesson, 2000). That is, competitive exclusion will only occur for pairs of
93
species which are insufficiently dissimilar in niches relative to their differences in competitive
94
abilities (Fig. 1C; Mayfield & Levine, 2010). This interplay of trait-similarity and trait-
95
hierarchy in determining competitive outcomes between species pairs is relatively unexplored.
96
Nonetheless, it was anticipated by Abrams (1983): “What is needed instead is a broader
97
definition of limiting similarity. The concept should be represented as a relationship between
98
the difference in competitive ability and the maximum similarity that will permit coexistence.
99
Such a relationship has the potential to be different for every different pair of species.”
100
101
Biological invasions, which often lead to intense competitive interactions, are choice settings
102
for investigating competition (Shea & Chesson, 2002). For instance, many classical invasion
103
hypotheses (empty niche, enemy escape, novel weapons etc.) essentially attribute invasion
104
outcomes to niche dissimilarities and competitive ability differences between invader and
105
native species (MacDougall et al., 2009). This framework of modern coexistence theory has
106
been used to identify the trait values of exotic plant species which confer competitive
107
advantages and facilitate invasion success (Gross et al., 2015) – but its potential to explain
108
invasions in other taxa is untapped.
109
110
Ecological literature on the ants (Hymenoptera: Formicidae) is replete with studies identifying
111
competition as a strong driver of community structure (Cerda, Arnan, & Retana, 2013) as well
112
as reports of exotic species competitively excluding native ones (Holway, 1999). Many ant
113
5
communities also show patterns of phylogenetic clustering in the presence of invasive ant
114
species (Lessard et al., 2009), and theory suggests that such patterns may emerge if community
115
assembly is driven by environmental filtering, or alternatively by competitive hierarchies.
116
However, it is difficult to distinguish these two processes solely on the basis of phylogenetic
117
relationships (Cadotte & Tucker, 2017). In most cases it is also hard to identify the species
118
which compete most with invasive species, or which are most susceptible to displacement,
119
especially when phylogenetic associations between invaders and resident species are
120
ambiguous (Lessard et al., 2009). Such limitation in inferring contemporary ecological
121
mechanisms from phylogenetic patterns of evolutionary history can be addressed with a focus
122
on species’ traits, which govern their abiotic and biotic interactions in real time (Wong,
123
Guénard & Lewis, 2019).
124
125
Here, we test trait-based hypotheses from classical niche theory and modern coexistence theory
126
empirically. We focus on the invasion of the non-native Red Imported Fire Ant (Solenopsis
127
invicta) in ant communities of wetland habitats in Hong Kong (reported in Wong, Guénard &
128
Lewis, 2020). In these relatively homogenous landscapes, communities are more likely to be
129
structured by competition as opposed to other mechanisms such as environmental filtering
130
(Keddy, 1992). There is some disagreement as to whether S. invicta competes strongly with
131
resident ant species during invasion. While some studies report competitive exclusion by S.
132
invicta (Porter & Savignano, 1990; Gotelli & Arnett, 2000), others contend that altered abiotic
133
conditions under anthropogenic disturbances – which happen to favour S. invicta – are directly
134
responsible for the decline of resident species (King & Tschinkel, 2008). To this end, trait-
135
based tests for theoretical mechanisms of competition in a system with low levels of
136
environmental variation may clarify the interactions between S. invicta and other species.
137
138
We integrate trait-based and co-occurrence analyses to investigate whether trait-similarity
139
and/or trait-hierarchy determine how S. invicta affect other ant species. There are two
140
advantages to this approach. First, it allows for detecting potentially varying relationships at
141
the fine ecological scales (species pairs) where competition unfolds (Abrams, 1983). Second,
142
it allows for developing and testing more specific predictions about assembly processes than
143
would be possible with standalone co-occurrence analyses (Veech, 2014). We first use a
144
network of species’ spatial associations (co-occurrences) to quantify negative associations
145
between S. invicta and other ants across multiple plots. Next, for distinct morphological traits
146
that regulate ant physiology and behaviour, we use non-directional and directional measures of
147
species’ trait differences as proxies for species’ niche dissimilarities (absolute dissimilarity)
148
6
and competitive ability differences (hierarchical difference) respectively (after Kunstler et al.,
149
2012; Carmona et al., 2019a). Integrating species’ trait differences and co-occurrences then
150
allows us to test three hypotheses on the likelihood and nature of pairwise competitive
151
exclusion between S. invicta and all resident species (Fig. 1).
152
153
If competitive exclusion is always driven by trait-similarity, absolute dissimilarity alone will
154
determine co-occurrence relationships, with decreasing absolute dissimilarity leading to more
155
negative co-occurrence (Fig. 1A). Alternatively, if competitive exclusion is always driven by
156
trait-hierarchy, hierarchical difference alone will determine co-occurrence relationships, with
157
larger hierarchical difference leading to more negative co-occurrence (Fig. 1B). Finally, if both
158
mechanisms operate, we expect an interaction of absolute dissimilarity and hierarchical
159
difference to determine co-occurrence relationships. Specifically, we expect absolute
160
dissimilarity to modulate the effect of hierarchical difference, such that hierarchical difference
161
determines co-occurrence relationships only if absolute dissimilarity is sufficiently low (Fig.
162
1C).
163
7
164
Figure 1. The trait-similarity and trait-hierarchy hypotheses of competition predict different outcomes for species
165
co-occurrences separately and in combination. Panels show hypothetical relationships between three ant species
166
and the invader S. invicta for one trait (left) and the corresponding pairwise co-occurrence relationships (right) as
167
predicted under specific hypotheses. In each panel, species in red experience competitive exclusion and negative
168
co-occurrence with S. invicta (i.e., they are not found in the same plots), with thicker lines indicating stronger
169
relationships; species in black can co-occur with S. invicta in the same plots. A: If competitive exclusion is driven
170
entirely by trait-similarity for all pairs of species (MacArthur & Levins, 1967), decreasing absolute dissimilarity
171
(i.e. increasing overlap) between a species’ range of trait values and that of S. invicta increases the strength of the
172
negative co-occurrence relationship, while increasing absolute dissimilarity (decreasing overlap) promotes co-
173
occurrence. B: If competitive exclusion is driven only by trait-hierarchy (e.g., Kunstler et al., 2012) and species’
174
mean trait values (T) correspond to their competitive abilities along a directional axis, then a larger hierarchical
175
difference (T1-T2) between a species and S. invicta increases the strength of the negative co-occurrence
176
relationship, while a smaller hierarchical difference promotes co-occurrence. C: Trait-similarity and trait-
177
hierarchy may jointly determine species co-occurrences because niche dissimilarities and competitive hierarchies
178
interact to determine competitive outcomes across different species pairs (Abrams, 1983; Chesson, 2000). The
179
likelihood of competitive exclusion (and strength of the negative co-occurrence relationship) between a species
180
and S. invicta increases with increasing hierarchical difference in competitive ability; however, this competitive
181
effect can also be counteracted and overcome by a large absolute dissimilarity in trait space, promoting co-
182
occurrence.
183
Sp.1
Sp.2
Sp.3
S. invicta
Sp.1
Sp.2
Sp.3
S. invicta
Sp.1
Sp.2
Sp.3
S. invicta
Trait 1
TSp.1 TSp.2 TSp.3 TS. invicta
Trait 1
TSp.1 TSp.2 TSp.3 TS. invicta
Trait 1
Sp.1 Sp.2 Sp.3 S. invicta
A. Trait-Similarity (MacArthur & Levins, 1967)
B. Trait-Hierarchy (e.g. Kunstler et al., 2012)
C. Trait-Similarity & Trait-Hierarchy (Abrams, 1983; Chesson, 2000)
Low
comp.
ability
High
comp.
ability
Low
comp.
ability
High
comp.
ability
8
MATERIAL AND METHODS
184
185
Ant sampling and environmental variables
186
To maximise the likelihood of detecting community patterns reflecting biotic assembly
187
processes such as interspecific competition (de Bello et al., 2012), we characterized ant
188
communities at fine spatial scales in a relatively homogenous landscape (Wong et al., 2020).
189
We selected two wetland reserves in Hong Kong – Lok Ma Chau (22.512°N, 114.063°E) and
190
Mai Po (22.485°N, 114.036°E) – which have been conserved for >35 years, and which contain
191
networks of exposed grass bunds (width ≤5 m) separating individual ponds. Most ant species
192
present were native, but high densities of S. invicta were also recorded at multiple locations in
193
pilot surveys conducted from 2015 to 2017. In 2018 we selected 61 plots from these locations,
194
including 24 plots where S. invicta were present.
195
196
From April to September 2018, we sampled the local ant community at each plot in a 4 ´ 4 m
197
quadrat, using six pitfall traps which were exposed for 48 hours (Wong et al., 2020). The
198
maximum distance between any two traps in each plot was 5.65 m, a higher sampling density
199
(i.e., traps / m2) than in previous studies characterising ant communities (Parr, 2008). We
200
sampled at such fine spatial scales to enhance the detection of species’ occurrence patterns
201
driven by biotic interactions, as most ant species in the region forage within 5 m of their nests
202
(Eguchi, Bui & Yamane, 2004) and S. invicta forage within 4 m of their nests (Weeks, Wilson
203
& Vinson, 2004). For the same reasons, a minimum distance of at least 20 m between individual
204
plots facilitated independent observations.
205
206
All specimens were sorted into morphospecies and subsequently identified to species (Wong
207
et al., 2020). We compiled a matrix of ant species’ occurrences (i.e., presence/absence data)
208
across all 61 plots, and classified plots as either ‘S. invicta-present’ or ‘S. invicta-absent’ based
209
on the presence of S. invicta workers in traps at each plot.
210
211
In addition to characterizing the ant community at each plot, we estimated the percentage of
212
ground cover, and obtained data on the NDVI and mean annual temperature from local climate
213
models at 30 m resolution (see Morgan & Guénard, 2019). We later used these data to check
214
whether species’ preferences for particular physical properties influenced their co-occurrences
215
(further below).
216
217
Building co-occurrence networks
218
9
Co-occurrence networks document all pairwise co-occurrence relationships (i.e., network
219
links) between species (i.e., network nodes) within a species pool. We used odds ratios to build
220
the network (after Lane et al., 2014); this approach can incorporate signals of asymmetry in
221
co-occurrence relationships (Araújo et al., 2011). We summarized the presence and absence of
222
species pairs in 2*2 contingency tables and calculated the strength of co-occurrence
223
relationships as their asymmetrical odds ratios (Lane et al., 2014). For example, given a species
224
pair A & B, the odds ratio for indication of B by A (ORAB) measures how the probability of
225
B’s presence at a plot changes under the presence of A in the same plot, and vice versa for
226
ORBA:
227
𝑂𝑅!" =
𝑁(𝐵 = 1 𝑎𝑛𝑑 𝐴 = 1) + 0.5
𝑁(𝐵 = 0 𝑎𝑛𝑑 𝐴 = 1) + 0.5
𝑁(𝐵 = 1) + 0.5
𝑁 (𝐵 = 0) + 0.5
𝑂𝑅"! =
𝑁(𝐴 = 1 𝑎𝑛𝑑 𝐵 = 1) + 0.5
𝑁(𝐴 = 0 𝑎𝑛𝑑 𝐵 = 1) + 0.5
𝑁(𝐴 = 1) + 0.5
𝑁 (𝐴 = 0) + 0.5
228
where N represents the number of plots. We applied Haldane’s correction and added 0.5 to all
229
components to avoid odds ratios becoming infinity or undefined (Agresti, 2018). We further
230
log-transformed the odds ratios in subsequent analyses such that they could be compared
231
arithmetically (Agresti, 2018). All species were included in the analyses.
232
233
Null models to assess co-occurrence relationships
234
To examine whether species were primarily associated with negative co-occurrence
235
relationships within networks, we quantified their weighted degree – the sum of strengths (i.e.,
236
log-transformed odds ratios) of all co-occurrence relationships in the network. For each
237
species, we only considered co-occurrence relationships which indicated how that species
238
affected the odds ratios of other species being present in the same plots. For instance, the
239
weighted degree of species A considered 𝑂𝑅!" but not 𝑂𝑅"!.
240
241
Since any observed co-occurrence relationships could be driven by random associations
242
(Gotelli, 2000), we used null models to compare their observed weighted degree to random
243
expectation. Sampling plots were spatially distributed across two general localities – Lok Ma
244
Chau and Mai Po (Wong et al., 2020) – and randomly shuffling species occurrences across the
245
whole matrix could result in unrealistic null communities if the localities had different species
246
pools. Thus, we randomly generated compositional data for each of the two localities, and then
247
combined the two matrices to form one null matrix. To generate random matrices we used the
248
fixed-fixed algorithm (“quasiswap” in R-package vegan), which is robust to Type-I errors and
249
suitable for heterogenous compositional data (Gotelli, 2000).
250
10
251
We generated 1,000 null matrices, and calculated odds ratios to build null networks. We
252
calculated the Standardized Effect Size for weighted degree (SESWD) (Gurevitch et al., 1992),
253
defined as
254
𝑆𝐸𝑆#$ = 𝑂𝑏𝑠𝑒𝑟𝑣𝑒𝑑 𝑣𝑎𝑙𝑢𝑒 − 𝑀𝑒𝑎𝑛%&'' #$
𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝑑𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛𝑠%&'' #$
255
A species was primarily associated with negative co-occurrence relationships compared with
256
random expectation if SESWD was less than zero.
257
258
In order to test whether competition by trait-similarity and/or trait-hierarchy explained co-
259
occurrences between S. invicta and resident ant species, we obtained the SES for all pairwise
260
co-occurrence relationships involving S. invicta (SESsinv). Here we considered pairwise co-
261
occurrence relationships which indicated how the presence of other species within plots were
262
affected by the presence of S. invicta, but not vice versa. We obtained mean and standard
263
deviations of odds ratios of each considered relationship from 1,000 null networks to calculate
264
SESsinv for each resident species. A negative SESsinv value indicates that a species has a more
265
negative co-occurrence relationship with S. invicta than expected by chance.
266
267
Trait measurements and trait ranges of species
268
We assembled an individual-level trait dataset comprising data for seven morphological traits
269
(body size, and six size-corrected traits: head width, eye width, mandible length, scape length,
270
pronotum width, leg length). These traits regulate ant physiology and behaviour and are
271
hypothesized to impact performance and fitness (see Table 1 in Wong et al., 2020). The dataset
272
comprised trait measurements collected from at least 10 individual workers of every species
273
(N=319 individual ants), including different subcastes (minor and major workers) of
274
polymorphic species such as S. invicta (details in Wong et al., 2020). Body size was log-
275
transformed to reduce right skewness. For each trait, we built species-level probability density
276
functions (Carmona et al., 2019b) to calculate trait probability distributions (the curves in Fig.
277
1A). These distributions – or trait ranges – reflect the probabilities of observing different trait
278
values within individual species; they were subsequently used to quantify absolute
279
dissimilarities between species in trait (niche) space; see below.
280
281
Species’ trait differences, phylogenetic dissimilarity, and environmental preferences
282
For each of the seven traits, we quantified differences between S. invicta and each resident ant
283
species with a non-directional measure of niche dissimilarity (Absolute Dissimilarity, AD),
284
11
and a directional measure of competitive ability difference (Hierarchical Difference, HD). We
285
measured AD as the proportion of a resident species’ trait probability density function which
286
did not overlap with S. invicta’s trait probability density function (i.e., the proportion of trait
287
space exclusive to the resident species’ trait range) (Carmona et al., 2019b). AD values range
288
from 0 (when a resident species’ trait range is identical to that of S. invicta) to 1 (no overlap
289
with trait range of S. invicta; e.g. Sp. 1 in Fig. 1A). We measured HD as 𝑇)*+,-+. − 𝑇). -%0-,12,
290
where 𝑇 is the mean trait value for the given species (after Kunstler et al., 2012).
291
292
Additionally, to assess whether phylogenetic relationships or differing environmental
293
preferences led to non-independence in co-occurrence relationships between S. invicta and
294
each resident species, we quantified their phylogenetic dissimilarity (as pairwise distances
295
between species in phylogenetic trees) as well as their dissimilarities in environmental
296
preferences in terms of NDVI, ground cover and temperature (see Supporting Information).
297
298
Statistical analyses
299
To determine whether pairwise co-occurrences between S. invicta and resident species were
300
determined by trait-similarity, trait-hierarchy, or both mechanisms, we used multiple linear
301
regression to test whether the SESsinv for each species-pair was best predicted by AD, HD, or
302
an interaction of AD and HD. Our objective here was to use species’ trait differences to proxy
303
their niche and competitive ability differences, rather than to understand the effect of different
304
traits per se. Therefore, rather than using a full model, we built one model for each trait, with
305
AD, HD and a two-way interaction term (AD*HD) as predictors. For the trait Mandible Length,
306
AD and HD were highly correlated (Pearson’s r > 0.7), suggesting that their effects could not
307
be separated (Dormann, 2013); thus, we excluded this trait from subsequent analyses.
308
309
For any trait models that detected a significant effect from the interaction of AD and HD, we
310
used the Johnson-Neyman procedure (Johnson & Neyman, 1936) to calculate the ‘zone of
311
significance’, that is, the range of values of AD at which HD influenced SESsinv significantly
312
(or vice versa). We controlled for false discovery rates using the procedure described in Esarey
313
and Sumner (2017). We also checked whether the results of any models detecting significant
314
effects were invariable to the use of different density-thresholds to classify S. invicta-present
315
plots (see Supporting Information).
316
317
In addition to the individual trait models, we built separate models for phylogenetic
318
dissimilarity and dissimilarities in species’ environmental preferences to determine whether
319
12
these factors predicted co-occurrences between S. invicta and resident species. We built one
320
model with phylogenetic dissimilarity as the sole predictor, and three additional models – each
321
using environmental-preference dissimilarity in either NDVI, temperature or ground cover as
322
a sole predictor. Environmental variables were also added to the trait and phylogenetic
323
dissimilarity models as covariates if they were found to be significant.
324
325
Regression analyses were conducted in R (R Development Core Team, 2017). Before the
326
analyses, we standardized the variables such that their relative importance could be assessed
327
based on coefficient estimates (Schielzeth, 2010). We re-analysed our data with robust linear
328
regressions to test whether our results were driven by statistical outliers.
329
330
331
RESULTS
332
We recorded 29 ant species including S. invicta (Fig. 2), which occurred in 39% of the sampled
333
plots. Within the co-occurrence network, S. invicta was the species most strongly characterized
334
by negative co-occurrence relationships with other species (SESall=-3.62, Fig. 2A). Four other
335
species were characterized by significant negative (SESall<-1.96) co-occurrence relationships,
336
and two by significant positive (SESall>1.96) co-occurrence relationships (Fig. 2A). Of the 28
337
resident species, pairwise co-occurrences with S. invicta were positive (SESsinv>0) for nine
338
species and negative (SESsinv<0) for 19 species (Fig. 2B). Of these, one positive and seven
339
negative co-occurrence relationships were significant (Fig. 2B).
340
341
We found little evidence to suggest that either trait-similarity or trait-hierarchy solely
342
determined species co-occurrences. On their own, both AD and HD were poor predictors of
343
co-occurrence relationships between S. invicta and the 28 resident species (i.e., SESsinv) in
344
separate models for six traits (Tables S1 & S3).
345
346
An interaction between niche dissimilarities and competitive ability differences best predicted
347
co-occurrence relationships between S. invicta and the 28 resident species. Among different
348
models for the six traits (Table S1), the most parsimonious model was that for pronotum width
349
incorporating AD, HD and an interaction term (AD*HD), which explained 37% of the variation
350
in SESsinv (Table 1). Here, the interaction term (AD*HD) significantly explained co-occurrence
351
relationships between S. invicta and the resident species (Table 1); removing the interaction
352
term and only retaining the main effects of AD and HD significantly reduced model
353
performance, as indicated by a Chi-square test (ΔAICc = 8.04, ΔAdjusted-R2 = 30.32, p <
354
13
0.001). A significant interaction between AD and HD was also consistently observed in all
355
other models for pronotum width which used different density thresholds to classify S. invicta-
356
present plots (Table S3).
357
358
In the model (Table 1; Fig. 3), the positive – and marginally significant – effect of HD on
359
SESsinv indicated that resident species with relatively wider or narrower pronotums than S.
360
invicta tended to be positively or negatively associated with it, respectively. However, the
361
significant negative effect of the interaction between AD and HD meant that the positive effect
362
of HD on SESsinv was reinforced when AD was low, and counteracted when AD was high.
363
364
Based on the model, we further estimated the magnitudes of niche dissimilarities (AD) between
365
resident species and S. invicta at which competitive ability differences (HD) significantly
366
influenced their associations. Applying the Johnson-Neyman procedure revealed that co-
367
occurrence relationships between resident species and S. invicta were significantly affected by
368
HD when AD<0.37 or AD>0.95. There were 10 species for which AD<0.37 and three species
369
for which AD>0.95 in pronotum width with respect to S. invicta (Fig. 3).
370
371
In models based on other traits, main effects of AD and HD as well as their interacting effects
372
were either non-significant or not consistently significant across different density thresholds of
373
S. invicta presence (Tables S1 & S3). Phylogenetic and environmental-preference
374
dissimilarities were also not significant predictors in any models (Table S2).
375
14
376
Figure 2. Of the 29 ant species sampled across 61 plots, the invasive ant S. invicta is most negatively associated
377
with all other species. Plots show: (A) the degree to which each of the 29 species – including the invader S. invicta
378
(in bold) – is characterised by positive (blue) or negative (red) associations within a co-occurrence network
379
containing all species (SESall); (B) the degree to which each of the 28 resident species displays positive or negative
380
associations with the invader S. invicta (SESsinv). Dashed lines indicate critical values for statistical significance
381
of co-occurrence relationships (i.e., SES<-1.96 or >1.96). Ant species are grouped under four subfamilies:
382
Myrmicinae (Myr), Formicinae (For), Dolichoderinae (Dol) and Ponerinae (Pon).
383
384
385
386
Table 1. Multiple linear regression model for pronotum width. For this trait, a non-directional measure of niche
387
dissimilarity (Absolute Dissimilarity, AD), a directional measure of competitive ability difference (Hierarchical
388
Difference, HD), and their two-way interaction (AD*HD) determine pairwise co-occurrence relationships
389
between the invader S. invicta and 28 ant species (Fig. 2B: SESsinv). Bold value indicates statistical significance
390
(p<0.05). ‘JN intervals’ indicate the range of AD values at which the effects of HD are significant, as identified
391
from the Johnson-Neyman procedure.
392
Independent variable
β
P
JN intervals
AD
-0.47
0.21
<0.37; >0.95
HD
1.08
0.06
-
AD*HD
-1.47
0.002
-
R2=0.37
393
394
15
395
Figure 3. A response-surface showing how niche dissimilarity (Absolute Dissimilarity) modulates the effect of
396
competitive ability difference (Hierarchical Difference) in determining resident ant species’ co-occurrence
397
relationships with the invader S. invicta. The response-surface shows the predicted pairwise co-occurrence
398
relationship between a given ant species and S. invicta (SESsinv) for the trait pronotum width, based on the multiple
399
linear regression model in Table 1. Pairwise co-occurrence relationships (SESsinv) vary from negative (red) to
400
positive (blue), with SESsinv<-1.96 and SESsinv>1.96 indicating significant negative or positive relationships
401
respectively; contour lines illustrate how predicted SESsinv changes across the response-surface. Coloured points
402
on the response-surface show the observed SESsinv for individual resident ant species (N=28) (full names of
403
species shown in Fig. 2). On the x-axis, increasing values indicate decreasing overlap between a given species’
404
range of pronotum width values and that of S. invicta. On the y-axis, a positive or negative value indicates that a
405
given species has a relatively wider or narrower pronotum than S. invicta, respectively. The masked area in the
406
centre of the response-surface corresponds to the range of Absolute Dissimilarity (0.37-0.95) where the positive
407
effect of Hierarchical Difference on SESsinv is counteracted, as calculated from the Johnson-Neyman procedure.
408
409
410
411
412
413
16
DISCUSSION
414
We found that an interaction between trait-similarity and trait-hierarchy largely determined
415
spatial associations (co-occurrences) between the invasive species S. invicta and 28 other ant
416
species. These results suggest that trait-similarity and trait-hierarchy are interactive rather than
417
discrete mechanisms of competitive exclusion, as predicted from theory (Chesson, 2000). We
418
also found that a simple model of species co-occurrences, incorporating the interaction of trait-
419
similarity and trait-hierarchy, broadly consolidated predictions of different theoretical rules of
420
community assembly (discussed further below). Our study demonstrates that trait-based co-
421
occurrence analyses uncover unique evidence that can help explain the outcomes of community
422
assembly and biological invasions.
423
424
The overall pattern of pronounced negative co-occurrences between the abundant S. invicta
425
and many other species (Fig. 2) strongly identifies S. invicta as an influential component of the
426
network. Abundant species with many negative associations are often strong competitors
427
(Calatayud et al., 2020). Previous studies (Gotelli & Arnett, 2000; LeBrun, Plowes & Gilbert,
428
2012) considered S. invicta to competitively exclude other ant species on the basis of negative
429
co-occurrence patterns similar to those we observed. However, we appreciate that such patterns
430
may also be generated by other ecological processes (Brazeau & Schamp, 2019). Thus, in order
431
to strengthen inferences for particular assembly processes which could be at play, we explicitly
432
scrutinized species’ co-occurrence relationships in light of their ecological differences (i.e.,
433
traits) within the context of classical and contemporary theories on interspecific competition
434
(Fig. 1).
435
436
Trait-similarity and trait-hierarchy jointly determine species’ co-occurrences
437
We found that no single mechanism of competitive exclusion (trait-similarity or trait-
438
hierarchy) was sufficient to explain patterns of co-occurrences between S. invicta and the 28
439
resident ant species. However, incorporating the interactive effects of both mechanisms
440
markedly improved explanatory power for a model based on the morphological trait, pronotum
441
width (Table 1). Consistent with the basic principles of modern coexistence theory (Fig. 1),
442
these results indicate that competitive outcomes among the ant species are unlikely to depend
443
on niche dissimilarities alone, but on the relative magnitudes of these in relation to differences
444
in their competitive abilities (Chesson, 2000). Competitive hierarchies in individual traits are
445
known to structure some communities (e.g., plant height, Kunstler et al., 2016; Perez-Ramos
446
et al., 2019) but are unexplored for most taxa. Our finding that differences in pronotum width
447
significantly predict species’ associations (Table 1) highlights the potential importance of this
448
17
frequently measured ‘functional’ trait (Parr et al., 2017) to competitive interactions among ant
449
species. Given that the pronotums of ant workers contain the musculature powering load-
450
bearing abilities (Keller et al., 2014), one testable hypothesis is that the relatively wider
451
pronotums in S. invicta reflect a competitive advantage over other ant species (Fig. 3) through
452
the more efficient capture, removal and transport of food resources. Notably, exploitative
453
interspecific resource competition among ants is especially intense in less heterogenous
454
habitats (Gibb, 2005), such as the one studied.
455
456
Community assembly via trait differences: four rules
457
The trait model incorporating the interaction term (AD*HD) reconciled the varying co-
458
occurrence patterns between S. invicta and individual ant species to the varying nature of each
459
pair’s trait differences (i.e., in terms of trait-similarity and trait-hierarchy) (Fig. 3).
460
Furthermore, the distinct ways by which species’ trait differences determine their co-
461
occurrences as reflected in the model are strikingly consistent with predictions under different
462
theoretical rules of community assembly. With reference to Fig. 3, our ecological interpretation
463
of the model identifies four alternative rules which determine the pattern of co-occurrence
464
between a given ant species and S. invicta across the landscape. Each rule is distinguished by
465
the specific magnitudes of niche dissimilarities (AD) and competitive ability differences (HD)
466
between paired species. The rules are: (I) competitive exclusion at HD<0 and AD<0.37, leading
467
to negative co-occurrence; (II) approximate competitive equivalence and coexistence at HD>0
468
and AD<0.37, leading to non-negative co-occurrence; (III) sufficiently large niche
469
dissimilarity and coexistence at AD=0.37–0.95, leading to non-negative co-occurrence; and
470
(IV) environmental filtering at AD>0.95, leading to negative co-occurrence.
471
472
Rules I and II apply to species which are largely similar to S. invicta in niches and trait values
473
(AD<0.37). Here the model predicts increasingly negative co-occurrences with increasingly
474
negative HD (Fig. 3: left unmasked area: SES becomes negative as HD becomes negative).
475
These results suggest that for ant species which have similar trait values to S. invicta,
476
interspecific competition with S. invicta is likely to be intense, such that large differences in
477
species’ competitive abilities drive exclusion, resulting in significant negative pairwise co-
478
occurrences (e.g., Kunstler et al., 2012) (Rule I). However, for some species, small differences
479
in competitive abilities with S. invicta (competitive equivalence) may facilitate coexistence
480
with S. invicta in the fashion of neutral-like dynamics (Scheffer & van Nes, 2006) (Rule II).
481
This is evident from the model, which predicts that the likelihood of co-occurrence for S.
482
invicta and a similar species (AD<0.37) does not differ significantly from the null expectation
483
18
(i.e., indicating coexistence is plausible) when HD becomes less negative (Fig. 3: left
484
unmasked area: -1.96<SESsinv <1.96).
485
486
In contrast to Rules I and II which apply to species sharing high niche similarity with S. invicta
487
and competing intensely, Rule III applies to species which are largely dissimilar (AD=0.37–
488
0.95) from S. invicta in niches and trait values – to the extent that such niche dissimilarity may
489
sufficiently mitigate any negative effects of competitive imbalances (e.g., individual plant traits
490
in Perez-Ramos et al., 2019). Hence, for these species, differences in competitive abilities do
491
not influence co-occurrences with S. invicta significantly (Fig. 3: masked area: SESsinv does
492
not significantly respond to HD). In addition, if niche dissimilarities are sufficiently large,
493
coexistence is plausible, and the likelihood of these species occurring in the same plots as S.
494
invicta generally does not differ significantly from null expectations (Fig. 3: masked area: -
495
1.96<SESsinv<1.96).
496
497
Rules I, II and III above concern interspecific competition, which we initially predicted to be
498
an important driver of the ant species’ co-occurrences given the relatively homogeneous
499
landscape. Less anticipated was an additional rule (IV), which likely relates to environmental
500
factors. Rule IV applies to the minority of species which are most dissimilar from S. invicta in
501
niches and trait values (Fig. 3: right unmasked area). For any species with such peak
502
dissimilarity from S. invicta (AD>0.95), the model inherently predicts significant negative co-
503
occurrence (SESsinv<-1.96) with S. invicta (Fig. 3). The extensive dissimilarities in trait values
504
between these species and S. invicta, and the low likelihood of co-occurrence, may result from
505
environmental filtering by unmeasured factors that vary across the plots (e.g., community-
506
weighted means in ant species’ pronotum widths respond to gradients of soil fertility in Fichaux
507
et al., 2019). If such trait-based environmental filtering occurs, directional differences in trait
508
values could further reinforce its deterministic effects – this would explain the increasingly
509
negative co-occurrence patterns observed with increasing HD (Fig. 3: right unmasked area).
510
511
In sum, different assembly rules collectively account for the co-occurrences of the invader S.
512
invicta and the 28 resident ant species across the landscape, highlighting the multifaceted
513
nature of community assembly. These findings broadly address the context-dependent nature
514
of the impacts of S. invicta invasions on native ants in the collective literature (Porter &
515
Savignano, 1990; Gotelli & Arnett, 2000; King & Tschinkel, 2008).
516
517
19
Abundant species, ranging from ants and beetles to trees and corals, often display negative and
518
positive spatial associations with many other species (Calatayud et al., 2020). The ‘trait-based
519
co-occurrence’ approach used in this study can provide insight into these ubiquitous patterns.
520
Our parsimonious, single-trait model encompassing species’ trait differences (in terms of trait-
521
similarity, trait-hierarchy, and their interacting effects) (Table 1; Fig. 3) reveals that an
522
abundant species competes intensely with a subset of similar species, may coexist with species
523
that are sufficiently different, and is further unlikely to co-occur with other species of different
524
environmental requirements. This provides a realistic view of community assembly as a
525
dynamic and multifaceted process acting varyingly on different pairs or sets of species
526
(Abrams, 1983) – an alternative to the problematic notion of assembly as occurring via a static
527
and discrete set of environmental and biotic ‘filters’ (Cadotte & Tucker, 2017).
528
529
Additional factors likely also influence co-occurrences between S. invicta and the 28 resident
530
species, since the best individual trait model explained 37% of the variance (Table 1; Table
531
S3). Also, if pronotum width was the only trait determining competitive outcomes, with wider
532
pronotums indicating superior competitive abilities, we would expect ant species with
533
AD<0.37 and HD>0 relative to S. invicta to exclude and show negative co-occurrences with S.
534
invicta, instead of the non-negative co-occurrences predicted by the model (Fig. 3). In addition
535
to the morphological traits measured in this study, other traits of ant species such as colony
536
size or relative levels of intra- and interspecific aggression could potentially affect interspecific
537
competition (Arnan, Cerdá & Retana, 2012). More broadly, the odds of competitive exclusion
538
– and consequently precise patterns of species’ co-occurrences – are likely to depend on a net
539
difference in competitive ability across multiple trait axes (Kraft et al., 2015). Thus, we
540
envisage that the understanding of assembly processes such as interspecific competition in
541
ecological communities can be enhanced by explicitly assessing trait-similarity, trait-hierarchy,
542
and their interaction across diverse suites of morphological, physiological and behavioural
543
traits. One could also extend the approaches used in this study into multi-dimensional trait
544
space, where there is some evidence for the strong stabilizing effects of species’ dissimilarities
545
(Kraft et al., 2015), but competitive hierarchies or the interacting effects of trait-similarity and
546
trait-hierarchy are less explored.
547
548
Trait-based co-occurrence: a framework for investigating community assembly
549
This study has shown that understanding of community assembly processes can be enhanced
550
via a hypothesis-driven framework incorporating species’ trait differences and co-occurrence
551
networks. Evidently, one advantage of such an approach is that it allows for the detection and
552
20
consolidation of multiple assembly processes and their interactions, across fine ecological
553
scales (species pairs and community subsets); whereas these processes may fail to be
554
represented in coarser community-wide metrics such as functional or phylogenetic
555
overdispersion and clustering (Mayfield & Levine, 2010). While experimental manipulations
556
and mesocosm studies can be invaluable for understanding the precise mechanisms underlying
557
community dynamics, their applicability decreases with increasing ecological, spatial and
558
temporal scales (Levin, 1992). Data on species’ traits, abundances and distributions across
559
multiple scales are increasingly collected and shared (Gallagher et al., 2019). For most species-
560
rich ecological communities, we suggest that the trait-based co-occurrence approach represents
561
an efficient and promising avenue for investigating assembly processes, and for identifying
562
particular interactions, species and traits that are important determinants of community
563
structure.
564
565
566
AKNOWLEDGEMENTS
567
We thank Carlos Carmona and Christopher Terry for their comments on a previous version of
568
the manuscript. This work was supported by a National Geographic Grant (60-16) and a
569
University of Oxford Clarendon Scholarship to MKLW.
570
571
572
REFERENCES
573
Abrams, P. (1983). The theory of limiting similarity. Annu. Rev. Ecol. Syst., 14, 359-376.
574
Agresti, A. (2018). An introduction to categorical data analysis. Second Edition. Wiley, New
575
Jersey.
576
Araújo, M. B., Rozenfeld, A., Rahbek, C., & Marquet, P. A. (2011). Using species co‐
577
occurrence networks to assess the impacts of climate change. Ecography, 34, 897-908.
578
Arnan, X., Cerdá, X., & Retana, J. (2012). Distinctive life traits and distribution along
579
environmental gradients of dominant and subordinate Mediterranean ant species.
580
Oecologia, 170, 489-500.
581
Brazeau, H. A., & Schamp, B. S. (2019). Examining the link between competition and negative
582
co‐occurrence patterns. Oikos, 128, 1358-1366.
583
Cadotte, M. W., & Tucker, C. M. (2017). Should environmental filtering be abandoned?.
584
Trends Ecol. Evol., 32, 429-437.
585
21
Calatayud, J., Andivia, E., Escudero, A., Melián, C. J., Bernardo-Madrid, R., Stoffel, M., et al.
586
(2020). Positive associations among rare species and their persistence in ecological
587
assemblages. Nat. Ecol. Evol., 4, 40-45.
588
Carmona, C. P., de Bello, F., Azcárate, F. M., Mason, N. W., & Peco, B. (2019a). Trait
589
hierarchies and intraspecific variability drive competitive interactions in Mediterranean
590
annual plants. J. Ecol., 107, 2078-2089.
591
Carmona, C. P., de Bello, F., Mason, N. W., & Lepš, J. (2019b). Trait probability density
592
(TPD): measuring functional diversity across scales based on TPD with R. Ecology, 100,
593
e02876.
594
Carpenter, S. R. (1996). Microcosm experiments have limited relevance for community and
595
ecosystem ecology. Ecology, 77, 677-680.
596
Cerda, X., Arnan, X., & Retana, J. (2013). Is competition a significant hallmark of ant
597
(Hymenoptera: Formicidae) ecology. Myrmecol. News, 18, 131-147.
598
Chesson, P. (2000). Mechanisms of maintenance of species diversity. Annu. Rev. Ecol. Syst.,
599
31, 343-366.
600
de Bello, F., Price, J. N., Münkemüller, T., Liira, J., Zobel, M., Thuiller, W., et al.. (2012).
601
Functional species pool framework to test for biotic effects on community assembly.
602
Ecology, 93, 2263-2273.
603
Dormann, C. F., Elith, J., Bacher, S., Buchmann, C., Carl, G., Carré, G., et al. (2013).
604
Collinearity: a review of methods to deal with it and a simulation study evaluating their
605
performance. Ecography, 36, 27-46.
606
Eguchi, K., Bui, T. V., & Yamane, S. (2004). A preliminary study on foraging distance and
607
nesting sites of ants in Indo-Chinese lowland vegetation. Sociobiology, 43, 445-457.
608
Esarey, J. & Sumner, J.L. (2017). Marginal effects in interaction models: determining and
609
controlling the false positive rate. Comp. Polit. Stud.,. 51, 1144-1176.
610
Fichaux, M., Béchade, B., Donald, J., Weyna, A., Delabie, J. H. C., Murienne, J., et al. (2019).
611
Habitats shape taxonomic and functional composition of Neotropical ant assemblages.
612
Oecologia, 189, 501-513.
613
Gallagher, R. V., Falster, D. S., Maitner, B. S., Salguero-Gomez, R., Vandvik, V., Pearse, W.
614
D., et al. (2019). The open traits network: Using open science principles to accelerate
615
trait-based science across the tree of life. EcoEvoRxiv. doi: 10.32942/osf.io/kac45
616
Gibb, H. (2005). The effect of a dominant ant, Iridomyrmex purpureus, on resource use by ant
617
assemblages depends on microhabitat and resource type. Austral Ecol., 30, 856-867.
618
Gotelli, N. J., & Arnett, A. E. (2000). Biogeographic effects of red fire ant invasion. Ecol. Lett.,
619
3, 257-261.
620
22
Gotelli, N. J. (2000). Null model analysis of species co‐occurrence patterns. Ecology, 81, 2606-
621
2621.
622
Grainger, T. N., Levine, J. M., & Gilbert, B. (2019). The invasion criterion: A common
623
currency for ecological research. Trends Ecol. Evol., 39, 925-935.
624
Gross, N., Liancourt, P., Butters, R., Duncan, R. P., & Hulme, P. E. (2015). Functional
625
equivalence, competitive hierarchy and facilitation determine species coexistence in
626
highly invaded grasslands. New Phytol., 206, 175-186.
627
Gurevitch, J., Morrow, L.L., Wallace, A. & Walsh, J.S. (1992). A meta-analysis of field
628
experiments on competition. Am. Nat., 140, 539-572.
629
Herben, T., & Goldberg, D. E. (2014). Community assembly by limiting similarity vs.
630
competitive hierarchies: testing the consequences of dispersion of individual traits. J.
631
Ecol., 102, 156-166.
632
Holway, D. A. (1999). Competitive mechanisms underlying the displacement of native ants by
633
the invasive Argentine ant. Ecology, 80, 238-251.
634
Hutchinson, G. E. (1959). Homage to Santa Rosalia or why are there so many kinds of
635
animals?. Am. Nat., 93, 145-159.
636
Johnson, P. O., & Neyman, J. (1936). Tests of certain linear hypotheses and their application
637
to some educational problems. Stat. Res. Mem., 1, 57-93.
638
Keddy, P. A. (1992). Assembly and response rules: two goals for predictive community
639
ecology. J. Veg. Sci., 3, 157-164.
640
Keller, R. A., Peeters, C., & Beldade, P. (2014). Evolution of thorax architecture in ant castes
641
highlights trade-off between flight and ground behaviors. Elife, 3, e01539.
642
King, J. R., & Tschinkel, W. R. (2008). Experimental evidence that human impacts drive fire
643
ant invasions and ecological change. PNAS, 105, 20339-20343.
644
Kraft, N. J., Godoy, O., & Levine, J. M. (2015). Plant functional traits and the multidimensional
645
nature of species coexistence. PNAS, 112, 797-802.
646
Kunstler, G., Falster, D., Coomes, D. A., Hui, F., Kooyman, R. M., Laughlin, D. C., et al.
647
(2016). Plant functional traits have globally consistent effects on competition. Nature,
648
529, 204-207.
649
Kunstler, G., Lavergne, S., Courbaud, B., Thuiller, W., Vieilledent, G., Zimmermann, N. E.,
650
et al. (2012). Competitive interactions between forest trees are driven by species' trait
651
hierarchy, not phylogenetic or functional similarity: implications for forest community
652
assembly. Ecol. Lett., 15, 831-840.
653
23
Lane, P. W., Lindenmayer, D. B., Barton, P. S., Blanchard, W., & Westgate, M. J. (2014).
654
Visualization of species pairwise associations: a case study of surrogacy in bird
655
assemblages. Ecol. Evol., 4, 3279-3289.
656
LeBrun, E. G., Plowes, R. M., & Gilbert, L. E. (2012). Imported fire ants near the edge of their
657
range: disturbance and moisture determine prevalence and impact of an invasive social
658
insect. J. Anim. Ecol., 81, 884-895.
659
Lessard, J. P., Fordyce, J. A., Gotelli, N. J., & Sanders, N. J. (2009). Invasive ants alter the
660
phylogenetic structure of ant communities. Ecology, 90, 2664-2669.
661
Levin, S. A. (1992). The problem of pattern and scale in ecology: the Robert H. MacArthur
662
award lecture. Ecology, 73, 1943-1967.
663
MacArthur, R., & Levins, R. (1967). The limiting similarity, convergence, and divergence of
664
coexisting species. Am. Nat., 101, 377-385.
665
MacDougall, A. S., Gilbert, B., & Levine, J. M. (2009). Plant invasions and the niche. J. Ecol.,
666
97, 609-615.
667
Mayfield, M. M., & Levine, J. M. (2010). Opposing effects of competitive exclusion on the
668
phylogenetic structure of communities. Ecol. Lett., 13, 1085-1093.
669
Mittelbach, G. G., & McGill, B. J. (2019). Community ecology. Second edition. Oxford
670
University Press, Oxford.
671
Morgan, B., & Guénard, B. (2019). New 30 m resolution Hong Kong climate, vegetation, and
672
topography rasters indicate greater spatial variation than global grids within an urban
673
mosaic. Earth Syst. Sci. Data, 11, 1083-1098.
674
Parr, C. L. (2008). Dominant ants can control assemblage species richness in a South African
675
savanna. J. Anim. Ecol., 77, 1191-1198.
676
Parr, C. L., Dunn, R. R., Sanders, N. J., Weiser, M. D., Photakis, M., Bishop, T. R., et al.
677
(2017). GlobalAnts: A new database on the geography of ant traits (Hymenoptera:
678
Formicidae). Insect Conserv. Divers., 10, 5-20.
679
Pérez-Ramos, I. M., Matías, L., Gómez-Aparicio, L., & Godoy, Ó. (2019). Functional traits
680
and phenotypic plasticity modulate species coexistence across contrasting climatic
681
conditions. Nat. Commun., 10, 1-11.
682
Porter, S. D., & Savignano, D. A. (1990). Invasion of polygyne fire ants decimates native ants
683
and disrupts arthropod community. Ecology, 71, 2095-2106.
684
R Development Core Team. (2017). R: a language and environment for statistical computing.
685
Available at: [https://www.r-project.org/].
686
Scheffer, M., & van Nes, E. H. (2006). Self-organized similarity, the evolutionary emergence
687
of groups of similar species. PNAS, 103, 6230-6235.
688
24
Schielzeth, H. (2010). Simple means to improve the interpretability of regression coefficients.
689
Methods Ecol. Evol., 1, 103-113.
690
Schoener, T. W. (1974). Resource partitioning in ecological communities. Science, 185, 27-
691
39.
692
Shea, K., & Chesson, P. (2002). Community ecology theory as a framework for biological
693
invasions. Trends Ecol. Evol., 17, 170-176.
694
Siepielski, A. M., & McPeek, M. A. (2010). On the evidence for species coexistence: a critique
695
of the coexistence program. Ecology, 91, 3153-3164.
696
Veech, J. A. (2014). The pairwise approach to analysing species co‐occurrence. J. Biogeogr.,
697
41, 1029-1035.
698
Weeks Jr, R. D., Wilson, L. T., & Vinson, S. B. (2004). Resource partitioning among colonies
699
of polygyne red imported fire ants (Hymenoptera: Formicidae). Environ. Entomol., 33,
700
1602-1608.
701
Wong, M. K. L., Guénard, B., & Lewis, O. T. (2019). Trait‐based ecology of terrestrial
702
arthropods. Biol. Rev., 94, 999-1022.
703
Wong, M. K. L., Guénard, B., & Lewis, O. T. (2020). The cryptic impacts of invasion:
704
functional homogenization of tropical ant communities by invasive fire ants. Oikos. doi:
705
10.1111/oik.06870.
706
| 2020 | Trait-similarity and trait-hierarchy jointly determine co-occurrences of resident and invasive ant species | 10.1101/2020.02.05.935858 | [
"Wong Mark K. L.",
"Tsang Toby P. N.",
"Lewis Owen T.",
"Guénard Benoit"
] | creative-commons |
1
Decoding across sensory modalities reveals common
1
supramodal signatures of conscious perception
2
3
Gaëtan Sanchez1,2,3, Thomas Hartmann1,2, Marco Fuscà1,2 , Gianpaolo Demarchi1,2 and
4
Nathan Weisz1,2
5
6
1 – Paris-Lodron Universität Salzburg, Centre for Cognitive Neuroscience and Division of Physiological
7
Psychology, Hellbrunnerstraße 34, 5020 Salzburg, Austria
8
2 – Center for Mind/Brain Sciences (CIMeC), University of Trento, via delle Regole 101, 38123 Mattarello (TN),
9
Italy
10
3 – Lyon Neuroscience Research Center (CRNL), Inserm U1028, CNRS UMR5292, University Lyon 1, Centre
11
Hospitalier Le Vinatier - Bât. 452, 95 boulevard Pinel, 69675 Bron, France
12
* Corresponding author: gaetan.sanchez@inserm.fr
13
Keywords
14
consciousness; perception; near-threshold stimulation; multivariate analysis; decoding
15
analysis; magnetoencephalography
16
Abstract
17
An increasing number of studies highlight common brain regions and processes in
18
mediating conscious sensory experience. While most studies have been performed in the
19
visual modality, it is implicitly assumed that similar processes are involved in other sensory
20
modalities. However, the existence of supramodal neural processes related to conscious
21
perception has not been convincingly shown so far. Here, we aim to directly address this issue
22
by investigating whether neural correlates of conscious perception in one modality can predict
23
conscious perception in a different modality. In two separate experiments, we presented
24
participants with successive blocks of near-threshold tasks involving tactile, visual or auditory
25
stimuli during the same magnetoencephalography (MEG) acquisition. Using decoding
26
analysis in the post-stimulus period between sensory modalities, our first experiment
27
uncovered supramodal spatio-temporal neural activity patterns predicting conscious
28
perception of the feeble stimulation. Strikingly, these supramodal patterns included activity in
29
primary sensory regions not directly relevant to the task (e.g. neural activity in visual cortex
30
predicting conscious perception of auditory near-threshold stimulation). We carefully replicate
31
our results in a control experiment that furthermore show that the relevant patterns are
32
independent of the type of report (i.e. whether conscious perception was reported by pressing
33
or withholding a button-press). Using standard paradigms for probing neural correlates of
34
conscious perception, our findings reveal a common signature of conscious access across
35
sensory modalities and illustrate the temporally late and widespread broadcasting of neural
36
representations, even into task-unrelated primary sensory processing regions.
37
38
2
39
Introduction
40
While the brain can process an enormous amount of sensory information in parallel,
41
only some information can be consciously accessed, playing an important role in the way we
42
perceive and act in our surrounding environment. An outstanding goal in cognitive
43
neuroscience is thus to understand the relationship between neurophysiological processes
44
and conscious experiences. However, despite tremendous research efforts, the precise brain
45
dynamics that enable certain sensory information to be consciously accessed remain
46
unresolved. Nevertheless, progress has been made in research focusing on isolating neural
47
correlates of conscious perception (1), in particular suggesting that conscious perception - at
48
least if operationalized as reportability (2) - of external stimuli crucially depends on the
49
engagement of a widely distributed brain network (3). To study neural processes underlying
50
conscious perception, neuroscientists often expose participants to near-threshold (NT) stimuli
51
that are matched to their individual perceptual thresholds (4). In NT experiments, there is a
52
trial-to-trial variability in which around 50% of the stimuli at NT-intensity are consciously
53
perceived. Because of the fixed intensity, the physical differences between stimuli within the
54
same modality can be excluded as a determining factor leading to reportable sensation (5).
55
Despite numerous methods used to investigate conscious perception of external events, most
56
studies target a single sensory modality. However, any specific neural pattern identified as a
57
correlate of consciousness needs evidence that it generalizes to some extent, e.g. across
58
sensory modalities. We argue that this has not been convincingly shown so far.
59
In the visual domain, it has been shown that reportable conscious experience is present
60
when primary visual cortical activity extends towards hierarchically downstream brain areas
61
(6), requiring the activation of frontoparietal regions in order to become fully reportable (7).
62
Nevertheless, a recent MEG study using a visual masking task revealed early activity in
63
primary visual cortices as the best predictor for conscious perception (8). Other studies have
64
shown that neural correlates of auditory consciousness relate to the activation of fronto-
65
temporal rather than fronto-parietal networks (9, 10). Additionally, recurrent processing
66
3
between primary, secondary somatosensory and premotor cortices have been suggested as
67
potential neural signatures of tactile conscious perception (11, 12). Indeed, recurrent
68
processing between higher and lower order cortical regions within a specific sensory system
69
is theorized to be a marker of conscious processing (6, 13, 14). Moreover, alternative theories
70
such as the global workspace framework (15) extended by Dehaene et al. (16) postulates that
71
the frontoparietal engagement aids in ‘broadcasting’ relevant information throughout the brain,
72
making it available to various cognitive modules. In various electrophysiological experiments,
73
it has been shown that this process is relatively late (~300 ms), and could be related to
74
increased evoked brain activity after stimulus onset such as the so-called P300 signal (17–
75
19). Such late brain activities seem to correlate with perceptual consciousness and could
76
reflect the global broadcasting of an integrated stimulus making it conscious. Taken together,
77
theories and experimental findings argue in favor of various ‘signatures’ of consciousness from
78
recurrent activity within sensory regions to a global broadcasting of information with
79
engagement of fronto-parietal areas. Even though usually implicitly assumed, it is so far
80
unclear whether similar spatio-temporal neural activity patterns are linked to conscious access
81
across different sensory modalities.
82
In the current study, we investigated conscious perception in different sensory systems
83
using multivariate analysis on MEG data. Our working assumption is that brain activity related
84
to conscious access has to be independent from the sensory modality: i.e. supramodal
85
consciousness-related neural processes need to exhibit spatio-temporal generalization. Such
86
a hypothesis is most ideally tested applying decoding methods to electrophysiological signals
87
recorded while probing conscious access in different sensory modalities. The application of
88
multivariate pattern analysis (MVPA) to EEG/MEG measurements offers increased sensitivity
89
in detecting experimental effects distributed over space and time (20–23). MVPA is often used
90
in combination with a searchlight method (24, 25), which involves sliding a small spatial
91
window over the data to reveal areas containing decodable information. The combination of
92
both methods provides spatio-temporal detection of optimal decodability, determining where,
93
when and for how long a specific pattern is present in brain activity. Such multivariate decoding
94
4
analyses have been proposed as an alternative in consciousness research, complementing
95
other conventional univariate approaches in order to identify neural activity predictive of
96
conscious experience at the single trial level (26).
97
Here, we acquired MEG data while each participant performed three different standard
98
NT tasks on three sensory modalities with the aim of characterizing supramodal brain
99
mechanisms of conscious perception. In the first experiment we show how neural patterns
100
related to perceptual consciousness can be generalized over space and time within and –most
101
importantly- between different sensory systems by using classification analysis on source-
102
level reconstructed brain activity. In an additional control experiment, we replicate the main
103
findings and exclude the possibility that our observed patterns are due to response preparation
104
/ selection.
105
106
Materials and Methods
107
Participants
108
Twenty-five healthy volunteers took part in the initial experiment conducted in Trento
109
and twenty-one healthy volunteers took part in the control experiment performed in Salzburg.
110
All participants presented normal or corrected-to-normal vision and no neurological or
111
psychiatric disorders. Three participants for the initial experiment and one participant for the
112
control experiment were excluded from the analysis due to excessive artifacts in the MEG data
113
leading to an insufficient number of trials per condition after artifact rejection (less than 30
114
trials for at least one condition). Additionally, within each experiment six participants were
115
discarded from the analysis because false alarms rate exceeded 30% and/or near-threshold
116
detection rate was over 85% or below 15% for at least one sensory modality (due to threshold
117
identification failure and difficulty to use response button mapping during the control
118
experiment, also leaving less than 30 trials for at least one relevant condition in one sensory
119
modality: detected or undetected). The remaining 16 participants (11 females, mean age: 28.8
120
years; SD: 3.4 years) for the initial experiment and 14 participants (9 females, mean age: 26.4
121
5
years; SD: 6.4 years) for the control experiment, reported normal tactile and auditory
122
perception. The ethics committee of the University of Trento and University of Salzburg
123
respectively, approved the experimental protocols that were used with the written informed
124
consent of each participant.
125
126
Stimuli
127
To ensure that the participant did not hear any auditory cues caused by the piezo-
128
electric stimulator during tactile stimulation, binaural white noise was presented during the
129
entire experiment (training blocks included). Auditory stimuli were presented binaurally using
130
MEG-compatible tubal in-ear headphones (SOUNDPixx, VPixx technologies, Canada). Short
131
bursts of white noise with a length of 50 ms were generated with Matlab and multiplied with a
132
Hanning window to obtain a soft on- and offset. Participants had to detect short white noise
133
bursts presented near hearing threshold (27). The intensity of such transient target auditory
134
stimuli was determined prior to the experiment in order to emerge from the background
135
constant white noise stimulation. Visual stimuli were Gabor ellipsoid (tilted 45°) back-projected
136
on a translucent screen by a Propixx DLP projector (VPixx technologies, Canada) at a refresh
137
rate of 180 frames per second. The stimuli were presented 50 ms in the center of the screen
138
at a viewing distance of 110 cm. Tactile stimuli were delivered with a 50 ms stimulation to the
139
tip of the left index finger, using one finger module of a piezo-electric stimulator (Quaerosys,
140
Schotten, Germany) with 2 × 4 rods, which can be raised to a maximum of 1 mm. The module
141
was attached to the finger with tape and the participant’s left hand was cushioned to prevent
142
any unintended pressure on the module (28). For the control experiment (conducted in another
143
laboratory; i.e. Salzburg), visual, auditory and tactile stimulation setups were identical but we
144
used a different MEG/MRI vibrotactile stimulator system (CM3, Cortical Metrics).
145
146
Task and design
147
The participants performed three blocks of a NT perception task. Each block included
148
three separate runs (100 trials each) for each sensory modality, tactile (T), auditory (A) and
149
6
visual (V). A short break (~1 min) separated each run and longer breaks (~4 min) were
150
provided to the participants after each block. Inside a block, runs alternated in the same order
151
within subject and were pseudo-randomized across subjects (i.e. subject 1 = TVA-TVA-TVA;
152
subject 2 = VAT-VAT-VAT; …). Participants were asked to fixate on a central white dot in a
153
grey central circle at the center of the screen throughout the whole experiment to minimize
154
eye movements.
155
A short training run with 20 trials was conducted to ensure that participants had
156
understood the task. Then, in three different training sessions prior to the main experiment,
157
participants’ individual perceptual thresholds (tactile, auditory and visual) were determined in
158
the shielded room. For the initial experiment, a 1-up/1-down staircase procedure with two
159
randomly interleaved staircases (one up- and one downward) was used with fixed step sizes.
160
For the control experiment we used a Bayesian active sampling protocol to estimate
161
psychometric slope and threshold for each participant (60).
162
The main experiment consisted of a detection task (Figure 1A). At the beginning of
163
each run, participants were told that on each trial a weak stimulus (tactile, auditory or visual
164
depending on the run) could be presented at random time intervals. 500 ms after the target
165
stimulus onset, participants were prompted to indicate whether they had felt the stimulus with
166
an on-screen question mark (maximal response time: 2 s). Responses were given using MEG-
167
compatible response boxes with the right index finger and the middle finger (response button
168
mapping was counterbalanced among participants). Trials were then classified into hits
169
(detected) and misses (undetected stimulus) according to the participants’ answers. Trials with
170
no response were rejected. Catch (above perceptual threshold stimulation intensity) and Sham
171
(absent stimulation) trials were used to control false alarms and correct rejection rates across
172
the experiment. Overall, there were 9 runs with 100 trials each (in total 300 trials for each
173
sensory modality). Each trial started with a variable interval (1.3–1.8 s, randomly-distributed)
174
followed by an experimental near-threshold stimulus (80 per run), a sham stimulus (10 per
175
run) or a catch stimulus (10 per run) of 50 ms each. Each run lasted for approximately 5 min.
176
The whole experiment lasted for ~1h.
177
7
Identical timing parameters were used in the control experiment. However, a specific
178
response screen design was used to control for motor response mapping. For each trial the
179
participants must use a different response mapping related to circle’s color surrounding the
180
question mark during response screen. Two colors (blue or yellow) were used and presented
181
randomly after each trial during the control experiment. One color was associated to the
182
following response mapping rule: “press the button only if there is a stimulation” (for near-
183
threshold condition: “detected”) and the other color was associated to the opposite response
184
mapping: “press a button only if there is no stimulation” (for near-threshold condition:
185
“undetected”). The association between one response mapping and a specific color (blue or
186
yellow) was fixed for a single participant but was predefined randomly across different
187
participant. Importantly, by delaying the response-mapping to after the stimulus presentation
188
in a -for the individual- unpredictable manner, neural patterns during relevant periods
189
putatively cannot be confounded by response selection / preparation. Both experiments were
190
programmed in Matlab using the open source Psychophysics Toolbox (61).
191
192
MEG data acquisition and preprocessing
193
MEG was recorded at a sampling rate of 1kHz using a 306-channel (204 first order
194
planar gradiometers, 102 magnetometers) VectorView MEG system for the first experiment in
195
Trento, and Triux MEG system for the control experiment in Salzburg (Elekta-Neuromag Ltd.,
196
Helsinki, Finland) in a magnetically shielded room (AK3B, Vakuumschmelze, Hanau,
197
Germany). Before the experiments, individual head shapes were acquired for each participant
198
including fiducials (nasion, pre-auricular points) and around 300 digitized points on the scalp
199
with a Polhemus Fastrak digitizer (Polhemus, Vermont, USA). Head positions of the individual
200
relative to the MEG sensors were continuously controlled within a run using five coils. Head
201
movements did not exceed 1 cm within and between blocks.
202
Data were analyzed using the Fieldtrip toolbox (62) and the CoSMoMVPA toolbox (63)
203
in combination with MATLAB 8.5 (MathWorks Natick, MA). First, a high-pass filter at 0.1 Hz
204
(FIR filter with transition bandwidth 0.1Hz) was applied to the continuous data. Then the data
205
8
were segmented from 1000 ms before to 1000 ms after target stimulation onset and down-
206
sampled to 512 Hz. Trials containing physiological or acquisition artifacts were rejected. A
207
semi-automatic artifact detection routine identified statistical outliers of trials and channels in
208
the datasets using a set of different summary statistics (variance, maximum absolute
209
amplitude, maximum z-value). These trials and channels were removed from each dataset.
210
Finally, the data were visually inspected and any remaining trials and channels with artifacts
211
were removed manually. Across subjects, an average of 5 channels (± 2 SD) were rejected.
212
Bad channels were excluded from the whole data set. A detailed report of remaining number
213
of trials per condition for each participant can be found in supplementary material (see SI
214
Appendix Table S1). Finally, in all further analyses and within each sensory modality for each
215
subject, an equal number of detected and undetected trials was randomly selected to prevent
216
any bias across conditions (64).
217
218
Source analyses
219
Neural activity evoked by stimulus onset was investigated by computing event-related
220
fields (ERF). For all source-level analyses, the preprocessed data was 30Hz lowpass-filtered
221
and projected to source-level using an LCMV beamformer analysis (65). For each participant,
222
realistically shaped, single-shell headmodels (66) were computed by co-registering the
223
participants’ headshapes either with their structural MRI or – when no individual MRI was
224
available (3 participants and 2 participants, for the initial experiment and the control
225
experiment respectively) – with a standard brain from the Montreal Neurological Institute (MNI,
226
Montreal, Canada), warped to the individual headshape. A grid with 1.5 cm resolution based
227
on an MNI template brain was morphed into the brain volume of each participant. A common
228
spatial filter (for each grid point and each participant) was computed using the leadfields and
229
the common covariance matrix, taking into account the data from both conditions (detected
230
and undetected; or catch and sham) for each sensory modality separately. The covariance
231
window for the beamformer filter calculation was based on 200 ms pre- to 500 ms post-
232
stimulus. Using this common filter, the spatial power distribution was then estimated for each
233
9
trial separately. The resulting data were averaged relative to the stimulus onset in all
234
conditions (detected, undetected, catch and sham) for each sensory modality. Only for
235
visualization purposes a baseline correction was applied to the averaged source-level data by
236
subtracting a time-window from 200 ms pre-stimulus to stimulus onset. Based on a significant
237
difference between event-related fields of the two conditions over time for each sensory
238
modality, the source localization was performed restricted to specific time-windows of interest.
239
All source images were interpolated from the original resolution onto an inflated surface of an
240
MNI template brain available within the Caret software package (67). The respective MNI
241
coordinates and labels of localized brain regions were identified with an anatomical brain atlas
242
(AAL atlas; (68)) and a network parcellation atlas (29).
243
244
Multivariate Pattern Analysis (MVPA) decoding
245
MVPA decoding was performed for the period 0 to 500 ms after stimulus onset based
246
on normalized (z-scored) single trial source data downsampled to 100Hz (i.e. time steps of 10
247
ms). We used multivariate pattern analysis as implemented in CoSMoMVPA (63) in order to
248
identify when and what kind of common network between sensory modality is activated during
249
the near-threshold detection task. We defined two classes for the decoding related to the task
250
behavioral outcome (detected and undetected). For decoding within the same sensory
251
modality, single trial source data were randomly assigned to one of two chunks (half of the
252
original data).
253
For decoding of all sensory modalities together, single trial source data were pseudo-
254
randomly assigned to one of the two chunks with half of the original data for each sensory
255
modality in each chunk. Data were classified using a 2-fold cross-validation procedure, where
256
a Bayes-Naive classifier predicted trial conditions in one chunk after training on data from the
257
other chunk. For decoding between different sensory modality, single trial source data of one
258
modality were assigned to one testing chunk and the trials from other modalities were
259
assigned to the training chunk. The number of target categories (e.g. detected / undetected)
260
10
was balanced in each training partition and for each sensory modality. Training and testing
261
partitions always contained different sets of data.
262
First, the temporal generalization method was used to explore the ability of each
263
classifier across different time points in the training set to generalize to every time point in the
264
testing set (21). In this analysis we used local neighborhoods features in time space (time
265
radius of 10ms: for each time step we included as additional features the previous and next
266
time sample data point). We generated temporal generalization matrices of task decoding
267
accuracy (detected/undetected), mapping the time at which the classifier was trained against
268
the time it was tested. Generalization of decoding accuracy over time was calculated for all
269
trials and systematically depended on a specific between or within sensory modality decoding.
270
The reported average accuracy of the classifier for each time point corresponds to the group
271
average of individual effect-size: the ability of classifiers to discriminate ‘detected’ from
272
‘undetected’ trials. We summarized time generalization by keeping only significant accuracy
273
for each sensory modalities decoding. Significant classifiers’ accuracies were normalized
274
between 0 and 1:
275
𝑦𝑡 =
𝑥𝑡−𝑚𝑖𝑛(𝑥)
𝑚𝑎𝑥(𝑥)−𝑚𝑖𝑛(𝑥)
(1)
276
Where 𝑥 is a variable of all significant decoding accuracies and 𝑥𝑡 is a given significant
277
accuracy at time 𝑡. Normalized accuracies (𝑦𝑡) were then averaged across significant testing
278
time and decoding conditions. The number of significant classifier generalization across
279
testing time points and the relevant averaged normalized accuracies were reported along
280
training time dimension (see Figure 3B and 5B). For all significant time points previously
281
identified we performed a ‘searchlight’ analysis across brain sources and time neighborhood
282
structure. In this analysis we used local neighborhoods features in source and time space. We
283
used a time radius of 10ms and a source radius of 3 cm. All significant searchlight accuracy
284
results were averaged over time and only the maximum 10% significant accuracy were
285
reported on brain maps for each sensory modality decoding condition (Figure 4) or for all
286
conditions together (Figure 5C).
287
11
Finally, we applied the same type of analysis to all sensory modalities by taking all
288
blocks together with detected and undetected NT trials (equalized within each sensory
289
modality). For the control experiment, we equalized trials based on the 2x2 design with
290
detection report (“detected” or “undetected”) and type of response (“button press = response”
291
or “no response”), so that we get the same number of trials inside each category (i.e. class)
292
for each sensory modality. We performed similar decoding analysis by using different classes
293
definition: either “detected vs. undetected” or “response vs. no response” (SI Appendix, Figure
294
S3B and C).
295
296
Statistical analysis
297
Detection rates for the experimental trials were statistically compared to those from the
298
catch and sham trials, using a dependent-samples T-Test. Concerning the MEG data, the
299
main statistical contrast was between trials in which participants reported a stimulus detection
300
and trials in which they did not (detected vs. undetected).
301
The evoked response at the source level was tested at the group level for each of the
302
sensory modalities. To eliminate polarity, statistics were computed on the absolute values of
303
source-level event-related responses. Based on the global average of all grid points, we first
304
identified relevant time periods with maximal difference between conditions (detected vs.
305
undetected) by performing group analysis with sequential dependent T-tests between 0 and
306
500 ms after stimulus onset using a sliding window of 30 ms with 10ms overlap. P-values were
307
corrected for multiple comparisons using Bonferroni correction. Then, in order to derive the
308
contributing spatial generators of this effect, the conditions ‘detected’ and ‘undetected’ were
309
contrasted for the specific time periods with group statistical analysis using nonparametric
310
cluster-based permutation tests with Monte Carlo randomization across grid points controlling
311
for multiple comparisons (69).
312
The multivariate searchlight analysis results discriminating between conditions were
313
tested at the group level by comparing the resulting individual accuracy maps against chance
314
level (50%) using a non-parametric approach implemented in CoSMoMVPA (63) adopting
315
12
10.000 permutations to generate a null distribution. P-values were set at p<0.005 for cluster
316
level correction to control for multiple comparisons using a threshold-free method for clustering
317
(70), which has been used and validated for MEG/EEG data (38, 71). The time generalization
318
results at the group level were thresholded using a mask with corrected z-score>2.58 (or
319
pcorrected<0.005) (Figure 3A and 5A). Time points exceeding this threshold were identified and
320
reported for each training data time course to visualize how long time generalization was
321
significant over testing data (Figure 3B and 5B). Significant accuracy brain maps resulting
322
from the searchlight analysis on previously identified time points were reported for each
323
decoding condition. The maximum 10% of averaged accuracies were depicted for each
324
significant decoding cluster on brain maps (Figure 4 and 5).
325
326
327
Results
328
Behavior
329
We investigated participants’ detection rate for NT, Sham and Catch trials separately
330
for the initial and the control experiment. During the initial experiment participants had to wait
331
for a response screen and press a button on each trial to report their perception (Figure 1A).
332
During the control experiment, however a specific response screen was used to control for
333
motor response mapping. At each trial the participants must use a different response mapping
334
related to circle’s color surrounding the question mark during response screen (see Figure
335
1C). For the initial experiment and across all participants (N = 16), detection rates for NT
336
experimental trials were: 50% (SD: 11%) for auditory runs, 56% (SD: 12%) for visual runs and
337
55% (SD: 8%) for tactile runs. The detection rates for the catch trials were 92% (SD: 11%) for
338
auditory runs, 90% (SD: 12%) for visual runs and 96% (SD: 5%) for tactile runs. The mean
339
false alarm rates in sham trials were 4% (SD: 4%) for auditory runs, 4% (SD: 4%) for visual
340
runs and 4% (SD: 7%) for tactile runs (Figure 1B). Detection rates of NT experimental trials in
341
all sensory modality significantly differed from those of catch trials (auditory: T15 = −14.44, p
342
13
< 0.001; visual: T15 = −9.47, p < 0.001; tactile: T15 = −20.16, p < 0.001) or sham trials
343
(auditory: T15 = 14.66, p < 0.001; visual: T15 = 16.99, p < 0.001; tactile: T15 = 20.66, p <
344
0.001). Similar results were observed for the control experiment across all participants (N =
345
14), detection rates for NT experimental trials were: 52% (SD: 17%) for auditory runs, 43%
346
(SD: 17%) for visual runs and 42% (SD: 12%) for tactile runs. The detection rates for the catch
347
trials were 97% (SD: 2%) for auditory runs, 95% (SD: 5%) for visual runs and 95% (SD: 4%)
348
for tactile runs. The mean false alarm rates in sham trials were 11% (SD: 4%) for auditory
349
runs, 7% (SD: 6%) for visual runs and 7% (SD: 6%) for tactile runs (Figure 1B). Detection rates
350
of NT experimental trials in all sensory modality significantly differed from those of catch trials
351
(auditory: T13 = −9.64, p < 0.001; visual: T13 = −10.78, p < 0.001; tactile: T13 = −14.75, p <
352
0.001) or sham trials (auditory: T13 = 7.85, p < 0.001; visual: T13 = 6.24, p < 0.001; tactile:
353
T13 = 9.75, p < 0.001). Overall the behavioral results are comparable to other studies (27,
354
28). Individual reaction-times and performances are reported in supplementary materials (see
355
SI Appendix Table S2).
356
357
14
358
Figure 1. Experimental designs and behavioral results. (A-B) Initial experiment; (C-D) Control experiment; (A)
359
After a variable inter-trial interval between 1.3-1.8 s during which participants fixated on a central white dot, a
360
tactile/auditory/visual stimulus (depending on the run) was presented for 50 ms at individual perceptual intensity.
361
After 500 ms, stimulus presentation was followed by an on-screen question mark, and participants indicated their
362
perception by pressing one of two buttons (i.e. stimulation was ‘present’ or ‘absent’) with their right hand. (B & D)
363
The group average detection rates for NT stimulation were around 50% across the different sensory modalities.
364
Sham trials in white (no stimulation) and Catch trials in dark (high intensity stimulation) were significantly different
365
from the NT condition in grey within the same sensory modality for both experiments. Error bars depict the standard
366
deviation. (C) Identical timing parameters were used in the control experiment; however, a specific response screen
367
design was used to control for motor response mapping. Each trial the participants must use a different response
368
mapping related to circle’s color surrounding the question mark during response screen. Two colors (blue or yellow)
369
were used and presented randomly during the control experiment. One color was associated to the following
370
response mapping rule: “press the button only if there is a stimulation” (for near-threshold condition: “detected”)
371
and the other color was associated to the opposite response mapping: “press a button only if there is no stimulation”
372
(for near-threshold condition: “undetected”). The association between one response mapping and a specific color
373
(blue or yellow) was fixed for a single participant but was predefined randomly across different participant.
374
375
15
376
Event-related neural activity
377
To compare poststimulus processing for ‘detected’ and ‘undetected’ trials, evoked
378
responses were calculated at the source level for the initial experiment. As a general pattern
379
over all sensory modalities, source-level event-related fields (ERF) averaged across all brain
380
sources show that stimuli reported as detected resulted in pronounced post-stimulus neuronal
381
activity, whereas unreported stimuli did not (Figure 2A). Similar general patterns were
382
observed for the control experiment with identical univariate analysis (see SI Appendix Figure
383
S2). ERFs were significantly different over the averaged time-course with specificity
384
dependent on the sensory modality targeted by the stimulation. Auditory stimulations reported
385
as detected elicit significant differences compared to undetected trials first between 190 and
386
210 ms, then between 250 and 425ms and finally between 460 and 500 ms after stimulus
387
onset (Figure 2A – left panel). Visual stimulation reported as detected elicits a large increase
388
of ERF amplitude compared to undetected trials from 230-250ms and from 310-500 ms after
389
stimulus onset (Figure 2A – middle panel). Tactile stimulation reported as detected elicits an
390
early increase of ERF amplitude between 95 and 150 ms then a later activation between 190
391
and 425 ms after stimulus onset (Figure 2A – right panel). Source localization of these specific
392
time periods of interest were performed for each modality (Figure 2B). The auditory condition
393
shows significant early source activity mainly localized to bilateral auditory cortices, superior
394
temporal sulcus and right inferior frontal gyrus, whereas the late significant component was
395
mainly localized to right temporal gyrus, bilateral precentral gyrus, left inferior and middle
396
frontal gyrus. A large activation can be observed for the visual conditions including primary
397
visual areas, fusiform and calcarine sulcus and a large fronto-parietal network activation
398
including bilateral inferior frontal gyrus, inferior parietal sulcus and cingulate cortex. The early
399
contrast of tactile evoked response shows a large difference in the brain activation including
400
primary and secondary somatosensory areas, but also a large involvement of right frontal
401
activity. The late contrast of tactile evoked response presents brain activation including left
402
16
frontal gyrus, left inferior parietal gyrus, bilateral temporal gyrus and supplementary motor
403
area.
404
405
406
Figure 2. NT trials event-related responses for different sensory modalities: auditory (left panel), tactile
407
(middle panel) and visual (right panel). (A) Source-level absolute value (baseline-corrected for visualization
408
purpose) of group event-related average (solid line) and standard error of the mean (shaded area) in the detected
409
(red) and undetected (blue) condition for all brain sources. Significant time windows are marked with bottom solid
410
lines (black line: pBonferroni-corrected < 0.05) for the contrast detected vs. undetected trials. The relative source
411
localization maps are represented in part B for the averaged time period. (B) Source reconstruction of the significant
412
time period marked in part A for the contrast detected vs. undetected trials, masked at pcluster-corrected < 0.05.
413
414
Decoding and multivariate searchlight analysis across time and brain regions
415
We investigated the generalization of brain activation over time within and between the
416
different sensory modalities. To this end, we performed a multivariate analysis of
417
reconstructed brain source-level activity from the initial experiment. Time generalization
418
analysis presented as a time-by-time matrix between 0 and 500 ms after stimulus onset shows
419
significant decoding accuracy for each condition (Figure 3A). As can be seen on the black
420
cells located on the diagonal in Figure 3A, cross-validation decoding was performed within the
421
same sensory modality. However, off-diagonal red cells of Figure 3A represent decoding
422
17
analysis between different sensory modality. Inside each cell, data reported along the diagonal
423
(dashed line) reveal average classifiers accuracy for a specific time point used for the training
424
and testing procedure, whereas off-diagonal data reveal a potential classifier ability to
425
generalize decoding based on different training and testing time points procedure. Indeed, we
426
observed the ability of the same classifier trained on a specific time point to generalize its
427
decoding performance over several time points (see off-diagonal significant decoding inside
428
each cell of Figure 3A). In order to appreciate this result, we computed the average duration
429
of significant decoding on testing time points based on the different training time points (Figure
430
3B). On average, decoding within the same modality, the classifier generalization starts after
431
200 ms and we observed significant maximum classification accuracy after 400 ms (see Figure
432
3B - top panel).
433
Early differences specific to the tactile modality have been grasped by the classification
434
analysis by showing significant decoding accuracy already after 100 ms without strong time
435
generalization for this sensory modality, where auditory and visual conditions show only
436
significant decoding starting around 250-300 ms after stimulus onset. Such an early dynamic
437
specific to the tactile modality could explain off-diagonal accuracy for all between modalities
438
decoding where the tactile modality was involved (Figure 3A). Interestingly, time generalization
439
analysis concerning between sensory modality decoding (red cells in Figure 3A) revealed
440
significant maximal generalization at around 400 ms (see Figure 3B - bottom panel). In
441
general, the time-generalization analysis revealed time-clusters restricted to late brain activity
442
with maximal decoding accuracy on average after 300 ms for all conditions. The similarity of
443
this time-cluster over all three sensory modalities suggests the generality of such brain
444
activation.
445
Restricted to the respective significant time clusters (Figure 3A), we investigated the
446
underlying brain sources resulting from the searchlight analysis within and between conditions
447
(Figure 4). The decoding within the same sensory modality revealed higher significant
448
accuracy in relevant sensory cortex for each specific modality condition (see Figure 4; brain
449
plots on diagonal). In addition, auditory modality searchlight decoding revealed also a strong
450
18
involvement of visual cortices (Figure 4: first row, first column), while somatosensory modality
451
decoding revealed parietal regions involvement such as precuneus (Figure 4: third row, third
452
column). However, decoding searchlight analysis between different sensory modalities
453
revealed higher decoding accuracy in fronto-parietal brain regions in addition to diverse
454
primary sensory regions (see Figure 4; brain plots off diagonal).
455
456
Figure 3. Time-by-time generalization analysis within and between sensory modality (for NT trials). 3x3
457
matrices of decoding results represented over time (from stimulation onset to 500 ms after). (A) Each cell presents
458
the result of the searchlight MVPA with time-by-time generalization analysis where classifier accuracy was
459
significantly above chance level (50%) (masked at pcorrected<0.005). For each temporal generalization matrix, a
460
classifier was trained at a specific time sample (vertical axis: training time) and tested on all time samples (horizontal
461
axis: testing time). The black dotted line corresponds to the diagonal of the temporal generalization matrix, i.e., a
462
classifier trained and tested on the same time sample. This procedure was applied for each combination of sensory
463
modality, i.e. presented on the first row is decoding analysis performed by classifiers trained on the auditory
464
modality and tested on auditory, visual or tactile (1st, 2nd and 3rd column respectively) for the two classes: detected
465
and undetected trials. The cells contoured with black line axes (on the diagonal) correspond to within the same
466
sensory modality decoding, whereas the cells contoured with red line axes correspond to between different
467
modalities decoding. (B) Summary of average time-generalization and decoding performance over time for all
468
within modality analysis (top panel: average based on the 3 black cells of part A) and between modalities analysis
469
(bottom panel: average based on the 6 red cells of part A). For each specific training time point on the x-axis the
470
average duration of classifier’s ability to significantly generalize on testing time points was computed and reported
471
19
on the y-axis. Additionally, normalized average significant classifiers accuracies over all testing time for a specific
472
training time point is represented as a color scale gradient.
473
474
475
Figure 4. Spatial distribution of significant searchlight MVPA decoding within and between sensory
476
modality. Source brain maps for average decoding accuracy restricted to the related time-generalization significant
477
time-by-time cluster (cf. Figure 3A). Brain maps were thresholded by only showing 10% maximum significant
478
decoding accuracy for each respective time-by-time cluster. Dark solid lines separate all between sensory modality
479
decoding brain maps from the cross-validation within one sensory modality decoding analysis on the diagonal.
480
481
Decoding and multivariate searchlight analysis over all sensory modalities
482
We further investigated the decoding generalizability of brain activity patterns across
483
all sensory modalities in one analysis by decoding detected versus undetected trials over all
484
blocks together (Figure 5A). Initially, we performed this specific analysis with data from the
485
first experiment and separately with data from the control experiment in order to replicate our
486
findings and control for potential motor response bias (see SI Appendix Figure S3). By
487
20
delaying the response-mapping to after the stimulus presentation in a random fashion during
488
the control experiment, neural patterns during relevant periods putatively cannot be
489
confounded by response selection / preparation. Importantly, analysis performed on the
490
control experiment used identical data in SI Appendix figure S3 B and C, but only trials
491
assignation (i.e. 2 classes definition) for decoding was different: “detected versus undetected”
492
(SI Appendix, Figure S3B) or “response versus no response” (SI Appendix, Figure S3C). Only
493
decoding of conscious report (i.e. “detected versus undetected”) showed significant time-by-
494
time clusters (SI Appendix, Figure S3 A&B). This result rules out a confounding influence of
495
the motor report and again strongly suggests the existence of a common supramodal pattern
496
related to conscious perception.
497
We investigated the similarity of time-generalization results by merging data from both
498
experiments (see Figure 5A). We tested for significant temporal dynamics of brain activity
499
patterns across all our data, taking into account that less stable or similar patterns would not
500
survive group statistics. Overall the ability for one classifier to generalize across time seems
501
to increase linearly after a critical time point around 100ms. We show that whereas the early
502
patterns (<250ms) are rather short-lived, temporal generalizability increases showing stability
503
values after ~350ms (Figure 5B). To follow-up on potential generators underlying these
504
temporal patterns, we depicted the searchlight results from three specific time-windows (W1,
505
W2 and W3) regarding the time-generalization decoding and the distribution of normalized
506
accuracy over time (Figure 5C). W1 from stimulation onset to 250ms depicts the first significant
507
searchlight decoding found in this analysis; W2 from 250ms to 350ms depicts the first
508
generalization period where decoding accuracy is low; finally W3 from 350ms to 500ms
509
depicts the second time-generalization period where higher decoding accuracy were found
510
(Figure 5B). The depiction of the results highlights precuneus, insula, anterior cingulate cortex,
511
frontal and parietal regions mainly involved during the first significant time-window (W1), while
512
the second time-window (W2) main significant cluster is located over left precentral motor
513
cortices. Interestingly the late time-window (W3) shows stronger decoding over primary
514
sensory cortices where accuracy are the highest: lingual and calcarine sulcus, superior
515
21
temporal and Heschl gyrus and right postcentral gyrus (Figure 5C). The sources depicted by
516
the searchlight analysis, suggest strong overlaps with functional brain networks related to
517
attention and saliency detection (29), especially during the earliest time periods (W1 and W2)
518
(see SI Appendix, Figure S4).
519
520
521
522
Figure 5. Time-by-time generalization and brain searchlight decoding analysis across all sensory
523
modalities (for NT trials). Compiled results for both initial and control experiments. (A) Decoding results
524
represented over time (from stimulation onset to 500 ms after. Result of the searchlight MVPA with time-by-time
525
generalization analysis of “detected” versus “undetected” trials across all sensory modalities. Figure shows the
526
time-clusters where classifier accuracy was significantly above chance level (50%) (masked at pcorrected<0.005).
527
The black dotted line corresponds to the diagonal of the temporal generalization matrix, i.e., a classifier trained and
528
tested on the same time sample. Horizontal black lines separate time windows (W1, W2 and W3) (B) Summary of
529
average time-generalization and decoding performance over time (A). For each specific training time point on the
530
x-axis the average duration of classifier’s ability to significantly generalize on testing time points was computed and
531
reported on the y-axis. Additionally, normalized average significant classifiers accuracies over all testing time for a
532
22
specific training time point is represented as a color scale gradient. Based on this summary three time windows
533
were depicted to explore spatial distribution of searchlight decoding (W1 : [0 250]ms ; W2 : [250 350]ms ; W3 : [350
534
500]ms). (C) Spatial distribution of significant searchlight MVPA decoding for the significant time clusters depicted
535
in (A) and (B). Brain maps were thresholded by only showing 10% maximum significant (pcorrected<0.005) decoding
536
accuracy for each respective time-by-time cluster.
537
538
539
Discussion
540
For a neural process to be a strong contender as a neural correlate of consciousness,
541
it should show some generalization e.g. across sensory modalities. This has –despite being
542
implicitly assumed- never been directly tested. To pursue this important issue, we investigated
543
a standard NT experiment targeting three different sensory modalities in order to explore
544
common spatio-temporal brain activity related to conscious perception using multivariate and
545
searchlight analysis. Our findings focusing on the post-stimulus evoked responses are in line
546
with previous studies for each specific sensory modality, showing stronger brain activation
547
when the stimulation was reported as perceived (27, 28, 30). Importantly by exploiting the
548
advantages of decoding, we provide for the first time direct evidence of common
549
electrophysiological correlates of conscious access across sensory modalities.
550
551
ERF time-course differences across sensory modalities
552
Our first results suggest significant temporal and spatial differences when univariate
553
contrast between ‘detected’ and ‘undetected’ trials were used to investigate sensory-specific
554
evoked responses. At the source level, the global group average activity revealed different
555
significant time periods according to the sensory modality targeted where modulations of
556
evoked responses related to detected trials can be observed (Figure 2A). In the auditory and
557
visual modalities, we found mainly significant differences after 200 ms. In the auditory domain,
558
perception- and attention-modulated sustained responses around 200 ms from sound onset
559
were found in bilateral auditory and frontal regions using MEG (31, 32). Using MEG, a previous
560
23
study confirmed awareness-related effects from 240 to 500 ms after target presentation during
561
visual presentation (33).
562
Our results show early differences in the transient responses (for the contrast detected
563
versus undetected) for the somatosensory domain compared to the other sensory modalities,
564
and have been previously identified using EEG at around 100 and 200 ms (34). Moreover,
565
previous MEG studies have shown early brain signal amplitude modulation (<200ms) related
566
to tactile perception in NT tasks (28, 35, 36). Such differences are less pronounced regarding
567
the contrast between catch and sham trials across sensory modality (see SI Appendix Figure
568
S1). Early ERF difference for the tactile NT trials can be due to the experimental setup where
569
auditory and visual targets stimulation emerged from a background stimulation (constant white
570
noise and screen display) whereas tactile stimuli remain isolated transient sensory targets.
571
Despite these differences the time generalization analysis was able to grasp similar brain
572
activity occurring at different time scale across these three sensory modalities.
573
Source localizations performed with univariate contrasts for each sensory modality
574
suggest differences in network activation with some involvement of similar brain regions in late
575
time windows such as: inferior frontal gyrus, inferior parietal gyrus and supplementary motor
576
area. However, qualitatively similar topographic patterns observed in such analysis cannot
577
easily be interpreted as similar brain processes. The important question is whether these
578
neural activity patterns within a specific sensory modality can be used to decode subjective
579
report of the stimulation within a different sensory context. The multivariate decoding analysis
580
we performed in the next analysis aimed to answer this question.
581
582
Identification of common brain activity across sensory modalities
583
Multivariate decoding analysis was used to refine spatio-temporal similarity across
584
these different sensory systems. In general, stable characteristics of brain signals have been
585
proposed as a transient stabilization of distributed cortical networks involved in conscious
586
perception (37). Using the precise time resolution of MEG signal and time-generalization
587
analysis, we investigated the stability and time dynamics of brain activity related to conscious
588
24
perception across sensory systems. The presence of similar brain activity can be revealed
589
between modalities using such a technique, even if significant ERF modulation is distributed
590
over time. As expected, between-modality time-generalization analysis involving tactile runs
591
show off-diagonal significant decoding due to early significant brain activity for the tactile
592
modality (Figure 3A). This result suggests the existence of early but similar brain activity
593
patterns related to conscious perception in the tactile domain compared to auditory and visual
594
modalities.
595
Generally, decoding results revealed a significant time cluster starting around 300 ms
596
with high classifier accuracy that speaks in favor of a late neural response related to conscious
597
report. Actually, we observed the ability of the same classifier trained on specific time points
598
with a specific sensory modality condition to generalize its decoding performance over several
599
time points with the same or another sensory modality. This result speaks in favor of
600
supramodal brain activity patterns that are consistent and stable over time. In addition, the
601
searchlight analysis across brain regions provides an attempt to depict brain network
602
activation during these significant time-generalization clusters. Note that, as seen also in
603
multiple other studies using decoding (22, 23, 38, 39), the average accuracy can be relatively
604
low and yet remains significant at the group level. Note however that contrary to many other
605
cognitive neuroscientific studies using decoding (39, 40), we do not apply the practice of
606
"subaveraging" trials to create "pseudo"-single trials, which naturally boosts average decoding
607
accuracy (41). Also, the statistical rigor of our approach is underlined by the fact that the
608
reported decoding results are restricted to highly significant effects (Pcorrected<0.005; see
609
Methods section). Critically, we replicated our results -applying the identical very conservative
610
statistical thresholds- within a second control experiment when looking at conscious
611
perception report contrast independently from motor response activity (SI Appendix, Figure
612
S3). Our results conform to those of previous studies in underlying the importance of late
613
activity patterns as crucial markers of conscious access (7, 42) and decision-making
614
processes (10, 43).
615
25
Furthermore in this study, we explored the brain regions underlying time dynamics of
616
conscious report by using brain source searchlight decoding. Knowing the limitations of such
617
MEG analysis, especially using low spatial resolution (3cm), we restricted depiction of results
618
to the main 10% maximum decoding accuracy over all searchlight brain regions. Some of the
619
brain regions found in our searchlight analysis, namely deep brain structures such as the
620
insula and anterior cingulate cortex are shared with other functional brain networks such as
621
the salience network (44, 45). Also the superior frontal and parietal cortex have been
622
previously found to be activated by attention-demanding cognitive tasks (46). Hence, we would
623
like to emphasize that one cannot conclude from our study that the observed network identified
624
in figure 5C is exclusively devoted to conscious report. Brain networks identified in this study
625
share common brain regions and dynamics with the attentional and salience networks that
626
remain relevant mechanisms to performing a NT-task. Interestingly this part of the network
627
seems to be more involved during the initial part of the process, prior to motor brain region
628
involvement (Figure 5C and SI Appendix Figure S4).
629
Indeed, some brain regions involved in motor planning were identified with our analysis,
630
such as precentral gyrus, and could in principle relate to the upcoming button-press to report
631
the subjective perception of the stimulus. We specifically targeted such motor preparation bias
632
within the control experiment, in which the participant was unable to predict a priori how to
633
report a conscious percept (i.e. pressing or withholding a button press) until the response
634
prompt appeared. Importantly, we did not find any significant decoding when trials used for
635
the analysis where sorted under response type (e.g. with or without an actual button press
636
from the participant) compared to subjective report of detection (see SI Appendix, Figure S3
637
B and C). Such findings could speak in favor of generic motor planning (47) or decision
638
processes related activity in such forced-choice paradigms (48, 49).
639
640
Late involvement of all primary sensory cortices
641
Some within-modalities decoding results highlighted unspecific primary cortices
642
involvement while decoding was performed on another sensory modality. For instance, during
643
26
auditory near-threshold stimulation, the main decoding accuracy of neural activity predicting
644
conscious perception was found in auditory cortices but also in visual cortices (see Figure 4:
645
first row, first column). Interestingly, our final analysis revealed and confirmed that primary
646
sensory regions are strongly involved in decoding conscious perception across sensory
647
modalities. Moreover, such brain regions were mainly found during the last time period
648
investigated following the first main involvement of fronto-parietal areas (see Figure 5). These
649
important results suggest that sensory cortices from a specific modality contain sufficient
650
information to allow the decoding perceptual conscious access in another different sensory
651
modality. These results suggest a late active role of primary cortices over three different
652
sensory systems (Figure 5). One study reported efficient decoding of visual object categories
653
in early somatosensory cortex using fMRI and multivariate pattern analysis (50). Another fMRI
654
experiment suggested that sensory cortices appear to be modulated via a common
655
supramodal frontoparietal network, attesting to the generality of attentional mechanism toward
656
expected auditory, tactile and visual information (51). However, in our study we demonstrate
657
how local brain activity from different sensory regions reveal a specific dynamic allowing
658
generalization over time to decode the behavioral outcome of a subjective perception in
659
another sensory modality. These results speak in favor of intimate cross-modal interactions
660
between modalities in perception (52).
661
Finally, our results suggest that primary sensory regions remain important at late
662
latency after stimulus onset for resolving stimulus perception over different sensory modalities.
663
We propose that this network could enhance the processing of behaviorally relevant signals,
664
here the sensory targets. Although the integration of classically unimodal primary sensory
665
cortices into a processing hierarchy of sensory information is well established (53), some
666
studies suggest multisensory roles of primary cortical areas (54, 55).
667
Today it remains unknown how such multisensory responses could be related to an
668
individual’s unisensory conscious percepts in humans. Since sensory modalities are usually
669
interwoven in real life, our findings of a supramodal network that may subserve both conscious
670
27
access and attentional functions have a higher ecological validity than results from previous
671
studies on conscious perception for single sensory modality.
672
Actually, our results are in line with an ongoing debate in neuroscience asking to what
673
extent multisensory integration emerges already in primary sensory areas (55, 56). Animal
674
studies provided compelling evidence suggesting that the neocortex is essentially
675
multisensory (57). Here our findings speak in favor of a multisensory interaction in primary and
676
associative cortices. Interestingly a previous an fMRI study by using multivariate decoding
677
revealed distinct mechanisms governing audiovisual integration in primary and associative
678
cortices needed for spatial orienting and interactions in a multisensory world (58).
679
680
Conclusion
681
We successfully characterized common patterns over time and space suggesting
682
generalization of consciousness-related brain activity across different sensory NT tasks. Our
683
study paves the way for future investigation using techniques with more precise spatial
684
resolution such as functional magnetic resonance imaging to depict in detail the brain network
685
involved. However, to our knowledge this is the first study to report significant spatio-temporal
686
decoding across different sensory modalities near-threshold perception experiment. Indeed,
687
our results speak in favor of the existence of stable and supramodal brain activity patterns,
688
distributed over time and involving seemingly task-unrelated primary sensory cortices. The
689
stability of brain activity patterns over different sensory modalities presented in this study is,
690
to date, the most direct evidence of a common network activation leading to conscious access
691
(2). Moreover, our findings add to recent remarkable demonstrations of applying decoding and
692
time generalization methods to MEG (21–23, 59), and show a promising application of MVPA
693
techniques to source level searchlight analysis with a focus on the temporal dynamics of
694
conscious perception.
695
696
697
28
Acknowledgements
698
This work was supported by the European Research Council (WIN2CON, ERC StG 283404).
699
We thank Julia Frey for her great support during data collection.
700
701
Author contributions
702
G.S. and N.W. conceived the approach. G.S., G.P. and T.H. implemented the experiment.
703
G.S. and M.F. collected the data. G.S. analyzed the data. G.S. and N.W. wrote the manuscript.
704
All authors approved the current manuscript.
705
706
707
Resource sharing and data availability
708
Further information and requests for resources or data should be directed to and will be fulfilled
709
by the corresponding author.
710
711
References
712
1.
Crick F, Koch C (2003) A framework for consciousness. Nat Neurosci 6(2):119–126.
713
2.
Dehaene S, Changeux J-P (2011) Experimental and Theoretical Approaches to
714
Conscious Processing. Neuron 70(2):200–227.
715
3.
Naghavi HR, Nyberg L (2005) Common fronto-parietal activity in attention, memory,
716
and consciousness: Shared demands on integration? Consciousness and Cognition
717
14(2):390–425.
718
4.
Foley JM, Legge GE (1981) Contrast detection and near-threshold discrimination in
719
human vision. Vision Research 21(7):1041–1053.
720
5.
Ruhnau P, Hauswald A, Weisz N (2014) Investigating ongoing brain oscillations and
721
their influence on conscious perception – network states and the window to
722
consciousness. Front Psychol 5:1230.
723
6.
Lamme VAF (2006) Towards a true neural stance on consciousness. Trends in
724
Cognitive Sciences 10(11):494–501.
725
7.
Dehaene S, Changeux J-P, Naccache L, Sackur J, Sergent C (2006) Conscious,
726
preconscious, and subliminal processing: a testable taxonomy. Trends in Cognitive
727
Sciences 10(5):204–211.
728
8.
Andersen LM, Pedersen MN, Sandberg K, Overgaard M (2016) Occipital MEG Activity
729
in the Early Time Range (<300 ms) Predicts Graded Changes in Perceptual
730
Consciousness. Cereb Cortex 26(6):2677–2688.
731
9.
Brancucci A, Lugli V, Perrucci MG, Gratta CD, Tommasi L (2014) A frontal but not
732
parietal neural correlate of auditory consciousness. Brain Struct Funct 221(1):463–472.
733
10. Joos K, Gilles A, Van de Heyning P, De Ridder D, Vanneste S (2014) From sensation
734
29
to percept: The neural signature of auditory event-related potentials. Neuroscience &
735
Biobehavioral Reviews 42:148–156.
736
11. Auksztulewicz R, Spitzer B, Blankenburg F (2012) Recurrent Neural Processing and
737
Somatosensory Awareness. J Neurosci 32(3):799–805.
738
12. Auksztulewicz R, Blankenburg F (2013) Subjective Rating of Weak Tactile Stimuli Is
739
Parametrically Encoded in Event-Related Potentials. J Neurosci 33(29):11878–11887.
740
13. Tallon-Baudry C (2012) On the Neural Mechanisms Subserving Consciousness and
741
Attention. Front Psychol 2:397.
742
14. van Gaal S, Lamme VAF (2012) Unconscious High-Level Information Processing:
743
Implication for Neurobiological Theories of Consciousness. The Neuroscientist
744
18(3):287–301.
745
15. Baars BJ (2005) Global workspace theory of consciousness: toward a cognitive
746
neuroscience of human experience. Progress in Brain Research (Elsevier), pp 45–53.
747
16. Dehaene S, Charles L, King J-R, Marti S (2014) Toward a computational theory of
748
conscious processing. Current Opinion in Neurobiology 25:76–84.
749
17. Sergent C, Baillet S, Dehaene S (2005) Timing of the brain events underlying access to
750
consciousness during the attentional blink. Nat Neurosci 8(10):1391–1400.
751
18. Fisch L, et al. (2009) Neural “Ignition”: Enhanced Activation Linked to Perceptual
752
Awareness in Human Ventral Stream Visual Cortex. Neuron 64(4):562–574.
753
19. Melloni L, Schwiedrzik CM, Müller N, Rodriguez E, Singer W (2011) Expectations
754
Change the Signatures and Timing of Electrophysiological Correlates of Perceptual
755
Awareness. J Neurosci 31(4):1386–1396.
756
20. Cichy RM, Pantazis D, Oliva A (2014) Resolving human object recognition in space and
757
time. Nat Neurosci 17(3):455–462.
758
21. King J-R, Dehaene S (2014) Characterizing the dynamics of mental representations:
759
the temporal generalization method. Trends in Cognitive Sciences 18(4):203–210.
760
22. Tucciarelli R, Turella L, Oosterhof NN, Weisz N, Lingnau A (2015) MEG Multivariate
761
Analysis Reveals Early Abstract Action Representations in the Lateral Occipitotemporal
762
Cortex. J Neurosci 35(49):16034–16045.
763
23. Wutz A, Muschter E, van Koningsbruggen MG, Weisz N, Melcher D (2016) Temporal
764
Integration Windows in Neural Processing and Perception Aligned to Saccadic Eye
765
Movements. Current Biology 26(13):1659–1668.
766
24. Kriegeskorte N, Goebel R, Bandettini P (2006) Information-based functional brain
767
mapping. PNAS 103(10):3863–3868.
768
25. Haynes J-D, et al. (2007) Reading Hidden Intentions in the Human Brain. Current
769
Biology 17(4):323–328.
770
26. Sandberg K, Andersen LM, Overgaard M (2014) Using multivariate decoding to go
771
beyond contrastive analyses in consciousness research. Front Psychol 5:1250.
772
27. Leske S, et al. (2015) Prestimulus Network Integration of Auditory Cortex Predisposes
773
Near-Threshold Perception Independently of Local Excitability. Cereb Cortex
774
25(12):4898–4907.
775
28. Frey JN, et al. (2016) The Tactile Window to Consciousness is Characterized by
776
Frequency-Specific Integration and Segregation of the Primary Somatosensory Cortex.
777
Scientific Reports 6:20805.
778
29. Gordon EM, et al. (2016) Generation and Evaluation of a Cortical Area Parcellation
779
from Resting-State Correlations. Cereb Cortex 26(1):288–303.
780
30. Karns CM, Knight RT (2008) Intermodal Auditory, Visual, and Tactile Attention
781
Modulates Early Stages of Neural Processing. Journal of Cognitive Neuroscience
782
21(4):669–683.
783
31. Kauramäki J, et al. (2012) Two-Stage Processing of Sounds Explains Behavioral
784
Performance Variations due to Changes in Stimulus Contrast and Selective Attention:
785
An MEG Study. PLOS ONE 7(10):e46872.
786
32. Zoefel B, Heil P (2013) Detection of Near-Threshold Sounds is Independent of EEG
787
Phase in Common Frequency Bands. Front Psychol 4:262.
788
33. Wyart V, Tallon-Baudry C (2008) Neural Dissociation between Visual Awareness and
789
30
Spatial Attention. J Neurosci 28(10):2667–2679.
790
34. Ai L, Ro T (2014) The phase of prestimulus alpha oscillations affects tactile perception.
791
Journal of Neurophysiology 111(6):1300–1307.
792
35. Palva S, Linkenkaer-Hansen K, Näätänen R, Palva JM (2005) Early Neural Correlates
793
of Conscious Somatosensory Perception. J Neurosci 25(21):5248–5258.
794
36. Wühle A, Mertiens L, Rüter J, Ostwald D, Braun C (2010) Cortical processing of near-
795
threshold tactile stimuli: An MEG study. Psychophysiology 47(3):523–534.
796
37. Schurger A, Sarigiannidis I, Naccache L, Sitt JD, Dehaene S (2015) Cortical activity is
797
more stable when sensory stimuli are consciously perceived. PNAS 112(16):E2083–
798
E2092.
799
38. Turella L, et al. (2016) Beta band modulations underlie action representations for
800
movement planning. NeuroImage 136:197–207.
801
39. Cichy RM, Khosla A, Pantazis D, Oliva A (2017) Dynamics of scene representations in
802
the human brain revealed by magnetoencephalography and deep neural networks.
803
NeuroImage 153:346–358.
804
40. Kaiser D, Oosterhof NN, Peelen MV (2016) The Neural Dynamics of Attentional
805
Selection in Natural Scenes. J Neurosci 36(41):10522–10528.
806
41. Grootswagers T, Wardle SG, Carlson TA (2016) Decoding Dynamic Brain Patterns
807
from Evoked Responses: A Tutorial on Multivariate Pattern Analysis Applied to Time
808
Series Neuroimaging Data. Journal of Cognitive Neuroscience 29(4):677–697.
809
42. Sergent C, Dehaene S (2004) Neural processes underlying conscious perception:
810
Experimental findings and a global neuronal workspace framework. Journal of
811
Physiology-Paris 98(4–6):374–384.
812
43. Polich J (2007) Updating P300: An integrative theory of P3a and P3b. Clinical
813
Neurophysiology 118(10):2128–2148.
814
44. Kucyi A, Hodaie M, Davis KD (2012) Lateralization in intrinsic functional connectivity of
815
the temporoparietal junction with salience- and attention-related brain networks. Journal
816
of Neurophysiology 108(12):3382–3392.
817
45. Chen T, Cai W, Ryali S, Supekar K, Menon V (2016) Distinct Global Brain Dynamics
818
and Spatiotemporal Organization of the Salience Network. PLOS Biol 14(6):e1002469.
819
46. Menon V, Uddin LQ (2010) Saliency, switching, attention and control: a network model
820
of insula function. Brain Struct Funct 214(5–6):655–667.
821
47. Rubia K, et al. (2001) Mapping Motor Inhibition: Conjunctive Brain Activations across
822
Different Versions of Go/No-Go and Stop Tasks. NeuroImage 13(2):250–261.
823
48. Donner TH, Siegel M, Fries P, Engel AK (2009) Buildup of Choice-Predictive Activity in
824
Human Motor Cortex during Perceptual Decision Making. Current Biology 19(18):1581–
825
1585.
826
49. Mostert P, Kok P, de Lange FP (2015) Dissociating sensory from decision processes in
827
human perceptual decision making. Scientific Reports 5:18253.
828
50. Smith FW, Goodale MA (2015) Decoding Visual Object Categories in Early
829
Somatosensory Cortex. Cereb Cortex 25(4):1020–1031.
830
51. Langner R, et al. (2011) Modality-Specific Perceptual Expectations Selectively
831
Modulate Baseline Activity in Auditory, Somatosensory, and Visual Cortices. Cereb
832
Cortex 21(12):2850–2862.
833
52. Pooresmaeili A, et al. (2014) Cross-modal effects of value on perceptual acuity and
834
stimulus encoding. PNAS 111(42):15244–15249.
835
53. Felleman DJ, Essen DCV (1991) Distributed Hierarchical Processing in the Primate
836
Cerebral Cortex. Cereb Cortex 1(1):1–47.
837
54. Lemus L, Hernández A, Luna R, Zainos A, Romo R (2010) Do Sensory Cortices
838
Process More than One Sensory Modality during Perceptual Judgments? Neuron
839
67(2):335–348.
840
55. Liang M, Mouraux A, Hu L, Iannetti GD (2013) Primary sensory cortices contain
841
distinguishable spatial patterns of activity for each sense. Nature Communications
842
4:1979.
843
56. Vetter P, Smith FW, Muckli L (2014) Decoding Sound and Imagery Content in Early
844
31
Visual Cortex. Current Biology 24(11):1256–1262.
845
57. Ghazanfar AA, Schroeder CE (2006) Is neocortex essentially multisensory? Trends in
846
Cognitive Sciences 10(6):278–285.
847
58. Rohe T, Noppeney U (2016) Distinct Computational Principles Govern Multisensory
848
Integration in Primary Sensory and Association Cortices. Current Biology 26(4):509–
849
514.
850
59. King J-R, Pescetelli N, Dehaene S (2016) Brain Mechanisms Underlying the Brief
851
Maintenance of Seen and Unseen Sensory Information. Neuron 92(5):1122–1134.
852
60. Kontsevich LL, Tyler CW (1999) Bayesian adaptive estimation of psychometric slope
853
and threshold. Vision Research 39(16):2729–2737.
854
61. Brainard D (1997) The Psychophysics Toolbox. Spatial Vision:433–436.
855
62. Oostenveld R, et al. (2010) FieldTrip: Open Source Software for Advanced Analysis of
856
MEG, EEG, and Invasive Electrophysiological Data, FieldTrip: Open Source Software
857
for Advanced Analysis of MEG, EEG, and Invasive Electrophysiological Data.
858
Computational Intelligence and Neuroscience, Computational Intelligence and
859
Neuroscience 2011, 2011:e156869.
860
63. Oosterhof NN, Connolly AC, Haxby JV (2016) CoSMoMVPA: Multi-Modal Multivariate
861
Pattern Analysis of Neuroimaging Data in Matlab/GNU Octave. Front Neuroinform:27.
862
64. Gross J, et al. (2013) Good practice for conducting and reporting MEG research.
863
NeuroImage 65:349–363.
864
65. Veen BDV, Drongelen WV, Yuchtman M, Suzuki A (1997) Localization of brain
865
electrical activity via linearly constrained minimum variance spatial filtering. IEEE
866
Transactions on Biomedical Engineering 44(9):867–880.
867
66. Nolte G (2003) The magnetic lead field theorem in the quasi-static approximation and
868
its use for magnetoencephalography forward calculation in realistic volume conductors.
869
Phys Med Biol 48(22):3637.
870
67. Van Essen DC, et al. (2001) An Integrated Software Suite for Surface-based Analyses
871
of Cerebral Cortex. Journal of the American Medical Informatics Association 8(5):443–
872
459.
873
68. Tzourio-Mazoyer N, et al. (2002) Automated Anatomical Labeling of Activations in SPM
874
Using a Macroscopic Anatomical Parcellation of the MNI MRI Single-Subject Brain.
875
NeuroImage 15(1):273–289.
876
69. Maris E, Oostenveld R (2007) Nonparametric statistical testing of EEG- and MEG-data.
877
Journal of Neuroscience Methods 164(1):177–190.
878
70. Smith SM, Nichols TE (2009) Threshold-free cluster enhancement: Addressing
879
problems of smoothing, threshold dependence and localisation in cluster inference.
880
NeuroImage 44(1):83–98.
881
71. Pernet CR, Latinus M, Nichols TE, Rousselet GA (2015) Cluster-based computational
882
methods for mass univariate analyses of event-related brain potentials/fields: A
883
simulation study. Journal of Neuroscience Methods 250:85–93.
884
885
| 2019 | Decoding across sensory modalities reveals common supramodal signatures of conscious perception | 10.1101/115535 | [
"Sanchez Gaëtan",
"Hartmann Thomas",
"Fuscà Marco",
"Demarchi Gianpaolo",
"Weisz Nathan"
] | null |
1
Development of a LAG-3 Immunohistochemistry Assay for Melanoma
Lori Johnson, BS,* Bryan McCune, MD,* Darren Locke, PhD,†‡ Cyrus Hedvat, MD,
PhD,†‡ John B. Wojcik, MD, PhD,† Caitlin Schroyer, BS,* Jim Yan, PhD,*
Krystal Johnson, MD,* Angela Sanders-Cliette, MD,* Sujana Samala, MD,* Lloye M.
Dillon, PhD,† Steven Anderson, PhD,* and Jeffrey Shuster, PhD*
*Labcorp of America, Burlington, NC, USA; and †Bristol Myers Squibb, Princeton, NJ,
USA. ‡At the time the study was performed.
Corresponding author: Jeffrey Shuster, PhD, Companion Diagnostics, Labcorp Drug
Development, 100 Perimeter Park, Suite C, Morrisville, NC 27560, USA. Email:
jeffrey.shuster@labcorp.com. Phone: 919-388-5536
Current word count: 2921/3000 words
Current figure/table count: 4 figures, 4 tables. 4 supplemental figures, 3 supplemental
tables.
References: 32
2
Abstract (248/250)
Aims: A robust immunohistochemistry (IHC) assay was developed to detect
lymphocyte-activation gene 3 (LAG-3) expression by immune cells (ICs) in tumor
tissues. LAG-3 is an immuno-oncology target with demonstrable clinical benefit, and
there is a need for a standardized, well-characterized assay to measure its expression.
This study aims to describe LAG-3 scoring criteria and present the specificity,
sensitivity, analytical precision, and reproducibility of this assay.
Methods: The specificity of the assay was investigated by antigen competition and with
LAG3 knockout cell lines. A melanin pigment removal procedure was implemented to
prevent melanin interference in IHC interpretation. Formalin-fixed, paraffin-embedded
(FFPE) human melanoma samples with a range of LAG-3 expression levels were used
to assess the sensitivity and analytical precision of the assay with a ≥1% cutoff to
determine LAG-3–positivity. Interobserver and intraobserver reproducibility were
evaluated with 60 samples in intralaboratory studies and 70 samples in interlaboratory
studies.
Results: The LAG-3 IHC method demonstrated performance suitable for analysis of
LAG-3 IC expression in clinical melanoma samples. The pretreatment step effectively
removed melanin pigment that could interfere with interpretation. LAG-3 antigen
competition and analysis of LAG3 knockout cell lines indicated that the 17B4 antibody
clone binds specifically to LAG-3. The intrarun repeatability, interday, interinstrument,
interoperator, and interreagent lot reproducibility demonstrated a high scoring
concordance (>95%). The interobserver and intraobserver reproducibility and overall
3
interlaboratory and intralaboratory reproducibility also showed high scoring concordance
(>90%).
Conclusions: We have demonstrated that the assay reliably assesses LAG-3
expression in FFPE human melanoma samples by IHC.
Key Words: immunohistochemistry, melanoma, molecular pathology
4
Key messages (3-5 sentences maximum)
What is already known on this topic: Lymphocyte-activation gene 3 (LAG-3) is an
immune checkpoint receptor expressed on immune cells that limits T-cell activity and is
being actively explored as a target for immunotherapy.
What this study adds: An immunohistochemistry assay was developed to detect the
LAG-3 protein in formalin-fixed paraffin-embedded human tumor tissue specimens. This
study describes scoring criteria and shows the specificity, sensitivity, analytical
precision, and reproducibility of this assay as an aid to determine LAG-3 expression in
melanoma patients using a ≥1% expression on immune cells threshold.
How this study might affect research, practice or policy:
The study describes a key immuno-oncology checkpoint immunohistochemistry assay
that is robust and suitable for clinical trials. The assay was used in RELATIVITY-047
(NCT03470922), a phase 2/3 clinical trial that compared combined nivolumab and
relatlimab treatment with nivolumab monotherapy, to stratify patients based on the
percentage of LAG-3–positive immune cells within the tumor region. This assay is also
being used in several ongoing clinical trials evaluating clinical response to relatlimab.
5
INTRODUCTION
Immune checkpoint inhibitor–based therapies have greatly improved clinical outcomes
across multiple disease settings,[1, 2] including advanced melanoma,[3-5] non-small
cell lung cancer,[6, 7] squamous cell carcinoma of the head and neck,[8, 9] and
urothelial carcinoma,[10, 11] among others. However, given the multiple mechanisms of
immune evasion utilized by cancer cells, inhibition of a single immune checkpoint, such
as programmed death-1 (PD-1), may not be sufficient to overcome immune
suppression.[12, 13] Novel immuno-oncology (I-O) combinations, including dual
checkpoint inhibition, may be necessary to enhance efficacy and to improve the
durability of patient responses.
Lymphocyte-activation gene 3 (LAG-3, CD223) is a cell-surface molecule expressed on
activated CD4+ and CD8+ T cells, as well as other immune cells (ICs) including
regulatory T cells, natural killer cells, B cells, macrophages, and dendritic cells, and is
under investigation as an I-O therapy target.[13-17] The interaction of LAG-3 with its
ligands, the major histocompatibility complex II (MHCII), and fibrinogen-like protein 1
(FGL-1), recently discovered as a LAG-3 ligand, initiates an inhibitory signal.[13, 18, 19]
This signal can impair T-cell function, activation, and proliferation, decrease production
of and response to proinflammatory cytokines, and decrease the development of
memory T cells.
Preclinical data indicate that simultaneous activation of the LAG-3 and PD-1 pathways
in tumor-infiltrating lymphocytes results in greater T-cell exhaustion than either pathway
alone, and dual inhibition of these pathways may improve T-cell function and increase
immune response.[20] Furthermore, combined therapy with anti–LAG-3 and anti–PD-1
6
agents in fibrosarcoma and colorectal adenocarcinoma mouse models resulted in
synergistic antitumor activity.[16] The clinical efficacy of combining relatlimab, an anti–
LAG-3 antibody, with nivolumab, an anti–PD-1 agent, was previously demonstrated in
patients with previously untreated metastatic or unresectable melanoma by the
phase 2/3 RELATIVITY-047 clinical trial (NCT03470922).[21] RELATIVITY-047
demonstrated superior progression-free survival (PFS) for relatlimab combined with
nivolumab versus nivolumab monotherapy, regardless of LAG-3 expression.[21]
A robust immunohistochemistry (IHC) assay was developed to detect LAG-3 expression
by ICs. The assay was used to stratify patients enrolled in RELATIVITY-047, based on
the percentage of LAG-3–positive ICs with a morphological resemblance to
lymphocytes relative to all nucleated cells within the tumor region (tumor cells [TCs],
intratumoral stroma, and peritumoral stroma [the band of stromal elements directly
contiguous with the outer tumor margin]) in samples containing ≥100 viable TCs. This
assay is also being used in several ongoing clinical trials evaluating relatlimab. This
study presents the specificity, sensitivity, analytical precision, and reproducibility of this
assay as an aid to determine LAG-3 expression in melanoma patients using a ≥1% IC
expression threshold.
MATERIALS AND METHODS
Principles of the LAG-3 IHC assay
The LAG-3 IHC assay was developed using a mouse monoclonal antibody clone 17B4
that was made to a synthetic peptide corresponding to the 30–amino acid extra-loop of
the first immunoglobulin domain of LAG-3,
7
GPPAAAPGHPLAPGPHPAAPSSWGPRPRRY.[22] The assay was performed on
formalin-fixed paraffin-embedded (FFPE) tissue sections mounted on glass slides and
included pretreatment to remove endogenous melanin that could interfere with
interpretation of LAG-3 staining. Following pretreatment, slides were stained and
processed using the 17B4 primary antibody on a Leica BOND-III autostainer (Leica
Biosystems, Buffalo Grove, IL).
Materials
Tissue specimens
FFPE melanoma specimens and control tonsil tissues were obtained from commercial
vendors (Boca Biolistics, Pompano Beach, FL; BioIVT, Westbury, NY; and Avaden
Biosciences, Seattle, WA). Sections were cut from each tissue block at 4-µm thickness,
placed on positively charged slides, and dried for 1 hour at 60°C ± 2°C. Excepting
sample stability studies, all cut sections were tested within 2 months of sectioning.
Antibodies
All experiments were performed with monoclonal LAG-3 antibody 17B4 preparations
manufactured from hybridoma cultures for Labcorp, except for analysis of clustered
regularly interspaced short palindromic repeats (CRISPR)-engineered LAG-3 knockout
cell lines, for which a commercially available LAG-3 17B4 antibody was obtained from
LSBio (Cat. # LS-C18692) or as otherwise noted in the text.[22] For precision studies, 3
independent lots of antibody were produced from the 17B4 hybridoma. The working
concentration of the LAG-3 17B4 antibody was 2.5 µg/mL. The negative control
antibody, mouse monoclonal immunoglobulin G1 (IgG1) clone MOPC-21, was obtained
8
from Leica Biosystems (Cat. # PA0996). Further details on the staining and melanin
removal procedures are in the supplemental material and supplemental table 1.
Melanin scoring
To determine the efficacy of the melanin removal step of the protocol, the amount of
melanin pigment in the tumor region was scored on a scale of 0 to 4+. Definitions for
melanin pigment scoring expected on melanoma tissue–stained slides and indications
for the evaluability of the melanin interpretation in LAG-3 IHC assay scoring are
provided in supplemental table 2.
LAG-3 scoring
An overview of the LAG-3 scoring method is provided in supplemental figure 1.
Evaluation criteria for staining intensity of LAG-3–positive ICs consisted of weak (1+),
moderate (2+), and strong (3+) LAG-3–positive staining (supplemental table 3). In
addition to cell-surface expression, LAG-3 protein is also retained in intracellular
compartments.[23] Thus, LAG-3 IC positivity was quantified in cells that morphologically
resembled lymphocytes with punctate (perinuclear and/or Golgi pattern), cytoplasmic,
and/or membranous LAG-3 staining of any intensity above background (supplemental
figure 2). LAG-3–positive IC content in the tumor region was visually estimated by
microscopic examination by the study pathologists, following group alignment using a
reference slide set. A hematoxylin and eosin-stained slide for each melanoma sample
tested was reviewed by a pathologist to identify the overall tumor region and confirm the
presence of ≥100 TCs. Results were reported as the percentage of LAG-3–positive ICs
relative to all nucleated cells (ICs [lymphocytes and macrophages], stromal cells, and
TCs) within the overall tumor region. The tumor region included TCs, intratumoral
9
stroma, and peritumoral stroma (the band of stromal elements directly contiguous with
the outer tumor margin). Normal and/or adjacent uninvolved tissues were not included
(supplemental figure 3). The scoring scale was (in %) 0, 1, 2 ,3, 4, 5, 10, and further
increments of 10 up to 100. Samples with LAG-3–positive IC percentage scores of ≥1%
were reported as LAG-3–positive.
The methods for the generation of CRISPR-engineered LAG-3 knockout cell lines,
peptide inhibition assay, precision study measurements and reproducibility within the
same laboratory and across laboratories, and stability experiments are provided in the
supplemental material.
RESULTS
Components of the LAG-3 IHC assay
Primary antibody concentration and incubation times for assay components were
optimized for appropriate positive staining, staining intensity, and overall staining quality
of LAG-3 while minimizing nonspecific background staining. Antibody concentrations of
1.25 µg/mL, 2.5 µg/mL, 3.0 µg/mL, and 3.5 µg/mL were evaluated, and 2.5 µg/mL was
determined to be the optimal concentration.
Detection of LAG-3 in tissues using the 17B4 clone antibody
To investigate the ability of the LAG-3 IHC assay to detect LAG-3 IC expression in
human FFPE tissue samples, the assay was used to stain LAG-3 in commercially
procured human tonsil tissue. We hypothesized that if the LAG-3 IHC assay detected
LAG-3 IC expression, then staining would be present in lymphocytes, but not in
nonimmune regions, such as the crypt epithelium. Staining of the tonsil tissues using
10
the LAG-3 IHC assay revealed membranous/cytoplasmic staining of LAG-3 in
lymphocytes in germinal center and interfollicular regions, but no LAG-3 staining in the
crypt epithelium (figure 1A). Additionally, no staining was observed in the slide stained
with the mouse IgG isotype control.
The LAG-3 IHC assay was developed to include attenuation of melanin staining from
FFPE sections prior to IHC and to minimize the impact of melanin pigment on
interpretation of the assay. Examples of different levels of melanin pigmentation are
shown in supplemental figure 4. The efficacy of melanin removal from tissue samples
using the melanin removal procedure is shown in figures 1B and 1C. All melanoma
tissue samples selected for further investigation had acceptable negative control
staining and melanin pigmentation ≤1+. LAG-3 staining was consistent in bleached and
unbleached serial sections from the same tissue block (data not shown).
Specificity and sensitivity of the LAG-3 IHC assay
To investigate the specificity of the LAG-3 IHC assay, the LAG3 gene was disrupted by
CRISPR-mediated mutagenesis in COV434 cell lines. In total, 3 pooled cell lines were
derived, each with differing levels of LAG3 knockout (out-of-frame indel frequency =
71.02% in Cr1, 62.07% in Cr2, and 65.74% in Cr3) (figure 2A). The LAG-3 expression
of these cell lines was compared with parental COV434 cells to investigate the
specificity of the LAG-3 IHC assay. LAG-3 staining in parental COV434 cells was
markedly higher than each of the 3 LAG3 knockout cell lines, which each had staining
consistent with anticipated levels of residual LAG-3 expression based on the frequency
of alterations determined by next-generation sequencing (figure 2B). These data
11
suggest that the LAG-3 IHC assay is specific for the detection of LAG-3 protein
expression.
A peptide competition assay was performed using a synthetic LAG-3 peptide to further
investigate the specificity of the LAG-3 IHC assay. The percentage of LAG-3–positive
ICs in melanoma tissue was found to decrease from a starting staining level of 40% to
<1% following preincubation with increasing molar ratios of a LAG-3 peptide (table 1),
indicating that the LAG-3 peptide bound competitively to the 17B4 clone.
12
TABLE 1. LAG-3 IHC peptide competition validation results
Specimen
(peptide:antibody ratio)
% LAG-3–positive
ICs
Staining intensity
Melanoma LAG-3 mAb (0:1)
40
2+
Melanoma LAG-3 peptide (1:0)
0
N/A
Melanoma (1:1)
40
2+
Melanoma (2:1)
30
2+
Melanoma (5:1)
10–20
1–2+
Melanoma (10:1)
2
1+
Melanoma (30:1)
<1
1+
ICs, immune cells; IHC, immunohistochemistry; LAG-3, lymphocyte-activation gene 3;
mAb, monoclonal antibody; N/A, not applicable.
To determine the range of LAG-3 IC expression in melanoma specimens, 100
commercially procured melanoma samples were assessed using the LAG-3 IHC assay.
Of these 100 samples, 38 were positive for LAG-3 IC expression and 62 were negative,
using 1% expression as a cutoff value (figure 3). The range of IC expression in the
positive specimens was 1% to 40%, with a median of 3%. Of the positive cases, the
majority (36) had a LAG-3 IC staining intensity of 2+, 1 sample had a LAG-3 IC staining
intensity of 3+, and 1 sample had a LAG-3 IC staining intensity of 1+. Taken together,
13
these data indicate that the LAG-3 IHC assay detects varying levels of immune
infiltrates expressing LAG-3 in human FFPE melanoma samples. Figure 4 shows
representative tissue examples of staining from 0% to 30%.
Analytical precision of the LAG-3 IHC assay within the same laboratory
Twenty-four FFPE melanoma samples and 1 normal human tonsil tissue control sample
were stained on 2 different Leica BOND-III instruments and subsequently scored by 2
independent pathologists to establish the repeatability and reproducibility of the LAG-3
IHC assay. The intrarun repeatability, interday, interinstrument, interoperator, and
interreagent lot reproducibility all demonstrated a high concordance, with all point
estimates >95% in average negative agreement (ANA), average positive agreement
(APA), and overall percentage agreement (OPA) (table 2).
14
TABLE 2. Summary of precision study results
Evaluation
Percentage agreement (95% CI)
Intrarun repeatability
ANA: 98.5 (97.3–99.6)
APA: 98.6 (97.4–99.6)
OPA: 98.5 (97.3–99.6)
Interday reproducibility
ANA: 97.4 (96.4–98.4)
APA: 97.6 (96.6–98.5)
OPA: 97.5 (96.5–98.4)
Interinstrument reproducibility
ANA: 97.8 (96.8–98.6)
APA: 97.9 (97.0–98.7)
OPA: 97.8 (97.0–98.6)
Interoperator reproducibility
ANA: 97.8 (96.8–98.6)
APA: 97.9 (97.0–98.7)
OPA: 97.8 (96.9–98.7)
Interreagent lot reproducibility
ANA: 97.4 (96.6–98.2)
APA: 97.6 (96.8–98.3)
OPA: 97.5 (96.7–98.3)
ANA, average negative agreement; APA, average positive agreement; CI, confidence
interval; OPA, overall percentage agreement.
15
Interobserver and intraobserver reproducibility of the LAG-3 IHC assay within the same
laboratory
Evaluations of 60 melanoma samples performed by 3 independent pathologists from the
same laboratory and repeat evaluations of the same 60 melanoma samples by the
same pathologist were examined to determine the interobserver and intraobserver
reproducibility of the assay within the same laboratory. To determine the interobserver
reproducibility of the LAG-3 IHC assay, pairwise comparisons were made of the 180
diagnostic calls by the 3 pathologists: 91 were concordant for positive-to-positive calls,
and 77 were concordant for negative-to-negative calls. Disagreements occurred in 12
cases, all of which had LAG-3 scores around the 1% threshold (LAG-3–positive IC
content of 0%–1%), resulting in a lower point estimate and lower bound 95% confidence
interval (CI) for ANA compared with APA and OPA. Point estimates for ANA, APA, and
OPA were >90% with the lower bound 95% CIs >85% (table 3).
To determine intraobserver reproducibility of the LAG-3 IHC assay, the 60 samples
assessed in the interobserver reproducibility testing were reassessed by the same
pathologists, following a wash-out period. Among the 180 comparisons of diagnostic
calls between 2 reads by 3 pathologists, 89 were positive-to-positive concordant, 78
were negative-to-negative concordant, 8 were negative-to-positive discordant, and 5
were positive-to-negative discordant. Additionally, the point estimates and lower bound
95% CIs were >90% and >85%, respectively, in ANA, APA, and OPA (table 3).
16
TABLE 3. Percentage agreement and 95% CIs for interobserver and intraobserver
agreement within the same laboratory
Evaluation
Percentage agreement (95% CI)
Interobserver reproducibility
ANA: 92.8 (88.31–96.59)
APA: 93.8 (89.95–97.06)
OPA: 93.3 (89.44–96.66)
Intraobserver reproducibility
ANA: 92.31 (87.74–96.09)
APA: 93.19 (89.22–96.52)
OPA: 92.78 (88.89–96.11)
ANA, average negative agreement; APA, average positive agreement; CI, confidence
interval; OPA, overall percentage agreement.
Interlaboratory and intralaboratory reproducibility of the LAG-3 IHC assay
Two experiments were performed to assess interlaboratory reproducibility: interobserver
and intraobserver reproducibility, and overall interlaboratory and intralaboratory
reproducibility. First, to investigate the interobserver and intraobserver reproducibility of
the LAG-3 IHC assay between different laboratories, 70 melanoma LAG-3–prestained
cases were assessed by 3 pathologists at 3 separate laboratories. Second, to
determine overall interlaboratory and intralaboratory reproducibility, unstained slides
from 24 melanoma cases that had previously been shown to have a range of LAG-3
17
expression were tested at 3 separate laboratories. The interobserver and intraobserver
reproducibility and overall interlaboratory and intralaboratory reproducibility
demonstrated assay staining and scoring concordance with point estimates for all
studies at >90% in ANA, APA, and OPA and lower bound 95% CIs >85% (table 4).
18
TABLE 4. Percentage agreement and 95% CIs in the interlaboratory reproducibility
study
Evaluation
Percentage agreement (95% CI)
Intraobserver reproducibility
ANA: 92.1 (89.6–94.4)
APA: 94.2 (92.4–95.9)
OPA: 93.3 (91.3–95.2)
Interobserver reproducibility
ANA: 90.2 (88.7–91.7)
APA: 92.9 (91.7–94.0)
OPA: 91.8 (90.5–93.0)
Intralaboratory reproducibility
ANA: 95.1 (93.3–96.7)
APA: 96.0 (94.5–97.3)
OPA: 95.6 (94.0–97.1)
Interlaboratory reproducibility
ANA: 93.2 (91.9–94.4)
APA: 94.4 (93.4–95.5)
OPA: 93.9 (92.7–94.9)
ANA, average negative agreement; APA, average positive agreement; CI, confidence
interval; OPA, overall percentage agreement.
19
Slide stability experiments
To establish the stability of LAG-3 protein in unstained FFPE tissue sections on glass
slides for the LAG-3 IHC assay, the concordance of sectioned tissue samples stained
after different storage periods was measured. There was 100% concordance in scoring
(positive or negative) at all time points for slides stored at ambient temperatures or 2–
8°C. The LAG-3–positive IC staining intensity results for the tonsil tissue were 100%
concordant from baseline through month 18 at both 2–8°C and ambient temperatures,
with a decrease in LAG-3 IC staining intensity from 3+ to 2+ at month 24. Although
there was some slight variation (increase or decrease) in the percentage of LAG-3–
positive ICs for some melanoma samples during the course of testing (eg, a case
reported as 2% at week 2, 1% at week 4, and 2% at month 2), the LAG-3 score
(positive or negative) and LAG-3–positive IC staining intensity (1+, 2+, 3+) results were
100% concordant for individual samples tested at each time point and each
temperature. The small differences observed may be attributable to variations in the
density of ICs between tissue sections.
DISCUSSION
LAG-3 is a key immune checkpoint currently being investigated as an I-O therapy for
patients with solid tumors and hematological malignancies.[13, 16, 18, 21, 24-26] The
development of a robust LAG-3 IHC assay will enable the analysis of IC LAG-3 status in
the tumor microenvironment and the correlation between LAG-3 expression status and
response to LAG-3–directed oncology treatments. A robust LAG-3 IHC assay that is
suitable for clinical trials and clinical use for melanoma is described in this work. The
specificity of the assay was demonstrated using cell lines with LAG3 gene disruptions
20
and with a peptide antigen competition assay. LAG-3 scoring was reported as the
percentage of LAG-3–positive ICs (which morphologically resembled lymphocytes)
relative to all nucleated cells within the overall tumor region. A ≥1% cutoff was used to
determine LAG-3 positivity. Analytical precision was demonstrated for intrarun
repeatability, interday, interinstrument, interoperator, and interreagent lot reproducibility,
with concordance >95%. Pathologist interobserver and intraobserver reproducibility was
>90% in terms of ANA, APA, and OPA. LAG-3 was observed to be stable in unstained
tissues mounted on glass slides, with concordant staining observed in samples stored
at both 2–8°C and ambient temperatures for up to 24 months. These data demonstrate
that this assay can reproducibly determine the proportion of LAG-3–positive ICs within a
sample. Despite challenges associated with the scoring of ICs, the LAG-3 IHC assay
demonstrated a high level of interobserver reproducibility both within the same
laboratory and between independent laboratories.[27, 28]
A particular issue for the interpretation of IHC assays for melanoma tissues is the
presence of melanin pigment. Melanin pigmentation can interfere with IHC
interpretation, as it may obscure morphological features and is similar in color to the
chromogen 3,3’-diaminobenzidine tetrahydrochloride hydrate (DAB), which is commonly
used in IHC assays, including the LAG-3 IHC assay described here. The pretreatment
method described in this work removed melanin from samples without compromising
the LAG-3 antigen and resulted in no samples that could not be interpreted due to
excess melanin pigmentation.
One limitation of the studies presented in this work is that a number of preanalytical
factors may impact the performance of the LAG-3 IHC assay, including location of the
21
tissue assessed (ie, primary vs. metastatic),[29, 30] sample ischemia time, and fixation
methods.[31] Additionally, the design of the cut slide stability studies compared LAG-3
staining and IC expression with baseline (time 0), but did not include comparison with
other timepoints.
The assay described in this report was utilized to stratify patients based on LAG-3
expression in RELATIVITY-047 (NCT03470922), a phase 2/3 clinical trial in patients
with previously untreated metastatic or unresectable melanoma. The trial compared
combined nivolumab (anti–PD-1) and relatlimab (anti–LAG-3) treatment with nivolumab
monotherapy, and benefit of combination therapy was observed in comparison with
nivolumab monotherapy.[21] While the median PFS estimates were longer for patients
with LAG-3 expression ≥1% across both treatment groups, a benefit with the
combination therapy over nivolumab was observed regardless of LAG-3 expression.
[21]
Both the present report and RELATIVITY-047 determined LAG-3 positivity using a ≥1%
cutoff.[21] However, the prevalence of LAG-3 positivity observed in other sample sets or
patient populations may vary, meaning cutoff values for clinical utility will have to be
determined and validated in clinical studies. For instance, Dillon et al reported a higher
prevalence of LAG-3 positivity using a ≥1% cutoff in a different set of commercially
procured FFPE melanoma samples than in the melanoma samples used in this
report.[32] Dillon et al also reported a higher prevalence of LAG-3 positivity in gastric
and gastroesophageal cancer samples than in the melanoma samples used in this
report. The LAG-3 assay described in this manuscript is currently being utilized in a
number of clinical trials for multiple different tumor types.
22
In summary, a robust IHC assay for the determination of LAG-3 IC status in the tumor
microenvironment in solid tumor tissues has been developed.
23
Acknowledgments
The authors thank John Feder and Samantha Yost, both of Bristol Myers Squibb, for
generating the CRISPR knock-out cell lines. Medical writing and editorial support were
provided by Peter Harrison, PhD, and Matthew Weddig of Spark Medica Inc, funded by
Bristol Myers Squibb.
Competing Interests
BM, LJ, JY, CS, JS, and SA are employees of Labcorp. BM, LJ, JY, SA, and JS have
stock in Labcorp. KJ, AS-C, and SS are consultants/independent contractors of
Labcorp. LD and JW are employees of and have stock in Bristol Myers Squibb. CH has
stock in Bristol Myers Squibb. DL had stock in Bristol Myers Squibb at the time the
study was performed.
Funding
This study was supported by Bristol Myers Squibb.
Authors’ Contributions
LJ, JY, BM, and JS designed the studies. LJ led the laboratory operation and
procedures to provide stained slides to pathologists. BM was the lead pathologist for the
study. BM, AS-C, SS, and KJ analyzed and interpreted the IHC slides and provided
LAG-3 scores. JY provided statistical study design, data analyses, and interpretation.
CS performed peptide inhibition assay. SA reviewed the data and provided input on the
interpretation of the data. JS, CS, LJ, LD, CH, and JW provided input on data analysis
and interpretation. LD co-led LAG-3 IHC diagnostic development with Labcorp. LD, JW,
and CH developed the validation strategy, in partnership with Labcorp, and reviewed
24
and approved the experimental design and validation reports. JW and CH served as
pathology subject matter experts for LAG-3 IHC assay development. DL oversaw assay
verification and optimization experiments in support of assay transfer to Labcorp and
trained Labcorp staff on using the LAG-3 IHC assay. CH trained pathologists at Labcorp
on manual scoring of the LAG-3 IHC assay and developed the LAG-3 IHC scoring
algorithm and the assay scoring manual used at Labcorp. All authors contributed to
drafting, reviewed, and approved the manuscript.
Data availability statement
The datasets generated during and/or analyzed during the current study are not publicly
available but are available from the corresponding author on reasonable request.
25
REFERENCES
1.
Vaddepally RK, Kharel P, Pandey R, et al. Review of indications of FDA-
approved immune checkpoint inhibitors per NCCN Guidelines with the level of
evidence. Cancers (Basel) 2020;12(3):738.
2.
Guo L, Wei R, Lin Y, et al. Clinical and recent patents applications of PD-1/PD-L1
targeting immunotherapy in cancer treatment-current progress, strategy, and
future perspective. Front Immunol 2020;11:1508.
3.
Larkin J, Chiarion-Sileni V, Gonzalez R, et al. Combined nivolumab and
ipilimumab or monotherapy in previously untreated melanoma. N Engl J Med
2015;373(1):23–34.
4.
Ascierto PA, Del Vecchio M, Mandalá M, et al. Adjuvant nivolumab versus
ipilimumab in resected stage IIIB-C and stage IV melanoma (CheckMate 238): 4-
year results from a multicentre, double-blind, randomised, controlled, phase 3
trial. Lancet Oncol 2020;21(11):1465–77.
5.
Eggermont AMM, Blank CU, Mandala M, et al. Adjuvant pembrolizumab versus
placebo in resected stage III melanoma (EORTC 1325-MG/KEYNOTE-054):
distant metastasis-free survival results from a double-blind, randomised,
controlled, phase 3 trial. Lancet Oncol 2021;22(5):643–54.
6.
Hellmann MD, Paz-Ares L, Bernabe Caro R, et al. Nivolumab plus ipilimumab in
advanced non–small-cell lung cancer. N Engl J Med 2019;381(21):2020–31.
7.
Paz-Ares L, Ciuleanu T-E, Cobo M, et al. First-line nivolumab plus ipilimumab
combined with two cycles of chemotherapy in patients with non-small-cell lung
26
cancer (CheckMate 9LA): an international, randomised, open-label, phase 3 trial.
Lancet Oncol 2021;22(2):198–211.
8.
Ferris RL, Blumenschein Jr G, Fayette J, et al. Nivolumab for recurrent
squamous-cell carcinoma of the head and neck. N Engl J Med
2016;375(19):1856–67.
9.
Burtness B, Harrington KJ, Greil R, et al. Pembrolizumab alone or with
chemotherapy versus cetuximab with chemotherapy for recurrent or metastatic
squamous cell carcinoma of the head and neck (KEYNOTE-048): a randomised,
open-label, phase 3 study. Lancet 2019;394(10212):1915–28.
10.
Bellmunt J, de Wit R, Vaughn DJ, et al. Pembrolizumab as second-line therapy
for advanced urothelial carcinoma. N Engl J Med 2017;376(11):1015–26.
11.
Bajorin DF, Witjes JA, Gschwend J, et al. First results from the phase 3
CheckMate 274 trial of adjuvant nivolumab vs placebo in patients who underwent
radical surgery for high-risk muscle-invasive urothelial carcinoma (MIUC). J Clin
Oncol 2021;39(suppl 6):Abstract 391.
12.
Huang R-Y, Francois A, McGray AR, et al. Compensatory upregulation of PD-1,
LAG-3, and CTLA-4 limits the efficacy of single-agent checkpoint blockade in
metastatic ovarian cancer. Oncoimmunology 2017;6(1):e1249561.
13.
Long L, Zhang X, Chen F, et al. The promising immune checkpoint LAG-3: from
tumor microenvironment to cancer immunotherapy. Genes Cancer 2018;9(5–
6):176–89.
27
14.
Camisaschi C, Casati C, Rini F, et al. LAG-3 expression defines a subset of
CD4(+)CD25(high)Foxp3(+) regulatory T cells that are expanded at tumor sites. J
Immunol 2010;184(11):6545–51.
15.
Grosso JF, Kelleher CC, Harris TJ, et al. LAG-3 regulates CD8+ T cell
accumulation and effector function in murine self- and tumor-tolerance systems.
J Clin Invest 2007;117(11):3383–92.
16.
Woo S-R, Turnis ME, Goldberg MV, et al. Immune inhibitory molecules LAG-3
and PD-1 synergistically regulate T-cell function to promote tumoral immune
escape. Cancer Res 2012;72(4):917–27.
17.
Keane C, Law SC, Gould C, et al. LAG3: a novel immune checkpoint expressed
by multiple lymphocyte subsets in diffuse large B-cell lymphoma. Blood Adv
2020;4(7):1367-77.
18.
Workman CJ, Cauley LS, Kim I-J, et al. Lymphocyte activation gene-3 (CD223)
regulates the size of the expanding T cell population following antigen activation
in vivo. J Immunol 2004;172(9):5450–55.
19.
Wang J, Sanmamed MF, Datar I, et al. Fibrinogen-like protein 1 is a major
immune inhibitory ligand of LAG-3. Cell 2019;176(1-2):334–47.e12.
20.
Matsuzaki J, Gnjatic S, Mhawech-Fauceglia P, et al. Tumor-infiltrating NY-ESO-
1-specific CD8+ T cells are negatively regulated by LAG-3 and PD-1 in human
ovarian cancer. Proc Natl Acad Sci U S A 2010;107(17):7875–80.
21.
Tawbi HA, Schadendorf D, Lipson EJ, et al. Relatlimab and nivolumab versus
nivolumab in untreated advanced melanoma. N Engl J Med 2022;386(1):24–34.
28
22.
Baixeras E, Huard B, Miossec C, et al. Characterization of the lymphocyte
activation gene 3-encoded protein. A new ligand for human leukocyte antigen
class II antigens. J Exp Med 1992;176(2):327–37.
23.
Woo S-R, Li N, Bruno TC, et al. Differential subcellular localization of the
regulatory T-cell protein LAG-3 and the coreceptor CD4. Eur J Immunol
2010;40(6):1768–77.
24.
ClinicalTrials.gov. A Study to Assess Adjuvant Immunotherapy With Relatlimab
and Nivolumab Versus Nivolumab Alone After Complete Resection of Stage III-IV
Melanoma (RELATIVITY-098). 2022.
https://clinicaltrials.gov/ct2/show/NCT05002569: Accessed February 11
25.
ClinicalTrials.gov. A Study to Evaluate the Safety, Tolerability, and Efficacy of
Relatlimab in Relapsed or Refractory B-Cell Malignancies. 2022.
https://clinicaltrials.gov/ct2/show/NCT02061761: Accessed February 11
26.
Morgensztern D, Chaudhry A, Ianotti N, et al. 1359TiP RELATIVITY-104: First-
line relatlimab (RELA) + nivolumab (NIVO) with chemotherapy vs nivo with
chemotherapy in stage IV or recurrent non-small cell lung cancer (NSCLC): A
phase II, randomized, double-blind study. Annals of Oncology 2021;5:S1030.
27.
Adam J, Le Stang N, Rouquette I, et al. Multicenter harmonization study for PD-
L1 IHC testing in non-small-cell lung cancer. Ann Oncol 2018;29(4):953–58.
28.
Rimm DL, Han G, Taube JM, et al. A prospective, multi-institutional, pathologist-
based assessment of 4 immunohistochemistry assays for PD-L1 expression in
non-small cell lung cancer. JAMA Oncol 2017;3(8):1051–58.
29
29.
Peng L, Zhang Z, Zhao D, et al. Discordance of immunohistochemical markers
between primary and recurrent or metastatic breast cancer: A retrospective
analysis of 107 cases. Medicine (Baltimore) 2020;99(25):e20738.
30.
Rozenblit M, Huang R, Danziger N, et al. Comparison of PD-L1 protein
expression between primary tumors and metastatic lesions in triple negative
breast cancers. J Immunother Cancer 2020;8(2):e001558.
31.
Ramos-Vara JA, Miller MA. When tissue antigens and antibodies get along:
revisiting the technical aspects of immunohistochemistry--the red, brown, and
blue technique. Vet Pathol 2014;51(1):42–87.
32.
Dillon LM, Wojcik J, Desai K, et al. Abstract 1625: Distribution and prevalence of
LAG-3 expression in samples of melanoma and gastric/gastroesophageal
junction cancer. Cancer Res 2021;81(Suppl 13):1625–25.
30
Figures
FIGURE 1. Identification of LAG-3 in human tissues using the LAG-3 IHC assay. A,
Detection of LAG-3 in human tonsil tissue. Left-hand image depicts LAG-3 staining
pattern in tonsil tissue showing moderate to strong plasma membrane/cytoplasmic
staining in lymphocytes in germinal centers and interfollicular region. The crypt
epithelium is negative. No staining is seen with negative reagent control (right-hand
image). B, Staining of FFPE melanoma samples with negative reagent control (upper)
or LAG-3 antibody (lower) before (left) and after (right) melanin removal procedure at
10× magnification. C, Examples of LAG-3 staining in FFPE melanoma samples before
(upper) and after (lower) the melanin removal procedure at 20× magnification. FFPE,
formalin-fixed paraffin-embedded; IHC, immunohistochemistry; LAG-3, lymphocyte-
activation gene 3.
31
32
FIGURE 2. Detection of LAG-3 expression in parental COV434 cells and LAG-3–
disrupted COV434 cells. A, Bar charts showing NGS results from each of the pooled
CRISPR-engineered COV434 cell lines. B, IHC staining showing LAG-3 expression in
parental COV434 cells and the 3 pooled CRISPR-engineered COV434 cell lines. Tonsil
tissue was used as a positive/negative control for the IHC staining. CRISPR, clustered
regularly interspaced short palindromic repeats; IHC, immunohistochemistry; LAG-3,
lymphocyte-activation gene 3; NGS, next-generation sequencing; WT, wild type.
33
FIGURE 3. Detection of a range of LAG-3 expression levels using the LAG-3 IHC
assay. Bar chart showing scoring distribution across LAG-3–positive samples (defined
as those with LAG-3–positive IC content ≥1%) from a set of 100 commercially procured
human FFPE melanoma specimens. Of the 100 samples, 38 were LAG-3–positive and
62 were LAG-3–negative. FFPE, formalin-fixed paraffin-embedded; IC, immune cell;
IHC, immunohistochemistry; LAG-3, lymphocyte-activation gene 3.
33
34
FIGURE 4. Examples of a range of LAG-3 expression levels detected in melanoma
tissues using the LAG-3 IHC assay. Melanoma tissues showing a range of staining
(0%–30%) for LAG-3 examined at magnifications of 10× (left-hand image) and 20×
(right-hand image). IHC, immunohistochemistry; LAG-3, lymphocyte-activation gene 3.
35
| 2022 | Development of a LAG-3 Immunohistochemistry Assay for Melanoma | 10.1101/2022.02.25.481964 | [
"Johnson Lori",
"McCune Bryan",
"Locke Darren",
"Hedvat Cyrus",
"Wojcik John B.",
"Schroyer Caitlin",
"Yan Jim",
"Johnson Krystal",
"Sanders-Cliette Angela",
"Samala Sujana",
"Dillon Lloye M.",
"Anderson Steven",
"Shuster Jeffrey"
] | null |
1
A Mycobacterium tuberculosis effector targets mitochondrion, controls
energy metabolism and limits cytochrome c exit.
Marianne Martin1, Angelique deVisch3, Philippe Barthe3, Obolbek Turapov2, Talip Aydogan1,
Laurène Heriaud3, Jerome Gracy3, Galina V. Mukamolova2, François Letourneur1*, Martin Cohen-
Gonsaud3*
1 Laboratory of Pathogen Host Interactions (LPHI), CNRS, University of Montpellier, France.
2 Leicester Tuberculosis Research Group, Department of Respiratory Sciences, University of
Leicester, UK.
3 Centre de Biochimie Structurale, CNRS, INSERM, Université de Montpellier, France.
* Corresponding authors
Abstract
Host metabolism reprogramming is a key feature of Mycobacterium tuberculosis (Mtb) infec-
tion that enables the survival of this pathogen within phagocytic cells and modulates the
immune response facilitating the spread of the tuberculosis disease. Here, we demonstrate
that a previously uncharacterized secreted protein from Mtb, Rv1813c manipulates the host
metabolism by targeting mitochondria. When expressed in eukaryotic cells, the protein is
delivered to the mitochondrial intermembrane space and enhances host ATP production by
boosting the oxidative phosphorylation metabolic pathway. Furthermore, Rv1813c appears
to differentially modulate the host cell response to oxidative stress. Expression of Rv1813 in
host cells inhibits the release of cytochrome c from mitochondria, an early apoptotic event,
in response to short-term oxidative stress. However, Rv1813c expressing cells showed in-
creased sensitivity to prolonged stress. This study reveals a novel class of mitochondria tar-
geting effectors from Mtb and opens new research directions on host metabolic reprogram-
ming and apoptosis control.
Introduction
Mycobacterium tuberculosis (Mtb) encodes secreted virulence factors contributing to its successful
infection of host cells and its ability to actively replicate inside the phagosome (Hmama et al., 2015)
(Winden et al., 2019). After phagocytosis, Mtb blocks phagosomal maturation, escapes phago-
somes and subverts the host immune response. Several virulence factors (e.g. proteins, lipids) have
been already described to mediate such mechanisms, but corruption of host cell defense is clearly
multifactorial (Nicholson et al., 2021). It is estimated that over 20% of bacterial proteins have func-
tions outside the bacterial cytoplasm and are exported to their designated locations by protein
2
export systems (Kostakioti et al., 2005). Identification of secreted proteins remains a challenging
task. Data from various proteomic studies on secreted mycobacterial proteins have shown a weak
overlap for proteins identified as secreted in different studies (Målen et al., 2007; Lange et al., 2014).
As experiments were made in various culture conditions, it is not surprising that secretion patterns
differ from one experiment to another. Furthermore, the host cell environment also plays an im-
portant role in defining the secretion pattern, as recently revealed by studies focusing on the iden-
tification of secreted proteins during infection (Perkowski et al., 2017; Penn et al., 2018). To get a
broader view on the Mtb secretome, we used multidisciplinary approaches including bioinformatics,
structural and biochemical techniques, and cellular biology analyses. We identified putative Mtb
secreted proteins using proteins primary sequence analysis combined with structure modelling.
Among the selected targets, we studied the protein coded by the rv1813c gene which is only pre-
sent in mycobacterial pathogens. The Rv1813c protein has been used as vaccine adjuvant
(Bertholet et al., 2008) and displays immunogenicity properties (Liang et al., 2019). Rv1813c ex-
pression was reported to be MprA and DosR regulated (Bretl et al., 2012) and Mtb ΔRv1813c mu-
tant was attenuated in the low-dose aerosol model of murine tuberculosis (Bretl et al., 2012).
In this paper we describe extensive molecular and functional analyses of this protein. We showed
that Rv1813c defines a new class of effectors, with an original fold, addressed to mitochondria.
Mitochondrion plays critical functions not only supplying cells with energy but also contributing to
several cellular mechanisms including cell cycle, apoptosis, and signaling pathways. Metabolism
modulation dictates macrophage function and subsequent Mtb infection progression. Here, we
demonstrate for the first time that Rv1813c affects the mitochondrial metabolic functions and the
cell response to oxidative stress. Together these results suggest that Rv1813c might be a key reg-
ulator of the metabolic shift and apoptosis regulation occurring in Mtb infected macrophages
Results
Bioinformatic analysis of Mtb genome for identification of secreted proteins
Mtb possesses at least three different secretion systems (Feltcher et al., 2010). The general secre-
tion (Sec) and the twin-arginine translocation (Tat) pathways perform the bulk of protein export and
are both essential for growth. Proteins exported by the Sec pathway are distinguished by the pres-
ence of an N-terminal signal recognized by the SecA proteins before translocation. The Tat pathway
exports preproteins containing N-terminal signal peptides with a twin-arginine motif for binding to
the TatC protein. Mtb has also specialized export pathways that transport subsets of proteins. Five
specialized ESX export systems (ESX-1 to ESX-5) are present in Mtb with some of them essential
for virulence (Brodin et al., 2006)(Tran et al. 2021). The ESX systems are also referred to as Type VII
secretion systems (T7SS). Proteins secreted by T7SS lack Sec or Tat signal peptides, instead se-
cretion relies on a combination of a sequence and a structural motif (Daleke et al., 2012). We ana-
lyzed the predicted Mtb H37Rv proteome using an in-house Protein Analysis Toolkit (PAT) 2/1/21
3
11:50:00 AM. First, known signal peptides and/or structural features necessary for secretion were
predicted using SignalP v4.1 and PredTAT softwares. In addition, transmembrane segments were
inferred using either Uniprot annotations or the TmHMM prediction software. The number of pre-
dicted transmembrane segments and the position of the last transmembrane segment were also
analyzed to identify signals potentially missed by the other servers. To search for potential T7SS-
mediated secreted proteins, we first performed helix structure prediction within the first 80 residues
of each protein using Psipred (McGuffin et al., 2000), and then searched for the YxxxD/E motif in
between the two helices. These data were compared with various proteomic data and model data-
bases (ModBase, Interpro, GO). Using this approach, we identified 118 putative T7SS-, 124 putative
Tat-, and 350 putative Sec-mediated secreted Mtb proteins. Proteins to be studied further were
selected if they met one of the following criteria linked to a putative host-pathogen interaction: i) a
small domain of unknown function, ii) a protein/protein interaction domain, iii) a "eukaryotic" domain
(e.g., arrestin). Among the proteins identified as potentially secreted, we studied Rv1813c, a 143
amino-acid protein, not previously identified as secreted and comprising a predicted folded domain
of unknown function.
Rv1813c protein sequence features
Primary sequence analysis of the Rv1813c protein unambiguously identified a potential signal se-
quence (residues 1 to 28) with an upstream arginine repeat (residues 6-8) indicating that the protein
could be exported by the TAT export system (Fig. 1A). Homologous proteins are mostly found in
Actinobacteria (Mycobacterium, Nocardia and Streptomyces genera). In addition to Mtb, the protein
is present in various mycobacteria including Mycobacterium marinum (Mmar), Mycobacterium
avium, Mycobacterium ulcerans and Mycobacterium abscessus. Multiple paralogues exist within
the same bacteria. For instance, Mtb possesses only one orthologue (Rv1269c) whereas Mmar
harbors three paralogues (MMAR_1426, MMAR_2533 and MMAR_4153). The sequence homology
between these various proteins is high (between 45 to 70%), with a lower sequence identity for the
N-terminal part of the protein (Fig. 1A). Four cysteine residues are present and conserved. The last
four amino acid residues (140WACN143) composed a strictly conserved motif that includes one of the
conserved cysteines. Fold-recognition and modelling server @TOME2 previously used in many
studies for protein function identification even at low sequence identity (Turapov et al., 2014) did
not identify any close or distant Rv1813c structural homologues.
Rv1813c is secreted by M. tuberculosis in broth culture
Despite its use as vaccine adjuvant (Bertholet et al., 2008), its immunogenicity properties (Liang et
al., 2019) and a clear secretion signal sequence, no published proteomic to date have identified
Rv1813c as a secreted protein, possibly due to the small size of the protein. Western blot analyses
were carried out using a rabbit polyclonal antibody developed against recombinant Rv1813c (see
4
below). As shown in Fig. 2, Rv1813c was detected in the Mtb culture filtrate but not in any of the
cellular fractions including the cell wall. This result suggests that Rv1813c is secreted during active
growth in culture medium and is not bound to the bacteria cell wall.
Rv1813c defines a new protein family and a unique fold
The Rv1813c-coding sequence without the first codons corresponding to the protein signal peptide
(residues 1 to 27) was cloned into an Escherichia coli expression vector. The protein was over-
expressed as inclusion bodies, purified and refolded. Circular dichroism experiments demonstrated
that the recombinant protein was folded, and SEC-MALS analysis (Size Exclusion Chromatography
– Multi Angle Light Scattering) confirmed that the sample size matched the predicted folded mon-
omer size. Next, the purified protein was used for multidimensional NMR experiments. Preliminary
examination of [1H,15N]-HSQC spectrum revealed that 30 residues were unfolded (Supplementary
Fig. S1). A full multi-dimensional NMR study led to the protein three-dimensional structure resolu-
tion (Fig. 1B, Supplementary Table S1). Structure resolution demonstrated that the residues 28 to
57 were unfolded and that the protein possessed a 86 residues folded domain. This domain is
composed of two duplicate lobes facing each other, certainly inherited from a duplication despite
a lack of sequence homology. Each lobe is a series of three ß-strands with an hydrophobic surfaces
and an α-helix (ß/ß/α/ß). The four conserved cysteines are engaged in disulphide bonds but, note-
worthy, the two di-sulfide bonds are located in different parts of each lobe. The conserved WACN
motif cysteine is engaged in a disulfide bond linking the strands ß6 and ß4, while its tryptophan is
solvent exposed (as is also the second protein tryptophan). We hypothetically assumed that this
solvent exposed tryptophan might be important for the protein function (i.e. hydrophobic binding
or protein surface recognition). The overall structure defines a previously undescribed fold as both
Dali (Holm and Rosenström, 2010) and FATCAT (Ye and Godzik, 2004) servers failed to detect any
structural homologues. Consequently, sequence and structure comparison analysis did not bring
any indication on the potential biological function of the Rv1813c protein family.
Rv1813c is addressed to mitochondria in Dictyostelium discoideum.
To assess the function of Rv1813c in host cells, we first used the amoeba Dictyostelium dis-
coideum. This professional phagocyte is amenable to biochemical, cell biological and genetic ap-
proaches, and has proved to be an advantageous host cell model to analyze the virulence of several
pathogenic bacteria (Steinert, 2011; Müller-Taubenberger et al., 2013). Furthermore, the intracellu-
lar replication of Mmar has been extensively studied in D. discoideum and shows similarity to Mtb
replication (Cardenal-Muñoz et al., 2017), indicating that comparable molecular mechanisms are at
play in infected D. discoideum and mammalian host cells. We first analyzed the intracellular locali-
sation of Rv1813c when overexpressed in D. discoideum (ectopic expression). Though protein ex-
5
pression levels might differ from what is encountered during Mtb infection, ectopic expression al-
lows the advantageous analysis of individual secreted mycobacterial proteins without the interfer-
ence of other bacterial effectors. Rv1813c deleted of its predicted signal peptide (first 27 amino
acid residues) was tagged with a myc epitope at its N-terminus (myc-Rv1813c_P28-N143, here
after referred to as myc-Rv1813c) and stably expressed in D. discoideum. Confocal microscopy
analysis revealed colocalization in ring like structures of myc-Rv1813c coinciding with a mitochon-
drial outer membrane protein, Mitoporin (Troll et al., 1992) (Fig. 3A). Mitochondrial targeting was
also observed in cells expressing Rv1813c tagged at the C-terminus (Rv1813c-myc) but was lost
when Rv1813c was fused to GFP (data not shown). This specific targeting was independent of the
added myc-tag as staining with an anti-Rv1813c polyclonal antibody of untagged Rv1813c showed
identical results (Supplementary Fig. S2A). Mitoporin staining patterns were similarly observed in
both recipient (Ax2) and Rv1813c transfected cells (Supplementary Fig. S2A, S2B) excluding gross
mitochondrial morphological defects induced by Rv1813c expression in D. discoideum. In cells
labeled with mitotracker deep red, a specific dye accumulating inside mitochondria, myc-Rv1813c
surrounded labeled mitochondria and was mostly excluded from internal structures (Fig. 3B). This
result suggested that Rv1813c might be attached either to the internal or the cytosolic sides of
mitochondrial outer membranes. Interestingly, deletion of the unfolded N-terminus region of
Rv1813c (myc-Rv1813c_49-143) had no effect on Rv1813c localization whereas Rv1813c deprived
of the folded region (myc-Rv1813c_28-56) was not transported to mitochondria (Fig. 3C). Thus, the
Rv1813c folded domain, which does not contain any known mitochondrial targeting signals, was
sufficient to specifically direct this protein to mitochondrial outer membranes.
Rv1813c homologues are addressed to mitochondria in D. discoideum
Intracellular localization was next extended to members of the Rv1813c family in Mtb and Mmar.
All these proteins were detected in mitochondria, however some Rv1813c-like proteins affected
mitochondria morphology. Whereas overexpression of Rv1813c Mmar orthologs (MMA_1426 and
MMA_2533) did not induce any apparent morphological defects in mitochondria, cells expressing
Rv1269c or its Mmar ortholog MMA_4153 displayed mitochondria with aberrant shapes and sizes
(Supplementary Fig. S2C). In addition to mitochondria, MMA_4153 also localized to the cytosol.
Together these results indicated that mitochondrial targeting is a characteristic feature of the
Rv1813c family, and for some members, this localization leads to defective mitochondrial morphol-
ogy.
Rv1813c resides in the mitochondrial inter membrane space
Mitochondria are composed of two membranes, the outer and inner membranes, separated by an
inter membrane space (IMS). To determine more precisely the localization of Rv1813c within these
6
submitochondrial compartments, we next applied a biochemical approach. First, mitochondria en-
riched fractions (here after referred to as mitochondria) were obtained by subcellular fractionation
(see scheme Fig. 4A). As expected, Rv1813c was recovered from the mitochondrial fraction con-
firmed by Mitoporin enrichment (Fig. 4B). Next, Triton X-114 phase partitioning experiments re-
vealed that Rv1813c is not an integral membrane protein, in agreement with the absence of any
predicted transmembrane domains (Fig. 4C) and its exclusion from the Mtb cell wall (Fig. 2). Con-
sistently, Rv1813c was extracted from mitochondrial membranes by sodium carbonate treatment,
a characteristic feature of peripheral membrane proteins (Fig. 4D). Since Rv1813c was not released
from mitochondria by high salt washes (Fig. 4E) and was protected from proteinase K digestion of
intact mitochondria (Fig. 4F), we concluded that Rv1813c resides inside mitochondria. In addition,
Rv1813c was partially released from mitochondria upon the specific rupture of mitochondrial outer
membranes in hypotonic medium indicating that Rv1813c accumulates into the mitochondrial IMS
(Fig. 4G).
Rv1813c disturbs mitochondrial membrane potential but not ROS production
To assess whether Rv1813c mitochondrial localization might interfere with mitochondrial functions,
we monitored the mitochondrial membrane potential (ΔΨM), a key indicator of mitochondrial activ-
ity. We used the membrane-permeant JC-1 dye which accumulates in healthy mitochondria and
forms aggregates exhibiting a fluorescence emission shift from green (~529 nm) to red (~590 nm)
which can be easily followed by flow cytometry. Here we noticed that Rv1813c expressing cells
showed an increased JC1-1 red/green ratio, consistent with an elevated mitochondrial membrane
potential compared to recipient cells (Fig. 4H). Interestingly, this high ΔΨM was not associated with
an increased production of mitochondrial reactive oxygen species (ROS) as assayed by flow cy-
tometry analysis of MitoSox stained cells (Fig. 4I).
Rv1813c overexpression increases cell death upon oxidative stress in D. discoideum
In addition to energetic and metabolism regulatory functions, mitochondria play essential roles in
cell death induced in response to oxidative stress. To test whether Rv1813c might impede this
mitochondrial function, cells were treated with hydrogen peroxide and observed by phase contrast
microscopy. Samples with cells overexpressing Rv1813c showed increased number of shrank and
broken cells upon addition of 0.4 mM hydrogen peroxide for four hours compared to recipient cells
(Fig. 4J). Quantification of cell viability by propidium iodide (PI) incubation and flow cytometry anal-
ysis revealed that Rv1813c overexpression significantly increased oxidative stress sensitivity of
Dictyostelium (Fig. 4K). Together these data indicated that Rv1813c targeting to mitochondria re-
sults in deleterious mitochondrial functions in resting cells further amplified under stress conditions.
7
Rv1813c protein family members are addressed to mitochondria in mammalian cells
We next extended the analysis to mammalian host cells. Native and myc-Rv1813c were transiently
expressed in HeLa cells and their intracellular localization was determined by confocal microscopy.
As observed in Dictyostelium, Rv1813c was efficiently targeted to mitochondria (Fig. 5A and Sup-
plementary Fig. S3) without any detectable morphological effects (Fig. 5B). MMA_1436 and
MMA_2533, two Mmar orthologs of Rv1813c also localized to mitochondria in HeLa cells. In con-
trast, Rv1269c remained in the cytosol both in HeLa and HEK293 cells (Supplementary Fig. S3
and Fig. S4). For MMA_4153, the Mmar orthologs of Rv1269c, a faint mitochondrial staining was
detected in both cell types. Though mitochondrial targeting might be dependent upon the expres-
sion level of ectopic proteins (expression in HEK293 cells gives a better yield), Rv1269c and
MMA_4153 might be incorrectly/partially folded when expressed in mammalian cells, as observed
in heterologous expression in E. coli, preventing efficient mitochondrial localization.
Whereas the overall morphology of mitochondria was preserved upon Rv1813c ectopic expression,
transmission electronic microscopy (TEM) revealed some ultrastructural modifications. Hence,
Rv1813c expressing cells contained mitochondria with either normal or electron-dense matrix, and
the intra-cristae space appeared significantly enlarged compared to native HeLa cells, a modifica-
tion observed by in Mtb H37Rv-infected macrophages (Abarca-Rojano et al., 2003) (Fig. 5D, 5E
and 5F). More important defects were also observed upon expression of Rv1813c in HEK 293 cells
providing higher protein expression levels than in HeLa cells (Supplementary Fig. S5). Since cristae
membranes are enriched in proteins involved in oxidative phosphorylation, this particular ultrastruc-
ture might lead to mitochondrial energetic/metabolism consequences.
Rv1813c overexpression enhances cell metabolism and mitochondrial ROS production
The observed changes in mitochondrial ultrastructure triggered by Rv1813c prompted us to test
whether they were associated with energy metabolism disorders. Oxidative phosphorylation
(OXPHOS) and glycolysis were simultaneously analyzed in intact cells making use of an extracellular
flux analyzer (XF, Agilent Seahorse). In this assay, mitochondrial respiratory characteristics are eval-
uated by recording oxygen consumption rate (OCR) upon sequential chemical perturbation of se-
lected mitochondrial functions (as detailed in figure 6 legend). In Rv1813c transfected cells, basal
respiration, ATP-linked respiration, maximal respiratory capacity and reserve capacity were signif-
icantly increased compared to native HeLa cells (Fig. 6A). Glycolysis was also assayed using a
glycolysis stress test (Agilent Technologies) and measurements of extracellular acidification rates
(ECAR) in incubation media. This assay revealed similar glycolytic profiles in control and Rv1813c
expressing HeLa cells (Fig. 6B). Next, mitochondrial membrane potential was tested using flow
cytometry of JC-1 stained cells. In contrast to Dictyostelium, expression of Rv1813c in Hela cells
had no effect on ΔΨM in resting cells (Fig. 6C). However, these cells showed a slight but significant
increased mitochondrial ROS production (Fig. 6D). Together results of these assays indicated that
8
Rv1813c expression improves mitochondrial respiratory capacities without altering glycolytic func-
tions, driving cells into an energy activated state. This higher mitochondrial respiration was associ-
ated with increased mitochondrial free radical formation without changes in the mitochondrial mem-
brane potential.
Rv1813c promotes cell death in response to prolonged oxidative stress
We next assessed whether these mitochondrial alterations might alter the ability of Rv1813c ex-
pressing mammalian cells to cope with oxidative stress, a feature of Mtb infection. Cells were in-
cubated in medium supplemented with increasing amounts of hydrogen peroxide (0.075 to 0.3
mM). After 8h and 24h, cell death was monitored by PI and Annexin V staining followed by flow
cytometry analysis. A first set of experiments using Hela cells resulted only in minor effects of
Rv1813c (data not shown). However, making use of HEK293 as recipient cells revealed an important
increase of total Annexin V positive cells (early and late apoptosis) in response to 0.15 mM hydrogen
peroxide over time in Rv1813c overexpressing cells compared to recipient cells (Fig. 6E). This in-
crease was maximal after 24h at 0.15 mM hydrogen peroxide whereas doubling this concentration
resulted in similar massive cell death even in empty vector transfected recipient cells (Fig. 6F). Thus,
as observed in D. discoideum, Rv1813c expression in mammalian cells enhanced the sensitivity of
cells to oxidative stress.
Short-term oxidative stress induces Rv1813c translocation and delays in cytochrome c re-
lease from mitochondria
Cytochrome c (Cyt-c) release from mitochondria into the cytosol is an early event in apoptotic cell
death in response to hydrogen peroxide (Stridh et al., 1998). To monitor this event, cells were incu-
bated with hydrogen peroxide for only three hours. Cyt-c and Rv1813c localizations were then an-
alyzed by confocal microscopy and quantified. As expected, Cyt-c showed a diffuse cytosolic stain-
ing in 21% of HeLa cells upon addition of 0.1mM hydrogen peroxide (Fig. 7A,B). Rv1813c release
from mitochondria was also observed in cells overexpressing Rv1813c in response to hydrogen
peroxide treatments (Fig. 7C). In contrast, Cyt-c release from mitochondria into the cytosol was
reduced in Rv1813c expressing cells, with only 7.9% of cells displaying a cytosolic Cyt-c staining
upon oxidative stress conditions (Fig. 7A,B). Note that cells with cytosolic Cyt-c always showed
concomitant Rv1813c cytosolic localization. Strikingly, Rv1813c release from mitochondria was
more frequently observed than Cyt-c translocation leading to another cell population with Rv1813c
in the cytosol but Cyt-c still in mitochondria (Fig. 7D). Rv1813c mitochondrial exit in response to
oxidative stress might be necessary for Cyt-c exit. As a whole, these results strongly suggested
that Rv1813c inhibits efficient Cyt-c translocation and possibly early apoptotic associated events.
9
Discussion
Intracellular pathogens (i.e. Rickettsia, Legionella, Salmonella) disrupt mitochondrial function mainly
due to indirect effects (Spier et al., 2019)(Garaude, 2019). Few effectors are directly targeting the
organelle (Hicks and Galán, 2013). For instance, EspF effector from enteropathogenic E. coli in-
duces cell death (Hua et al., 2018). After injection to the intestinal epithelial cells, EspF effector is
targeted to the mitochondria via a mitochondrial import signal and promotes caspase mediated
apoptosis (Hua et al., 2018). Recently it was demonstrated that the MitF protein from L. pneumo-
philia alters mitochondria fission dynamics and a consequence promotes a Warburg-like phenotype
in macrophages (Escoll et al., 2017). Using bioinformatics screening and functional analysis, we
have identified Rv1813c from Mtb as a putative secreted protein and established that Rv1813c
belongs to a new protein family specifically addressed to mitochondria. Furthermore, we demon-
strated that ectopic expression of Rv1813c in D. discoideum induced strong functional defects in
eukaryotic cells. In Mycobrowser (https://mycobrowser.epfl.ch) and other Mtb databases, Rv1813c
is currently annotated as a “conserved hypothetical protein”. Despite being used as a vaccine ad-
juvant (Bertholet et al., 2008) and its high immunogenicity (Liang et al., 2019), no functional infor-
mation is available to our knowledge. The gene is non-essential for growth, however its deletion
impaired virulence in low dose murine model (Bretl et al., 2012). Yet the precise mechanism of this
attenuation is currently unknown.
The structure of Rv1813c solved here defines a new protein folding with no homology with any
structures solved to date (Fig 1). The small 9 kDa folded domain, which includes a highly conserved
C-terminal motif (140WACN143), is sufficient to specifically address the protein into mitochondria
where it accumulates the IMS (Fig. 3 and Fig. 4). Noteworthy the efficiency of this mitochondrial
targeting differs among the family members analyzed. The highly divergent primary sequences of
the N-terminal unfolded parts might be responsible for this difference, possibly by impacting the
whole protein dynamics and stability.
The only previously published study on Rv1813c reported that MprA and DosR regulate its expres-
sion (Bretl et al., 2012). DosR is a transcriptional regulator induced by host intracellular stimuli, such
as nitric oxide (NO), carbon monoxide (CO), and hypoxia (Bretl et al., 2012), while MrpA responses
to environmental stress and within the macrophage (Haydel and Clark-Curtiss, 2004;(Pan et al.,
2020) and is required during infection (Zahrt and Deretic, 2001). Reference transcriptomic studies
have revealed that Rv1813c is over-expressed (x2 and x4, 24h and 48h post-infection, respectively)
in activated infected macrophages (Schnappinger et al., 2006), and in BLAC mouse model (x5, x14
then x2, 7 days 14 days and 28 days post-infection, respectively) (Talaat et al., 2004). Here we
confirmed that the protein is constitutively secreted by Mtb in culture medium and its overexpres-
sion in host cells results in phenotypes linked to its localisation into the mitochondrial IMS. Hence,
we demonstrate that ectopically expressed Rv1813c i) enhances OXPHOS, ii) increases cell death
under prolonged oxidative stress, iii) inhibits cytochrome-c exit upon short-term oxidative stress.
10
While more in vivo experiments will be necessary to fully understand the function of Rv1813c, the
three main Rv1813c-dependent defects revealed in our study are clearly connected to important
host defence mechanisms against Mtb infections.
Does the secretion of this Mtb effector protein increase mitochondrial ATP production in host cells
bring any substantial intracellular replication advantages or help to prevent host cell defenses?
Efficient response to Mtb infection by macrophages relies on their activation leading to polarisation
toward an M1 profile (Shi et al., 2019). This is achieved by a metabolic reprogramming after Nf-kB
activation either by pathogen-associated molecular patterns (PAMPs) or IFNg. Nf-kB promotes the
expression of the inducible nitric oxide synthase (iNOS) and subsequent nitric oxide (NO) release.
Besides bactericidal activity, NO directly inactivates the electron transfer chain (ETC) proteins, trig-
gering a complex series of events, mainly dependent to reactive oxygen species (ROS) production
and metabolites balance changes (i.e. NAD/NADP ratio). When the Krebs cycle is consequently
blocked, citrate accumulates enhancing glycolysis and lipids biosynthesis. In addition, succinate
also accumulates leading to HIF-1a (Hypoxia-inducible factor-1) stabilisation, which results in a
complete metabolic switch similar to the Warburg effect in tumours (Shi et al., 2016). HIF-1a not
only promotes the expression of enzymes involved in glycolytic ATP production, but also induces
expression pattern leading to synthesis of important immune effectors, including inflammatory cy-
tokines and chemokines under normoxic conditions. Therefore, full M1 polarisation is essential for
defense to pathogenic infections, and appears as a target choice for Mtb (Wilson et al., 2019).
Very few studies have assessed the precise metabolic state of Mtb infected cells (Mohareer et al.,
2020). Recently, bioenergetic analyses have been performed on infected macrophages. XF exper-
iments and metabolites analysis using 13C-tracing in infected macrophages have revealed a de-
creased of cell energetic flux through glycolysis and the TCA cycle. Consequently, the total level of
ATP produced in Mtb infected cells 5 and 24 hours post infection is decreased (Cumming et al.,
2018). However, different effects are observed with BCG or dead Mtb, which do not gain access to
the cytosol or release effector proteins in the cytosol after phagocytosis (Simeone et al., 2012).
Hence, the glycolytic flux is enhanced with these two strains in contrast to virulent H37rv bacteria.
These results suggest a weaker metabolic macrophage response to virulent Mtb infection that
might be controlled by bacterial effectors leading to incomplete or delayed metabolic shift.
On the other hand, studies have proposed that maintaining the ATP production is beneficial for
Mtb, avoiding ROS production and apoptosis. For instance, a much higher ATP/ADP ration was
observed in H37Rv-infected cells compared to cells infected with avirulent H37Ra, a strain that
does not escape from phagosomes to access the cytosol (Jamwal et al., 2013)(Jamwal et al., 2016).
An elevated ATP/ADP ratio was further correlated to lower apoptosis rates observed in H37Rv-
infected cells (Jamwal et al., 2013)(Mehrotra et al., 2014). Together these data indicate that main-
taining a high ATP production might be beneficial to delay a deleterious full metabolic shift and/or
apoptosis of the host-cell. Comforting this hypothesis, secretion of Rv1813c could participate in
11
maintaining a higher ATP production within mitochondria during Mtb infection. Moreover, a recent
study indicates that the anti-mycobacterial drug Bedaquiline disturbs the host metabolism and in-
creases the macrophage resistance to various bacterial infection without direct inhibition of the
pathogens (Giraud-Gatineau et al., 2020). While some diverging results have been reported
(Belosludtsev et al., 2019) (Luo et al., 2020), it is well-established that host ATP production is re-
duced in Bedaquiline-treated eukaryotic cells and this may contribute to successful Mtb elimina-
tion.
Our results also indicate that Rv1813c displays both anti-apoptotic and pro-apoptotic effects upon
short-term and prolonged oxidative stress respectively. Mitochondrial proton leak generated from
the ETC is the major source of mitochondrial ROS. ROS excess results in multiple effects including
cytochrome-c translocation followed by caspase dependent apoptosis as well as inflammasome
activation (Jamwal et al., 2013). Our bioenergetics and microscopy experimental data indicate that
Rv1813c expression does not modify mitochondria numbers. Instead ETC and/or ATP synthase
boosted functions are likely to account for the observed ATP increased production. Accordingly,
Rv1813c ectopic expression induces a slight increase of ROS in resting cells (Fig. 6D). Among
multiple cellular effects, hydrogen peroxide dramatically increases ROS production, eventually
leading to cell death by apoptosis and/or necrosis. Due to higher basal ETC activity, Rv1813c ex-
pressing cells might already cope for elevated ROS levels, and might be more sensitive to any
further elevated ROS levels, induced for instance by hydrogen peroxide treatments. This would
result in the reduction of the ROS level below the threshold that normally triggers cell death. This
hypothesis might thus explain the increase sensitivity of Dictyostelium and mammalian cells to pro-
longed oxidative stress. Surprisingly, Rv1813c expression produces opposite effects upon short-
term oxidative stress, inhibiting the release of Cyt-c from mitochondria, which is fully consistent
with an anti-apoptotic function. Several mechanisms have been described to explain the exit of
Cyt-c from mitochondria during apoptosis (Garrido et al., 2006). Our data indicate that mitochon-
drial membrane rupture does not participate in the release of Cyt-c since mitochondria keep their
integrity to selectively retain Cyt-c but release Rv1813c (Fig. 7). Though we cannot exclude that
Rv1813c might inhibit a cardiolipins dependent mechanism (Barayeu et al., 2019), our working hy-
pothesis is that Rv1813c might interfere with the BAX/BAK pore formation in mitochondrial outer
membranes required for Cyt-c exit.
Overall, our results on Rv1813c pave the way to further characterisation on the establishment of
metabolic shift and apoptosis regulation occurring in Mtb infected macrophages
12
Detailed methods are provided in Supplementary Materials and Methods of this paper and
include the following:
o Purification of recombinant 6His-Rv1813c28-143 in E. coli
o Solution structure of Rv1813c28-143
o Antibodies
o Preparation of Mycobacterium tuberculosis culture
o Mycobacterial cell fractionation
o Protein Electrophoresis and Western Blot
o Cell culture and transfection conditions
o Mitochondria isolation and biochemical treatments
o Immunocytochemistry
o Flow cytometry analysis of JC-1, MitoSox and Annexin 5/PI stained cells
o MACS enrichment of CD4-Rv1813c transfected cells
o Extracellular flux analysis
o Transmission electron microscopy
AUTHOR CONTRIBUTION
MM, AdV, PB, YMB, OP, TA, LH, JG, CG, GM, FL and MCG performed experiments; MM, GM, CG,
ON, FL and MCG analysed the data; FL and MCG conceived this study; All authors contributed to
manuscript writing.
AKNOWLEDGEMENTS
Flow cytometry and microscopy analyses of uninfected cells were performed at the Montpellier RIO
imaging facility of the University of Montpellier, member of the national infrastructure France-
BioImaging, supported by the French National Research Agency (ANR-10-INBS-04, “Investments
for the future”). The CBS acknowledges support from the French Infrastructure for Integrated
Structural Biology (FRISBI) ANR-10-INSB-05-01. The following reagents were obtained through BEI
Resources, NIAID, NIH: Monoclonal Anti-M. tuberculosis GlnA (Gene Rv2220), Clone IT-58 (CBA5)
(produced in vitro), NR-44103; Polyclonal Anti-Mycobacterium tuberculosis FtsZ (Gene Rv2150c)
(antiserum rabbit).
DECLATION OF INTERESTS
The authors declare no competing interests.
13
References
Abarca-Rojano, E., Rosas-Medina, P., Zamudio-Cortéz, P., Mondragón-Flores, R., and Sánchez-
García, F.J. (2003). Mycobacterium tuberculosis virulence correlates with mitochondrial cyto-
chrome c release in infected macrophages. Scand J Immunol 58, 419–427.
Barayeu, U., Lange, M., Méndez, L., Arnhold, J., Shadyro, O.I., Fedorova, M., and Flemmig, J.
(2019). Cytochrome c autocatalyzed carbonylation in the presence of hydrogen peroxide and car-
diolipins. J Biol Chem 294, 1816–1830.
Belosludtsev, K.N., Belosludtseva, N.V., Talanov, E.Y., Tenkov, K.S., Starinets, V.S., Agafonov, A.V.,
Pavlik, L.L., and Dubinin, M.V. (2019). Effect of bedaquiline on the functions of rat liver mitochondria.
Biochimica et Biophysica Acta (BBA) - Biomembranes 1861, 288–297.
Bertholet, S., Ireton, G.C., Kahn, M., Guderian, J., Mohamath, R., Stride, N., Laughlin, E.M., Bald-
win, S.L., Vedvick, T.S., Coler, R.N., et al. (2008). Identification of Human T Cell Antigens for the
Development of Vaccines Against Mycobacterium Tuberculosis. J Immunol 181, 7948–7957.
Bretl, D.J., He, H., Demetriadou, C., White, M.J., Penoske, R.M., Salzman, N.H., and Zahrt, T.C.
(2012). MprA and DosR coregulate a Mycobacterium tuberculosis virulence operon encoding
Rv1813c and Rv1812c. Infect. Immun. 80, 3018–3033.
Brodin, P., Majlessi, L., Marsollier, L., de Jonge, M.I., Bottai, D., Demangel, C., Hinds, J., Neyrolles,
O., Butcher, P.D., Leclerc, C., et al. (2006). Dissection of ESAT-6 system 1 of Mycobacterium tu-
berculosis and impact on immunogenicity and virulence. Infect Immun 74, 88–98.
Cardenal-Muñoz, E., Barisch, C., Lefrançois, L.H., López-Jiménez, A.T., and Soldati, T. (2017).
When Dicty Met Myco, a (Not So) Romantic Story about One Amoeba and Its Intracellular Pathogen.
Front Cell Infect Microbiol 7, 529.
Cumming, B.M., Addicott, K.W., Adamson, J.H., and Steyn, A.J. (2018). Mycobacterium tuberculo-
sis induces decelerated bioenergetic metabolism in human macrophages. Elife 7.
Daleke, M.H., Ummels, R., Bawono, P., Heringa, J., Vandenbroucke-Grauls, C.M.J.E., Luirink, J.,
and Bitter, W. (2012). General secretion signal for the mycobacterial type VII secretion pathway.
Proc Natl Acad Sci U S A 109, 11342–11347.
Escoll, P., Song, O.-R., Viana, F., Steiner, B., Lagache, T., Olivo-Marin, J.-C., Impens, F., Brodin,
P., Hilbi, H., and Buchrieser, C. (2017). Legionella pneumophila Modulates Mitochondrial Dynamics
to Trigger Metabolic Repurposing of Infected Macrophages. Cell Host Microbe 22, 302-316.e7.
Feltcher, M.E., Sullivan, J.T., and Braunstein, M. (2010). Protein export systems of Mycobacterium
tuberculosis: novel targets for drug development? Future Microbiol 5, 1581–1597.
Garaude, J. (2019). Reprogramming of mitochondrial metabolism by innate immunity. Curr Opin
Immunol 56, 17–23.
Garrido, C., Galluzzi, L., Brunet, M., Puig, P.E., Didelot, C., and Kroemer, G. (2006). Mechanisms
of cytochrome c release from mitochondria. Cell Death Differ 13, 1423–1433.
Giraud-Gatineau, A., Coya, J.M., Maure, A., Biton, A., Thomson, M., Bernard, E.M., Marrec, J.,
Gutierrez, M.G., Larrouy-Maumus, G., Brosch, R., et al. (2020). The antibiotic bedaquiline acti-
vates host macrophage innate immune resistance to bacterial infection. ELife 9, e55692.
14
Haydel, S.E., and Clark-Curtiss, J.E. (2004). Global expression analysis of two-component system
regulator genes during Mycobacterium tuberculosis growth in human macrophages. FEMS Micro-
biol. Lett. 236, 341–347.
Hicks, S.W., and Galán, J.E. (2013). Exploitation of eukaryotic subcellular targeting mechanisms
by bacterial effectors. Nat Rev Microbiol 11, 316–326.
Hmama, Z., Peña‐Díaz, S., Joseph, S., and Av‐Gay, Y. (2015). Immunoevasion and immunosup-
pression of the macrophage by Mycobacterium tuberculosis. Immunological Reviews 264, 220–
232.
Holm, L., and Rosenström, P. (2010). Dali server: conservation mapping in 3D. Nucleic Acids Res
38, W545-549.
Hua, Y., Yan, K., and Wan, C. (2018). Clever Cooperation: Interactions Between EspF and Host
Proteins. Front. Microbiol. 9.
Jamwal, S., Midha, M.K., Verma, H.N., Basu, A., Rao, K.V.S., and Manivel, V. (2013). Characteriz-
ing virulence-specific perturbations in the mitochondrial function of macrophages infected with
Mycobacterium tuberculosis. Sci Rep 3, 1328.
Jamwal, S.V., Mehrotra, P., Singh, A., Siddiqui, Z., Basu, A., and Rao, K.V.S. (2016). Mycobacte-
rial escape from macrophage phagosomes to the cytoplasm represents an alternate adaptation
mechanism. Scientific Reports 6, 23089.
Kostakioti, M., Newman, C.L., Thanassi, D.G., and Stathopoulos, C. (2005). Mechanisms of Pro-
tein Export across the Bacterial Outer Membrane. J Bacteriol 187, 4306–4314.
Lange, S., Rosenkrands, I., Stein, R., Andersen, P., Kaufmann, S.H.E., and Jungblut, P.R. (2014).
Analysis of protein species differentiation among mycobacterial low-Mr-secreted proteins by nar-
row pH range Immobiline gel 2-DE-MALDI-MS. J Proteomics 97, 235–244.
Lee, J.-M., Cho, H.Y., Cho, H.J., Ko, I.-J., Park, S.W., Baik, H.-S., Oh, J.-H., Eom, C.-Y., Kim,
Y.M., Kang, B.S., et al. (2008). O2- and NO-sensing mechanism through the DevSR two-compo-
nent system in Mycobacterium smegmatis. J. Bacteriol. 190, 6795–6804.
Liang, Y., Zhang, X., Bai, X., Yang, Y., Gong, W., Wang, T., Ling, Y., Zhang, J., Wang, L., Wang,
J., et al. (2019). Immunogenicity and Therapeutic Effects of Latency-Associated Genes in a Myco-
bacterium Tuberculosis Reactivation Mouse Model. Hum Gene Ther Methods 30, 60–69.
Luo, M., Zhou, W., Patel, H., Srivastava, A.P., Symersky, J., Bonar, M.M., Faraldo-Gómez, J.D.,
Liao, M., and Mueller, D.M. (2020). Bedaquiline inhibits the yeast and human mitochondrial ATP
synthases. Communications Biology 3, 1–10.
Målen, H., Berven, F.S., Fladmark, K.E., and Wiker, H.G. (2007). Comprehensive analysis of ex-
ported proteins from Mycobacterium tuberculosis H37Rv. Proteomics 7, 1702–1718.
McGuffin, L.J., Bryson, K., and Jones, D.T. (2000). The PSIPRED protein structure prediction
server. Bioinformatics 16, 404–405.
Mehrotra, P., Jamwal, S.V., Saquib, N., Sinha, N., Siddiqui, Z., Manivel, V., Chatterjee, S., and
Rao, K.V.S. (2014). Pathogenicity of Mycobacterium tuberculosis is expressed by regulating met-
abolic thresholds of the host macrophage. PLoS Pathog. 10, e1004265.
Mohareer, K., Medikonda, J., Vadankula, G.R., and Banerjee, S. (2020). Mycobacterial Control of
Host Mitochondria: Bioenergetic and Metabolic Changes Shaping Cell Fate and Infection Out-
come. Front. Cell. Infect. Microbiol. 10.
15
Müller-Taubenberger, A., Kortholt, A., and Eichinger, L. (2013). Simple system--substantial share:
the use of Dictyostelium in cell biology and molecular medicine. Eur J Cell Biol 92, 45–53.
Nicholson, K.R., Mousseau, C.B., Champion, M.M., and Champion, P.A. (2021). The genetic pro-
teome: Using genetics to inform the proteome of mycobacterial pathogens. PLOS Pathogens 17,
e1009124.
Pan, F., Ye, T., Sun, P., Gui, S., Liang, B., Li, L., Zheng, D., Wang, J., Hesketh, R.L., Yang, L., et
al. (2020). Time Course of Lung Changes On Chest CT During Recovery From 2019 Novel Coro-
navirus (COVID-19) Pneumonia. Radiology 200370–200370.
Penn, B.H., Netter, Z., Johnson, J.R., Von Dollen, J., Jang, G.M., Johnson, T., Ohol, Y.M., Maher,
C., Bell, S.L., Geiger, K., et al. (2018). An Mtb-Human Protein-Protein Interaction Map Identifies a
Switch between Host Antiviral and Antibacterial Responses. Mol Cell 71, 637-648.e5.
Perkowski, E.F., Zulauf, K.E., Weerakoon, D., Hayden, J.D., Ioerger, T.R., Oreper, D., Gomez,
S.M., Sacchettini, J.C., and Braunstein, M. (2017). The EXIT Strategy: an Approach for Identifying
Bacterial Proteins Exported during Host Infection. MBio 8.
Schnappinger, D., Schoolnik, G.K., and Ehrt, S. (2006). Expression profiling of host pathogen in-
teractions: how Mycobacterium tuberculosis and the macrophage adapt to one another. Microbes
Infect 8, 1132–1140.
Shi, L., Eugenin, E.A., and Subbian, S. (2016). Immunometabolism in Tuberculosis. Front Immunol
7, 150.
Shi, L., Jiang, Q., Bushkin, Y., Subbian, S., and Tyagi, S. (2019). Biphasic Dynamics of Macro-
phage Immunometabolism during Mycobacterium tuberculosis Infection. MBio 10.
Simeone, R., Bobard, A., Lippmann, J., Bitter, W., Majlessi, L., Brosch, R., and Enninga, J. (2012).
Phagosomal rupture by Mycobacterium tuberculosis results in toxicity and host cell death. PLoS
Pathog 8, e1002507.
Spier, A., Stavru, F., and Cossart, P. (2019). Interaction between Intracellular Bacterial Pathogens
and Host Cell Mitochondria. Bacteria and Intracellularity 3–13.
Steinert, M. (2011). Pathogen-host interactions in Dictyostelium, Legionella, Mycobacterium and
other pathogens. Semin Cell Dev Biol 22, 70–76.
Stridh, H., Kimland, M., Jones, D.P., Orrenius, S., and Hampton, M.B. (1998). Cytochrome c re-
lease and caspase activation in hydrogen peroxide- and tributyltin-induced apoptosis. FEBS Lett
429, 351–355.
Talaat, A.M., Lyons, R., Howard, S.T., and Johnston, S.A. (2004). The temporal expression profile
of Mycobacterium tuberculosis infection in mice. Proc. Natl. Acad. Sci. U.S.A. 101, 4602–4607.
Troll, H., Malchow, D., Müller-Taubenberger, A., Humbel, B., Lottspeich, F., Ecke, M., Gerisch, G.,
Schmid, A., and Benz, R. (1992). Purification, functional characterization, and cDNA sequencing
of mitochondrial porin from Dictyostelium discoideum. J Biol Chem 267, 21072–21079.
Turapov, O., Waddell, S.J., Burke, B., Glenn, S., Sarybaeva, A.A., Tudo, G., Labesse, G., Young,
D.I., Young, M., Andrew, P.W., et al. (2014). Oleoyl coenzyme A regulates interaction of transcrip-
tional regulator RaaS (Rv1219c) with DNA in mycobacteria. J Biol Chem 289, 25241–25249.
16
Wilson, J.L., Mayr, H.K., and Weichhart, T. (2019). Metabolic Programming of Macrophages: Im-
plications in the Pathogenesis of Granulomatous Disease. Front Immunol 10, 2265.
Winden, V.J.C. van, Houben, E.N.G., and Braunstein, M. (2019). Protein Export into and across
the Atypical Diderm Cell Envelope of Mycobacteria. Microbiology Spectrum 7.
Ye, Y., and Godzik, A. (2004). Database searching by flexible protein structure alignment. Protein
Sci 13, 1841–1850.
Zahrt, T.C., and Deretic, V. (2001). Mycobacterium tuberculosis signal transduction system re-
quired for persistent infections. PNAS 98, 12706–12711.
FIGURE LEGENDS
Fig. 1 Rv1813c defines a new protein family (A) M. tuberculosis and M. marinum Rv1813c
homologues sequence alignments. The conservation between the members is high for the folded
domain (secondary structure of the Rv1813c structure are reported above the alignment). The N-
terminal unfolded domain is less conserved and seems to define two subclasses. The yellow num-
bers indicate the cysteine engaged in disulfide bridges. (B) Rv1813c structure determined by
multi-dimensional NMR. Three cartoon representations of the structure. Only 4 residues of N-
terminal unfolded part of the protein (residues 28-57) are represented. The cysteine residues, all
engaged in disulfide bridges are represented in yellow while the two solvent-exposed tryptophans
are represented in orange.
Fig. 2 Rv1813c is secreted by Mtb in broth culture
Rv1813 is detected in M. tuberculosis culture filtrate but not in cellular fractions. Lysate ob-
tained from M. tuberculosis H37Rv strain grown in Sauton!s medium to logarithmic phase
(OD580~0.7) were fractionated and probed with anti-Rv1813 polyclonal antibodies. Culture filtrates
were obtained from the same culture. Anti-FtsZ (FtsZ is a cytoplasmic protein), anti-GlnA (GlnA is
a membrane protein) and anti-RpfB (RpfB is a membrane and cell-wall anchored protein) antibodies
were used to confirm the purity of mycobacterial fractions.
Fig. 3 Mitochondrial localisation of Rv1813c in Dictyostelium.
Dictyostelium cells expressing the indicated constructs were fixed, processed for immunofluores-
cence, and analyzed by confocal microscopy (Airyscan). (A) myc-Rv1813c detected with a rabbit
polyclonal to Rv1813c colocalises with mitochondrial Mitoporin in ring shaped structures. (B)
Rv1813c is mostly excluded from the inside of mitochondria labeled with the Mitotracker deep red
dye. Close up of mitochondria are shown in the insert. (C) The Rv1813c folded domain alone is
efficiently targeted to mitochondria. Maximum projection of Z confocal sections of the full-length
17
protein (myc-Rv1813c), the sole structured (myc-Rv1813c_49-143) and the unfolded domain (myc-
Rv1813c_28-56) labeled with an anti-myc antibody. Cell contour is indicated by dotted lines. Bar,
5 µm.
Fig. 4 Biochemical analysis of Rv1813c mitochondrial localization and functional conse-
quences
(A) Fractionation scheme of differential centrifugation steps used to purify the Rv1813c enriched
fraction. (B) Fractions were analyzed by immunoblotting with antibody to Mitoporin (mitochondria),
EHD (endocytic vacuoles) and myc-tag (myc-Rv1813c). Rv1813c concentrates into the mitochon-
drial enriched fraction P1. (C) The mitochondrial fraction was fractionated by Triton X114 extraction.
The separated Triton X-114 (TX) and aqueous (W) phases were analyzed as above. Rv1813c is not
extracted by Triton X-114 indicating no insertion inside membranes. (D) Mitochondria were incu-
bated in sodium carbonate for 30 min. Rv1813c is mainly detected in the supernatant fraction (S)
after centrifugation at 100,000g of treated mitochondria, a characteristic of soluble and/or mem-
brane peripheral proteins. (E) Mitochondria were incubated in buffer ± 200 mM KCl for 30 min and
centrifuged at 16,000 g for 10 min. Rv1813c stays in P1 indicating that no washing out from mito-
chondria by this treatment. (F) Intact or Triton X100 treated mitochondria were subjected to pro-
teinase K digestion for 30 min and analyzed by immunoblotting. In intact mitochondria, Rv1813c is
protected from proteinase K digestion but sensitive to this treatment in presence of detergent con-
sistent with an intra-mitochondrial localization. (G) Mitochondria swelling was induced by hypotonic
buffer incubation for 30 min. Released proteins from broken outer membranes were recovered by
centrifugation at 16,000g for 10 min and analyzed by western blotting. Rv1813c is equally distrib-
uted in P1 and S1 fractions indicating localization both on internal mitochondrial membranes
and in the IMS. (H) Analysis of mitochondrial membrane potential by flow cytometry of JC-1 stained
cells. JC-1 Red/Green ratio were calculated and expressed as the % of this ratio in recipient cells
(Ax2). Rv1813c expression increases the mitochondrial membrane potential as revealed by an ele-
vated JC-1 Red/Green ratio. Values are means ± s.e.m. of three independent experiments. (I)
Analysis of mitochondrial ROS production by flow cytometry of MitoSox stained cells. MitoSox
fluorescence was expressed as the % of fluorescence in recipient cells (Ax2). No modification of
mitochondrial ROS production is observed upon Rv1813c expression. Values are means ± s.e.m.
of three independent experiments. (J) Phase contrast microscopy of wild type (Ax2) or Rv1813c
expressing cells (Ax2 + myc-Rv1813c) incubated in 0.4 mM hydrogen peroxide for 4 hours. White
arrows indicate cellular debris. Bar, 10 µm. (K) Graph of cell viability observed after hydrogen per-
oxide treatment for 4 hours determined by FACS analysis upon Propidium Iodide staining of dead
cells. Rv1813c increases cell sensitivity to oxidative stress. Values are means ± s.e.m. of four
independent experiments.
18
Fig. 5 Targeting of Rv1813c to mitochondria in mammalian cells
(A) Confocal microscopy analysis of HeLa cells transiently expressing myc-Rv1813. Cells were
fixed 48h post-transfection, processed for immunofluorescence, and analyzed by Airyscan micros-
copy. myc-Rv1813c was detected by a polyclonal antibody to RV1813c and shows strict col-
ocalization with mitochondrial Cytochrome c. Bar, 10 µm. (B) Quantitative analysis of mitochon-
dria morphology in HeLa cells transiently transfected with pCI-myc-Rv1813 or empty vector. Mito-
chondria morphology was manually identified by confocal microscopy and classified in one hun-
dred cells. Rv1813c expressing cells exhibit a normal mitochondrial morphology. (C) Schematic
ultrastructure of a single crista in mitochondria. (D, E) Representative mitochondria ultrastructure
determined by transmission electronic microscopy of MACS enriched HeLa cells transiently trans-
fected with pMACS-4-IRES-II Rv1813c (D) or vector alone (E). Black arrows indicate some en-
larged intra-cristae spaces in Rv1813c expressing cells. Spaces between adjacent cristae are
enlarged upon Rv1813c overexpression. Bars, 500 nm. (F) Bar graph of intra-cristae spaces meas-
urements for the indicated cell lines (100 random measurements each, **** p<0.0001 in student
test).
Fig. 6 Functional consequences of Rv1813c mitochondrial localisation
(A) Rv1813c expression enhances cell respiratory functions. HeLa cells were transiently trans-
fected with pMACS-4-IRES-II Rv1813c or vector alone and enriched to >94% through a double
magnetic cell sorting (MACS) procedure (see experimental section). Cell respiratory profiles (OCR)
were obtained using an extracellular flux analyser (Seahorse XF analyzer) and the mitochondrial
respiration test. After reaching basal respiration, cells were subjected to 1µM oligomycin to inhibit
the ATP synthase and measure the mitochondrial ATP-linked OCR, followed by 1µM FCCP (cya-
nide-4-[trifluoromethoxy]phenylhydrazone) to uncouple mitochondrial respiration and maximize
OCR, and finally 1µM antimycin A and 100nM rotenone to inhibit complex III and I in the ETC re-
spectively, and shut down respiration. In Rv1813c transfected cells, basal respiration, ATP-linked
respiration, maximal respiratory capacity and reserve capacity are significantly increased compared
to native HeLa cells. (B) Analysis of glycolytic functions. Extracellular acidification (ECAR) profiles
of the same MACS enriched transfected cells were determined simultaneously to OCR analysis
using the glycolysis stress test and the XF analyzer. After reaching non-glycolytic acidification,
10mM glucose was added, followed by 1µM oligomycin to inhibit the ATP synthase and induce
maximal glycolysis. Finally, 100 mM 2-deoxyglucose (2-DG) was added to shut down glycolysis.
19
This last injection resulted in a decreased ECAR confirming that the recorded ECAR was only due
to glycolysis. Rv1813c expressing cells and native Hela show similar in ECAR profiles.
In (C) and (D), HeLa cells were transiently transfected with pCI-myc-Rv1813c or empty vector and
analyzed 48h later by flow cytometry (C) Flow cytometry analysis of JC-1 stained cells to monitor
mitochondrial membrane potential. JC-1 Red/Green ratio were calculated and expressed as the %
of this ratio in HeLa cells. Values are means ± s.e.m. of three independent experiments. ns, not
significantly different Student’s t-test. In contrast to Dictyostelium, the mitochondrial membrane
potential is not enhanced in HeLa cells transiently expressing Rv1813c (D) Flow cytometry
analysis of MitoSox stained cells. MitoSox fluorescence was expressed as the % of fluorescence
in HeLa cells. Mitochondrial ROS production is slightly increased upon Rv1813c expression.
Values are means ± s.e.m. of three independent experiments, ** p≤0.01 Student’s t-test.
In (E) and (F), HEK293T cells were transiently transfected with pCI-myc-Rv1813c or empty vector
and analyzed 48h later by flow cytometry (E) Cell death analysis by flow cytometry of cells treated
with 0.15 mM hydrogen peroxide for 8h and 24h and stained with Annexin V and propidium iodide
(PI). Early apoptosis is characterized by no PI labeling but Annexin V cell staining due to the trans-
location of phosphatidylserine (PS) from the inner face of the plasma membrane to the cell surface
and. In contrast, late apoptosis and necrosis results in both Annexin V and PI positive staining.
Percentage of each characteristic population is indicated. Flow graphs are representative of three
independent experiments. (F) Quantification of Annexin V fluorescence in cells treated with increas-
ing amount of hydrogen peroxide for 24h shows that Rv1813c promotes cell death in response
to prolonged oxidative stress. Cells were analyzed as described in (E). Annexin V positive cells
include PI negative and positive stained cells. Values are means ± s.e.m. of three independent
experiments, ** p≤0.01 Student’s t-test.
Fig. 7 Defective Cyt-c release from mitochondria upon oxidative stress and massive Rv1813c
egress to the cytosol.
(A) Confocal microscopy analysis of HeLa cells transiently expressing myc-Rv1813.
48h post-transfection, cells were treated with 0.1 or 0.2mM hydrogen peroxide for three hours,
fixed, processed for immunofluorescence with anti-Cytochrome c (green) and anti-Rv1813c (red)
antibodies, and observed by confocal microscopy. Nuclei were stained with Hoechst (blue). White
arrows and white stars indicate cells with Cyt-c in mitochondria and cytosol respectively. Cyt-c
shows a diffuse cytosolic staining in 21% of HeLa cells submitted to a moderate oxidative stress
(0.1mM hydrogen peroxide). This effect is dose dependent, rising up to 25.7% at 0.2 mM hydrogen
peroxide. In contrast, Cyt-c translocation into the cytosol is highly reduced in cells overex-
pressing Rv1813c, with only 7.9% of cells displaying a cytosolic Cyt-c staining upon 0.1mM hy-
drogen peroxide treatment. Scale bar, 10 µm (B, C, D) Quantification of cells with Cyt-c in cytosol
20
(B), with Rv1813c in cytosol (C) and Rv1813c in cytosol but Cyt-c in mitochondria (D) upon incuba-
tion with hydrogen peroxide for three hours. Oxydative stress induces a massive exit of Rv1813c
from mitochondria and a subsequent partial inhibition of Cyt-c release in the cytosol. Values are
means ± s.e.m. of three independent experiments, with 100 cells analyzed for each condition, *
p<0.05 Student’s t-test.
SUPPLEMENTARY DATA LEGENDS
Fig. S1 1H-15N HSQC spectrum of Rv1813c
This spectrum was obtained at 800 MHz, 20°C and pH 6,8 with 0.3 mM 15N-uniformly labeled sam-
ple. Cross peak assignments are indicated using the one-letter amino acid code and number fol-
lowing the full-length protein sequence numbering.
Fig. S2 Confocal microscopy analysis of Rv1813c family members localisation in Dictyoste-
lium.
Dictyostelium cells expressing the indicated constructs were fixed, processed for immunofluores-
cence, and analyzed by confocal microscopy (Airyscan).
(A) Mitochondrial localisation of untagged Rv1813c expressed in Dictyostelium. Rv1813c was la-
beled with a rabbit pAb anti-Rv1813c antibody. Rv1813c colocalises with Mitoporin, a mitochon-
drial specific protein.
(B) Mitoporin localisation in untransfected recipient Ax2 cells. Cells were labeled with a mouse mAb
to mitoporin revealing characteristic ring shaped structures. A maximum projection of Z confocal
sections is shown on the right panel.
(C) Localization of different Rv1813c family members of M. tuberculosis and M. marinum expressed
in Dictyostelium and revealed by anti-myc labelling. Whereas M. marinum orthologs of Rv1813c
(MMA_1436 and MMA_2533) are addressed to mitochondria without any major morphological ef-
fects, Rv1269c and its ortholog MMA_4153 localization to mitochondria induces mitochondria mor-
phological defects. White arrows indicate so mitochondria with affected shapes. Bar, 5 µm.
Fig. S3 Rv1813c family localisation in HeLa cells
HeLa cells expressing the indicated constructs were fixed, processed for immunofluorescence, and
analyzed by confocal microscopy (Airyscan).
Cells were colabeled either with rabbit polyclonal anti-Rv1813c, mouse mAb anti-Cytochrome c,
and mitotracker deep red (upper panel) or anti-myc, rabbit anti-grp75 (mitochondria marker) and
mitotracker deep red (lower panels). Rv1813c and its M. marinum orthologs (MMA_1426 and
21
MMA_2533) are efficiently addressed to mitochondria in contrast to Rv1269c and its ortholog
MMA_4153. Bar, 10 µm.
Fig. S4 Rv1813c family localization in HEK293 cells
HEK293 cells expressing the indicated constructs were fixed, processed for immunofluorescence,
and analyzed by confocal microscopy (Airyscan).
Cells were labeled as described in Fig. S4. Rv1813c family members are targeted to mitochondria
in HEK293. As observed in HeLa cells, Rv1269c is barely detectable in mitochondria whereas some
faint mitochondrial staining is observed with MM_4153 expressed in HEK293. Bar, 10 µm.
Fig. S5 Mitochondrial ultrastructure in HEK293 expressing Rv1813c
Representative mitochondria ultrastructure determined by transmission electronic microscopy of
HEK293 cells transiently transfected with Rv1813c (right panel) or vector alone (left panel). Black
arrows indicate some enlarged intra-cristae spaces in Rv1813c expressing cells. Bars, 500 nm.
22
SUPPLEMENTARY TABLE 1
NMR and refinement statistics for RV1813 protein structures
NMR distance and dihedral constraints
Distance constraints
Total NOE
1516
Intra-residue
406
Inter-residue
Sequential (|i – j| = 1)
465
Medium-range (|i – j| < 4)
253
Long-range (|i – j| > 5)
392
Hydrogen bonds
84
Total dihedral angle restraints
f
82
y
82
Structure statistics
Violations (mean and s.d.)
Max. distance constraint violation (Å)
0.18 ± 0.03
Max. dihedral angle violation (º)
2.04 ± 0.48
Deviations from idealized geometry
Bond lengths (Å)
0.0118 ± 0.0002
Bond angles (º)
1.2010 ± 0.0214
Impropers (º)
1.3446 ± 0.0861
Ramachandran plot (%)
Most favoured region
84.7
Additionally allowed region
14.2
Generously allowed region
0.8
Disallowed region
0.3
Average pairwise r.m.s. deviation** (Å)
Backbone
0.66 ± 0.18
Heavy
1.26 ± 0.19
** “Pairwise r.m.s. deviation calculated among 20 refined structures for residues 31-116.”
23
Supplementary Materials and methods
Purification of recombinant 6His-Rv1813c28-143 in E. coli
E. coli BL21(DE3) strains containing pET::rv181328-143 vector were used to inoculate 1 L of LB medium
supplemented with ampicillin (100 μg/ml) and resulting cultures were incubated at 37 °C with shak-
ing until A600 reached ~0.5. Then, 1 mM final of isopropyl 1-thio-β-d-galactopyranoside was added
and growth was continued for 3 hr at 37 °C. The cells were harvested by centrifugation and the
resulting cell pellet was resuspended in buffer A (50 mM Tris-HCl pH 8.5, 150 mM NaCl, 2mM DTT).
Cells were then lysed by sonication and cell debris and insoluble materials were separated by cen-
trifugation. The pellet was then resuspended in buffer B (Buffer A + 8M Urea). After centrifugation
the supernatant was loaded into a HitrapTM IMAC HP column (Amersham biosciences), equilibrated
in buffer B and 4 % of buffer C (buffer B supplemented with 300 mM of imidazole). The column was
washed with successive applications of buffer B (approximately 30 ml in total) to remove all the
impurities and then buffer C was increased over 20 ml to 100%. Fractions containing the Rv1813c
proteins were identified by SDS-PAGE, then pooled and concentrated using a 5 K cut-off concen-
trator to a 2mg/ml concentration. The protein was dialysed against buffer A over-night at 4°C. The
refolded protein was very unstable until removal of the 6His tag using 3C protease (4h digestion at
4°C). The protein was then loaded to a Superdex 75 26/60 (Amersham biosciences) size exclusion
column, equilibrated in buffer 20 mM Na-Phosphate pH 6.2, 150 mM NaCl. Again, fractions con-
taining the protein were identified by SDS-PAGE, then pooled and stored at -20°C until required.
This protocol was carried out for all the non-labelled constructs of Rv1813c as well as for 15N and
15N -13C labelled constructs, except that the cultures were grown in a minimum media containing
15NH4Cl and 15NH4Cl/13C6-glucose as the sole nitrogen and carbon sources.
Solution structure of Rv1813c28-143
All NMR experiments were generally carried out at 25°C on Bruker Avance III 700 (1H-15N double
resonance experiments) or Avance III 500 (1H-13C-15N triple-resonance experiments) spectrometer
equipped with 5 mm z-gradient TCI cryoprobe, using the standard pulse sequences. NMR samples
consist of approximately 0.9 mM 15N- or 15N,13C-labeled protein dissolved in 25 mM NaCitrate, 150
mM NaCl (pH 5.6) with 10% D2O for the lock. 1H chemical shifts were directly referenced to the
methyl resonance of DSS, while 13C and 15N chemical shifts were referenced indirectly to the absolute
15N/1H or 13C/1H frequency ratios. All NMR spectra were processed and analyzed with GIFA. Back-
bone and Cβ resonance assignments were made using standard HNCA, HNCACB, CBCA(CO)NH,
HNCO, and HN(CA)CO experiments performed on the 15N,13C-labeled Rv1813c 28-143 sample. NOE
cross-peaks identified on 3D [1H, 15N] NOESY-HSQC (mixing time 120 ms) were assigned through
automated NMR structure calculations with CYANA 2.1, whereas NOE on 3D [1H,13C] NOESY-HSQC
24
were assigned manually. Backbone φ and ψ torsion angle constraints were obtained from a data-
base search procedure on the basis of backbone (15N, HN, 13C’, 13Cα, Hα, 13Cβ) chemical shifts using
the program TALOS+ (Shen et al., 2009). Hydrogen bond restraints were derived using standard
criteria on the basis of the amide 1H / 2H exchange experiments and NOE data. When identified, the
hydrogen bond was enforced using the following restraints: ranges of 1.8–2.0 Å for d(N-H,O), and
2.7–3.0 Å for d(N,O). The final list of restraints, from which values redundant with the covalent ge-
ometry has been eliminated. The 30 best structures (based on the final target penalty function val-
ues) were minimized with CNS 1.2 according the RECOORD procedure (Nederveen et al., 2005)
and analyzed with PROCHECK (Laskowski et al., 1993). The rmsds were calculated with MOLMOL
(Koradi et al., 1996). All statistics are given in Table 1. The chemical shift table was deposited in the
BMRB databank (accession number XXX) and the coordinates have been deposited in the PDB:
PDBXXX.
Antibodies
The following primary antibodies were used in this study: mouse anti-Myc (Invitrogen, #13-2500,
1∶200 for immunofluorescence, 1:500 for immunoblot), mouse anti-cytochrome c (clone 6H2.B4,
BD PharMingen, 1:500 for immunofluorescence), mouse anti-Dictyostelium Mitoporin (70-100-1;
1:2000 for immunofluorescence and immunoblot) (Troll et al., 1992), rabbit anti-Rv1813c raised
using recombinant Rv1813c (ProteoGenix SAS, Schiltigheim, France) (1:2000 for immunofluores-
cence, 1:5000 for immunoblot), rabbit anti-Grp75 (D13H4, XP #3593, Cell Signalling, 1:100 for im-
munofluorescence), rabbit anti-EHD (Dias et al., 2012; 1:4000 for immunoblot). Secondary antibod-
ies used for immunoblotting were horseradish peroxidase (HRP)-conjugated donkey anti-mouse
IgG (H+L) (#715-035-151) and HRP-conjugated donkey anti-rabbit IgG (H+L) (#715-035-152) (Jack-
son ImmunoResearch). Secondary antibodies used for immunofluorescence were Alexa-Fluor-568-
conjugated goat anti-mouse IgG (H+L) (#A11031), Alexa-Fluor-594-conjugated donkey anti-rabbit
IgG (H+L) (#A21207), Alexa-Fluor-488-conjugated goat anti-rabbit IgG (H+L) (#A11029) and Alexa-
Fluor-488-conjugated donkey anti-rabbit IgG (H+L) (#A21206) (ThermoFisher Scientific, Illkirsh,
France). All secondary antibodies were used at 1:500 for immunofluorescence. Prolong Golf Anti-
fade and Hoechst 33342 (#62249) were purchased from Molecular Probes (ThermoFisher Scientific,
Illkirsh, France).
Preparation of Mycobacterium tuberculosis culture
M. tuberculosis was grown in Middlebrook 7H9 liquid medium supplemented with 10% (v/v) Albu-
min-Dextrose Complex (ADC), 0.2% (v/v) glycerol and 0.1% Tween 80 (w/v), at 37°C in a roller in-
cubator. Bacterial growth was followed by measurement of absorbance at 580 nm, using a spec-
trophotometer, or by colony-forming unit (CFU) counting on 7H10 agar.
25
Mycobacterial cell fractionation
Mycobacteria cell fractionation was done as described else were (O.Turapov, Cell Report, 2018).
Briefly, cells were lysed in a buffer that contained 20 mM TrisHCl, pH 8.0, 150 mM NaCl, 20 mM
KCl, 10 mM MgCl2. Bacterial culture was homogenized with a Minilys homogenizer (Bertin Instru-
ments) using glass beads. A cocktail of proteinase/phosphatase inhibitors (Roche, UK) were used
in all buffers. Lysates were centrifuged for 1 hour at 27,000 x g, the pellets were washed in a car-
bonate buffer (pH 11) and used as a cell wall material. The supernatant was centrifuged again for 4
hours at 100,000 x g. The supernatants from this step was used as cytoplasmic fraction and the
pellets (membrane fractions) were washed once in carbonate buffer, pH 11 and twice in TBS buffer.
Proteins from cellular fractions were separated on SDS-PAGE. The purity of fractions was con-
firmed by the detection of diagnostic proteins as described below.
Protein Electrophoresis and Western Blot
Proteins were separated on 4%–20% gradient SERVA gels and transferred onto a nitrocellulose
membrane using a Trans-Blot® Turbo Transfer System (Bio-Rad) according to the manufacturer’s
instruction. SignalFire Elite ECL Reagent (Cell Signalling, UK) were used to visualize proteins on C-
DiGit Chemiluminescent Blot Scanner (LI-COR Biosciences), according to the manufacturer’s in-
structions. All the secondary antibody were from Cell Signalling, UK. Diagnostic proteins were used
for all the cellular fractions: GlnA (membrane protein), RpfB (membrane and cell wall protein) and
FtsZ was used as a cytoplasmic fractions marker.
Cell culture and transfection conditions
D. discoideum strain Ax2 was grown at 22oC in HL5c medium supplemented with 18 g/L Maltose
(Formedium, Norfolk, United Kingdom). For ectopic expression in Dictyostelium, Rv1813c family
coding sequences with Dictyostelium optimized codons (IDT, Integrated DNA Technologies, Inc.,
Coralville, Iowa 5224, USA) were cloned into pDXA-3C-myc (Manstein et al., 1995). Plasmids were
linearized by ScaI and transfected by electroporation as described (Cornillon et al., 2000). Clones
were selected in 5µg/mL G418.
HeLa (ATCC CRM-CCL-2) and HEK-293T (ATCC CRL-3216) cells were maintained in DMEM, high
glucose (Dulbecco's Modified Eagle Medium) containing 5% and 10% heat-inactivated foetal bo-
vine serum, respectively, and supplemented with GlutaMAXTM (Gibco Life Technologies), penicillin
(100 units/mL), and streptomycin (100 µg/mL). Transfections of HeLa and HEK-293T cells were
performed using JetPEITM transfection reagent (PolyPlus-Transfection, Ozyme, Saint Quentin,
France), according to the manufacturer. Cells plated one day before transfection were incubated
with JetPEITM -DNA complexes (N/P=5), and after 5h the medium was changed. All assays were
performed 48h post-transfection.
26
For confocal microscopy analysis, HeLa or HEK-293T cells were seeded on glass coverslips coated
with 0.001% poly-L-Lysine (# P4707, Sigma). For localization, Rv1813c family coding sequences
with human optimized codons were cloned into the mammalian expression vector pCI (a kind gift
of Dr. Solange Desagher, IGMM, Montpellier, France). Cells on glass coverslips were transfected in
a 24-well culture plate and analysed 48h later. For mitochondrial membrane potential, mitochondrial
ROS and oxidative stress studies, cells were transfected on 6-well culture plates. After 24h, resus-
pended cells were pooled and plated either on glass coverslips for confocal microscopy or on 6-
well culture plates at a density of 2-3.105 cells/well for FACS analysis. For extracellular flux analysis,
HeLa cells seeded into five 100-mm tissue culture dishes were transfected with Rv1813c DNA
cloned into pMACS 4-IRESII vector (Miltenyi Biotec, France), allowing Rv1813c co-expression with
a truncated CD4 surface marker. After 24h, EDTA resuspended cells were pooled and CD4 positive
cells selected through magnetic cell sorting (MACS) (see below).
Mitochondria isolation and biochemical treatments
Mitochondria were isolated as described (Aubry and Klein, 2006). Briefly Dictyostelium cells were
washed in ice-cold buffer A (20 mM HEPES pH7, 1 mM EDTA, 250 mM Sucrose, proteinase inhib-
itors), resuspended at a cell density of 3x108 cells/mL, and broken with a ball bearing homogenizer
(8.02 mm bore, 8.002 mm ball; 20 strokes). Unbroken cells were removed by low speed centrifuga-
tion (5 min, 1500 g). The supernatant was next centrifuged for 15 min at 16,000 g. The pellet was
resuspended in buffer A and the centrifugation repeated to yield the enriched mitochondria fraction.
For further subcellular fractionation, this fraction was further centrifuged at 100,000 g for 1h. Triton
X-114 phase fractionation was performed as described (Bordier, 1981). Briefly, mitochondria were
incubated for 20 min at 4oC in 10 mM Tris-HCl pH7.4, 150 mM NaCl and 1% Triton X-114. Samples
were loaded on a 6% sucrose cushion, incubated at 30oC for 3 min for condensation, and centri-
fuged at 300 g for 3 min at room temperature. Supernatants were adjusted to 1% Triton X-114 and
the procedure repeated. Detergent and aqueous phases were analysed by western blotting.
For Carbonate extraction of integral membrane proteins, mitochondria were incubated for 30 min
at 4oC in 0.1 M Na2CO3 pH11.5 and centrifuged for 30 min at 100,000 g. Pellets were resuspended
in buffer A. Proteins in resuspended pellets and supernatants were precipitated with 15% TCA and
resuspended in SDS page loading buffer. For high salt washes, intact mitochondria were incubated
in 10 mM Tris-HCl pH7.3, 250 mM Sucrose, 200 mM KCl and incubated for 30 min at 4oC. Mito-
chondria were then centrifuged for 10 min at 16,000 g. Pellets and supernatants were treated as
above. For proteinase K digestions of mitochondrial peripheral membrane proteins, mitochondria
in 20 mM HEPES pH7, 250 mM Sucrose, 100 mM KCl, 2 mM MgCl2, 1mM KH2PO4 were incubated
with 100 µg/mL proteinase K for 30 min at 4oC ± 1% Triton X100. Samples were then treated with
TCA for protein precipitation. To break selectively mitochondrial outer membranes, mitochondria
were resuspended in hypotonic buffer (2 mM HEPES pH7, 5 mM KCL, proteinase inhibitors) for 30
27
min at room temperature. After centrifugation at 16,000 g for 10 min, pellets and supernatants were
treated with TCA as above.
Immunocytochemistry
Dictyostelium cells were applied on glass coverslips for 3h, and then fixed with 4% paraformalde-
hyde for 30 min, washed and permeabilized for 2 min in -20oC methanol. Cells were incubated with
the indicated antibodies for 1h, washed, and then stained with appropriate fluorescent secondary
antibodies for 30 min. After three washes, coverslips were mounted in Mowiol. Mammalian cells
were cultured on glass coverslips and fixed with 4% paraformaldehyde in phosphate-buffered sa-
line (PBS) for 20 min. Cells were washed in Tris-buffered saline (TS; 25mM Tris pH7.4, 150mM NaCl)
for 10 min. After permeabilization with 0.2% Triton X-100 in TS for 4 min, non-specific binding was
blocked with 0.2% gelatin from cold water fish skin (Sigma-Aldrich, France) in TS for 30 min. Cells
were incubated with primary antibodies in blocking buffer for 1h and were then washed 3 times
with 0.008% TritonX-100 in TS for 10 minutes. Cells were incubated for 30 minutes with Alexa-
Fluor-labelled secondary antibodies in blocking buffer. After rinsing in washing buffer, cell nuclei
were stained with 1 µg/ml Hoechst in TS for 5 minutes. Finally, coverslips were mounted with Pro-
long Gold Antifade (#P36934 Thermo Fisher Scientific). Slides were examined under a Leica TCS
SPE confocal microscope equipped with a 40X/1.15 or 63X/1.33 ACS APO oil-immersion objective
or a Zeiss LSM880 AiryScan confocal microscope equipped with a 40X/1.4 or 63x/1.4 Oil Plan-
apochromat DIC objective. Fluorescence images were adjusted for brightness, contrast and colour
balance by using the ImageJ software.
Flow cytometry analysis of JC-1, MitoSox and Annexin 5/PI stained cells
Dictyostelium cells were washed in incubation buffer (2 mM Na2HPO4, 15 mM KH2PO4, 310 µM CaCl2,
500 µM MgCl2, 1.35 mM KCl, 1.8% Maltose, pH6). Cells (5x105) were incubated either in 5 µM Mi-
toSox Red or 2 µM JC-1 dye (ThermoFisher Scientific, Illkirsh, France) for 30 min at 22oC with shak-
ing, and then washed twice before FACS analysis. As positive control of JC-1 staining, 5 µM car-
bonyl cyanide m-chlorophenyl hydrazone (CCCP) was added to cells during JC-1 cell incubation.
For MitoSox red staining of HeLa cells, 2.5x105 cells resuspended in CPBS buffer (PBS, 2.67 mM
KCl, 0.5 mM MgCl2, 0.7 mM CaCl2 and 0.1% glucose) were incubated in 5 µM MitoSox red. After
20 min at 37oC with shaking, cells were washed twice in CPBS buffer before FACS analysis. JC-1
staining of HeLa cells was made according to the manufacturer recommendations. Briefly, cells
cultured in 6-well culture plates (2.5 x105/well) were incubated at 37oC in culture medium supple-
mented with 2 µM JC1. After 30min, cells were washed, resuspended in PBS, and directly analysed
by flow cytometry.To detect cell death upon oxidative stress, 2.5x105 Hela cells were resuspended
in 50 µL Annexin V buffer (10 mM HEPES pH7, 140mM NaCl, 2.5 mM CaCl2) and incubated at room
28
temperature with 5 µL Annexin V-FITC (eBioscience, Vienna, Austria) and 10 µL propidium iodide
(stock 0.1 mg/mL). After 15 min, cells were washed once in PBS before FACS analysis.
MACS enrichment of CD4-Rv1813c transfected cells
MACS enrichment of transfected cells was done with MACSelect Transfected Cell Selection kit
from Miltenyi Biotec, according to the supplier. Briefly, HeLa cells were transfected with empty
pMACS4-IRESII or pMACS4-IRESII-Rv1813c plasmids allowing expression of truncated CD4 cell
surface marker alone or in combination with Rv1813c respectively. After 24h, ~107 cells were
washed, dissociated in ice cold PBS containing 5 mM EDTA, centrifuged at 200 g for 10 minutes
at 4°C, and resuspended in 320 µl ice-cold de-gassed PBS supplemented with 0.5% bovine serum
albumin and 5 mM EDTA (PBE). Magnetic labelling of the transfected cells was achieved by incu-
bating cells with 80 μl of anti-CD4 coupled MACSelect MicroBeads on ice for 15 minutes. Volume
was adjusted to 2 ml with PBE and cells were subjected to magnetic separation using LS column
(Miltenyi Biotec) and MACS separator. After three washes with 3 ml of PBE, cells were flushed out
with 5 ml of PBE. To increase the purity of the magnetically labelled fraction, magnetic separation
was repeated once on a second LS column. After the final wash, cells were flushed out with 5ml of
cell culture medium, counted and seeded at a density of 1.85x104 cells/well on XF96 cell culture
microplates (Seahorse, Agilent Technologies, France) previously coated with 0.1mg/ml poly-D-Ly-
sine (#P7280, SIGMA) or on glass coverslips to evaluate the level of MACS enrichment of trans-
fected cells by immunofluorescence. Cells were incubated at 37°C and analysed 24h later using
the Seahorse XF96 extracellular flux analyser or by confocal microscopy.
Extracellular flux analysis
Cells plated the day before on XF96 cell culture microplates were washed with pre-warmed cell
culture medium 5h before analysis to eliminate dead cells. Extracellular Flux analysis was performed
using Seahorse XF Extracellular Flux analyser, allowing simultaneous measurement of oxygen con-
sumption rate (OCR) and extracellular acidification rate (ECAR). Mitochondrial respiration and gly-
colytic function of the cells were measured using Cell Mito Stress Test Kit (#103015-100) and Cell
Glycolysis Stress Test Kit (#103020-100), respectively (Agilent Technologies, France). Cells were
incubated in Seahorse XF DMEM pH7.4 (#103575-100, Agilent) supplemented with 1 mM sodium
pyruvate, 2 mM glutamine and with 10 mM glucose (Cell Mito Stress Test Kit) or without glucose
(Cell Glycolysis Stress Test Kit) in a 37°C incubator without CO2 for 1h prior to the assay. After
calibration and three initial measurements at baseline, different perturbing chemicals corresponding
to each kit were sequentially injected, and three successive measurements were taken after each
injection.
Transmission electron microscopy
29
MACS enriched cells on glass coverslips were successively fixed with 2.5% gluteraldehyde in 0.1
M cacodylate buffer pH 7.4, washed with cacodylate buffer, post-fixed in 1% osmium tetroxide in
cacodylate buffer, washed with distilled water, and finally incubated in 1% uranyl acetate. Dehy-
dration was performed through acetonitrile series. Samples were impregnated first in epon 118:
acetonitrile 50:50, and twice in 100% epon. After overnight polymerization at 60˚C, coverslips were
detached by thermal shock with liquid nitrogen. Polymerization was then prolonged for 48h at 60˚C.
Ultrathin sections of 70 nm were cut with a Leica UC7 ultramicrotome (Leica microsystems), coun-
terstained with lead citrate and uranyl acetate prepared in ethanol. Sections were observed in a
Jeol 1200 EXII transmission electron microscope. All chemicals were from Electron Microscopy
Sciences (USA) and solvents were from Sigma. Images were processed using the Fiji software.
TT TT
Rv1813c
1 1 0 2 0 3 0 4 0 5 0 6 0 7 0
Rv1813c
RRR A L YGAIAY G G
M T N MAAA L AL I VP V I PS
I .. L T G GA GL...G L T DAHLANGSMSEVMMSEIAGLPIPP IH A AS
Mmar1426
RRR A L YGAIAY G G
M T N LIVV L AL L L A I PN
M .. L A T AA GL...G L SP GAHLYDDSI........TGRIVAP TY G VN
Mmar2533
RRR A L YGAIAY G G
M LA A V A L G I I PN
T....R I T T GAT GLMFIG A T S GANMDRAVMSEMG..MLPEGPVPL VH A AF
Rv1269c
RRR A L YGAIAY G G
M T T VAVA V AA V AP A GN
T MI L F G AT AT...T T APA.....................NA DV S SW
Mmar4153
RRR A L YGAIAY G G
T S VAVA V A L AP A A G
.M TN R L S AT TAT...T T V ......................DA DQ S D SW
�
TT TT
Rv1813c
8 0 9 0 1 0 0 1 1 0 1 2 0 1 3 0 1 4 0
Rv1813c
AE A CG CKV F CGAVA GG G T A DA L GG I WACN
K T A L VS T Y A V R N
AWHQR P R QV EK DKT R R YNGSK Q T L RR ED N E R V .
Mmar1426
AE A CG CKV F CGAVA GG G T A DA L GG I WACN
R TRA LK LSS M R G W T
SWNNR Q SS L VEG VR FDGSARH V R RQ ED RF E N
Mmar2533
AE A CG CKV F CGAVA GG G T A DA L GG I WACN
K SR LK L N N Y A I R G W N
ARRFT FG QA Q LDS I R YNNLK Q S W LS QQ D V .
Rv1269c
AE A CG CKV F CGAVA GG G T A DA L GG I WACN
R TRA VK LTS T Y A L K G Y T
SWDYP A AT S YSD A ANDRA Q V P LA MK T D .
Mmar4153
AE A CG CKV F CGAVA GG G T A DA L GG I WACN
R TRA VK LTT T F A L K G Y T
ASHYP A AT L YSD A ADGKT E V P LS MK S D .
1 1 2 2�
β1
β2
α1
β3
β4
β5
α2
β6
A
B
Pro54
Pro54
Pro54
Cter
Cter
Cter
1
1
2
2
SIGNAL PEPTIDE
17
20
25
35
48
63
75
11
Membrane
Cytoplasm
Culture FIltrate
Cell Wall
Anti-FtsZ
17
20
25
35
48
63
75
11
Membrane
Cytoplasm
Culture FIltrate
Cell Wall
17
20
25
35
48
63
75
11
Membrane
Cytoplasm
Culture FIltrate
Cell Wall
Anti-GlnA
Membrane
Cytoplasm
Culture FIltrate
Cell Wall
17
20
25
35
48
63
75
11
Figure 3
A
C
B
myc-Rv1813c
medial plan
Mitotracker
merge
Zoom
myc-Rv1813c
Mitoporin
medial plan
merge
Max projection
myc-Rv1813c
myc-Rv1813c-49-143
myc-Rv1813c-28-56
Figure 4
A
P1
S1
P2
S2
16,000 g
100,000 g
cell disruption
1,500 g
P0
S0
E
myc-
Rv1813c
S
Na2CO3
P
C
myc-
Rv1813c
W
TX
Triton X114
-
+
-
+
+
+
-
-
myc-
Rv1813c
PK
Triton
B
myc-
Rv1813c
S0
P1
S1
P2
S2
Mitoporin
EHD
KCl
P1
S1
-
+
-
+
myc-
Rv1813c
F
D
H
NT
P1
S1
Hypotonic
Medium
myc-
Rv1813c
I
0
20
40
60
80
100
MitoSox fluorescence (% of control)
Ax2
Ax2 + myc-Rv1813c
J
K
G
0
20
40
60
80
100
% viable cells
Hydrogen peroxide (mM)
Ax2
Ax2 +
myc-
Rv1813c
0
0.2
0.4
0.6
0.8
0
20
40
60
80
100
120
140
160
JC-1 Red/Green ratio (% of control)
Ax2
Ax2 + myc-Rv1813c
Ax2 +
myc-Rv1813c
Ax2
untreated
0.4 mM H2O2
Mitoporin
Mitoporin
Mitoporin
Mitoporin
Mitoporin
IMS
Matrix
Crista
intra-crista
space
Cytosol
HeLa + Rv1813c
HeLa + Rv1813c
HeLa HeLa +
Rv1813c
0
20
40
60
80
Intra-crista space (nm)
****
Figure 5
Cytochrome c
myc-Rv1813c
merge
Max projection
C
D
F
E
A
B
Punctate Intermediate
0
20
40
60
% cells
HeLa
HeLa +
Rv1813c
Filamentous
HeLa
0
20
40
60
80
100
120
MitoSox fluorescence (% of control)
HeLa
HeLa + myc-Rv1813c
**
0
1
2
3
4
5
6
7
8
9
ECAR (mpH/min/µg protein)
0
10
20
30
40
50
60
70
80
Oligomycin
Glucose
2-DG
Time (min)
FCCP
0
5
10
15
20
25
0
10
20
30
40
50
60
70
80
Time (min)
HeLa
HeLa +
myc-
Rv1813c
Rotenone + antimycin
Oligomycin
A
B
C
D
E
F
0.15
0.3
**
0
10
20
30
40
50
60
70
80
90
100
0
0.075
% Annexin V positive cells
ns
ns
HEK293
HEK293 +
Rv1813c
Hydrogen peroxide (mM)
ns
0
20
40
60
80
100
JC-1 red/green ratio (% of control)
HeLa
HeLa + myc-Rv1813c
100
101 102
103 104
105
100
101 102
103 104
105 100
101 102
103 104
105
100
101 102
103 104
105
Annexin V
Propidium Iodide
HEK293
8h H2O2
8h H2O2
24h H2O2
24h H2O2
HEK293 + Rv1813c
0.7%
3.1%
2.7%
93.5%
0.5%
8.9%
6.3%
84.3%
3%
43.1%
6.1%
47.7%
5.2%
73.2%
5.5%
16.1%
HeLa
HeLa +
myc-
Rv1813c
Figure 6
Figure 7
A
C
D
B
HeLa
HeLa +
Rv1813c
0
5
10
15
20
25
30
35
% Cells with cytochorme c in cytosol
*
*
0.1
0.2
Hydrogen peroxide (mM)
% Cells with Rv1813c in cytosol
0
5
10
15
20
25
30
35
40
45
HeLa
untreated
0.1 mM H2O2
0.2 mM H2O2
Cytochrome c
HeLa + Rv1813c
Rv1813c
Cytochrome c
merge
0
5
10
15
20
25
30
35
40
45
% Cells with Rv1813c in cytosol
and Cyt-c in mitochondria
0.1
0.2
Hydrogen peroxide (mM)
0.1
0.2
Hydrogen peroxide (mM)
124 122 120 118 116 114 112 110 108
PPM
126
124
120 118 116 114 112 110 108
122
126
Figure S2
A
B
C
Ax2 + Rv1813c
Rv1813c
pAb anti-Rv1813c
Max projection
Max projection
Ax2 Mitoporin
Max projection
myc-MMA_1426
myc-MMA_2533
myc-Rv1269c
myc-MMA_4153
Max projection
Mitoporin
Ax2 Mitoporin
medial plan
merge
anti-myc
anti-grp75
merge
mitotracker
myc-Rv1813c
Rv1813c
myc-Rv1269c
anti-Rv1813c
anti-cytochrome c
merge
myc-MMA_1426
myc-MMA_2533
myc-MMA_4153
mitotracker
Empty vector
Empty vector
anti-myc
anti-grp75
merge
mitotracker
myc-Rv1813c
Rv1813c
anti-Rv1813c
anti-cytochrome c
merge
mitotracker
myc-Rv1269c
myc-MMA_1426
myc-MMA_2533
myc-MMA_4153
Figure S5
HEK293+
myc-Rv1813c
HEK293
| 2021 | A effector targets mitochondrion, controls energy metabolism and limits cytochrome c exit | 10.1101/2021.01.31.428746 | [
"Martin Marianne",
"deVisch Angelique",
"Barthe Philippe",
"Turapov Obolbek",
"Aydogan Talip",
"Heriaud Laurène",
"Gracy Jerome",
"Mukamolova Galina V.",
"Letourneur François",
"Cohen-Gonsaud Martin"
] | null |
The fine-scale recombination rate variation and associations
with genomic features in a butterfly
Aleix Palahí i Torres1,∗, Lars Höök1, Karin Näsvall1, Daria Shipilina1, Christer Wiklund2, Roger Vila3, Peter Pruisscher1 and Niclas
Backström1
1Evolutionary Biology Program, Department of Ecology and Genetics (IEG), Uppsala University, Sweden. Norbyvägen 18d, SE-752 36, Uppsala, Sweden
2Department of Zoology: Division of Ecology, Stockholm University, Svante Arrhenius väg 18B, SE-106 91 Stockholm, Sweden
3Butterfly Diversity and Evolution Lab, Institut de Biologia Evolutiva (CSIC-UPF), Barcelona, Spain
∗Corresponding author: aleix.palahi@ebc.uu.se
Abstract
Genetic recombination is a key molecular mechanism that has profound implications on both micro- and macro-evolutionary processes.
However, the determinants of recombination rate variation in holocentric organisms are poorly understood, in particular in Lepidoptera (moths
and butterflies). The wood white butterfly (Leptidea sinapis) shows considerable intraspecific variation in chromosome numbers and is a
suitable system for studying regional recombination rate variation and its potential molecular underpinnings. Here, we developed a large whole-
genome resequencing data set from a population of wood whites to obtain high-resolution recombination maps using linkage disequilibrium
information. The analyses revealed that larger chromosomes had a bimodal recombination landscape, potentially due to interference between
simultaneous chiasmata. The recombination rate was significantly lower in subtelomeric regions, with exceptions associated with segregating
chromosome rearrangements, showing that fissions and fusions can have considerable effects on the recombination landscape.
There
was no association between the inferred recombination rate and base composition, supporting a negligible influence of GC-biased gene
conversion in butterflies. We found significant but variable associations between the recombination rate and the density of different classes
of transposable elements (TEs), most notably a significant enrichment of SINEs in genomic regions with higher recombination rate. Finally,
the analyses unveiled significant enrichment of genes involved in farnesyltranstransferase activity in recombination cold-spots, potentially
indicating that expression of transferases can inhibit formation of chiasmata during meiotic division. Our results provide novel information
about recombination rate variation in holocentric organisms and has particular implications for forthcoming research in population genetics,
molecular/genome evolution and speciation.
Keywords: Lepidoptera, recombination rate, wood white, linkage disequilibrium, transposable elements, Leptidea
Introduction
The meiotic division process allows sexually reproducing or-
ganisms to generate haploid gametes. During meiosis, double-
strand breaks are induced into DNA, and a proportion (e.g. 5%
in Arabidopsis thaliana, 10% in Mus musculus) of those are re-
solved as crossovers (Choi and Henderson 2015, Moens et al.
2002), leading to novel combinations of maternal and paternal
chromosome segments. It is well established that the frequency
and genomic distribution of crossovers can influence both micro-
and macro-evolutionary processes – detailed knowledge about
the recombination landscape is therefore key for understand-
ing the relative importance of different proximate and ultimate
mechanisms affecting genome evolution, generation and main-
tenance of genetic diversity, adaptation and speciation (Dapper
and Payseur 2017, Stapley et al. 2017, Peñalba and Wolf 2020).
The frequency of recombination events resolved as crossovers
(from here on recombination) can vary both at the inter- (Stap-
ley et al. 2017, Smukowski and Noor 2011) and intra-specific
level (Samuk et al. 2020), as well as between individuals within
populations (Johnston et al. 2016) and within individuals over
time (Stapley et al. 2017, Peñalba and Wolf 2020). It is well
established that the recombination rate has a genetic component,
but the rate can also be influenced by environmental factors (i.e.,
recombination is partly plastic; Peñalba and Wolf 2020). Recom-
bination rate can also vary considerably across chromosomes
and chromosome regions in many species, and mapping this
variation may shed light on the mechanistic control of where
in the genome recombination is initiated. The reasons for such
regional recombination rate variation have been studied in detail
in a few organism groups and a handful of consensuses have
been reached. First, the size of a chromosome can affect the
recombination rate, mainly because correct segregation seems
to be dependent on at least one recombination event per chro-
mosome arm in many species (Pardo-Manuel de Villena and
Sapienza 2001, Smith and Nambiar 2020). A lower recombi-
nation rate on the sex-chromosomes than on the autosomes is
also a commonly observed pattern and this difference can poten-
tially be attributed to low sequence homology as a consequence
of general degeneration of sex-limited chromosomes (Y or W)
(e.g. Bergero and Charlesworth 2009). Within chromosomes,
the location of recombination events might be determined by
preferential initiation of double-strand breaks close to the telom-
1
2
The fine-scale recombination rate variation and associations with genomic features in a butterfly
eres and where the chromatin in general is open and accessible
(e.g. Haenel et al. 2018, Gray and Cohen 2016). Conversely, the
recombination rate is usually suppressed within and around
the highly heterochromatic centromeres (Dapper and Payseur
2017, Stapley et al. 2017). Furthermore, physical interference
between multiple chiasmata may lead to regional differences in
the recombination rate (Gray and Cohen 2016, Peñalba and Wolf
2020). In mammals, the gene PRDM9 mediates recombination
by binding to specific sequence motifs, which explains that most
recombination events occur in a limited portion of the genome,
i.e. recombination hot-spots (e.g. Myers et al. 2005, Grey et al.
2011). However, most vertebrate and all evertebrate lineages
lack a functional copy of PRDM9 and recombination initiation
must hence be mediated by other factors in these species.
We expect the regional variation in recombination rate to be
associated with genomic features, the efficiency of selection and
the levels of genetic diversity (Dolgin and Charlesworth 2003,
Petrov et al. 2011). Such regional variation in the efficiency of
selection can for example affect the distribution of transposable
elements (TEs), which tend to accumulate in regions of low re-
combination rate in both animals (Bartolomé et al. 2002) and
plants (Xu and Du 2014). In addition, the frequency of recombi-
nation events can also affect nucleotide composition as a direct
consequence of GC-biased gene conversion (gBGC), a process
that facilitates the fixation of strong (G and C) over weak nu-
cleotides (A and T) during the double-strand break repair step
(Duret and Galtier 2009). Although genome-wide estimates of
the recombination rate and large-scale variation landscapes have
been obtained for many species, detailed recombination maps
are still mainly limited to model organisms and domesticated
species, and little is known about recombination rate variation
in natural populations. Hence, a broader taxonomic sampling
will be needed to get a more complete picture of what drives
recombination rate variation within genomes and between lin-
eages. This applies not the least to holocentric organisms which
have chromosomes without localized centromeres (Suomalainen
1953), such as Lepidoptera (butterflies and moths), where the
research on causes and consequences of recombination rate vari-
ation is still in its infancy.
Here, we used a large whole-genome re-sequencing data set
from a Swedish population of the wood white butterfly (Leptidea
sinapis) to characterize the fine-scale variation in recombination
rate and assess potential associations between recombination
rate, nucleotide composition and genomic features. So far, the
knowledge about regional recombination rate variation in Lep-
idoptera is restricted to a handful of pedigree based genetic
maps (Davey et al. 2017, Shipilina et al. 2022, Smolander et al.
2022) and this is therefore a spearheading attempt to describe
the fine-scale variation in recombination rate and the potential
associations with genomic features in a butterfly. The wood
white is widely distributed across western Eurasia and shows
extreme intraspecific variation in chromosome numbers, with
an increasing number of chromosomes in a cline-like pattern
from 2n ∼ 56 − 60 in the northern (Scandinavia) and eastern
(central Asia) parts of the distribution range to 2n ∼ 106 − 110 in
the south-western (Iberia) part of the distribution range (Dinc˘a
et al. 2011, Lukhtanov et al. 2018). Hence, some wood white
populations differ significantly from the ancestral Lepidoptera
chromosome number of n = 31 (Robinson 1971, Ahola et al. 2014).
Previous studies suggest that recurrent chromosome fissions and
fusions underlie this variation (Dinc˘a et al. 2011, Talla et al. 2017)
and that at least a handful of fission/fusion polymorphisms seg-
regate in the Swedish population (Höök et al. 2022). The rapid
karyotype evolution in wood whites provides a unique system
for characterizing the regional variation in recombination rate
in a natural population of a holocentric species, and combining
it with investigating the effects of segregating chromosome re-
arrangements on the recombination landscape. Given the key
role of recombination in both micro- and macro-evolutionary
processes, our results will be important for understanding molec-
ular mechanisms and evolutionary forces affecting genome evo-
lution and divergence processes in holocentric organisms in
general.
Results
Demographic history inferences
The demographic trajectories inferred separately for each chro-
mosome jointly revealed the demographic history of the Swedish
population of Leptidea sinapis. These demographic trajectories
shared three main features. Firstly, a maximum effective popula-
tion size (Ne) of around 106 approximately 10,000 generations
before present (BP), preceded by a period of high Ne and a
slight increase matching with the end of the Last Glacial Pe-
riod. After this time point, Ne started to decline exponentially
until stabilizing about 100 generations BP. In the most recent
past, Ne remained constant. Contemporary estimates of Ne os-
cillate between 103 and 2 ∗ 104 for the different chromosomes
(Supplementary Figure 1).
Recombination rate variation and distribution
The estimated genome-wide recombination rate was 7.37
cM/Mb, with measurements for individual chromosomes rang-
ing between 3.5 - 15.3 cM/Mb. There was a marginally non-
significant (Spearman ρ = -0.292, p = 0.06) negative association
between the recombination rate and chromosome length (Sup-
plementary Figure 2). Autosomes (7.65 cM/Mb) showed on
average a higher recombination rate than the Z-chromosomes
(7.03 cM/Mb), although this difference was not significant
(Wilcoxon’s test, W = 41, p = 0.92).
Visual inspection of the variation in recombination rate re-
vealed a considerably reduced recombination rate towards chro-
mosome ends (Figure 1A). A formal analysis showed that the
recombination rate in the subtelomeres (last 5 100 kb windows
at each end of the chromosomes) was significantly reduced (2.46
cM / Mb) compared to the 100 kb windows located in proximal
positions of the chromosomes i.e., outside of the subtelomeric
regions (Wilcoxon’s test, W = 1,310,856, p-value = 3.18 ∗ 10−66).
It should be noted that this does not necessarily reflect a low
recombination rate in the telomeres, but rather in the subtelom-
eric regions, since telomeric repeats were not covered in the
wood white genome assembly (Höök et al. 2022). We also an-
alyzed each chromosome separately and found a reduction in
recombination rate in subtelomeric regions in 50 out of the 58
chromosome ends (29 chromosome pairs, 2n = 58; Supplemen-
tary Table 1). Out of the 8 chromosome ends that did not show a
reduced recombination rate, four match observations of segregat-
ing fission/fusion polymorphisms – these involve chromosome
pairs 18 and 25, 11 and 26, and 5 and 27, respectively (Höök et
al. 2022).
Besides a significantly reduced recombination rate in sub-
telomeric regions in almost all chromosomes, we also found
that regional variation differed in other respects. First, some
chromosomes showed an obvious unimodal distribution for the
Palahí et al.
3
Figure 1 Recombination rate estimations and hot-spot determination. (A) The 1Mb-scale estimates of recombination rate in 1Mb
windows are shown in cM/Mb for all chromosomes, ordered by decreasing length. (B) Detailed recombination map for chromo-
some 24, with all oscillations in recombination rate inferred by pyrho in grey. The orange line represents the 10x background re-
combination rate, the threshold used to determine the minimum recombination rate of local hot-spots. Red lines underneath the
plot indicate the presence of a recombination hot-spot according to our defining parameters. See Materials & Methods for a more
detailed description of the parameters.
4
The fine-scale recombination rate variation and associations with genomic features in a butterfly
recombination rate, with a maximum value in the central region
and a progressive decrease towards the terminal regions. This
was a frequent (but not exclusive) pattern for the shorter chro-
mosomes (Figure 1A). In contrast, the recombination rate was
bimodally distributed along some chromosomes, with a central
region of reduced recombination in addition to the reduction in
subtelomeric regions. This was the most commonly observed
pattern for the larger chromosomes (Figure 1).
Recombination hot-spot and cold-spot identification
Informed by coalescent simulations, we developed thresholds
for identification recombination hot- and cold spots, which take
into account both specific demographic history of the popula-
tion of interest and the potential stochasticity of our method of
recombination rate inference (see methods for further details).
Based on these thresholds, a total of 3,124 recombination hot-
spots were classified (Figure 1B). The hot-spots had an average
length of 1,656 bp and the mean recombination rate within hot-
spots was 94.1± 62.5 cM/Mb. The highest estimated rate in any
hotspot was 708 cM/Mb. The average recombination rate for
hot-spots represented an approximate 13-fold increase over the
genome-wide recombination rate, but hot-spots only constituted
5.2 Mb (0.87%) of the genome. Recombination hot-spots were
found at a significantly lower frequency in the terminal 10%
regions of the chromosomes (5% on each end); Wilcoxon’s W
= 220.5; p = 4.94 ∗ 10−4 (Supplementary Figure 3A). A higher
density of recombination hot-spots was detected on the three
Z-chromosomes (six hot-spots / Mb) when compared to the
autosomes (five hot-spots / Mb), but the difference was not
significant (Student’s t-test, t = -0.85556, df = 27, p = 0.20. Permu-
tation analysis revealed a significantly lower LINE (p = 0.02) and
LTR density (p > 0.001) and a higher SINE density (p < 0.001) in
the hot-spots, while DNA transposon density did not deviate
significantly (p = 0.49) from the genome wide average (Figure 2).
There was no enrichment of functional gene categories or spe-
cific sequence motifs in the hot-spots (p < 0.05, Supplementary
Table 2).
We also identified 1,283 recombination cold-spots, i.e., re-
gions with considerably reduced recombination rate as com-
pared to the genomic average (see methods). The average length
of the cold-spots was 30 kb, and 70 of the cold-spots were longer
than 100 kb. Despite the lower frequency of cold-spots than
hot-spots, they represented a substantially larger proportion of
the genome (38.2 Mb, 6.37%). As expected given the overall
reduced recombination rate in subtelomeric regions, cold-spots
were particularly abundant in these regions. In particular, sub-
telomeric regions contained 29.6% of the cold-spots in total,
while representing only 4.8% of the genome. The cold-spots lo-
cated outside subtelomeric regions had an average length of 25.6
kb and 46 were longer than 100 kb. This translates to a signifi-
cant enrichment in the frequency of cold-spots in the 10% most
terminal chromosomal regions compared to the more central
positions of the chromosome (Wilcoxon’s W = 3; p = 5.62 ∗ 10−4)
(Supplementary Figure 3B). Similar to the observation for hot-
spots, there was no significant difference in cold-spot frequency
between the autosomes (1.8 cold-spots / Mb) and the three Z-
chromosomes (2.5 cold-spots / Mb) (Student’s t-test, t = -0.74483,
df = 27, p-value = 0.231. Permutation analyses revealed a sig-
nificantly lower transposon density in cold-spots, consistent
across all classes (p < 0.01; Figure 3). There was also a significant
enrichment of genes related to transtransferase activity in the
cold-spot regions compared to the genome in general (Supple-
mentary Table 3), but no enrichment of specific sequence motifs
in the cold-spots (p > 0.05).
Association between recombination rate and base com-
position
To investigate potential effects of gBGC on the base composi-
tion in the wood white genome, we assessed the relationship
between the recombination rate and nucleotide composition
using a window-based approach. The analysis showed that
the recombination rate (averaged over 1 Mb windows) was not
significantly correlated with the GC content in the genome in
general (Spearman ρ = -0.07, p = 0.10) (Figure 4A).
Since we observed a significant reduction in recombination
rate in subtelomeric regions, we investigated if those regions
had deviating base composition. The analysis showed that there
was a significantly higher GC content in subtelomeric regions
(33.46%) compared to proximal chromosome regions (32.61%)
(Wilcoxon’s W = 1,109,810; p = 5.73 ∗ 10−22).
The GC content was also slightly lower within hot-spots
(32.43%) and their 5 kb flanking regions (32.45%) as compared
to the genome-wide estimates (32.65%).
Associations between recombination rate and genomic
features
We used multiple regression to investigate the relative effect of
different explanatory variables (GC content, gene density, DNA
transposon, SINE, LINE and LTR retrotransposon densities). The
regression model revealed an overall significant association be-
tween recombination rate and the explanatory variables (F(6) =
10.77, df = 603, p = 2.13 ∗ 10−11), but it explained a marginal part
of the total variation in the recombination rate (R2 = 0.10, AdjR2
= 0.09) and only SINE and LINE density were significant ex-
planatory variables (Table 1).
The recombination rate was not significantly associated with
genome-wide gene density (Spearman ρ = -0.04, p = 0.17; Figure
4B). However, when partitioning the data and running analyses
for different gene elements separately, we observed a signifi-
cantly lower (Wilcoxon’s W = 1.5971e−10, p-value < 2.2∗10−16)
recombination rate within exons (5.5 cM/Mb) and in 5’ UTR
regions (6.3 cM/Mb; Wilcoxon’s W = 1.5869 ∗ 108, p-value <
2.2∗10−16) and a significantly higher recombination rate in in-
trons (Wilcoxon’s W = 7.3462∗1012, p-value < 2.2∗10−16) com-
pared to intergenic regions (7.5 cM/Mb) (Figure 5).
The associations between recombination rate and the den-
sities of different TE classes varied considerably. For all four
classes analyzed, we found a significant correlation with the
recombination rate, but the direction and strength of these asso-
ciations varied. DNA transposons (Spearman ρ = 0.09, p = 0.03)
and SINEs (Spearman ρ = 0.29, p = 3.31∗10−13) were positively
associated, and LTRs (Spearman ρ = -0.11, p = 8.50∗10−3) and
LINEs (Spearman ρ = -0.19, p = 3.01∗10−6) negatively associated
with the recombination rate (Figure 6).
Differences between classes of transposable elements were
not restricted to the overall associations with recombination
rate.
The average recombination rate within each class of
TEs varied as well. In LINEs (6.6 cM / Mb; Wilcoxon’s W =
5.233*1012, p-value < 2.2*10-16) and DNA transposons (6.8 cM /
Mb; Wilcoxon’s W = 2.6104*1012, p-value < 2.2*10-16) the recom-
bination rate was significantly lower, and in SINEs (7.4 cM / Mb;
Wilcoxon’s W = 1.2212*1012, p-value < 2.2*10-16) and LTRs (7.8
cM / Mb; Wilcoxon’s W = 7.0468*1011, p-value < 2.2*10-16) the
Palahí et al.
5
Figure 2 Density of TEs in hot-spots. The bars show the distribution of sampled TE density means for 3,124 random 1,656 bp ge-
nomic windows (number and average length of the hot-spots). The red line indicates the observed mean TE densities in the defined
hot-spot windows.
Figure 3 Density of TEs in cold-spots. Histogram shows the distribution of permuted TE density means for 1,283 random 30 kb
genomic windows (number and average length of the cold-spots). The red lines indicate the the observed mean TE densities in the
defined cold-spot windows.
recombination rate was significantly higher than the genomic
average rate (Figure 5).
Discussion
Genome-wide distribution of the recombination rate in
wood whites
The genome-wide rate of recombination has been shown to vary
considerably between different insect species, from compara-
tively low in Diptera (< 1 cM / Mb, Beye et al. 2006) to excep-
tionally high in honeybees (19 cM / Mb; Beye et al. 2006). We
found that the genome-wide recombination rate in the Swedish
wood white population was 7.37 cM / Mb. This is slightly higher
than estimates from other lepidopterans like Bombyx mori (4.6
cM / Mb), Heliconius melpomene (5.5 cM / Mb) and H. erato (6
cM / Mb) (Tobler et al. 2005; Yasukochi 1998; Jiggins et al. 2005)
and substantially higher than in vertebrates for which estimates
range between 0.16 cM / Mb in the Atlantic trout to 3.17 cM /
Mb in chicken (Beye et al. 2006).
Although not significant at the 5% level, we found that the
recombination rate was negatively associated with chromosome
size. A clearly deviating rate was found for chromosome 16
(15.3 cM / Mb), which explains why the association between
chromosome size and recombination was non-significant. Nega-
tive associations between recombination rate and chromosome
length have repeatedly been observed in different taxonomic
groups, for example yeast (Kaback et al. 1992), humans and ro-
dents (Jensen-Seaman et al. 2004), birds (Backström et al. 2010),
cattle (Mouresan et al. 2019) and butterflies (Martin et al. 2019,
Shipilina et al. 2022). Such a relationship is expected given
that crossovers are necessary for correct segregation of chromo-
somes during meiosis (Pardo-Manuel de Villena and Sapienza
2001, Smith and Nambiar 2020). Our analysis also showed that
longer chromosomes tended to have a bimodal recombination
6
The fine-scale recombination rate variation and associations with genomic features in a butterfly
Figure 4 Associations between the recombination rate and base composition (A) and gene density (B). Correlation between parame-
ters was calculated using 1 Mb genomic windows.
Table 1 Linear regression model with six explanatory variables
included. Explanatory variables that are significantly asso-
ciated with recombination rate variation at a 1 Mb scale are
highlighted in bold.
Estimate
Std. error
t
p-value
GC
27.5203
24.6317
1.117
0.264
Gene density
-6.6062
6.7692
-0.976
0.329
DNA
7.2352
16.7373
0.432
0.666
LINE
-30.4439
7.2939
-4.174
3.44e−05
SINE
104.8162
16.0648
6.525
1.45e−10
LTR
32.1472
23.9983
1.340
0.181
rate distribution with a reduced rate at the chromosome center
and towards the chromosome ends. This pattern is in line with
findings in other taxa, both organisms with defined centromeres
where recombination is reduced (Dapper and Payseur 2017), and
holocentric species such as Caenorhabditis elegans, in which the
recombination rate has been shown to increase with the relative
distance from the center of the chromosomes (Prachumwat et al.
2004). The reduced recombination rate in the center of chromo-
somes in monocentric species is often a direct consequence of the
lack of crossing-over events in the centromeres. This explanation
is obviously not valid in holocentric lineages. Since the pattern
is restricted to larger chromosomes, a potential explanation to
the reduced rate in chromosome centers could be the occurrence
of multiple chiasmata on larger chromosomes. For example,
observations in Psylla foersteri suggest that longer chromosomes
can accommodate the formation of two simultaneous chiasmata,
while shorter chromosomes only have one (Nokkala et al. 2004).
In the cases where two chiasmata are formed in a single chromo-
some, crossover interference may prevent those from forming
near each other and tend to drive them towards opposite ends of
the chromosome (Otto and Payseur 2019). An additional, but not
mutually exclusive explanation, is the formation of the “meiotic
bouquet”, a stage in early meiosis characterized by the aggrega-
tion of chromosome ends close to the nuclear membrane, which
can drive the crossovers towards distal positions and reduce
the recombination rate in the center of the larger chromosomes
Figure 5 Recombination rate estimates in different gene ele-
ments and classes of transposable elements. The genome-wide
recombination rate (7.37 cM/Mb) is indicated with the hori-
zontal black dashed line. Orange bars indicate the 95% confi-
dence intervals.
(Scherthan et al. 1996, Haenel et al. 2018).
Besides a reduced recombination rate in the center of larger
chromosomes, we also found a reduced recombination rate to-
wards the very ends of chromosomes. This pattern was observed
for almost all chromosomes (see exceptions below), irrespec-
tive of chromosome size and type. Note that the telomeres,
which in Lepidoptera are 6-8 kb long tandem repeats of the mo-
tif (TTAGG)n (Okazaki et al. 1993, Sasaki and Fujiwara 2000)
were not assembled in the reference genome. Such decrease in
the recombination rate in the subtelomeric regions of the chro-
mosomes has previously been observed in Heliconius butterflies
(Martin et al. 2019), and also in some other organism groups
such as yeast (DuBois et al. 2002, Barton et al. 2008) and fly-
catchers (Kawakami et al. 2013). A potential explanation for this
pattern is that crossover initiation is prevented near the telom-
eres to minimize the risk for ectopic recombination between
non-homologous repeat sequences during meiosis (Smith and
Nambiar 2020). As mentioned above, our results showed that
the reduced recombination rate towards chromosome ends was
not ubiquitous across all chromosomes; eight chromosome ends
(one for each of the following chromosomes: 5, 6, 10, 11, 16, 25,
Palahí et al.
7
27 and 29) did not show a significant reduction in the recombi-
nation rate as compared to each respective intra-chromosomal
level. Four of these exceptions (one chromosome end for each
of chromosomes 5, 11, 25 and 27) coincide with recently identi-
fied fission and fusion polymorphisms segregating in the wood
white population in Scandinavia (Höök et al. 2022). These re-
sults show that fission/fusion events can have immediate effects
on the distribution of crossover events within and between chro-
mosomes.
Characterization of recombination hot-spots and cold-
spots
The total number of hot-spots (n = 3,124) identified in the wood
whites is equivalent to what has previously been observed in
for example Ficedula flycatchers using a comparable approach
(Kawakami et al. 2013). The density of hot-spots was, however,
much lower than in humans (n = 25,000 - 50,000 hot-spots in
an approximately five times larger genome) (Myers et al. 2005).
Although the specific thresholds for defining hot-spots vary be-
tween studies, they all rely on the comparison of the background
and local recombination rates, making the results reasonably
comparable. We found that the distribution of hot-spots was
similar between wood whites and humans, as hot-spots occurred
mostly outside of genes (McVean et al. 2004). This is in contrast
to birds, which show an enrichment of hot-spots within genic
and regulatory regions (Kawakami et al. 2013; Singhal et al. 2015;
Smeds et al. 2016). We found that the frequency of recombina-
tion hot-spots and cold-spots was relatively similar in the center
of chromosomes, but that the number of hot-spots decreased,
and cold-spots were more frequent towards chromosome ends.
A higher occurrence of recombination cold-spots in terminal
regions of the chromosomes has previously been observed in
yeast, which seem to lack crossovers close to chromosome ends
altogether (Su et al. 2000; Barton et al. 2003). This observation is
also in line with the observed decrease in average recombination
rate close to chromosome ends in the wood whites and again
suggests that the recombination machinery is partly blocked
from accessing the very ends of chromosomes.
We found that the recombination landscape in the wood
white was highly variable. This is in line with observations in
other organisms like humans, birds (Singhal et al. 2015) and
dogs (Axelsson et al. 2012). However, other insects like for
example D. melanogaster (Comeron et al. 2012), for which de-
tailed recombination maps are available, generally show less
pronounced recombination hot-spots. While the hot-spot lo-
cations in humans largely are determined by the presence of
sequence motifs associated with PRDM9 binding (Grey et al.
2011), little is known about what drives crossovers to occur at
specific locations in organisms that lack a functional copy of
PRDM9. In order to get preliminary information about potential
mechanistic underpinnings of recombination rate variation in
wood whites, we therefore assessed if specific sequence motifs
or gene categories were enriched in recombination hot-spots
and cold-spots. The analyses did not reveal any associations
for sequence motifs, and we found no enrichment of specific
gene categories in hot-spots. In cold-spots, however, there was
a significant enrichment of genes with functions associated to
transferase activity. Particularly interesting is the case of farne-
syltranstransferase activity, as farnesylation is a key step for the
correct attachment between the spindle and the kinetochores
in humans (Moudgil et al. 2015). It is therefore tempting to
speculate that the active expression of farnesyltranstransferase
might block the recombination machinery close to those genes.
However, since farnesyltranstransferases are located in only 13
different cold-spot regions, other forces must also underlie the
absence of recombination in many cold-spots.
Associations between recombination, nucleotide com-
position and gene content
The high-density recombination maps developed here, allowed
us to investigate potential associations between the local recom-
bination rate and different genomic features. Such information
can be used to deduce the effects of recombination on base com-
position and/or potential regulatory mechanisms modulating
the recombination landscape. A potential driver of a positive
association between recombination and nucleotide composition
is GC-biased gene conversion (gBGC), i.e., the fixation bias fa-
voring “strong” alleles (G and C) over “weak” alleles (A and
T) during meiotic recombination (Duret and Galtier 2009). This
process mimics directional selection and can lead to deviating
nucleotide composition between regions experiencing different
recombination rates. The analysis showed that the local recom-
bination rate was not associated with nucleotide composition
(GC-content) in the wood whites. We also found that recombina-
tion hot-spots had marginally lower GC-content (the opposite is
usually observed when biased fixation is a considerable force;
Kawakami et al. 2013). This is in line with a limited effect of
gBGC in Leptidea butterflies (Boman et al. 2021) and stays in
contrast to findings in several other systems like humans (Fuller-
ton et al. 2001, Meunier and Duret 2004), mice (Clément and
Arndt 2013), flycatchers (Kawakami et al. 2013) and fruit flies
(Marais et al. 2001), as well as plants (Muyle et al. 2011) and
yeast (Gerton et al. 2000, Kiktev et al. 2018). As far as we are
aware, there are no other studies that have analyzed the strength
of gBGC in butterflies outside the Leptidea genus. Hence, it is
premature to draw conclusions regarding the impact of gBGC
on nucleotide composition in Lepidoptera in general.
In many organisms, for example mouse (Paigen et al. 2008)
and different plant species (Gaut et al. 2007, Tiley Burleigh 2015),
recombination occurs more frequently in gene-dense genomic
regions, but we did not find such an association in the wood
whites. However, our data showed that the recombination rate
was significantly reduced in exons and 5’ UTR regions compared
to the introns and intergenic regions. This is in line with findings
in other insects (Wallberg et al. 2015; Jones et al. 2019), as well as
in humans (McVean et al. 2004), where recombination hot-spots
mainly occur in the vicinity of, but not within, coding and regula-
tory regions. The small but significantly elevated recombination
rate in introns compared to intergenic regions is consistent with
findings in the holocentric nematode C. elegans(Prachumwat
et al. 2004), but in contrast to the observations in for example
Drosophila (Carvalho Clark 1999) and humans (Comeron Kre-
itman 2000). Taken together, this indicates that recombination
occurs within genes in butterflies, but that crossovers are partly
inhibited in coding sequences which might lead to a slightly
elevated rate in introns.
Different TE classes show contrasting association pat-
terns with the recombination rate
Potential associations between recombination and TE densities
have mainly been investigated in organisms with defined cen-
tromeres (Kent et al. 2017), while investigations in holocentric
species are scarce (but see Lavoie et al. 2013, Baril Hayward
2022, Smolander et al. 2022). To investigate the potential asso-
8
The fine-scale recombination rate variation and associations with genomic features in a butterfly
Figure 6 Association between the recombination rate (1 Mb scale) and the proportion of four different TE classes; (A) DNA trans-
posons, (B) LTRs, (C) SINEs and (D) LINEs. The abundance of TEs was calculated as the fraction of each 1 Mb window occupied by
each specific element.
ciations between the recombination rate and genomic features
in the wood white, we used density information for TEs previ-
ously identified in the species (Höök et al. 2022). The analysis
revealed that associations between the recombination rate and
the abundance of TEs varied considerably depending on the TE
class. DNA transposons and SINEs were positively associated,
while LINEs and LTRs were negatively associated with the re-
combination rate. Under the assumption that TE insertions in
general are slightly deleterious we would expect a negative cor-
relation between the recombination rate and the abundance of
TEs (Kent et al. 2017), as a consequence of more efficient purging
of deleterious insertions in regions with a higher recombination
rate (Bartolomé et al. 2002, Wright et al. 2003). However, given
that recombination is initiated by a double-strand break, it is
possible that certain types of TEs are used as a template for the
repairing process, driving them to higher frequencies in regions
of high recombination rate (Onozawa et al. 2014). SINEs have
for example have been shown to use DNA breaks to integrate
back into the genome after replication (Singer 1982). Potential
associations between TEs and the recombination rate are hence
expected to depend on the occurrence of specific classes of TE
in the focal study system. For example, Alu elements (a sub-
family of SINEs) in humans have been shown to accumulate in
regions with elevated recombination (Witherspoon et al. 2009)
and SINEs are strongly positively associated with the recom-
bination rate in the painted lady (Vanessa cardui) (Shipilina et
al. 2022). Similarly, DNA transposons are associated with high
recombination rate in C. elegans (Duret et al. 2000). The causal-
ity of such associations between the variation in recombination
rate and the abundance of TEs is not easy to establish. In cases
where TE proliferation has deleterious fitness effects, we expect
a negative association between TE abundance and recombina-
tion rate. However, presence of Alu elements in humans has
been shown to lead to an increase in the local recombination
rate, possibly a consequence of that the Alu elements mimic
the action of short recombinogenic motifs (Witherspoon et al.
2009). It is also possible that other underlying factors affect the
TE and recombination rate distributions similarly – both the
TE proliferation and the recombination initiation machinery for
example seem to target open chromatin more easily (Kawakami
et al. 2013). In the wood whites, the average recombination rate
estimates within TE classes did not deviate considerably from
the genome-wide average. However, these comparatively minor
differences in the recombination rate can indicate differences
in the selective pressure against insertion of specific families of
TEs. This does not seem to be related with the length of the TEs,
as SINEs – which are considerably shorter than LINEs and LTR
elements – showed an average recombination rate between the
longer types.
In summary, the different TEs showed different associations
with the local recombination rate. These results are consistent
with findings in other studies and may point toward similar
determinants in holocentric organisms compared to those with
defined centromeres. However, the causality needs further study,
for example by detailed characterization of cross-over regions in
large pedigrees.
Palahí et al.
9
Materials and methods
Genome assembly
The wood white genome assembly used as reference was devel-
oped for another study (Höök et al. 2022). In brief, one mated
adult female wood white was caught in Sweden and kept in
the lab for egg laying. From the offspring, one male pupa was
sampled and flash frozen in liquid nitrogen. The sample was
divided to create a 10X Genomics Chromium Genome-library
and a Dovetail HiC-library from the same individual. For 10X
sequencing, DNA was extracted using a modified HMW salt
extraction method (Aljanabi and Martinez 1997). Tissue for HiC-
sequencing was disrupted in liquid nitrogen. Library prepa-
rations, sequencing and genome assembly was performed by
NGI Stockholm. Sequencing was performed on Illumina No-
vaSeq6000 with a 2x151 setup. 10X linked reads were assem-
bled with 10X Genomics Supernova v2.1.0 (Weisenfeld et al.
2017). HiC reads were processed with Juicer v1.6 (Durand et al.
2016a) and used for scaffolding the 10X assembly with 3DDNA
v.180922 (Dudchenko et al. 2017). Resulting assemblies were re-
viewed with Juicebox v1.11.08 (Durand et al. 2016b). In addition
to minor corrections to the initial assembly, two chromosome
sized scaffolds were merged. The assembly was finalized with
the script ‘run-asm-pipeline-post-review.sh’ from the 3DDNA
pipeline v.180922 (Dudchenko et al. 2017).
Gene annotation
Gene annotation lift-over was performed by aligning wood
white protein queries generated by Talla et at.(2017) to the cur-
rent version of reference genome, using spaln 2.4.0 (Iwata and
Ghoto 2012) with the parameters -Q7 -LS -O7 -S3.
TE annotation
Repetitive element consensus sequences were predicted de novo
using RepeatModeler 1.0.11 (Bao et al. 2015). Transposable
elements characterized as unknown were submitted to CENSOR
(Bao et al. 2015) for annotation, where any hits with a score
< 200 were removed. All predicted sequences were matched
against gene annotations using diamond blast 2.0.4 (Buchfink
et al. 2021), to correct for annotation errors (bitscore > 100).
Transposable elements were then annotated in the genome with
RepeatMasker 4.1.0, using the predicted library of consensus
sequences in wood white and previously characterized TEs in
Heliconius melpomene in the RepeatMasker library 4.0.8 (Bao et al.
2015).
Sampling of individuals
Adult male wood whites were collected across the distribution
range in Sweden during June and July 2020 (Supplementary
Table 4). Sex was determined in situ based on two sexually
dimorphic characters; the presence of a black apical spot on
the forewing and the white coloration of the ventral part of the
antennae in males. Sampled individuals were directly preserved
in ethanol and frozen at -20ºC.
DNA extraction
DNA was extracted following two different protocols. In both
cases, the dissected tissue was digested overnight in Laird’s
buffer and homogenized with 20µl of proteinase K (20mg/ml,
>600 mAU/ml), followed by incubation with RNase A at 37ºC
for 30 minutes. DNA was extracted from thoraces using salt
extraction; 300 of NaCl (5M) was added, followed by centrifuga-
tion for 15 minutes at 13,000 revolutions per minute (rpm). Three
washing steps were completed with one volume of 70% ethanol
and centrifuging for five minutes at maximum speed. The re-
maining pellet was air-dried and then resuspended in 30 µl of
MilliQ H2O. For the abdomens, a phenol-chloroform extraction
protocol was used. Two cycles of phenol:chloroform:isoamyl
alcohol (25:24:1) addition and centrifugation for five minutes at
13,000 rpm were completed, plus a third cleaning cycle using
only chloroform. Precipitation of DNA was achieved by adding
2x volumes isopropanol + 0.1x 3M NaAc, incubating at -18ºC
overnight and centrifuging for 15 minutes at 13,000 rpm. The
final pellet was resuspended with 30 µl of MilliQ H2O. DNA pu-
rity was assessed with NanoDrop, and concentration measured
with Qubit DNA Broad Range.
Sequencing
To capture the genetic variation in the population in Sweden, 84
individuals from different geographic regions and the highest
DNA quality were selected for analysis. Library preparation
for all 84 samples using the TruSeq PCR-free kit followed by
multiplexing, and sequencing on two NovaSeq 6000 S4 lanes
with 2x150 bp reads, were performed at the National Genomics
Infrastructure (NGI), Stockholm.
Read trimming
Illumina sequencing adapters were trimmed by eliminating the
first fifteen base pairs (bp) on each end of the raw reads with
CutAdapt 1.9.1 (Martin 2011), filtered on Q-score < 30 and a min-
imum length of 30 bp. Read quality after cleaning was assessed
with FastQC (Andrews 2010). Before filtering, an average of 4.3
million reads per sample were obtained, and 2.5% were filtered
out.
Mapping and filtering
For each individual, paired-end reads were mapped to the refer-
ence genome with bwa v0.7.17 (Li and Durbin 2009). Samtools
v1.10 (Li et al. 2009) was used to select reads with paired infor-
mation. MarkDuplicatesSpark as implemented in GATK v4.1.4.1
(McKenna et al. 2010) was used to eliminate duplicated regions
with the –remove-sequencing-duplicates option.
Variant calling and filtering
The tool HaplotypeCaller in GATK v4.1.4.1 (McKenna et al. 2010)
was used for variant calling. Each chromosome for each individ-
ual was processed separately, and the resulting 84 files for each
chromosome were grouped and converted into a VCF file with
the GATK v4.1.4.1 tools Combine_gVCF and Genotype_gVCF
(McKenna et al. 2010), respectively. Total variant count was
obtained with the stats option in bcftools v1.10 (Li 2011). The
variants were filtered to have a minimum minor allele count
(MAC) of two, a per-site depth between 10 and 50, minimum per
site quality of 30, and < 20% per-site missing data with vcftools
0.1.15 (Danecek et al. 2011). Additionally, all insertions and dele-
tions were removed with the –remove-indels option. The num-
ber of remaining sites in each chromosome was counted again
with the stats option in bcftools v1.10 (Li 2011). Initial variant
calling resulted in a total of 51,189,479 markers along the genome
(11,055,543 indels + 40,133,936 SNPs), of which 10,565,404 SNPs
remained after filtering. This represents a genome-wide aver-
age of ∼17.6 SNPs/kb, given the ∼ 600 Mb total length of the
reference genome.
10
The fine-scale recombination rate variation and associations with genomic features in a butterfly
Inference of the demographic history
SMC++ (Terhorst et al. 2017) was used to infer the demographic
history of each individual chromosome, using a set of six “dis-
tinguished” individuals. These six individuals were selected
among the 10 with a higher average sequencing coverage for
the variants after filtering, so that they constituted a good rep-
resentation of the geographic distribution of the species. The
per-base mutation rate was set at 2.9 x 10-9 per generation, an es-
timate based on mutation frequency in H. melpomene (Keightley
et al. 2015). Known invariable regions such as centromeres must
be masked before inferring the demographic history, as they
can interfere with the signal. Since Lepidoptera are holocentric
and lack defined centromeres, a cut-off value of 150 kb was set
instead, so that any longer invariable region was considered
as missing data and discarded for the demographic inference.
The demographic trajectories were inferred for the last 5 million
generations, as defaulted by the program.
Recombination rate estimation
The chromosome-specific demographic trajectories were used
together with the VCF files to obtain high-resolution recombi-
nation maps using pyrho (Spence and Song 2019). An algo-
rithm implemented in the software LDpop (Kamm et al. 2016)
was used to compute a table of two-loci likelihoods under the
coalescent with recombination using the chromosome-specific
demographic trajectories as input. The same mutation rate as
before was used, together with the parameter values n = 168
(twice the number of diploid individuals), and N = 210 (25%
larger than n, as recommended by the manual). A relative tol-
erance (--decimate_rel_tol) value of 0.1 was used, together
with the --approx flag, recommended for large datasets. Differ-
ent window sizes (maximum distance between SNPs) and the
block penalty (determinant of the smoothness of the curve) were
tested for each chromosome, and the most appropriate were
selected based on the correlation between the data in the likeli-
hood tables and simulated data at different scales (1 bp, 10 kb
and 100 kb). For all chromosomes, the best block_penalty was 25,
and the best suitable window_size ranged between 50 and 100.
The look-up table, together with the final VCF file for each chro-
mosome were used to infer the local recombination rate using
the most appropriate parameter values and the --fast_missing
flag. The per-base, per-generation recombination rate between
each pair of SNP markers was obtained in the end. A positive
association between the recombination rate and marker density
was observed at a 1Mb scale (see Supplementary Information
and Supplementary Figure 4 for further discussion).
Distribution of recombination rate variation and iden-
tification of hot-spots and cold-spots
Regional recombination rate estimates were obtained in win-
dows of two different sizes (100 kb and 1 Mb) with a custom
script by calculating the weighted mean on each interval be-
tween markers, accounting for their length and recombination
rate. We used simulation-based approach to establish thresholds
for hot-spot identification. By performing coalescent simulation
with the flat recombination landscape but taking into account
previously inferred population history and genetic drift and we
obtained levels of recombination rate variation, which are due to
our imputation strategy. msprime 1.1.1 (Kelleher et al. 2016) was
used to simulate 99 independent sequences, each 100 kb long,
and the VCF files for each sequence were concatenated. The
parameters for the simulations included a flat recombination
landscape (7.37 cM/Mb, same as the obtained genome-wide
rate) and the same mutation rate used for SMC++, together with
a demographic trajectory that approximately reflects the inferred
trajectories from the empirical data; exponential growth in the
interval 104-106 generations BP, exponential decline 102-104 gen-
erations BP, and stable population in the last 102 generations. At
each time point, Ne was calculated as the average of the estimates
across chromosomes. The simulated genomic sequences were
analyzed in pyrho according to the steps described earlier. The
resulting regional recombination rates predominantly oscillated
in the range 2-8 cM/Mb, with a maximum of 20 cM/Mb, a 3-fold
higher rate compared to the genome-wide average (Supplemen-
tary Figure 5). Recombination peaks occurring at boundaries
between independently obtained VCF files were omitted. There-
fore, simulation allowed us to establish lower bound for hot-spot
threshold. For the final (more conservative) threshold we choose
regions with a recombination rate higher than 25 cM/Mb, be-
tween 750 and 10,000 bp long, and showing a 10-fold increase
over the regional background recombination rate (the mean rate
in the focal 100 kb window and the two flanking windows, in
total 300 kb). To avoid biases resulting from erroneously called
variants, hot-spots that included less than four markers (i.e.,
three intervals) were discarded. Recombination cold-spots were
identified as regions with a recombination rate 10-fold lower
than the genome-wide average, and including 4 markers with
no length limitations.
For the assessment of the distribution of recombination hot-
spots and cold-spots, we defined the subtelomeric regions as the
last 500 kb on each chromosome end. The rest of the genome
was considered proximal. Flanking regions to recombination
hot-spots were defined as the 5 kb segments on each side of each
accepted hot-spot.
Association of recombination rate with GC content,
gene density and TE classes
A multiple linear regression model was constructed to disen-
tangle the explanatory potential of each genomic feature in the
variation of recombination rate at a 1 Mb scale, using the lm
function in base R (R Core Team 2021). The linear model had
the recombination rate as response variable, and the explanatory
variables included the GC content, the gene density, and the rela-
tive abundance of four TE classes (DNA transposons, and LINE,
SINE and LTR retrotransposons). Potential associations between
the regional recombination rate and the different genomic fea-
tures were also analyzed with Spearman’s rank correlation tests
using the corr option in base R (R Core Team 2021).
BEDTools 2.29.2 (Quinlan and Hall 2010) maskfasta option
was used to select all annotated exons and TEs. Base composi-
tion for specific regions (corrected for masked positions) was
obtained with BEDTools nuc option. Gene and TE densities were
calculated as the proportion of a region covered by the annotated
sequences of each category.
To assess potential variation in the recombination rate within
specific genomic features, we also estimated the recombination
rate within each TE class, different regions in protein coding
genes (exons, introns and upstream regions) and intergenic se-
quence. Since 5’ UTR regions were not included in the annota-
tion file, we used the 100 bp upstream of the first exon of each
gene – a conservative selection to represent the 5’ UTR (Chen
et al. 2011, 2014). In order to avoid biases due to differential
selective pressures, only the intervals between markers from the
pyrho output that were positioned completely within the feature
Palahí et al.
11
were considered - i.e. for a given exon, if an interval between
markers overlapped completely with the predicted exon, it was
retained; if the overlap was only partial (including part of the
exon sequence but also part of a neighboring intron or UTR), it
was discarded. This overlap was checked with BEDTools 2.29.2
(Quinlan and Hall 2010) intersect option using the flag -f 1.
Permutation test to assess TE density in the cold- and
hot-spots
To assess the TE density in the cold and hotspot regions we ran a
permutation test in R v4.2.1 (R Core Team, 2013), by resampling
10,000 means of windows reflecting the count and average length
of the outlier windows and assessing where in this distribution
of resampled means the value of the hot- and cold-spot windows
appears. We then performed a two-sided test by assessing the
number of means that exceeded the original difference.
Gene Set Enrichment Analysis (GSEA) for the genes present
in the cold- and hot-spot regions We assessed enrichment of
functional categories in the cold and hot spot regions using
topGO v2.44.0 (Alexa and Rahnenfuhrer 2021) in R v4.2.1 (R
Core Team, 2013). We used the annotated gene set with gene
ontology (GO) terms associated to molecular function. To assess
significance, we used the Fisher’s exact test and the default
algorithm (“weight01”) accounting for the hierarchical structure
of the GO-terms (Alexa et al. 2006). We adjusted the p-values
with Benjamini-Hochberg’s method of multiple test correction
(p.adjust(x, method = "fdr")). We used HOMER v4.11 (Heinz
et al. 2010) to assess motif enrichment in the cold and hot spot
windows.
Data access
All
raw
data
generated
in
this
study
has
been
sub-
mitted
to
the
European
Nucleotide
Archive
(ENA;
https://www.ebi.ac.uk/ena/browser/home)
under
ac-
cession number PRJEB56690. Command lines are available on
the group’s GitHub page (https://github.com/EBC-butterfly-
genomics-team).
Competing Interests Statement
The authors declare that they have no conflict of interest.
Acknowledgements
This work was supported by a research grant from the Swedish
Research Council (Vetenskapsrådet Grant ID: 2019-04791) to
N.B. The authors acknowledge support from the National Ge-
nomics Infrastructure in Stockholm funded by Science for Life
Laboratory, the Knut and Alice Wallenberg Foundation and
the Swedish Research Council, and SNIC/Uppsala Multidis-
ciplinary Center for Advanced Computational Science for as-
sistance with massively parallel sequencing and access to the
UPPMAX computational infrastructure. This work was also sup-
ported by NBIS/SciLifeLab long-term bioinformatics support
(WABI). R.V. was supported by grant PID2019-107078GB-I00,
funded by MCIN/AEI/10.13039/ 501100011033.
References
Ahola V, Lehtonen R, Somervuo P, Salmela L, Koskinen P, Rastas P,
Välimäki N, Paulin L, Kvist J, Wahlberg N et al. 2014. The Glanville
fritillary genome retains an ancient karyotype and reveals selective
chromosomal fusions in Lepidoptera. Nature Communications 5: 4737.
doi:10.1038/ncomms5737.
Aljanabi SM, Martinez I. 1997. Universal and rapid salt-extraction of
high quality genomic DNA for PCR-based techniques. Nucleic Acids
Research 25: 4692–4693.
Andrews
S.
2010.
FastQC:
a
quality
control
tool
for
high
throughput
sequence
data.
Available
online
at:
http://www.bioinformatics.babraham.ac.uk/projects/fastqc.
Axelsson E, Webster MT, Ratnakumar A, The LUPA Consortium,
Ponting CP, Lindblad-Toh K. 2012. Death of PRDM9 coincides with
stabilization of recombination landscape in the dog genome. Genome
Research 22: 51 – 63.
Backström N, Forstmeier W, Schielzeth H, Mellenius H, Nam K,
Bolund E, Webster MT, Öst T, Schneider M, Kempenaers B. et al. 2010.
The recombination landscape of the zebra finch Taeniopygia guttata
genome. Genome Research 20: 485-495.
Bao W, Kojima KK, Kohany O. 2015. Repbase Update, a database
of repetitive elements in eukaryotic genomes.
Mobile DNA 6: 11.
doi:10.1186/s13100-015-0041-9.
Baril T, Hayward A. 2022. Migrators within migrators: exploring
transposable element dynamics in the monarch butterfly, Danaus plexip-
pus. Mobile DNA 13: 5. doi:10.1186/s13100-022-00263-5.
Bartolomé C, Maside X, Charlesworth B. 2002. On the abundance
and distribution of transposable elements in the genome of Drosophila
melanogaster. Molecular Biology and Evolution 19: 926–937.
Barton AB, Su Y, Lamb J, Barber D, Kaback DB. 2003. A function for
subtelomeric DNA in Saccharomyces cerevisiae. Genetics 165: 929 – 934.
Barton AB, Pekosz MR, Kurvathi RS, Kaback DB. 2008. Meiotic
recombination at the ends of chromosomes in Saccharomyces cerevisiae.
Genetics 179: 1221 – 1235.
Bergero R, Charlesworth D. 2009. The evolution of restricted recom-
bination in sex chromosomes. Trends in Ecology and Evolution 24: 94 –
102.
Beye M, Gattermeier I, Hasselmann M, Gempe T, Schioett M, Baines
JF, Schlipalius D, Mougel F, Emore C, Rueppell O et al. 2006. Exception-
ally high levels of recombination across the honey bee genome. Genome
Research 16: 1339 – 1344.
Boman J, Mugal CF, Backström N. 2021. The effects of GC-biased
gene conversion on patterns of genetic diversity among and across
butterfly genomes.
Genome Biology and Evolution 13:
evab064.
doi:10.1093/gbe/evab064.
Buchfink B, Reuter K, Drost HG. 2021. Sensitive protein alignments
at tree-of-life scale using DIAMOND. Nature Methods 18: 366 – 368.
Carvalho A, Clark A. 1999. Intron size and natural selection. Nature
401: 344.
Chen C, Lin H, Pan C, Chen F. 2011. The length evolution of 5’
untranslated regions - the stochastic model revisited. IEEE International
Conference on Bioinformatics and Engineering: 162 – 166.
Chen X, Lü X, Zhang Y. 2014. Characterization of protein phos-
phatase 5 from three Lepidopteran insects:
Helicoverpa armigera,
Mythimna separata and Plutella xylostella.
PLoS One 9:
e97437.
doi:10.1371/journal.pone.0097437
Choi K, Henderson IR. 2015. Meiotic recombination hot-spots – a
comparative view. Plant Journal 83: 52 – 61.
Clément Y, Arndt PF. 2013. Meiotic recombination strongly influences
GC-content evolution in short regions in the mouse genome. Molecular
Biology and Evolution 30: 2612 – 2618.
Comeron JM, Kretman M. 2000. The correlation between intron
length and recombination in Drosophila: dynamic equilibrium between
mutational and selective forces. Genetics 156: 1175 – 1190.
Comeron JM, Kreitman M. 2002. Population, evolutionary and ge-
nomic consequences of interference selection. Genetics 161: 389 – 410.
Comeron JM, Ratnappan R, Bailin S. 2012. The many landscapes of
recombination in Drosophila melanogaster. PLoS Genetics 8: e1002905.
doi:10.1371/journal.pgen.1002905
Danecek P, Auton A, Abecasis G, Albers CA, Banks E, DePristo
MA, Handsaker RE, Lunter G, Marth GT, Sherry ST et al. 2011. 1000
Genomes Project Analysis Group, The variant call format and VCFtools.
Bioinformatics 27: 2156 – 2158.
Dantart J. 2003. Leptidea sinapis and L. reali, sibling species as yet
poorly known in Catalonia. Cynthia: Bulletin of the Catalan Butterfly
Monitoring Scheme 3: 5 – 6.
12
The fine-scale recombination rate variation and associations with genomic features in a butterfly
Dapper AL, Payseur BA. 2017.
Connecting theory and data
to understand recombination rate evolution.
Philosophical Trans-
actions of the Royal Society B: Biological Sciences 372:
1 – 11.
doi:10.1098/rstb.2016.0469.
Davey JW, Barker SL, Rastas PM, Pinharanda A, Martin SH, Durbin
R, McMillan WO, Merrill RM, Jiggins CD. 2017. No evidence for mainte-
nance of a sympatric Heliconius species barrier by chromosomal inver-
sions. Evolution Letters 1: 138 – 154.
Dinc˘a V, Lukhtanov VA, Talavera G, Vila R. 2011. Unexpected layers
of cryptic diversity in wood white Leptidea butterflies. Nature Commu-
nications 2: e324. doi:10.1038/ncomms1329.
Dolgin ES, Charlesworth B. 2008. The effects of recombination rate
on the distribution and abundance of transposable elements. Genetics
178: 2169 – 2177.
DuBois ML, Haimberger ZW, McIntosh MW, Gottschling DE. 2002. A
quantitative assay for telomere protection in Saccharomyces cerevisiae.
Genetics 161: 995 – 1013.
Dudchenko O, Batra SS, Omer AD, Nyquist SK, Hoeger M, Du-
rand NC, Shamim MS, Machol I, Lander ES, Aiden AP, Aiden EL. 2017.
De novo assembly of the Aedes aegypti genome using Hi-C yields
chromosome-length scaffolds. Science 356: 92 – 95.
Durand NC, Shamim MS, Machol I, Rao SS, Huntley MH, Lander
ES, Aiden EL. 2016a. Juicer provides a one-click system for analyzing
loop-resolution Hi-C experiments. Cell Systems 3: 95 – 98.
Durand NC, Robinson JT, Shamim MS, Machol I, Mesirov JP, Lander
ES, Aiden EL. 2016b. Juicebox provides a visualization system for Hi-C
contact maps with unlimited zoom. Cell Systems 3: 99 – 101.
Duret L, Galtier N. 2009. Biased gene conversion and the evolution
of mammalian genomic landscapes. Annual Review of Genomics and
Human Genetics 10: 285 – 311.
Duret L, Marais G, Biémont C. 2000. Transposons but not retrotrans-
posons are located preferentially in regions of high recombination rate
in Caenorhabditis elegans. Genetics 156: 1661 – 1669.
Fang L, Ye J, Li N, Zhang Y, Li S, Wong G, Wang J. 2008. Positive cor-
relation between recombination rate and nucleotide diversity is shown
under domestication selection in the chicken genome. China Science
Bulletin 53: 746 – 750.
Friberg M, Bergman M, Kullberg J, Wahlberg N, Wiklund C. 2008.
Niche separation in space and time between two sympatric sister species
– a case of ecological pleiotropy. Evolutionary Ecology 22: 1 – 18.
Fullerton SM, Bernardo Carvalho A, Clark AG. 2001. Local rates of
recombination are positively correlated with GC content in the human
genome. Molecular Biology and Evolution 18: 1139 – 1142.
Gaut B, Wright S, Rizzon C, Dvorak J, Anderson LK. 2007. Recombi-
nation: an underappreciated factor in the evolution of plant genomes.
Nature Reviews Genetics 8: 77 – 84.
Gerton JL, DeRisi J, Shroff R, Lichten M, Brown PO, Petes TD. 2000.
Global mapping of meiotic recombination hot-spots and cold-spots in the
yeast Saccharomyces cerevisiae. Proceedings of the National Academy
of Sciences 97: 11383 - 11390. doi:10.1073/pnas.97.21.11383.
Gray S, Cohen PE. 2016. Control of meiotic crossovers: from doubre-
strand break formation to designation. Annual Reviews Genetics 50: 175
– 210.
Grey C, Barthès P, Chauveau-Le Friec G, Langa F, Baudat F, De Massy
B. 2011. Mouse PRDM9 DNA-binding specificity determines sites of
histone H3 lysine 4 trimethylation for initiation of meiotic recombination.
PLos Biology 9: e1001176. doi:10.1371/journal.pbio.1001176.
Günther T, Malmström H, Svensson EM, Omrak A, Sánchez-Quinto
F, Kılınç GM, Krzewinska M, Eriksson G, Fraser M, Edlund H et al.
2018. Population genomics of Mesolithic Scandinavia: Investigating
early postglacial migration routes and high-latitude adaptation. PLoS
Biology 16: e2003703. doi:10.1371/journal.pbio.2003703.
Haenel Q, Laurentino TG, Roesti M, Berner D. 2018. Meta-analysis
of chromosome-scale crossover rate variation in eukaryotes and its sig-
nificance to evolutionary genomics. Molecular Ecology 27: 2477–2497.
Heinz S, Benner C, Spann N, Bertolino E, Lin YC, Laslo P, Cheng JX,
Murre C, Singh H, Glass CK. 2010. Simple Combinations of Lineage-
Determining Transcription Factors Prime cis-Regulatory Elements Re-
quired for Macrophage and B Cell Identities. Moleular Cell 38:576 –
589.
Hill WG, Robertson A. 1966. The effect of linkage on limits to artificial
selection. Genetic Research 8: 269 – 294.
Höök L, Näsvall K, Vila R, Wiklund C, Backström N. 2022. High-
density linkage maps and chromosome level genome assemblies unveil
direction and frequency of extensive structural rearrangements in wood
white butterflies (Leptidea sp.). [Submitted]
Iwata H, Gotoh O, 2012. Benchmarking spliced alignment programs
including Spaln2, an extended version of Spaln that incorporates ad-
ditional species-specific features.
Nucleic Acids Research 40: e161.
doi:10.1093/nar/gks708.
Jacobs GS, Sluckin TJ, Kivisild T. 2016. Refining the use of linkage
disequilibrium as a robust signature of selective sweeps. Genetics 203:
1807 – 1825.
Jeffcoate S, Joy J. 2011. Evidence-based national recovery plan for
Leptidea sinapis (wood white butterfly) in Southern Britain. Journal of
Insect Conservation 15: 759 – 763.
Jensen-Seaman MI, Furey TS, Payseur BA, Lu Y, Roskin KM, Chen CF,
Thomas MA, Haussler D, Jacob HJ. 2004. Comparative recombination
rates in the rat, mouse, and human genomes. Genome research 14: 528 –
538.
Jiggins CD, Mavarez J, Beltrán M, McMillan WO, Spencer Johnston
J, Bermingham E. 2005. A genetic linkage map of the mimetic butterfly
Heliconius melpomene. Genetics 171: 557 – 570.
Johnston SE, Berenos C, Slate J, Pemberton JM. 2016. Conserved
genetic architecture underlying individual recombination rate variation
in a wild population of soay sheep (Ovis aries). Genetics 203: 583 – 598.
Jones JC, Wallberg A, Christmas MJ, Kapheim KM, Webster MT. 2019.
Extreme differences in recombination rate between the genomes of a
solitary and a social bee. Molecular Biology and Evolution 36: 2277 –
2291.
Kaback DB, Guacci V, Barber D, Mahon JW. 1992. Chromosome
size-dependent control of meiotic recombination. Science 256: 228 – 232.
Kamm JA, Spence JP, Chan J, Song YS. 2016. Two-locus likelihoods
under variable population size and fine-scale recombination rate estima-
tion. Genetics 203: 1381 – 1399.
Kawakami T, Mugal CF, Suh A, Nater A, Burri R, Smeds L, Ellegren
H. 2013. Whole-genome patterns of linkage disequilibrium across fly-
catcher populations clarify the causes and consequences of fine-scale
recombination rate variation in birds. Molecular Ecology 26: 4158 – 4172.
Keightley PD, Pinharanda A, Ness RW, Simpson F, Dasmahapatra
KK, Mallet J, Davey JW, Jiggins CD. 2015. Estimation of the sponta-
neous mutation rate in Heliconius melpomene. Molecular Biology and
Evolution 32: 239 – 243.
Kelleher J, Etheridge AM, McVean G. 2016. Efficient coalescent simu-
lation and genealogical analysis for large sample sizes. PLoS Computa-
tional Biology 12: e1004842. doi:10.1371/journal.pcbi.1004842.
Kent TV, Uzunovi´c J, Wright SI. 2017. Coevolution between trans-
posable elements and recombination. Philosophical Transactions of the
Royal Society B 372: 20160458. doi:10.1098/rstb.2016.0458.
Kiktev DA, Sheng Z, Lobachev KS, Petes TD. 2018. GC content
elevates mutation and recombination rates in the yeast Saccharomyces
cerevisiae. Proceedings of the National Academy of Sciences 115: E7109
– E7118. doi: 10.1073/pnas.1807334115.
Kim Y, Nielsen R, 2004. Linkage disequilibrium as a signature of
selective sweeps. Genetics 167: 1513 – 1524.
Lavoie CA, Platt RN, Novick PA, Counterman BA, Ray DA. 2013.
Transposable element evolution in Heliconius suggests genome diversity
within Lepidoptera. Mobile DNA 4: 21. doi:10.1186/1759-8753-4-21.
Li H, Durbin R. 2009. Fast and accurate short read alignment with
Burrows-Wheeler Transform. Bioinformatics 25: 1754 – 1760.
Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, Marth G,
Abecasis G, Durbin R, 1000 Genome Project Data Processing Subgroup.
2009. The Sequence alignment/map (SAM) format and SAMtools. Bioin-
formatics 25: 2078 – 2079.
Li H. 2011. A statistical framework for SNP calling, mutation discov-
ery, association mapping and population genetical parameter estimation
from sequencing data. Bioinformatics 27: 2987 – 2993.
Li H, Durbin R. 2011. Inference of human population history from
individual whole-genome sequences. Nature 475: 493 – 496.
Palahí et al.
13
Lin Z, Li W. 2011. Evolution of 5’ Untranslated region length and
gene expression reprogramming in yeasts. Molecular Biology and Evo-
lution 29: 81 – 89.
Lukhtanov VA, Dinc˘a V, Friberg M, Šíchová J, Olofsson M, Vila R,
Marec F, Wiklund C. 2018. Versatility of multivalent orientation, inverted
meiosis, and rescued fitness in holocentric chromosomal hybrids. Pro-
ceedings of the National Academy of Sciences of the United States of
America 115: E9610 – E9619. doi:10.1073/pnas.1802610115.
Marais G, Charlesworth B. Genome evolution: recombination speeds
up adaptive evolution. Current biology 13: R68 – R70.
Marais G, Mouchiroud D, Duret L. 2001. Does recombination im-
prove selection on codon usage? Lessons from nematode and fly com-
plete genomes. Proceedings of the National Academy of Sciences 98:
5688 – 5692. doi:10.1073/pnas.091427698.
Margus A, Lindström L. 2020. Prolongued diapause has sex-specific
fertility and fitness costs. Evolutionary Ecology 34: 41 – 57.
Martin M. 2011. Cutadapt removes adapter sequences from high-
throughput sequencing reads. EMBnet.journal 17: 10 – 12.
Martin SH, Davey JW, Salazar C, Jiggins CD. 2019. Recombination
rate variation shapes barriers to introgression across butterfly genomes.
PLos Biology 17: e2006288. doi:10.1371/journal.pbio.2006288.
Mather N, Traves SM, Ho SYW. 2020. A practical introduction to
sequentially Markovian coalescent methods for estimating demographic
history from genomic data. Ecology and Evolution 10: 579 – 589.
Maynard Smith J, Haigh J. 1974.
The hitch-hiking effect of a
favourable gene. Genome Research 23: 23 – 35.
McKenna A, Hanna M, Banks E, Sivachenko A, Cibulskis K, Kernyt-
sky A, Garimella K, Altshuler D, Gabriel S, Daly M et al. 2010. The
Genome Analysis Toolkit: a MapReduce framework for analyzing next-
generation DNA sequencing data. Genome Research 20: 1297 – 1303.
McVean GAT, Myers SR, Hunt S, Deloukas P, Bentley DR, Donnelly
P. 2004. The fine-scale structure of recombination rate variation in the
human genome. Science 304: 581 – 584.
Meunier J, Duret L. 2004. Recombination drives the evolution of
GC-content in the human genome. Molecular Biology and Evolution 21:
984 – 990.
Moens PB, Kolas NK, Tarsounas M, Marcon E, Cohen PE, Spy-
ropoulos B. 2002. The time course and chromosomal localization of
recombination-related proteins at meiosis in the mouse are compatible
with models that can resolve the early DNA-DNA interactions without
reciprocal recombination. Journal of Cell Science 115: 1611-1622.
Moudgil DK, Westcott N, Famulski JK, Patel K, Macdonald D, Hang
H, Chan GKT. 2015. A novel role of farnesylation in targeting a mitotic
checkpoint protein, human Spindly, to kinetochores. Journal in Cell
Biology 208: 881 – 896.
Mouresan EF, González-Rodríguez A, Cañas-Álvarez JJ, Munilla S,
Altarriba J, Díaz C, Baró JA, Molina A, López-Buesa P, Piedrafita J et
al. 2019. Mapping recombination rate on the autosomal chromosomes
based on the persistency of linkage disequilibrium phase among au-
tochthonous beef cattle populations in Spain. Frontiers in Genetics 10:
1170. doi:10.3389/fgene-2019.01170.
Muyle A, Serres-Giardi L, Ressayre A, Escobar J, Glémin S. 2011.
GC-biased gene conversion and selection affect GC content in the Oryza
genus (rice). Molecular Biology and Evolution 28: 2695 – 2706.
Myers S, Bottolo L, Freeman C, McVean G, Donelly P. 2005.
A
fine-scale map of recombination rates and hot-spots across the human
genome. Science 310: 321 – 324.
Nokkala S, Kuznetsova VG, Maryanska-Nadachowska A, Nokkala C.
2004. Holocentric chromosomes in meiosis. I. Restriction of the number
of chiasmata in bivalents. Chromosome Research 12: 733 – 739.
Okazaki S, Tsuchida K, Maekawa H, Ishikawa H, Fujiwara H. 1993.
Identification of a pentanucleotide telomeric sequence, (TTAGG)n, in
the silkworm Bombyx mori and in other insects. Molecular and Cellular
Biology 13: 1424 – 1432.
Otto SP, Payseur BA. 2019. Crossover interference: shedding light on
the evolution of recombination. Annual Review of Genetics 53: 19 – 44.
Onozawa M, Zhang Z, Kim YJ, Goldberg L, Varga T, Leif Bergsagel P,
Kuehl WM, Aplan PD. 2014. Repair of DNA double-strand breaks by
templated nucleotide sequence insertions derived from distant regions
of the genome. Proceedings of the National Academy of Sciences 111:
7729 – 7734.
Paigen K, Szatkiewicz JP, Sawyer K, Leahy N, Parvanov ED, Ng
SH, Graber JH, Broman KW, Petkov PM. 2008.
The recombina-
tional anatomy of a mouse chromosome. PLoS Genetics 4: e1000119.
doi:10.1371/journal.pgen.1000119.
Pardo-Manuel de Villena F, Sapienza C. 2001. Recombination is pro-
portional to the number of chromosome arms in mammals. Mammalian
Genome 12: 318 – 322.
Peñalba JV, Wolf JBW. 2020. From molecules to populations: appre-
ciating and estimating recombination rate variation. Nature Reviews
Genetics 21: 476 – 492.
Petrov DA, Fiston-Lavier AS, Lipatov M, Lenkov K, González J.
2011. Population genomics of transposable elements in Drosophila
melanogaster. Molecular Biology and Evolution 28: 1633 – 1644.
Powell JA. 1989. Synchronized, mass-emergence of a Yucca moth,
Proxodus Y-inversus (Lepidoptera: Prodoxidae) after 16 and 17 years in
diapause. Oecologia 81: 490 – 493.
Prachumwat A, DeVincentis L, Palopoli MF. 2004. Intron size cor-
relates positively with recombination rate in Caenorhabditis elegans.
Genetics 166: 1585 – 1590.
Quinlan AR, Hall IM. 2010. BEDTools: a flexible suite of utilities for
comparing genomic features. Bioinformatics 26: 841 – 842.
Robinson R. 1971. Lepidoptera genetics. 1st ed. Pergamon Press,
Oxford.
R Core Team. 2021. R: A language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna, Austria.
Available online at: https://www.R-project.org/.
Samuk K, Manzano-Winkler B, Ritz KR, Noor MAF. 2020. Natu-
ral selection shapes variation in genome-wide recombination rate in
Drosophila pseudoobscura. Current Biology 30: 1517 – 1528.
Sasaki T, Fujiwara H. 2000. Detection and distribution patterns of
telomerase activity in insects. European Journal of Biochemistry 267:
3025 – 3031.
Scherthan H, Weich S, Schwegler H, Heyting C, Härle M, Cremer
T. 1996. Centromere and telomere movements during early meiotic
prophase of mouse and man are associated with the onset of chromo-
some pairing. Journal of Cell Biology 134: 1109 – 1125.
Shen C, Li X, Zhang R, Lin Z. 2017. Genome-wide recombination
rate variation in a recombination map of cotton. PLoS ONE 12: e0188682.
doi:10.1371/journal.pone.0188682.
Shipilina D, Näsvall K, Höök L, Vila R, Talavera G, Backström N.
2022. Linkage mapping and genome annotation give novel insights into
gene family expansions and regional recombination rate variation in
the painted lady (Vanessa cardui) butterfly. Genomics 114: 110481. doi:
10.1016/j.ygeno.2022.110481
Singer MF. 1982. SINEs and LINEs: highly repeated short and long
interspersed sequences in mammalian genomes. Cell 28: 433 – 434.
Singhal S, Leffler EM, Sannareddy K, Turner I, Venn O, Hooper
DM, Strand AI, Li Q, Raney B, Balakrishnan CN et al. 2015. Stable
recombination hot-spots in birds. Science 350: 928 – 932.
Smeds L, Mugal CF, Qvarnström A, Ellegren H. 2016.
High-
resolution mapping of crossover and non-crossover recombination
events by whole-genome re-sequencing of an avian pedigree. PLOS
Genetics 12: e1006044. doi:10.1371/journal.pgen.1006044.
Smith GR, Nambiar M. 2020. New solutions to old problems: mecha-
nisms of meiotic crossover control. Trends in Genetics 36: 337 – 346.
Smolander O-P, Blande D, Ahola V, Rastas P, Tanskanen J, Kam-
monen JI, Oostra V, Pellegrini L, Ikonen S, Dallas T, et al.
2022.
Improved chromosome-level genome assembly of the Glanville frit-
illary butterfly (Melitaea cinxia) integrating Pacific Biosciences long
reads and a high-density linkage map, GigaScience 11:
giab097.
doi:10.1093/gigascience/giab097.
Smukowski CS, Noor MA. 2011. Recombination rate variation in
closely related species. Heredity 107: 496 – 508.
Spence JP, Song YS. 2019. Inference and analysis of population-
specific fine-scale recombination maps across 26 diverse human popula-
tions. Science Advances 5: eaaw9206. doi:10.1126/sciadv.aaw9206.
Stapley J, Feulner PGD, Johnston SE, Santure AW, Smadja CM. 2017.
Variation in recombination frequency and distribution across eukaryotes:
patterns and processes. Philosophical Transactions of the Royal Society
B 372: 20160455. doi:10.1098/rstb.2016.0455.
14
The fine-scale recombination rate variation and associations with genomic features in a butterfly
Su Y, Barton AB, Kaback DB. 2000. Decreased meiotic reciprocal re-
combination in subtelomeric regions in Saccharomyces cerevisiae. Chro-
mosoma 109: 467 – 475.
Suomalainen E. 1953. The kinetochore and the bivalent structure in
the Lepidoptera. Hereditas 39: 88 – 96.
Talla V, Suh A, Kalsoom F, Dinc˘a V, Vila R, Friberg M, Wiklund C,
Backström B. 2017. Rapid increase in genome size as a consequence of
transposable element hyperactivity in wood-white (Leptidea) butterflies.
Genome Biology and Evolution 9: 2491 – 2505.
Talla V, Johansson A, Dinca V, Vila R, Friberg M, Wiklund C, Back-
ström N. 2019. Lack of gene flow: Narrow and dispersed differentiation
island in a triplet of Leptidea butterfly species. Molecular Ecology 28:
3756 – 3770.
Terhorst J, Kamm J, Song Y. 2017. Robust and scalable inference of
population history from hundreds of unphased whole genomes. Nature
Genetics 49: 303 – 309.
Tiley GP, Burleigh JG. 2015.
The relationship of recombination
rate, genome structure, and patterns of molecular evolution across an-
giosperms. BMC Evolutionary Biology 15: 194. doi:10.1186/s12862-015-
0473-3.
Tobler A, Kapan D, Flanagan NS, Gonzalez C, Peterson E, Jiggins CD,
Johnston JS, Heckel DG, McMillan WO. 2005. First-generation linkage
map of the warningly colored butterfly Heliconius erato. Heredity 94:
408 – 417.
Wallberg A, Glémin S, Webster MT. 2015.
Extreme recombina-
tion frequencies shape genome variation and evolution in the honey-
bee, Apis mellifera. PLOS Genetics 11: e1005189. doi: 10.1371/jour-
nal.pgen.1005189
Weisenfeld NI, Kumar V, Shah P, Church DM, Jaffe DB. 2017. Direct
determination of diploid genome sequences. Genome Research 27: 757-
767.
Witherspoon DJ, Watkins WS, Zhang Y, Xin J, Tolpinrud WL, Hedges
DJ, Batzer MA, Borde LB. 2009. Alu repeats increase local recombination
rates. BMC Genomics 10: 530.
Wright SI, Agrawal N, Bureau TE. 2003. Effects of recombination rate
and gene density on transposable element distributions in Arabidopsis
thaliana. Genome Research 13: 1897 – 1903.
Xu YX, Du JC. 2014. Young but not relatively old retrotransposoons
are preferentially located in gene-rich euchromatic regions in tomato
(Solanum lycopersicum) plants. The Plant Journal 80: 582 – 591.
Yasukochi Y. 1998. A dense genetic map of the silkworm, Bombyx mori,
covering all chromosomes based on 1018 molecular markers. Genetics
150: 1513 – 1525.
| 2022 | The fine-scale recombination rate variation and associations with genomic features in a butterfly | 10.1101/2022.11.02.514807 | [
"i Torres Aleix Palahí",
"Höök Lars",
"Näsvall Karin",
"Shipilina Daria",
"Wiklund Christer",
"Vila Roger",
"Pruisscher Peter",
"Backström Niclas"
] | creative-commons |
1
Short Title: Dominance interaction between fruits and shoots
1
2
The role of auxin and sugar signaling in dominance inhibition of inflorescence growth by
3
fruit load
4
5
Marc Goetz, Maia Rabinovich and Harley M. Smith1,2
6
CSIRO Agriculture and Food, Private Bag 2, Glen Osmond, South Australia 5064, Australia
7
1Author for contact: harley.smith@csiro.au
8
2Senior author.
9
10
The author responsible for distribution of materials integral to the findings presented in this
11
article in accordance with the policy described in the Instructions for Authors
12
(www.plantphysiol.org) is:
13
Harley M. Smith (harley.smith@csiro.au).
14
15
One-sentence summary: Dominance inhibition of inflorescence shoot growth by fruit load is
16
involves auxin and sugar signaling during the end of flowering transition.
17
18
Author contributions: H.M.S. conceived the project, supervised the experiments, performed
19
the in situ hybridization experiments, wrote the manuscript and agrees to serve as the
20
author responsible for contact and communication; M.G. quantified soluble sugar in active
21
and quiescent apices and fruits; M.R. performed DR5:GUS analyses in active and quiescent
22
apices.
23
2
ABSTRACT
24
Dominance inhibition of shoot growth by fruit load is a major factor that regulates shoot
25
architecture and limits yield in agriculture and horticulture crops. In annual plants, the
26
inhibition of inflorescence growth by fruit load occurs at a late stage of inflorescence
27
development termed the end of flowering transition. Physiological studies show that this
28
transition is mediated by production and export of auxin from developing fruits in close
29
proximity to the inflorescence apex. In the meristem, cessation of inflorescence growth is
30
controlled in part by the age dependent pathway, which regulates the timing of arrest. Here,
31
results show that the end of flowering transition is a two-step process in which the first
32
stage is characterized by a cessation of inflorescence growth, while immature fruit continue
33
to develop. At this stage, dominance inhibition of inflorescence growth by fruit load
34
correlates with a selective dampening of auxin transport in the apical region of the stem.
35
Subsequently, an increase in auxin response in the vascular tissues of the apical stem where
36
developing fruits are attached marks the second stage for the end of flowering transition.
37
Similar to the vegetative and floral transition, the end of flowering transition correlates with
38
a change in sugar signaling and metabolism in the inflorescence apex. Taken together, our
39
results suggest that during the end of flowering transition, dominance inhibition of
40
inflorescence shoot growth by fruit load is mediated by auxin and sugar signaling.
41
3
INTRODUCTION
42
Understanding how growing units in a shoot system are regulated, including apical and
43
lateral buds, as well as fruits, is key to developing elite breeding lines and management tools
44
aimed at optimizing plant architecture and increasing yield in agriculture and horticulture
45
crops (Teichmann and Muhr, 2015; Guo et al., 2020). The activity and development of apical
46
and lateral buds, as well as fruits, is controlled by light, temperature, hormone, sugar and
47
nutrient signaling (Montgomery, 2008; Pfeiffer et al., 2017; Barbier et al., 2019). Moreover,
48
these endogenous and environmental signalling pathways facilitate communication between
49
the growing shoot apex and lateral sinks (meristems or fruits) to ensure plants adopt the
50
appropriate architecture and reproductive capacity based on carbohydrate, nutrient and
51
water availability (Walker and Bennett, 2018; Barbier et al., 2019).
52
53
The correlative or dominance inhibition hypothesis predicts that sinks with a high growth
54
potential inhibit the growth of younger or subordinate organs with a lower sink activity
55
(Bangerth, 1989; Smith and Samach, 2013; Walker and Bennett, 2018). As a result, the
56
dominant sink is able direct water, assimilates and nutrients required for growth and
57
development. Dominance inhibition occurs among fruits within an inflorescence or between
58
apical and axillary buds within a shoot. Interestingly, in perennial tree crops, dominance
59
inhibition between shoots and fruits is highly plastic, as changes in the growth potential
60
between these competing sinks can change over the course of season. For example, a high
61
rate of immature fruit abscission in late spring/early summer correlates with the outgrowth
62
of preformed vegetative shoots in avocado (Salazar-García et al., 2013). Therefore, it has
63
been hypothesized that dominance exerted by the vegetative shoot with a high growth
64
potential results in the abscission of developing fruitlets, which have a low growth potential.
65
However, as ‘retained’ avocado fruitlets enter a phase of rapid growth and the sink potential
66
increases, dominance between shoots and fruits switches and shoot growth is inhibited by
67
the developing fruit (Salazar-García et al., 1998; Ziv et al., 2014). Dominance inhibition of
68
shoot growth by fruit load is problematic when trees maintain a high crop load, as this
69
condition significantly reduces canopy growth resulting in a severe reduction in flowering
70
and yield the following year (Samach and Smith, 2013; Smith and Samach, 2013). Therefore,
71
dynamic dominance interaction between developing fruits and shoots are of significant
72
interest, as fruit abscission and the inhibition of shoot growth by fruit load significantly
73
4
reduces yield in tree crops (Samach and Smith, 2013; Smith and Samach, 2013; Sawicki et al.,
74
2015).
75
76
Auxin is major regulator of dominance inhibition of lateral buds by the growing shoot apex,
77
termed apical dominance (Barbier et al., 2019; Schneider et al., 2019), as well as among
78
developing fruits within an inflorescence shoot (Bangerth et al., 2000; Smith and Samach,
79
2013; Walker and Bennett, 2018). In both cases, dominance inhibition is initiated by the
80
biosynthesis and basipetal transport of auxin from the growing shoot apex or dominant fruit.
81
For apical dominance, the polar auxin transport system (PATS) channels auxin basipetally in
82
the stem in association with the vascular tissues (Galweiler et al., 1998). A local auxin
83
transport system called the connective auxin transport system (CATS) also distributes this
84
hormone in stem tissues (Bennett et al., 2016; van Rongen et al., 2019). Together,
85
movement of auxin via the PATS and CATS indirectly inhibits bud outgrowth (Barbier et al.,
86
2019). The canalization hypothesis predicts that a high stream of auxin channelled
87
basipetally in the stem from the dominant shoot apex indirectly dampens auxin transport
88
out of the lateral bud, which prevents release (Muller and Leyser, 2011). The second-
89
messenger hypothesis reasons that high auxin concentration in the stem promotes
90
strigolactone (SL) biosynthesis and this hormone moves into the bud to inhibit growth
91
(Rameau et al., 2014; Barbier et al., 2019) . In the bud, SL acts in part to dampen auxin
92
transport to prevent bud outgrowth (Crawford et al., 2010; Shinohara et al., 2013).
93
Furthermore, this mobile hormone is implicated in suppressing auxin biosynthesis and
94
response genes in the bud (Wang et al., 2020). SL also functions to regulate key bud
95
dormancy related transcription factors including D53/SUPPRESSOR OF MAX2-LIKE 6, 7 and 8
96
(Jiang et al., 2013; Zhou et al., 2013; Soundappan et al., 2015; Wang et al., 2015; Wang et al.,
97
2020), as well as BRANCHED1 (BRC1)/TEOSINTE BRANCHED1 (TB1) (Aguilar-Martinez et al.,
98
2007; Braun et al., 2012; Wang et al., 2020). Interestingly, studies in Cucumis sativus suggest
99
that BRC1/TB1 prevents bud release in part by repressing transcription of a polar auxin
100
transporter gene involved in branching (Shen et al., 2019). Finally, bud dormancy is
101
maintained in part through the suppression of cell division and ribosome production
102
(Gonzalez-Grandio et al., 2013), as well as the upregulation of abscisic acid (ABA) and
103
jasmonic acid (JA) (Gonzalez-Grandio et al., 2017; Dong et al., 2019).
104
105
5
An underlying factor in dominance interaction is the ability of a developing sink to
106
maintain a high growth potential via uptake and metabolism of sugars, including sucrose
107
(Eveland and Jackson, 2012; Barbier et al., 2015; Pfeiffer et al., 2017). For example, the
108
growth potential of shoot and root apices, as well as developing fruits, are dependent
109
upon invertase activity, which functions to metabolize sucrose to glucose and fructose
110
(Ruan et al., 2012; Bihmidine et al., 2013). In addition, sugar catabolic pathways
111
mediated by glycolysis/the tricarboxylic acid and oxidative pentose phosphate pathway
112
also regulate shoot growth (Wang et al., 2021). The demand of growing sinks for
113
carbohydrates is due to the fact that sugars are key drivers of cell division and
114
differentiation required for growth (Ruan et al., 2012; Sablowski and Carnier Dornelas,
115
2014). Indeed, sugar availability plays a role branching (Mason et al., 2014; Barbier et
116
al., 2015), as wells as meristem activity (Wu et al., 2005; Pfeiffer et al., 2016). While
117
sugars are essential for energy and cell wall biosynthesis, glucose and sucrose also
118
function as signals that regulate plant developmental programs (Eveland and Jackson,
119
2012; Barbier et al., 2015), including the vegetative phase transition (Yang et al., 2013;
120
Yu et al., 2013). In addition to sucrose and glucose, trehalose 6-phosphate (T6P)
121
functions as a sugar signal that regulates growth in response to sucrose availability
122
(Nunes et al., 2013; Lastdrager et al., 2014; Baena-Gonzalez and Lunn, 2020). For
123
example, the T6P pathway regulates the vegetative and floral transition in response to
124
the sugar availability to ensure sufficient carbohydrates are accessible to support
125
reproductive development (Wahl et al., 2013; Ponnu et al., 2020). In addition, T6P plays
126
a role in regulating branching and bud outgrowth in response to decapitation (Satoh-
127
Nagasawa et al., 2006; Fichtner et al., 2017). Taken together, sugar signaling and
128
metabolism are key drivers of plant growth and developmental processes.
129
130
In annual plants, inflorescence growth and fruit development coexist for a definite period of
131
time before inflorescence growth ceases (Bleecker and Patterson, 1997; Nooden and
132
Penney, 2001; Gonzalez-Suarez et al., 2020). This developmental transition is referred to as
133
the “end of flowering” phase transition (Gonzalez-Suarez et al., 2020), which is confined to
134
the later stage of inflorescence development (Balanza et al., 2018; Gonzalez-Suarez et al.,
135
2020; Ware et al., 2020). Inflorescence growth cessation is mediated by fruit load, as
136
removing these seed-bearing structures restores flower and fruit production. The inhibition
137
of growth appears to be a separate step from senescence, which usually follows arrest
138
6
(Bleecker and Patterson, 1997; Nooden and Penney, 2001; Wuest et al., 2016; Wang et al.,
139
2020; Ware et al., 2020). A recent hypothesis predicts that inflorescence apices acquire a
140
competency to undergo growth cessation late in inflorescence development (Ware et al.,
141
2020). Once inflorescences acquire this competency, export of auxin from developing fruits
142
induces growth cessation. Competency for inflorescence arrest involves the FRUITFUL
143
(FUL)/APETALA2 (AP2) age dependent module, which indirectly regulates stem-cell
144
homeostasis through the WUSCHEL (WUS) transcription factor (Balanza et al., 2018;
145
Martinez-Fernandez et al., 2020). Interestingly, transcript levels for ABA signaling and
146
response genes associated with lateral bud dormancy are higher in arrested inflorescence
147
meristems at the end of flowering compared to active meristems during the growing phase
148
of inflorescence development (Wuest et al., 2016). In addition, the FUL/AP1 module appears
149
to directly regulate ABA response genes at the end of flowering transition (Martinez-
150
Fernandez et al., 2020). Lastly, experimental studies indicate that JA may also play a role in
151
the end of flowering phase transition (Kim et al., 2013).
152
153
Here, our results show that the end of flowering phase transition is a two-stage process that
154
involves auxin. The first stage is marked by the selective dampening of auxin transport in the
155
apical region of the inflorescence. Further, the transition from the first to the second stage is
156
accompanied by an increase in auxin response in the vascular tissues where developing
157
fruits are attached to the stem. Together, the dampening of auxin transport followed by an
158
increase in auxin response in the apical region of the stem may function to prevent
159
canalization required for flower production and development. Consistent with previous
160
studies showing that sugar metabolism and signaling regulate the vegetative and flower
161
transitions, the first stage of the end of flowering transition is associated with a significant
162
reduction in sugar signaling and metabolism. We propose that inhibition of inflorescence
163
growth by fruit load is regulated by auxin and sugar signaling for end of flowering transition
164
in annual plants.
165
166
RESULTS
167
Characterization of the end of flowering phase transition
168
To better understand the end of flowering phase transition, the inflorescence arrest
169
phenotype was characterized. During the growing phase of inflorescence development, the
170
7
inflorescence meristem produces floral meristems, which give rise to flowers and floral
171
organs, respectively (Fig. 1A). As flowers develop into fruits, the subtending internodes
172
elongate, which separates the siliques. Characterization of the inflorescence arrest
173
phenotype indicated that the end of flowering phase can be divided into two stages. During
174
the first stage, only the apical bud, which consists of the inflorescence meristem, young
175
unopened flower primordia and the immediate subtending internodes, transitioned to a
176
quiescent state (Fig. 1B). In contrast, 4-6 mature flowers with developing fruits attached to
177
elongated pedicels continued to develop (Fig. 1B). For the purposes of this study, we defined
178
the first stage of growth cessation stage as quiescent 1 (Q1). The end of flowering phase was
179
completed at the quiescent 2 (Q2) stage, when growth at the inflorescences apex completely
180
ceased, including the last set of fruits to develop (Fig. 1C).
181
182
In actively growing Arabidopsis inflorescences, the shoot meristem allocates cells that give
183
rise to flowers and internodes (Serrano-Mislata and Sablowski, 2018). The gradual decline in
184
meristem size indicates that meristem activity decreases during inflorescence development
185
(Balanza et al., 2018; Wang et al., 2020). To further support this hypothesis, the average
186
length for the last 30 internodes produced on the primary stem was determined. Results
187
showed that over the course of inflorescence development, internode length gradually
188
declined (Fig. 1D). The steady decline in internode development indicates that the meristem
189
allocates fewer and fewer cells to support stem growth due to a gradual decrease in
190
meristem activity.
191
192
Transition to the Q1 stage is associated with a cessation of meristem activity
193
To evaluate the effect of fruit load on growth processes in the inflorescence apex, a series of
194
mRNA in situ hybridizations were performed with genes that control meristem activity. To
195
demonstrate that cessation of inflorescence growth occurred at the Q1 stage, the cell
196
division marker, CYCLIN DEPENDENT KINASE B1;1 (CDKB1;1) was used as a marker to assess
197
whether the shoot apex was active (Segers et al., 1996). Results showed that CDKB1;1 was
198
expressed in inflorescence and flower meristems, as well as the vasculature of actively
199
growing inflorescence apices (Fig. 2A). In contrast, transcripts for CDKB1;1 were not readily
200
detected in Q1 shoot apices (Fig. 2B). The HISTONE H4 gene, which also serves as a cell
201
division marker (Krizek, 1999; Gaudin et al., 2000), was not expressed in Q1 apices compared
202
8
to active inflorescence apices (data not shown). SHOOTMERISTEMLESS (STM) is a regulator
203
of shoot meristem identity (Long et al., 1996). Therefore, to determine if the fate of the
204
inflorescence meristem cells had changed during Q1, the expression pattern of STM was
205
examined. Results showed that STM was expressed in both active and the Q1 inflorescence
206
and floral meristems (Fig. 2C and D).
207
208
To investigate the impact of fruit load on stem cell homeostasis, the expression pattern for
209
WUSCHEL (WUS) was evaluated in active and Q1 inflorescence apices. In active inflorescence
210
apices, WUS was expressed in the central domain of the inflorescence meristem (Fig. 2E)
211
(Laux et al., 1996; Clark et al., 1997). In contrast to active inflorescence apices, WUS
212
expression was not detected in Q1 meristems (Fig. 2F). MONOPTEROS (MP)/AUXIN
213
RESPONSE FACTOR (ARF5) encodes an auxin response factor that is expressed in the
214
periphery of the shoot meristem where it controls auxin mediated leaf and flower
215
formation, as well as vascular development (Przemeck et al., 1996; Hardtke and Berleth,
216
1998; Schuetz et al., 2008). In active inflorescence apices MP/ARF5 expression was detected
217
in peripheral region of the shoot meristem and vascular tissues of the inflorescence apex
218
(Fig. 2G). Interestingly, the expression pattern of MP/ARF5 was altered in Q1 inflorescence
219
apices, as the mRNA localized to the subapical region of the inflorescence meristem (Fig.
220
2H). Further, MP/ARF5 expression was no longer detected in the periphery of the
221
inflorescence meristem, as well as the quiescent floral meristems and vasculature tissues of
222
the stem and pedicels (Fig. 2H). Taken together, the expression studies show that key
223
determinants of cell division, stem cell homeostasis and auxin-mediated organogenesis are
224
suppressed at the Q1 stage. However, meristem identity is maintained in Q1 meristems, as
225
indicated by the expression of STM.
226
227
Selective inhibition of auxin transport in the apical inflorescence stem correlates with
228
arrest
229
We speculated that the end of flowering transition involved dominance inhibition of
230
inflorescence growth by fruit load. Moreover, dominance inhibition was predicted to
231
correlate with the selective inhibition of auxin transport in the apical region of the stem
232
below the inflorescence apex. To test this hypothesis, basipetal auxin transport was
233
measured in two sets of stem segments during inflorescence development using 14C-indole-
234
9
3-acetic acid (14C-IAA). First, auxin transport was determined in apical stem (AS) segments
235
(Fig. 3A and B), from the stem region just below the inflorescence apex to the site of stem
236
where developing fruits were attached. The region of the stem were developing fruits are
237
attached was referred as the zone of fruit development (ZFD; Fig. 3A and B, white box). In
238
basal stem (BS) segments, auxin transport was also measured below the ZFD (Fig. 3A and B).
239
240
In this analysis, auxin transport was measured early in inflorescence development before
241
fruit set (BFS) and results showed that these stem segments transported an average 20.8
242
fmoles 14C-IAA (Fig. 3C). After 10 fruit set, a significant decline in radiolabeled IAA transport
243
occurred in AS and BS segments (Fig. 3C). The level of 14C-IAA transport was maintained in
244
AS and BS segments up to the time in which 20 fruit set. At the 35-fruit set time point, an
245
apparent further decline in radiolabeled IAA transport occurred (Fig. 3C). At the Q1 stage,
246
auxin transport was severely reduced in AS segments, as these segments transported an
247
average 1.3 fmoles of 14C-IAA (Fig. 3C). Interestingly, the average amount of 14C-IAA
248
transported in Q1 BS segments was 8.1 fmoles, which was similar to the transport capacity
249
of 35 BS segments. In addition, there was an apparent trend in which BS segments displayed
250
higher 14C-IAA transport capacity than AS segments when shoots produced 10, 20 and 35
251
fruits (Fig. 3C). Lastly, removing fruit at the Q2 stage, restored inflorescence growth and IAA
252
transport in AS segments (Supplemental Fig. S1). In summary, these results indicate that
253
fruit load modulates auxin transport and supports the hypothesis that inflorescence growth
254
arrest correlates with a selective dampening of auxin transport in the AS segments at the Q1
255
stage.
256
257
Auxin response increase is primarily associated with Q2 stage of growth cessation
258
To further characterize the role of auxin in meristem arrest, auxin response was examined in
259
active and arrested inflorescences at the Q1 and Q2 stages using the synthetic DR5 auxin
260
responsive promoter fused to the reporter gene beta-glucuronidase (GUS) (Ulmasov et al.,
261
1997). In active inflorescence apices, DR5:GUS activity was detected primarily in young floral
262
buds (Fig. 4A). In addition, DR5:GUS expression was also detected in fruits (Fig. 4A) and
263
pedicels, particularly at the base of the developing fruits (Fig. 4B). At the Q1 stage, DR5:GUS
264
was detected in the dormant floral buds but to a lesser extent compared to active
265
inflorescence apices (Fig. 4C). In addition, DR5:GUS was detected in the last set of
266
10
developing fruits (Fig. 4C). Interestingly, in approximately 37% of Q1 DR5:GUS
267
inflorescences, GUS activity was also apparent in the ZFD and throughout most the pedicels
268
of developing fruits (Fig. 4C and D). GUS activity in the remaining 63% of Q1 DR5:GUS
269
inflorescences was below the level of detection similar to actively growing inflorescences
270
(data not shown). At the Q2 stage, auxin response was detected primarily in the
271
inflorescence stem and pedicels just below the inflorescence apex where the last set of fruit
272
developed (Fig. 4E and F). The “stripe-like” pattern of DR5:GUS activity in the stem suggests
273
that auxin response was induced in the vascular tissue of the main stem and pedicels (Fig.
274
4F). To test this hypothesis, histological experiments were performed in active and Q2
275
inflorescence stems below the inflorescence apex where the last 2-3 fruits had set. Results
276
showed that DR5:GUS was not detected in cross sections through the ZFD during active
277
inflorescence growth (Fig. 4G). In contrast, DR5:GUS activity was readily observed in vascular
278
cells of the Q2 stems where the last set of fruit developed (Fig. 4H). Taken together, results
279
show that auxin response increases in the vascular tissues of the apical stem, which
280
corresponds to the site where the last set of fruit completed their developmental program
281
during the transition from Q1 to Q2.
282
283
Carbohydrate status is reduced in Q1 inflorescence apices
284
As sugar signaling and metabolism plays a critical role in developmental phase transitions
285
(Poethig, 2013; Wang, 2014), and influences auxin signaling and transport (Le et al., 2010;
286
Lilley et al., 2012; Sairanen et al., 2012; Barbier et al., 2015; Lauxmann et al., 2016), we
287
examined the carbohydrate status of inflorescence apices after the transition to the Q1
288
stage. The expression patterns of key genes involved in sugar signaling, metabolism and
289
transport were investigated. TREHALOSE 6-PHOSPHATE SYNTHASE 1 (TPS1) is an essential
290
enzyme that catalyzes T6P from glucose-6-phophate and UDP-glucose (Lastdrager et al.,
291
2014; Baena-Gonzalez and Lunn, 2020). As expression of TPS1 correlates with T6P levels
292
during inflorescence development (Wahl et al., 2013), this biosynthetic gene was used as a
293
marker to assess the T6P pathway in Q1 shoot apices. Results showed that TPS1 was
294
primarily expressed in the vascular system of active inflorescence apices, as well as the
295
flanks of the inflorescence meristem (Fig. 5A). In contrast to actively growing inflorescence
296
apices, TPS1 was not detected in the vascular system or inflorescence meristems in Q1
297
apices (Fig. 5B). Invertases are key enzymes involved in regulating sink activity in meristems
298
11
and fruits (Ruan et al., 2012; Bihmidine et al., 2013). The CYTOSOLIC INVERTASE 1 (CINV1)
299
and related genes in Oryza sativa (OsCYT-INV1), Lotus japonicas (LjINV1) and Solanum
300
lycopersicum (N16) are required for growth and carbon partitioning (Lou et al., 2007; Qi et
301
al., 2007; Jia et al., 2008; Barratt et al., 2009; Welham et al., 2009; Barnes and Anderson,
302
2018; Leskow et al., 2020). In active inflorescence apices, CINV1 was expressed in
303
inflorescence and flower meristems, vascular cells and young floral organ primordia (Fig. 5C).
304
Similar to the results obtained with TPS1, CINV1 expression was not detected in the Q1
305
apices, (Fig. 5D) indicating that sucrose metabolism and sink activity is highly reduced in
306
arrested meristems. To investigate a possible effect of inflorescence growth arrest by fruit
307
load on carbohydrate partitioning, the SUCROSE TRANSPORTER 2 (SUC2) was examined, as
308
this transporter is expressed in the vasculature of inflorescence stems (Truernit and Sauer,
309
1995; Gottwald et al., 2000). Results showed that SUC2 was expressed in the vasculature
310
tissues of active and Q1 apices (Fig. 5E and F). While SUC2 is expressed in Q1 apices, the
311
decrease in CINV1 and TPS1 expression indicates that carbohydrate status was reduced
312
when shoot apices transition to the Q1 stage.
313
314
To further investigate a role for sugar metabolism and signaling in the end of flowering
315
phase transition, glucose, fructose and sucrose were measured in inflorescence apices
316
before and after 15 fruit set, as well as the Q1 stage. In addition, these sugars were
317
measured in developing fruits when 15 fruits set and at the Q1 stage. Results showed that
318
the levels of glucose and fructose were similar in active inflorescence apices before and after
319
15 fruit set (Fig. 6A and B). In contrast, the levels of these monosaccharides were
320
significantly reduced in Q1 inflorescence apices. (Fig. 6A and B). In active and Q1
321
inflorescence apices, the level of sucrose was similar, indicating that sucrose transport was
322
not affected at the Q1 stage (Fig. 6C), which is consistent with the expression of SUC2 in
323
arrested inflorescences. Together, these results indicate that sucrose metabolism but not
324
transport is significantly reduced when inflorescence apices transition from an active to a
325
quiescent state. In developing fruits, the levels of glucose, fructose and sucrose were
326
significantly higher compared to active inflorescence apices (Fig. 6A and B). These results
327
indicate that developing fruits have a higher sink potential than active inflorescence apices
328
during inflorescence development. Interestingly, developing fruits at Q1 had the highest
329
levels of glucose and fructose indicating that sucrose metabolism is increased in fruits at the
330
12
end of flowering phase (Fig. 6A and B). Taken together, results from above indicate that a
331
change in sugar signaling and metabolism in arrested inflorescence apices and developing
332
fruits is associated when active inflorescence apices transition to the Q1 stage during the
333
end of flowering transition.
334
335
336
337
DISCUSSION
338
Dominance interaction between fruits and shoot apices is a major factor that influences
339
shoot architecture and yield (Bangerth, 1989; Smith and Samach, 2013; Walker and Bennett,
340
2018). During inflorescence development, the decline in meristem size and activity,
341
correlates with a decrease in stem cell renewal based on WUS expression (Balanza et al.,
342
2018; Wang et al., 2020). As the growth potential of the inflorescence apex declines, and the
343
meristem become competent for arrest, the end of flowering phase is initiated (Ware et al.,
344
2020). To further extend these studies, our results showed that the end of flowering phase
345
transition is a two-step process. At the Q1 stage, a cessation of growth selectively occurs in
346
the inflorescence apex, while the remaining immature fruits continue to develop. This is
347
supported by expression studies indicating that stem cell renewal and auxin mediated
348
organogenesis in the inflorescence meristem, as well as cell division, are significantly
349
reduced in the Q1 shoot apex. The end of flowering phase is completed at the Q2 stage
350
when fruit growth and development is completed.
351
352
Results from a recent study show that production and export of auxin from developing fruits
353
at a late stage of inflorescence development promotes inflorescence arrest (Ware et al.,
354
2020). The end of flowering model proposed by Ware et al., 2020 predicts that auxin export
355
from developing fruits induces inflorescence arrest by disrupting polar auxin transport in the
356
inflorescence stem. Results from our study show that growing fruits impact auxin transport
357
from the inflorescence apex. First, auxin transport in the apical stem was the highest before
358
fruit set. However, after 10 fruits were produced, a significant decline in auxin transport
359
occurred in AS and BS segments. Second, after 20 fruits were produced, an apparent gradual
360
decrease in auxin transport primarily occurred in AS segments. At the Q1 stage, the
361
inhibition of inflorescence growth correlates with a selective dampening of auxin transport
362
13
in the apical region of the stem below the shoot apex, while transport below the ZFD is
363
functional. Furthermore, an increase in auxin response in the vascular tissues in the ZFD is
364
initiated at Q1 and reaches a maximum at Q2. Finally, removal of fruits at the Q2 stage
365
restores growth and auxin transport in the apical region of the stem. Taken together, we
366
propose that dampening of auxin transport together with an increase in auxin response
367
functions to maintain growth cessation by preventing auxin canalization from the
368
inflorescence apex until the seeds in the developing fruits fully mature.
369
370
Sugar signaling plays an essential role in plant developmental phase transitions (Bolouri
371
Moghaddam and Van den Ende, 2013; Poethig, 2013). Flowering is a major phase transition
372
that is regulated by florigen, FLOWERING LOCUS T (FT), and the age-dependent pathway
373
mediated by the microRNA156 (miR156)/ SQUAMOSA PROMOTER BINDING PROTEIN-LIKE
374
(SPL) module (Turck et al., 2008; Srikanth and Schmid, 2011; Wang, 2014). Experimental
375
studies show that T6P pathway regulates FT and the miR156/SPL module in leaves and shoot
376
apices, respectively (Wahl et al., 2013). In addition, the suppression of miR156 during the
377
vegetative phase transition is mediated by sugars, including glucose and sucrose (Yang et al.,
378
2013; Yu et al., 2013), as well as the T6P pathway (Ponnu et al., 2020). In our study, we show
379
that glucose and fructose levels, as well as TSP1 expression, are significantly reduced in Q1
380
apices compared to actively growing inflorescence apices. As the biosynthesis of T6P is
381
dependent upon glucose and TSP1 expression, our results indicate that the T6P pathway is
382
highly reduced in Q1 apices. Therefore, we propose that the suppression of sugar signaling
383
mediated by glucose and the T6P pathway is involved in the end of flowering phase
384
transition.
385
386
The reduction in glucose and fructose but not sucrose in Q1 apices at the end of flowering
387
indicates that sucrose metabolism is selectively inhibited in inflorescence apices but not
388
developing fruits. Consistent with this view, CINV1 expression is suppressed when
389
inflorescence apices transition to the Q1 stage. In contrast, developing fruits at Q1 stage
390
display an increase in sucrose metabolism compared to growing fruits at an earlier stage of
391
inflorescence development. This is supported by the fact that the levels of glucose and
392
fructose are higher, while sucrose is lower in developing fruits at the Q1 stage. Therefore, it
393
is tempting to speculate that end of flowering phase transition not only functions to repress
394
14
inflorescence growth, but also acts to further increase the growth potential of fruits. The
395
increase in the growth potential of developing fruit at the Q1 stage may be mediated by the
396
end of flowering-competence factors.
397
398
Expanding on the model proposed by Ware et al., 2020, we propose that inflorescence
399
growth is maintained by: (1) the basipetal auxin transport system in the apical domain of the
400
inflorescence stem and (2) sugar signaling and metabolism in the inflorescence apex. During
401
the active stage of inflorescence development, the shoot system can support both shoot and
402
fruit growth. However, with the continuous decline in meristem activity and auxin transport
403
out of the inflorescence apex, a competence juncture is reached in which the apical bud
404
including the inflorescence and floral meristems, immature flowers and subtending
405
internodes, can no longer maintain growth, as fruits continue to develop. At this
406
competence juncture, export of auxin from developing fruits selectively impairs auxin
407
transport in the apical stem below the inflorescence apex, which induces inflorescence
408
arrest. We propose that impairment of auxin transport in the apical inflorescence stem
409
suppresses sugar signaling and metabolism in the inflorescence apex, which negatively
410
impacts stem-cell renewal and organogenesis. This is supported by the fact that auxin
411
mediated organogenesis and stem cell renewal is dependent upon sugar availability and
412
signaling (Lauxmann et al., 2016; Pfeiffer et al., 2016). Further, the increase in auxin
413
response in the vascular tissues of the apical stem from the Q1 to the Q2 stage acts to
414
maintain growth arrest by dampening auxin export from the inflorescence meristem and
415
immature floral buds at the Q2 stage.
416
417
In fruit tree crops, inhibition of shoot growth by fruit load is a major driver of biennial or
418
alternate bearing, which is a major challenge for fruit tree crop industries worldwide
419
(Samach and Smith, 2013; Smith and Samach, 2013). Due to significant challenges and
420
barriers associated with the usage of genetic and molecular manipulations in fruit tree crops,
421
Arabidopsis may serve as a model system to understand the physiological basis of shoot
422
growth arrest in response to fruit load. Translational research from Arabidopsis to fruit tree
423
crops can be utilized to develop new innovative tools to limit the impact of fruits on shoot
424
growth in order to maximize yield and reduce seasonal variation.
425
426
15
MATERIALS AND METHODS
427
Plant materials and growth conditions
428
The Arabidopsis thaliana Columbia-0 (Col-0) accession was used to characterize
429
inflorescence arrest in response to fruit load. Auxin response was evaluated during
430
inflorescence development using the DR5:GUS system (Ulmasov et al., 1997). Plants were
431
grown at 220C under long day growth conditions, 16-hour light/8-hour dark.
432
433
Internode measurements
434
The length of the last 30 internodes were measured in the primary inflorescence after the
435
end of flowering transition was completed in 30 plants. The average length in mm and
436
standard deviation for each internode was determined.
437
438
Gene expression analyses
439
To examine the expression pattern of key genes that regulate meristem activity and sugar
440
signaling, in situ hybridization was performed using a standard method of fixation,
441
sectioning and mRNA hybridization as previously described (Jackson, 2001; Chuck et al.,
442
2002). Active inflorescence apices were harvested after 5-10 fruits were produced.
443
Quiescent apices were harvested at the Q1 stage of the end of flowering transition.
444
Synthesis of UTP-digoxigenin anti-sense probes were previously described for STM (Long et
445
al., 1996), WUS (Yadav et al., 2009), MP (Zhao et al., 2010) and TPS1 (Zhao et al., 2010).
446
Primer sequences were used to PCR amplify CDKB1;1, CINV1 and SUC2 DNA fragments for
447
the synthesis of UTP-digoxigenin antisense probes. The sequences for CDKB1;1, CINV1 and
448
SUC2
primers
were
CDKB1;1-F
(CGAGATGGACGAAGAAGGTATTCCACC),
CDKB1;1-R
449
(GAAATAATACGACTCACTATAGGGACTCGTGAGAAGATCAACTCCTTGAGGTG),
CINV1-F
450
(CCGATGGAGATGGCAGAGAGG),
CINV1-R
451
(GAAATAATACGACTCACTATAGGGACTGGCCAAGACGCAGATCGCTTGATGAC),
SUC2-F
452
(CTGAGTCATGCGATCTCTACTGCG)
and
SUC2-R
453
(GAAATAATACGACTCACTATAGGGACTCTTACCGCTGCCGCAATCGCTCC).
The
method
to
454
visualize GUS activity in DR5:GUS inflorescence shoots was described previously (Sundaresan
455
et al., 1995; Springer et al., 2000). DR5:GUS was evaluated in inflorescences during the active
456
period of inflorescence development when apices produced 10-20 fruits, as well as the Q1
457
and Q2 stages.
458
16
459
14C-IAA transport assay
460
Basipetal auxin transport was measured in inflorescence stems using a 14C-IAA protocol
461
previously described (Lewis and Muday, 2009). Briefly, 20 mm stem segments were
462
harvested from inflorescences before fruit set. After the shoot apex produced 10, 20 and 35
463
fruits and at the Q1 and Q2 stages of development, AS and BS segments were isolated from
464
each inflorescence as described in the results section. To measure auxin transport in each
465
stem segment, the apical end of the stem was placed in 20 µL of auxin transport buffer (100
466
nM 14C-IAA, 0.05% MES, pH 5.7). To measure movement mediated by diffusion, a separate
467
set of stem segments were isolated and the apical end of each stem was placed in auxin
468
transport buffer containing naphthylphthalamic acid (NPA) to a final concentration of 10
469
µM. After 10 hours of auxin transport at 220C, each segment was removed from the auxin
470
transport buffer (+/- NPA) and a 5 mm section at the apical end was cleaved and discarded.
471
Next, each stem segment was transferred to an Eppendorf tube and ground in scintillation
472
fluid using a plastic pestle. For each sample, the extract from was transferred to a single
473
scintillation vial containing 20 mL of scintillation fluid and 14C d.p.m. was determined before
474
conversion to fmoles. Five biological replicates were used to calculate the mean and
475
standard deviation. Analysis of variance and Tukey’s honest significant difference analysis
476
was performed using standard statistical packages in R.
477
478
Sugar measurements
479
For sugar extraction, ~100 mg of inflorescence apices and developing fruit were collected
480
during inflorescence development in triplicate. After collection, the material was freeze-
481
dried for 12 h and the dry weight for each sample was determined. Each sample was ground
482
with a mortar and pestle in 1.0 mL of 80% ethanol. Next, samples were incubated at 800C for
483
30 minutes to extract soluble sugars. After the insoluble material was pelleted at 10,000 xg
484
and the supernatant decanted, the tissues were re-extracted two more times with 80%
485
ethanol. After combining and mixing the three separate 80% ethanol extracts, 650 mL of the
486
soluble extract was placed in an Eppendorf tube and dried in a Gene miVac Quattro (SP
487
Industries, Warminster, PA, USA) for 1.5 h at 550C. Each dried sample was resuspended in 20
488
µL sterile H2O. Glucose, fructose and sucrose were separated by High Performance Liquid
489
Chromatography using the Sugar-Pak cation-exchange column (Waters, Rydalmere, NSW,
490
17
AUS). The Aglient Technologies 1200 G1362A infinity refractive index detector (Santa Clara,
491
California, USA) was used to identify and quantify separated sugars in each of the samples by
492
comparison to the glucose, fructose and sucrose standards. Analysis of variance and Tukey’s
493
honest significant difference analysis was performed using standard statistical packages in R.
494
495
496
497
498
ACKNOWLEDGMENTS
499
We thank Tom Bennett (University of Leeds, UK) for discussions and reviewing the
500
manuscript. We also thank Kate Tepper and Rhys Webber for the maintenance of
501
Arabidopsis plants and Dr Tom Guilfoyle for providing DR5:GUS seed.
502
503
FIGURE LEGENDS
504
Figure 1. Characterization of the end of flowering transition. (A) An active inflorescence
505
shoot apex containing numerous flowers at different stages of development. (B) Q1
506
inflorescence shoot apex with a white arrow pointing at the compact quiescent apex, which
507
consists of young unopened flower buds. The asterisks mark mature flowers with developing
508
fruit attached to elongated pedicels. (C) Image inflorescence shoot apex at the Q2 stage of
509
development, in which growth at the apex, including the fruits, ceased. (D) Average
510
internode length was determined for the last 30 internodes produced after the end of
511
flowering transition. Note: internode 30 is the last internode to elongate before Q1 stage
512
arrest.
513
514
Figure 2. Expression patterns for key genes that control shoot growth. Gene expression
515
patterns were shown for (A, C, E and G) active and (B, D, F and H) Q1 apices. (A and B)
516
CDKB1;1 is a mitotic regulator expressed in shoot meristem (Segers et al., 1996). (C and D)
517
STM is a shoot meristem identity gene (Long et al., 1996). (E and F) WUS is key regulator of
518
stem cell homeostasis (Laux et al., 1996). (G and H) MP regulates flower formation and
519
vascular development (Hardtke and Berleth, 1998). (A) The length of the bar is 50 µm. (H)
520
Arrow points at the MP expression in the sub-apical region of the meristem and pith cells in
521
the Q1 apex.
522
18
523
Figure 3. 14C-IAA transport in stem segments during inflorescence development. Images of
524
(A) active and (B) Q1 inflorescence shoots. The white box marks the region of the stem
525
where developing fruits are attached. This region is referred as the zone of fruit
526
development (ZFD). IAA transport was determined in apical stem (AS) segments and basal
527
stem (BS) segments relative to the ZFD. (C) Radiolabeled IAA transport was measured during
528
inflorescence development starting at a time just before the first fruit set (BFS). After fruit
529
set, 14C-IAA was determined in apical stem (AS) and basal stem (BS) segments when 10, 20
530
and 35 fruits were produced, as well as the Q1 stage. The light colour boxes represent
531
control stem segments in which 14C-IAA transport was measured in the presence of
532
naphthyphthalamic acid (NPA), an inhibitor of polar auxin transport. The letters above the
533
bars determine whether differences in 14C-IAA transport were statistically significant using
534
analysis of variance, Tukey’s honest significant difference.
535
536
Figure 4. A change in auxin response in apical inflorescence stems at end of flowering
537
transition. DR5:GUS staining patterns in (A) active, (C) Q1 and (E) Q2 shoot apices. Close up
538
of stems where developing fruits are attached in (B) active, (D) Q1 and (F) Q2 shoot apices.
539
Histological cross sections in the apical stem for an (G) active and (H) Q2 stage inflorescence.
540
Note: inset of vascular bundle displayed in upper right corner of each image. In actively
541
growing shoots, the section shown in (G) was through the region of the stem where the
542
developing fruits were attached. The section displayed in (H) was in a region of the stem
543
where the last fruits set.
544
545
Figure 5. Expression patterns for sugar signaling, metabolism and transport genes. Gene
546
expression patterns were determined in (A, C and E) active and (B, D and F) Q1 apices. (A
547
and B) TPS1, which is expressed in the vasculature of the inflorescence stem, was used as a
548
marker to assess the T6P pathway (Wahl et al., 2013; Ponnu et al., 2020). (C and D) CINV1 is
549
an invertase, which is required for growth (Barratt et al., 2009). (E and F) SUC2 is a sucrose
550
transporter expressed in the vascular tissues of inflorescences (Truernit and Sauer, 1995;
551
Gottwald et al., 2000). The length of the bar is 50 µm.
552
553
19
Figure 6. Sugar content in shoot apices and developing fruits. Sugar levels determined in
554
shoot apices before fruit set (BFS Apex), after 15 fruit set (15FS Apex) and at the Q1 stage
555
(Q1 Apex). Sugar levels were also determined in developing fruits when 15 fruits set (15FS
556
Fruit) and at the Q1 stage (Q1 Fruit). The levels of (A) Glucose, (B) fructose and (C) sucrose
557
were measured from dry weight (DW) tissue. The letters above the bars determine whether
558
differences in sugar levels were statistically significant using analysis of variance, Tukey’s
559
honest significant difference.
560
20
LITERATURE CITED
561
Aguilar-Martinez JA, Poza-Carrion C, Cubas P (2007) Arabidopsis BRANCHED1 acts as
562
an integrator of branching signals within axillary buds. Plant Cell 19: 458-472
563
Baena-Gonzalez E, Lunn JE (2020) SnRK1 and trehalose 6-phosphate - two ancient
564
pathways converge to regulate plant metabolism and growth. Curr Opin Plant Biol 55:
565
52-59
566
Balanza V, Martinez-Fernandez I, Sato S, Yanofsky MF, Kaufmann K, Angenent GC,
567
Bemer M, Ferrandiz C (2018) Genetic control of meristem arrest and life span in
568
Arabidopsis by a FRUITFULL-APETALA2 pathway. Nat Commun 9: 565
569
Bangerth F (1989) Dominance among fruits/sinks and the search for a correlative signal.
570
Physiologia Plantarum 76: 608-614
571
Bangerth F, Li CJ, Gruber J (2000) Mutual interaction of auxin and cytokinin in regulating
572
correlative dominance. Plant Growth Regulation 32: 205-217
573
Barbier F, Peron T, Lecerf M, Perez-Garcia MD, Barriere Q, Rolcik J, Boutet-Mercey
574
S, Citerne S, Lemoine R, Porcheron B, Roman H, Leduc N, Le Gourrierec J,
575
Bertheloot J, Sakr S (2015) Sucrose is an early modulator of the key hormonal
576
mechanisms controlling bud outgrowth in Rosa hybrida. J Exp Bot 66: 2569-2582
577
Barbier FF, Dun EA, Kerr SC, Chabikwa TG, Beveridge CA (2019) An Update on the
578
Signals Controlling Shoot Branching. Trends Plant Sci 24: 220-236
579
Barbier FF, Lunn JE, Beveridge CA (2015) Ready, steady, go! A sugar hit starts the race to
580
shoot branching. Curr Opin Plant Biol 25: 39-45
581
Barnes WJ, Anderson CT (2018) Cytosolic invertases contribute to cellulose biosynthesis
582
and influence carbon partitioning in seedlings of Arabidopsis thaliana. Plant J 94: 956-
583
974
584
Barratt DH, Derbyshire P, Findlay K, Pike M, Wellner N, Lunn J, Feil R, Simpson C,
585
Maule AJ, Smith AM (2009) Normal growth of Arabidopsis requires cytosolic
586
invertase but not sucrose synthase. Proc Natl Acad Sci U S A 106: 13124-13129
587
Bennett T, Hines G, van Rongen M, Waldie T, Sawchuk MG, Scarpella E, Ljung K,
588
Leyser O (2016) Connective Auxin Transport in the Shoot Facilitates Communication
589
between Shoot Apices. PLoS Biol 14: e1002446
590
Bihmidine S, Hunter CT, 3rd, Johns CE, Koch KE, Braun DM (2013) Regulation of
591
assimilate import into sink organs: update on molecular drivers of sink strength. Front
592
Plant Sci 4: 177
593
Bleecker AB, Patterson SE (1997) Last exit: senescence, abscission, and meristem arrest in
594
Arabidopsis. Plant Cell 9: 1169-1179
595
Bolouri Moghaddam MR, Van den Ende W (2013) Sugars, the clock and transition to
596
flowering. Front Plant Sci 4: 22
597
Braun N, de Saint Germain A, Pillot JP, Boutet-Mercey S, Dalmais M, Antoniadi I, Li
598
X, Maia-Grondard A, Le Signor C, Bouteiller N, Luo D, Bendahmane A,
599
Turnbull C, Rameau C (2012) The pea TCP transcription factor PsBRC1 acts
600
downstream of Strigolactones to control shoot branching. Plant Physiol 158: 225-238
601
Chuck G, Muszynski M, Kellogg E, Hake S, Schmidt RJ (2002) The control of spikelet
602
meristem identity by the branched silkless1 gene in maize. Science 298: 1238-1241
603
Clark SE, Williams RW, Meyerowitz EM (1997) The CLAVATA1 gene encodes a putative
604
receptor kinase that controls shoot and floral meristem size in Arabidopsis. Cell 89:
605
575-585
606
Crawford S, Shinohara N, Sieberer T, Williamson L, George G, Hepworth J, Muller D,
607
Domagalska MA, Leyser O (2010) Strigolactones enhance competition between
608
shoot branches by dampening auxin transport. Development 137: 2905-2913
609
21
Dong Z, Xiao Y, Govindarajulu R, Feil R, Siddoway ML, Nielsen T, Lunn JE, Hawkins
610
J, Whipple C, Chuck G (2019) The regulatory landscape of a core maize
611
domestication module controlling bud dormancy and growth repression. Nat Commun
612
10: 3810
613
Eveland AL, Jackson DP (2012) Sugars, signalling, and plant development. J Exp Bot 63:
614
3367-3377
615
Fichtner F, Barbier FF, Feil R, Watanabe M, Annunziata MG, Chabikwa TG, Hofgen
616
R, Stitt M, Beveridge CA, Lunn JE (2017) Trehalose 6-phosphate is involved in
617
triggering axillary bud outgrowth in garden pea (Pisum sativum L.). Plant J 92: 611-
618
623
619
Galweiler L, Guan C, Muller A, Wisman E, Mendgen K, Yephremov A, Palme K (1998)
620
Regulation of polar auxin transport by AtPIN1 in Arabidopsis vascular tissue. Science
621
282: 2226-2230
622
Gaudin V, Lunness PA, Fobert PR, Towers M, Riou-Khamlichi C, Murray JA, Coen E,
623
Doonan JH (2000) The expression of D-cyclin genes defines distinct developmental
624
zones in snapdragon apical meristems and is locally regulated by the Cycloidea gene.
625
Plant Physiol 122: 1137-1148
626
Gonzalez-Grandio E, Pajoro A, Franco-Zorrilla JM, Tarancon C, Immink RG, Cubas P
627
(2017) Abscisic acid signaling is controlled by a BRANCHED1/HD-ZIP I cascade in
628
Arabidopsis axillary buds. Proc Natl Acad Sci U S A 114: E245-E254
629
Gonzalez-Grandio E, Poza-Carrion C, Sorzano CO, Cubas P (2013) BRANCHED1
630
promotes axillary bud dormancy in response to shade in Arabidopsis. Plant Cell 25:
631
834-850
632
Gonzalez-Suarez P, Walker CH, Bennett T (2020) Bloom and bust: understanding the
633
nature and regulation of the end of flowering. Curr Opin Plant Biol 57: 24-30
634
Gottwald JR, Krysan PJ, Young JC, Evert RF, Sussman MR (2000) Genetic evidence for
635
the in planta role of phloem-specific plasma membrane sucrose transporters. Proc Natl
636
Acad Sci U S A 97: 13979-13984
637
Guo W, Chen L, Herrera-Estrella L, Cao D, Tran LP (2020) Altering Plant Architecture
638
to Improve Performance and Resistance. Trends Plant Sci 25: 1154-1170
639
Hardtke CS, Berleth T (1998) The Arabidopsis gene MONOPTEROS encodes a
640
transcription factor mediating embryo axis formation and vascular development.
641
EMBO J 17: 1405-1411
642
Jackson D (2001) In-situ hybridization in plants. In SJ Gurr, MJ McPherson, DJ Bowles, eds,
643
Molecular Plant Pathology: A Practical Approach, Vol 1. Oxford University Press,
644
Oxford, pp 163-174
645
Jia L, Zhang B, Mao C, Li J, Wu Y, Wu P, Wu Z (2008) OsCYT-INV1 for alkaline/neutral
646
invertase is involved in root cell development and reproductivity in rice (Oryza sativa
647
L.). Planta 228: 51-59
648
Jiang L, Liu X, Xiong G, Liu H, Chen F, Wang L, Meng X, Liu G, Yu H, Yuan Y, Yi W,
649
Zhao L, Ma H, He Y, Wu Z, Melcher K, Qian Q, Xu HE, Wang Y, Li J (2013)
650
DWARF 53 acts as a repressor of strigolactone signalling in rice. Nature 504: 401-405
651
Kim J, Dotson B, Rey C, Lindsey J, Bleecker AB, Binder BM, Patterson SE (2013) New
652
clothes for the jasmonic acid receptor COI1: delayed abscission, meristem arrest and
653
apical dominance. PLoS One 8: e60505
654
Krizek BA (1999) Ectopic expression of AINTEGUMENTA in Arabidopsis plants results in
655
increased growth of floral organs. Dev Genet 25: 224-236
656
Lastdrager J, Hanson J, Smeekens S (2014) Sugar signals and the control of plant growth
657
and development. J Exp Bot 65: 799-807
658
22
Laux T, Mayer KF, Berger J, Jurgens G (1996) The WUSCHEL gene is required for shoot
659
and floral meristem integrity in Arabidopsis. Development 122: 87-96
660
Lauxmann MA, Annunziata MG, Brunoud G, Wahl V, Koczut A, Burgos A, Olas JJ,
661
Maximova E, Abel C, Schlereth A, Soja AM, Blasing OE, Lunn JE, Vernoux T,
662
Stitt M (2016) Reproductive failure in Arabidopsis thaliana under transient
663
carbohydrate limitation: flowers and very young siliques are jettisoned and the
664
meristem is maintained to allow successful resumption of reproductive growth. Plant
665
Cell Environ 39: 745-767
666
Le CS, Schmelz EA, Chourey PS (2010) Sugar levels regulate tryptophan-dependent auxin
667
biosynthesis in developing maize kernels. Plant Physiol 153: 306-318
668
Leskow CC, Conte M, Del Pozo T, Bermudez L, Lira BS, Gramegna G, Baroli I, Burgos
669
E, Zavallo D, Kamenetzky L, Asis R, Gonzalez M, Fernie AR, Rossi M, Osorio S,
670
Carrari F (2020) The cytosolic invertase NI6 affects vegetative growth, flowering,
671
fruit set and yield in tomato. J Exp Bot
672
Lewis DR, Muday GK (2009) Measurement of auxin transport in Arabidopsis thaliana. Nat
673
Protoc 4: 437-451
674
Lilley JL, Gee CW, Sairanen I, Ljung K, Nemhauser JL (2012) An endogenous carbon-
675
sensing pathway triggers increased auxin flux and hypocotyl elongation. Plant Physiol
676
160: 2261-2270
677
Long JA, Moan EI, Medford JI, Barton MK (1996) A member of the KNOTTED class of
678
homeodomain proteins encoded by the STM gene of Arabidopsis. Nature 379: 66-69
679
Lou Y, Gou JY, Xue HW (2007) PIP5K9, an Arabidopsis phosphatidylinositol
680
monophosphate kinase, interacts with a cytosolic invertase to negatively regulate
681
sugar-mediated root growth. Plant Cell 19: 163-181
682
Martinez-Fernandez I, Menezes de Moura S, Alves-Ferreira M, Ferrandiz C, Balanza V
683
(2020) Identification of Players Controlling Meristem Arrest Downstream of the
684
FRUITFULL-APETALA2 Pathway. Plant Physiol 184: 945-959
685
Mason MG, Ross JJ, Babst BA, Wienclaw BN, Beveridge CA (2014) Sugar demand, not
686
auxin, is the initial regulator of apical dominance. Proc Natl Acad Sci U S A 111:
687
6092-6097
688
Montgomery BL (2008) Right place, right time: Spatiotemporal light regulation of plant
689
growth and development. Plant Signal Behav 3: 1053-1060
690
Muller D, Leyser O (2011) Auxin, cytokinin and the control of shoot branching. Ann Bot
691
107: 1203-1212
692
Nooden LD, Penney JP (2001) Correlative controls of senescence and plant death in
693
Arabidopsis thaliana (Brassicaceae). J Exp Bot 52: 2151-2159
694
Nunes C, O'Hara LE, Primavesi LF, Delatte TL, Schluepmann H, Somsen GW, Silva
695
AB, Fevereiro PS, Wingler A, Paul MJ (2013) The trehalose 6-phosphate/SnRK1
696
signaling pathway primes growth recovery following relief of sink limitation. Plant
697
Physiol 162: 1720-1732
698
Pfeiffer A, Janocha D, Dong Y, Medzihradszky A, Schone S, Daum G, Suzaki T, Forner
699
J, Langenecker T, Rempel E, Schmid M, Wirtz M, Hell R, Lohmann JU (2016)
700
Integration of light and metabolic signals for stem cell activation at the shoot apical
701
meristem. Elife 5
702
Pfeiffer A, Wenzl C, Lohmann JU (2017) Beyond flexibility: controlling stem cells in an
703
ever changing environment. Curr Opin Plant Biol 35: 117-123
704
Poethig RS (2013) Vegetative phase change and shoot maturation in plants. Curr Top Dev
705
Biol 105: 125-152
706
23
Ponnu J, Schlereth A, Zacharaki V, Dzialo MA, Abel C, Feil R, Schmid M, Wahl V
707
(2020) The trehalose 6-phosphate pathway impacts vegetative phase change in
708
Arabidopsis thaliana. Plant J 104: 768-780
709
Przemeck GK, Mattsson J, Hardtke CS, Sung ZR, Berleth T (1996) Studies on the role of
710
the Arabidopsis gene MONOPTEROS in vascular development and plant cell
711
axialization. Planta 200: 229-237
712
Qi X, Wu Z, Li J, Mo X, Wu S, Chu J, Wu P (2007) AtCYT-INV1, a neutral invertase, is
713
involved in osmotic stress-induced inhibition on lateral root growth in Arabidopsis.
714
Plant Mol Biol 64: 575-587
715
Rameau C, Bertheloot J, Leduc N, Andrieu B, Foucher F, Sakr S (2014) Multiple
716
pathways regulate shoot branching. Front Plant Sci 5: 741
717
Ruan YL, Patrick JW, Bouzayen M, Osorio S, Fernie AR (2012) Molecular regulation of
718
seed and fruit set. Trends Plant Sci 17: 656-665
719
Sablowski R, Carnier Dornelas M (2014) Interplay between cell growth and cell cycle in
720
plants. J Exp Bot 65: 2703-2714
721
Sairanen I, Novak O, Pencik A, Ikeda Y, Jones B, Sandberg G, Ljung K (2012) Soluble
722
carbohydrates regulate auxin biosynthesis via PIF proteins in Arabidopsis. Plant Cell
723
24: 4907-4916
724
Salazar-García S, Garner LC, Lovatt CJ (2013) Reproductive Biology. In B Schaffer, BN
725
Wolstenholme, AW Whiley, eds, The Avocado, Ed 2nd. CABI, Wallingford
726
Oxfordshire OX10 8DE
727
UK, pp 118-167
728
Salazar-García S, Lord EM, Lovatt CJ (1998) Inflorescence and flower development of the
729
‘Hass’ avocado (Persea americana Mill.) during ‘on’ an d’off’ crop years. Journal of
730
the American Society for Horticulture Science 123: 537-544
731
Samach A, Smith HM (2013) Constraints to obtaining consistent annual yields in perennials.
732
II: Environment and fruit load affect induction of flowering. Plant Sci 207: 168-176
733
Satoh-Nagasawa N, Nagasawa N, Malcomber S, Sakai H, Jackson D (2006) A trehalose
734
metabolic enzyme controls inflorescence architecture in maize. Nature 441: 227-230
735
Sawicki M, Ait Barka E, Clement C, Vaillant-Gaveau N, Jacquard C (2015) Cross-talk
736
between environmental stresses and plant metabolism during reproductive organ
737
abscission. J Exp Bot 66: 1707-1719
738
Schneider A, Godin C, Boudon F, Demotes-Mainard S, Sakr S, Bertheloot J (2019) Light
739
Regulation of Axillary Bud Outgrowth Along Plant Axes: An Overview of the Roles
740
of Sugars and Hormones. Front Plant Sci 10: 1296
741
Schuetz M, Berleth T, Mattsson J (2008) Multiple MONOPTEROS-dependent pathways
742
are involved in leaf initiation. Plant Physiol 148: 870-880
743
Segers G, Gadisseur I, Bergounioux C, de Almeida Engler J, Jacqmard A, Van Montagu
744
M, Inze D (1996) The Arabidopsis cyclin-dependent kinase gene cdc2bAt is
745
preferentially expressed during S and G2 phases of the cell cycle. Plant J 10: 601-612
746
Serrano-Mislata A, Sablowski R (2018) The pillars of land plants: new insights into stem
747
development. Curr Opin Plant Biol 45: 11-17
748
Shen J, Zhang Y, Ge D, Wang Z, Song W, Gu R, Che G, Cheng Z, Liu R, Zhang X
749
(2019) CsBRC1 inhibits axillary bud outgrowth by directly repressing the auxin efflux
750
carrier CsPIN3 in cucumber. Proc Natl Acad Sci U S A 116: 17105-17114
751
Shinohara N, Taylor C, Leyser O (2013) Strigolactone can promote or inhibit shoot
752
branching by triggering rapid depletion of the auxin efflux protein PIN1 from the
753
plasma membrane. PLoS Biol 11: e1001474
754
24
Smith HM, Samach A (2013) Constraints to obtaining consistent annual yields in perennial
755
tree crops. I: Heavy fruit load dominates over vegetative growth. Plant Sci 207: 158-
756
167
757
Soundappan I, Bennett T, Morffy N, Liang Y, Stanga JP, Abbas A, Leyser O, Nelson
758
DC (2015) SMAX1-LIKE/D53 Family Members Enable Distinct MAX2-Dependent
759
Responses to Strigolactones and Karrikins in Arabidopsis. Plant Cell 27: 3143-3159
760
Springer PS, Holding DR, Groover A, Yordan C, Martienssen RA (2000) The essential
761
Mcm7 protein PROLIFERA is localized to the nucleus of dividing cells during the
762
G(1) phase and is required maternally for early Arabidopsis development.
763
Development 127: 1815-1822
764
Srikanth A, Schmid M (2011) Regulation of flowering time: all roads lead to Rome. Cell
765
Mol Life Sci 68: 2013-2037
766
Sundaresan V, Springer P, Volpe T, Haward S, Jones JD, Dean C, Ma H, Martienssen R
767
(1995) Patterns of gene action in plant development revealed by enhancer trap and
768
gene trap transposable elements. Genes Dev 9: 1797-1810
769
Teichmann T, Muhr M (2015) Shaping plant architecture. Front Plant Sci 6: 233
770
Truernit E, Sauer N (1995) The promoter of the Arabidopsis thaliana SUC2 sucrose-H+
771
symporter gene directs expression of beta-glucuronidase to the phloem: evidence for
772
phloem loading and unloading by SUC2. Planta 196: 564-570
773
Turck F, Fornara F, Coupland G (2008) Regulation and identity of florigen: FLOWERING
774
LOCUS T moves center stage. Annu Rev Plant Biol 59: 573-594
775
Ulmasov T, Murfett J, Hagen G, Guilfoyle TJ (1997) Aux/IAA proteins repress expression
776
of reporter genes containing natural and highly active synthetic auxin response
777
elements. Plant Cell 9: 1963-1971
778
van Rongen M, Bennett T, Ticchiarelli F, Leyser O (2019) Connective auxin transport
779
contributes to strigolactone-mediated shoot branching control independent of the
780
transcription factor BRC1. PLoS Genet 15: e1008023
781
Wahl V, Ponnu J, Schlereth A, Arrivault S, Langenecker T, Franke A, Feil R, Lunn JE,
782
Stitt M, Schmid M (2013) Regulation of flowering by trehalose-6-phosphate
783
signaling in Arabidopsis thaliana. Science 339: 704-707
784
Walker CH, Bennett T (2018) Forbidden fruit: Dominance relationships and the control of
785
shoot architecture. Annual Plant Reviews 1: 217-254
786
Wang JW (2014) Regulation of flowering time by the miR156-mediated age pathway. J Exp
787
Bot 65: 4723-4730
788
Wang L, Wang B, Jiang L, Liu X, Li X, Lu Z, Meng X, Wang Y, Smith SM, Li J (2015)
789
Strigolactone Signaling in Arabidopsis Regulates Shoot Development by Targeting
790
D53-Like SMXL Repressor Proteins for Ubiquitination and Degradation. Plant Cell
791
27: 3128-3142
792
Wang L, Wang B, Yu H, Guo H, Lin T, Kou L, Wang A, Shao N, Ma H, Xiong G, Li X,
793
Yang J, Chu J, Li J (2020) Transcriptional regulation of strigolactone signalling in
794
Arabidopsis. Nature 583: 277-281
795
Wang M, Perez-Garcia MD, Daviere JM, Barbier F, Oge L, Gentilhomme J, Voisine L,
796
Peron T, Launay-Avon A, Clement G, Baumberger N, Balzergue S, Macherel D,
797
Grappin P, Bertheloot J, Achard P, Hamama L, Sakr S (2021) Axillary bud
798
outgrowth in rose is controlled by sugar metabolic and signalling pathways. J Exp Bot
799
Wang Y, Kumaishi K, Suzuki T, Ichihashi Y, Yamaguchi N, Shirakawa M, Ito T (2020)
800
Morphological and Physiological Framework Underlying Plant Longevity in
801
Arabidopsis thaliana. Front Plant Sci 11: 600726
802
25
Ware A, Walker CH, Simura J, Gonzalez-Suarez P, Ljung K, Bishopp A, Wilson ZA,
803
Bennett T (2020) Auxin export from proximal fruits drives arrest in temporally
804
competent inflorescences. Nat Plants 6: 699-707
805
Welham T, Pike J, Horst I, Flemetakis E, Katinakis P, Kaneko T, Sato S, Tabata S,
806
Perry J, Parniske M, Wang TL (2009) A cytosolic invertase is required for normal
807
growth and cell development in the model legume, Lotus japonicus. J Exp Bot 60:
808
3353-3365
809
Wu X, Dabi T, Weigel D (2005) Requirement of homeobox gene STIMPY/WOX9 for
810
Arabidopsis meristem growth and maintenance. Curr Biol 15: 436-440
811
Wuest SE, Philipp MA, Guthorl D, Schmid B, Grossniklaus U (2016) Seed Production
812
Affects Maternal Growth and Senescence in Arabidopsis. Plant Physiol 171: 392-404
813
Yadav RK, Girke T, Pasala S, Xie M, Reddy GV (2009) Gene expression map of the
814
Arabidopsis shoot apical meristem stem cell niche. Proc Natl Acad Sci U S A 106:
815
4941-4946
816
Yang L, Xu M, Koo Y, He J, Poethig RS (2013) Sugar promotes vegetative phase change in
817
Arabidopsis thaliana by repressing the expression of MIR156A and MIR156C. Elife
818
2: e00260
819
Yu S, Cao L, Zhou CM, Zhang TQ, Lian H, Sun Y, Wu J, Huang J, Wang G, Wang JW
820
(2013) Sugar is an endogenous cue for juvenile-to-adult phase transition in plants.
821
Elife 2: e00269
822
Zhao Z, Andersen SU, Ljung K, Dolezal K, Miotk A, Schultheiss SJ, Lohmann JU
823
(2010) Hormonal control of the shoot stem-cell niche. Nature 465: 1089-1092
824
Zhou F, Lin Q, Zhu L, Ren Y, Zhou K, Shabek N, Wu F, Mao H, Dong W, Gan L, Ma
825
W, Gao H, Chen J, Yang C, Wang D, Tan J, Zhang X, Guo X, Wang J, Jiang L,
826
Liu X, Chen W, Chu J, Yan C, Ueno K, Ito S, Asami T, Cheng Z, Wang J, Lei C,
827
Zhai H, Wu C, Wang H, Zheng N, Wan J (2013) D14-SCF(D3)-dependent
828
degradation of D53 regulates strigolactone signalling. Nature 504: 406-410
829
Ziv D, Zviran T, Zezak O, Samach A, Irihimovitch V (2014) Expression profiling of
830
FLOWERING LOCUS T-like gene in alternate bearing 'Hass' avocado trees suggests
831
a role for PaFT in avocado flower induction. PLoS One 9: e110613
832
833
Figure 1. Characterization of the end of flowering transition. (A) An active inflorescence shoot
apex containing numerous flowers at different stages of development. (B) Q1 inflorescence
shoot apex with a white arrow pointing at the compact quiescent apex, which consists of
young unopened flower buds. The asterisks mark mature flowers with developing fruit
attached to elongated pedicels. (C) Image inflorescence shoot apex at the Q2 stage of
development, in which growth at the apex, including the fruits, ceased. (D) Average internode
length was determined for the last 30 internodes produced after the end of flowering
transition. Note: internode 30 is the last internode to elongate before Q1 stage arrest.
Figure 2. Expression patterns for key genes that control shoot growth. Gene expression
patterns were shown for (A, C, E and G) active and (B, D, F and H) Q1 apices. (A and B) CDKB1;1
is a mitotic regulator expressed in shoot meristem (Segers et al., 1996). (C and D) STM is a
shoot meristem identity gene (Long et al., 1996). (E and F) WUS is key regulator of stem cell
homeostasis (Laux et al., 1996). (G and H) MP regulates flower formation and vascular
development (Hardtke and Berleth, 1998). (A) The length of the bar is 50 µm. (H) Arrow points
at the MP expression in the sub-apical region of the meristem and pith cells in the Q1 apex.
Figure 3. 14C-IAA transport in stem segments during inflorescence development. Images of
(A) active and (B) Q1 inflorescence shoots. The white box marks the region of the stem where
developing fruits are attached. This region is referred as the zone of fruit development (ZFD).
IAA transport was determined in apical stem (AS) segments and basal stem (BS) segments
relative to the ZFD. (C) Radiolabeled IAA transport was measured during inflorescence
development starting at a time just before the first fruit set (BFS). After fruit set, 14C-IAA was
determined in apical stem (AS) and basal stem (BS) segments when 10, 20 and 35 fruits were
produced, as well as the Q1 stage. The light colour boxes represent control stem segments in
which 14C-IAA transport was measured in the presence of naphthyphthalamic acid (NPA), an
inhibitor of polar auxin transport. The letters above the bars determine whether differences
in 14C-IAA transport were statistically significant using analysis of variance, Tukey’s honest
significant difference.
Figure 4. A change in auxin response in apical inflorescence stems at end of flowering
transition. DR5:GUS staining patterns in (A) active, (C) Q1 and (E) Q2 shoot apices. Close up
of stems where developing fruits are attached in (B) active, (D) Q1 and (F) Q2 shoot apices.
Histological cross sections in the apical stem for an (G) active and (H) Q2 stage inflorescence.
Note: inset of vascular bundle displayed in upper right corner of each image. In actively
growing shoots, the section shown in (G) was through the region of the stem where the
developing fruits were attached. The section displayed in (H) was in a region of the stem
where the last fruits set.
Figure 5. Expression patterns for sugar signaling, metabolism and transport genes. Gene
expression patterns were determined in (A, C and E) active and (B, D and F) Q1 apices. (A and
B) TPS1, which is expressed in the vasculature of the inflorescence stem, was used as a marker
to assess the T6P pathway (Wahl et al., 2013; Ponnu et al., 2020). (C and D) CINV1 is an
invertase, which is required for growth (Barratt et al., 2009). (E and F) SUC2 is a sucrose
transporter expressed in the vascular tissues of inflorescences (Truernit and Sauer, 1995;
Gottwald et al., 2000). The length of the bar is 50 µm.
Figure 6. Sugar content in shoot apices and developing fruits. Sugar levels determined in
shoot apices before fruit set (BFS Apex), after 15 fruit set (15FS Apex) and at the Q1 stage (Q1
Apex). Sugar levels were also determined in developing fruits when 15 fruits set (15FS Fruit)
and at the Q1 stage (Q1 Fruit). The levels of (A) Glucose, (B) fructose and (C) sucrose were
measured from dry weight (DW) tissue. The letters above the bars determine whether
differences in sugar levels were statistically significant using analysis of variance, Tukey’s
honest significant difference.
Parsed Citations
Aguilar-Martinez JA, Poza-Carrion C, Cubas P (2007) Arabidopsis BRANCHED1 acts as an integrator of branching signals within axillary
buds. Plant Cell 19: 458-472
Google Scholar: Author Only Title Only Author and Title
Baena-Gonzalez E, Lunn JE (2020) SnRK1 and trehalose 6-phosphate - two ancient pathways converge to regulate plant metabolism
and growth. Curr Opin Plant Biol 55: 52-59
Google Scholar: Author Only Title Only Author and Title
Balanza V, Martinez-Fernandez I, Sato S, Yanofsky MF, Kaufmann K, Angenent GC, Bemer M, Ferrandiz C (2018) Genetic control of
meristem arrest and life span in Arabidopsis by a FRUITFULL-APETALA2 pathway. Nat Commun 9: 565
Google Scholar: Author Only Title Only Author and Title
Bangerth F (1989) Dominance among fruits/sinks and the search for a correlative signal. Physiologia Plantarum 76: 608-614
Google Scholar: Author Only Title Only Author and Title
Bangerth F, Li CJ, Gruber J (2000) Mutual interaction of auxin and cytokinin in regulating correlative dominance. Plant Growth
Regulation 32: 205-217
Google Scholar: Author Only Title Only Author and Title
Barbier F, Peron T, Lecerf M, Perez-Garcia MD, Barriere Q, Rolcik J, Boutet-Mercey S, Citerne S, Lemoine R, Porcheron B, Roman H,
Leduc N, Le Gourrierec J, Bertheloot J, Sakr S (2015) Sucrose is an early modulator of the key hormonal mechanisms controlling bud
outgrowth in Rosa hybrida. J Exp Bot 66: 2569-2582
Google Scholar: Author Only Title Only Author and Title
Barbier FF, Dun EA, Kerr SC, Chabikwa TG, Beveridge CA (2019) An Update on the Signals Controlling Shoot Branching. Trends Plant
Sci 24: 220-236
Google Scholar: Author Only Title Only Author and Title
Barbier FF, Lunn JE, Beveridge CA (2015) Ready, steady, go! A sugar hit starts the race to shoot branching. Curr Opin Plant Biol 25: 39-
45
Google Scholar: Author Only Title Only Author and Title
Barnes WJ, Anderson CT (2018) Cytosolic invertases contribute to cellulose biosynthesis and influence carbon partitioning in
seedlings of Arabidopsis thaliana. Plant J 94: 956-974
Google Scholar: Author Only Title Only Author and Title
Barratt DH, Derbyshire P, Findlay K, Pike M, Wellner N, Lunn J, Feil R, Simpson C, Maule AJ, Smith AM (2009) Normal growth of
Arabidopsis requires cytosolic invertase but not sucrose synthase. Proc Natl Acad Sci U S A 106: 13124-13129
Google Scholar: Author Only Title Only Author and Title
Bennett T, Hines G, van Rongen M, Waldie T, Sawchuk MG, Scarpella E, Ljung K, Leyser O (2016) Connective Auxin Transport in the
Shoot Facilitates Communication between Shoot Apices. PLoS Biol 14: e1002446
Google Scholar: Author Only Title Only Author and Title
Bihmidine S, Hunter CT, 3rd, Johns CE, Koch KE, Braun DM (2013) Regulation of assimilate import into sink organs: update on
molecular drivers of sink strength. Front Plant Sci 4: 177
Google Scholar: Author Only Title Only Author and Title
Bleecker AB, Patterson SE (1997) Last exit: senescence, abscission, and meristem arrest in Arabidopsis. Plant Cell 9: 1169-1179
Google Scholar: Author Only Title Only Author and Title
Bolouri Moghaddam MR, Van den Ende W (2013) Sugars, the clock and transition to flowering. Front Plant Sci 4: 22
Google Scholar: Author Only Title Only Author and Title
Braun N, de Saint Germain A, Pillot JP, Boutet-Mercey S, Dalmais M, Antoniadi I, Li X, Maia-Grondard A, Le Signor C, Bouteiller N, Luo
D, Bendahmane A, Turnbull C, Rameau C (2012) The pea TCP transcription factor PsBRC1 acts downstream of Strigolactones to
control shoot branching. Plant Physiol 158: 225-238
Google Scholar: Author Only Title Only Author and Title
Chuck G, Muszynski M, Kellogg E, Hake S, Schmidt RJ (2002) The control of spikelet meristem identity by the branched silkless1 gene
in maize. Science 298: 1238-1241
Google Scholar: Author Only Title Only Author and Title
Clark SE, Williams RW, Meyerowitz EM (1997) The CLAVATA1 gene encodes a putative receptor kinase that controls shoot and floral
meristem size in Arabidopsis. Cell 89: 575-585
Google Scholar: Author Only Title Only Author and Title
Crawford S, Shinohara N, Sieberer T, Williamson L, George G, Hepworth J, Muller D, Domagalska MA, Leyser O (2010) Strigolactones
enhance competition between shoot branches by dampening auxin transport. Development 137: 2905-2913
Google Scholar: Author Only Title Only Author and Title
Dong Z, Xiao Y, Govindarajulu R, Feil R, Siddoway ML, Nielsen T, Lunn JE, Hawkins J, Whipple C, Chuck G (2019) The regulatory
landscape of a core maize domestication module controlling bud dormancy and growth repression. Nat Commun 10: 3810
Google Scholar: Author Only Title Only Author and Title
Eveland AL, Jackson DP (2012) Sugars, signalling, and plant development. J Exp Bot 63: 3367-3377
Google Scholar: Author Only Title Only Author and Title
Fichtner F, Barbier FF, Feil R, Watanabe M, Annunziata MG, Chabikwa TG, Hofgen R, Stitt M, Beveridge CA, Lunn JE (2017) Trehalose
6-phosphate is involved in triggering axillary bud outgrowth in garden pea (Pisum sativum L.). Plant J 92: 611-623
Google Scholar: Author Only Title Only Author and Title
Galweiler L, Guan C, Muller A, Wisman E, Mendgen K, Yephremov A, Palme K (1998) Regulation of polar auxin transport by AtPIN1 in
Arabidopsis vascular tissue. Science 282: 2226-2230
Google Scholar: Author Only Title Only Author and Title
Gaudin V, Lunness PA, Fobert PR, Towers M, Riou-Khamlichi C, Murray JA, Coen E, Doonan JH (2000) The expression of D-cyclin
genes defines distinct developmental zones in snapdragon apical meristems and is locally regulated by the Cycloidea gene. Plant
Physiol 122: 1137-1148
Google Scholar: Author Only Title Only Author and Title
Gonzalez-Grandio E, Pajoro A, Franco-Zorrilla JM, Tarancon C, Immink RG, Cubas P (2017) Abscisic acid signaling is controlled by a
BRANCHED1/HD-ZIP I cascade in Arabidopsis axillary buds. Proc Natl Acad Sci U S A 114: E245-E254
Google Scholar: Author Only Title Only Author and Title
Gonzalez-Grandio E, Poza-Carrion C, Sorzano CO, Cubas P (2013) BRANCHED1 promotes axillary bud dormancy in response to shade
in Arabidopsis. Plant Cell 25: 834-850
Google Scholar: Author Only Title Only Author and Title
Gonzalez-Suarez P, Walker CH, Bennett T (2020) Bloom and bust: understanding the nature and regulation of the end of flowering.
Curr Opin Plant Biol 57: 24-30
Google Scholar: Author Only Title Only Author and Title
Gottwald JR, Krysan PJ, Young JC, Evert RF, Sussman MR (2000) Genetic evidence for the in planta role of phloem-specific plasma
membrane sucrose transporters. Proc Natl Acad Sci U S A 97: 13979-13984
Google Scholar: Author Only Title Only Author and Title
Guo W, Chen L, Herrera-Estrella L, Cao D, Tran LP (2020) Altering Plant Architecture to Improve Performance and Resistance. Trends
Plant Sci 25: 1154-1170
Google Scholar: Author Only Title Only Author and Title
Hardtke CS, Berleth T (1998) The Arabidopsis gene MONOPTEROS encodes a transcription factor mediating embryo axis formation
and vascular development. EMBO J 17: 1405-1411
Google Scholar: Author Only Title Only Author and Title
Jackson D (2001) In-situ hybridization in plants. In SJ Gurr, MJ McPherson, DJ Bowles, eds, Molecular Plant Pathology: A Practical
Approach, Vol 1. Oxford University Press, Oxford, pp 163-174
Google Scholar: Author Only Title Only Author and Title
Jia L, Zhang B, Mao C, Li J, Wu Y, Wu P, Wu Z (2008) OsCYT-INV1 for alkaline/neutral invertase is involved in root cell development and
reproductivity in rice (Oryza sativa L.). Planta 228: 51-59
Google Scholar: Author Only Title Only Author and Title
Jiang L, Liu X, Xiong G, Liu H, Chen F, Wang L, Meng X, Liu G, Yu H, Yuan Y, Yi W, Zhao L, Ma H, He Y, Wu Z, Melcher K, Qian Q, Xu HE,
Wang Y, Li J (2013) DWARF 53 acts as a repressor of strigolactone signalling in rice. Nature 504: 401-405
Google Scholar: Author Only Title Only Author and Title
Kim J, Dotson B, Rey C, Lindsey J, Bleecker AB, Binder BM, Patterson SE (2013) New clothes for the jasmonic acid receptor COI1:
delayed abscission, meristem arrest and apical dominance. PLoS One 8: e60505
Google Scholar: Author Only Title Only Author and Title
Krizek BA (1999) Ectopic expression of AINTEGUMENTA in Arabidopsis plants results in increased growth of floral organs. Dev Genet
25: 224-236
Google Scholar: Author Only Title Only Author and Title
Lastdrager J, Hanson J, Smeekens S (2014) Sugar signals and the control of plant growth and development. J Exp Bot 65: 799-807
Google Scholar: Author Only Title Only Author and Title
Laux T, Mayer KF, Berger J, Jurgens G (1996) The WUSCHEL gene is required for shoot and floral meristem integrity in Arabidopsis.
Development 122: 87-96
Google Scholar: Author Only Title Only Author and Title
Lauxmann MA, Annunziata MG, Brunoud G, Wahl V, Koczut A, Burgos A, Olas JJ, Maximova E, Abel C, Schlereth A, Soja AM, Blasing
OE, Lunn JE, Vernoux T, Stitt M (2016) Reproductive failure in Arabidopsis thaliana under transient carbohydrate limitation: flowers
and very young siliques are jettisoned and the meristem is maintained to allow successful resumption of reproductive growth. Plant
and very young siliques are jettisoned and the meristem is maintained to allow successful resumption of reproductive growth. Plant
Cell Environ 39: 745-767
Google Scholar: Author Only Title Only Author and Title
Le CS, Schmelz EA, Chourey PS (2010) Sugar levels regulate tryptophan-dependent auxin biosynthesis in developing maize kernels.
Plant Physiol 153: 306-318
Google Scholar: Author Only Title Only Author and Title
Leskow CC, Conte M, Del Pozo T, Bermudez L, Lira BS, Gramegna G, Baroli I, Burgos E, Zavallo D, Kamenetzky L, Asis R, Gonzalez M,
Fernie AR, Rossi M, Osorio S, Carrari F (2020) The cytosolic invertase NI6 affects vegetative growth, flowering, fruit set and yield in
tomato. J Exp Bot
Google Scholar: Author Only Title Only Author and Title
Lewis DR, Muday GK (2009) Measurement of auxin transport in Arabidopsis thaliana. Nat Protoc 4: 437-451
Google Scholar: Author Only Title Only Author and Title
Lilley JL, Gee CW, Sairanen I, Ljung K, Nemhauser JL (2012) An endogenous carbon-sensing pathway triggers increased auxin flux and
hypocotyl elongation. Plant Physiol 160: 2261-2270
Google Scholar: Author Only Title Only Author and Title
Long JA, Moan EI, Medford JI, Barton MK (1996) A member of the KNOTTED class of homeodomain proteins encoded by the STM gene
of Arabidopsis. Nature 379: 66-69
Google Scholar: Author Only Title Only Author and Title
Lou Y, Gou JY, Xue HW (2007) PIP5K9, an Arabidopsis phosphatidylinositol monophosphate kinase, interacts with a cytosolic invertase
to negatively regulate sugar-mediated root growth. Plant Cell 19: 163-181
Google Scholar: Author Only Title Only Author and Title
Martinez-Fernandez I, Menezes de Moura S, Alves-Ferreira M, Ferrandiz C, Balanza V (2020) Identification of Players Controlling
Meristem Arrest Downstream of the FRUITFULL-APETALA2 Pathway. Plant Physiol 184: 945-959
Google Scholar: Author Only Title Only Author and Title
Mason MG, Ross JJ, Babst BA, Wienclaw BN, Beveridge CA (2014) Sugar demand, not auxin, is the initial regulator of apical
dominance. Proc Natl Acad Sci U S A 111: 6092-6097
Google Scholar: Author Only Title Only Author and Title
Montgomery BL (2008) Right place, right time: Spatiotemporal light regulation of plant growth and development. Plant Signal Behav 3:
1053-1060
Google Scholar: Author Only Title Only Author and Title
Muller D, Leyser O (2011) Auxin, cytokinin and the control of shoot branching. Ann Bot 107: 1203-1212
Google Scholar: Author Only Title Only Author and Title
Nooden LD, Penney JP (2001) Correlative controls of senescence and plant death in Arabidopsis thaliana (Brassicaceae). J Exp Bot 52:
2151-2159
Google Scholar: Author Only Title Only Author and Title
Nunes C, O'Hara LE, Primavesi LF, Delatte TL, Schluepmann H, Somsen GW, Silva AB, Fevereiro PS, Wingler A, Paul MJ (2013) The
trehalose 6-phosphate/SnRK1 signaling pathway primes growth recovery following relief of sink limitation. Plant Physiol 162: 1720-1732
Google Scholar: Author Only Title Only Author and Title
Pfeiffer A, Janocha D, Dong Y, Medzihradszky A, Schone S, Daum G, Suzaki T, Forner J, Langenecker T, Rempel E, Schmid M, Wirtz M,
Hell R, Lohmann JU (2016) Integration of light and metabolic signals for stem cell activation at the shoot apical meristem. Elife 5
Google Scholar: Author Only Title Only Author and Title
Pfeiffer A, Wenzl C, Lohmann JU (2017) Beyond flexibility: controlling stem cells in an ever changing environment. Curr Opin Plant Biol
35: 117-123
Google Scholar: Author Only Title Only Author and Title
Poethig RS (2013) Vegetative phase change and shoot maturation in plants. Curr Top Dev Biol 105: 125-152
Google Scholar: Author Only Title Only Author and Title
Ponnu J, Schlereth A, Zacharaki V, Dzialo MA, Abel C, Feil R, Schmid M, Wahl V (2020) The trehalose 6-phosphate pathway impacts
vegetative phase change in Arabidopsis thaliana. Plant J 104: 768-780
Google Scholar: Author Only Title Only Author and Title
Przemeck GK, Mattsson J, Hardtke CS, Sung ZR, Berleth T (1996) Studies on the role of the Arabidopsis gene MONOPTEROS in
vascular development and plant cell axialization. Planta 200: 229-237
Google Scholar: Author Only Title Only Author and Title
Qi X, Wu Z, Li J, Mo X, Wu S, Chu J, Wu P (2007) AtCYT-INV1, a neutral invertase, is involved in osmotic stress-induced inhibition on
lateral root growth in Arabidopsis. Plant Mol Biol 64: 575-587
Google Scholar: Author Only Title Only Author and Title
Rameau C, Bertheloot J, Leduc N, Andrieu B, Foucher F, Sakr S (2014) Multiple pathways regulate shoot branching. Front Plant Sci 5:
741
Google Scholar: Author Only Title Only Author and Title
Ruan YL, Patrick JW, Bouzayen M, Osorio S, Fernie AR (2012) Molecular regulation of seed and fruit set. Trends Plant Sci 17: 656-665
Google Scholar: Author Only Title Only Author and Title
Sablowski R, Carnier Dornelas M (2014) Interplay between cell growth and cell cycle in plants. J Exp Bot 65: 2703-2714
Google Scholar: Author Only Title Only Author and Title
Sairanen I, Novak O, Pencik A, Ikeda Y, Jones B, Sandberg G, Ljung K (2012) Soluble carbohydrates regulate auxin biosynthesis via
PIF proteins in Arabidopsis. Plant Cell 24: 4907-4916
Google Scholar: Author Only Title Only Author and Title
Salazar-García S, Garner LC, Lovatt CJ (2013) Reproductive Biology. In B Schaffer, BN Wolstenholme, AW Whiley, eds, The Avocado,
Ed 2nd. CABI, Wallingford
Google Scholar: Author Only Title Only Author and Title
Oxfordshire OX10 8DE
UK, pp 118-167
Salazar-García S, Lord EM, Lovatt CJ (1998) Inflorescence and flower development of the 'Hass' avocado (Persea americana Mill.)
during 'on' an d'off' crop years. Journal of the American Society for Horticulture Science 123: 537-544
Google Scholar: Author Only Title Only Author and Title
Samach A, Smith HM (2013) Constraints to obtaining consistent annual yields in perennials. II: Environment and fruit load affect
induction of flowering. Plant Sci 207: 168-176
Google Scholar: Author Only Title Only Author and Title
Satoh-Nagasawa N, Nagasawa N, Malcomber S, Sakai H, Jackson D (2006) A trehalose metabolic enzyme controls inflorescence
architecture in maize. Nature 441: 227-230
Google Scholar: Author Only Title Only Author and Title
Sawicki M, Ait Barka E, Clement C, Vaillant-Gaveau N, Jacquard C (2015) Cross-talk between environmental stresses and plant
metabolism during reproductive organ abscission. J Exp Bot 66: 1707-1719
Google Scholar: Author Only Title Only Author and Title
Schneider A, Godin C, Boudon F, Demotes-Mainard S, Sakr S, Bertheloot J (2019) Light Regulation of Axillary Bud Outgrowth Along
Plant Axes: An Overview of the Roles of Sugars and Hormones. Front Plant Sci 10: 1296
Google Scholar: Author Only Title Only Author and Title
Schuetz M, Berleth T, Mattsson J (2008) Multiple MONOPTEROS-dependent pathways are involved in leaf initiation. Plant Physiol 148:
870-880
Google Scholar: Author Only Title Only Author and Title
Segers G, Gadisseur I, Bergounioux C, de Almeida Engler J, Jacqmard A, Van Montagu M, Inze D (1996) The Arabidopsis cyclin-
dependent kinase gene cdc2bAt is preferentially expressed during S and G2 phases of the cell cycle. Plant J 10: 601-612
Google Scholar: Author Only Title Only Author and Title
Serrano-Mislata A, Sablowski R (2018) The pillars of land plants: new insights into stem development. Curr Opin Plant Biol 45: 11-17
Google Scholar: Author Only Title Only Author and Title
Shen J, Zhang Y, Ge D, Wang Z, Song W, Gu R, Che G, Cheng Z, Liu R, Zhang X (2019) CsBRC1 inhibits axillary bud outgrowth by
directly repressing the auxin efflux carrier CsPIN3 in cucumber. Proc Natl Acad Sci U S A 116: 17105-17114
Google Scholar: Author Only Title Only Author and Title
Shinohara N, Taylor C, Leyser O (2013) Strigolactone can promote or inhibit shoot branching by triggering rapid depletion of the auxin
efflux protein PIN1 from the plasma membrane. PLoS Biol 11: e1001474
Google Scholar: Author Only Title Only Author and Title
Smith HM, Samach A (2013) Constraints to obtaining consistent annual yields in perennial tree crops. I: Heavy fruit load dominates over
vegetative growth. Plant Sci 207: 158-167
Google Scholar: Author Only Title Only Author and Title
Soundappan I, Bennett T, Morffy N, Liang Y, Stanga JP, Abbas A, Leyser O, Nelson DC (2015) SMAX1-LIKE/D53 Family Members Enable
Distinct MAX2-Dependent Responses to Strigolactones and Karrikins in Arabidopsis. Plant Cell 27: 3143-3159
Google Scholar: Author Only Title Only Author and Title
Springer PS, Holding DR, Groover A, Yordan C, Martienssen RA (2000) The essential Mcm7 protein PROLIFERA is localized to the
nucleus of dividing cells during the G(1) phase and is required maternally for early Arabidopsis development. Development 127: 1815-
1822
Google Scholar: Author Only Title Only Author and Title
Srikanth A, Schmid M (2011) Regulation of flowering time: all roads lead to Rome. Cell Mol Life Sci 68: 2013-2037
Google Scholar: Author Only Title Only Author and Title
Sundaresan V, Springer P, Volpe T, Haward S, Jones JD, Dean C, Ma H, Martienssen R (1995) Patterns of gene action in plant
development revealed by enhancer trap and gene trap transposable elements. Genes Dev 9: 1797-1810
Google Scholar: Author Only Title Only Author and Title
Teichmann T, Muhr M (2015) Shaping plant architecture. Front Plant Sci 6: 233
Google Scholar: Author Only Title Only Author and Title
Truernit E, Sauer N (1995) The promoter of the Arabidopsis thaliana SUC2 sucrose-H+ symporter gene directs expression of beta-
glucuronidase to the phloem: evidence for phloem loading and unloading by SUC2. Planta 196: 564-570
Google Scholar: Author Only Title Only Author and Title
Turck F, Fornara F, Coupland G (2008) Regulation and identity of florigen: FLOWERING LOCUS T moves center stage. Annu Rev Plant
Biol 59: 573-594
Google Scholar: Author Only Title Only Author and Title
Ulmasov T, Murfett J, Hagen G, Guilfoyle TJ (1997) Aux/IAA proteins repress expression of reporter genes containing natural and
highly active synthetic auxin response elements. Plant Cell 9: 1963-1971
Google Scholar: Author Only Title Only Author and Title
van Rongen M, Bennett T, Ticchiarelli F, Leyser O (2019) Connective auxin transport contributes to strigolactone-mediated shoot
branching control independent of the transcription factor BRC1. PLoS Genet 15: e1008023
Google Scholar: Author Only Title Only Author and Title
Wahl V, Ponnu J, Schlereth A, Arrivault S, Langenecker T, Franke A, Feil R, Lunn JE, Stitt M, Schmid M (2013) Regulation of flowering
by trehalose-6-phosphate signaling in Arabidopsis thaliana. Science 339: 704-707
Google Scholar: Author Only Title Only Author and Title
Walker CH, Bennett T (2018) Forbidden fruit: Dominance relationships and the control of shoot architecture. Annual Plant Reviews 1:
217-254
Google Scholar: Author Only Title Only Author and Title
Wang JW (2014) Regulation of flowering time by the miR156-mediated age pathway. J Exp Bot 65: 4723-4730
Google Scholar: Author Only Title Only Author and Title
Wang L, Wang B, Jiang L, Liu X, Li X, Lu Z, Meng X, Wang Y, Smith SM, Li J (2015) Strigolactone Signaling in Arabidopsis Regulates
Shoot Development by Targeting D53-Like SMXL Repressor Proteins for Ubiquitination and Degradation. Plant Cell 27: 3128-3142
Google Scholar: Author Only Title Only Author and Title
Wang L, Wang B, Yu H, Guo H, Lin T, Kou L, Wang A, Shao N, Ma H, Xiong G, Li X, Yang J, Chu J, Li J (2020) Transcriptional regulation
of strigolactone signalling in Arabidopsis. Nature 583: 277-281
Google Scholar: Author Only Title Only Author and Title
Wang M, Perez-Garcia MD, Daviere JM, Barbier F, Oge L, Gentilhomme J, Voisine L, Peron T, Launay-Avon A, Clement G, Baumberger
N, Balzergue S, Macherel D, Grappin P, Bertheloot J, Achard P, Hamama L, Sakr S (2021) Axillary bud outgrowth in rose is controlled by
sugar metabolic and signalling pathways. J Exp Bot
Google Scholar: Author Only Title Only Author and Title
Wang Y, Kumaishi K, Suzuki T, Ichihashi Y, Yamaguchi N, Shirakawa M, Ito T (2020) Morphological and Physiological Framework
Underlying Plant Longevity in Arabidopsis thaliana. Front Plant Sci 11: 600726
Google Scholar: Author Only Title Only Author and Title
Ware A, Walker CH, Simura J, Gonzalez-Suarez P, Ljung K, Bishopp A, Wilson ZA, Bennett T (2020) Auxin export from proximal fruits
drives arrest in temporally competent inflorescences. Nat Plants 6: 699-707
Google Scholar: Author Only Title Only Author and Title
Welham T, Pike J, Horst I, Flemetakis E, Katinakis P, Kaneko T, Sato S, Tabata S, Perry J, Parniske M, Wang TL (2009) A cytosolic
invertase is required for normal growth and cell development in the model legume, Lotus japonicus. J Exp Bot 60: 3353-3365
Google Scholar: Author Only Title Only Author and Title
Wu X, Dabi T, Weigel D (2005) Requirement of homeobox gene STIMPY/WOX9 for Arabidopsis meristem growth and maintenance. Curr
Biol 15: 436-440
Google Scholar: Author Only Title Only Author and Title
Wuest SE, Philipp MA, Guthorl D, Schmid B, Grossniklaus U (2016) Seed Production Affects Maternal Growth and Senescence in
Arabidopsis. Plant Physiol 171: 392-404
Google Scholar: Author Only Title Only Author and Title
Yadav RK, Girke T, Pasala S, Xie M, Reddy GV (2009) Gene expression map of the Arabidopsis shoot apical meristem stem cell niche.
Proc Natl Acad Sci U S A 106: 4941-4946
Google Scholar: Author Only Title Only Author and Title
Yang L, Xu M, Koo Y, He J, Poethig RS (2013) Sugar promotes vegetative phase change in Arabidopsis thaliana by repressing the
expression of MIR156A and MIR156C. Elife 2: e00260
Google Scholar: Author Only Title Only Author and Title
Yu S, Cao L, Zhou CM, Zhang TQ, Lian H, Sun Y, Wu J, Huang J, Wang G, Wang JW (2013) Sugar is an endogenous cue for juvenile-to-
adult phase transition in plants. Elife 2: e00269
Google Scholar: Author Only Title Only Author and Title
Zhao Z, Andersen SU, Ljung K, Dolezal K, Miotk A, Schultheiss SJ, Lohmann JU (2010) Hormonal control of the shoot stem-cell niche.
Nature 465: 1089-1092
Google Scholar: Author Only Title Only Author and Title
Zhou F, Lin Q, Zhu L, Ren Y, Zhou K, Shabek N, Wu F, Mao H, Dong W, Gan L, Ma W, Gao H, Chen J, Yang C, Wang D, Tan J, Zhang X,
Guo X, Wang J, Jiang L, Liu X, Chen W, Chu J, Yan C, Ueno K, Ito S, Asami T, Cheng Z, Wang J, Lei C, Zhai H, Wu C, Wang H, Zheng N,
Wan J (2013) D14-SCF(D3)-dependent degradation of D53 regulates strigolactone signalling. Nature 504: 406-410
Google Scholar: Author Only Title Only Author and Title
Ziv D, Zviran T, Zezak O, Samach A, Irihimovitch V (2014) Expression profiling of FLOWERING LOCUS T-like gene in alternate bearing
'Hass' avocado trees suggests a role for PaFT in avocado flower induction. PLoS One 9: e110613
Google Scholar: Author Only Title Only Author and Title
| 2021 | The role of auxin and sugar signaling in dominance inhibition of inflorescence growth by fruit load | 10.1101/2021.02.12.430977 | [
"Goetz Marc",
"Rabinovich Maia",
"Smith Harley M."
] | creative-commons |
Article
Substrate mediatedelastic coupling betweenmotile cells
modulates inter-cellinteractions and enhances cell-cell contact
Subhaya Bose1, Kinjal Dasbiswas1,†and Arvind Gopinath2,‡
Preprint
1
Departmentof Physics, University of California Merced,Merced,CA USA.
2
Departmentof Bioengineering,University of California Merced,Merced, CA USA.
*
Correspondence: ‡agopinath@ucmerced.edu, †kdasbiswas@ucmerced.edu
Abstract:
The mechanical micro-environmentof cells and tissuesinfluences key aspectsof cell structure and
function including cell motility. For proper tissuedevelopment, cells need to migrate,interact with
other neighbouring cells and form contacts,each of which require the cell to exertphysical forces.
Cells are known to exert contractile forces on underlying soft substrates. These stressesresult in
substratedeformation that can affectmigratory behavior of cells as well as provide an avenue for
cells tosenseeachother and coordinatetheir motion. The role of substratemechanics,particularly its
stiffness,in such biological processesis therefore a subjectof active investigation. Recent progress
in experimental techniques have enabled key insights into pairwise mechanical interactions that
control cell motility when they move on compliant softsubstrates.Analysis and modeling ofsuch
systemsis however still in its nascentstages.Motivated by the role modeling is expectedto play in
interpreting,informing andguiding experiments,we build a biophysical modelfor cell migrationand
cell-cellinteractions. Our focus is on situationshighly relevant to tissueengineering and regenerative
medicine -when substratetraction stressesinduced by motile cells enable substratedeformation and
serve as amedium of communication. Using a generalizable agent-basedmodel,we compute key
metrics ofcell motile behavior such asthe number of cell-cellcontactsover a given time, dispersion
of cell trajectories,and probability of permanentcell contact,and analyze how thesedepend on a cell
motility parameterand onsubstratestiffness.Our results provide a framework towards modeling the
manner in which cells may senseeachother mechanically via the substrateand usethis information
to generatecoordinated movements acrossmuch longer length scales. Our results also provide a
foundation to analyze experimentson thephenomenonknown as durotaxis where single cells move
preferentially towards regions of high stiffnesson patterned substrates.
2 of 16
1. Introduction
Many eukaryotic cells move by crawling, that is by adhering to and exerting mechan-
ical stresses and local forces on their extracellular matrix (ECM) that they then actively
deform (see for instance [1–4] and references therein). Existing approaches to modeling
collective cell motility focus on direct (steric and adhesive) cell-cell interactions or focus
at the single cell level on cell-substrate interactions [2] such as the details of focal adhe-
sions that are crucial to generating traction stresses in both adherent and motile cells [5].
Experiments strongly indicate however that cells cultured on soft, elastic, biocompatible
substrates can respond to each other even when not in direct contact [3,4]. Such interactions
can arise in cell culture experiments, with cells on the surface of synthetic hydrogels such as
polyacrylamide, which are linearly elastic, through mutual and active deformations of the
gel by the cells. These mechanically derived non-contact cell-cell interactions are even more
relevant and act over longer ranges in the biological extracellular matrix (ECM) comprising
collagen or fibrin, where cells can interact by remodeling and reorienting the fibers in the
ECM [6–8]. Even without such cell–matrix feedback, the presence of deformations has been
shown recently to guide the migration of other cells without requiring chemotactic cues [9].
Mechanical non-local interactions between cells offer advantages compared to chemi-
cal means. Specifically, mechanical signaling and mechanosensing of neighbouring cells
is typically faster and longer-ranged than chemical signaling. Chemical interactions are
limited by diffusion rates while mechanical interactions propagate near instantaneously
for purely elastic deformations [10]. Indeed, this crucially allows cells to not just sense
each other but also to synchronize their behaviour. For instance, substrate deformation-
mediated long-range interactions has been clearly demonstrated in heart muscle cells that
synchronize their beating without direct contact [11,12], as well as at a subcellular level
between myofibrils within a single heart muscle cell [13]. Cell communications via sensing
of substrate or matrix deformation are particularly important in sparse, non-confluent cell
cultures or tissue that occur in a number of biologically relevant situations. Apart from
beating cardiomyocytes, examples of such situations include wound healing involving
fibroblasts [14], sprouting blood vessels comprising endothelial cells [15], and migration
of mesenchymal cells in zebrafish embryo before the formation of confluent epithelial
tissue [16]. In all these cases, cells are not in direct contact but exert traction forces on the
surrounding mechanical medium and concomitantly sense deformations caused by nearby
cells. Such interactions therefore crucially depend on the stiffness of the substrate, and
can be probed by experiments that vary the stiffness of the hydrogel substrate on which
the cells are cultured [17,18]. These aspects influence not only motility response at the
single cell level but also strongly impact collective behavior including directed motility
and subsequent spatial self-organization.
On the other hand, while substrate-mediated cell-cell elastic interactions have been
considered for the organization of adherent cells in a variety of mechanobiological contexts
[19,20] (the physical basis of such modeling is reviewed in Ref. [21]), their effect on collective
cell motility, which in principle is always present, have not been carefully modeled. Here,
we present a simple biophysical agent–based model and computational results that focus
on how substrate mediate mechanical communication allows two cells to sense each other
and impacts their collective and relative motility. Our approach provides a foundation
for the study of more general cell interactions that include both mechanical and chemical
signalling, and also serves as a starting point for future studies of mechanical substrate
based interactions in multi-cellular systems such as growing tissue and confluent sheets.
2. Experimental observations motivate model for cell elastic interactions
Many eukaryotic cells use contractile localized forces generated by their actomyosin
cytoskeleton to adhere to and move on their substrates. Such traction forces typically cause
measurable deformations in the underlying substrates in cell culture experiment [5], and
have a spatially dipolar pattern [22]. A cell typically acts as a force dipole exerting – a pair
of equal and opposite forces – on the elastic medium. The dipolar pattern arises due to the
3 of 16
initial
final
M
F
contours
X
Z
Y
A
B
Cell A
Cell B
Elastic substrate
<latexit sha1_base64="YMvcfB29NQEDn9ZkuTDIR8ovUw4=">AB6XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqseiF49V7Ae0oWy2k3bpZhN
2N0IJ/QdePCji1X/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZLGLVCahGwSU2DTcCO4lCGgUC28H4dua3n1BpHstHM0nQj+hQ8pAzaqz0QFS/XHGr7hxklXg5qUCORr/81RvELI1QGiao1l3PTYyfUWU4Ezgt9VKNCWVjOsSupZJGqP1sfumUnFlQMJY2ZK
GzNXfExmNtJ5Ege2MqBnpZW8m/ud1UxNe+xmXSWpQsWiMBXExGT2NhlwhcyIiSWUKW5vJWxEFWXGhlOyIXjL6+S1kXVq1Xd+8tK/SaPowgncArn4MEV1OEOGtAEBiE8wyu8OWPnxXl3PhatBSefOY/cD5/ADRajSU=</latexit>r
Radial
distance
Cell B
Cell A
Figure 1. Schematic of the cell-cell mechanical interactions model: (A) Two cells A and B cultured
on the surface of thick elastic substrate can sense each other and interact at long range (when the
inter-cell distance r is longer than typical cell sizes, here depicted by dashed red circles) through
mechanical deformations of the underlying substrate. These deformations arise as the cells exert
traction stresses on the underlying elastic material. Here the cells are restricted to move on the surface
of the substrate. (B) We study with our computational model how a motile cell (M, Cell A, pink)
moves in the presence of a fixed central cell (Cell B, yellow). This two cell system on a substrate
(schematic shown as a top view) also mimics scenarios where a motile cell may encounter an elastic
impurity or obstacle on the medium. Shown as blue circles are contours of constant elastic potential
(in simplified form) that determine the inter-cell elastic force experienced by the motile cell B as a
result of the elastic deformations of the medium by both cells A and B. Also shown (in black) is a
representative simulated trajectory of the motile cell which starts outside the area of influence of the
stationary cell.
fact that no external forces are present on the system, and the cell, as a whole, moves on
its own accord. The net effect of these stresses is to contract or pull in the elastic material
comprising the substrate towards the cell.
In Ref. [3], it was shown that endothelial cells cultured on hydrogel substrates of
varying stiffness change their motile behavior in the presence of traction stresses exerted
on the substrate by neighbouring but non-contacting cells. In particular, it was shown that
pairs of cells on softer gels, showed reduced collective migration in comparison to isolated
cells. The number of contacts two cells made over specific periods of time by extending
their pseudopodia towards each other was also measured and found to depend sensitively
on substrate stiffness. Remarkably, the cells made stable contacts on very soft gels (Elastic
modulus, 500 Pa), whereas they made repeated contacts and withdrawals on substrates of
intermediate compliance (Elastic modulus, 2500 � 5500 Pa).
Motivated by this experiment, we here model the motility characteristics of a two
cell system and demonstrate how elastic deformations induced in the substrate allow
cells to respond to each other. We consider a pair of cells that each adhere to, and exert
stresses on the underlying substrate thereby deforming it as shown in Fig. 1. As mentioned
above, adherent and motile cells generate a contractile stress on the substrate. Here, the
contractility P of each cell, is minimally described by a physical model of force dipoles –a
pair of equal and opposite forces exerted on the substrate, and is thus a tensorial quantity
[19]. Such modeling is inspired by the theory of deformations induced by inclusions in
materials [23]. Unlike passive material inclusions, cells can actively regulate their force
production in response to external mechano-chemical cues from the substrate, including
the presence of other cells. Such complicating feedback effects in cell–cell interactions has
also been theoretically considered [24,25], but we ignore these for simplicity here, and we
treat P as an intrinsic cell property that is independent of underlying substrate matrix
strain and stiffness.
4 of 16
To simplify our study, we assume that one of the cells is motile (Cell A) and the other
is stationary (Cell B). The stationary cell B is nonetheless alive in that it still deforms the
substrate. The resulting deformation field, or equivalently the substrate mediated elastic
potential, is sensed by the other, distant, motile cell A. The interaction potential between
the cells in turn creates a mechanical force on the motile cell A. For polarized and elongated
cells, the deformations have a dipolar spatial pattern (described in Appendix A). However,
here we consider a simplified scenario that is valid when cells reorient very fast in the time
for them to translate and migrate (Appendix A, §3). This implies that the directions of
the dipole axis (of both cells A and B) fluctuates rapidly as cell A moves resulting in an
effectively isotropic, attractive interaction potential that decays with distance as ⇠ 1/r3
(iso–surfaces shown as blue circles in Fig. 1 B). Analysis of this model interaction provides
us insight into attractive potentials strongly influence cell motility.
The motile cell is considered to move diffusively with an effective diffusion coeffi-
cient, while also being acted upon by an elastic interaction force from the stationary cell.
Although, polarized cells may propel themselves persistently along their body axis, we
consider more isotropic cells here which extend their pseudopodia in different directions
randomly, and are thus described adequately by a diffusive process. Such a simple effective
Langevin equation is commonly used to describe elastically coupled motile active particles
[26] and swarming bacteria [27] but has not been studied previously in conjunction with
this specific type of interactions that arise on an elastic substrate.
We note that the model can be easily generalized (as derived in Appendix A) to
describe a pair of motile cells since the interactions are pairwise and reciprocal. The
interaction potential is not isotropic and depends on both the inter-cell distance as well
as on the instantaneous alignment of the cells’ dipole axes. Thus the force on each cell
(related to the gradient of the potential) depends on not just the relative positions of the
cells.but additionally on the direction of the contractile dipoles exerted by cells A and
B. Truly spherical dipoles embedded in an elastic medium do not interact mechanically
[23], unless cell-substrate feedback effects occur [25]. Furthermore, cell-cell interactions
in a fibrous, nonlinear elastic medium can be longer ranged [28] and have a power law
character, ⇠ 1/ra, where a < 3 [29]. The interaction of disk-like cells on top of a thick
substrate (semi-infinite geometry) is also more complicated [30]. We choose the isotropic,
attractive 1/r3 potential as the simplest attractive interaction with the same distance
dependence as the dipolar interaction, with the objective of testing how such a potential
can affect cell motility. Motivating future work, we show how the conclusions from the
simpler potential remain qualitatively valid even as specifics of cell trajectories change
when the more general dipolar potential is used. This model highlighted in this work,
although very simplified both in its description of cell contractility and motility, can thus
capture key aspects of motility and contact formation, as we now describe.
3. Materials and Methods
3.1. Model for two-cell interactions
The model used to analyze the two-cell system is an agent-based stochastic model.
We start with the stochastic Langevin equation for the dynamics of the moving cell A in
the presence of a second cell B fixed at the origin as illustrated in Figure 1(A). Details of
the model and the simplifications involved may be found in Appendix A. Starting from
the more general model where both cells A and B can move, we now fix cell B and thus
set rB = 0. In other words, we choose the center of cell B to be the origin from which the
position of cell A and its distance relative to B is measured. Writing r = rB � rA, we write
the equation for r(t) where t is the time,
dr
dt = �µT
∂W
∂r +
p
2Deff hhh(t)
(1)
5 of 16
where Deff is the effective translational diffusivity quantifying the random motion of the
moving cell in the absence of the fixed cell, and hhh is a random white noise term whose
components satisfy
hhi(t)hj(t0)i = d(t � t0)dij.
Note that h - the active noise term - has units of t�1/2. The mobility µT in equation (1)
quantifies the effective friction from the medium and is inversely proportional to the cell
size s and inversely proportional the the viscosity at the surface. Here it is assumed that
the cells moving on a wet surface and that the fluid nature of the surface provides a viscous
resistance opposing cell motion.
The two-cell potential W derives from the elastic interactions communicated via the
linear deformation of the substrate (Appendix A, Equation A5) and is given by,
W
=
1
2k(s � r)2, when 0 r s, and
(2)
=
� P2
E
f(n)
r3 ,
when r > s.
(3)
Numerical solutions to equation (1) are obtained with varying initial conditions for cell
A as explained subsequently. To ease the computational analysis, we work in scaled
dimensionless units. We choose cell size (diameter) s (see Fig 1), diffusion time s2/D0,
and thermal energy kBT – with T corresponding to the temperature of the cell/substrate
system – as our length, time and energy scales respectively. Equations (1-3) may then be
rewritten as
dr⇤
dt⇤ = �dW⇤
dr⇤ +
p
2DT hhh⇤
T,
(4)
where the potential in scaled form is
W⇤
=
1
2ksteric(1 � r⇤)2, when 0 r⇤ 1, and
(5)
=
� a
r⇤3 ,
when r⇤ > 1.
(6)
Superscripts ⇤ in equations (4)-(6) denote non-dimensional quantities. Henceforth, we will
drop this subscript for clarity. Thus the dynamics may be followed as a function of three
dimensionless numbers (parameters)
a ⌘
✓ P2f(n)
EkBTs3
◆
,
DT ⌘
✓ Deff
D0
◆
,
and ksteric ⌘
✓ ks2
kBT
◆
.
(7)
3.2. Dimensionless parameters quantifying cell motion and interactions
The parameters that emerge in equations (1)-(7) and typical of the two-cell scenario
studied here are summarized in Table 1. Following Ref. [3], we are interested in substrates
that are linearly elastic with the Young’s modulus E ranging from 0.5 kPa to 33 kPa, well
within the range of 0.1-100 kPa appropriate for tissues and bio-compatible materials [18].
The effective diffusion coefficients exhibited by cells in experiments [3] include the random
noisy motion as the cells explore territory and a contribution due to short-time deterministic
motion. We explore values in the range 3µm2/minute to 50 µm2/minute. Time scales
are estimated from experiments as well and 250 seconds in real time correspond to a
dimensionless time duration of unity.
Scaled non-dimensional parameters relevant to the simulation may be calculated
from dimensional quantities as explained earlier. Three scaled parameters determine the
dynamics of the two-cell system: DT, a and ksteric. Values used in the computations are
listed in Table 2. The self avoidance parameter ksteric is chosen such that the cells don’t
overlap and is computed based on the time step used in the simulations. This allows us to
control the stability of the simulation and its accuracy.
6 of 16
3.3. Numerical solution and tracking cell trajectories
Equations (4)-(7) are solved for the dynamics of the moving cell with appropriate
boundary and initial conditions. The Langevin equation (4) is an example of stochastic
differential equations; here we solve this equation using the explicit half-order Euler-
Maruyama method one of us has used recently in similar problems involving bacteria cells
moving in light fields [27] and in simulations of active Brownian particles [26].
Table 1. Biophysical parameters characterizing the two-cell (typical values from [3,31,32]).
Quantity
Interpretation
Experimental values
s
Cell size
10-100 µm
T
Temperature
250 C
D0
Thermal Diffusivity
25 µm2/min
Deff
Effective Diffusivity
3 � 50 µm2/min
E
Young’s modulus
0.5 � 33 kPa
n
Poisson ratio
0.3 - 0.5
P
Contractility
10�14 Nm
Table 2. Simulation parameters and their meaning.
Parameter
Interpretation
Definition
Simulation values
DT
Diffusivity
Deff/D0
0.1-10
a
Cell-cell interaction
P2f(n)/(EkBTs3)
0.1-100
ksteric
Self-avoidance
ks2/kBT
103 � 104
Given the position of cell A at time t, r(t), its subsequent location at time t + dt, r(t + dt),
follows,
r(t + dt) = r(t) �
✓∂W
∂r |r(t)
◆
dt +
p
2DTdt w,
(8)
where w is a random two-dimensional vector with components each drawn at every time
step from a normal distribution with mean zero and standard deviation of unity.
We simulated several trajectories of cell A ((n = 1000) trajectories, diameter s = 1 in
scaled units), under the influence of the central stationary cell B (also having diameter
s = 1). The simulations were conducted in two different geometries as described below.
To study the contact frequency between two-cells and explore the systematically
explore the role of the elastic potential, we simulated cell A moving in a confined square
box of size 12s with the stationary cell B at the center of the box. Cells reflect from the box
surface when they encounter it and thus are restricted to remain within the simulation
domain.
In order to calculate the number of contact in due course of the simulation, we define
a contact radius 1.5s from the centre of the stationary cell, and we consider a contact
if the centre of the test cell lies within the contact radius. The cell can come out of the
contact radius and re-enter, increasing the number of contacts. The time step used in these
simulations is dt = 0.0001 and total number of steps in this simulation is 107, i.e. a cell
trajectories were followed for a total time of T = 1000.
On the other hand for calculating cell dispersivities, and specifically the mean squared
displacement (MSD) of cell A, we used periodic boundary conditions and a periodic
potential. This corresponds to cell A moving in a periodic domain and interacting with
a regular square lattice of multiple stationary cells (images of B) separated uniformly by
7 of 16
a distance 12s. The time step used to integrate equation (7) in these simulations is also
dt = 0.0001 and total number of steps in this simulation is 107, i.e. a cell trajectories were
followed for a total time of T = 1000. The mean square displacement MSD was calculated
by tracking trajectories of cell A (the same as tracking n = 100 cells). As before, cell A is
initialized randomly inside the same square box of length of 12s, but outside the contact
radius. Cells that move out of the domain are reintroduced into the domain in a manner
that respects periodic boundary conditions and the appropriate symmetries.
In this case since r ⌘ xex + yey is the relative distance between the cells, the mean
square displacement is calculated by the equation,
MSD(t) = 1
n
n
Â
a=1
h[xa(tR + t) � xa(tR)]2 + [ya(tR + t) � ya(tR)]2i
(9)
where t is the delay time, and the summation is over each cell trajectory (indexed by a)
and extends over the full number of trajectories n = 100. The delay time is varied and the
averages are obtained by choosing different values of the reference time tR as is normally
done. The MSD given by equation (9) is thus an average over time and also an average
over realized cell trajectories.
-5
0
5
-6
-4
-2
0
2
4
6
Number of contacts
0
1000
2000
3000
4000
5000
6000
7000
8000
0.1
1
5
10
20
100
0.1
1
5
10
20
100
8000
7000
6000
5000
4000
3000
2000
1000
0
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRV
CyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGN
izCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
<latexit sha1_base64="7brGteJTxOG/ubQVZlufQYTATc=">AB8HicbVBNSwMxEJ
2tX7V+VT16CRbBU9kVUS9CUQ8eK/RL2qVk02wbmSXJCuUZX+Fw+KePXnePfmLZ70NYHA4/3ZpiZF8ScaeO6305hZXVtfaO4Wdra3tndK+8ftHSUKEKbJOKR6gRYU84kbRpmO3E
imIRcNoOxrdTv/1ElWaRbJhJTH2Bh5KFjGBjpce7ftrI0DXy+uWKW3VnQMvEy0kFctT75a/eICKJoNIQjrXuem5s/BQrwinWamXaBpjMsZD2rVUYkG1n84OztCJVQYojJQtadBM/T2
RYqH1RAS2U2Az0oveVPzP6yYmvPJTJuPEUEnmi8KEIxOh6fdowBQlhk8swUQxeysiI6wMTajkg3BW3x5mbTOqt5F1X04r9Ru8jiKcATHcAoeXEIN7qEOTSAg4Ble4c1Rzovz7nzMW
wtOPnMIf+B8/gBbw492</latexit>DT = 1
<latexit sha1_base64="Cue/mGefYH6DOZApKvlDIARi+RY=">AB8HicbVBNSwMxEJ
2tX7V+VT16CRbBU9ktUr0IRT14rNAvaZeSTbNtaJdkqxQlv4KLx4U8erP8ea/MW3oK0PBh7vzTAzL4g508Z1v53c2vrG5lZ+u7Czu7d/UDw8aukoUYQ2ScQj1QmwpxJ2jTMcNqJ
FcUi4LQdjG9nfvuJKs0i2TCTmPoCDyULGcHGSo93/bQxRdeo0i+W3LI7B1olXkZKkKHeL371BhFJBJWGcKx13Nj46dYGUY4nRZ6iaYxJmM8pF1LJRZU+n84Ck6s8oAhZGyJQ2aq78
nUiy0nojAdgpsRnrZm4n/ed3EhFd+ymScGCrJYlGYcGQiNPseDZixPCJZgoZm9FZIQVJsZmVLAheMsvr5JWpexVy+7DRal2k8WRhxM4hXPw4BJqcA91aAIBAc/wCm+Ocl6cd+dj0
Zpzsplj+APn8wdR493</latexit>DT = 2
<latexit sha1_base64="vTFNT+F7LjdOFlVcVpPFGjA+tTk=">AB8HicbVBNSwMxEJ
2tX7V+VT16CRbBU9kVq16Eoh48VuiXtEvJptk2NMkuSVYoS3+Fw+KePXnePfmLZ70NYHA4/3ZpiZF8ScaeO6305uZXVtfSO/Wdja3tndK+4fNHWUKEIbJOKRagdYU84kbRhmOG3H
imIRcNoKRrdTv/VElWaRrJtxTH2B5KFjGBjpce7XlqfoGtU6RVLbtmdAS0TLyMlyFDrFb+6/YgkgkpDONa647mx8VOsDCOcTgrdRNMYkxEe0I6lEguq/XR28ASdWKWPwkjZkgbN1N8
TKRZaj0VgOwU2Q73oTcX/vE5iwis/ZTJODJVkvihMODIRmn6P+kxRYvjYEkwUs7ciMsQKE2MzKtgQvMWXl0nzrOxdlN2H81L1JosjD0dwDKfgwSVU4R5q0ACAp7hFd4c5bw4787Hv
DXnZDOH8AfO5w9h0496</latexit>DT = 5
-5
0
5
-6
-4
-2
0
2
4
6
<latexit sha1_base64="sX6ceXyDkDrHKl56EUzFwVjyo0=">ACAHicdVDLSsNAFJ3UV62vqgsXbgaL4EJCEmtbBaGoC5cV+oImhMl02g6dPJiZCVk46+4caGIWz/DnX/jpK2gogcunDnXube40WMC
mkYH1puYXFpeSW/Wlhb39jcKm7vtEUYc0xaOGQh73pIEYD0pJUMtKNOEG+x0jHG19lfueOcEHDoCknEXF8NAzogGIkleQW92zEohGCF/D0GNrn127STNXDdIslQz+rVaxyBRq6YVRNy8yIVS2flKGplAwlMEfDLb7b/RDHPgkZkiInmlE0kQlxQzkhbsWJAI4TEakp6iAfKJcJLpASk8VEofDkKuKpBwqn6fSJAvxMT3VKeP5Ej89jLxL68Xy0HNSWgQxZIEePbRIGZQhjBLA/YpJ1iyiSIc6p2hXiEOMJSZVZQIXxdCv8
nbUs3K7pxWy7VL+dx5ME+OABHwARVUAc3oAFaAIMUPIAn8Kzda4/ai/Y6a81p85ld8APa2yd9YJRr</latexit>� = 5, DT = 1
-5
0
5
-6
-4
-2
0
2
4
6
=100 DT=5
1
1
4
4
<latexit sha1_base64="FXK5L+m+Qc3TiguW5FtJdbaOGCo=">ACAnicbVDLSgMxFM34rPU16krcBIvgQkpGfCEIRV24rNAXdIYhk2ba0ExmSDJCGYobf8WNC0Xc+hXu/BvTdhbaeuDCyTn3kntPkHCmN
ELf1tz8wuLScmGluLq2vrFpb203VJxKQusk5rFsBVhRzgSta6Y5bSWS4ijgtBn0b0Z+84FKxWJR04OEehHuChYygrWRfHvXxTzpYXgFHYSOoHt562e1oXme+nYJldEYcJY4OSmBHFXf/nI7MUkjKjThWKm2gxLtZVhqRjgdFt1U0QSTPu7StqECR1R52fiEITwSgeGsTQlNByrvycyHCk1iALTGWHdU9PeSPzPa6c6vPAyJpJU0EmH4UphzqGozxgh0lKNB8YgolkZldIelhiok1qROCM3yLGkcl52zMro/KVWu8zgKYA/
sg0PgHNQAXegCuqAgEfwDF7Bm/VkvVjv1sekdc7KZ3bAH1ifPwQ8lJ0=</latexit>� = 100, DT = 5
3
3
2
2
<latexit sha1_base64="5JYanCdHzdygT4Y909yu7tGkuIg=">ACAXicdVDLSgMxFM3UV62vUTeCm2ARXEjJ1KGtglDUhcsKfUE7DJk0bUMzD5KMUIa68VfcuFDErX/hzr8x01ZQ0QMXTs65l9x7vIgzq
RD6MDILi0vLK9nV3Nr6xuaWub3TlGEsCG2QkIei7WFJOQtoQzHFaTsSFPsepy1vdJn6rVsqJAuDuhpH1PHxIGB9RrDSkmvudTGPhieQwsdw+7ZlZvUJ+nLNfOocFopFe0SRAWEylbRSkmxbJ/Y0NJKijyYo+a791eSGKfBopwLGXHQpFyEiwUI5xOct1Y0giTER7QjqYB9ql0kukFE3iolR7sh0JXoOBU/T6RYF/Kse/pTh+rofztpeJfXidW/YqTsCKFQ3I7KN+zKEKYRoH7DFBieJjTARTO8KyRALTJQOLadD+LoU/k+
axYJVKqAbO1+9mMeRBfvgABwBC5RBFVyDGmgAu7A3gCz8a98Wi8GK+z1owxn9kFP2C8fQLqi5Sh</latexit>� = 10, DT = 1
<latexit sha1_base64="VqCrxFTWefFUQRo7dsSW1orLZI=">ACAXicdVDLSgMxFM3UV62vqhvBTbAILmTIjENbBaGoC5cV+oK2lEyaUMzD5KMUIa68VfcuFDErX/hzr8x01ZQ0QMXTs65l9x73Igzq
RD6MDILi0vLK9nV3Nr6xuZWfnunIcNYEFonIQ9Fy8WSchbQumK01YkKPZdTpvu6DL1m7dUSBYGNTWOaNfHg4B5jGClpV5+r4N5NMTwHNroGHbOrnpJbaJfVi9fQOZpuWg7RYhMhEqWbaXELjknDrS0kqIA5qj28u+dfkhinwaKcCxl20KR6iZYKEY4neQ6saQRJiM8oG1NA+xT2U2mF0zgoVb60AuFrkDBqfp9IsG+lGPf1Z0+VkP520vFv7x2rLxyN2FBFCsakNlHXsyhCmEaB+wzQYniY0wEUzvCskQC0yUDi2nQ/i6FP5
PGrZpFU104xQqF/M4smAfHIAjYIESqIBrUAV1QMAdeABP4Nm4Nx6NF+N1pox5jO74AeMt0/sHZSi</latexit>� = 20, DT = 1
A
Figure 2. The number of cell–cell contact events measured in a fixed interval of time depends
strongly on the elastic interaction parameter. A contact event is identified as cell A coming within
a prescribed contact radius of cell B with cell A initialized randomly in a certain area around cell
B. Thus the number of contact is be interpreted as the average number of contacts of the two cells.
The number of simulation runs conducted were 50 for each combination of DT and a. The dashed
curves are guides to the eye illustrating the trends seen with increasing values of a. Diffusion is the
major factor in governing the number of contacts for low values of a. For higher a, the attractive
potential increases the probability of the cell to stay near the contact radius and controls the number
of contacts. Trajectories for highlighted data points (1)-(4) are shown on the right. The box plots
show the distribution of contact numbers. The lower and upper bounds of the box are the first and
the third quartiles respectively, while the line in middle is the median. The lower and upper limits
of the dashed lines are the minimum and maximum number of contacts observed for cells for each
combination of a and DT. The simulation was run for a total time of T = 1000 and updates in cell
position were made every dt = 0.001.
The mobility of cell A reflects the properties of the microenvironment created by cell
B and by the substrate. The mean square displacement in (9) is written as a function of the
delay time t that may be interpreted as an effective observation time over which the cell
motion is observed. For instance, a cell that moves with constant speed for small times
(say ⇠ T1) and undergoes a diffusive random walk when observed over long times (say ⇠
8 of 16
T2) will exhibit different slopes for t < T1 and for t > T2. The exponent characterizing the
dependence of the MSD on the delay time provides information as to whether the motion is
sub-diffusive (exponent < 1), diffusive (exponent = 1), or super-diffusive (exponent > 1).
It is constructive to study the expected MSD for cell A in the absence of cell B. In
this particular case, since A is purely diffusive, the MSD has the simple form valid for
diffusion in two dimensions MSD(t) = 4DTt. Deviations from this expression arise due to
the mechanically induced inter-cell interaction and thus quantify the extent to which cell B
perturbs the dispersion of cell A. For instance transient or persistent trapping of cell A will
result in the MSD scaling sub-linearly with t.
0
1000
2000
3000
4000
5000
6000
7000
8000
0.5
0.2
0.1
1
2
5
10
Number of contacts
0.1
0.2
0.5
1.0
2.0
10.0
8000
7000
6000
5000
4000
3000
2000
1000
0
Effective translational diffusivity
5.0
1
2
3
4
-5
0
5
-6
-4
-2
0
2
4
6
=10 DT=0.1
2
<latexit sha1_base64="a1ohGl7mT+YWUTrN1SID4wGcAfA=">ACBHicdVDLSgMxFM3UV62vUXe6CRbBhZSkFNsKQlEXLiu0tAZSiZN29DMgyQjlKHgxl9x40JB3PoR7vwbM20FT1w4eSce8m9x4sEV
xqhDyuzsLi0vJdza2tb2xu2ds7NyqMJWVNGopQtj2imOABa2quBWtHkhHfE6zljS5Sv3XLpOJh0NDjiLk+GQS8zynRuraew4R0ZDAM4jRMXROL7tJY2JeqIC7dh4VEIY5gSXD5BhlSrlSKuQJxaBnkwR71rvzu9kMY+CzQVRKkORpF2EyI1p4JNck6sWEToiAxYx9CA+Ey5yfSGCTw0Sg/2Q2kq0HCqfp9IiK/U2PdMp0/0UP32UvEvrxPrfsVNeBDFmgV09lE/FlCHMA0E9rhkVIuxIYRKbnaFdEgkodrEljMhfF0K/ye
tYgGXChfl/K183keWbAPDsARwKAMauAK1ETUHAHsATeLburUfrxXqdtWas+cwu+AHr7RNZlJU7</latexit>� = 10, DT = 0.1
-5
0
5
-6
-4
-2
0
2
4
6
=1 DT=0.1
1
3
<latexit sha1_base64="RAZ8AY3g78fqOrdBOgY30DI3NCc=">ACAXicbVDLSsNAFL3xWesr6kZwM1gEF1KSIiqCUNSFywp9QRPCZDph04ezEyEurGX3HjQhG3/oU7/8ZJ24W2Hrhw5px7mXuPn
3AmlWV9GwuLS8srq4W14vrG5ta2ubPblHEqCG2QmMei7WNJOYtoQzHFaTsRFIc+py1/cJP7rQcqJIujuhom1A1xL2IBI1hpyTP3HcyTPkZXqGKdIOfy1svqo/zlmSWrbI2B5ok9JSWYouaZX043JmlI0U4lrJjW4lyMywUI5yOik4qaYLJAPdoR9MIh1S62fiCETrShcFsdAVKTRWf09kOJRyGPq6M8SqL2e9XPzP6QquHAzFiWpohGZfBSkHKkY5XGgLhOUKD7UBPB9K6I9LHAROnQijoEe/bkedKslO2zsn
V/WqpeT+MowAEcwjHYcA5VuIMaNIDAIzDK7wZT8aL8W58TFoXjOnMHvyB8fkDjZKUYQ=</latexit>� = 20, DT = 2
-5
0
5
-6
-4
-2
0
2
4
6
=20 DT=10
4
<latexit sha1_base64="8NixRt3VOpY3OoM9bE3TxbwgoX8=">ACAnicbVDLSgMxFM34rPU16krcBIvgQkqmiIogFHXhskJf0BmGTJpQzOZIckIZShu/BU3LhRx61e4829M21lo64ELJ
+fcS+49QcKZ0gh9WwuLS8srq4W14vrG5ta2vbPbVHEqCW2QmMeyHWBFORO0oZnmtJ1IiqOA01YwuBn7rQcqFYtFXQ8T6kW4J1jICNZG8u19F/Okj+EVrKAT6F7e+l9ZF4O8u0SKqMJ4DxclICOWq+/eV2Y5JGVGjCsVIdByXay7DUjHA6KrqpogkmA9yjHUMFjqjyskJI3hklC4MY2lKaDhRf09kOFJqGAWmM8K6r2a9sfif10l1eOFlTCSpoJMPwpTDnUMx3nALpOUaD40BPJzK6Q9LH
ERJvUiYEZ/bkedKslJ2zMro/LVWv8zgK4AcgmPgHNQBXegBhqAgEfwDF7Bm/VkvVjv1se0dcHKZ/bAH1ifPwDOlJo=</latexit>� = 20, DT = 10
<latexit sha1_base64="yfmWjMnUvNtJIcrJd1h/ivdYe8=">AB8X
icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1ItQ9OKxgv3ANpTJdtMu3WzC7kYof/CiwdFvPpvPlv3LY5aOuDgcd7M8zMCxLBtXHdb6ewsrq2vlHcLG1t
7+zulfcPmjpOFWUNGotYtQPUTHDJGoYbwdqJYhgFgrWC0e3Ubz0xpXksH8w4YX6EA8lDTtFY6bGLIhkiuSZer1xq+4MZJl4OalAjnqv/NXtxzSNmD
RUoNYdz02Mn6EynAo2KXVTzRKkIxywjqUSI6b9bHbxhJxYpU/CWNmShszU3xMZRlqPo8B2RmiGetGbiv95ndSEV37GZIaJul8UZgKYmIyfZ/0uWLU
iLElSBW3txI6RIXU2JBKNgRv8eVl0jyrehdV9/68UrvJ4yjCERzDKXhwCTW4gzo0gIKEZ3iFN0c7L8678zFvLTj5zCH8gfP5AzNDj/M=</latexit>� = 1
<latexit sha1_base64="v+50geBnlGfYijhr/Qpw8AHi0KI=">AB8ni
cbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1ItQ9OKxgrWFNJTJdtMu3WzC7kYoT/DiwdFvPprvPlv3LY5aOuDgcd7M8zMC1PBtXHdb6e0srq2vlHerGxt7+zuV
fcPHnWSKcpaNBGJ6oSomeCStQw3gnVSxTAOBWuHo9up35iSvNEPphxyoIYB5JHnKxkt9FkQ6RXBP7Vrbt2dgSwTryA1KNDsVb+6/YRmMZOGCtTa9z
UBDkqw6lgk0o30yxFOsIB8y2VGDMd5LOTJ+TEKn0SJcqWNGSm/p7IMdZ6HIe2M0Yz1IveVPzP8zMTXQU5l2lmKTzRVEmiEnI9H/S54pRI8aWIFXc3kroE
BVSY1Oq2BC8xZeXyeNZ3buou/fntcZNEUcZjuAYTsGDS2jAHTShBRQSeIZXeHOM8+K8Ox/z1pJTzBzCHzifP6QJkC0=</latexit>� = 10
<latexit sha1_base64="MgogXpwrP2/PI/J64WxN78zV9j0=">AB8n
icbVBNS8NAEJ3Ur1q/qh69LBbBU0mKqBeh6MVjBfsBaSib7aZdutmE3YlQSn+GFw+KePXePfuG1z0NYHA4/3ZpiZF6ZSGHTdb6ewtr6xuVXcLu3s
7u0flA+PWibJNONlshEd0JquBSKN1Gg5J1UcxqHkrfD0d3Mbz9xbUSiHnGc8iCmAyUiwShaye9SmQ4puSE1t1euFV3DrJKvJxUIEejV/7q9hOWxV
whk9QY3NTDCZUo2CST0vdzPCUshEdcN9SRWNugsn85Ck5s0qfRIm2pZDM1d8TExobM45D2xlTHJplbyb+5/kZRtfBRKg0Q67YlGUSYIJmf1P+kJz
hnJsCWVa2FsJG1JNGdqUSjYEb/nlVdKqVb3LqvtwUanf5nEU4QRO4Rw8uI63EMDmsAgWd4hTcHnRfn3flYtBacfOY/sD5/AGljpAu</latexit>� = 20
B
<latexit sha1_base64="AE4U7J6tq1m9/R7/tgBxel2cIUI=">ACBHicdVDLSsNAFJ3UV62vqMtuBovgQkJSQ1sLQlEXLiu0tdCEMJlO26GTBzMToYQu3Pgrblwo4taPcOfOGkrq
OiBgXPuZc79/gxo0Ka5oeW1peWV3Lrxc2Nre2d/TdvY6IEo5JG0cs4l0fCcJoSNqSka6MSco8Bm58cXmX9zS7igUdiSk5i4ARqGdEAxkry9KDWDxC8Axax9CpO/VL21NVWkalqeXTO0VinbFVWaZtUqWxkpV+0TG1pKyVACzQ9/d3pRzgJSCgxQ0L0LDOWboq4pJiRacFJBIkRHqMh6SkaoAIN50dMYWHSunDQcTVCyWcqd8nUhQIMQl81RkgORK/vUz8y+slclBzUx
rGiSQhni8aJAzKCGaJwD7lBEs2UQRhTtVfIR4hjrBUuRVUCF+Xwv9Jp2xYFcO8tkuN80UceVAEB+AIWKAKGuAKNEbYHAHsATeNbutUftRXudt+a0xcw+AHt7ROZzpWD</latexit>� = 1, DT = 0.1
Figure 3. The number of cell–cell contact events in a fixed interval of time (T = 1000) plotted here as
a function of the scaled effective diffusivity, DT, which represents the random motility of cell B. Here
we show how the number of cell–cell contact varies for three different elastic interaction strength
values, a, corresponding to substrates with three different stiffness. The highlighted points numbered
from (1)-(4), show representative cell trajectories over long times and highlight how varying a and
DT can yield states where the cells are in close proximity most of the time (low DT, high a) or states
where cells interact rarely (high DT, low a). Interpretation of the box plots is the same as in Figure 2.
The simulation was run for a total time of T = 1000 and updates in cell position were made every
dt = 0.001.
4. Results
4.1. Cell-cell contact frequency shows biphasic dependence on matrix elastic interactions
Motivated by experiments which show that two cells make repeated contact and
withdrawals on soft substrates, with contact frequency dependent on the substrate stiffness,
we measure the total number of contacts of the motile cell (A) with the stationary cell (B) in
our model simulations. As indicated earlier, the simulated cells are initialized randomly
inside the box, but outside of a pre-defined contact radius around the stationary cell. The
total number of contacts between the cells is counted over a fixed period of time i.e. T =
1000. It should be remembered that the cells are confined to stay within the square domain
during the course of the simulation.
Cell A’s movement is governed by an attractive elastic potential induced by the
stationary, central cell and its own random motion, described as an effective diffusion.
Additionally when the cell encounters the bounding wall of the square domain, it reflects
(moves away) from it. Overall, random noise encapsulated in the diffusion coefficient
causes A to move towards or away from B in an unbiased manner. The attractive potential
W being isotropic and spatially varying suggests that there is a critical radius of influence
(dependent on both a and DT) within which forces due to the attractive potential dominate
diffusion and significantly influence the trajectory of cell A. This effect results in the cell
getting closer to cell B, eventually entering this zone of influence.
9 of 16
To carefully study how elastic interactions (a) and random diffusion (DT) each influ-
ence this process, we first systematically calculated the number of contacts by a, while
keeping DT constant at three different values, DT = 1,2,5. (Figure 2). As illustrated by the
dotted lines which serve as a guide to the eye, the behavior is highly non-monotonic. For
small a, the number of contacts increases with increasing a, then reduces to 1 at high a.
The position of the peak increases with increasing DT. The initial increase in contacts is
due to the increased directional movement of the test cells towards the central cell. The
decrease in the number of contacts for very high values of a is expected since the attractive
potential is strong enough to overcome the effect of diffusion. In this case, the motile
cell is unable to move away from and makes stable contact with the stationary cell. For
a = 5 and DT = 1 (trajectory 1), the test cell spends most of the time exploring space rather
than near the stationary cell, which also reduces the number of contacts. Increasing a to
10 (trajectory 2) the radius of influence increases, increasing the duration of contact and
thereby increasing contacts. On further increasing a to 20 (trajectory 3), the test cell is
tightly adhered to the stationary cell which allows only one single contact. Note that the
statistics for the high DT and low a regime are influenced by the confinement. Cells in this
particular limit frequently escape the region of influence and wander away only to return
again after encountering the wall and diffusing away. For instance, the number of contacts
for DT = 5 and a = 0.1, combines the effect of repeated escapes from the region of influence
and repeated returns due to confinement. Since the size of the box is fixed, the increase in
number of contacts with DT for a = 0.1 is still a signature of diffusive effects dominating
the attractive potential.
We next investigated the effect of increasing diffusivity on the number of contacts
for constant a (1, 10 and 20). Results from this set of simulations are shown in Figure
3. The red dotted line serves as a guide to the eye highlighting the trend observed. We
see a steady increase in cell-cell contacts with diffusivity. Without diffusion, the test cell
shows unidirectional motion towards the central cell and remains in contact throughout
the simulation. Increasing diffusion increases the chance of test cell to go out of the radius
of influence and come back again (trajectories 3 and 4).
Overall combining the results shown in Figures 2 and 3, we conclude that the number
of contacts is maximized at an optimal value of the elastic interaction strength. If the elastic
strength is too high or too low, the cell either makes stable contact or is too motile to make
too many contacts. This optimal value scales with the diffusivity, which is a measure of the
cell motility in our model.
4.2. Cell motility characteristics depend on elastic interactions
To quantify the long-time statistics of the motility of cell A in the elastic potential
field generated by cell B, we analyze the mean squared displacement (MSD) as given by
equation (9) from simulation. The metric MSD measured in terms of a delay time t contains
information about the short time mobility of a cell, the long time mobility of the cell, and
additionally provides signatures of capture and trapping effects. Specifically, the slope
of the mean square displacement can be used to extract effective exponents that provides
insight on the relative importance of diffusion and elastic attractive interactions.
We plot the MSD in Fig. 4 for DT = 2 and a = 0.1,1,5,10,20,100. For a = 0.1,1,5,10,
we find that the slope is close to 1, which suggests diffusion drives the motion of the cell
and the attractive potential is not strong enough to influence the movement of the cell.
For higher a, we observe a transition towards sub-diffusive behavior at t ⇠ 0.5. At a = 20
(green line), the curve shows a significant decrease in slope at t = 2, the time scale for
which a test cell in average encounters the central cell for the first time and stays in contact
for a while, as shown by trajectory 3, Figure 3. The slope then increases again, but remains
less than 1 suggesting a sub-diffusive behavior in the long run. At a = 100 (blue line),
the MSD saturates after initial diffusion to a zero slope which suggests that the motion is
bounded, and it can only explore the circumference of the stationary cell.
10 of 16
10-2
10-1
100
101
102
10-2
100
102
104
MSD
=0.1
=1
=5
=10
=20
=100
69
70
71
72
73
74
540
560
580
MSD
=0.1
=1
=5
<latexit sha1_base64="zjtCf8Hr8dE1aetXG1h2l1nUV0k=">AB+HicbVDLSgNBEOz1GeMjqx69DAbBU9gVUS9C0IvHCOYByRJ6J7PJkNnZWZWiEu+xIsHRbz6Kd7
8GyePgyYWdFNUdTM9FaCa+N5387K6tr6xmZhq7i9s7tXcvcPGjrJFGV1mohEtULUTHDJ6oYbwVqpYhiHgjXD4e3Ebz4ypXkiH8woZUGMfckjTtFYqeuWOijSARJyTXzPq3hdt2z7FGSZ+HNShjlqXfer0toFjNpqECt276XmiBHZTgVbFzsZJqlSIfYZ21LJcZMB/n08DE5sUqPRImyJQ2Zqr83coy1HsWhnYzRDPSiNxH/89qZia6CnMs0M0zS2UN
RJohJyCQF0uOKUSNGliBV3N5K6AVUmOzKtoQ/MUvL5PGWcW/qHj35+XqzTyOAhzBMZyCD5dQhTuoQR0oZPAMr/DmPDkvzrvzMRtdceY7h/AHzucPxTCRNA=</latexit>� = 100.0
<latexit sha1_base64="bR3mqEHla/OljzIiyR3OipFU52U=">AB9XicbVBNSwMxEJ2tX7V+VT16CRbBU9ktol6EohePFewHtGuZTdM2NJtdkqxSlv4PLx4U8ep/8ea
/MW3oK0PZni8N0MmL4gF18Z1v53cyura+kZ+s7C1vbO7V9w/aOgoUZTVaSQi1QpQM8ElqxtuBGvFimEYCNYMRjdTv/nIlOaRvDfjmPkhDiTvc4rGSg8dFPEQCbkiFbfsdosl2cgy8TLSAky1LrFr04voknIpKECtW57bmz8FJXhVLBJoZNoFiMd4YC1LZUYMu2ns6sn5MQqPdKPlC1pyEz9vZFiqPU4DOxkiGaoF72p+J/XTkz/0k+5jBPDJ0/1E8
EMRGZRkB6XDFqxNgSpIrbWwkdokJqbFAFG4K3+OVl0qiUvfOye3dWql5nceThCI7hFDy4gCrcQg3qQEHBM7zCm/PkvDjvzsd8NOdkO4fwB87nD91kMo=</latexit>� = 20.0
<latexit sha1_base64="KOqRLIldbIVDcYZdyZUNX8U206c=">AB9XicbVBNSwMxEJ31s9avqkcvwSJ4Krsi6kUoevFYwX5Au5bZNuGZrNLklXK0v/hxYMiXv0v3vw
3pu0etPXBDI/3ZsjkBYng2rjut7O0vLK6tl7YKG5ube/slvb2GzpOFWV1GotYtQLUTHDJ6oYbwVqJYhgFgjWD4c3Ebz4ypXks780oYX6EfclDTtFY6aGDIhkgIVfEcytut1S2fQqySLyclCFHrVv6vRimkZMGipQ67bnJsbPUBlOBRsXO6lmCdIh9lnbUokR0342vXpMjq3SI2GsbElDpurvjQwjrUdRYCcjNAM9703E/7x2asJLP+MySQ2TdPZQmAp
iYjKJgPS4YtSIkSVIFbe3EjpAhdTYoIo2BG/+y4ukcVrxzivu3Vm5ep3HUYBDOIT8OACqnALNagDBQXP8ApvzpPz4rw7H7PRJSfOYA/cD5/ANvukMk=</latexit>� = 10.0
<latexit sha1_base64="ncuhdQ/iZoQBY6/XDnj01Mzs/tg=">AB9HicbVDJSgNBEK2JW4xb1KOXxiB4GmbE7SIEvXiMYBZIhlDT6Uma9Cx29wTCkO/w4kERr36MN/
GTjIHTXxQ8Hiviqp6fiK40o7zbRVWVtfWN4qbpa3tnd298v5BQ8WpKxOYxHLlo+KCR6xuZasFYiGYa+YE1/eDf1myMmFY+jRz1OmBdiP+IBp6iN5HVQJAMk5IZc2E63XHFsZwayTNycVCBHrVv+6vRimoYs0lSgUm3XSbSXodScCjYpdVLFEqRD7LO2oRGTHnZ7OgJOTFKjwSxNBVpMlN/T2QYKjUOfdMZoh6oRW8q/ue1Ux1cexmPklSziM4XBak
gOibTBEiPS0a1GBuCVHJzK6EDlEi1yalkQnAX14mjTPbvbSdh/NK9TaPowhHcAyn4MIVOEealAHCk/wDK/wZo2sF+vd+pi3Fqx85hD+wPr8AXC6kJM=</latexit>� = 5.0
<latexit sha1_base64="0IyVoHMjXwAhpJhBT8qlQuC9CJk=">AB9HicbVBNS8NAEJ34WetX1aOXxSJ4ComIehGKXjxWsB/QhjLZbtqlm03c3RK6e/w4kERr/4Yb/4
bt20O2vpg4PHeDPzwlRwbTzv21lZXVvf2CxsFbd3dvf2SweHdZ1kirIaTUSimiFqJrhkNcONYM1UMYxDwRrh4G7qN4ZMaZ7IRzNKWRBjT/KIUzRWCto0j4SckN81+uUyp7rzUCWiZ+TMuSodkpf7W5Cs5hJQwVq3fK91ARjVIZTwSbFdqZinSAPdayVGLMdDCeHT0hp1bpkihRtqQhM/X3xBhjrUdxaDtjNH296E3F/7xWZqLrYMxlmhkm6XxRlAl
iEjJNgHS5YtSIkSVIFbe3EtpHhdTYnIo2BH/x5WVSP3f9S9d7uChXbvM4CnAMJ3AGPlxBe6hCjWg8ATP8ApvztB5cd6dj3nripPHMEfOJ8/aqKQjw=</latexit>� = 1.0
<latexit sha1_base64="UipcCNiwk3tyJxC6qK4wvY+qE=">AB9HicbVBNS8NAEJ34WetX1aOXxSJ4ComIehGKXjxWsB/QhjLZbtqlm03c3RK6e/w4kERr/4Yb/4
bt20O2vpg4PHeDPzwlRwbTzv21lZXVvf2CxsFbd3dvf2SweHdZ1kirIaTUSimiFqJrhkNcONYM1UMYxDwRrh4G7qN4ZMaZ7IRzNKWRBjT/KIUzRWCto0j4SckM81+Uyp7rzUCWiZ+TMuSodkpf7W5Cs5hJQwVq3fK91ARjVIZTwSbFdqZinSAPdayVGLMdDCeHT0hp1bpkihRtqQhM/X3xBhjrUdxaDtjNH296E3F/7xWZqLrYMxlmhkm6XxRlAl
iEjJNgHS5YtSIkSVIFbe3EtpHhdTYnIo2BH/x5WVSP3f9S9d7uChXbvM4CnAMJ3AGPlxBe6hCjWg8ATP8ApvztB5cd6dj3nripPHMEfOJ8/aqCQjw=</latexit>� = 0.1
<latexit sha1_base64="ncuhdQ/iZoQBY6/XDnj01Mzs/tg=">AB9HicbVDJSgNBEK2JW4xb1KOXxiB4GmbE7SIEvXiMYBZIhlDT6Uma9Cx29wTCkO/w4kERr36MN/
GTjIHTXxQ8Hiviqp6fiK40o7zbRVWVtfWN4qbpa3tnd298v5BQ8WpKxOYxHLlo+KCR6xuZasFYiGYa+YE1/eDf1myMmFY+jRz1OmBdiP+IBp6iN5HVQJAMk5IZc2E63XHFsZwayTNycVCBHrVv+6vRimoYs0lSgUm3XSbSXodScCjYpdVLFEqRD7LO2oRGTHnZ7OgJOTFKjwSxNBVpMlN/T2QYKjUOfdMZoh6oRW8q/ue1Ux1cexmPklSziM4XBak
gOibTBEiPS0a1GBuCVHJzK6EDlEi1yalkQnAX14mjTPbvbSdh/NK9TaPowhHcAyn4MIVOEealAHCk/wDK/wZo2sF+vd+pi3Fqx85hD+wPr8AXC6kJM=</latexit>� = 5.0
<latexit sha1_base64="0IyVoHMjXwAhpJhBT8qlQuC9CJk=">AB9HicbVBNS8NAEJ34WetX1aOXxSJ4ComIehGKXjxWsB/QhjLZbtqlm03c3RK6e/w4kERr/4Yb/4
bt20O2vpg4PHeDPzwlRwbTzv21lZXVvf2CxsFbd3dvf2SweHdZ1kirIaTUSimiFqJrhkNcONYM1UMYxDwRrh4G7qN4ZMaZ7IRzNKWRBjT/KIUzRWCto0j4SckN81+uUyp7rzUCWiZ+TMuSodkpf7W5Cs5hJQwVq3fK91ARjVIZTwSbFdqZinSAPdayVGLMdDCeHT0hp1bpkihRtqQhM/X3xBhjrUdxaDtjNH296E3F/7xWZqLrYMxlmhkm6XxRlAl
iEjJNgHS5YtSIkSVIFbe3EtpHhdTYnIo2BH/x5WVSP3f9S9d7uChXbvM4CnAMJ3AGPlxBe6hCjWg8ATP8ApvztB5cd6dj3nripPHMEfOJ8/aqKQjw=</latexit>� = 1.0
<latexit sha1_base64="UipcCNiwk3tyJxC6qK4wvY+qE=">AB9HicbVBNS8NAEJ34WetX1aOXxSJ4ComIehGKXjxWsB/QhjLZbtqlm03c3RK6e/w4kERr/4Yb/4
bt20O2vpg4PHeDPzwlRwbTzv21lZXVvf2CxsFbd3dvf2SweHdZ1kirIaTUSimiFqJrhkNcONYM1UMYxDwRrh4G7qN4ZMaZ7IRzNKWRBjT/KIUzRWCto0j4SckM81+Uyp7rzUCWiZ+TMuSodkpf7W5Cs5hJQwVq3fK91ARjVIZTwSbFdqZinSAPdayVGLMdDCeHT0hp1bpkihRtqQhM/X3xBhjrUdxaDtjNH296E3F/7xWZqLrYMxlmhkm6XxRlAl
iEjJNgHS5YtSIkSVIFbe3EtpHhdTYnIo2BH/x5WVSP3f9S9d7uChXbvM4CnAMJ3AGPlxBe6hCjWg8ATP8ApvztB5cd6dj3nripPHMEfOJ8/aqCQjw=</latexit>� = 0.1
Mean square displacement,
Delay time,
<latexit sha1_base64="yGbOJNxjeEKB5Z2Sb3L8DaMe/10=">AB/XicbVDLSgMxFM
3UV62v8bFzEyxC3ZQZEXVZ1IUboaJ9QGcomTRtQ5PMkGSEOgz+ihsXirj1P9z5N2baWjrgcDhnHu5JyeIGFXacb6twsLi0vJKcbW0tr6xuWVv7zRVGEtMGjhkoWwHSBFGBWloqhlpR5
IgHjDSCkaXmd96IFLRUNzrcUR8jgaC9ilG2khdey/xONJDyZObu6s0rXgaxUdu+xUnQngPHFzUgY56l37y+uFOZEaMyQUh3XibSfIKkpZiQtebEiEcIjNCAdQwXiRPnJH0KD43Sg/
1Qmic0nKi/NxLElRrzwExmUdWsl4n/eZ1Y98/9hIo1kTg6aF+zKAOYVYF7FJsGZjQxCW1GSFeIgkwtoUVjIluLNfnifN46p7WnVuT8q1i7yOItgHB6ACXHAGauAa1EDYPAInsEreL
OerBfr3fqYjhasfGcX/IH1+QNc9JUo</latexit>MSD(�)
<latexit sha1_base64="lCedxjNk/DYMPs6qZXTtQUTGNXI=">AB63icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPRi8cK9gPaUDbTbt0swm7E6GE/gUvHhTx6h/y5r9
x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wSTu9zvPHFtRKwecZpwP6IjJULBKOZSH2k6qNbcujsHWSVeQWpQoDmofvWHMUsjrpBJakzPcxP0M6pRMlnlX5qeELZhI54z1JFI278bH7rjJxZUjCWNtSObq74mMRsZMo8B2RhTHZtnLxf+8XorhjZ8JlaTIFVsClNJMCb542QoNGc
op5ZQpoW9lbAx1ZShjadiQ/CWX14l7Yu6d1V3Hy5rjdsijKcwCmcgwfX0IB7aEILGIzhGV7hzYmcF+fd+Vi0lpxi5hj+wPn8ASLTjk0=</latexit>�
Figure 4. Mean square displacement (MSD) as a function of the delay time interval t (calculated
from Equation 9), for the motile cell A is shown. Here we explore the variation in the MSD for
various values of substrate-mediated elastic interactions, a. The diffusivity DT is held constant for
these simulations with DT = 2. Other diffusivities were explored (results not shown). At low elastic
interaction strengths, a, corresponding to stiff substrates, the cell shows a purely diffusive trajectory,
whereas at higher values of a, the motile cell is captured by the strong attractive interaction from
the stationary cell, resulting in a flattening of the MSD (blue curve). At an intermediate interaction
regime (green curve), the motile cell makes repeated contact with the fixed cell but is never fully
captured.
4.3. Elastic interactions lead to effective capture of motile cell
Taken together, our simulations suggest that strongly attractive elastic interactions
can lead to stable contact between initially distant cells. We next explore the statistics of
this “capture” process. Capture mechanisms underlying and influencing these statistics
are potentially relevant for timescales of contact formation between initially well-separated
motile cells that then form confluent monolayers, such as in mesenchymal–to–epithelial
transitions during tissue morphogenesis [33].
Figures 2 and 3 suggest that the motile cell A (as it explores space and samples
the potential field over its various trajectories) is attracted to the stationary cell with the
attracting force increasing with decreasing distance r. Acting in tandem and superposed on
this aspect of the motion is diffusion that allows A to wander away from B multiple times.
In order to understand how parameters a and DT affect this phenomenon, we tracked
the number of cells inside the contact radius over the course of the simulation. The
probability of cells inside the contact radius reached a steady state at time t < 100 for all
parameters (Figure 5A). Keeping a constant and increasing DT the probability of cells being
inside the contact radius decreases (Figure 5B). The steady-state probability PSS increases
with increase in a for constant DT (Figure 5C). To understand the relationship between PSS
and both a and DT, we investigated PSS for the ratio a/DT and showed that they remain
constant for this ratio.
Plotting PSS vs a/DT, the strength of the elastic interactions relative to the diffusivity,
we find that the data can be collapsed into a single master curve (Figure 5D). The collapse of
our data and the master curve plotted in Figure 5D is expected since our model steady state
is a thermal equilibrium with effective temperature set by the value of DT; the competition
between attractive interactions and noise meanwhile dictates how many cells are captured
vs. how many can escape.
11 of 16
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9AuaUCbt0swm7G6GE/g
0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu56bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0n
kSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9Aphy
T948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jy
HXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIb
H0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj
04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQ
Wf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR
5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv
6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/r
auFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1
hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAri
CKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948
aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/
3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj
04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvT
k/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhT
uoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9A
uaUCbt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu5
6bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj
5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCe
UCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEw
zalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctn
juAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA="
>AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQ
Wf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aR
bKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHX
GFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAri
CKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyh
NnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFT
mp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=<
/latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr
6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/L
BjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+
KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9AuaUCb
bt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu56bmCBDZTgVbFryU
80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj5zCL/gfHwDszuRdw=</late
xit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhN
nJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5Q
tmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latex
it>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">A
AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1
ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIK
clCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPv
XYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29
eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJ
BkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUV
jbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHL
YBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPO
YUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YO
YJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9A
uaUCbt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu5
6bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj
5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCe
UCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEw
zalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctn
juAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQ
PdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr
6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/
/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZL
B4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaP
WA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0
S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnD
qlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJX
bL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnF
eCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/e
x7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyh
NnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFT
mp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=<
/latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr
6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/L
BjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+
KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9AuaUCb
bt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu56bmCBDZTgVbFryU80SpG
Mcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJ
JBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDn
jbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">A
AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ld
W9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCF
HrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUk
iJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1
xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBk
zO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNs
du2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHLY
BA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPOYUrZ
MaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YOYJUc3c
roQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9A
uaUCbt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu5
6bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj
5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA
=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdH
dFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMop
MatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZur
viYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu
3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1
_base64="rxvQPdcIbH0VPSVp
lj1yBJIN9tA=">AB7XicbVD
LSgNBEOz1GeMr6tHLYBA8hV0R
9Rj04jGCeUCyhNnJBkzO7PM9
AphyT948aCIV/Hm3/jJNmDJh
Y0FXdHdFiRQWf/bW1ldW9/
YLGwVt3d29/ZLB4cNq1PDeJ1p
qU0ropZLoXgdBUreSgyncSR5M
xrdTv3mEzdWaPWA4SHMR0o0R
eMopMatW5m7aRbKvsVfwayTIK
clCFHrVv6vQ0S2OukElqbTvw
EwzalAwySfFTmp5QtmIDnjbU
UVjbsNsdu2EnDqlR/rauFJIZu
rviYzG1o7jyHXGFId20ZuK/3n
tFPvXYSZUkiJXbL6on0qCmkxf
Jz1hOEM5doQyI9ythA2poQxdQ
EUXQrD48jJpnFeCy4p/f1Gu3u
RxFOAYTuAMAriCKtxBDerA4BG
e4RXePO29eO/ex7x1xctnjuAP
vM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyh
NnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFT
mp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=<
/latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ
0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmR
VCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQu
lWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJG
XLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eG
GNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhC
CBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kD
jSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9
AuaUCbt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpq
ECtu56bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8
+a8L1oLTj5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jG
CeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqb
TvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/
ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=
">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdF
iRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW
5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG
1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFO
AYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUC
yhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalA
wySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8
fzk2PSA=</latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeM
r6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6Y
NV/LBjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrl
M8sknS+KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
Figure 5. Capture statistics of motile cell. (A) Probability that cell B is inside contact radius as a
function of time. (B and C) The dependence of steady state capture probability, Pss, i.e. the fraction
of cells captured within the contact radius after a long time interval, on simulation parameters. (B)
shows the dependence on diffusivity,DT at different values of the elastic interaction parameter, a,
whereas (C) shows the dependence on a for different values of DT. (D) The steady state capture
probability, Pss, data can be collapsed into a single master curve, when plotted vs. the key parameter,
alpha/DT, the strength of the elastic interactions relative to the diffusivity. This is expected since our
model steady state is a thermal equilibrium with effective temperature set by the noisy cell motility,
DT, and the competition between attractive interactions and noise dictates the number of cells (cell
trajectories) captured vs. the number that escape.
This further justifies the notion introduced earlier of a radius of influence, that is, the
distance from the stationary cell at which its elastic attractive tendency approximately
balances the random noisy movements of the motile cell. Here we use a simple balance to
estimate this radius of influence. Working in dimensionless units, we note that the dipolar
interaction potential fall off as a/r3, while the effective temperature – a measure of the
randomizing force – scales as kBT = µTDT. Balancing these yields,
rI ⇠
✓
a
µTDT
◆ 1
3
,
(10)
which explicitly shows the importance of the a/DT parameter. Thus, a stronger a from
deformations exerted by the stationary cell (corresponding to softer substrate stiffness, or
higher contractility) and lower random movements of the motile cell, DT, leads to a larger
radius of influence. This in turn implies that the probability of being captured within the
contact radius increases because the stationary cell can influence motile cells over a larger
area.
12 of 16
4.4. Future work and perspectives: Anisotropic cell-cell elastic interactions
For polarized cells, that orient their cytoskeletal fibers and contractility along some
principal axis, the cell-cell interaction potential is not isotropic. The individual cells on
an elastic medium behave as force dipoles, with interaction potential energy having both
attractive and repulsive regions that depend on mutual orientation of the two cells and
their separation vector [19], as detailed in Appendix A. The force experienced by the motile
cell has both radial and tangential components depending on its position and orientation
relative to the central cell, and its direction is sensitive to the Poisson’s ratio of the elastic
medium [34]. Thus, trajectories of cell A interacting with stationary cell B when the fully
anisotropic interaction potential (Equation A1 and A2, Appendix A) is included will differ
from trajectories observed in isotropic potentials. The difference arises in part due to an
additional torque that reorients cell A to preferentially align with cell B as it moves towards
it. Nonetheless, qualitative nature of the capture process and the observation of an effective
region of influence will still remain valid.
X
Y
B Fixed cell
B
Cell A initially aligned
normal to dipole axis
Cell A initially aligned
along dipole axis
Figure 6. Dipolar cell orientation and trajectoryThe equilibrium orientation of contractile cells fixed
in position, but free to reorient, and that are uniformly distributed in a square box of size 10s, are
depicted by two arrows (red) pointing towards each other. Each cell is influenced by the central
stationary cell B (green) and not by each other. Two possible trajectories of cell A (blue and black) are
recorded for DT = 0.1, a = 40 for total time T = 500 with time steps of dt = 0.001. The cells did not
have any self propulsion or rotational diffusion. The Poisson’s ratio n of the substrate was considered
0.3 for this simulation
To illustrate this we simulated the equilibrium orientation of uniformly spaced
(pinned) test dipolar cells on a square lattice which are kept fixed in a square box of
length 10s. The Poisson’s ratio of the simulated substrate is 0.3 and a is 40. Results are
shown in Figure 6. None of the cells overlap with the central stationary cell; they may
rotate to reorient their dipole axis but are restricted from translating. We re-iterate that
the cells on the lattice do not mutually interact with each other, but are only meant to
illustrate the interaction of a test dipolar cell A placed at different spatial locations with
the central stationary cell B. We note that fixed cells adjust the axis of their contractile
dipoles in accordance to the potential field due to cell B (the dipole axis of B is fixed).
Superposed on this are two trajectories corresponding to two cells that are freed from
constraints and allowed to rotate and translate in response to the two-cell potential and
thermal noise. The two cells start from their equilibrium orientation - i.e, they are first
held pinned and allowed to reorient until the dipole axis attains a static value and then the
pinning constraint is removed. Cells in the close vicinity of the central cell’s orientation
axis exhibit a nearly linear motion to the pole of the fixed cell (trajectory in black). Cells
away from the orientation axis take a longer route to come in contact with the central cell
(trajectory in blue). The common attribute in both trajectories is that they prefer to adhere
to the central cell’s pole, that is cell A as it moves towards B also continuously reorients in
a manner that brings it into alignment with the cell B’s polar axis (the axis of the dipole).
13 of 16
5. Discussion
Using our model for cell contractility and motility, we computed several metrics of
experimental relevance such as number of cell–cell contacts, the mean square displacement
of a motile cell in the presence of elastic deformations induced by a cell in its vicinity, and
associated capture statistics resulting from attractive interactions between two such cells.
In each case, we predict how the computed metric depends on the elastic properties of the
substrate, captured in the interaction parameter, a ⇠ 1/E, and on cell motility, captured by
the effective diffusivity, DT.
Similar to the observations for pairs of endothelial cells mechanically interacting
through the compliant substrates [3], we find that the motility and number of cell-cell
contacts are lowered at large a, corresponding to softer substrates. This is because the
elastic deformations of the substrate, and therefore, the cell–cell attractive interactions are
stronger compared to the random motility. As observed in experiments, we also find that at
intermediate interaction strength, the cells can make repeated contacts and withdrawals as
shown in the contact number measurements. For very stiff substrates, that is low interaction
strength, we find the cell remains diffusive and can migrate away from the stationary cell
and does not make frequent contacts. Our findings would therefore suggest an optimal
substrate stiffness at which contact frequency is maximal. These trends are also reflected
in the MSD measurements. Unlike the experiment, we don’t find diffusive MSD for the
strongly attractive case, but the MSD turns subdiffusive, suggesting perhaps that such high
interaction strengths were not probed in experiment.
Biologically, such altered motility and contact formation could be relevant for forming
stable adhesive contacts between cells and tissue development, including that of blood
vessels during vasculogenesis [35]. We made several simplifying assumptions in the model
(stated in section 2), including using a purely attractive and isotropic potential instead
of the dipolar potential relevant for elongated and motile cells. Fig. 6 illustrates how
the position and orientation of the motile cell with respect to the stationary cell leads
to qualitatively different trajectories when the interaction potential is dipolar. Such an
anisotropic potential is expected to lead to end–to–end alignment and contact formation of
a pair of cells. With multiple cells, larger scale structures such as chains and networks of
cells can result [19]. The influence of cellular motility on these structures will be the topic
of a future study. The advantages of complementing experimental studies with modeling
approaches as discussed in this paper is that hard to realize parameter regimes may be
easily investigated. Furthermore, the role of different physical parameters may be clearly
studied in isolation; a feature hard to achieve in an experimental setting.
In summary, our results illustrate how cell–cell mechanical interactions can lead to
their mutual contact formation without requiring specific chemical factors to guide their
motility, and how the substrate stiffness is an important control parameter in guiding cell
motility and forming multi-cellular structures. The computational framework introduced
and analyzed here can be extended to study durotaxis – that is, the modification of cell
motility by variations in substrate elasticity at the single cell or tissue level and the motion
of cells towards higher stiffness regions [36,37]. Understanding the mechanistic aspects of
cell-cell interactions as done here has implications for regenerative medicine and tissue
engineering and will guide and inform experiments exploring how cells communicate with
each other in the process of organizing and moving collectively.
Author Contributions: Conceptualization, K.D. and A.G.; methodology, A.G. and K.D.; software,
A.G and S.B.; validation, S.B., K.D and A.G.; investigation, S.B.; resources, K.D. and A.G; writing,
S.B., K.D and A.G. All authors have read and agreed to the published version of the manuscript.
Funding: AG acknowledges funding from NSF-MCB-2026782. SB, KD and AG also acknowledge
funding from the National Science Foundation: NSF-CREST: Center for Cellular and Biomolecular
Machines (CCBM) at the University of California, Merced: NSF-HRD-1547848.
Institutional Review Board Statement: Not applicable.
14 of 16
Data Availability Statement: Data is contained within the article or supplementary material.
Conflicts of Interest: The authors declare no conflict of interest.
Abbreviations
The following abbreviations are used in this manuscript:
MSD
Mean Square Displacement
Appendix A Model for a moving cell interacting with a stationary cell via substrate
elasticity
The flat substrate is treated as being semi-infinite (Figure 1) and comprised of a
linearly elastic, isotropic gel-like material with Young’s modulus E and Poisson’s ratio n,
that capture its stiffness and compressibility respectively. The minimal model that describes
the deformations created by cells exerting contractile forces on the substrate is a point-like
force dipole [31]. Two identical dipolar cells denoted by A and B move in the upper plane
(chosen to be the x-y plane, see Figure 1). Cell A is allowed to move and its dynamics
is specified completely by its location on the substrate rA(t) and by its self-propulsion
direction eA(t). Cell B is held fixed at point rB. As a result of the contractile dipoles exerted
on the substrate the cells communicate elastically. The potential WAB characterizing this
elastic interaction between the two cells is given by
WAB(r) = P2eB
j eB
i ∂j∂lGAB
ik (r)eA
k eA
l ,
with r = rA � rB,
(A1)
where P is the strength of the force dipole capturing the contractile stresses exerted by a
cell on the medium. In writing (A1), we have made the plausible assumption that cells
orient their cytoskeletal structures such as stress fibers and exert their traction primarily
along their motility axis, such that the force dipole tensor, which captures the moment of
their force distribution, is assumed to be, Pij = Peiej. The tensor
GAB
ij (r) = 1 + n
pE
(1 � n)dij
r + nrirj
r3
�
,
(A2)
is the Green’s function that captures the displacement in the elastic medium at the location
of one cell (dipole) caused by the application of a point force at the location of the other
[38]. The partial derivatives in (A1) on the right hand side are taken with respect to relative
position vector r. Standard Einstein notation has been chosen in writing the form of WAB
and the derivatives in equations (A1) and (A2).
To obtain the force and torque balance equations that govern the dynamics of cell
A, we make the simplifying assumption that the cells move in an overdamped fashion.
This implies that hydrodynamic interactions between cells are ignored, and that each cell
feels a resisting viscous frictional drag/torque that is proportional to its velocity/rotation
rate. Conversely, when acted on by a force F or a torque T, a cell in this overdamped
environment will move with velocity µTF or rotate at a rate µRT respectively. Here, µT and
µR are appropriate mobility terms that depend on the cell size.
The micro-dynamics of cell A moving on the substrate is governed by the Langevin
equations for the translation and rotary motion of cell. Recognizing that the elastic interac-
tion generates (extra) forces and torques that act on each cell, and including the effects of
fluctuating time dependent forces xxxT(t) and torques xxxR(t) originating from thermal noise,
we can write the equations for the position and orientation of cell A in the presence of cell
B as
∂rA
∂t
=
v0eA � µT
∂WAB
∂rA
+ µTxxxT(t),
and
(A3)
∂eA
∂t
=
�µR
✓
eA ⇥ ∂WAB
∂eA
◆
+ µRxxxR(t).
(A4)
15 of 16
In an equilibrium situation, the random forces and torques are white noise terms and
are related to one another by the equipartition and fluctuation-dissipation theorems:
hxxxT(t)xxxT(t0)i = (2kBT/µT)dddd(t � t0) where ddd is the Kronecker delta function. For active
cells however, these restrictions do not hold; these terms are set by active internal cell
responses to the substrate properties. Equations (A1-A4) are used in the results illustrated
in Figure 5.
In the bulk of the paper and for results presented in Figures 1-4, we use an isotropic
version of the potential in equation (A1) that ignores orientational dynamics that are in
general present for highly elongated cells. This assumes a separation of scales between the
time over which cells reorient and the dipole axis changes and the time for the center of
the cell to move significantly such as when the rotation noise in (A4) is significant. In this
limit, one can average over the rapid reorientations of the cells and replace eB
j eB
i by dij and
eA
k eA
l by dkl. Equation (A1) then reduces to the simpler form that we employ in the main
discussion of the paper and implement as a numerical simulation,
WAB(r) = P2∂i∂kGAB
ik (r) = P2
E
f(n)
r3
(A5)
with the function f(n) = (1 � n2)/p dependent solely on the Poisson ratio, and hence
fixed in the simulation. Furthermore, since the dipole axis of cell A reorients in time scales
much faster than its slower rate of translation, the voeA term in (A3) simplifies to a time
fluctuating variable with a mean that is roughly zero but with a non-zero variance. Thus
its net effect may be incorporated by appropriately modifying the translational diffusivity.
For an isotropic symmetric potential as here, the equation that needs to be solved is then
∂r
∂t = �µT
∂WAB
∂r
+ µTxxxT
⇤ (t),
(A6)
with the modified random force xxxT
⇤ reflecting an effective translational diffusivity Deff differ-
ent from the thermal diffusivity D0, through a relation, hxxxT
⇤ (t)xxxT
⇤ (t0)i = (Deff/µ2
T)dddd(t � t0).
We define the dimensionless number DT ⌘ Deff/D0. Consistent with this, we choose
µT = Deff/kBT.
References
1.
Ladoux, B.; Mège, R.M. Mechanobiology of collective cell behaviours. Nature Reviews Molecular Cell Biology 2017, 18, 743–757.
2.
Alert, R.; Trepat, X. Physical Models of Collective Cell Migration. Annual Review of Condensed Matter Physics 2020, 11, 77–101.
3.
Reinhart-King, C.A.; Dembo, M.; Hammer, D.A. Cell-Cell Mechanical Communication through Compliant Substrates. Biophysical
Journal 2008, 95, 6044 – 6051.
4.
Winer JP, Oake S, J.P.. Non-Linear Elasticity of Extracellular Matrices Enables Contractile Cells to Communicate Local Position
and Orientation. PLoS ONE 2009, 4, e6382.
5.
Balaban, N.Q.; Schwarz, U.S.; Riveline, D.; Goichberg, P.; Tzur, G.; Sabanay, I.; Mahalu, D.; Safran, S.; Bershadsky, A.; Addadi, L.;
Geiger, B. Force and focal adhesion assembly: a close relationship studied using elastic micropatterned substrates. Nature Cell
Biology 2001, 3, 466. Article.
6.
Guo, C.L.; Ouyang, M.; Yu, J.Y.; Maslov, J.; Price, A.; Shen, C.Y. Long-range mechanical force enables self-assembly of epithelial
tubular patterns. Proceedings of the National Academy of Sciences 2012, 109, 5576–5582.
7.
Wang, H.; Abhilash, A.; Chen, C.S.; Wells, R.G.; Shenoy, V.B. Long-Range Force Transmission in Fibrous Matrices Enabled by
Tension-Driven Alignment of Fibers. Biophysical Journal 2014, 107, 2592 – 2603.
8.
Natan, S.; Koren, Y.; Shelah, O.; Goren, S.; Lesman, A. Long-range mechanical coupling of cells in 3D fibrin gels. Molecular Biology
of the Cell 2020, 31, 1474–1485. PMID: 32374653.
9.
Pakshir, P.; Alizadehgiashi, M.; Wong, B.; Coelho, N.M.; Chen, X.; Gong, Z.; Shenoy, V.B.; McCulloch, C.A.; Hinz, B. Dynamic
fibroblast contractions attract remote macrophages in fibrillar collagen matrix. Nature Communications 2019, 10, 1850.
10.
Yuval, J.; Safran, S.A.
Dynamics of elastic interactions in soft and biological matter.
Phys.
Rev.
E 2013, 87, 042703.
doi:10.1103/PhysRevE.87.042703.
11.
Tang, X.; Bajaj, P.; Bashir, R.; Saif, T.A. How far cardiac cells can see each other mechanically. Soft Matter 2011, 7, 6151–6158.
12.
Nitsan, I.; Drori, S.; Lewis, Y.E.; Cohen, S.; Tzlil, S. Mechanical communication in cardiac cell synchronized beating. Nature
Physics 2016, 12, 472–477.
16 of 16
13.
Dasbiswas, K.; Majkut, S.; Discher, D.E.; Safran, S.A. Substrate stiffness-modulated registry phase correlations in cardiomyocytes
map structural order to coherent beating. Nature Communications 2015, 6, 6085.
14.
Tomasek, J.J.; Gabbiani, G.; Hinz, B.; Chaponnier, C.; Brown, R.A. Myofibroblasts and mechano-regulation of connective tissue
remodelling. Nature Reviews Molecular Cell Biology 2002, 3, 349–363.
15.
Korff, T.; Augustin, H. Tensional forces in fibrillar extracellular matrices control directional capillary sprouting. Journal of Cell
Science 1999, 112, 3249–3258.
16.
Liu, Z.; Woo, S.; Weiner, O.D. Nodal signaling has dual roles in fate specification and directed migration during germ layer
segregation in zebrafish. Development 2018, 145.
17.
Pelham, R.J.; Wang, Y.l. Cell locomotion and focal adhesions are regulated by substrate flexibility. Proceedings of the National
Academy of Sciences 1997, 94, 13661–13665.
18.
Discher, D.E.; Janmey, P.; Wang, Y.l. Tissue Cells Feel and Respond to the Stiffness of Their Substrate. Science 2005, 310, 1139–1143.
doi:10.1126/science.1116995.
19.
Bischofs, I.B.; Schwarz, U.S. Cell organization in soft media due to active mechanosensing. Proceedings of the National Academy of
Sciences 2003, 100, 9274–9279.
20.
van Oers, R.; Rens, E.; LaValley, D.; Reinhart-King, C.; Merks, R. Mechanical Cell-Matrix Feedback Explains Pairwise and
Collective Endothelial Cell Behavior In Vitro. PLoS Comput Biol 2014, 10, e1003774.
21.
Schwarz, U.S.; Safran, S.A. Physics of adherent cells. Rev. Mod. Phys. 2013, 85, 1327–1381. doi:10.1103/RevModPhys.85.1327.
22.
Mandal, K.; Wang, I.; Vitiello, E.; Orellana, L.A.C.; Balland, M. Cell dipole behaviour revealed by ECM sub-cellular geometry.
Nature Communications 2014, 5, 5749.
23.
Mura, T. Micromechanics of Defects in Solids; TheHague: Martnius Nijhoff, 1991.
24.
Ben-Yaakov, D.; Golkov, R.; Shokef, Y.; Safran, S.A. Response of adherent cells to mechanical perturbations of the surrounding
matrix. Soft Matter 2015, 11, 1412–1424.
25.
Golkov, R.; Shokef, Y. Shape regulation generates elastic interaction between living cells. New Journal of Physics 2017, 19, 063011.
26.
Fily, Y.; Subramanian, P.; Schneider, T.M.; Chelakkot, R.; Gopinath, A. Buckling instabilities and spatio-temporal dynamics of
active elastic filaments. J. Roy. Soc. Interface 2019, 17, 20190794.
27.
Yang, J.; Arratia, P.E.; Patteson, A.E.; Gopinath, A. Quenching active swarms: effects of light exposure on collective motility in
swarming Serratia marcescens. J. Roy. Soc. Interface 2019, 16, 1620180960.
28.
Notbohm, J.; Lesman, A.; Rosakis, P.; Tirrell, D.A.; Ravichandran, G. Microbuckling of fibrin provides a mechanism for cell
mechanosensing. Journal of The Royal Society Interface 2015, 12, 20150320.
29.
Xu, X.; Safran, S.A. Nonlinearities of biopolymer gels increase the range of force transmission. Phys. Rev. E 2015, 92, 032728.
30.
Golkov, R.; Shokef, Y. Elastic interactions between anisotropically contracting circular cells. Phys. Rev. E 2019, 99, 032418.
31.
Bischofs, I.B.; Safran, S.A.; Schwarz, U.S. Elastic interactions of active cells with soft materials. Phys. Rev. E 2004, 69, 021911.
32.
Phillips, R.; Kondev, J.; Theriot, J. Physical Biology of the Cell; Garland Science, Taylor & Francis Group: New York, 2008.
33.
Pei, D.; Shu, X.; Gassama-Diagne, A.; Thiery, J.P. Mesenchymal–epithelial transition in development and reprogramming. Nature
Cell Biology 2019, 21, 44–53. doi:10.1038/s41556-018-0195-z.
34.
Bischofs, I.B.; Schwarz, U.S. Effect of Poisson Ratio on Cellular Structure Formation. Phys. Rev. Lett. 2005, 95, 068102.
35.
Califano, J.P.; Reinhart-King, C.A. A Balance of Substrate Mechanics and Matrix Chemistry Regulates Endothelial Cell Network
Assembly. Cellular and Molecular Bioengineering 2008, 1, 122.
36.
Sunyer, R.; Conte, V.; Escribano, J.; Elosegui-Artola, A.; Labernadie, A.; Valon, L.; Navajas, D.; García-Aznar, J.M.; Muñoz, J.J.;
Roca-Cusachs, P.; Trepat, X. Collective cell durotaxis emerges from long-range intercellular force transmission. Science 2016,
353, 1157–1161.
37.
Novikova, E.A.; Raab, M.; Discher, D.E.; Storm, C. Persistence-Driven Durotaxis: Generic, Directed Motility in Rigidity Gradients.
Phys. Rev. Lett. 2017, 118, 078103.
38.
Landau, L.D.; Lifshitz, E.M. Theory of Elasticity; Vol. 7, Course of Theoretical Physics, Pergamon Press: London, 1959.
2 of 16
1. Introduction
Many eukaryotic cells move by crawling, that is by adhering to and exerting mechan-
ical stresses and local forces on their extracellular matrix (ECM) that they then actively
deform (see for instance [1–4] and references therein). Existing approaches to modeling
collective cell motility focus on direct (steric and adhesive) cell-cell interactions or focus
at the single cell level on cell-substrate interactions [2] such as the details of focal adhe-
sions that are crucial to generating traction stresses in both adherent and motile cells [5].
Experiments strongly indicate however that cells cultured on soft, elastic, biocompatible
substrates can respond to each other even when not in direct contact [3,4]. Such interactions
can arise in cell culture experiments, with cells on the surface of synthetic hydrogels such as
polyacrylamide, which are linearly elastic, through mutual and active deformations of the
gel by the cells. These mechanically derived non-contact cell-cell interactions are even more
relevant and act over longer ranges in the biological extracellular matrix (ECM) comprising
collagen or fibrin, where cells can interact by remodeling and reorienting the fibers in the
ECM [6–8]. Even without such cell–matrix feedback, the presence of deformations has been
shown recently to guide the migration of other cells without requiring chemotactic cues [9].
Mechanical non-local interactions between cells offer advantages compared to chemi-
cal means. Specifically, mechanical signaling and mechanosensing of neighbouring cells
is typically faster and longer-ranged than chemical signaling. Chemical interactions are
limited by diffusion rates while mechanical interactions propagate near instantaneously
for purely elastic deformations [10]. Indeed, this crucially allows cells to not just sense
each other but also to synchronize their behaviour. For instance, substrate deformation-
mediated long-range interactions has been clearly demonstrated in heart muscle cells that
synchronize their beating without direct contact [11,12], as well as at a subcellular level
between myofibrils within a single heart muscle cell [13]. Cell communications via sensing
of substrate or matrix deformation are particularly important in sparse, non-confluent cell
cultures or tissue that occur in a number of biologically relevant situations. Apart from
beating cardiomyocytes, examples of such situations include wound healing involving
fibroblasts [14], sprouting blood vessels comprising endothelial cells [15], and migration
of mesenchymal cells in zebrafish embryo before the formation of confluent epithelial
tissue [16]. In all these cases, cells are not in direct contact but exert traction forces on the
surrounding mechanical medium and concomitantly sense deformations caused by nearby
cells. Such interactions therefore crucially depend on the stiffness of the substrate, and
can be probed by experiments that vary the stiffness of the hydrogel substrate on which
the cells are cultured [17,18]. These aspects influence not only motility response at the
single cell level but also strongly impact collective behavior including directed motility
and subsequent spatial self-organization.
On the other hand, while substrate-mediated cell-cell elastic interactions have been
considered for the organization of adherent cells in a variety of mechanobiological contexts
[19,20] (the physical basis of such modeling is reviewed in Ref. [21]), their effect on collective
cell motility, which in principle is always present, have not been carefully modeled. Here,
we present a simple biophysical agent–based model and computational results that focus
on how substrate mediate mechanical communication allows two cells to sense each other
and impacts their collective and relative motility. Our approach provides a foundation
for the study of more general cell interactions that include both mechanical and chemical
signalling, and also serves as a starting point for future studies of mechanical substrate
based interactions in multi-cellular systems such as growing tissue and confluent sheets.
2. Experimental observations motivate model for cell elastic interactions
Many eukaryotic cells use contractile localized forces generated by their actomyosin
cytoskeleton to adhere to and move on their substrates. Such traction forces typically cause
measurable deformations in the underlying substrates in cell culture experiment [5], and
have a spatially dipolar pattern [22]. A cell typically acts as a force dipole exerting – a pair
of equal and opposite forces – on the elastic medium. The dipolar pattern arises due to the
3 of 16
initial
final
M
F
contours
X
Z
Y
A
B
Cell A
Cell B
Elastic substrate
<latexit sha1_base64="YMvcfB29NQEDn9ZkuTDIR8ovUw4=">AB6XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqseiF49V7Ae0oWy2k3bpZhN
2N0IJ/QdePCji1X/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZLGLVCahGwSU2DTcCO4lCGgUC28H4dua3n1BpHstHM0nQj+hQ8pAzaqz0QFS/XHGr7hxklXg5qUCORr/81RvELI1QGiao1l3PTYyfUWU4Ezgt9VKNCWVjOsSupZJGqP1sfumUnFlQMJY2ZK
GzNXfExmNtJ5Ege2MqBnpZW8m/ud1UxNe+xmXSWpQsWiMBXExGT2NhlwhcyIiSWUKW5vJWxEFWXGhlOyIXjL6+S1kXVq1Xd+8tK/SaPowgncArn4MEV1OEOGtAEBiE8wyu8OWPnxXl3PhatBSefOY/cD5/ADRajSU=</latexit>r
Radial
distance
Cell B
Cell A
Figure 1. Schematic of the cell-cell mechanical interactions model: (A) Two cells A and B cultured
on the surface of thick elastic substrate can sense each other and interact at long range (when the
inter-cell distance r is longer than typical cell sizes, here depicted by dashed red circles) through
mechanical deformations of the underlying substrate. These deformations arise as the cells exert
traction stresses on the underlying elastic material. Here the cells are restricted to move on the surface
of the substrate. (B) We study with our computational model how a motile cell (M, Cell A, pink)
moves in the presence of a fixed central cell (Cell B, yellow). This two cell system on a substrate
(schematic shown as a top view) also mimics scenarios where a motile cell may encounter an elastic
impurity or obstacle on the medium. Shown as blue circles are contours of constant elastic potential
(in simplified form) that determine the inter-cell elastic force experienced by the motile cell B as a
result of the elastic deformations of the medium by both cells A and B. Also shown (in black) is a
representative simulated trajectory of the motile cell which starts outside the area of influence of the
stationary cell.
fact that no external forces are present on the system, and the cell, as a whole, moves on
its own accord. The net effect of these stresses is to contract or pull in the elastic material
comprising the substrate towards the cell.
In Ref. [3], it was shown that endothelial cells cultured on hydrogel substrates of
varying stiffness change their motile behavior in the presence of traction stresses exerted
on the substrate by neighbouring but non-contacting cells. In particular, it was shown that
pairs of cells on softer gels, showed reduced collective migration in comparison to isolated
cells. The number of contacts two cells made over specific periods of time by extending
their pseudopodia towards each other was also measured and found to depend sensitively
on substrate stiffness. Remarkably, the cells made stable contacts on very soft gels (Elastic
modulus, 500 Pa), whereas they made repeated contacts and withdrawals on substrates of
intermediate compliance (Elastic modulus, 2500 � 5500 Pa).
Motivated by this experiment, we here model the motility characteristics of a two
cell system and demonstrate how elastic deformations induced in the substrate allow
cells to respond to each other. We consider a pair of cells that each adhere to, and exert
stresses on the underlying substrate thereby deforming it as shown in Fig. 1. As mentioned
above, adherent and motile cells generate a contractile stress on the substrate. Here, the
contractility P of each cell, is minimally described by a physical model of force dipoles –a
pair of equal and opposite forces exerted on the substrate, and is thus a tensorial quantity
[19]. Such modeling is inspired by the theory of deformations induced by inclusions in
materials [23]. Unlike passive material inclusions, cells can actively regulate their force
production in response to external mechano-chemical cues from the substrate, including
the presence of other cells. Such complicating feedback effects in cell–cell interactions has
also been theoretically considered [24,25], but we ignore these for simplicity here, and we
treat P as an intrinsic cell property that is independent of underlying substrate matrix
strain and stiffness.
4 of 16
To simplify our study, we assume that one of the cells is motile (Cell A) and the other
is stationary (Cell B). The stationary cell B is nonetheless alive in that it still deforms the
substrate. The resulting deformation field, or equivalently the substrate mediated elastic
potential, is sensed by the other, distant, motile cell A. The interaction potential between
the cells in turn creates a mechanical force on the motile cell A. For polarized and elongated
cells, the deformations have a dipolar spatial pattern (described in Appendix A). However,
here we consider a simplified scenario that is valid when cells reorient very fast in the time
for them to translate and migrate (Appendix A, §3). This implies that the directions of
the dipole axis (of both cells A and B) fluctuates rapidly as cell A moves resulting in an
effectively isotropic, attractive interaction potential that decays with distance as ⇠ 1/r3
(iso–surfaces shown as blue circles in Fig. 1 B). Analysis of this model interaction provides
us insight into attractive potentials strongly influence cell motility.
The motile cell is considered to move diffusively with an effective diffusion coeffi-
cient, while also being acted upon by an elastic interaction force from the stationary cell.
Although, polarized cells may propel themselves persistently along their body axis, we
consider more isotropic cells here which extend their pseudopodia in different directions
randomly, and are thus described adequately by a diffusive process. Such a simple effective
Langevin equation is commonly used to describe elastically coupled motile active particles
[26] and swarming bacteria [27] but has not been studied previously in conjunction with
this specific type of interactions that arise on an elastic substrate.
We note that the model can be easily generalized (as derived in Appendix A) to
describe a pair of motile cells since the interactions are pairwise and reciprocal. The
interaction potential is not isotropic and depends on both the inter-cell distance as well
as on the instantaneous alignment of the cells’ dipole axes. Thus the force on each cell
(related to the gradient of the potential) depends on not just the relative positions of the
cells.but additionally on the direction of the contractile dipoles exerted by cells A and
B. Truly spherical dipoles embedded in an elastic medium do not interact mechanically
[23], unless cell-substrate feedback effects occur [25]. Furthermore, cell-cell interactions
in a fibrous, nonlinear elastic medium can be longer ranged [28] and have a power law
character, ⇠ 1/ra, where a < 3 [29]. The interaction of disk-like cells on top of a thick
substrate (semi-infinite geometry) is also more complicated [30]. We choose the isotropic,
attractive 1/r3 potential as the simplest attractive interaction with the same distance
dependence as the dipolar interaction, with the objective of testing how such a potential
can affect cell motility. Motivating future work, we show how the conclusions from the
simpler potential remain qualitatively valid even as specifics of cell trajectories change
when the more general dipolar potential is used. This model highlighted in this work,
although very simplified both in its description of cell contractility and motility, can thus
capture key aspects of motility and contact formation, as we now describe.
3. Materials and Methods
3.1. Model for two-cell interactions
The model used to analyze the two-cell system is an agent-based stochastic model.
We start with the stochastic Langevin equation for the dynamics of the moving cell A in
the presence of a second cell B fixed at the origin as illustrated in Figure 1(A). Details of
the model and the simplifications involved may be found in Appendix A. Starting from
the more general model where both cells A and B can move, we now fix cell B and thus
set rB = 0. In other words, we choose the center of cell B to be the origin from which the
position of cell A and its distance relative to B is measured. Writing r = rB � rA, we write
the equation for r(t) where t is the time,
dr
dt = �µT
∂W
∂r +
p
2Deff hhh(t)
(1)
5 of 16
where Deff is the effective translational diffusivity quantifying the random motion of the
moving cell in the absence of the fixed cell, and hhh is a random white noise term whose
components satisfy
hhi(t)hj(t0)i = d(t � t0)dij.
Note that h - the active noise term - has units of t�1/2. The mobility µT in equation (1)
quantifies the effective friction from the medium and is inversely proportional to the cell
size s and inversely proportional the the viscosity at the surface. Here it is assumed that
the cells moving on a wet surface and that the fluid nature of the surface provides a viscous
resistance opposing cell motion.
The two-cell potential W derives from the elastic interactions communicated via the
linear deformation of the substrate (Appendix A, Equation A5) and is given by,
W
=
1
2k(s � r)2, when 0 r s, and
(2)
=
� P2
E
f(n)
r3 ,
when r > s.
(3)
Numerical solutions to equation (1) are obtained with varying initial conditions for cell
A as explained subsequently. To ease the computational analysis, we work in scaled
dimensionless units. We choose cell size (diameter) s (see Fig 1), diffusion time s2/D0,
and thermal energy kBT – with T corresponding to the temperature of the cell/substrate
system – as our length, time and energy scales respectively. Equations (1-3) may then be
rewritten as
dr⇤
dt⇤ = �dW⇤
dr⇤ +
p
2DT hhh⇤
T,
(4)
where the potential in scaled form is
W⇤
=
1
2ksteric(1 � r⇤)2, when 0 r⇤ 1, and
(5)
=
� a
r⇤3 ,
when r⇤ > 1.
(6)
Superscripts ⇤ in equations (4)-(6) denote non-dimensional quantities. Henceforth, we will
drop this subscript for clarity. Thus the dynamics may be followed as a function of three
dimensionless numbers (parameters)
a ⌘
✓ P2f(n)
EkBTs3
◆
,
DT ⌘
✓ Deff
D0
◆
,
and ksteric ⌘
✓ ks2
kBT
◆
.
(7)
3.2. Dimensionless parameters quantifying cell motion and interactions
The parameters that emerge in equations (1)-(7) and typical of the two-cell scenario
studied here are summarized in Table 1. Following Ref. [3], we are interested in substrates
that are linearly elastic with the Young’s modulus E ranging from 0.5 kPa to 33 kPa, well
within the range of 0.1-100 kPa appropriate for tissues and bio-compatible materials [18].
The effective diffusion coefficients exhibited by cells in experiments [3] include the random
noisy motion as the cells explore territory and a contribution due to short-time deterministic
motion. We explore values in the range 3µm2/minute to 50 µm2/minute. Time scales
are estimated from experiments as well and 250 seconds in real time correspond to a
dimensionless time duration of unity.
Scaled non-dimensional parameters relevant to the simulation may be calculated
from dimensional quantities as explained earlier. Three scaled parameters determine the
dynamics of the two-cell system: DT, a and ksteric. Values used in the computations are
listed in Table 2. The self avoidance parameter ksteric is chosen such that the cells don’t
overlap and is computed based on the time step used in the simulations. This allows us to
control the stability of the simulation and its accuracy.
6 of 16
3.3. Numerical solution and tracking cell trajectories
Equations (4)-(7) are solved for the dynamics of the moving cell with appropriate
boundary and initial conditions. The Langevin equation (4) is an example of stochastic
differential equations; here we solve this equation using the explicit half-order Euler-
Maruyama method one of us has used recently in similar problems involving bacteria cells
moving in light fields [27] and in simulations of active Brownian particles [26].
Table 1. Biophysical parameters characterizing the two-cell (typical values from [3,31,32]).
Quantity
Interpretation
Experimental values
s
Cell size
10-100 µm
T
Temperature
250 C
D0
Thermal Diffusivity
25 µm2/min
Deff
Effective Diffusivity
3 � 50 µm2/min
E
Young’s modulus
0.5 � 33 kPa
n
Poisson ratio
0.3 - 0.5
P
Contractility
10�14 Nm
Table 2. Simulation parameters and their meaning.
Parameter
Interpretation
Definition
Simulation values
DT
Diffusivity
Deff/D0
0.1-10
a
Cell-cell interaction
P2f(n)/(EkBTs3)
0.1-100
ksteric
Self-avoidance
ks2/kBT
103 � 104
Given the position of cell A at time t, r(t), its subsequent location at time t + dt, r(t + dt),
follows,
r(t + dt) = r(t) �
✓∂W
∂r |r(t)
◆
dt +
p
2DTdt w,
(8)
where w is a random two-dimensional vector with components each drawn at every time
step from a normal distribution with mean zero and standard deviation of unity.
We simulated several trajectories of cell A ((n = 1000) trajectories, diameter s = 1 in
scaled units), under the influence of the central stationary cell B (also having diameter
s = 1). The simulations were conducted in two different geometries as described below.
To study the contact frequency between two-cells and explore the systematically
explore the role of the elastic potential, we simulated cell A moving in a confined square
box of size 12s with the stationary cell B at the center of the box. Cells reflect from the box
surface when they encounter it and thus are restricted to remain within the simulation
domain.
In order to calculate the number of contact in due course of the simulation, we define
a contact radius 1.5s from the centre of the stationary cell, and we consider a contact
if the centre of the test cell lies within the contact radius. The cell can come out of the
contact radius and re-enter, increasing the number of contacts. The time step used in these
simulations is dt = 0.0001 and total number of steps in this simulation is 107, i.e. a cell
trajectories were followed for a total time of T = 1000.
On the other hand for calculating cell dispersivities, and specifically the mean squared
displacement (MSD) of cell A, we used periodic boundary conditions and a periodic
potential. This corresponds to cell A moving in a periodic domain and interacting with
a regular square lattice of multiple stationary cells (images of B) separated uniformly by
7 of 16
a distance 12s. The time step used to integrate equation (7) in these simulations is also
dt = 0.0001 and total number of steps in this simulation is 107, i.e. a cell trajectories were
followed for a total time of T = 1000. The mean square displacement MSD was calculated
by tracking trajectories of cell A (the same as tracking n = 100 cells). As before, cell A is
initialized randomly inside the same square box of length of 12s, but outside the contact
radius. Cells that move out of the domain are reintroduced into the domain in a manner
that respects periodic boundary conditions and the appropriate symmetries.
In this case since r ⌘ xex + yey is the relative distance between the cells, the mean
square displacement is calculated by the equation,
MSD(t) = 1
n
n
Â
a=1
h[xa(tR + t) � xa(tR)]2 + [ya(tR + t) � ya(tR)]2i
(9)
where t is the delay time, and the summation is over each cell trajectory (indexed by a)
and extends over the full number of trajectories n = 100. The delay time is varied and the
averages are obtained by choosing different values of the reference time tR as is normally
done. The MSD given by equation (9) is thus an average over time and also an average
over realized cell trajectories.
-5
0
5
-6
-4
-2
0
2
4
6
Number of contacts
0
1000
2000
3000
4000
5000
6000
7000
8000
0.1
1
5
10
20
100
0.1
1
5
10
20
100
8000
7000
6000
5000
4000
3000
2000
1000
0
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRV
CyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGN
izCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
<latexit sha1_base64="7brGteJTxOG/ubQVZlufQYTATc=">AB8HicbVBNSwMxEJ
2tX7V+VT16CRbBU9kVUS9CUQ8eK/RL2qVk02wbmSXJCuUZX+Fw+KePXnePfmLZ70NYHA4/3ZpiZF8ScaeO6305hZXVtfaO4Wdra3tndK+8ftHSUKEKbJOKR6gRYU84kbRpmO3E
imIRcNoOxrdTv/1ElWaRbJhJTH2Bh5KFjGBjpce7ftrI0DXy+uWKW3VnQMvEy0kFctT75a/eICKJoNIQjrXuem5s/BQrwinWamXaBpjMsZD2rVUYkG1n84OztCJVQYojJQtadBM/T2
RYqH1RAS2U2Az0oveVPzP6yYmvPJTJuPEUEnmi8KEIxOh6fdowBQlhk8swUQxeysiI6wMTajkg3BW3x5mbTOqt5F1X04r9Ru8jiKcATHcAoeXEIN7qEOTSAg4Ble4c1Rzovz7nzMW
wtOPnMIf+B8/gBbw492</latexit>DT = 1
<latexit sha1_base64="Cue/mGefYH6DOZApKvlDIARi+RY=">AB8HicbVBNSwMxEJ
2tX7V+VT16CRbBU9ktUr0IRT14rNAvaZeSTbNtaJdkqxQlv4KLx4U8erP8ea/MW3oK0PBh7vzTAzL4g508Z1v53c2vrG5lZ+u7Czu7d/UDw8aukoUYQ2ScQj1QmwpxJ2jTMcNqJ
FcUi4LQdjG9nfvuJKs0i2TCTmPoCDyULGcHGSo93/bQxRdeo0i+W3LI7B1olXkZKkKHeL371BhFJBJWGcKx13Nj46dYGUY4nRZ6iaYxJmM8pF1LJRZU+n84Ck6s8oAhZGyJQ2aq78
nUiy0nojAdgpsRnrZm4n/ed3EhFd+ymScGCrJYlGYcGQiNPseDZixPCJZgoZm9FZIQVJsZmVLAheMsvr5JWpexVy+7DRal2k8WRhxM4hXPw4BJqcA91aAIBAc/wCm+Ocl6cd+dj0
Zpzsplj+APn8wdR493</latexit>DT = 2
<latexit sha1_base64="vTFNT+F7LjdOFlVcVpPFGjA+tTk=">AB8HicbVBNSwMxEJ
2tX7V+VT16CRbBU9kVq16Eoh48VuiXtEvJptk2NMkuSVYoS3+Fw+KePXnePfmLZ70NYHA4/3ZpiZF8ScaeO6305uZXVtfSO/Wdja3tndK+4fNHWUKEIbJOKRagdYU84kbRhmOG3H
imIRcNoKRrdTv/VElWaRrJtxTH2B5KFjGBjpce7XlqfoGtU6RVLbtmdAS0TLyMlyFDrFb+6/YgkgkpDONa647mx8VOsDCOcTgrdRNMYkxEe0I6lEguq/XR28ASdWKWPwkjZkgbN1N8
TKRZaj0VgOwU2Q73oTcX/vE5iwis/ZTJODJVkvihMODIRmn6P+kxRYvjYEkwUs7ciMsQKE2MzKtgQvMWXl0nzrOxdlN2H81L1JosjD0dwDKfgwSVU4R5q0ACAp7hFd4c5bw4787Hv
DXnZDOH8AfO5w9h0496</latexit>DT = 5
-5
0
5
-6
-4
-2
0
2
4
6
<latexit sha1_base64="sX6ceXyDkDrHKl56EUzFwVjyo0=">ACAHicdVDLSsNAFJ3UV62vqgsXbgaL4EJCEmtbBaGoC5cV+oImhMl02g6dPJiZCVk46+4caGIWz/DnX/jpK2gogcunDnXube40WMC
mkYH1puYXFpeSW/Wlhb39jcKm7vtEUYc0xaOGQh73pIEYD0pJUMtKNOEG+x0jHG19lfueOcEHDoCknEXF8NAzogGIkleQW92zEohGCF/D0GNrn127STNXDdIslQz+rVaxyBRq6YVRNy8yIVS2flKGplAwlMEfDLb7b/RDHPgkZkiInmlE0kQlxQzkhbsWJAI4TEakp6iAfKJcJLpASk8VEofDkKuKpBwqn6fSJAvxMT3VKeP5Ej89jLxL68Xy0HNSWgQxZIEePbRIGZQhjBLA/YpJ1iyiSIc6p2hXiEOMJSZVZQIXxdCv8
nbUs3K7pxWy7VL+dx5ME+OABHwARVUAc3oAFaAIMUPIAn8Kzda4/ai/Y6a81p85ld8APa2yd9YJRr</latexit>� = 5, DT = 1
-5
0
5
-6
-4
-2
0
2
4
6
=100 DT=5
1
1
4
4
<latexit sha1_base64="FXK5L+m+Qc3TiguW5FtJdbaOGCo=">ACAnicbVDLSgMxFM34rPU16krcBIvgQkpGfCEIRV24rNAXdIYhk2ba0ExmSDJCGYobf8WNC0Xc+hXu/BvTdhbaeuDCyTn3kntPkHCmN
ELf1tz8wuLScmGluLq2vrFpb203VJxKQusk5rFsBVhRzgSta6Y5bSWS4ijgtBn0b0Z+84FKxWJR04OEehHuChYygrWRfHvXxTzpYXgFHYSOoHt562e1oXme+nYJldEYcJY4OSmBHFXf/nI7MUkjKjThWKm2gxLtZVhqRjgdFt1U0QSTPu7StqECR1R52fiEITwSgeGsTQlNByrvycyHCk1iALTGWHdU9PeSPzPa6c6vPAyJpJU0EmH4UphzqGozxgh0lKNB8YgolkZldIelhiok1qROCM3yLGkcl52zMro/KVWu8zgKYA/
sg0PgHNQAXegCuqAgEfwDF7Bm/VkvVjv1sekdc7KZ3bAH1ifPwQ8lJ0=</latexit>� = 100, DT = 5
3
3
2
2
<latexit sha1_base64="5JYanCdHzdygT4Y909yu7tGkuIg=">ACAXicdVDLSgMxFM3UV62vUTeCm2ARXEjJ1KGtglDUhcsKfUE7DJk0bUMzD5KMUIa68VfcuFDErX/hzr8x01ZQ0QMXTs65l9x7vIgzq
RD6MDILi0vLK9nV3Nr6xuaWub3TlGEsCG2QkIei7WFJOQtoQzHFaTsSFPsepy1vdJn6rVsqJAuDuhpH1PHxIGB9RrDSkmvudTGPhieQwsdw+7ZlZvUJ+nLNfOocFopFe0SRAWEylbRSkmxbJ/Y0NJKijyYo+a791eSGKfBopwLGXHQpFyEiwUI5xOct1Y0giTER7QjqYB9ql0kukFE3iolR7sh0JXoOBU/T6RYF/Kse/pTh+rofztpeJfXidW/YqTsCKFQ3I7KN+zKEKYRoH7DFBieJjTARTO8KyRALTJQOLadD+LoU/k+
axYJVKqAbO1+9mMeRBfvgABwBC5RBFVyDGmgAu7A3gCz8a98Wi8GK+z1owxn9kFP2C8fQLqi5Sh</latexit>� = 10, DT = 1
<latexit sha1_base64="VqCrxFTWefFUQRo7dsSW1orLZI=">ACAXicdVDLSgMxFM3UV62vqhvBTbAILmTIjENbBaGoC5cV+oK2lEyaUMzD5KMUIa68VfcuFDErX/hzr8x01ZQ0QMXTs65l9x73Igzq
RD6MDILi0vLK9nV3Nr6xuZWfnunIcNYEFonIQ9Fy8WSchbQumK01YkKPZdTpvu6DL1m7dUSBYGNTWOaNfHg4B5jGClpV5+r4N5NMTwHNroGHbOrnpJbaJfVi9fQOZpuWg7RYhMhEqWbaXELjknDrS0kqIA5qj28u+dfkhinwaKcCxl20KR6iZYKEY4neQ6saQRJiM8oG1NA+xT2U2mF0zgoVb60AuFrkDBqfp9IsG+lGPf1Z0+VkP520vFv7x2rLxyN2FBFCsakNlHXsyhCmEaB+wzQYniY0wEUzvCskQC0yUDi2nQ/i6FP5
PGrZpFU104xQqF/M4smAfHIAjYIESqIBrUAV1QMAdeABP4Nm4Nx6NF+N1pox5jO74AeMt0/sHZSi</latexit>� = 20, DT = 1
A
Figure 2. The number of cell–cell contact events measured in a fixed interval of time depends
strongly on the elastic interaction parameter. A contact event is identified as cell A coming within
a prescribed contact radius of cell B with cell A initialized randomly in a certain area around cell
B. Thus the number of contact is be interpreted as the average number of contacts of the two cells.
The number of simulation runs conducted were 50 for each combination of DT and a. The dashed
curves are guides to the eye illustrating the trends seen with increasing values of a. Diffusion is the
major factor in governing the number of contacts for low values of a. For higher a, the attractive
potential increases the probability of the cell to stay near the contact radius and controls the number
of contacts. Trajectories for highlighted data points (1)-(4) are shown on the right. The box plots
show the distribution of contact numbers. The lower and upper bounds of the box are the first and
the third quartiles respectively, while the line in middle is the median. The lower and upper limits
of the dashed lines are the minimum and maximum number of contacts observed for cells for each
combination of a and DT. The simulation was run for a total time of T = 1000 and updates in cell
position were made every dt = 0.001.
The mobility of cell A reflects the properties of the microenvironment created by cell
B and by the substrate. The mean square displacement in (9) is written as a function of the
delay time t that may be interpreted as an effective observation time over which the cell
motion is observed. For instance, a cell that moves with constant speed for small times
(say ⇠ T1) and undergoes a diffusive random walk when observed over long times (say ⇠
8 of 16
T2) will exhibit different slopes for t < T1 and for t > T2. The exponent characterizing the
dependence of the MSD on the delay time provides information as to whether the motion is
sub-diffusive (exponent < 1), diffusive (exponent = 1), or super-diffusive (exponent > 1).
It is constructive to study the expected MSD for cell A in the absence of cell B. In
this particular case, since A is purely diffusive, the MSD has the simple form valid for
diffusion in two dimensions MSD(t) = 4DTt. Deviations from this expression arise due to
the mechanically induced inter-cell interaction and thus quantify the extent to which cell B
perturbs the dispersion of cell A. For instance transient or persistent trapping of cell A will
result in the MSD scaling sub-linearly with t.
0
1000
2000
3000
4000
5000
6000
7000
8000
0.5
0.2
0.1
1
2
5
10
Number of contacts
0.1
0.2
0.5
1.0
2.0
10.0
8000
7000
6000
5000
4000
3000
2000
1000
0
Effective translational diffusivity
5.0
1
2
3
4
-5
0
5
-6
-4
-2
0
2
4
6
=10 DT=0.1
2
<latexit sha1_base64="a1ohGl7mT+YWUTrN1SID4wGcAfA=">ACBHicdVDLSgMxFM3UV62vUXe6CRbBhZSkFNsKQlEXLiu0tAZSiZN29DMgyQjlKHgxl9x40JB3PoR7vwbM20FT1w4eSce8m9x4sEV
xqhDyuzsLi0vJdza2tb2xu2ds7NyqMJWVNGopQtj2imOABa2quBWtHkhHfE6zljS5Sv3XLpOJh0NDjiLk+GQS8zynRuraew4R0ZDAM4jRMXROL7tJY2JeqIC7dh4VEIY5gSXD5BhlSrlSKuQJxaBnkwR71rvzu9kMY+CzQVRKkORpF2EyI1p4JNck6sWEToiAxYx9CA+Ey5yfSGCTw0Sg/2Q2kq0HCqfp9IiK/U2PdMp0/0UP32UvEvrxPrfsVNeBDFmgV09lE/FlCHMA0E9rhkVIuxIYRKbnaFdEgkodrEljMhfF0K/ye
tYgGXChfl/K183keWbAPDsARwKAMauAK1ETUHAHsATeLburUfrxXqdtWas+cwu+AHr7RNZlJU7</latexit>� = 10, DT = 0.1
-5
0
5
-6
-4
-2
0
2
4
6
=1 DT=0.1
1
3
<latexit sha1_base64="RAZ8AY3g78fqOrdBOgY30DI3NCc=">ACAXicbVDLSsNAFL3xWesr6kZwM1gEF1KSIiqCUNSFywp9QRPCZDph04ezEyEurGX3HjQhG3/oU7/8ZJ24W2Hrhw5px7mXuPn
3AmlWV9GwuLS8srq4W14vrG5ta2ubPblHEqCG2QmMei7WNJOYtoQzHFaTsRFIc+py1/cJP7rQcqJIujuhom1A1xL2IBI1hpyTP3HcyTPkZXqGKdIOfy1svqo/zlmSWrbI2B5ok9JSWYouaZX043JmlI0U4lrJjW4lyMywUI5yOik4qaYLJAPdoR9MIh1S62fiCETrShcFsdAVKTRWf09kOJRyGPq6M8SqL2e9XPzP6QquHAzFiWpohGZfBSkHKkY5XGgLhOUKD7UBPB9K6I9LHAROnQijoEe/bkedKslO2zsn
V/WqpeT+MowAEcwjHYcA5VuIMaNIDAIzDK7wZT8aL8W58TFoXjOnMHvyB8fkDjZKUYQ=</latexit>� = 20, DT = 2
-5
0
5
-6
-4
-2
0
2
4
6
=20 DT=10
4
<latexit sha1_base64="8NixRt3VOpY3OoM9bE3TxbwgoX8=">ACAnicbVDLSgMxFM34rPU16krcBIvgQkqmiIogFHXhskJf0BmGTJpQzOZIckIZShu/BU3LhRx61e4829M21lo64ELJ
+fcS+49QcKZ0gh9WwuLS8srq4W14vrG5ta2vbPbVHEqCW2QmMeyHWBFORO0oZnmtJ1IiqOA01YwuBn7rQcqFYtFXQ8T6kW4J1jICNZG8u19F/Okj+EVrKAT6F7e+l9ZF4O8u0SKqMJ4DxclICOWq+/eV2Y5JGVGjCsVIdByXay7DUjHA6KrqpogkmA9yjHUMFjqjyskJI3hklC4MY2lKaDhRf09kOFJqGAWmM8K6r2a9sfif10l1eOFlTCSpoJMPwpTDnUMx3nALpOUaD40BPJzK6Q9LH
ERJvUiYEZ/bkedKslJ2zMro/LVWv8zgK4AcgmPgHNQBXegBhqAgEfwDF7Bm/VkvVjv1se0dcHKZ/bAH1ifPwDOlJo=</latexit>� = 20, DT = 10
<latexit sha1_base64="yfmWjMnUvNtJIcrJd1h/ivdYe8=">AB8X
icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1ItQ9OKxgv3ANpTJdtMu3WzC7kYof/CiwdFvPpvPlv3LY5aOuDgcd7M8zMCxLBtXHdb6ewsrq2vlHcLG1t
7+zulfcPmjpOFWUNGotYtQPUTHDJGoYbwdqJYhgFgrWC0e3Ubz0xpXksH8w4YX6EA8lDTtFY6bGLIhkiuSZer1xq+4MZJl4OalAjnqv/NXtxzSNmD
RUoNYdz02Mn6EynAo2KXVTzRKkIxywjqUSI6b9bHbxhJxYpU/CWNmShszU3xMZRlqPo8B2RmiGetGbiv95ndSEV37GZIaJul8UZgKYmIyfZ/0uWLU
iLElSBW3txI6RIXU2JBKNgRv8eVl0jyrehdV9/68UrvJ4yjCERzDKXhwCTW4gzo0gIKEZ3iFN0c7L8678zFvLTj5zCH8gfP5AzNDj/M=</latexit>� = 1
<latexit sha1_base64="v+50geBnlGfYijhr/Qpw8AHi0KI=">AB8ni
cbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1ItQ9OKxgrWFNJTJdtMu3WzC7kYoT/DiwdFvPprvPlv3LY5aOuDgcd7M8zMC1PBtXHdb6e0srq2vlHerGxt7+zuV
fcPHnWSKcpaNBGJ6oSomeCStQw3gnVSxTAOBWuHo9up35iSvNEPphxyoIYB5JHnKxkt9FkQ6RXBP7Vrbt2dgSwTryA1KNDsVb+6/YRmMZOGCtTa9z
UBDkqw6lgk0o30yxFOsIB8y2VGDMd5LOTJ+TEKn0SJcqWNGSm/p7IMdZ6HIe2M0Yz1IveVPzP8zMTXQU5l2lmKTzRVEmiEnI9H/S54pRI8aWIFXc3kroE
BVSY1Oq2BC8xZeXyeNZ3buou/fntcZNEUcZjuAYTsGDS2jAHTShBRQSeIZXeHOM8+K8Ox/z1pJTzBzCHzifP6QJkC0=</latexit>� = 10
<latexit sha1_base64="MgogXpwrP2/PI/J64WxN78zV9j0=">AB8n
icbVBNS8NAEJ3Ur1q/qh69LBbBU0mKqBeh6MVjBfsBaSib7aZdutmE3YlQSn+GFw+KePXePfuG1z0NYHA4/3ZpiZF6ZSGHTdb6ewtr6xuVXcLu3s
7u0flA+PWibJNONlshEd0JquBSKN1Gg5J1UcxqHkrfD0d3Mbz9xbUSiHnGc8iCmAyUiwShaye9SmQ4puSE1t1euFV3DrJKvJxUIEejV/7q9hOWxV
whk9QY3NTDCZUo2CST0vdzPCUshEdcN9SRWNugsn85Ck5s0qfRIm2pZDM1d8TExobM45D2xlTHJplbyb+5/kZRtfBRKg0Q67YlGUSYIJmf1P+kJz
hnJsCWVa2FsJG1JNGdqUSjYEb/nlVdKqVb3LqvtwUanf5nEU4QRO4Rw8uI63EMDmsAgWd4hTcHnRfn3flYtBacfOY/sD5/AGljpAu</latexit>� = 20
B
<latexit sha1_base64="AE4U7J6tq1m9/R7/tgBxel2cIUI=">ACBHicdVDLSsNAFJ3UV62vqMtuBovgQkJSQ1sLQlEXLiu0tdCEMJlO26GTBzMToYQu3Pgrblwo4taPcOfOGkrq
OiBgXPuZc79/gxo0Ka5oeW1peWV3Lrxc2Nre2d/TdvY6IEo5JG0cs4l0fCcJoSNqSka6MSco8Bm58cXmX9zS7igUdiSk5i4ARqGdEAxkry9KDWDxC8Axax9CpO/VL21NVWkalqeXTO0VinbFVWaZtUqWxkpV+0TG1pKyVACzQ9/d3pRzgJSCgxQ0L0LDOWboq4pJiRacFJBIkRHqMh6SkaoAIN50dMYWHSunDQcTVCyWcqd8nUhQIMQl81RkgORK/vUz8y+slclBzUx
rGiSQhni8aJAzKCGaJwD7lBEs2UQRhTtVfIR4hjrBUuRVUCF+Xwv9Jp2xYFcO8tkuN80UceVAEB+AIWKAKGuAKNEbYHAHsATeNbutUftRXudt+a0xcw+AHt7ROZzpWD</latexit>� = 1, DT = 0.1
Figure 3. The number of cell–cell contact events in a fixed interval of time (T = 1000) plotted here as
a function of the scaled effective diffusivity, DT, which represents the random motility of cell B. Here
we show how the number of cell–cell contact varies for three different elastic interaction strength
values, a, corresponding to substrates with three different stiffness. The highlighted points numbered
from (1)-(4), show representative cell trajectories over long times and highlight how varying a and
DT can yield states where the cells are in close proximity most of the time (low DT, high a) or states
where cells interact rarely (high DT, low a). Interpretation of the box plots is the same as in Figure 2.
The simulation was run for a total time of T = 1000 and updates in cell position were made every
dt = 0.001.
4. Results
4.1. Cell-cell contact frequency shows biphasic dependence on matrix elastic interactions
Motivated by experiments which show that two cells make repeated contact and
withdrawals on soft substrates, with contact frequency dependent on the substrate stiffness,
we measure the total number of contacts of the motile cell (A) with the stationary cell (B) in
our model simulations. As indicated earlier, the simulated cells are initialized randomly
inside the box, but outside of a pre-defined contact radius around the stationary cell. The
total number of contacts between the cells is counted over a fixed period of time i.e. T =
1000. It should be remembered that the cells are confined to stay within the square domain
during the course of the simulation.
Cell A’s movement is governed by an attractive elastic potential induced by the
stationary, central cell and its own random motion, described as an effective diffusion.
Additionally when the cell encounters the bounding wall of the square domain, it reflects
(moves away) from it. Overall, random noise encapsulated in the diffusion coefficient
causes A to move towards or away from B in an unbiased manner. The attractive potential
W being isotropic and spatially varying suggests that there is a critical radius of influence
(dependent on both a and DT) within which forces due to the attractive potential dominate
diffusion and significantly influence the trajectory of cell A. This effect results in the cell
getting closer to cell B, eventually entering this zone of influence.
9 of 16
To carefully study how elastic interactions (a) and random diffusion (DT) each influ-
ence this process, we first systematically calculated the number of contacts by a, while
keeping DT constant at three different values, DT = 1,2,5. (Figure 2). As illustrated by the
dotted lines which serve as a guide to the eye, the behavior is highly non-monotonic. For
small a, the number of contacts increases with increasing a, then reduces to 1 at high a.
The position of the peak increases with increasing DT. The initial increase in contacts is
due to the increased directional movement of the test cells towards the central cell. The
decrease in the number of contacts for very high values of a is expected since the attractive
potential is strong enough to overcome the effect of diffusion. In this case, the motile
cell is unable to move away from and makes stable contact with the stationary cell. For
a = 5 and DT = 1 (trajectory 1), the test cell spends most of the time exploring space rather
than near the stationary cell, which also reduces the number of contacts. Increasing a to
10 (trajectory 2) the radius of influence increases, increasing the duration of contact and
thereby increasing contacts. On further increasing a to 20 (trajectory 3), the test cell is
tightly adhered to the stationary cell which allows only one single contact. Note that the
statistics for the high DT and low a regime are influenced by the confinement. Cells in this
particular limit frequently escape the region of influence and wander away only to return
again after encountering the wall and diffusing away. For instance, the number of contacts
for DT = 5 and a = 0.1, combines the effect of repeated escapes from the region of influence
and repeated returns due to confinement. Since the size of the box is fixed, the increase in
number of contacts with DT for a = 0.1 is still a signature of diffusive effects dominating
the attractive potential.
We next investigated the effect of increasing diffusivity on the number of contacts
for constant a (1, 10 and 20). Results from this set of simulations are shown in Figure
3. The red dotted line serves as a guide to the eye highlighting the trend observed. We
see a steady increase in cell-cell contacts with diffusivity. Without diffusion, the test cell
shows unidirectional motion towards the central cell and remains in contact throughout
the simulation. Increasing diffusion increases the chance of test cell to go out of the radius
of influence and come back again (trajectories 3 and 4).
Overall combining the results shown in Figures 2 and 3, we conclude that the number
of contacts is maximized at an optimal value of the elastic interaction strength. If the elastic
strength is too high or too low, the cell either makes stable contact or is too motile to make
too many contacts. This optimal value scales with the diffusivity, which is a measure of the
cell motility in our model.
4.2. Cell motility characteristics depend on elastic interactions
To quantify the long-time statistics of the motility of cell A in the elastic potential
field generated by cell B, we analyze the mean squared displacement (MSD) as given by
equation (9) from simulation. The metric MSD measured in terms of a delay time t contains
information about the short time mobility of a cell, the long time mobility of the cell, and
additionally provides signatures of capture and trapping effects. Specifically, the slope
of the mean square displacement can be used to extract effective exponents that provides
insight on the relative importance of diffusion and elastic attractive interactions.
We plot the MSD in Fig. 4 for DT = 2 and a = 0.1,1,5,10,20,100. For a = 0.1,1,5,10,
we find that the slope is close to 1, which suggests diffusion drives the motion of the cell
and the attractive potential is not strong enough to influence the movement of the cell.
For higher a, we observe a transition towards sub-diffusive behavior at t ⇠ 0.5. At a = 20
(green line), the curve shows a significant decrease in slope at t = 2, the time scale for
which a test cell in average encounters the central cell for the first time and stays in contact
for a while, as shown by trajectory 3, Figure 3. The slope then increases again, but remains
less than 1 suggesting a sub-diffusive behavior in the long run. At a = 100 (blue line),
the MSD saturates after initial diffusion to a zero slope which suggests that the motion is
bounded, and it can only explore the circumference of the stationary cell.
10 of 16
10-2
10-1
100
101
102
10-2
100
102
104
MSD
=0.1
=1
=5
=10
=20
=100
69
70
71
72
73
74
540
560
580
MSD
=0.1
=1
=5
<latexit sha1_base64="zjtCf8Hr8dE1aetXG1h2l1nUV0k=">AB+HicbVDLSgNBEOz1GeMjqx69DAbBU9gVUS9C0IvHCOYByRJ6J7PJkNnZWZWiEu+xIsHRbz6Kd7
8GyePgyYWdFNUdTM9FaCa+N5387K6tr6xmZhq7i9s7tXcvcPGjrJFGV1mohEtULUTHDJ6oYbwVqpYhiHgjXD4e3Ebz4ypXkiH8woZUGMfckjTtFYqeuWOijSARJyTXzPq3hdt2z7FGSZ+HNShjlqXfer0toFjNpqECt276XmiBHZTgVbFzsZJqlSIfYZ21LJcZMB/n08DE5sUqPRImyJQ2Zqr83coy1HsWhnYzRDPSiNxH/89qZia6CnMs0M0zS2UN
RJohJyCQF0uOKUSNGliBV3N5K6AVUmOzKtoQ/MUvL5PGWcW/qHj35+XqzTyOAhzBMZyCD5dQhTuoQR0oZPAMr/DmPDkvzrvzMRtdceY7h/AHzucPxTCRNA=</latexit>� = 100.0
<latexit sha1_base64="bR3mqEHla/OljzIiyR3OipFU52U=">AB9XicbVBNSwMxEJ2tX7V+VT16CRbBU9ktol6EohePFewHtGuZTdM2NJtdkqxSlv4PLx4U8ep/8ea
/MW3oK0PZni8N0MmL4gF18Z1v53cyura+kZ+s7C1vbO7V9w/aOgoUZTVaSQi1QpQM8ElqxtuBGvFimEYCNYMRjdTv/nIlOaRvDfjmPkhDiTvc4rGSg8dFPEQCbkiFbfsdosl2cgy8TLSAky1LrFr04voknIpKECtW57bmz8FJXhVLBJoZNoFiMd4YC1LZUYMu2ns6sn5MQqPdKPlC1pyEz9vZFiqPU4DOxkiGaoF72p+J/XTkz/0k+5jBPDJ0/1E8
EMRGZRkB6XDFqxNgSpIrbWwkdokJqbFAFG4K3+OVl0qiUvfOye3dWql5nceThCI7hFDy4gCrcQg3qQEHBM7zCm/PkvDjvzsd8NOdkO4fwB87nD91kMo=</latexit>� = 20.0
<latexit sha1_base64="KOqRLIldbIVDcYZdyZUNX8U206c=">AB9XicbVBNSwMxEJ31s9avqkcvwSJ4Krsi6kUoevFYwX5Au5bZNuGZrNLklXK0v/hxYMiXv0v3vw
3pu0etPXBDI/3ZsjkBYng2rjut7O0vLK6tl7YKG5ube/slvb2GzpOFWV1GotYtQLUTHDJ6oYbwVqJYhgFgjWD4c3Ebz4ypXks780oYX6EfclDTtFY6aGDIhkgIVfEcytut1S2fQqySLyclCFHrVv6vRimkZMGipQ67bnJsbPUBlOBRsXO6lmCdIh9lnbUokR0342vXpMjq3SI2GsbElDpurvjQwjrUdRYCcjNAM9703E/7x2asJLP+MySQ2TdPZQmAp
iYjKJgPS4YtSIkSVIFbe3EjpAhdTYoIo2BG/+y4ukcVrxzivu3Vm5ep3HUYBDOIT8OACqnALNagDBQXP8ApvzpPz4rw7H7PRJSfOYA/cD5/ANvukMk=</latexit>� = 10.0
<latexit sha1_base64="ncuhdQ/iZoQBY6/XDnj01Mzs/tg=">AB9HicbVDJSgNBEK2JW4xb1KOXxiB4GmbE7SIEvXiMYBZIhlDT6Uma9Cx29wTCkO/w4kERr36MN/
GTjIHTXxQ8Hiviqp6fiK40o7zbRVWVtfWN4qbpa3tnd298v5BQ8WpKxOYxHLlo+KCR6xuZasFYiGYa+YE1/eDf1myMmFY+jRz1OmBdiP+IBp6iN5HVQJAMk5IZc2E63XHFsZwayTNycVCBHrVv+6vRimoYs0lSgUm3XSbSXodScCjYpdVLFEqRD7LO2oRGTHnZ7OgJOTFKjwSxNBVpMlN/T2QYKjUOfdMZoh6oRW8q/ue1Ux1cexmPklSziM4XBak
gOibTBEiPS0a1GBuCVHJzK6EDlEi1yalkQnAX14mjTPbvbSdh/NK9TaPowhHcAyn4MIVOEealAHCk/wDK/wZo2sF+vd+pi3Fqx85hD+wPr8AXC6kJM=</latexit>� = 5.0
<latexit sha1_base64="0IyVoHMjXwAhpJhBT8qlQuC9CJk=">AB9HicbVBNS8NAEJ34WetX1aOXxSJ4ComIehGKXjxWsB/QhjLZbtqlm03c3RK6e/w4kERr/4Yb/4
bt20O2vpg4PHeDPzwlRwbTzv21lZXVvf2CxsFbd3dvf2SweHdZ1kirIaTUSimiFqJrhkNcONYM1UMYxDwRrh4G7qN4ZMaZ7IRzNKWRBjT/KIUzRWCto0j4SckN81+uUyp7rzUCWiZ+TMuSodkpf7W5Cs5hJQwVq3fK91ARjVIZTwSbFdqZinSAPdayVGLMdDCeHT0hp1bpkihRtqQhM/X3xBhjrUdxaDtjNH296E3F/7xWZqLrYMxlmhkm6XxRlAl
iEjJNgHS5YtSIkSVIFbe3EtpHhdTYnIo2BH/x5WVSP3f9S9d7uChXbvM4CnAMJ3AGPlxBe6hCjWg8ATP8ApvztB5cd6dj3nripPHMEfOJ8/aqKQjw=</latexit>� = 1.0
<latexit sha1_base64="UipcCNiwk3tyJxC6qK4wvY+qE=">AB9HicbVBNS8NAEJ34WetX1aOXxSJ4ComIehGKXjxWsB/QhjLZbtqlm03c3RK6e/w4kERr/4Yb/4
bt20O2vpg4PHeDPzwlRwbTzv21lZXVvf2CxsFbd3dvf2SweHdZ1kirIaTUSimiFqJrhkNcONYM1UMYxDwRrh4G7qN4ZMaZ7IRzNKWRBjT/KIUzRWCto0j4SckM81+Uyp7rzUCWiZ+TMuSodkpf7W5Cs5hJQwVq3fK91ARjVIZTwSbFdqZinSAPdayVGLMdDCeHT0hp1bpkihRtqQhM/X3xBhjrUdxaDtjNH296E3F/7xWZqLrYMxlmhkm6XxRlAl
iEjJNgHS5YtSIkSVIFbe3EtpHhdTYnIo2BH/x5WVSP3f9S9d7uChXbvM4CnAMJ3AGPlxBe6hCjWg8ATP8ApvztB5cd6dj3nripPHMEfOJ8/aqCQjw=</latexit>� = 0.1
<latexit sha1_base64="ncuhdQ/iZoQBY6/XDnj01Mzs/tg=">AB9HicbVDJSgNBEK2JW4xb1KOXxiB4GmbE7SIEvXiMYBZIhlDT6Uma9Cx29wTCkO/w4kERr36MN/
GTjIHTXxQ8Hiviqp6fiK40o7zbRVWVtfWN4qbpa3tnd298v5BQ8WpKxOYxHLlo+KCR6xuZasFYiGYa+YE1/eDf1myMmFY+jRz1OmBdiP+IBp6iN5HVQJAMk5IZc2E63XHFsZwayTNycVCBHrVv+6vRimoYs0lSgUm3XSbSXodScCjYpdVLFEqRD7LO2oRGTHnZ7OgJOTFKjwSxNBVpMlN/T2QYKjUOfdMZoh6oRW8q/ue1Ux1cexmPklSziM4XBak
gOibTBEiPS0a1GBuCVHJzK6EDlEi1yalkQnAX14mjTPbvbSdh/NK9TaPowhHcAyn4MIVOEealAHCk/wDK/wZo2sF+vd+pi3Fqx85hD+wPr8AXC6kJM=</latexit>� = 5.0
<latexit sha1_base64="0IyVoHMjXwAhpJhBT8qlQuC9CJk=">AB9HicbVBNS8NAEJ34WetX1aOXxSJ4ComIehGKXjxWsB/QhjLZbtqlm03c3RK6e/w4kERr/4Yb/4
bt20O2vpg4PHeDPzwlRwbTzv21lZXVvf2CxsFbd3dvf2SweHdZ1kirIaTUSimiFqJrhkNcONYM1UMYxDwRrh4G7qN4ZMaZ7IRzNKWRBjT/KIUzRWCto0j4SckN81+uUyp7rzUCWiZ+TMuSodkpf7W5Cs5hJQwVq3fK91ARjVIZTwSbFdqZinSAPdayVGLMdDCeHT0hp1bpkihRtqQhM/X3xBhjrUdxaDtjNH296E3F/7xWZqLrYMxlmhkm6XxRlAl
iEjJNgHS5YtSIkSVIFbe3EtpHhdTYnIo2BH/x5WVSP3f9S9d7uChXbvM4CnAMJ3AGPlxBe6hCjWg8ATP8ApvztB5cd6dj3nripPHMEfOJ8/aqKQjw=</latexit>� = 1.0
<latexit sha1_base64="UipcCNiwk3tyJxC6qK4wvY+qE=">AB9HicbVBNS8NAEJ34WetX1aOXxSJ4ComIehGKXjxWsB/QhjLZbtqlm03c3RK6e/w4kERr/4Yb/4
bt20O2vpg4PHeDPzwlRwbTzv21lZXVvf2CxsFbd3dvf2SweHdZ1kirIaTUSimiFqJrhkNcONYM1UMYxDwRrh4G7qN4ZMaZ7IRzNKWRBjT/KIUzRWCto0j4SckM81+Uyp7rzUCWiZ+TMuSodkpf7W5Cs5hJQwVq3fK91ARjVIZTwSbFdqZinSAPdayVGLMdDCeHT0hp1bpkihRtqQhM/X3xBhjrUdxaDtjNH296E3F/7xWZqLrYMxlmhkm6XxRlAl
iEjJNgHS5YtSIkSVIFbe3EtpHhdTYnIo2BH/x5WVSP3f9S9d7uChXbvM4CnAMJ3AGPlxBe6hCjWg8ATP8ApvztB5cd6dj3nripPHMEfOJ8/aqCQjw=</latexit>� = 0.1
Mean square displacement,
Delay time,
<latexit sha1_base64="yGbOJNxjeEKB5Z2Sb3L8DaMe/10=">AB/XicbVDLSgMxFM
3UV62v8bFzEyxC3ZQZEXVZ1IUboaJ9QGcomTRtQ5PMkGSEOgz+ihsXirj1P9z5N2baWjrgcDhnHu5JyeIGFXacb6twsLi0vJKcbW0tr6xuWVv7zRVGEtMGjhkoWwHSBFGBWloqhlpR5
IgHjDSCkaXmd96IFLRUNzrcUR8jgaC9ilG2khdey/xONJDyZObu6s0rXgaxUdu+xUnQngPHFzUgY56l37y+uFOZEaMyQUh3XibSfIKkpZiQtebEiEcIjNCAdQwXiRPnJH0KD43Sg/
1Qmic0nKi/NxLElRrzwExmUdWsl4n/eZ1Y98/9hIo1kTg6aF+zKAOYVYF7FJsGZjQxCW1GSFeIgkwtoUVjIluLNfnifN46p7WnVuT8q1i7yOItgHB6ACXHAGauAa1EDYPAInsEreL
OerBfr3fqYjhasfGcX/IH1+QNc9JUo</latexit>MSD(�)
<latexit sha1_base64="lCedxjNk/DYMPs6qZXTtQUTGNXI=">AB63icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPRi8cK9gPaUDbTbt0swm7E6GE/gUvHhTx6h/y5r9
x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wSTu9zvPHFtRKwecZpwP6IjJULBKOZSH2k6qNbcujsHWSVeQWpQoDmofvWHMUsjrpBJakzPcxP0M6pRMlnlX5qeELZhI54z1JFI278bH7rjJxZUjCWNtSObq74mMRsZMo8B2RhTHZtnLxf+8XorhjZ8JlaTIFVsClNJMCb542QoNGc
op5ZQpoW9lbAx1ZShjadiQ/CWX14l7Yu6d1V3Hy5rjdsijKcwCmcgwfX0IB7aEILGIzhGV7hzYmcF+fd+Vi0lpxi5hj+wPn8ASLTjk0=</latexit>�
Figure 4. Mean square displacement (MSD) as a function of the delay time interval t (calculated
from Equation 9), for the motile cell A is shown. Here we explore the variation in the MSD for
various values of substrate-mediated elastic interactions, a. The diffusivity DT is held constant for
these simulations with DT = 2. Other diffusivities were explored (results not shown). At low elastic
interaction strengths, a, corresponding to stiff substrates, the cell shows a purely diffusive trajectory,
whereas at higher values of a, the motile cell is captured by the strong attractive interaction from
the stationary cell, resulting in a flattening of the MSD (blue curve). At an intermediate interaction
regime (green curve), the motile cell makes repeated contact with the fixed cell but is never fully
captured.
4.3. Elastic interactions lead to effective capture of motile cell
Taken together, our simulations suggest that strongly attractive elastic interactions
can lead to stable contact between initially distant cells. We next explore the statistics of
this “capture” process. Capture mechanisms underlying and influencing these statistics
are potentially relevant for timescales of contact formation between initially well-separated
motile cells that then form confluent monolayers, such as in mesenchymal–to–epithelial
transitions during tissue morphogenesis [33].
Figures 2 and 3 suggest that the motile cell A (as it explores space and samples
the potential field over its various trajectories) is attracted to the stationary cell with the
attracting force increasing with decreasing distance r. Acting in tandem and superposed on
this aspect of the motion is diffusion that allows A to wander away from B multiple times.
In order to understand how parameters a and DT affect this phenomenon, we tracked
the number of cells inside the contact radius over the course of the simulation. The
probability of cells inside the contact radius reached a steady state at time t < 100 for all
parameters (Figure 5A). Keeping a constant and increasing DT the probability of cells being
inside the contact radius decreases (Figure 5B). The steady-state probability PSS increases
with increase in a for constant DT (Figure 5C). To understand the relationship between PSS
and both a and DT, we investigated PSS for the ratio a/DT and showed that they remain
constant for this ratio.
Plotting PSS vs a/DT, the strength of the elastic interactions relative to the diffusivity,
we find that the data can be collapsed into a single master curve (Figure 5D). The collapse of
our data and the master curve plotted in Figure 5D is expected since our model steady state
is a thermal equilibrium with effective temperature set by the value of DT; the competition
between attractive interactions and noise meanwhile dictates how many cells are captured
vs. how many can escape.
11 of 16
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9AuaUCbt0swm7G6GE/g
0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu56bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0n
kSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9Aphy
T948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jy
HXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIb
H0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj
04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQ
Wf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR
5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv
6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/r
auFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1
hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAri
CKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948
aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/
3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj
04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvT
k/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhT
uoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9A
uaUCbt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu5
6bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj
5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCe
UCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEw
zalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctn
juAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA="
>AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQ
Wf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aR
bKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHX
GFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAri
CKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyh
NnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFT
mp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=<
/latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr
6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/L
BjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+
KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9AuaUCb
bt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu56bmCBDZTgVbFryU
80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj5zCL/gfHwDszuRdw=</late
xit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhN
nJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5Q
tmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latex
it>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">A
AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1
ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIK
clCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPv
XYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29
eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJ
BkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUV
jbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHL
YBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPO
YUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YO
YJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9A
uaUCbt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu5
6bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj
5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCe
UCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEw
zalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctn
juAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQ
PdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr
6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/
/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZL
B4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaP
WA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0
S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnD
qlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJX
bL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnF
eCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/e
x7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyh
NnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFT
mp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=<
/latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr
6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/L
BjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+
KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9AuaUCb
bt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu56bmCBDZTgVbFryU80SpG
Mcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJ
JBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDn
jbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">A
AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ld
W9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCF
HrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUk
iJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1
xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBk
zO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNs
du2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHLY
BA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPOYUrZ
MaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YOYJUc3c
roQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9A
uaUCbt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpqECtu5
6bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8+a8L1oLTj
5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA
=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdH
dFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMop
MatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZur
viYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu
3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1
_base64="rxvQPdcIbH0VPSVp
lj1yBJIN9tA=">AB7XicbVD
LSgNBEOz1GeMr6tHLYBA8hV0R
9Rj04jGCeUCyhNnJBkzO7PM9
AphyT948aCIV/Hm3/jJNmDJh
Y0FXdHdFiRQWf/bW1ldW9/
YLGwVt3d29/ZLB4cNq1PDeJ1p
qU0ropZLoXgdBUreSgyncSR5M
xrdTv3mEzdWaPWA4SHMR0o0R
eMopMatW5m7aRbKvsVfwayTIK
clCFHrVv6vQ0S2OukElqbTvw
EwzalAwySfFTmp5QtmIDnjbU
UVjbsNsdu2EnDqlR/rauFJIZu
rviYzG1o7jyHXGFId20ZuK/3n
tFPvXYSZUkiJXbL6on0qCmkxf
Jz1hOEM5doQyI9ythA2poQxdQ
EUXQrD48jJpnFeCy4p/f1Gu3u
RxFOAYTuAMAriCKtxBDerA4BG
e4RXePO29eO/ex7x1xctnjuAP
vM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyh
NnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFT
mp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=<
/latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ
0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmR
VCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQu
lWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6YNV/LBjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJG
XLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eG
GNizCiJXGeCdmAWvan4n9fObHwdjrlM8sknS+KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhC
CBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kD
jSePHQ=</latexit>�
10-2
100
102
104
/DT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Pss
<latexit sha1_base64="Eg/FY6atVM2mXI+fOHUdfGCm6s=">AB83icbVBNS8NAEJ3Ur1q/qh69LBbBU01E1GNRDx4r9
AuaUCbt0swm7G6GE/g0vHhTx6p/x5r9x2+ag1QcDj/dmJkXJoJr47pfTmFldW19o7hZ2tre2d0r7x+0dJwqypo0FrHqhKiZ4JI1DTeCdRLFMAoFa4fj25nfmRK81g2zCRhQYRDyQecorGS76NIRnh218sa01654lbdOchf4uWkAjnqvfKn349pGjFpq
ECtu56bmCBDZTgVbFryU80SpGMcsq6lEiOmg2x+85ScWKVPBrGyJQ2Zqz8nMoy0nkSh7YzQjPSyNxP/87qpGVwHGZdJapiki0WDVBATk1kApM8Vo0ZMLEGquL2V0BEqpMbGVLIheMsv/yWt86p3WXUfLiq1mzyOIhzBMZyCB1dQg3uoQxMoJPAEL/DqpM6z8
+a8L1oLTj5zCL/gfHwDszuRdw=</latexit>�/DT
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jG
CeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqb
TvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/
ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
Steady state probability,
D
100
101
102
103
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, DT = 2
= 1, DT = 1
= 1, DT = 0.5
= 5, DT = 2
= 20, DT = 5
= 10, DT = 2
= 10, DT = 1
= 20, DT = 1
A
Probability of cell inside
contact radius
Time
B
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=
">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUCyhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdF
iRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW
5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalAwySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG
1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFO
AYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8fzk2PSA=</latexit>Pss
C
0
2
4
6
8
10
10-1
100
= 0.1
= 1
= 5
= 10
= 20
= 100
Diffusivity
10-1
100
101
102
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
DT = 0.1
DT = 0.2
DT = 1
DT = 2
DT = 5
DT = 10
Steady state probability
<latexit sha1_base64="rxvQPdcIbH0VPSVplj1yBJIN9tA=">AB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Rj04jGCeUC
yhNnJBkzO7PM9AphyT948aCIV/Hm3/jJNmDJhY0FXdHdFiRQWf/bW1ldW9/YLGwVt3d29/ZLB4cNq1PDeJ1pqU0ropZLoXgdBUreSgyncSR5MxrdTv3mEzdWaPWA4SHMR0o0ReMopMatW5m7aRbKvsVfwayTIKclCFHrVv6vQ0S2OukElqbTvwEwzalA
wySfFTmp5QtmIDnjbUVjbsNsdu2EnDqlR/rauFJIZurviYzG1o7jyHXGFId20ZuK/3ntFPvXYSZUkiJXbL6on0qCmkxfJz1hOEM5doQyI9ythA2poQxdQEUXQrD48jJpnFeCy4p/f1Gu3uRxFOAYTuAMAriCKtxBDerA4BGe4RXePO29eO/ex7x1xctnjuAPvM8
fzk2PSA=</latexit>Pss
Elastic interaction parameter,
<latexit sha1_base64="W4NsZ3UdMd3JSqHJ0nbSZyXsADw=">AB7XicbVDLSgNBEOz1GeM
r6tHLYBA8hV0R9Rj04jGCeUCyhN7JbDJmdmaZmRVCyD948aCIV/Hm3/jJNmDJhY0FXdHdFqeDG+v63t7K6tr6xWdgqbu/s7u2XDg4bRmWasjpVQulWhIYJLlndcitYK9UMk0iwZjS8nfrNJ6Y
NV/LBjlIWJtiXPOYUrZMaHRTpALulsl/xZyDLJMhJGXLUuqWvTk/RLGHSUoHGtAM/teEYteVUsEmxkxmWIh1in7UdlZgwE45n107IqVN6JFbalbRkpv6eGNizCiJXGeCdmAWvan4n9fObHwdjrl
M8sknS+KM0GsItPXSY9rRq0YOYJUc3croQPUSK0LqOhCBZfXiaN80pwWfHvL8rVmzyOAhzDCZxBAFdQhTuoQR0oPMIzvMKbp7wX7937mLeuePnMEfyB9/kDjSePHQ=</latexit>�
Figure 5. Capture statistics of motile cell. (A) Probability that cell B is inside contact radius as a
function of time. (B and C) The dependence of steady state capture probability, Pss, i.e. the fraction
of cells captured within the contact radius after a long time interval, on simulation parameters. (B)
shows the dependence on diffusivity,DT at different values of the elastic interaction parameter, a,
whereas (C) shows the dependence on a for different values of DT. (D) The steady state capture
probability, Pss, data can be collapsed into a single master curve, when plotted vs. the key parameter,
alpha/DT, the strength of the elastic interactions relative to the diffusivity. This is expected since our
model steady state is a thermal equilibrium with effective temperature set by the noisy cell motility,
DT, and the competition between attractive interactions and noise dictates the number of cells (cell
trajectories) captured vs. the number that escape.
This further justifies the notion introduced earlier of a radius of influence, that is, the
distance from the stationary cell at which its elastic attractive tendency approximately
balances the random noisy movements of the motile cell. Here we use a simple balance to
estimate this radius of influence. Working in dimensionless units, we note that the dipolar
interaction potential fall off as a/r3, while the effective temperature – a measure of the
randomizing force – scales as kBT = µTDT. Balancing these yields,
rI ⇠
✓
a
µTDT
◆ 1
3
,
(10)
which explicitly shows the importance of the a/DT parameter. Thus, a stronger a from
deformations exerted by the stationary cell (corresponding to softer substrate stiffness, or
higher contractility) and lower random movements of the motile cell, DT, leads to a larger
radius of influence. This in turn implies that the probability of being captured within the
contact radius increases because the stationary cell can influence motile cells over a larger
area.
12 of 16
4.4. Future work and perspectives: Anisotropic cell-cell elastic interactions
For polarized cells, that orient their cytoskeletal fibers and contractility along some
principal axis, the cell-cell interaction potential is not isotropic. The individual cells on
an elastic medium behave as force dipoles, with interaction potential energy having both
attractive and repulsive regions that depend on mutual orientation of the two cells and
their separation vector [19], as detailed in Appendix A. The force experienced by the motile
cell has both radial and tangential components depending on its position and orientation
relative to the central cell, and its direction is sensitive to the Poisson’s ratio of the elastic
medium [34]. Thus, trajectories of cell A interacting with stationary cell B when the fully
anisotropic interaction potential (Equation A1 and A2, Appendix A) is included will differ
from trajectories observed in isotropic potentials. The difference arises in part due to an
additional torque that reorients cell A to preferentially align with cell B as it moves towards
it. Nonetheless, qualitative nature of the capture process and the observation of an effective
region of influence will still remain valid.
X
Y
B Fixed cell
B
Cell A initially aligned
normal to dipole axis
Cell A initially aligned
along dipole axis
Figure 6. Dipolar cell orientation and trajectoryThe equilibrium orientation of contractile cells fixed
in position, but free to reorient, and that are uniformly distributed in a square box of size 10s, are
depicted by two arrows (red) pointing towards each other. Each cell is influenced by the central
stationary cell B (green) and not by each other. Two possible trajectories of cell A (blue and black) are
recorded for DT = 0.1, a = 40 for total time T = 500 with time steps of dt = 0.001. The cells did not
have any self propulsion or rotational diffusion. The Poisson’s ratio n of the substrate was considered
0.3 for this simulation
To illustrate this we simulated the equilibrium orientation of uniformly spaced
(pinned) test dipolar cells on a square lattice which are kept fixed in a square box of
length 10s. The Poisson’s ratio of the simulated substrate is 0.3 and a is 40. Results are
shown in Figure 6. None of the cells overlap with the central stationary cell; they may
rotate to reorient their dipole axis but are restricted from translating. We re-iterate that
the cells on the lattice do not mutually interact with each other, but are only meant to
illustrate the interaction of a test dipolar cell A placed at different spatial locations with
the central stationary cell B. We note that fixed cells adjust the axis of their contractile
dipoles in accordance to the potential field due to cell B (the dipole axis of B is fixed).
Superposed on this are two trajectories corresponding to two cells that are freed from
constraints and allowed to rotate and translate in response to the two-cell potential and
thermal noise. The two cells start from their equilibrium orientation - i.e, they are first
held pinned and allowed to reorient until the dipole axis attains a static value and then the
pinning constraint is removed. Cells in the close vicinity of the central cell’s orientation
axis exhibit a nearly linear motion to the pole of the fixed cell (trajectory in black). Cells
away from the orientation axis take a longer route to come in contact with the central cell
(trajectory in blue). The common attribute in both trajectories is that they prefer to adhere
to the central cell’s pole, that is cell A as it moves towards B also continuously reorients in
a manner that brings it into alignment with the cell B’s polar axis (the axis of the dipole).
13 of 16
5. Discussion
Using our model for cell contractility and motility, we computed several metrics of
experimental relevance such as number of cell–cell contacts, the mean square displacement
of a motile cell in the presence of elastic deformations induced by a cell in its vicinity, and
associated capture statistics resulting from attractive interactions between two such cells.
In each case, we predict how the computed metric depends on the elastic properties of the
substrate, captured in the interaction parameter, a ⇠ 1/E, and on cell motility, captured by
the effective diffusivity, DT.
Similar to the observations for pairs of endothelial cells mechanically interacting
through the compliant substrates [3], we find that the motility and number of cell-cell
contacts are lowered at large a, corresponding to softer substrates. This is because the
elastic deformations of the substrate, and therefore, the cell–cell attractive interactions are
stronger compared to the random motility. As observed in experiments, we also find that at
intermediate interaction strength, the cells can make repeated contacts and withdrawals as
shown in the contact number measurements. For very stiff substrates, that is low interaction
strength, we find the cell remains diffusive and can migrate away from the stationary cell
and does not make frequent contacts. Our findings would therefore suggest an optimal
substrate stiffness at which contact frequency is maximal. These trends are also reflected
in the MSD measurements. Unlike the experiment, we don’t find diffusive MSD for the
strongly attractive case, but the MSD turns subdiffusive, suggesting perhaps that such high
interaction strengths were not probed in experiment.
Biologically, such altered motility and contact formation could be relevant for forming
stable adhesive contacts between cells and tissue development, including that of blood
vessels during vasculogenesis [35]. We made several simplifying assumptions in the model
(stated in section 2), including using a purely attractive and isotropic potential instead
of the dipolar potential relevant for elongated and motile cells. Fig. 6 illustrates how
the position and orientation of the motile cell with respect to the stationary cell leads
to qualitatively different trajectories when the interaction potential is dipolar. Such an
anisotropic potential is expected to lead to end–to–end alignment and contact formation of
a pair of cells. With multiple cells, larger scale structures such as chains and networks of
cells can result [19]. The influence of cellular motility on these structures will be the topic
of a future study. The advantages of complementing experimental studies with modeling
approaches as discussed in this paper is that hard to realize parameter regimes may be
easily investigated. Furthermore, the role of different physical parameters may be clearly
studied in isolation; a feature hard to achieve in an experimental setting.
In summary, our results illustrate how cell–cell mechanical interactions can lead to
their mutual contact formation without requiring specific chemical factors to guide their
motility, and how the substrate stiffness is an important control parameter in guiding cell
motility and forming multi-cellular structures. The computational framework introduced
and analyzed here can be extended to study durotaxis – that is, the modification of cell
motility by variations in substrate elasticity at the single cell or tissue level and the motion
of cells towards higher stiffness regions [36,37]. Understanding the mechanistic aspects of
cell-cell interactions as done here has implications for regenerative medicine and tissue
engineering and will guide and inform experiments exploring how cells communicate with
each other in the process of organizing and moving collectively.
Author Contributions: Conceptualization, K.D. and A.G.; methodology, A.G. and K.D.; software,
A.G and S.B.; validation, S.B., K.D and A.G.; investigation, S.B.; resources, K.D. and A.G; writing,
S.B., K.D and A.G. All authors have read and agreed to the published version of the manuscript.
Funding: AG acknowledges funding from NSF-MCB-2026782. SB, KD and AG also acknowledge
funding from the National Science Foundation: NSF-CREST: Center for Cellular and Biomolecular
Machines (CCBM) at the University of California, Merced: NSF-HRD-1547848.
Institutional Review Board Statement: Not applicable.
14 of 16
Data Availability Statement: Data is contained within the article or supplementary material.
Conflicts of Interest: The authors declare no conflict of interest.
Abbreviations
The following abbreviations are used in this manuscript:
MSD
Mean Square Displacement
Appendix A Model for a moving cell interacting with a stationary cell via substrate
elasticity
The flat substrate is treated as being semi-infinite (Figure 1) and comprised of a
linearly elastic, isotropic gel-like material with Young’s modulus E and Poisson’s ratio n,
that capture its stiffness and compressibility respectively. The minimal model that describes
the deformations created by cells exerting contractile forces on the substrate is a point-like
force dipole [31]. Two identical dipolar cells denoted by A and B move in the upper plane
(chosen to be the x-y plane, see Figure 1). Cell A is allowed to move and its dynamics
is specified completely by its location on the substrate rA(t) and by its self-propulsion
direction eA(t). Cell B is held fixed at point rB. As a result of the contractile dipoles exerted
on the substrate the cells communicate elastically. The potential WAB characterizing this
elastic interaction between the two cells is given by
WAB(r) = P2eB
j eB
i ∂j∂lGAB
ik (r)eA
k eA
l ,
with r = rA � rB,
(A1)
where P is the strength of the force dipole capturing the contractile stresses exerted by a
cell on the medium. In writing (A1), we have made the plausible assumption that cells
orient their cytoskeletal structures such as stress fibers and exert their traction primarily
along their motility axis, such that the force dipole tensor, which captures the moment of
their force distribution, is assumed to be, Pij = Peiej. The tensor
GAB
ij (r) = 1 + n
pE
(1 � n)dij
r + nrirj
r3
�
,
(A2)
is the Green’s function that captures the displacement in the elastic medium at the location
of one cell (dipole) caused by the application of a point force at the location of the other
[38]. The partial derivatives in (A1) on the right hand side are taken with respect to relative
position vector r. Standard Einstein notation has been chosen in writing the form of WAB
and the derivatives in equations (A1) and (A2).
To obtain the force and torque balance equations that govern the dynamics of cell
A, we make the simplifying assumption that the cells move in an overdamped fashion.
This implies that hydrodynamic interactions between cells are ignored, and that each cell
feels a resisting viscous frictional drag/torque that is proportional to its velocity/rotation
rate. Conversely, when acted on by a force F or a torque T, a cell in this overdamped
environment will move with velocity µTF or rotate at a rate µRT respectively. Here, µT and
µR are appropriate mobility terms that depend on the cell size.
The micro-dynamics of cell A moving on the substrate is governed by the Langevin
equations for the translation and rotary motion of cell. Recognizing that the elastic interac-
tion generates (extra) forces and torques that act on each cell, and including the effects of
fluctuating time dependent forces xxxT(t) and torques xxxR(t) originating from thermal noise,
we can write the equations for the position and orientation of cell A in the presence of cell
B as
∂rA
∂t
=
v0eA � µT
∂WAB
∂rA
+ µTxxxT(t),
and
(A3)
∂eA
∂t
=
�µR
✓
eA ⇥ ∂WAB
∂eA
◆
+ µRxxxR(t).
(A4)
15 of 16
In an equilibrium situation, the random forces and torques are white noise terms and
are related to one another by the equipartition and fluctuation-dissipation theorems:
hxxxT(t)xxxT(t0)i = (2kBT/µT)dddd(t � t0) where ddd is the Kronecker delta function. For active
cells however, these restrictions do not hold; these terms are set by active internal cell
responses to the substrate properties. Equations (A1-A4) are used in the results illustrated
in Figure 5.
In the bulk of the paper and for results presented in Figures 1-4, we use an isotropic
version of the potential in equation (A1) that ignores orientational dynamics that are in
general present for highly elongated cells. This assumes a separation of scales between the
time over which cells reorient and the dipole axis changes and the time for the center of
the cell to move significantly such as when the rotation noise in (A4) is significant. In this
limit, one can average over the rapid reorientations of the cells and replace eB
j eB
i by dij and
eA
k eA
l by dkl. Equation (A1) then reduces to the simpler form that we employ in the main
discussion of the paper and implement as a numerical simulation,
WAB(r) = P2∂i∂kGAB
ik (r) = P2
E
f(n)
r3
(A5)
with the function f(n) = (1 � n2)/p dependent solely on the Poisson ratio, and hence
fixed in the simulation. Furthermore, since the dipole axis of cell A reorients in time scales
much faster than its slower rate of translation, the voeA term in (A3) simplifies to a time
fluctuating variable with a mean that is roughly zero but with a non-zero variance. Thus
its net effect may be incorporated by appropriately modifying the translational diffusivity.
For an isotropic symmetric potential as here, the equation that needs to be solved is then
∂r
∂t = �µT
∂WAB
∂r
+ µTxxxT
⇤ (t),
(A6)
with the modified random force xxxT
⇤ reflecting an effective translational diffusivity Deff differ-
ent from the thermal diffusivity D0, through a relation, hxxxT
⇤ (t)xxxT
⇤ (t0)i = (Deff/µ2
T)dddd(t � t0).
We define the dimensionless number DT ⌘ Deff/D0. Consistent with this, we choose
µT = Deff/kBT.
References
1.
Ladoux, B.; Mège, R.M. Mechanobiology of collective cell behaviours. Nature Reviews Molecular Cell Biology 2017, 18, 743–757.
2.
Alert, R.; Trepat, X. Physical Models of Collective Cell Migration. Annual Review of Condensed Matter Physics 2020, 11, 77–101.
3.
Reinhart-King, C.A.; Dembo, M.; Hammer, D.A. Cell-Cell Mechanical Communication through Compliant Substrates. Biophysical
Journal 2008, 95, 6044 – 6051.
4.
Winer JP, Oake S, J.P.. Non-Linear Elasticity of Extracellular Matrices Enables Contractile Cells to Communicate Local Position
and Orientation. PLoS ONE 2009, 4, e6382.
5.
Balaban, N.Q.; Schwarz, U.S.; Riveline, D.; Goichberg, P.; Tzur, G.; Sabanay, I.; Mahalu, D.; Safran, S.; Bershadsky, A.; Addadi, L.;
Geiger, B. Force and focal adhesion assembly: a close relationship studied using elastic micropatterned substrates. Nature Cell
Biology 2001, 3, 466. Article.
6.
Guo, C.L.; Ouyang, M.; Yu, J.Y.; Maslov, J.; Price, A.; Shen, C.Y. Long-range mechanical force enables self-assembly of epithelial
tubular patterns. Proceedings of the National Academy of Sciences 2012, 109, 5576–5582.
7.
Wang, H.; Abhilash, A.; Chen, C.S.; Wells, R.G.; Shenoy, V.B. Long-Range Force Transmission in Fibrous Matrices Enabled by
Tension-Driven Alignment of Fibers. Biophysical Journal 2014, 107, 2592 – 2603.
8.
Natan, S.; Koren, Y.; Shelah, O.; Goren, S.; Lesman, A. Long-range mechanical coupling of cells in 3D fibrin gels. Molecular Biology
of the Cell 2020, 31, 1474–1485. PMID: 32374653.
9.
Pakshir, P.; Alizadehgiashi, M.; Wong, B.; Coelho, N.M.; Chen, X.; Gong, Z.; Shenoy, V.B.; McCulloch, C.A.; Hinz, B. Dynamic
fibroblast contractions attract remote macrophages in fibrillar collagen matrix. Nature Communications 2019, 10, 1850.
10.
Yuval, J.; Safran, S.A.
Dynamics of elastic interactions in soft and biological matter.
Phys.
Rev.
E 2013, 87, 042703.
doi:10.1103/PhysRevE.87.042703.
11.
Tang, X.; Bajaj, P.; Bashir, R.; Saif, T.A. How far cardiac cells can see each other mechanically. Soft Matter 2011, 7, 6151–6158.
12.
Nitsan, I.; Drori, S.; Lewis, Y.E.; Cohen, S.; Tzlil, S. Mechanical communication in cardiac cell synchronized beating. Nature
Physics 2016, 12, 472–477.
16 of 16
13.
Dasbiswas, K.; Majkut, S.; Discher, D.E.; Safran, S.A. Substrate stiffness-modulated registry phase correlations in cardiomyocytes
map structural order to coherent beating. Nature Communications 2015, 6, 6085.
14.
Tomasek, J.J.; Gabbiani, G.; Hinz, B.; Chaponnier, C.; Brown, R.A. Myofibroblasts and mechano-regulation of connective tissue
remodelling. Nature Reviews Molecular Cell Biology 2002, 3, 349–363.
15.
Korff, T.; Augustin, H. Tensional forces in fibrillar extracellular matrices control directional capillary sprouting. Journal of Cell
Science 1999, 112, 3249–3258.
16.
Liu, Z.; Woo, S.; Weiner, O.D. Nodal signaling has dual roles in fate specification and directed migration during germ layer
segregation in zebrafish. Development 2018, 145.
17.
Pelham, R.J.; Wang, Y.l. Cell locomotion and focal adhesions are regulated by substrate flexibility. Proceedings of the National
Academy of Sciences 1997, 94, 13661–13665.
18.
Discher, D.E.; Janmey, P.; Wang, Y.l. Tissue Cells Feel and Respond to the Stiffness of Their Substrate. Science 2005, 310, 1139–1143.
doi:10.1126/science.1116995.
19.
Bischofs, I.B.; Schwarz, U.S. Cell organization in soft media due to active mechanosensing. Proceedings of the National Academy of
Sciences 2003, 100, 9274–9279.
20.
van Oers, R.; Rens, E.; LaValley, D.; Reinhart-King, C.; Merks, R. Mechanical Cell-Matrix Feedback Explains Pairwise and
Collective Endothelial Cell Behavior In Vitro. PLoS Comput Biol 2014, 10, e1003774.
21.
Schwarz, U.S.; Safran, S.A. Physics of adherent cells. Rev. Mod. Phys. 2013, 85, 1327–1381. doi:10.1103/RevModPhys.85.1327.
22.
Mandal, K.; Wang, I.; Vitiello, E.; Orellana, L.A.C.; Balland, M. Cell dipole behaviour revealed by ECM sub-cellular geometry.
Nature Communications 2014, 5, 5749.
23.
Mura, T. Micromechanics of Defects in Solids; TheHague: Martnius Nijhoff, 1991.
24.
Ben-Yaakov, D.; Golkov, R.; Shokef, Y.; Safran, S.A. Response of adherent cells to mechanical perturbations of the surrounding
matrix. Soft Matter 2015, 11, 1412–1424.
25.
Golkov, R.; Shokef, Y. Shape regulation generates elastic interaction between living cells. New Journal of Physics 2017, 19, 063011.
26.
Fily, Y.; Subramanian, P.; Schneider, T.M.; Chelakkot, R.; Gopinath, A. Buckling instabilities and spatio-temporal dynamics of
active elastic filaments. J. Roy. Soc. Interface 2019, 17, 20190794.
27.
Yang, J.; Arratia, P.E.; Patteson, A.E.; Gopinath, A. Quenching active swarms: effects of light exposure on collective motility in
swarming Serratia marcescens. J. Roy. Soc. Interface 2019, 16, 1620180960.
28.
Notbohm, J.; Lesman, A.; Rosakis, P.; Tirrell, D.A.; Ravichandran, G. Microbuckling of fibrin provides a mechanism for cell
mechanosensing. Journal of The Royal Society Interface 2015, 12, 20150320.
29.
Xu, X.; Safran, S.A. Nonlinearities of biopolymer gels increase the range of force transmission. Phys. Rev. E 2015, 92, 032728.
30.
Golkov, R.; Shokef, Y. Elastic interactions between anisotropically contracting circular cells. Phys. Rev. E 2019, 99, 032418.
31.
Bischofs, I.B.; Safran, S.A.; Schwarz, U.S. Elastic interactions of active cells with soft materials. Phys. Rev. E 2004, 69, 021911.
32.
Phillips, R.; Kondev, J.; Theriot, J. Physical Biology of the Cell; Garland Science, Taylor & Francis Group: New York, 2008.
33.
Pei, D.; Shu, X.; Gassama-Diagne, A.; Thiery, J.P. Mesenchymal–epithelial transition in development and reprogramming. Nature
Cell Biology 2019, 21, 44–53. doi:10.1038/s41556-018-0195-z.
34.
Bischofs, I.B.; Schwarz, U.S. Effect of Poisson Ratio on Cellular Structure Formation. Phys. Rev. Lett. 2005, 95, 068102.
35.
Califano, J.P.; Reinhart-King, C.A. A Balance of Substrate Mechanics and Matrix Chemistry Regulates Endothelial Cell Network
Assembly. Cellular and Molecular Bioengineering 2008, 1, 122.
36.
Sunyer, R.; Conte, V.; Escribano, J.; Elosegui-Artola, A.; Labernadie, A.; Valon, L.; Navajas, D.; García-Aznar, J.M.; Muñoz, J.J.;
Roca-Cusachs, P.; Trepat, X. Collective cell durotaxis emerges from long-range intercellular force transmission. Science 2016,
353, 1157–1161.
37.
Novikova, E.A.; Raab, M.; Discher, D.E.; Storm, C. Persistence-Driven Durotaxis: Generic, Directed Motility in Rigidity Gradients.
Phys. Rev. Lett. 2017, 118, 078103.
38.
Landau, L.D.; Lifshitz, E.M. Theory of Elasticity; Vol. 7, Course of Theoretical Physics, Pergamon Press: London, 1959.
| 2021 | Substrate mediated elastic coupling between motile cells modulates inter–cell interactions and enhances cell–cell contact | 10.1101/2021.03.06.434234 | [
"Bose Subhaya",
"Dasbiswas Kinjal",
"Gopinath Arvind"
] | creative-commons |
Evolutionary trajectory of organelle-derived nuclear DNAs in the
1
Triticum/Aegilops complex species
2
Zhibin Zhanga#, Jing Zhaoa#, Juzuo Lia#, Jinyang Yaoa, Bin Wanga, Yiqiao Mab, Ning Lia, Tianya
3
Wanga, Hongyan Wangc, Bao Liua and Lei Gonga*
4
5
a Key Laboratory of Molecular Epigenetics of the Ministry of Education (MOE), Northeast
6
Normal University, Changchun 130024, China.
7
b Jilin Academy of Vegetable and Flower Science, Changchun, 130033, China.
8
c Laboratory of Plant Epigenetics and Evolution, School of Life Science, Liaoning University,
9
Shenyang, 110036, China.
10
11
# These authors contributed equally to this work.
12
* Corresponding author:
13
gongl100@nenu.edu.cn (LG)
14
15
16
Abstract
17
Organelle-derived nuclear DNAs, nuclear plastid DNAs (NUPTs) and nuclear
18
mitochondrial DNAs (NUMTs), have been identified in plants. Most, if not all, genes
19
residing in NUPTs/NUMTs (NUPGs/NUMGs) are known to be inactivated and
20
pseudogenized. However, the role of epigenetic control in silencing NUPGs/NUMGs
21
and the dynamic evolution of NUPTs/NUMTs with respect to organismal phylogeny
22
remain barely explored. Based on the available nuclear and organellar genomic
23
resources of the Triticum/Aegilops complex species, we investigated the evolutionary
24
fates of NUPTs/NUMTs in terms of their epigenetic silencing and their dynamic
25
occurrence rates in the nuclear diploid genomes and allopolyploid subgenomes.
26
NUPTs and NUMTs possessed similar genomic atlas, including preferential
27
integration to the transposable element-rich intergenic regions and generating
28
sequence variations in the nuclear genome. The global transcriptional silencing of
29
NUPGs/NUMGs with disrupted and intact open reading frames can be mainly
30
attributed to their repressive chromatin states, namely high levels of DNA methylation
31
and low levels of active histone modifications. Phylogenomic analyses suggested that
32
the species-specific and gradual accumulation of NUPTs/NUMTs accompanied the
33
speciation processes. Moreover, based on further pan-genomic analyses, we found
34
significant subgenomic asymmetry in the NUPT/NUMT occurrence, which
35
accumulated during allopolyploid wheat evolution. Our findings provide novel
36
insights into the dynamic evolutionary fates of organelle-derived nuclear DNA in
37
plants.
38
Key words: NUPTs/NUMTs; Evolution trajectory; Epigenetics; Subgenome
39
asymmetry; Triticum/Aegilops complex species
40
Introduction
41
In higher plants, mitochondria and plastids originated from the endosymbiotic α-
42
proteobacteria- and cyanobacteria-like prokaryotes, respectively (McFadden 1999;
43
Osteryoung and Nunnari 2003; Archibald 2015). Owing to the same cellular
44
environment, extensive inter-compartmental DNA transfer among nuclear,
45
mitochondrial, and plastid genomes occurred during the course of evolution in higher
46
plants (Martin, et al. 2002; Keeling and Palmer 2008; Kleine, et al. 2009; Downie
47
and Jansen 2015). Among those DNA transfer events, the frequency of transfer of the
48
chloroplast/mitochondrial DNA to the nucleus was much higher than that of the other
49
transfer events, such as DNA transfer from the nucleus to organelle genomes or
50
between organelle genomes (Martin, et al. 1998; Kleine, et al. 2009; Sloan, et al.
51
2018; Zhao, et al. 2019). The nuclear plastid DNA (NUPT) and nuclear
52
mitochondrial DNA (NUMT) refer to the organellar DNA derived from the plastids
53
and mitochondria, respectively, which have already been incorporated in the nuclear
54
DNA (Leister 2005). NUPTs/NUMTs occur frequently and continuously; for
55
example, more than 200 deciphered plant genomes, including Arabidopsis, rice,
56
maize, and wheat, possess NUPTs/NUMTs with varying abundance (Michalovová, et
57
al. 2013; Zhang, et al. 2020). Moreover, changes in external environment and the
58
switch of a developmental stage can lead to dramatic changes in the frequency of
59
NUPT/NUMT occurrence in the nuclear genome (Sheppard, et al. 2008; Caro, et al.
60
2010; Cheng and Ivessa 2010; Wang, Lloyd, et al. 2012). Several potential
61
mechanisms underpinning the transfer and integration of plastid DNA (ptDNA) and
62
mitochondrial DNA (mtDNA) into the nuclear genome have been posited: (i) direct
63
physical association of the nucleus with the organelle, (ii) formation of the tubular
64
extensions from the organelle membranes for the DNA transfer (Leister 2005), and
65
(iii) occurrence of double-strand breaks (DSBs) in the nuclear genome, facilitating
66
ptDNA /mtDNA integration between those breaks by the non-homologous end joining
67
pathway or through homologous recombination (Kleine, et al. 2009; Hazkani-Covo,
68
et al. 2010; Portugez, et al. 2018).
69
Several studies have reported that after the integration into the nuclear
70
genome, NUPTs/NUMTs could generate nucleotide mutations (Huang, et al. 2005),
71
amplified together with the hosting transposable elements (TEs) (VanBuren and
72
Ming 2013), or get fragmented because of TE insertion (Michalovová, et al. 2013).
73
The foregoing processes are prone to inducing interruptions in the open reading
74
frames (ORFs) of those organellar genes residing in NUPTs/NUMTs
75
(NUPGs/NUMGs). However, the inactivation and pseudogenization of
76
NUPGs/NUMGs after their integration into the nuclear genome is still an
77
underexplored area of study (Park, et al. 2020). Considering the importance of
78
epigenetic modifications in gene activity (Feng and Jacobsen 2011; Zhang, et al.
79
2018), the role of epigenetic silencing in the transcriptional inactivation of
80
NUPGs/NUMGs should be explored. Three exemplary studies on the inactivation of
81
NUPGs/NUMGs independently reported NUPTs/NUMTs as the alien nuclear genetic
82
materials (like TEs) to be silenced through genomic defense (Zhang, et al. 2020), the
83
important role of DNA methylation against NUPTs to maintain genome stability
84
(Yoshida, et al. 2019), and the possible role of epigenetic modification in the
85
silencing of a NUMT fragment in Arabidopsis (Fields, et al. 2022). However, these
86
studies did not compare the DNA methylation patterns of NUPGs/NUMGs with the
87
indigenous nuclear genes; moreover, the generality of their conclusions remained
88
unclear.
89
Besides diploid divergence and speciation, the ubiquitous whole-genome
90
duplication or polyploidization has played a pivotal role in the evolution and
91
speciation of angiosperms (Adams and Wendel 2005; Jiao, et al. 2011; Van de Peer,
92
et al. 2017). Theoretically, in a polyploid plant, multiple nuclear subgenomes are
93
merged into the same nucleus; thus, more sites are available for NUPTs/NUMTs.
94
Based on the established subgenome dominance embodied as asymmetric expression,
95
epigenetic modification, structural variation and TE dynamics (Pont and Salse
96
2017; Bird, et al. 2018; Li, et al. 2021), whether the evolutionary dynamics of
97
NUPTs/NUMTs is related to these features of subgenome asymmetry poses an
98
intriguing question.
99
The Triticum/Aegilops complex consists of 31 species including 14 diploid, 11
100
allotetraploid, and 6 allohexaploid species (Ogihara, et al. 2016). Around 7 million
101
years ago (MYA), ancient Triticum and Aegilops species diverged into two diploid
102
lineages, namely A- and B-lineages; thereafter, the D-lineage species were derived
103
from the homoploid hybridization between A- and B-lineage species. Common wheat
104
(Triticum aestivum) harboring three distinct subgenomes, A (from T. urartu in A
105
lineage), B (from an unknown species related to Aegilops speltoides in B lineage), and
106
D (from Ae. tauschii in D lineage), was eventually developed via two distinct rounds
107
of allopolyploidization events (Marcussen, et al. 2014; Levy and Feldman 2022; Li,
108
et al. 2022; Xiao, et al. 2022). Similarly, the Triticum/Aegilops complex species
109
encompass a reticulate evolutionary trajectory involving diploid speciation,
110
allopolyploidization, and crop domestication and improvement. Moreover, since a
111
series of high-quality nuclear and organelle genome assemblies in the
112
Triticum/Aegilops complex species have been recently published (Avni, et al. 2017;
113
Luo, et al. 2017; Consortium, et al. 2018; Ling, et al. 2018; Maccaferri, et al.
114
2019; Walkowiak, et al. 2020; Wang, et al. 2020; Fu 2021; Li, et al. 2022), these
115
genomic resources established the basis for systematic investigation of the dynamic
116
evolution of NUPTs/NUMTs at the phylogenomic scale (Liang, et al. 2018).
117
In this study, we investigated the evolutionary fates of NUPTs/NUMTs by
118
delineating the genome-wide atlas of NUPTs/NUMTs in the diploid and allopolyploid
119
Triticum/Aegilops species. Using common wheat as a reference, we also characterized
120
the mutational features, expression profiles, and the role of epigenetic modification in
121
the silencing of NUPGs/NUMGs. We constructed a phylogenomic- and pan-genomic-
122
based pipeline to analyze the evolution pattern of the genomic/subgenomic
123
NUPTs/NUMTs during the diploid speciation and allopolyploid evolution of the
124
Triticum/Aegilops species. Our results provide novel insights into the dynamic
125
evolutionary fates of organelle-derived nuclear DNAs in plants.
126
127
Results
128
The landscapes of NUPTs/NUMTs in the Triticum/Aegilops complex species
129
The genomic/subgenomic sequences of interest in the Triticum/Aegilops complex
130
species were identified. A total of 1,860–2,954 NUPT (1.26–3.35 Mb; 0.026%–
131
0.057%) and 2,440–4,787 NUMT (3.56–8.12 Mb; 0.084%–0.180%) high-confidence
132
sequences were identified (Figure 1B and C and Figure S3B and C; see Materials
133
and Methods), and NUMT proportion was higher than that of NUPTs (Figure 1D;
134
Mann–Whitney U test, p value < 0.001). The proportion of NUPTs and NUMTs
135
varied across different species lineages: NUPTs, D lineage (2,585–2,954, 2.00–3.35
136
Mb) > B lineage (2,018–2,213, 1.33–1.95 Mb) ≈ A lineage (1,860–2,226, 1.27–1.78
137
Mb); NUMTs, D lineage (3,250–4,787, 5.39–8.21 Mb) ≈ B lineage (except Ae.
138
speltoides; 3,078–3,142, 6.71–6.98 Mb) > A lineage (2,367–2,818, 4.29–4.81 Mb)
139
(Figure 1B and C and Figure S3B and C).
140
The genomic distribution of both NUPTs and NUMTs was conserved across the
141
genomes/subgenomes. The majority of NUPTs/NUMTs were located in the intergenic
142
regions (70.9%–88.6% for NUPTs; 76.3%–90.0% for NUMTs) and were especially
143
enriched near the Gypsy (25.8%–33.8% for NUPTs; 29.3%–36.0% for NUMTs),
144
CACTA (19.7%–24.4% for NUPTs; 18.9%–22.6% for NUMTs), and Copia (12.3%–
145
18.4% for NUPTs; 12.5%–19.0% for NUMTs) TEs (Figure 1E–F). The distribution
146
of NUPTs/NUMTs, relative to the protein-coding genes and TEs present in the
147
chromosomes, was further compared (Figure 1G) based on the IWGSC RefSeq 1.0
148
genome (T. aestivum Chinese Spring variety). The gene density gradually decreased
149
from the telomeric region to the centromeric region, whereas TEs exhibited opposite
150
trends in all three subgenomes (Figure 1G). Intriguingly, neither NUPTs nor NUMTs
151
showed a distribution similar to that of genes or TEs. The NUPT/NUMT distribution
152
patterns within different subgenomes were distinct, suggesting that large-scale
153
subgenomic/species-specific integration of the plastid/mitochondrial DNA happened
154
during the evolution of the Triticum/Aegilops complex species.
155
Additionally, the occurrence and the extent of the second
156
amplification/duplication events of NUPTs/NUMTs were investigated. Intriguingly,
157
similar to other indigenous genic duplication events, the results showed a large scale
158
of endo-nuclear replication events for both NUPTs and NUMTs (see Materials and
159
Methods). A total of 169–544 (9.0%–23.4%) NUPTs and 165–920 (6.6%–28.7%)
160
NUMTs had at least one duplication event, where the major duplication classes were
161
of dispersed duplication (43.8%–70.0% for NUPTs; 20.7%–72.9% for NUMTs) and
162
tandem duplication (25.7%–51.5% for NUPTs; 14.6%–68.9% for NUMTs), followed
163
by proximal duplication (1.5%–12.8% for NUPTs; 4.6%–16.5% for NUMTs).
164
Segmental duplication events occurred only for five species/subgenomes, 4.4% for
165
NUMTs and none for NUPTs. The endo-nuclear replication results of NUPTs/NUMTs
166
suggested their potential genetic effects on reshaping the nuclear genomic structure in
167
the Triticum/Aegilops complex species.
168
169
Genetic variations resulting in the loss of the coding ability of NUPGs/NUMGs
170
After aligning to the corresponding organelle genomes, all NUPTs in each of the
171
genome/subgenome covered the whole chloroplast genome regions, from 2x depth of
172
inverted repeat region b (IRb) in Triticum dicoccoides B-subgenome to 65x depth of
173
large single-copy (LSC) region in Aegilops sharonensis for genetic variability (Figure
174
2A). The results suggested that the ubiquitous chloroplast DNA (cpDNA) sequences
175
could transfer into the nuclear genome. During genomic evolution, NUPTs/NUMTs
176
could generate a wide range of genetic variations, such as single-nucleotide
177
polymorphisms (SNPs) and insertion/deletions (InDels). A detailed analysis of
178
respective SNPs and InDels in NUPTs among all the genomes/subgenomes with
179
reference to their indigenous chloroplast genomic sequences revealed the following:
180
(i) a total of 12.7%–21.5% SNP sites (70% non-redundant SNPs in total), among
181
which respective density ranged from 32 SNPs/kb in IRb in Thinopyrum elongatum to
182
314 SNPs/kb in LSC regions in Ae. bicornis (Figure 2B); (ii) a total of 1.6%–15.4%
183
InDels (35.0% non-redundant InDels in total), among which the respective density
184
ranged from 2 InDels/kb of IRa (inverted repeat region a) regions in T. durum B-
185
subgenome to 274 InDels/kb in LSC regions in Ae. sharonensis (Figure 2C); (iii)
186
SNPs and InDels were highly similar among NUPTs of all the genomes/subgenomes
187
in terms of their types, transitions, and transversions for SNPs and InDels of different
188
lengths. For SNPs, the proportion of transitions (especially G to A and C to T) was
189
larger than that of transversions, which is consistent with the results of a previous
190
study (Noutsos, et al. 2005). For InDels, the proportions of 1-bp variations (both
191
insertion and deletion) were significantly overwhelming compared with the other
192
length classes (Figure 2D–E); (iv) In NUPTs of all the genomes/subgenomes, the
193
proportions of conserved SNPs (0.10%) and InDels (0.05%) were very low, and
194
29.1% of SNPs and 55.1% InDels were genome-specific (Figure 2F and G); (v)
195
mutations within the same genome/subgenome origination (A, B, D, and S) were
196
highly shared, resulting in the phylogenomic-mimic clustering pattern, especially for
197
SNPs (Figure 2F and I). NUMTs exhibited similar mutation patterns as NUPTs;
198
moreover, they had certain conserved genomic variations as well as a high proportion
199
of species-specific variations (Figure S4).
200
Based on the SNP/InDel results, the genetic fate of the NUPG/NUMG ORF
201
disruptions, via fragmentation and premature and frameshift mutations, was explored
202
to determine the intact and disrupted ORFs based on the maintenance or loss of the
203
original coding ability of ORFs, respectively. Genes with intact and disrupted ORFs
204
were named as NUPGs/NUMGs and d-NUPGs/NUMGs, respectively. All organellar
205
genes were analyzed for their susceptibility to integration in the nuclear genome. The
206
NUPT/NUMT alignment with respective chloroplast/mitochondrial genomes
207
identified 234 (T. dicoccoides B-subgenome)–1,170 (Ae. sharonensis) and 152 (Ae.
208
speltoides)–395 (Ae. bicornis) sequences harboring both NUPGs/NUMGs and d-
209
NUPGs/NUMGs. Among them, d-NUPGs ranged from 216 (45.3%; T. urartu) to 890
210
(76.1%; Ae. sharonensis), and d-NUMGs ranged from 173 (48.5%; Ae. tauschii) to
211
358 (90.6%; Ae. bicornis) (Figure 3A and Figure S5A). Among the proportions
212
occupied by d-NUPGs, Sitopsis genomes, except that of Ae searsii, were the largest
213
(69.1%–76.0%), followed by A- (except T. urartu, 56.7%–60.0%), B- (47.8%–
214
54.7%), and D (47.5%–50.8%)-genomes/subgenomes (Figure 3A). The proportions
215
of d-NUMGs in the different genomes/subgenomes existed in a similar order as that
216
of NUPGs but with an overall higher probability of disrupted ORFs (except Ae.
217
searsii and T. dicoccoides A subgenomes) (Figure S5A). The frequency of the intact
218
NUPG/NUMG ORFs was determined for different organellar genes (Figure 3B and
219
Figure S5B). In reference to the indigenous chloroplast genes, almost all NUPGs,
220
including rpl22, atpB, ndhF, and rpoC2, lost their respective coding ability (based on
221
the median of function-retention frequency), whereas more than three-quarters of
222
NUPGs, including psbl, petN, psaJ, and psbN, maintained their respective coding
223
ability (Figure 3B).
224
225
Possible transcriptional silencing of NUPGs/NUMGs by repressive epigenetic
226
modifications
227
The eventual coding abilities of NUPGs/NUMGs still depend on their transcriptional
228
status. Accordingly, based on the PacBio SMRT RNA-seq data, we further analyzed
229
whether those NUPGs/NUMGs were transcribed in the nuclear genome of hexaploid
230
wheat (see Materials and Methods). We found that 0.04%–2.5% and 0.02%–0.08%
231
of transcripts/isoforms included at least one chloroplast and mitochondrial annotated
232
gene, respectively (Table 1). Furthermore, almost all transcripts/isoforms were
233
transcribed from the chloroplast/mitochondrion rather than from NUPGs/NUMGs
234
based on the similarity assessment (Table 1), suggesting the global transcriptional
235
silencing of the intact NUPGs/NUMGs after their insertion into the nuclear genome.
236
Considering the importance of epigenetic regulation in transcription, we
237
investigated whether epigenetic regulation can contribute to the foregoing
238
transcriptional silencing. Accordingly, we characterized and compared the epigenetic
239
signal intensities (DNA methylation in the CG, CHG, and CHH context and six
240
histone modifications) among NUPTs/NUMTs, NUPGs/NUMGs, protein-coding
241
genes (PC-genes), transposons (such as Gypsy, Copia, and CACTA transposons), and
242
their up/downstream flanking regions (+3 kb) (Figure 4). Notably, divergent signal
243
patterns were generated for all the aforementioned epigenetic makers between PC-
244
genes and NUPGs/NUMGs (Figure 4). Specifically, for DNA methylation,
245
NUPGs/NUMGs with flanking regions were highly methylated; their signal fluctuated
246
across the body, and flanking regions were not as remarkable as that for PC-genes
247
(Figure 4A). For the CG and CHG context, the methylation levels of NUMGs were
248
comparable to those of TEs, whereas those of NUMTs, NUPTs, and NUPGs were
249
slightly lower than those of TEs but significantly higher than those of PC-genes
250
(Figure 4A). The methylation levels of the CHH context were similar between
251
NUPGs and NUMGs, which were lower than those of NUPTs/NUMTs and TEs
252
(Figure 4A). For the euchromatin markers (H3K4me3, H3K27me3, H3K36me3, and
253
H3K9Ac), the signal intensities of PC-genes were significantly higher than those of
254
other genomic features, especially in gene-body regions, whereas NUPGs/NUMGs
255
exhibited the lowest transcriptional activation signal (Figure 4B). For the
256
heterochromatin makers, the signal intensities of H3K27me1 in NUPGs and NUMGs
257
were higher than those in PC-genes and lower than those in TEs and NUPTs/NUMTs,
258
whereas those of the NUPGs and NUMGs reached the bottom for H3K9me2 marker
259
(Figure 4C). These results suggested that the global transcriptional silencing of
260
NUPGs/NUMGs was mainly attributed to their specific chromatin states, high DNA
261
methylation level, and low level of active epigenetic modifications.
262
263
The gradual relaxation of epigenetic repression in NUPTs/NUMTs
264
Considering that most of the NUPTs/NUMTs are located in the silent chromatin
265
region, we further investigated the tempo of establishing the current epigenetic status
266
in the alien NUPTs/NUMTs gradually after their insertion. Three possible scenarios
267
were proposed, which included (i) gradual heterochromatinization, (ii) immediate
268
heterochromatinization maintained over time, and (iii) immediate silencing followed
269
by gradually relaxed heterochromatinization. To determine the scenario of the case,
270
we categorized NUPTs/NUMTs into three classes based on their insertion time, which
271
was estimated by their sequence similarity with respective donor segments in the
272
chloroplast/mitochondrial genome: young, similarity ≥ 98%; intermedium, 94% <
273
similarity < 98%; and old, similarity ≤ 94%. The epigenetic signal intensities (DNA
274
methylation in the CG, CHG, and CHH context and six histone modifications) of
275
those categorized NUPTs/NUMTs and their up/downstream flanking regions (+3 kb)
276
were characterized and compared as mentioned above.
277
Regarding DNA methylation, we observed the following: (i) the overall
278
hierarchical order of CG DNA methylation levels was “old ≈ intermedium > young”
279
and “old > intermedium ≈ young” for NUPT and NUMT body regions, respectively
280
(Figure 5A); (ii) the overall hierarchical order of CHG DNA methylation levels was
281
also “old > intermedium ≈ young” for NUMT body regions but was “intermedium >
282
young > old” for NUPT body regions; (iii) the hierarchical order “old < intermedium
283
< young” was observed for both of NUPT and NUMT flanking regions in both CG
284
and CHG contexts; (iv) the highest DNA methylation level was detected in flanking
285
regions of old NUPTs/NUMTs in the CHH context. For the epigenetic histone
286
modification, except for H3K4me3, the signal intensity of the other four euchromatic
287
markers in body regions of NUPTs and NUMTs increased with their insertion time
288
(Figure 5B). Contrary to the active chromatic markers, the two heterochromatic
289
markers existed in the following opposite trend: the signal intensities of old NUPTs
290
and NUMTs were the lowest in both body and flanking regions (Figure 5C).
291
Interestingly, compared with the young and intermedium NUPTs/NUMTs, the old
292
ones were allocated away from TEs but close to PC-genes (Figure 5D; Tukey–
293
Kramer test after Kruskal–Wallis rank sum test, p values < 2.2e-16 for both NUPTs
294
and NUMTs). These results suggested that the contextual epigenetic modifications
295
surrounding the nuclear insertion sites underpinned the repressive chromatin status of
296
young (event intermedium) NUPTs/NUMTs, whereas such epigenetic regulation can
297
be gradually relaxed during the course of evolution.
298
299
The gradual accumulation of species-specific NUPTs/NUMTs in the
300
Triticum/Aegilops complex species
301
To investigate the evolution of NUPTs/NUMTs at the phylogenic scale, we further
302
strictly identified homologous NUPTs/NUMTs (abbreviated as homo-
303
NUPTs/NUMTs) among the Triticum/Aegilops complex species and constructed their
304
polymorphism matrix. For two arbitrary NUPTs/NUMTs derived from different
305
genomes, they were classified into an identical homo-NUPT/NUMT group if they had
306
similar body and flanking regions and were located in synteny genomic regions
307
(Figures S1–S2; see Materials and Methods). Then, we characterized the dynamic
308
evolutionary history for the homo-NUPT/NUMT group. Taking the NUPT as an
309
example, among seven diploid Triticum/Aegilops and one outgroup species (Th.
310
elongatum), we identified 968 highly confident homo-NUPT groups in which the
311
majority of groups (807; 83.4%) were species-specific (defined as a specific group).
312
However, only 122 (12.6%) groups were shared by at least two species (defined as a
313
shared group) (Figure 6A and B). Following the parsimony criteria, we labeled the
314
dynamic InDels of respective homo-NUPT groups in the phylogenetic speciation tree
315
(Li, et al. 2022). Our findings are as follows: (i) the relative insertion frequency of
316
homo-NUPTs increased gradually (on each node) from the ancestral node (3.0; 7.3
317
MYA) to the present node (35.8; less than 1 MYA) for shared groups; (ii) the relative
318
insertion frequencies in the D-lineage species (18.3–25.4) were higher than those in
319
A- (T. urartu, 15.0) and B-lineage species (Ae. speltoides, 17.3) (Figure 3C), which
320
was consistent with aforementioned NUPT content in D-lineage species (Figure 1A
321
and B). Notably, neither species-specific nor shared deletion of NUPTs was detected
322
based on the current 968 homo-NUPT groups, which indicated gradual accumulation
323
during the evolution of the Triticum/Aegilops complex species. Similarity analysis
324
between NUPTs and respective original chloroplast sequences at each node/tip also
325
supported the accuracy of our current phylogeny-based method (Figure 6C).
326
Specifically, older (younger) NUPT groups, which were shared by more (less)
327
species, had lower (higher) sequence similarity. Additionally, the similarity of a given
328
species-specific group was higher than that of its nearest shared group (i.e., the
329
similarity of the Ae. tauschii-specific group was less than that of the 5-shared group).
330
We then performed pairwise homo-NUPT comparisons between Ae.
331
longissima/Ae. sharonensis (the species at the base of the phylogenetic tree) and each
332
of the rest species. The number of homo-NUPTs decreased from 1,607 (60.2%, Ae.
333
longissima vs. Ae. sharonensis) to 205 pairs (7.7%, Ae. longissima vs. Th. elongatum)
334
as the divergent time increased, whereas Ae. longissima-specific NUPTs increased
335
from 1,061 to 2,450 (29.8% to 92.3%; Figure 6D and E), which was expected. Even
336
though both unaligned flanking regions and non-syntenic NUPTs also contributed to
337
the content of species-specific NUPTs, their proportions were found to be only 5.2%–
338
17.5% among comparisons across different divergence times. Notably, the proportion
339
of real species-specific insertion increased with divergence time (from 82.5% to
340
94.8%), whereas the proportion of non-syntenic NUPTs showed an opposite trend
341
(Figure 6G). With Ae. sharonensis as a comparison anchor, we observed similar
342
results (Figure 6F and G). Similar to NUPTs, NUMTs exhibited the species-specific
343
characteristics, and their accumulation gradually increased during the differentiation
344
of the Triticum/Aegilops complex diploid species (Figure S6).
345
346
Asymmetric ptDNA/mtDNA integration into subgenomes during the evolution of
347
allopolyploid wheat
348
Contrary to the single-origin nuclear and cytoplasmic genomes in the
349
Triticum/Aegilops complex diploid species, the uniparental inheritance of maternal
350
organellar genome (B-genome origin) with multi-origin nuclear subgenomes in wild
351
and domesticated allopolyploid Triticum species (B- and A-subgenomes in
352
allotetraploid wheat and B-, A-, and D-subgenomes in allohexaploid wheat) allowed
353
us to determine whether ptDNA/mtDNA subgenome integration was asymmetric
354
during the evolution trajectory of allopolyploid wheat (because the real B-genome
355
parent for allotetraploid wheat is still controversial, allotetraploidy process was not
356
considered in our study).
357
We first compared the profiles of NUPTs/NUMTs in A- and B-subgenomes of
358
wild and domesticated allopolyploid wheat, respectively. As shown in Figure 7A, we
359
defined the dynamic index (DI) as the ratio of NUPTs/NUMTs (novel integration into
360
respective subgenomes) that occurred in the stage before compared with after
361
domestication. Accordingly, a significantly higher DI value of a certain subgenome
362
than that of its counterpart represented asymmetric ptDNA/mtDNA integration into
363
different subgenomes in the domestication process. The DI values between A- and B-
364
subgenomes were compared in two different manners, by considering or without
365
considering the corresponding diploid species (Figure 7B and C; T. urartu and Ae.
366
speltoides for A- and B-subgenomes, respectively). When we only considered Chinese
367
Spring as the representative domesticated allohexaploid wheat, the comparison
368
revealed that the DI of NUPTs in the B-subgenome was significantly higher than that
369
in the A-subgenome (0.182 vs. 0.110 if considering diploid species, p value = 4,595e-
370
6, Figure 7B; 0.145 vs. 0.110 if excluding diploid species, p value = 0.0013, Figure
371
7C; Fisher’s exact test). When considering more hexaploid wheat genomes, we found
372
the DI differences mostly supported such subgenomic asymmetry of ptDNA/mtDNA
373
integration into B-subgenome (Figure 7D and E; except NUMTs of Norin61 at
374
diploid-including manner, although not all comparisons were statistically significant).
375
Besides foregoing DI comparison, we performed a pairwise comparison of species-
376
specific and -shared NUPTs/NUMTs for paired wild and domesticated allotetraploid
377
wheat species (T. dicoccoides vs. T. durum), which consistently revealed more
378
species-specific NUPTs/NUMTs after domestication in B-subgenome than in A-
379
subgenome (Figure 7F). For hexaploidy, based on the comparison between T. durum
380
and numeric T. aestivum genomes, we observed increased dominance of
381
polymorphism in B-subgenome (Figure 7G; Mann–Whitney U test, p value = 1.985e-
382
05 and 8.505e-05 for NUMTs and NUPTs, respectively).
383
Finally, to characterize any subgenomic asymmetric accumulation of
384
NUPTs/NUMTs in wheat at the hexaploidy level, we investigated subgenomic
385
polymorphisms of NUPTs/NUMTs in 12 hexaploid wheat genomes from the pan-
386
genomic viewpoint. First, the pan-NUPTs were constructed based on homo-NUPTs
387
for each subgenome (see Materials and Methods and Figure S2), which revealed
388
their relative abundance as follows: D-subgenome (2,032) > B-subgenome (1,762) >
389
A-subgenome (1,509); the relative abundance of core-NUPTs in the three subgenomes
390
was consistently ranked as follows: D- (1,890) > B- (1,358) > A- subgenome (1,308)
391
(Figure 8A). We also calculated the NUPT polymorphism ratio based on the number
392
of core- and pan-NUPTs and found that this ratio was highest in B-subgenome as well
393
(Figure 8B; χ2 test and post hoc test, p value < 0.01). Furthermore, based on the
394
sequenced genomes of wild and domesticated allotetraploid wheat (Zavitan and
395
Svevo, for A- and B-subgenomes) and two Ae. tauschii accessions (AL8/78 and AY61
396
for D-subgenome), we further estimated the gain and loss of NUPTs during the
397
improvement process for each subgenome (Figure 8C and 8D; Materials and
398
Methods). As shown in Figure 8D, we also found that both gain and loss of NUPTs
399
preferentially occurred in B-subgenome (χ2 test and post hoc test, p value < 0.01),
400
wherein the gain of NUPTs was significantly higher than the loss of NUPTs (χ2 test, p
401
value < 0.01). We then performed a pairwise comparison among 12 genomes to
402
characterize the shared ratio of homo-NUPTs (Figure 8E), which revealed that the
403
three subgenomes showed significantly different abundance in the shared homo-
404
NUPTs, as follows: B-subgenome < A-subgenome < D-subgenome (Figure 8F;
405
Tukey–Kramer test after Kruskal–Wallis rank sum test, p value < 0.01). Similar to the
406
results of NUMTs (Figure S7), all these results suggested the subgenomic asymmetry
407
of NUPT/NUMT polymorphism during allohexaploidy and the improvement process
408
of wheat.
409
410
Discussion
411
Integration of organellar DNA (both mitochondrial and/or chloroplast DNA) into the
412
nuclear genome has been identified in many eukaryotes from fungi and plants to
413
mammals, which affects the genome structure and genetic diversity and further
414
promotes evolution (Leister 2005; Kleine, et al. 2009; Sloan, et al. 2018).
415
Nevertheless, the transcriptional expression and epigenetic state of organelle genes
416
inside NUPTs/NUMTs and the evolutionary dynamics of NUPTs/NUMTs at the
417
phylogenic scale are poorly explored. Accordingly, in the Triticum/Aegilops complex
418
species with abundant NUPTs/NUMTs and distinct evolutionary trajectories
419
(Marcussen, et al. 2014; Glémin, et al. 2019; Zhang, et al. 2020), we determined
420
the genetic mutation, transcriptional expression, and epigenetic status of
421
NUPTs/NUMTs and their phylogenomic and pan-genomic insertion characteristics
422
during diploid speciation, polyploidization, and domestication.
423
Transcriptional silencing and epigenetic control of NUPGs/NUMGs
424
A previous study showed that organelle-derived nuclear genes are always inactivated,
425
lose their original function, and are pseudogenized (Kleine, et al. 2009). Consistently,
426
we found most NUPGs/NUMGs identified in the Triticum/Aegilops complex species
427
were pseudogenized after the accumulation of genetic mutations that cause ORF
428
interruption. However, the NUPGs/NUMGs maintaining intact ORFs facilitated the
429
determination of the potential epigenetic regulation underlying respective
430
inactivation. Accordingly, our methylome and ChIP-seq analyses showed that ORF-
431
intact NUPGs/NUMGs were significantly distinct from endogenous nuclear genes but
432
similar to TEs in terms of their epigenetic modifications (Figure 4). Consistent with
433
previous findings that showed that NUPTs/NUMTs and TEs might share a similar
434
homology-dependent DNA methylation mechanism to maintain nuclear genome
435
stability (Maumus and Quesneville 2014; Yoshida, et al. 2019), we confirmed that
436
NUPGs/NUMGs did not possess the epigenetic properties of actively transcribed
437
genes. Furthermore, the full-length transcriptomic analysis confirmed that almost all
438
NUPGs/NUMGs are transcriptionally silent compared with their counterparts in the
439
organelles (Table 1). Accordingly, as a novel input to the fate of NUPGs/NUMGs,
440
ORF-intact NUPGs/NUMGs can still be transcriptionally silenced under epigenetic
441
control.
442
Another well-known fate of NUPGs/NUMGs is functional maintenance in
443
encoding proteins targeting back to the original endosymbionts, such as
444
proteobacteria-like and cyanobacteria-like prokaryotes, which involves exemplary
445
rbcS encoding subunits of the chloroplast RuBisco complex and cytochrome c
446
encoding subunits of the mitochondrial enzyme complex of oxidative phosphorylation
447
(Blier, et al. 2001; Rand, et al. 2004; Andersson and Backlund 2008). How those
448
ancient NUPGs/NUMGs escaped from foregoing transcriptional silencing is an
449
intriguing question. Given that certain NUPGs/NUMGs identified in rice species were
450
integrated into euchromatic regions (Wang and Timmis 2013), the NUPGs/NUMGs
451
generating foregoing functional genes possibly integrated into the euchromatic
452
regions with a low load of epigenetic silencing within their eukaryotic ancestors.
453
More epigenetic and evolutionary data with the basal eukaryotic species are required
454
to test this hypothesis.
455
456
Species-specific and continuous insertion mechanisms for NUPTs/NUMTs
457
Previous studies have compared the genomic composition and characteristics of
458
NUPTs/NUMTs within various species (Michalovová, et al. 2013; Zhao, et al.
459
2019). However, as they rarely distinguish shared- and species-specific
460
NUPTs/NUMTs, making characterization of the dynamics of NUPTs/NUMTs in the
461
evolution of closely related species is difficult (Liang, et al. 2018). In the present
462
study, we used a phylogenomic-based method to identify homo- and species-specific
463
NUPT/NUMT groups among diploid Triticum/Aegilops complex species for the
464
robust investigation of NUPTs/NUMTs at the evolutionary scale. The results
465
suggested that the species-specific insertion of NUPTs/NUMTs, having significantly
466
higher abundance than homo-NUPTs/NUMTs, mainly contributed to NUPT/NUMT
467
polymorphism within the last ~7 million years. Moreover, by using the phylogenomic
468
method, we also estimated and compared the relative insertion rates of those
469
NUPTs/NUMTs (Richly and Leister 2004; Leister 2005). Interestingly, we observed
470
a gradual increase in the relative insertion frequency of homo-NUPTs from the
471
ancestral node to the present node (Figure 6B and S6B). The relevant results showed
472
that the insertion rate was related to the number of NUPTs/NUMTs (for NUPTs, D-
473
lineage species > Ae. speltoides > T. urartu; for NUMTs, D-lineage species > T.
474
urartu > Ae. speltoides), indicating that the evolution of NUPTs/NUMTs occurred
475
gradually in diploid Triticum/Aegilops complex species.
476
Notably, the relative abundance of homo-NUPTs/NUMTs was low in the
477
respective species. Such a phenomenon could be explained by two possible events,
478
the involvement of the loss of homology at the insertion site and limited synteny. TE
479
insertion, rapid mutation, and deletion could all cause the former homology loss of the
480
insertion site; considering the non-synteny feature of TEs in the three subgenomes of
481
common wheat (Wicker, et al. 2018), the majority of ptDNA/mtDNA inserted near
482
TEs lacked collinearity between the highly diverged species.
483
Integration asymmetry of ptDNA/mtDNA among subgenomes during the
484
evolution trajectory of allopolyploid wheat
485
Besides empirical evidence for subgenomic dominance (bias in gene retention, gene
486
diversity, TE dynamics, and epigenetic modifications) in allopolyploids (Pont and
487
Salse 2017; Li, et al. 2021; Levy and Feldman 2022), we observed and defined the
488
novel subgenomic dominance regarding the asymmetric integration of
489
ptDNA/mtDNA into different subgenomes (majorly biased integration into the B-
490
subgenome). As both A- and D-subgenomes were less fractionated and B-subgenome
491
was more fractionated during allopolyploid wheat evolution (Pont, et al. 2013; El
492
Baidouri, et al. 2017; Pont and Salse 2017), the more plastic subgenome increased
493
NUPT/NUMT polymorphism. We assume that a subgenome with more TE dynamics
494
may generate more TE-related DSBs, especially during meiosis, which eventually
495
enhances the dynamics of NUPTs/NUMTs. This assumption is based on the following
496
lines of evidence: (i) the molecular basis for the integration of ptDNA/mtDNA
497
necessitates nuclear DSBs (Hazkani-Covo, et al. 2010), (ii) TE transportation
498
generally produces DSBs (Gorbunova and Levy 1999; Gasior, et al. 2006; Hedges
499
and Deininger 2007), and (iii) ptDNA/mtDNA preferably inserted in TE-related
500
regions in the present Triticum/Aegilops case. Moreover, B-subgenome, which is the
501
largest subgenome and contains more abundant TEs in the wheat genome
502
(Consortium, et al. 2018), can carry a stronger genetic load, including
503
ptDNA/mtDNA insertions. Notably, a comprehensive introgression into the B-
504
subgenome of allopolyploid wheat (Walkowiak, et al. 2020; Zhou, et al. 2020;
505
Wang, et al. 2022) may also be a potential contributor to enhance the genomic
506
diversity and further NUPT/NUMT polymorphism. Additional molecular and
507
evolutionary evidence is required to validate these speculations in the future.
508
Taken together, the present systematic analyses of the whole-genome atlas of
509
NUPTs/NUMTs in the Triticum/Aegilops complex species reveal their repressed
510
epigenetic status, species-specificity, gradual accumulation, and asymmetric
511
subgenome integration in allopolyploid species. The study provides new insights into
512
the evolution of nuclear organellar DNAs in plants.
513
514
Materials and Methods
515
Sequence resources of nuclear, chloroplast and mitochondria
516
All genome sequence resources of nuclear, chloroplast and mitochondria were
517
obtained from previous publications and/or NCBI website
518
(https://www.ncbi.nlm.nih.gov/). Nuclear genome sequences included Thinopyrum
519
elongatum, Triticum urartu (Ling, et al. 2018), five Aegilops tauschii accessions
520
(AL8/78, AY17, AY61, T093 and XJ02) (Luo, et al. 2017; Zhou, et al. 2021), five
521
Sitopsis species (Ae. speltoides, Ae. searsii, Ae. bicornis, Ae. sharonensis and Ae.
522
longissima) (Li, et al. 2022), T. dicoccoides (Avni, et al. 2017), T. durum
523
(Maccaferri, et al. 2019), T. aestivum of semi-wild Zang1817 (Guo, et al. 2020) and
524
extra 11 accessions (Chinese Spring, ArinaLrFor, Jagger, Julius, LongReach Lancer,
525
CDC Landmark, Mace, Norin61, Spelta, CDC Stanley and SY Mattis) (Consortium,
526
et al. 2018; Walkowiak, et al. 2020). Chloroplast genome sequences were obtained
527
from NCBI website, including Th. elongatum (NC_043841), T. urartu (MG958555),
528
Ae. tauschii (MG958544), Ae. speltoides (MG958553), Ae. searsii (NC_024815), Ae.
529
bicornis (NC_024831), Ae. sharonensis (NC_024815), Ae. longissima (MG958549),
530
T. dicoccoides (MG958552), T. durum (MG958545) and T. aestivum (MG958554).
531
Mitochondria genome sequence of T. aestivum (NC_036024) was also obtained from
532
NCBI. For polyploid nuclear genomes, they were split to different subgenome
533
sequences to separate database files for further identification of NUPTs/NUMTs.
534
535
Identification of NUPTs/NUMTs and intra-genomic duplication NUPTs/NUMTs
536
BLAST based method was used to identify NUPTs/NUMTs. For NUPTs, all query
537
chloroplast sequences were aligned to corresponding or related (for example, for
538
identification of NUPTs in T. aestivum A-subgenome, the chloroplast genome of T.
539
urartu was treated as query sequence) genomes/subgenomes using Blastn. While for
540
NUMTs, the mitochondria genome of T. aestivum was aligned to each
541
genomes/subgenomes using Blastn. The parameters of Blastn were set as “-evalue 1e-
542
10 -dust no -penalty -2 -word_size 9”. To obtain high confident NUPTs/NUMTs, the
543
raw NUPT/NUMT hits were further filtered based on the following criteria: (i) the
544
length of Blast hit is larger than 100bp; (ii) the similarity of Blast hit is larger than
545
90%; (iii) if the distance between adjacent NUPTs/NUMTs is less than 1kb, they were
546
merged into one NUPT/NUMT. The annotation of the genomic regions for
547
NUPTs/NUMTs was performed using ChIPseeker (Yu, et al. 2015). The nearest
548
genomic features (such as gene and different types of TEs) of NUPTs/NUMTs were
549
determined using Bedtools (https://bedtools.readthedocs.io/en/latest/index.html).
550
The intra-genomic duplication of NUPTs/NUMTs after their insertion in
551
nuclear genome were identified based on previous methods with modification (Liang,
552
et al. 2018): (i) all vs. all Blastn of NUPTs/NUMTs in each genome/subgenome was
553
performed for checking candidate NUPT/NUMT pairs; (ii) for each duplicated
554
NUPT/NUMT pair, the 5’ and 3’ flanking regions of 500bp length were extracted for
555
further pairwise alignment using Blastn. The candidate NUPT/NUMT pair was
556
retained if both of the flanking regions were well aligned; (iii) MCscanX (Wang,
557
Tang, et al. 2012) was performed to classify maintained candidate NUPT/NUMT
558
pairs to duplication categories, including dispersed, tandem, proximal and segmental
559
duplication NUPTs/NUMTs.
560
561
Detection of homo-NUPTs/NUMTs groups among genomes/subgenomes
562
24,239 conserved genomic regions (CGRs) among eight diploid genomes (including
563
seven Triticum/ Aegilops species and Th. elongatum) were constructed based on our
564
previous pipeline (Li, et al. 2022). In brief: first, the genic and flanking 20kb regions
565
of all diploid genomes were aligned to the B-subgenome sequence of IWGSC RefSeq
566
1.0 (backbone sequence) using the nucmer module of MUMmer v3.9 (Kurtz, et al.
567
2004) with the parameters --mum -c 90 -l 40. Second, the best one-to-one query-
568
reference alignments for each diploid genome were obtained based on delta-filter and
569
show-coords module. Then, Bedtools intersect module was used to identify original
570
CGRs among all species based on the alignment regions on the backbone sequence.
571
Finally, the original CGRs from backbone genome were re-aligned to each diploid
572
genome sequence for detection of final CGRs through minimap2 (v2.17) (Li 2018).
573
According to above 24,239 CGR markers, we also constructed 22,763, 23,287 and
574
22,531 CGR markers for A-, B- and D-subgenomes of 12 T. aestivum genomes,
575
respectively. The CGR marker showed remarkable syntenic relationships among
576
different genomes and therefore provided robust anchors for further homo-
577
NUPTs/NUMTs detection. The CGR markers were ranked in each
578
genome/subgenome.
579
The method for detection of homo-NUPTs/NUMTs between a given pair of
580
genome/subgenomes was mimic to intra-genomic duplication NUPTs/NUMTs. A pair
581
of NUPTs/NUMTs from two different genome/subgenomes were defined as homo-
582
NUPT/NUMT if: (i) their NUPT/NUMT bodies were well aligned; (ii) their
583
NUPT/NUMT flanking regions (500 bp) were well aligned; (iii) they were located in
584
synteny regions, i.e., the rank difference between the NUPT/NUMT-nearest CGR
585
markers was less than 20. Except homo-NUPTs/NUMTs, the remaining
586
NUPTs/NUMTs in each genome/subgenome were defined as genome/subgenome-
587
specific NUPTs/NUMTs. For given genome/subgenome-specific NUPT/NUMT, if its
588
flanking regions have syntenic locus in the other genome/subgenome, we defined
589
such locus as NUPT/NUPT-related homologous locus (NHL). The homo-
590
NUPT/NUMT pairs and NUPT/NUMT-NHL pairs were used for further identification
591
of homo-NUPT/NUMT groups.
592
For the eight diploid genomes, we performed all possible pairwise comparisons
593
and obtained 28 combinations of homo-NUPT/NUMT pairs and NUPT/NUMT-NHL
594
pairs. We used the Python model networkx to concatenate all related homo-
595
NUPT/NUMT pairs and NUPT/NUMT-NHL pairs and generate candidate homo-
596
NUPT/NUMT groups among eight species. If a given homo-NUPT/NUMT group
597
satisfied (i) including exact eight members, with each species providing one member
598
and (ii) possessing only two types of members namely NUPT/NUMT and NHL, it
599
will be maintained as a diploid-level homo-NUPT/NUMT group for further analysis.
600
For 12 hexaploid wheat genomes, we performed same analysis to obtain A-, B- and
601
D-subgenome-level homo-NUPT/NUMT groups (also defined as A-, B- and D-
602
subgenome pan-NUPTs/NUMTs), respectively. The ideogram of the pipeline on
603
identification of homo-NUPT/NUMT groups were shown in Figure S2.
604
605
Variant calling for NUPTs/NUMTs and coding ability checking for organelle-
606
derived genes
607
For each genome/subgenome, NUPTs/NUMTs were first aligned to corresponding
608
organelle genome sequences and generated related Bam files using Minimap2.
609
Second, Samtools mpileup module (https://www.samtool.org) was performed to
610
produce the pileup files which including the base information for each nucleotide site
611
of organelle genome sequences. Custom Python scripts was used to parse the pileup
612
file and obtain the variant information (including SNP and InDels) and insertion times
613
(depth) for each non-overlapping 1kb nucleotide region.
614
For NUPTs/NUMTs which covered the complete gene body regions of genes
615
in organelle genome sequences, we aligned them to corresponding coding sequences
616
of organelle genes by MAFFT (https://mafft.cbrc.jp/alignment/software/). Custom
617
Python script was used to search variant sites (including SNPs and InDels) occurred
618
in CDS regions which changed the coding ability of organelle genes. Six possible
619
destinies of organelle-derived genes might be happened: (i) identical to original
620
organelle gene (same); (ii) SNP/InDel introduced variations that cause amino acid
621
changes, but ORF region maintained and have more than 50% sequence similarity to
622
the source gene (normal); (iii) SNP/InDel introduced variations that cause amino acid
623
changes, but ORF region maintained and have less than 50% sequence similarity to
624
the source gene (new); (iv) SNP/InDel-induced premature; (v) SNP/InDel-induced
625
loss of initial and stop codons (fragment) and (vi) frame shift (the sequence length of
626
alignment region is not multiple of three). The first two types maintained the intact
627
ORF of source organelle genes were defined as NUPT/NUMT-related genes
628
(NUPGs/NUMGs), whereas the last three types with disrupted ORF were defined as
629
d-NUPGs/NUMGs.
630
631
RNA-seq data analysis
632
Considering the powerful potency for identification of full-length RNA transcripts,
633
the PacBio SMRT RNA-seq data were used to determine whether an expressed gene
634
is derived from chloroplast/mitochondria or nuclei. Previous published full-length
635
transcript datasets of hexaploidy wheat (Dong, et al. 2015; Wei, et al. 2019;
636
Athiyannan, et al. 2022), which assembled based on long sequencing reads from
637
Pacbio SMRT platform, were download from NCBI SRA and GEO database
638
(including ERR6022024, ERR6022025, ERR6022026, ERR6022027, ERR6022028,
639
ERR6022029, SRR3018829 and GSE118474). The transcript datasets were aligned to
640
the reference genome and transcriptome of both nuclear (IWGSC reference genome
641
V1.0) and organelle (MG958554 for chloroplast and NC_036024 for mitochondria) of
642
heaxploid wheat using Minimap2. If the editing distance (based on the number of
643
SNPs and InDels) of a given transcript to the NUPG/NUMG is less than that to the
644
organelle gene, such transcript is inferred as NUPG/NUMG -derived, otherwise it is
645
inferred as chloroplast/mitochondria-derived transcript/isoform.
646
647
Methylome and ChIP-seq data analysis
648
The hexaploid wheat methylome data (SRR6792673, SRR6792681, SRR6792684,
649
SRR6792687, SRR6792688 and SRR6792689) and ChIP-seq data including ASY1
650
(ERR464976), DMC1 (ERR4649761), H3K4me3 (ERR4649763), H3K27me3
651
(SRR10300747), H3K27me3 (SRR6350666), H3K9ac (SRR6350667), H3K36me3
652
(SRR6350670), H3K27me1 (ERR4649762), H3K9me2 (ERR4649764) were
653
download from NCBI SRA database (Consortium, et al. 2018; Tock, et al. 2021). A
654
combined reference genome data was constructed through merging IWGSC RefSeq
655
1.0, chloroplast (MG958554) and mitochondria (NC_036024) genome sequences for
656
further short reads alignment.
657
For methylome data analysis, after filtering out adaptors and low-quality data
658
by Trimmomatic (Bolger et al., 2014), the bisulfite-treated short reads were first
659
aligned to the combined reference genome using bismark (Krueger and Andrews
660
2011) with default parameters. Second, bismark_methylation_extractor and
661
bismark2bedGraph modules were performed to generate bedGraph files for CG, CHG
662
and CHH genomic context. Then, bedGraph files were converted to bigWig files
663
using bedGraphToBigWig script
664
(https://www.encodeproject.org/software/bedgraphtobigwig/). Finally, deepTools
665
(Ramírez, et al. 2014) computeMatrix module was used to calculate the methylation
666
level of different genomic features (including genes, TEs, NUPTs/NUMTs and
667
NUPGs/NUMGs) and flanking regions (3000 bp) in CG, CHG and CHH genomic
668
context.
669
For ChIP-seq data analysis, after filtering out adaptors and low-quality data by
670
Trimmomatic, all short reads datasets were first aligned to the combined reference
671
genome using Bowtie2 (Langmead and Salzberg 2012) with default parameters.
672
Second, the aligned reads which satisfied (i) proper pair (ii) MAPQ large than 2 and
673
(iii) less than 6 mismatches were maintained. The filtered Bam files were converted to
674
bigwig file using deepTools bamCoverage module. Finally, deepTools computeMatrix
675
module was used to calculate the reads density of different genomic features
676
(including genes, TEs and NUPTs/NUMTs and NUPGs/NUMGs) and flanking
677
regions (3000 bp).
678
679
680
Acknowledgements
681
This work was supported by the National Natural Science Foundation of China
682
(NSFC #31970238), the China Postdoctoral Science Foundation (#2021M700749),
683
the Young Scientific and Technological Talents Supporting Project of Jilin Province
684
(#QT202119) and the Fundamental Research Fund for Central Universities.
685
686
Author Contributions
687
Z.B.Z., B.L., and L.G. designed the research. Z.B.Z., B.L., and L.G. performed the
688
research. B.W. and Y.Q.M. collected and preprocessed all sequencing data, Z.B.Z.,
689
J.Z., J.Z.L., J.T.Y., N.L., and T.Y.W. analyzed data. Z.B.Z., H.Y.W., B.L., and L.G.
690
wrote the manuscript.
691
692
Competing interest statement
693
The authors declare no competing interests.
694
695
References
696
Adams KL, Wendel JF. 2005. Polyploidy and genome evolution in plants. Current
697
Opinion in Plant Biology 8:135-141.
698
Andersson I, Backlund A. 2008. Structure and function of Rubisco. Plant Physiology
699
and Biochemistry 46:275-291.
700
Archibald JM. 2015. Endosymbiosis and eukaryotic cell evolution. Current Biology
701
25:R911-R921.
702
Athiyannan N, Abrouk M, Boshoff WH, Cauet S, Rodde N, Kudrna D, Mohammed
703
N, Bettgenhaeuser J, Botha KS, Derman SS. 2022. Long-read genome
704
sequencing of bread wheat facilitates disease resistance gene cloning. Nature
705
Genetics 54:227-231.
706
Avni R, Nave M, Barad O, Baruch K, Twardziok SO, Gundlach H, Hale I, Mascher
707
M, Spannagl M, Wiebe K. 2017. Wild emmer genome architecture and diversity
708
elucidate wheat evolution and domestication. Science 357:93-97.
709
Bird KA, VanBuren R, Puzey JR, Edger PP. 2018. The causes and consequences of
710
subgenome dominance in hybrids and recent polyploids. New Phytologist
711
220:87-93.
712
Blier PU, Dufresne F, Burton RS. 2001. Natural selection and the evolution of
713
mtDNA-encoded peptides: evidence for intergenomic co-adaptation. TRENDS in
714
Genetics 17:400-406.
715
Caro P, Gómez J, Arduini A, González-Sánchez M, González-García M, Borrás C,
716
Viña J, Puertas MJ, Sastre J, Barja G. 2010. Mitochondrial DNA sequences are
717
present inside nuclear DNA in rat tissues and increase with age. Mitochondrion
718
10:479-486.
719
Cheng X, Ivessa AS. 2010. The migration of mitochondrial DNA fragments to the
720
nucleus affects the chronological aging process of Saccharomyces cerevisiae.
721
Aging Cell 9:919-923.
722
Consortium IWGS, Appels R, Eversole K, Stein N, Feuillet C, Keller B, Rogers J,
723
Pozniak CJ, Choulet F, Distelfeld A. 2018. Shifting the limits in wheat research
724
and breeding using a fully annotated reference genome. Science 361:eaar7191.
725
Dong L, Liu H, Zhang J, Yang S, Kong G, Chu JS, Chen N, Wang D. 2015. Single-
726
molecule real-time transcript sequencing facilitates common wheat genome
727
annotation and grain transcriptome research. BMC Genomics 16:1-13.
728
Downie SR, Jansen RK. 2015. A comparative analysis of whole plastid genomes from
729
the Apiales: expansion and contraction of the inverted repeat, mitochondrial to
730
plastid transfer of DNA, and identification of highly divergent noncoding
731
regions. Systematic Botany 40:336-351.
732
El Baidouri M, Murat F, Veyssiere M, Molinier M, Flores R, Burlot L, Alaux M,
733
Quesneville H, Pont C, Salse J. 2017. Reconciling the evolutionary origin of
734
bread wheat (Triticum aestivum). New Phytologist 213:1477-1486.
735
Feng S, Jacobsen SE. 2011. Epigenetic modifications in plants: an evolutionary
736
perspective. Current Opinion in Plant Biology 14:179-186.
737
Fields PD, Waneka G, Naish M, Schatz MC, Henderson IR, Sloan DB. 2022.
738
Complete Sequence of a 641-kb Insertion of Mitochondrial DNA in the
739
Arabidopsis thaliana Nuclear Genome. Genome Biology and Evolution
740
14:evac059.
741
Fu Y-B. 2021. Characterizing chloroplast genomes and inferring maternal divergence
742
of the Triticum–Aegilops complex. Scientific Reports 11:1-15.
743
Gasior SL, Wakeman TP, Xu B, Deininger PL. 2006. The human LINE-1
744
retrotransposon creates DNA double-strand breaks. Journal of Molecular Biology
745
357:1383-1393.
746
Glémin S, Scornavacca C, Dainat J, Burgarella C, Viader V, Ardisson M, Sarah G,
747
Santoni S, David J, Ranwez V. 2019. Pervasive hybridizations in the history of
748
wheat relatives. Science Advances 5:eaav9188.
749
Gorbunova V, Levy AA. 1999. How plants make ends meet: DNA double-strand
750
break repair. Trends in Plant Science 4:263-269.
751
Guo W, Xin M, Wang Z, Yao Y, Hu Z, Song W, Yu K, Chen Y, Wang X, Guan P. 2020.
752
Origin and adaptation to high altitude of Tibetan semi-wild wheat. Nature
753
Communications 11:1-12.
754
Hazkani-Covo E, Zeller RM, Martin W. 2010. Molecular poltergeists: mitochondrial
755
DNA copies (numts) in sequenced nuclear genomes. PLoS Genetics 6:e1000834.
756
Hedges D, Deininger P. 2007. Inviting instability: transposable elements, double-
757
strand breaks, and the maintenance of genome integrity. Mutation
758
Research/Fundamental and Molecular Mechanisms of Mutagenesis 616:46-59.
759
Huang CY, Grunheit N, Ahmadinejad N, Timmis JN, Martin W. 2005. Mutational
760
decay and age of chloroplast and mitochondrial genomes transferred recently to
761
angiosperm nuclear chromosomes. Plant Physiology 138:1723-1733.
762
Jiao Y, Wickett NJ, Ayyampalayam S, Chanderbali AS, Landherr L, Ralph PE,
763
Tomsho LP, Hu Y, Liang H, Soltis PS. 2011. Ancestral polyploidy in seed plants
764
and angiosperms. Nature 473:97-100.
765
Keeling PJ, Palmer JD. 2008. Horizontal gene transfer in eukaryotic evolution. Nature
766
Reviews Genetics 9:605-618.
767
Kleine T, Maier UG, Leister D. 2009. DNA transfer from organelles to the nucleus:
768
the idiosyncratic genetics of endosymbiosis. Annual Review of Plant Biology
769
60:115-138.
770
Krueger F, Andrews SR. 2011. Bismark: a flexible aligner and methylation caller for
771
Bisulfite-Seq applications. Bioinformatics 27:1571-1572.
772
Kurtz S, Phillippy A, Delcher AL, Smoot M, Shumway M, Antonescu C, Salzberg SL.
773
2004. Versatile and open software for comparing large genomes. Genome
774
Biology 5:1-9.
775
Langmead B, Salzberg SL. 2012. Fast gapped-read alignment with Bowtie 2. Nature
776
Methods 9:357-359.
777
Leister D. 2005. Origin, evolution and genetic effects of nuclear insertions of
778
organelle DNA. Trends in Genetics 21:655-663.
779
Levy AA, Feldman M. 2022. Evolution and origin of bread wheat. The Plant Cell.
780
Li H. 2018. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics
781
34:3094-3100.
782
Li L-F, Zhang Z-B, Wang Z-H, Li N, Sha Y, Wang X-F, Ding N, Li Y, Zhao J, Wu Y.
783
2022. Genome sequences of five Sitopsis species of Aegilops and the origin of
784
polyploid wheat B subgenome. Molecular Plant 15:488-503.
785
Li M, Sun W, Wang F, Wu X, Wang J. 2021. Asymmetric epigenetic modification and
786
homoeolog expression bias in the establishment and evolution of allopolyploid
787
Brassica napus. New Phytologist 232:898-913.
788
Liang B, Wang N, Li N, Kimball RT, Braun EL. 2018. Comparative genomics reveals
789
a burst of homoplasy-free numt insertions. Molecular Biology and Evolution
790
35:2060-2064.
791
Ling H-Q, Ma B, Shi X, Liu H, Dong L, Sun H, Cao Y, Gao Q, Zheng S, Li Y. 2018.
792
Genome sequence of the progenitor of wheat A subgenome Triticum urartu.
793
Nature 557:424-428.
794
Luo M-C, Gu YQ, Puiu D, Wang H, Twardziok SO, Deal KR, Huo N, Zhu T, Wang L,
795
Wang Y. 2017. Genome sequence of the progenitor of the wheat D genome
796
Aegilops tauschii. Nature 551:498-502.
797
Maccaferri M, Harris NS, Twardziok SO, Pasam RK, Gundlach H, Spannagl M,
798
Ormanbekova D, Lux T, Prade VM, Milner SG. 2019. Durum wheat genome
799
highlights past domestication signatures and future improvement targets. Nature
800
Genetics 51:885-895.
801
Marcussen T, Sandve SR, Heier L, Spannagl M, Pfeifer M, International Wheat
802
Genome Sequencing Consortium, Jakobsen KS, Wulff BB, Steuernagel B,
803
Mayer KF. 2014. Ancient hybridizations among the ancestral genomes of bread
804
wheat. Science 345:1250092.
805
Martin W, Rujan T, Richly E, Hansen A, Cornelsen S, Lins T, Leister D, Stoebe B,
806
Hasegawa M, Penny D. 2002. Evolutionary analysis of Arabidopsis,
807
cyanobacterial, and chloroplast genomes reveals plastid phylogeny and
808
thousands of cyanobacterial genes in the nucleus. Proceedings of the National
809
Academy of Sciences 99:12246-12251.
810
Martin W, Stoebe B, Goremykin V, Hansmann S, Hasegawa M, Kowallik KV. 1998.
811
Gene transfer to the nucleus and the evolution of chloroplasts. Nature 393:162-
812
165.
813
Maumus F, Quesneville H. 2014. Ancestral repeats have shaped epigenome and
814
genome composition for millions of years in Arabidopsis thaliana. Nature
815
Communications 5:1-9.
816
McFadden GI. 1999. Endosymbiosis and evolution of the plant cell. Current Opinion
817
in Plant Biology 2:513-519.
818
Michalovová M, Vyskot B, Kejnovsky E. 2013. Analysis of plastid and mitochondrial
819
DNA insertions in the nucleus (NUPTs and NUMTs) of six plant species: size,
820
relative age and chromosomal localization. Heredity 111:314-320.
821
Noutsos C, Richly E, Leister D. 2005. Generation and evolutionary fate of insertions
822
of organelle DNA in the nuclear genomes of flowering plants. Genome Research
823
15:616-628.
824
Ogihara Y, Takumi S, Handa H. 2016. Advances in Wheat Genetics: From Genome to
825
Field: Proceedings of the 12th International Wheat Genetics Symposium:
826
Springer Nature.
827
Osteryoung KW, Nunnari J. 2003. The division of endosymbiotic organelles. Science
828
302:1698-1704.
829
Park S, An B, Park S. 2020. Recurrent gene duplication in the angiosperm tribe
830
Delphinieae (Ranunculaceae) inferred from intracellular gene transfer events and
831
heteroplasmic mutations in the plastid matK gene. Scientific Reports 10:1-11.
832
Pont C, Murat F, Guizard S, Flores R, Foucrier S, Bidet Y, Quraishi UM, Alaux M,
833
Doležel J, Fahima T. 2013. Wheat syntenome unveils new evidences of
834
contrasted evolutionary plasticity between paleo‐and neoduplicated subgenomes.
835
The Plant Journal 76:1030-1044.
836
Pont C, Salse J. 2017. Wheat paleohistory created asymmetrical genomic evolution.
837
Current Opinion in Plant Biology 36:29-37.
838
Portugez S, Martin WF, Hazkani-Covo E. 2018. Mosaic mitochondrial-plastid
839
insertions into the nuclear genome show evidence of both non-homologous end
840
joining and homologous recombination. BMC Evolutionary Biology 18:1-10.
841
Ramírez F, Dündar F, Diehl S, Grüning BA, Manke T. 2014. deepTools: a flexible
842
platform for exploring deep-sequencing data. Nucleic Acids Research 42:W187-
843
W191.
844
Rand DM, Haney RA, Fry AJ. 2004. Cytonuclear coevolution: the genomics of
845
cooperation. Trends in Ecology & Evolution 19:645-653.
846
Richly E, Leister D. 2004. NUPTs in sequenced eukaryotes and their genomic
847
organization in relation to NUMTs. Molecular Biology and Evolution 21:1972-
848
1980.
849
Sheppard AE, Ayliffe MA, Blatch L, Day A, Delaney SK, Khairul-Fahmy N, Li Y,
850
Madesis P, Pryor AJ, Timmis JN. 2008. Transfer of plastid DNA to the nucleus is
851
elevated during male gametogenesis in tobacco. Plant Physiology 148:328-336.
852
Sloan DB, Warren JM, Williams AM, Wu Z, Abdel-Ghany SE, Chicco AJ, Havird JC.
853
2018. Cytonuclear integration and co-evolution. Nature Reviews Genetics
854
19:635-648.
855
Tock AJ, Holland DM, Jiang W, Osman K, Sanchez-Moran E, Higgins JD, Edwards
856
KJ, Uauy C, Franklin FCH, Henderson IR. 2021. Crossover-active regions of the
857
wheat genome are distinguished by DMC1, the chromosome axis, H3K27me3,
858
and signatures of adaptation. Genome Research 31:1614-1628.
859
Van de Peer Y, Mizrachi E, Marchal K. 2017. The evolutionary significance of
860
polyploidy. Nature Reviews Genetics 18:411-424.
861
VanBuren R, Ming R. 2013. Organelle DNA accumulation in the recently evolved
862
papaya sex chromosomes. Molecular Genetics and Genomics 288:277-284.
863
Walkowiak S, Gao L, Monat C, Haberer G, Kassa MT, Brinton J, Ramirez-Gonzalez
864
RH, Kolodziej MC, Delorean E, Thambugala D. 2020. Multiple wheat genomes
865
reveal global variation in modern breeding. Nature 588:277-283.
866
Wang D, Lloyd AH, Timmis JN. 2012. Environmental stress increases the entry of
867
cytoplasmic organellar DNA into the nucleus in plants. Proceedings of the
868
National Academy of Sciences 109:2444-2448.
869
Wang D, Timmis JN. 2013. Cytoplasmic organelle DNA preferentially inserts into
870
open chromatin. Genome Biology and Evolution 5:1060-1064.
871
Wang H, Sun S, Ge W, Zhao L, Hou B, Wang K, Lyu Z, Chen L, Xu S, Guo J. 2020.
872
Horizontal gene transfer of Fhb7 from fungus underlies Fusarium head blight
873
resistance in wheat. Science 368:eaba5435.
874
Wang Y, Tang H, DeBarry JD, Tan X, Li J, Wang X, Lee T-h, Jin H, Marler B, Guo H.
875
2012. MCScanX: a toolkit for detection and evolutionary analysis of gene
876
synteny and collinearity. Nucleic Acids Research 40:e49-e49.
877
Wang Z, Wang W, Xie X, Wang Y, Yang Z, Peng H, Xin M, Yao Y, Hu Z, Liu J. 2022.
878
Dispersed emergence and protracted domestication of polyploid wheat
879
uncovered by mosaic ancestral haploblock inference. Nature Communications
880
13:1-14.
881
Wei J, Cao H, Liu J-d, Zuo J-h, Fang Y, Lin C-T, Sun R-z, Li W-l, Liu Y-x. 2019.
882
Insights into transcriptional characteristics and homoeolog expression bias of
883
embryo and de-embryonated kernels in developing grain through RNA-Seq and
884
Iso-Seq. Functional & Integrative Genomics 19:919-932.
885
Wicker T, Gundlach H, Spannagl M, Uauy C, Borrill P, Ramírez-González RH, De
886
Oliveira R, Mayer KF, Paux E, Choulet F. 2018. Impact of transposable elements
887
on genome structure and evolution in bread wheat. Genome Biology 19:1-18.
888
Xiao J, Liu B, Yao Y, Guo Z, Jia H, Kong L, Zhang A, Ma W, Ni Z, Xu S. 2022.
889
Wheat genomic study for genetic improvement of traits in China. Science China
890
Life Sciences:1-58.
891
Yoshida T, Furihata HY, To TK, Kakutani T, Kawabe A. 2019. Genome defense
892
against integrated organellar DNA fragments from plastids into plant nuclear
893
genomes through DNA methylation. Scientific Reports 9:1-11.
894
Yu G, Wang L-G, He Q-Y. 2015. ChIPseeker: an R/Bioconductor package for ChIP
895
peak annotation, comparison and visualization. Bioinformatics 31:2382-2383.
896
Zhang G-J, Dong R, Lan L-N, Li S-F, Gao W-J, Niu H-X. 2020. Nuclear integrants of
897
organellar DNA contribute to genome structure and evolution in plants.
898
International Journal of Molecular Sciences 21:707.
899
Zhang H, Lang Z, Zhu J-K. 2018. Dynamics and function of DNA methylation in
900
plants. Nature Reviews Molecular Cell Biology 19:489-506.
901
Zhao N, Grover CE, Chen Z, Wendel JF, Hua J. 2019. Intergenomic gene transfer in
902
diploid and allopolyploid Gossypium. BMC Plant Biology 19:1-18.
903
Zhou Y, Bai S, Li H, Sun G, Zhang D, Ma F, Zhao X, Nie F, Li J, Chen L. 2021.
904
Introgressing the Aegilops tauschii genome into wheat as a basis for cereal
905
improvement. Nature Plants 7:774-786.
906
Zhou Y, Zhao X, Li Y, Xu J, Bi A, Kang L, Xu D, Chen H, Wang Y, Wang Y-g. 2020.
907
Triticum population sequencing provides insights into wheat adaptation. Nature
908
Genetics 52:1412-1422.
909
910
911
Table and Figure legends
912
Table 1. Detection of nuclear plastid DNA/nuclear mitochondrial DNA-derived
913
or organelle-derived transcripts
914
Resource
Tissue
Num.
transcripts
Origin of transcripts
(chloroplast genes or NUPGs)
Origin of transcripts
(mitochondrial genes or NUMGs)
chloroplast
NUPG
ambiguous
mitochondrial
NUM
G
ambiguous
Kariega
Dusk
87,246
1,291
0
0
22
0
0
Kariega
Flag leaves
82,292
777
0
1
27
0
0
Kariega
Root
80,622
32
0
0
35
0
0
Kariega
Seedling
47,930
687
1
0
11
0
0
Kariega
Spike
64,761
364
0
0
29
0
0
Kariega
Grain
12,474
51
0
0
2
0
0
Zhou8425B
Grain
51,459
49
0
1
8
0
1
Xiaoyan81
Grain
197,709
4927
3
6
152
0
0
915
916
917
918
Figure 1. Genomic landscape of NUPTs/NUMTs in the Triticum/Aegilops complex
919
species. (A) Phylogenetic tree topology in the Triticum/Aegilops complex species
920
from the study of Li et al. Whole-genome statistics of NUPTs including (B) numbers,
921
(C) total length, (E) the proportion of the distribution of different genomic features,
922
and (F) the proportion of nearest transposon types. (D) The proportion of the total
923
length of NUPTs and NUMTs; each point represents a given genome (diploid) or
924
subgenome (allopolyploid). (G) The relative proportion of genes, transposable
925
elements, and NUPTs/NUMTs in five different chromosome regions of the IWGSC
926
RefSeq 1.0 genome; the division of chromosome regions is based on the work of
927
Consortium et al. (H) The proportion of different types of duplicated NUPTs/NUMTs
928
in each genome/subgenome, including dispersed, tandem, proximal, and segmental
929
NUPTs/NUMTs. The corresponding genomic characteristics of NUMTs are shown in
930
Figure S3.
931
932
933
934
Figure 2. Characteristics of genetic variations in NUPTs compared with the
935
chloroplast genome. The density feature of NUPTs in different regions of the
936
chloroplast genome based on non-overlapping 1-kb windows, including (A) insertion
937
frequency (depth), (B) single nuclear polymorphism (SNP) density, and (C) InDel
938
density. NUPTs were first aligned to the chloroplast genomes and then examined for
939
each feature. LSC: large single-copy region; SSC: small single-copy region; IRa:
940
repeat region a; and IRb: repeat region b. The vertical dashed lines represent the
941
junction of the above regions. (D) The proportion of different types of SNPs. The
942
error bars indicate the standard deviation among different genomes/subgenomes. (E)
943
The proportion of different types of insertions and deletions. (F)–(G) The UpSet plot
944
based on the intersection matrix of SNPs (F) and InDels (G) in each variation site
945
among genomes/subgenomes. (H)–(I) The neighbor-joining tree topology based on
946
the intersection matrix of SNPs (H) and InDels (I) used in (F) and (G). The
947
corresponding information on NUMTs is shown in Figure S4.
948
949
950
951
952
Figure 3. The genetic fate of genes in NUPTs. (A) Different genetic fates of genes in
953
NUPTs. Three groups including lost start and terminal codons (fragment), single-
954
nucleotide polymorphism/InDel induced premature (premature), and frameshift were
955
defined as d-NUPT genes (NUPGs) (disruption of open reading frames [ORFs]),
956
whereas those that maintained original (same) and larger than 50% amino acid
957
similarity (normal) with chloroplast gene sequences were defined as NUPGs
958
(maintaining intact ORFs). (B) The proportion of intact ORFs (NUPGs) for each
959
chloroplast gene among different genomes/subgenomes. The box plot on the top panel
960
shows the median proportion of each chloroplast gene among the 13
961
genomes/subgenomes. The heatmap on the bottom panel gives detailed information
962
on each chloroplast gene in each genome/subgenome. The corresponding information
963
on NUMTs is shown in Figure S5.
964
965
966
967
Figure 4. Epigenetic profiling of different genomic features. The read density (for
968
chromatin immunoprecipitation sequencing data, measured as a Counter per Million
969
reads [CPM] value) and methylation signal (for methylome data) of the body and the
970
flanking 3-kb regions of different epigenetic categories including (A) DNA
971
methylation (including the CG, CHG, and CHH context), (B) representative
972
euchromatin markers (H3K4me3, H4K27me3, H3K36me3, and H3K9ac), and (C)
973
heterochromatin markers (H3K27me1 and H3K9me2) were investigated for protein-
974
coding genes, transposable elements (Gypsy, Copia, and CACTA), NUPTs/NUMTs,
975
and NUPGs/NUMGs, respectively. The regions between dashed lines indicate body
976
region; for protein-coding genes, “start” means the start codon position, whereas
977
“end” means the terminal codon position.
978
979
980
981
Figure 5. Epigenetic profiling of different types of NUPTs/NUMTs. The read
982
density (for chromatin immunoprecipitation sequencing data and measured as a CPM
983
value) and methylation signal (for methylome data) of the body and the flanking 3-kb
984
regions of different epigenetic categories including (A) DNA methylation (including
985
the CG, CHG, and CHH context), (B) representative euchromatin markers
986
(H3K4me3, H4K27me3, H3K36me3, and H3K9ac), and (C) heterochromatin markers
987
(H3K27me1 and H3K9me2) were investigated for young, medium, and old
988
NUPTs/NUMTs. The regions between dashed lines indicate body regions; for protein-
989
coding genes, “start” means the start codon position, whereas “end” means the
990
terminal codon position. (D) Comparisons of the distance between the three
991
NUPT/NUMT groups and adjacent transposable elements and genes. Left panel,
992
NUPTs; right panel, NUMTs. Different letters represent different distances among the
993
three NUPT/NUMT groups (Tukey–Kramer test after Kruskal–Wallis rank sum test, p
994
< 2.2e-16).
995
996
997
998
Figure 6. Evolution of NUPTs/NUMTs during species differentiation among
999
diploids from the Triticum/Aegilops complex species. (A) The number of different
1000
types of homo-NUPT/NUMT groups. For each homo-NUPT/NUMT group, “shared”
1001
means the homo-NUPTs/NUMTs shared by two or more species under each node of
1002
the phylogenetic tree (from 2 to 8 species, see B), “specific” means only one species
1003
including NUPTs/NUMTs, whereas “others” represents the remaining types. (B)
1004
Phylogeny-based statistics of NUPTs. The ideograms of shared and specific homo-
1005
NUPT/NUMT groups are drawn near each node and tip, respectively. Blue and green
1006
numbers represent the number and relative insertion ratio (insertion number between
1007
two adjacent nodes divided by the evolution time between the corresponding two
1008
adjacent nodes) of NUPTs in each node/tip. (C) Statistics of NUPT similarity
1009
(sequence similarity between NUPTs and corresponding DNA fragments in
1010
chloroplast genome sequence) for the shared and species-specific homo-
1011
NUPT/NUMT groups. (D) Example circus plots of homo-NUPT pairs between Ae.
1012
sharonensis and Ae. longissima (diverged 0.98 MYA) and between Ae. sharonensis
1013
and Ae. speltoides (diverged 7.28 MYA). The numbers indicate the homo-NUPT pairs
1014
in each comparison. (E)–(F) Change patterns of homo-NUPT pairs and species-
1015
specific NUPTs over the divergence time, with Ae. longissima (E) and Ae.
1016
sharonensis (F) considered as the base (anchor) species, respectively. For each point,
1017
the X-axis indicates the divergence time between the base species and one of the rest
1018
species. (G) The proportion of different types of non-homo-NUPTs in the base
1019
(anchor) species when compared with different non-base species. “Flanking variation”
1020
means that a given NUPT has a synteny counterpart in the non-base species but the
1021
flanking regions are not aligned with each other (the loss of homology at the insertion
1022
site). “Synteny lost” means a given NUPT has a counterpart with flanking regions
1023
matched but lost synteny relationship. The corresponding information of NUMTs is
1024
shown in Figure S6. Aspe: Ae. speltoides; Tura: T. urartu; Atau: Ae. tauschii; Asea:
1025
Ae. searsii; Abic: Ae. bicornis; Asha: Ae. sharonrnsis; Alon: Ae. longissima.
1026
1027
1028
1029
Figure 7. Subgenomic asymmetry of ptDNA/mtDNA integration during
1030
tetraploid domestication and allohexaploid processes. (A) The schematic diagram
1031
for calculating the dynamic index (DI) for NUTPs/NUMTs. The alphabet near each
1032
node/tip indicates the number of NUTPs/NUMTs in each evolution node. The dashed
1033
line indicates the boundary before and after tetraploid domestication. (B)–(C)
1034
Comparing the DI of NUPTs between A- and B-subgenomes based on (B) diploid-
1035
including and (C) diploid-excluding manners. Whether the DI in A-subgenome is
1036
significantly different from that in B-subgenome is tested using the χ2 test. In these
1037
two manners, IWGSC RefSeq 1.0 (Chinese Spring) was used as the representative
1038
common wheat genome (T. aestivum). (D)–(E) Comparing the DI of NUPTs/NUMTs
1039
between A- and B-subgenomes based on (D) diploid-including and (E) diploid-
1040
excluding manners, using different resources of common wheat genomes. The
1041
significant results of the χ2 test are shown as ** (p < 0.01) and * (p < 0.05). (F)
1042
Statistics of the ratio of the shared to the specific NUPTs/NUMTs based on the
1043
number of homo-NUPT/NUMT pairs between T. dicoccoides (before domestication)
1044
and T. durum (after domestication) (χ2 test, p < 0.05). (G) Comparisons of the ratio of
1045
specific NUPTs/NUMTs between A- and B-subgenomes based on the number of
1046
homo-NUPT/NUMT pairs between T. durum (before hexaploidy) and T. aestivum
1047
(after hexaploidy) using different resources of bread wheat genomes. p values were
1048
calculated based on the pairwise Mann–Whitney U test (p < 0.01).
1049
1050
1051
1052
Figure 8. Subgenomic asymmetry of NUPT polymorphism among bread wheat
1053
accessions. (A) The profiling of pan-NUPTs and core-NUPTs among 12 hexaploid
1054
wheat genomes using pan-genome-based analysis according to 1,509 (A-subgenome),
1055
1,752 (B-subgenome), and 2,032 (D-subgenome) homo-NUPT groups. (B)
1056
Comparisons of the NUPT polymorphism ratio ((Npan-NUPTs − Ncore-NUPTs)/Npan-NUPTs)
1057
among A-, B-, and D-subgenomes. Alphabets indicate the results of multiple
1058
comparisons based on the χ2 and post hoc tests. The numbers indicate the difference
1059
between the numbers of pan-NUPTs and core-NUPTs. (C) Ideogram of gain and loss
1060
patterns of NUPTs/nuclear mitochondrial DNAs (NUMTs) among hexaploid genomes
1061
based on two outgroup genomes for each of the three subgenomes. Taking A-
1062
subgenome as an example: the outgroup genomes are A-subgenomes of T.
1063
dicoccoides and T. durum. For each homo-NUPT group, if NUPTs did not occur in
1064
both outgroup genomes but existed in hexaploidy genomes (at least one of 12
1065
genomes), it is treated as a gain group; if NUPTs occurred in both outgroup genomes
1066
but did not exist in at least one hexaploidy genome, it is treated as a lost group. (D)
1067
The proportion of gain and loss homo-NUPT groups among the three subgenomes in
1068
hexaploid wheat species. Alphabets indicate the results of multiple comparisons based
1069
on the χ2 and post hoc tests. The number of gain and loss groups is also shown. (E)
1070
Pairwise comparisons of homo-NUPT pairs among 12 genomes of the three
1071
subgenomes. The number in each cell represents the proportion of homo-NUPTs to
1072
the total number of NUPTs for each comparison. (F) Summary of homo-NUPT ratios
1073
among the three subgenomes based on 12 hexaploid genome datasets. Alphabets
1074
indicate the results of multiple comparisons based on the Kruskal–Wallis rank sum
1075
test and Tukey–Kramer test (p < 2.2e-16). The corresponding information on NUMTs
1076
is shown in Figure S7.
1077
1078
| 2022 | Evolutionary trajectory of organelle-derived nuclear DNAs in the complex species | 10.1101/2022.12.04.519011 | [
"Zhang Zhibin",
"Zhao Jing",
"Li Juzuo",
"Yao Jinyang",
"Wang Bin",
"Ma Yiqiao",
"Li Ning",
"Wang Tianya",
"Wang Hongyan",
"Liu Bao",
"Gong Lei"
] | null |
1
Cooperative NF-κB and Notch1 signaling promotes macrophage-mediated MenaINV
1
expression in breast cancer
2
3
Camille L. Duran1,2, George S. Karagiannis2,3,4, Xiaoming Chen1,2, Ved P. Sharma1,2,3, David
4
Entenberg1,2,3, John S. Condeelis2,3,5,6*, Maja H. Oktay1,2,3*
5
6
1Department of Pathology, Albert Einstein College of Medicine / Montefiore Medical Center,
7
Bronx, NY, USA
8
2Gruss-Lipper Biophotonics Center, Albert Einstein College of Medicine / Montefiore Medical
9
Center, Bronx, NY, USA
10
3Integrated Imaging Program, Albert Einstein College of Medicine / Montefiore Medical Center,
11
Bronx, NY, USA
12
4Department of Microbiology and Immunology, Albert Einstein College of Medicine / Montefiore
13
Medical Center, Bronx, NY, USA
14
5Department of Cell Biology, Albert Einstein College of Medicine / Montefiore Medical Center,
15
Bronx, NY, USA
16
6Department of Surgery, Albert Einstein College of Medicine / Montefiore Medical Center, Bronx,
17
NY, USA
18
19
* Co-Corresponding Authors:
20
21
John S. Condeelis
22
Albert Einstein College of Medicine
23
1301 Morris Park Avenue
24
Price Building Room 220
25
Bronx, NY, 10461
26
Tel: 718-678-1112
27
Email: john.condeelis@einsteinmed.edu
28
29
Maja H. Oktay
30
Albert Einstein College of Medicine
31
1301 Morris Park Avenue
32
Price Building Room 206
33
Bronx, NY, 10461
34
Tel: 718-678-1117
35
Email: moktay@montefiore.org
36
37
38
Running title: NF-κB and Notch1 signaling promote MenaINV expression
39
40
Conflicts of Interest: The authors declare no potential conflicts of interest.
41
2
Abstract
42
Metastasis is a multistep process that leads to the formation of clinically detectable tumor foci at
43
distant organs and frequently patient demise. Only a subpopulation of breast cancer cells within
44
the primary tumor can disseminate systemically and cause metastasis. To disseminate, cancer
45
cells must express MenaINV, an isoform of the actin-regulatory protein Mena encoded by the
46
ENAH gene that endows tumor cells with transendothelial migration activity allowing them to
47
enter and exit the blood circulation. We have previously demonstrated that MenaINV mRNA and
48
protein expression is induced in cancer cells by macrophage contact. In this study, we
49
discovered the precise mechanism by which macrophages induce MenaINV expression in
50
tumor cells. We examined the promoter of the human and mouse ENAH gene and discovered a
51
conserved NF-κB transcription factor binding site. Using live imaging of an NF-κB activity
52
reporter and staining of fixed tissues from mouse and human breast cancer we further
53
determined that for maximal induction of MenaINV in cancer cell NF-κB needs to cooperate with
54
the Notch1 signaling pathway. Mechanistically, Notch1 signaling does not directly increase
55
MenaINV expression, but it enhances and sustains NF-κB signaling through retention of p65, an
56
NF-κB transcription factor, in the nucleus of tumor cells, leading to increased MenaINV
57
expression. In mice, these signals are augmented following chemotherapy treatment and
58
abrogated upon macrophage depletion. Targeting Notch1 signaling in vivo decreased NF-κB
59
signaling and MenaINV expression in the primary tumor and decreased metastasis. Altogether,
60
these data uncover mechanistic targets for blocking MenaINV induction that should be explored
61
clinically to decrease cancer cell dissemination and improve survival of patients with metastatic
62
disease.
63
Keywords: TMEM doorways, MenaINV, breast cancer, NF-κB, Notch1
64
3
Introduction
65
Breast cancer is the second leading cause of cancer-related mortality in women in the US.
66
Since the majority of breast cancer mortality is due to metastases, understanding the
67
mechanisms that drive metastases is fundamental for the development of anti-metastatic
68
therapies to improve the survival of patients with metastases.
69
The cell-biological program called “epithelial-to-mesenchymal transition” (EMT)(1, 2), during
70
which cancer cells lose epithelial polarity and cell to cell cohesion(2, 3) is, in most instances,
71
required for the onset of the metastatic cascade. The EMT program has been associated with
72
heterotypic interactions of cancer cells with stromal and immune cells (e.g. macrophages), as
73
well as with modified extracellular matrix, a hallmark of cancer progression and metastasis(2, 4,
74
5). During EMT, mRNAs encoding various proteins undergo alternative splicing(6), including
75
mRNA for Mammalian enabled (Mena). EMT-induced alternative splicing of mRNA that encodes
76
Mena, a protein involved in regulation of actin dynamics, results in a decrease of the non-
77
metastatic isoform, Mena11a(6-9). However, the generation of dissemination-competent cancer
78
cells requires an additional step: tumor cell-macrophage collisions, which lead to an increase in
79
the expression of the MenaINV isoform(10). MenaINV enhances invasive cell motility(9, 11) and
80
sensitizes cells to receptor tyrosine kinase (RTK) growth factors. These properties enable
81
cancer cells to engage in a paracrine EGF-CSF1 signaling loop with tumor-associated
82
macrophages (TAMs) and establish streaming migration with macrophages towards HGF-
83
secreting endothelial cells (ECs)(12-15).
84
MenaINV expression in breast tumor cells is crucial for the formation of invadopodia,
85
invasive protrusions required for cancer cell intravasation through portals on blood vessels
86
called tumor microenvironment of metastasis (TMEM) doorways, and for extravasation at
87
metastatic sites(16). Indeed, in vivo loss-of-function studies in Mena knockout mice, in which
88
MenaINV expression is also eliminated, demonstrate a reduction in cell invasion, motility,
89
intravasation, and metastatic dissemination in several mouse models(9, 17-19). Our recent
90
4
studies demonstrate an increased density of MenaINV cancer cells, as well as cancer stem cells
91
within 200 µm of TMEM doorways, where the most cancer cell-macrophage collisions occur,
92
indicating that increased cancer cell-macrophage contact may be responsible for endowing
93
cancer cells with both MenaINV and stem phenotypes.
94
We previously showed that MenaINV mRNA and protein expression in cancer cells involves
95
macrophage-directed Notch1 signaling(10), and that macrophage expression of Jagged1 (a
96
Notch signaling ligand) is critical for tumor cell intravasation(20). However, the promoter for the
97
ENAH gene, which encodes Mena, does not contain binding sites for transcription factors in the
98
Notch1 pathway. Thus, the mechanism of how macrophages induce expression of MenaINV,
99
the protein required for tumor cell metastasis remains unidentified.
100
We have found, and independent reports confirm, that there is a κB binding site within the
101
ENAH promoter, conserved from mouse to human(21). κB binding sites are used by
102
transcription factors in the NF-κB signaling pathway, especially p65 (also known as RelA), to
103
drive expression of target genes. There are numerous reports demonstrating Notch-mediated
104
enhancement, and context-dependent activation, of NF-κB signaling in cancer(22-24). Thus,
105
Notch1 signaling may activate the ENAH promoter indirectly via NF-κB signaling.
106
The NF-κB signaling pathway is known to play a major role in the progression of many
107
cancers through promotion of processes such as EMT, proliferation, invasion, and resistance to
108
cell death. NF-κB signaling can be activated by many factors produced within the TME,
109
including proinflammatory cytokines such as TNFα and IL1β, growth factors, and oxidative
110
stressors(25). It has become increasingly apparent that while NF-κB signaling can control a
111
myriad of pro-invasive and pro-metastatic phenotypes, the downstream consequences of NF-κB
112
activation are extraordinarily context dependent, and can, for example, enhance or inhibit
113
apoptosis or tumor growth depending on the environment or stimulus(26-32).
114
As expression of MenaINV is essential for induction of an intravasation and extravasation-
115
competent- competent phenotype in breast cancer cells which endows tumor cells the ability to
116
5
metastasize, it is critical to determine if Notch1 signaling, caused by a juxtacrine, macrophage-
117
tumor cell interaction, can promote NF-κB signaling, and subsequently contribute to increased
118
MenaINV expression in vivo. Thus, here we investigated the hypothesis that macrophage-
119
cancer cell interactions induce MenaINV expression in breast cancer cells through cooperation
120
between Notch1 and NF-κB signaling. Understanding the mechanism by which tumor cells
121
acquire MenaINV expression and its associated metastasis-inducing phenotypes is critical to aid
122
the discovery of targetable signals to decrease metastatic burden and improve survival in breast
123
cancer patients.
124
125
Materials and Methods
126
Cell lines and reagents
127
The MDA-MB-231 (231) human breast cancer cell line was purchased from ATCC, and the
128
identity of the line was re-confirmed by STR profiling (Laragen Corp.), after expansion and
129
passaging. The 6DT1 murine breast cancer cell line was generously provided by Dr. Lalage
130
Wakefield, NCI. The MDA-MB-231 and 6DT1 cell lines were maintained in 10% FBS in DMEM
131
with antibiotics. The BAC1.2F5 macrophage cell line was generously provided by Dr. Richard
132
Stanley, Albert Einstein College of Medicine, and was maintained in 10% FBS in α-MEM with
133
3,000 units/ml CSF-1. All cells were maintained at 37°C in a 5% CO2 incubator, and were
134
shown to be mycoplasma-free (Sigma LookOut Mycoplasma PCR detection kit, cat# MO0035-
135
1KT). DAPT was reconstituted in 100% ethanol to a stock concentration of 20 mg/ml, aliquoted
136
and stored at -20C (Sigma, cat# D5942). DHMEQ (MedChemExpress, cat# HY-14645) was
137
reconstituted in DMSO, aliquoted and stored at -20C. C87 (Millipore Sigma, cat# 530796) was
138
reconstituted in DMSO and stored at -80C. SAHM1 (Millipore Sigma, cat# 491002) was
139
reconstituted to 50 mg/ml in DMSO and stored at -20C, clodronate liposomes (Encapsula Nano
140
Sciences, cat# CLD-8901) were used as previously described(16). Jagged1 (Anaspec, cat# AS-
141
6
61298) Jagged1 scrambled (Anaspec, cat# AS-64239) were reconstituted in DMSO, aliquoted
142
and stored at -20C and used at 80 M/ml. Recombinant TNFα (Thermo Fisher, cat# PHC3015)
143
was reconstituted at 0.1mg/ml in water, aliquoted and stored at -80C, and used at 10 ng/ml.
144
Active TGF was used at 5 and 10 ng/ml (abcam, cat# ab50036), LiCl was reconstituted in
145
water, aliquoted and stored at -20C, and used at 25 and 50 mM (Sigma Aldrich, cat# L9650),
146
and Jagged1/2 blocking antibodies and IgG isotype control (Biolegend, cat#s 130902, 131001,
147
400902) were used at 20 M/ml.
148
The MenaINV antibodies were generated by Covance, as previously described(10), and used at
149
0.25 g/ml concentration for immunofluorescence staining. The p65 antibody used at 1:1000 for
150
western blotting and staining (Cell Signaling Technology, cat# 8242S). Lamin A/C was used at
151
1:1000 for western blotting (Cell Signaling Technology, cat# 2032S), GAPDH was used at
152
1:10,000 for western blotting (Abcam, cat# ab8245), and Iba1 was used at 1:6,000 for staining
153
(Wako, cat# 019-19741).
154
155
Design of the NF-κB activity reporter
156
The GFP-p65 CDS was cloned out of the addgene plasmid (cat# 23255) by PCR, creating AfeI
157
and PacI restriction enzyme cut sites, and ligated into the pT3-neo-Ef1a-GFP sleeping beauty
158
vector from addgene (cat# 69134), cutting out the GFP sequence from the pT3-neo-Ef1a-GFP
159
vector. Positive clones (pT3-neo-GFP-p65) were confirmed by sequencing. MDA-MB-231 and
160
6DT1 cells at 60% confluency were transiently transfected with 5.4 g of pT3-neo-GFP-p65 and
161
0.6 g of the transposase SB100 (addgene, cat# 34879) using 24 l of Lipofectamine 2000
162
(Invitrogen). Stable MDA-MB-231/GFP-p65 and 6DT1/GFP-p65 cell lines were created by
163
maintaining cells in 700 g/ml G418, for 2 weeks. Expression of GFP-p65 was confirmed by
164
western blotting and immunofluorescence staining using p65 and GFP antibodies and visual
165
7
examination for GFP fluorescence. Cells were then flow sorted for top 90-95% of cells
166
expressing GFP.
167
Tumor cell and macrophage co-culture assay
168
MDA-MB-231 tumor cells (231) and BAC1.2F5 macrophages were co-cultured as previously
169
described(10). In brief, 231 cells were seeded at 50% confluency in a 6-well plate and serum
170
starved (0.5% FBS) overnight. The next morning, macrophages were seeded in the wells at a
171
1:5 ratio (231:macrophages), in media containing 0.5% FBS and 3000 units/ml CSF-1. At this
172
point, any additional treatments or inhibitors were also added. Cells were allowed to incubate for
173
4 hours at 37°C in a 5% CO2 incubator before trypsinizing 231 tumor cells and making RNA or
174
protein extracts.
175
176
mRNA isolation and qPCR
177
Total RNA was isolated from tumor cells using RNA Mini Plus Kit (Qiagen, cat# 74134). cDNA
178
was synthesized from 1g total RNA using iScript cDNA synthesis (BioRad, cat# 1708891)
179
following manufacturer’s instructions. Quantitative RT-PCR (qPCR) was performed with Power
180
SYBR Green PCR Master Mix (applied biosystems, Thermo Fisher Scientific, cat# 4367659)
181
using a QuantStudio 3 real-time PCR instrument (applied biosystems, Thermo Fisher Scientific).
182
Expression of mRNA was normalized to human GAPDH expression levels as the endogenous
183
control. The following primers were used: human GAPDH 5’- CGACCACTTTGTCAAGCTCA -3’,
184
5’- CCCTGTTGCTGTAGCCAAAT-3’; human MenaINV 5’- GATTCAAGACCATCAGGTTGTG -
185
3’, 5’- TACATCGCAAATTAGTGCTGTC -3’; human Hes1 5’- GTGAAGCACCTCCGGAAC -3’,
186
5’- GTCACCTCGTTCATGCACTC -3’; human IL-6 5’- AGCCACTCACCTCTTCAGAAC -3’, 5’-
187
GCAAGTCTCCTCATTGAATCCAG -3’; mouse MenaINV 5’- AGAGGATGCCAATGTCTTCG -3’,
188
5’- TTAGTGCTGTCCTGCGTAGC -3’; and mouse GAPDH 5’-
189
CATGTTCCAGTATGACTCCCTC -3’, 5’- GGCCTCACCCCATTTGATGT -3’.
190
191
8
Live epifluorescence and analysis
192
GFP-p65 expressing tumor cells (MDA-MB-231/GFP-p65, 6DT1/GFP-p65) were seeded at 30%
193
onto glass-bottom dishes (Mattek, cat# P35G-1.4-14-C) and serum starved overnight (0.5%
194
FBS in DMEM). The next morning, the media was replaced with imaging media (0.5% FBS in L-
195
15) and equilibrated at the heated (37°C) microscope for 2 hours. Cells were imaged live using
196
an Olympus epifluorescence microscope with coolSNAP HQ2 CCD camera using a 40x
197
objective. One 10x10 mosaic was captured and designated as time=0 and baseline GFP-p65
198
nuclear localization and the live imaging quickly paused. Any treatment (0.1-10 ng/ml TNFα,
199
80um Jagged1, or controls) were then added and imaging was immediately resumed and
200
continued without interruption for 4 hours, taking an image of the same field approximately
201
every 2.5 minutes. Timelapse movies were processed and analyzed using FIJI/ImageJ (NIH).
202
To quantify the nuclear GFP-p65 localization over time, at time zero, a circular ROI was placed
203
inside the nucleus and intensity of GFP signal was measured and designated as baseline GFP-
204
p65 nuclear localization. The intensity of the GFP signal within the nuclear ROI was measured
205
in each frame throughout the entire time course, moving the ROI (maintaining the same ROI
206
size) only if the cell/nucleus moved in the frame. Forty-five cells were measured for each
207
treatment, with three replicate dishes per treatment.
208
209
Cell fractionation and western blotting
210
Cells were seeded into 6-well plates and serum starved overnight in 0.5% FBS in DMEM. Next
211
morning, cells were treated with TNFα, Jagged1, or control (DMSO) at concentrations and times
212
indicated in the figure legends. At the end of the treatment cells were trypsinized and
213
cytoplasmic and nuclear fractions were separated and extracted using the NE-PER kit (Thermo-
214
Fisher Scientific, cat# 78833) and stored at -80C. Before western blotting, protein extracts were
215
diluted at a 1:1 ratio in 2x laemmli sample buffer containing 2% 2-mercaptoethanol and boiled at
216
100C for 5 minutes. Protein extracts were separated using a 10% sodium dodecyl sulfate
217
9
polyacrylamide gel and transferred to immobilon polyvinylidene difluoride membranes (EMD
218
Millipore). After blocking for one hour at room temperature in odyssey blocking buffer (LI-COR
219
Biosciences), membranes were incubated with antibodies directed against p65 (1:1000, Cell
220
Signaling Technologies, cat# 8242), Lamin A/C (1:1000, Cell Signaling Technologies, cat#
221
2032), or GAPDH (1:10,000, abcam, ab8245), rotating overnight at 4°C. Membranes were
222
washed three times for five minutes with 0.1% Tween-20 in TBS before incubating for one hour
223
with goat anti-mouse and goat anti-rabbit IRDye700CW-conjugated secondary antibodies (LI-
224
COR Biosciences). Following three five-minute washes with 0.1% Tween-20 in TBS,
225
membranes were scanned using a Classic Odyssey Infared Imager (LI-COR Biosciences).
226
Quantitative analysis of images from three experiments was performed using FIJI/Image J
227
software (NIH).
228
229
Animal Models
230
All procedures were conducted in accordance with National Institutes of Health regulations and
231
approved by the Albert Einstein College of Medicine Animal Use Committee. MDA-MB-231 cells
232
were injected into the mammary fat pad of SCID mice (NCI) as previously described(33).
233
Transgenic mice expressing the polyoma virus middle-T (PyMT) antigen under the control of the
234
mammary tumor virus long terminal repeat (MMTV-LTR)(34) were bred in house and result in
235
palpable tumors at approximately 6 weeks old. Patient derived xenograft (PDX) transplants of
236
HT17 tumor chunks into SCID mice have been previously described(19, 35).
237
238
In vivo treatments
239
Notch signaling inhibition in vivo using DAPT
240
Notch signaling inhibition in vivo using DAPT has been previously described(36). In brief, DAPT
241
(Sigma-Aldrich, cat# D5942) was reconstituted in 100% ethanol to a stock concentration of 20
242
mg/ml, then further diluted in corn oil to a final concentration of 2 mg/ml. Eight-week-old PyMT
243
10
mice bearing palpable tumors and separate cohort of SCID mice with tumors from orthotopically
244
xenographed MDA-MB-231 cells were given daily intraperitoneal injections of 10 mg/kg DAPT
245
or vehicle control (1:10 ethanol in corn oil) for 14 days. On day 15, the primary tumors were
246
collected from the mice and fixed in 10% formalin. Mice were weighed on day 1 and day 15 to
247
ensure no significant weight loss was suffered due to the DAPT treatment. Duodenums were
248
stained using the Periodic acid-Schiff (PAS) staining and demonstrated an increase in goblet
249
cell hyperplasia in the intestinal crypts (Supp. Fig. 6J(36)), consistent with successful Notch
250
signaling inhibition in vivo(37-39).
251
252
Macrophage depletion using clodronate liposomes
253
Macrophage depletion using clodronate liposomes in vivo has been previously described(16).
254
Briefly, tumor bearing mice were treated with a 200 l intraperitoneal injection of clodronate or
255
PBS liposomes (Encapsula Nano Sciences, cat# CLD-8901) every other day for two weeks.
256
After completion of the treatment, primary tumors were extracted from the mice and fixed in
257
10% formalin.
258
259
Paclitaxel and clodronate treatment
260
Paclitaxel treatment of mice in vivo has been previously described(19). Briefly, paclitaxel
261
(Sigma-Aldrich) was reconstituted to a concentration of 10 mg/ml in 1:1 ethanol:cremophor-EL
262
(Millipore, cat# 238470). Tumor bearing mice were treated intravenously with either 10 mg/kg
263
paclitaxel (total of 200 l) or 200 l vehicle control (1:1 ethanol:cremophor-EL) every five days
264
for two doses. Mice were randomly divided into four treatment groups: PBS liposomes and 1:1
265
ethanol:cremophor; PBS liposomes and paclitaxel; chlodronate liposomes and 1:1
266
ethanol:cremophor; and clodronate liposomes and paclitaxel. Treatment schemes are
267
diagramed in Fig. 6A and Supp. Fig. 7A.
268
11
269
Tissue fixing, staining, and analysis
270
Following treatments described above, mice were sacrificed, and all mammary tumors were
271
extracted and immersed in 10% formalin in a volume ratio of tumor to formalin of 1:7. Tissues
272
were fixed for 24 to 48 hours and embedded in paraffin, then processed for histological
273
examination. Paraffin blocks were cut into 10 µm thick sections and slides were deparaffinized
274
by melting at 60oC in an oven equipped with a fan for 60 minutes, followed by 2x xylene
275
treatment for 20 minutes. Slides were then rehydrated, and antigen retrieval was performed in 1
276
mM EDTA (pH 8.0) or 1x citrate buffer (pH 6.0) (Diagnostic BioSystems) at 97oC for 20 minutes
277
in a conventional steamer. Endogenous peroxidase was blocked by using 0.3% hydrogen
278
peroxide in water, followed by incubation of slides in a blocking buffer solution (10% FBS, 1%
279
BSA, 0.0025% fish skin gelatin in 0.05% PBST) for 60 minutes at room temperature. Slides then
280
were stained using the multiplex tyramide signal amplification (TSA) immunofluorescence
281
assay, using the Perkin Elmer Opal 4-color Fluorescent IHC kit, according to the manufacturer’s
282
instructions. The slides were stained with primary antibodies in sequence, against p65 (1:1000,
283
Cell Signaling Technology, cat #8242S), Iba1 (1:6,000, Wako, cat# 019-19741), and MenaINV
284
(1:1000, 0.25 μg/ml, see above). Slides were then washed three times in 0.05% PBST and
285
incubated with secondary HRP-conjugated antibodies in appropriate sequence, including anti-
286
rabbit and anti-chicken for 1 hour at room temperature. After washing three times with 0.05%
287
PBST, slides were incubated with biotinylated tyramide (Perkin Elmer; Opal 4-color Fluorescent
288
IHC kit) diluted at 1:50 in amplification buffer for 10 minutes. After washing, slides were
289
incubated with spectral DAPI for 5 minutes and mounted with ProLong Gold antifade reagent
290
(Life Technologies). The slides were imaged on the Pannoramic 250 Flash II digital whole slide
291
scanner, using a 20x 0.75NA objective lens. Tissue suitable for scanning was automatically
292
detected using intensity thresholding. Whole tissue images were uploaded in Pannoramic
293
Viewer version 1.15.4 (3DHISTECH).
294
12
295
To measure p65 expression, p65 nuclear localization, MenaINV expression, and MenaINV
296
expression associated with nuclear p65 in tissue section, a total of 10 different 40x fields were
297
acquired per mouse, avoiding necrotic areas in the center of the tumor and the peritumoral
298
stromal sheath at the rim of the tumor, which is devoid of tumor cells and infiltrated by
299
inflammatory cells. The MenaINV, p65, and Iba1 channels were each thresholded just above
300
background based upon intensity compared to the secondary antibody only control slide.
301
Thresholding was achieved by only using linear methods, namely contrast adjustment.
302
303
Statistics
304
GraphPad Prism 7 and Excel were used to generate graphs/plots and for statistical hypothesis
305
testing. Statistical significance was determined by either student’s t-test (normally distributed
306
paired or unpaired dataset) or a one-way ANOVA with Tukey’s or Dunnett’s multiple
307
comparisons test, as indicated in the figure legends. Statistical significance was defined as p-
308
value < 0.05.
309
310
Results
311
NF-κB signaling mediates induction of MenaINV expression.
312
We have previously found that MenaINV mRNA and protein expression are upregulated in
313
breast cancer cells upon their direct cell contact with macrophages through Notch1
314
signaling(10). However, we now discovered that the promoter sequence for the ENAH gene
315
does not contain RBP-J/CSL consensus binding sites, the transcription sites activated by
316
Notch1 signaling. This finding indicates that Notch1 works in concert with other macrophage-
317
mediated signals to induce MenaINV expression. We and others found consensus binding sites
318
for transcription factors in NF-κB, Wnt, and TGFβ signaling pathways(21) (Supp Fig. 1A). Out
319
of these three transcription binding sites only the κB site, located at -1070 and -850 in the ENAH
320
13
promoter, is conserved across the species we examined: H. sapiens, M. mulatta, M. musculus,
321
and R. norvegicus (Supp Fig. 1A). Neither TGFβ nor Wnt signaling induced MenaINV
322
expression in human triple negative breast cancer cells MDA-MB-231 (231) in response to
323
increasing doses of TGFβ and LiCl, activators of TGFβ and Wnt signaling, respectively(40, 41)
324
(Supp Fig. 1B &C).
325
To test whether NF-κB signaling promotes MenaINV expression, we cultured 231 cells in
326
the presence or absence of BAC1.2F5 macrophages with either TNFα, a potent activator of NF-
327
κB signaling, or vehicle control. Co-culture of 231 cells with macrophages caused a 5-fold
328
increase in MenaINV mRNA expression and treatment with TNFα led to a 1.7-fold increase in
329
MenaINV expression. Under both conditions the increase in MenaINV mRNA was significant
330
compared to the 231 cells cultured alone and treated with vehicle control (Fig. 1A). The addition
331
of TNFα to the 231-macrophage co-culture did not enhance MenaINV expression beyond the
332
level observed for 231-macrophage co-culture, indicating that the addition of TNFα is likely
333
redundant to any signals provided by macrophages.
334
To test if NF-κB signaling is involved in the macrophage-induced MenaINV expression,
335
we treated the 231-macrophage co-culture with the NF-κB signaling inhibitor, DHMEQ, and
336
found the macrophage-induced increase in MenaINV mRNA expression in the 231 cells was
337
abrogated back to the level observed for 231 cells cultured alone (Fig. 1B). We also found that
338
the γ-secretase inhibitor, DAPT, which attenuates Notch signaling, only partially blocked the
339
macrophage-induced increase in MenaINV mRNA expression. The addition of both DHMEQ
340
and DAPT to the 231-macrophage co-culture brought MenaINV mRNA level back to that
341
observed for 231 cells cultured alone (Fig. 1B). We ensured that the 231-macrophage co-
342
culture was effectively activating Notch1 signaling by examining activation of Hes transcription,
343
a transcriptional target of Notch1 signaling. Accordingly, we found that Hes mRNA levels were
344
3-fold higher in co-cultured cells compared to mono-cultured control 231 cells, and was
345
abrogated when DAPT was added (Supp Fig. 2).
346
14
These results demonstrate that macrophage induced MenaINV expression in tumor cells
347
requires the simultaneous activation of Notch1 and NF-κB signaling. Since NF-κB signaling is
348
required but not sufficient to induce MenaINV expression to the levels achieved by the
349
macrophage, we hypothesized that macrophages induce MenaINV expression in tumor cells
350
through cooperation of Notch1 and NF-κB.
351
Cooperation between Notch1 and NF-κB leading to enhanced and prolonged signaling
352
between the two pathways has previously been reported in other contexts(22). For example,
353
upon Notch1 activation, the Notch intracellular domain (NICD) translocates to the nucleus where
354
it can bind to the transcription factors of the NF-κB signaling pathway. This binding event blocks
355
nuclear export of p65, causing nuclear retention of NF-κB transcription factors, and allowing for
356
sustained and enhanced NF-κB signaling and transcription of NF-κB target genes(24). Thus, we
357
hypothesized that macrophage-activated Notch1 enhances NF-κB signaling which increases
358
MenaINV expression through prolonged nuclear retention of NF-κB transcription factor, p65
359
(Fig. 1C).
360
361
Notch1 prolongs and sustains NF-κB signaling leading to MenaINV expression.
362
To test the above hypothesis, we used an NF-κB reporter which allowed us to monitor,
363
using live cell imaging, the activation of NF-κB signaling via direct visualization of p65 cellular
364
localization. Briefly, we used a GFP sequence cloned upstream of the N-terminus of human p65
365
and cloned this GFP-p65 construct downstream of an EF-1α promoter in a sleeping beauty
366
transposon vector. When NF-κB signaling is inactive, endogenous and GFP-p65 are retained in
367
the cytosol, while upon NF-κB activation, endogenous and GFP-p65 are translocated to the
368
nucleus (Supp Fig. 3A). We overexpressed the NF-κB reporter in 231 cells and 6DT1 mouse
369
breast cancer carcinoma cells and tested several concentrations of TNFα in our system to
370
activate NF-κB signaling. We determined that 10 ng/ml induces translocation of GFP-p65 into
371
15
the nucleus within 30 minutes of the onset of treatment, while in the untreated cells, GFP-p65
372
was retained in the cytosol (Supp Fig. 3B-D).
373
To examine the levels of GFP-p65 compared to endogenous p65 and ensure there was
374
no aberrant activation of NF-κB signaling in GFP-p65 overexpressing cells, we made nuclear
375
and cytosolic extracts of 231 and 6DT1 GFP-p65 expressing cells treated with TNFα for 0, 10,
376
and 30 minutes and then probed for p65 using western blotting. Cellular fractionation
377
demonstrated the exogenous GFP-p65 was expressed at similar levels to endogenous p65 in
378
both 231 and 6DT1 cells (Supp Fig. 3E-H). Although TNFα treated cells compared to untreated
379
cells had significantly higher levels of nuclear p65, the amount of exogenous and endogenous
380
p65, both nuclear and cytosolic, was similar in untreated and TNFα treated cells (Supp Fig. 3E-
381
H). To ensure that NF-κB target genes were not aberrantly activated by the GFP-p65 reporter,
382
we treated wild type and GFP-p65 expressing 231 cells with TNFα and measured induction of
383
IL-6 mRNA expression, a cytokine potently expressed following NF-κB activation. We found that
384
both wild type and GFP-p65 expressing 231 cells expressed similar levels of IL-6 mRNA
385
following TNFα treatment, and the 231/GFP-p65 cells did not display any upregulation of IL-6
386
expression in untreated conditions, compared to wild type control 231 cells (Supp Fig. 3I).
387
These results indicated that the NF-κB reporter was functional and could be used to monitor NF-
388
κB signaling activation in both human and mouse mammary carcinoma cells.
389
To determine whether Notch1 signaling could potentiate NF-κB signaling, we incubated
390
GFP-p65 expressing 231 and 6DT1 tumor cells with vehicle control, TNFα, Jagged1 (Notch1
391
ligand expressed on macrophages)(36), or both TNFα and Jagged1 combined for four hours
392
and measured the intensity of green fluorescence signal in the nucleus over time. At time zero
393
(t=0) we acquired one pre-treatment image (Fig. 2A), initiated one of the above treatments, and
394
then continued time-lapse imaging for four hours (Movies 1-4). For the TNFα alone and
395
TNFα+Jagged1 treatment groups, TNFα was added after the pre-treatment image and after 10
396
minutes of imaging, the TNFα-containing media was washed out and replaced with minimal
397
16
media or Jagged1 containing media, respectively. Stills from the time lapse movies at 0, 17, and
398
240 minutes are shown in Figure 2A and the intensity of nuclear GFP-p65 signal at each time
399
point is quantified in Figure 2B. In the vehicle control-treated cells, GFP-p65 was retained in the
400
cytosol throughout the experiment, while in the TNFα treated cells, GFP-p65 robustly
401
translocated into the nucleus within 17 minutes of the onset of treatment and two hours later
402
shuttled back into the cytosol. In the Jagged1 treated cells, GFP-p65 shuttled into the nucleus
403
very slowly over the course of four hours of imaging, never reaching the amplitude seen in the
404
TNFα treated cells (Fig. 2B). The TNFα+Jagged1 treated cells demonstrated nuclear
405
translocation of GFP-p65 at 17 minutes, as was seen in the TNFα-only treated cells, followed by
406
nuclear retention of p65 throughout the four-hour time course (Fig. 2A & B). Similar results
407
were obtained using 6DT1 GFP-p65 cells (Supp Fig 4A, Movies 5-8).
408
To ensure that p65 nuclear translocation after treatment with TNFα and Jagged1 was
409
not an artifact of the exogenously expressed GFP-p65, we treated wild type 231 cells with
410
identical conditions and made nuclear and cytoplasmic extracts after 30 minutes and four hours
411
of treatment and found a similar pattern of nuclear and cytosolic localization of endogenous p65
412
to that shown in the time lapse movies with GFP-p65 (Fig. 2C & D). To investigate if the above
413
treatments lead to an increase in MenaINV expression, we treated wild type 231 cells
414
accordingly and isolated mRNA after one and four hours of treatment. After one hour, none of
415
the treatments had an effect on MenaINV mRNA expression. After four hours of treatment,
416
TNFα alone caused a small but significant increase in MenaINV mRNA expression, Jagged1
417
alone had a slight but not significant increase in MenaINV mRNA expression, while
418
TNFα+Jagged1 treatment led to a 2.5-fold increase in MenaINV mRNA expression (Fig. 2E).
419
Similar results were obtained with 6DT1 cells (Supp Fig 4B). These data indicate that the
420
treatment which resulted in the most robust and sustained activation of NF-κB signaling, as
421
indicated by sustained nuclear p65 localization (Fig. 2B-D), also led to the most robust induction
422
of MenaINV mRNA expression. In particular, the co-activation of Notch1 and NF-κB signaling
423
17
(TNFα+Jagged1 treatment group) had a synergistic effect on MenaINV expression, compared to
424
activation of each of the signaling pathway separately (Fig. 2E). Taken together these data
425
indicate that the cooperation of Notch1 and NF-κB signaling is required for appreciable induction
426
of MenaINV expression in vitro.
427
428
Macrophage-mediated induction of MenaINV expression in tumor cells requires NF-κB
429
and Notch1
430
To determine if the macrophage-mediated induction of MenaINV expression occurs
431
specifically via TNFα and Notch1, we treated 231-macrophage co-cultures with more specific
432
inhibitors, C87 and SAHM1, respectively. C87 is a small molecule inhibitor which directly binds
433
to TNFα and blocks TNFα-induced NF-κB signaling(42). SAHM1 is a MAML1 inhibitor which
434
prevents the NICD from binding to the transcriptional co-activator MAML1, leading to inhibition
435
of Notch1 signaling downstream from receptor activation(43). While blocking TNFα activity with
436
C87 almost completely abrogated the macrophage-induced expression of MenaINV, inhibition of
437
MAML1 led to only partial reduction of MenaINV expression (Fig. 3A). Inhibition of both TNFα
438
and MAML1 brought the macrophage-induced MenaINV expression to baseline levels observed
439
when cancer cells were cultured without macrophages.
440
We next aimed to determine the specific Notch1 ligands on macrophages involved in the
441
induction of MenaINV expression. While there are many Notch1 ligands, our recent study has
442
found that the macrophages used in our co-culture experiments (BAC1.2F5), which show robust
443
upregulation of MenaINV, primarily express Jagged1 and Jagged2, and an order of magnitude
444
lower mRNA expression levels of Dll1, Dll2, and Dll4(36). Therefore, we focused our studies
445
here on the role of Jagged1 and Jagged2 in the induction of MenaINV expression. We used
446
Jagged1 and Jagged2 blocking antibodies to prevent Notch1 signaling activation in response to
447
these macrophage-derived ligands. We found that blocking the Jagged1 ligand led to a modest
448
but significant decrease in MenaINV expression compared to the tumor cell-macrophage co-
449
18
cultured control group. Blocking the Jagged2 ligand did not significantly affect MenaINV mRNA
450
expression. Blocking both Jagged1 and Jagged2 ligands together, did not lead to a further
451
inhibition of MenaINV mRNA expression compared to blocking either ligand alone (Fig. 3B).
452
This partial effect of blocking Jagged1/Notch1 signaling on MenaINV mRNA expression is
453
consistent with the results seen with the more potent Notch1 inhibitors, DAPT and SAHM1.
454
Taken together, these data indicate that macrophage-mediated induction of MenaINV
455
expression occurs via TNFα and Jagged1. Furthermore, these data show that neither Notch1
456
nor NF-κB signaling on their own could fully account for the upregulation of MenaINV
457
expression.
458
459
Macrophage depletion decreases NF-κB signaling and MenaINV expression in vivo.
460
We next wanted to determine whether macrophages are required for NF-κB mediated
461
induction of MenaINV expression in cancer cells in vivo. We used two in vivo models of breast
462
cancer previously generated in our laboratory: patient derived xenografts (PDX) from triple
463
negative breast tumors (HT17) transplanted into SCID mice, and the autochthonous transgenic
464
MMTV-PyMT transplantation model (PyMT), where a single spontaneously developed tumor is
465
transplanted into the mammary fat pad of syngenic FVB mice(19, 35). The PyMT model fully
466
recapitulates the entire breast cancer development and progression process(44). To deplete
467
macrophages, we treated mice with clodronate liposomes (Fig. 4A, Supp Fig. 5A). Upon
468
completion of treatment, we harvested the tumors and stained slides from the paraffin
469
embedded tissues for the macrophage marker, Iba1 (to ensure our treatment decreased
470
macrophage density in the primary tumor) (Fig. 4B), MenaINV, p65, and DAPI (Fig. 4C and
471
Supp Fig. 5B). Treatment with clodronate liposomes, compared to control, decreased p65
472
expression in tumor cells (Fig. 4D and Supp Fig. 5C), and of the p65 that was expressed, less
473
of it was localized in the nucleus of the tumor cells (Fig. 4E and Supp Fig. 5D). This indicates
474
that macrophage depletion decreases expression as well as activation NF-κB signaling. Further,
475
19
we found a corresponding decrease in MenaINV expression in tumors of clodronate-treated
476
compared to control-treated, mice (Fig. 4F and Supp Fig. 5E). These data indicate that
477
macrophage-mediated NF-κB activation is associated with MenaINV expression in tumor cells in
478
vivo.
479
480
Inhibition of Notch signaling in vivo decreases activation of NF-κB and MenaINV
481
expression in tumor cells
482
To determine whether inhibition of Notch1 signaling affects NF-κB activity and MenaINV
483
expression in vivo, as observed in vitro, we treated mice bearing human breast cancer cell
484
xenografts (MBA-MB-231 cells injected into the mammary fat pad) and PyMT(34) breast tumors
485
with the γ-secretase inhibitor, DAPT, or control for two weeks (Fig. 5A and Supp Fig. 6A)(36).
486
Upon completion of treatment, we harvested the tumors and stained them for p65, MenaINV,
487
and DAPI (Fig. 5B and Supp Fig. 6B). We found a decrease in nuclear p65 (active NF-κB) in
488
mice treated with DAPT compared to control mice in both models of breast cancer (Fig. 5C and
489
Supp Fig 6C). Moreover, we observed a corresponding decrease in overall MenaINV
490
expression in mice treated with DAPT, compared to control mice, in both models (Fig. 5D and
491
Supp Fig. 6D). These results indicate that Notch1 inhibition in vivo decreases expression of
492
MenaINV in an NF-κB dependent manner.
493
494
Chemotherapy treatment enhances NF-κB activation and MenaINV expression through
495
macrophage recruitment
496
We have previously shown that chemotherapy induces recruitment of macrophages into
497
the tumor microenvironment, expression of MenaINV in transgenic and xenograft (human and
498
mouse) mammary breast carcinoma models, and expression of MenaINV in residual breast
499
cancer in patients after neoadjuvant treatment(19). We hypothesized that the chemotherapy-
500
mediated increase in MenaINV expression occurs via macrophage recruitment and subsequent
501
20
macrophage-mediated increase in NF-κB signaling. We tested this hypothesis by depleting the
502
macrophages using clodronate in chemotherapy treated and untreated mice.
503
Briefly, mice bearing HT17 human PDXs or syngenic mouse PyMT tumors were treated
504
with clodronate or control liposomes and either vehicle control (Ctrl) or paclitaxel (Ptx) as
505
outlined in Fig. 6A and Supp Fig. 7A. Upon completion of treatment we compared the fold
506
change in p65 nuclear localization (NF-κB activation) and MenaINV expression in tumors
507
among the treatment groups (Fig. 6B-D and Supp Fig. 7B-D). Paclitaxel treatment significantly
508
increased p65 nuclear localization (NF-κB activation) compared to vehicle control, while
509
treatment with clodronate liposomes not only abrogated this increase, but also decreased
510
nuclear p65 below the baseline of control animals which did not receive paclitaxel (Fig. 6C and
511
Supp Fig. 7C). These findings indicate that macrophages are required for NF-κB activation in
512
both chemotherapy treated and treatment-naïve animals. Furthermore, MenaINV expression in
513
the tumor cells followed the same trend as the NF-κB signaling activation: paclitaxel, compared
514
to control, increased MenaINV expression whereas clodronate abrogated paclitaxel-mediated
515
induction of MenaINV expression as well as lowered MenaINV expression below the baseline
516
observed in chemotherapy naïve animals (Fig. 6D and Supp Fig 7D).
517
To examine the relationship between p65 and MenaINV expressing cells, we measured
518
whether tumor cells expressing MenaINV also express p65 in the same cell, and found that
519
almost 90% of the MenaINV-hi expressing tumor cells also express p65, regardless of treatment
520
(Fig 6E). Of the tumor cells which express both p65 and MenaINV-hi, we found that the majority
521
(~65-70%) of tumor cells express p65 in the cytoplasm, indicating that NF-κB signaling is not
522
constitutively activated in these tumor cells under any treatment condition. (Fig. 6F).
523
To determine if active (nuclear p65) NF-κB signaling was associated with MenaINV
524
expression, we measured the average fold change in MenaINV expression in cells where p65
525
was localized in the nucleus in treated mice compared to control mice (Fig. 6G). We found in
526
the paclitaxel treatment group (where we had previously found the most robust NF-κB signaling
527
21
activation), MenaINV was more highly expressed when p65 was nuclear, whereas in the
528
clodronate treatment group (which has the lowest NF-κB activation), there was decreased
529
MenaINV expression associated with nuclear p65. These results indicate that in the treatment
530
group where NF-κB signaling is most robust and sustained, there is a concomitant upregulation
531
of MenaINV expression in tumor cells. Similar results were obtained using the PyMT model of
532
breast cancer (Supp Fig 7E). Taken together, these data demonstrate that macrophages are
533
required for NF-κB activation and associated MenaINV expression in vivo in both
534
chemotherapy-treated and chemotherapy naïve animals.
535
536
Discussion
537
We discovered here that the specific mechanism by which tumor cells acquire swift and
538
sustained expression of the metastasis-inducing protein, MenaINV, is via macrophage-mediated
539
co-operative NF-κB and Notch1 signaling. Although previously found to be involved in the
540
induction of MenaINV expression in response to macrophage and tumor cell contact, Notch1
541
signaling alone was unable to induce MenaINV expression (Fig. 7A). However, we determined
542
that MenaINV can be induced by tumor-associated macrophages directly through macrophage-
543
mediated activation of NF-κB, increasing expression of MenaINV by 1.5-fold (Fig. 7B).
544
MenaINV expression can be further enhanced to 2.5-fold when Notch1 is activated in addition to
545
NF-κB in tumor cells. Mechanistically, activation of Notch1 signaling in tumors cells by Jagged1-
546
expressing macrophages leads to prolonged nuclear retention of the NF-κB transcription factor
547
p65, and subsequent increase of MenaINV expression in tumor cells (Fig. 7C). Importantly, we
548
determined that the mechanism by which chemotherapy treatment enhances MenaINV
549
expression occurs also through increased macrophage recruitment and subsequent cooperative
550
Notch1 and NF-κB signaling in tumor cells.
551
The precise mechanism of MenaINV induction in breast cancer cells shown here are of
552
great translational significance for patients with metastases as previous studies have
553
22
demonstrated that only cancer cells expressing the MenaINV isoform of the actin regulatory
554
protein Mena, are capable of intravasating and metastasizing to secondary sites(12-15, 17, 19).
555
MenaINV is required for formation of mature invadopodia which increase invasive and
556
transendothelial migration capabilities of cancer cells(10, 18, 45). In addition, it was found that
557
the expression of MenaINV occurs in macrophage-rich areas associated with TMEM doorways,
558
increasing the likelihood that the MenaINV-expressing tumor cells will intravasate at TMEM
559
doorways(36). Furthermore, MenaINV expressing tumor cells show a dramatically increased
560
extravasation activity at distant sites, such as the lung, leading to highly efficient metastatic
561
seeding(16).Therefore, discovering the mechanisms by which MenaINV expression is increased
562
is important for understanding and targeting metastatic dissemination, which can occur not only
563
from primary tumors, but also from metastatic foci resulting in overwhelming metastatic burden
564
and patient demise(46-49).
565
Previous work demonstrated that macrophages induce MenaINV expression in tumor
566
cells through Notch1 signaling(10). However, the Notch intracellular domain (NICD) which is
567
cleaved from the intracellular portion of the receptor upon Notch1 activation, does not have DNA
568
binding activity but acts as a transcriptional co-activator, along with MAML1 and RBP-J (CSL),
569
to activate transcription of genes with RBP-J binding sites(50). We found that are no RBP-J
570
binding sites within the ENAH promoter, and while we did not look at distant enhancer elements
571
in this study, nonetheless, we determined that Notch1 could not induce MenaINV transcription
572
directly. However, we and others, found κB sites within the ENAH promoter, conserved from
573
mouse to human(21). Intriguingly, Notch1 and NF-κB signaling can crosstalk to enhance
574
signaling of both signaling pathways(22-24). Our findings of macrophage-mediated direct
575
induction of MenaINV expression by NF-κB, and indirect by Notch1 is supported by the fact that
576
macrophages can provide stimuli to induce both Notch1 (Jagged1) and NF-κB (TNFα) signaling.
577
Shin et al. reported that NICD can bind to the NF-κB transcription factor, p65, which
578
blocks p65 export from the nucleus, leading to sustained NF-κB signaling(24). Moreover, Field
579
23
et al., found that an initial NF-κB signaling surge, followed by activation of a second NF-κB-
580
independent signaling pathway, can lead to enhanced transcription of NF-κB target genes, and
581
even increased levels of alternative transcripts(51). Consistent with these observations, we also
582
found that Notch1 induces a prolonged nuclear retention of p65 and a subsequent surge in the
583
expression of the MenaINV isoform of Mena, indicating that prolonged nuclear retention,
584
through some still undiscovered mechanism, may affect alternative splicing. Intriguingly, there
585
are also several p300 binding sites within the ENAH promoter which are binding sites for
586
histone acetyltransferases. p65 has been found to promote strong activation of gene
587
transcription following engagement with p300 and histone acetyltransferase activity(52).
588
Indeed, Notch1-mediated prolonged nuclear retention of p65 can explain our data
589
showing that although NF-κB alone can induce MenaINV expression (1.5-fold), the level of
590
induction is below the one achieved by macrophage-cancer cell contact (5-fold). The most
591
robust activation of MenaINV expression was observed when Notch1 and NF-κB signaling were
592
co-activated, by either macrophages or by specific Notch1 and NF-κB activating stimuli.
593
Furthermore, NF-κB inhibitors blocked macrophage-mediated induction of MenaINV expression
594
confirming that NF-κB signaling causes the initial increase in MenaINV expression, but that
595
Notch1 signaling activation leads to the sustained NF-κB activity needed for the robust
596
MenaINV expression.
597
Overall, these discoveries have important clinical implications because they indicate that
598
increased intratumoral macrophage density, as encountered in certain clinical scenarios
599
including high macrophage densities associated with TMEM doorways or inflammatory breast
600
cancer, may affect disease progression. For example, several groups have shown that
601
chemotherapy administration increases the density of tumor associated macrophages
602
(TAMs)(19, 53, 54) and TMEM doorways(19). Thus, chemotherapy given pre-operatively to
603
patients with more advanced disease may lead to an increase in intratumoral macrophage
604
density and tumor cell dissemination via TMEM doorway activity(19). If chemotherapy fails to
605
24
eradicate the tumor completely, an increased density of TAMs may subsequently enhance NF-
606
κB signaling, which combined with Notch1 signaling, will increase MenaINV expression in
607
residual tumor cells (Fig. 7). Our data explain the previously observed increase in MenaINV
608
expression in the residual disease of some breast cancer patients who were treated with pre-
609
operative (neoadjuvant) chemotherapy(19). Since chemotherapy is also given to patients with
610
metastatic disease, one may speculate that macrophage recruitment and subsequent increase
611
in MenaINV expression may occur in metastatic nodules as well. Moreover, it has been reported
612
that chemotherapy not only increases MenaINV expression but also increases the density of
613
TMEM doorways(19, 55) and potentially increases tumor cell dissemination via the blood
614
circulation. Indeed, a recent study indicates that neoadjuvant chemotherapy in patients with
615
early breast cancer leads to an increase of disseminated tumor cells in the bone marrow and
616
subsequent worse overall survival(56). Therefore, the dismal five-year survival rate for breast
617
cancer patients with metastatic disease of approximately 26% may be due to chemotherapy-
618
induced cancer cell dissemination and increased cancer burden(54, 55, 57-60).
619
Paclitaxel is known to increase NF-κB signaling activation directly through binding to
620
TLR4 receptors(61). We report here an additional mechanism of chemotherapy-mediated
621
activation of NF-κB. This mechanism includes paclitaxel-mediated macrophage recruitment
622
which leads to sustained NF-κB activation, potentially through Notch1. Given that MenaINV is
623
critical for tumor cell invadopodium activation (which is involved in migration, invasion, and
624
intravasation), the increase in NF-κB signaling activation and associated MenaINV expression
625
following paclitaxel treatment observed here could explain previous studies that found that
626
paclitaxel treatment increases circulating tumor cells (CTCs)(19).
627
The common therapies for advanced cancers, in addition to neoadjuvant chemotherapy,
628
may include radiation. Interestingly, ionizing radiation is known to increase NF-κB signaling
629
which may lead to NF-κB-mediated radiation resistance(62, 63). Thus, different treatment
630
modalities for advanced cancer, while decreasing tumor mass, may inadvertently induce pro-
631
25
metastatic changes in tumor microenvironment. These changes are characterized by
632
macrophage recruitment, increased TMEM doorway density and activity(19), enhancement of
633
NF-κB and Notch1 signaling in a subset of tumor cells leading to MenaINV-expression in cancer
634
cells, followed by cancer cell dissemination through TMEM doorways, and ultimately increased
635
metastatic burden.
636
The clinical use of Notch1 and NF-κB inhibitors were abandoned in the treatment of solid
637
carcinoma due to toxicity when used systemically(64-66). Though the NF-κB signaling inhibitor
638
bortezomib, a proteosome inhibitor, has been approved for treatment of multiple myeloma in
639
patients who have failed two prior lines of therapy(67), there have not been many other
640
successful uses of these inhibitors. As our knowledge and drug discovery platforms have
641
improved over the last decades, it is important to revisit more specific inhibitors of Notch1 and
642
NF-κB pathways as these signaling pathways may be enhanced by our current standard of care
643
treatments.
644
645
Conclusions
646
In summary, we have found that macrophages enhance expression of MenaINV, a pro-
647
metastatic isoform of Mena, in breast cancer cells through Notch1-mediated prolongation of NF-
648
κB activation. This macrophage-mediated sustained NF-κB signaling is seen in vivo and is
649
enhanced by neoadjuvant chemotherapy. Thus, these findings underscore the need to further
650
investigate combining inhibitors of Notch1 and NF-κB with chemotherapy to decrease
651
chemotherapy-induced cancer cell dissemination and prolong survival of patients with advanced
652
breast cancer.
653
26
Declarations:
654
Ethics approval and consent to participate
655
All procedures were conducted in accordance with National Institutes of Health regulations and
656
approved by the Albert Einstein College of Medicine Animal Use Committee.
657
658
Consent for publication
659
Not applicable
660
661
Availability of data and materials
662
Data sharing is not applicable to this article as no datasets were generated or analysed during
663
the current study
664
665
Competing interests
666
The authors declare that they have no competing interests
667
668
Funding
669
This study was supported by grants from the NIH (R01 CA255153, F32 CA243350, K99
670
CA237851, an IRACDA fellowship, K12 GM102779), SIG OD019961, the Gruss-Lipper
671
Biophotonics Center, the Integrated Imaging Program, The Evelyn Gruss-Lipper Charitable
672
Foundation, and The Helen & Irving Spatz Foundation.
673
674
Authors’ contributions
675
Conceptualization - CLD, MHO, JSC, DE
676
Methodology - CLD, GSK, XC
677
Formal Analysis - CLD
678
Software – DE
679
27
Investigation – CLD, GSK, XC, VPS
680
Writing - CLD, JSC, DE, MHO
681
Funding Acquisition – CLD, JSC, MHO, DE, GSK
682
Supervision - MHO, JSC, DE
683
All authors read and approved the final manuscript.
684
685
Acknowledgements
686
We thank members of the Condeelis, Oktay, Entenberg, Cox, Segall, and Hodgson laboratories
687
for helpful discussions. This study was supported by grants from the NIH (R01 CA255153, F32
688
CA243350, K99 CA237851, an IRACDA fellowship, K12 GM102779), SIG OD019961, the
689
Gruss-Lipper Biophotonics Center, The Integrated Imaging Program, The Evelyn Gruss-Lipper
690
Charitable Foundation, and The Helen & Irving Spatz Foundation.
691
28
REFERENCES
692
1.
Zenklusen D, Larson DR, Singer RH. Single-RNA counting reveals alternative modes of
693
gene expression in yeast. Nat Struct Mol Biol. 2008;15(12):1263-71.
694
2.
Kalluri R, Weinberg RA. The basics of epithelial-mesenchymal transition. J Clin Invest.
695
2009;119(6):1420-8.
696
3.
Su S, Liu Q, Chen J, Chen J, Chen F, He C, et al. A positive feedback loop between
697
mesenchymal-like cancer cells and macrophages is essential to breast cancer metastasis. Cancer
698
Cell. 2014;25(5):605-20.
699
4.
Karagiannis GS, Poutahidis T, Erdman SE, Kirsch R, Riddell RH, Diamandis EP.
700
Cancer-associated fibroblasts drive the progression of metastasis through both paracrine and
701
mechanical pressure on cancer tissue. Mol Cancer Res. 2012;10(11):1403-18.
702
5.
Zhu Y, Knolhoff BL, Meyer MA, Nywening TM, West BL, Luo J, et al. CSF1/CSF1R
703
blockade reprograms tumor-infiltrating macrophages and improves response to T-cell checkpoint
704
immunotherapy in pancreatic cancer models. Cancer Res. 2014;74(18):5057-69.
705
6.
Shapiro IM, Cheng AW, Flytzanis NC, Balsamo M, Condeelis JS, Oktay MH, et al. An
706
EMT-driven alternative splicing program occurs in human breast cancer and modulates cellular
707
phenotype. PLoS Genet. 2011;7(8):e1002218.
708
7.
Gertler FB, Niebuhr K, Reinhard M, Wehland J, Soriano P. Mena, a relative of VASP
709
and Drosophila Enabled, is implicated in the control of microfilament dynamics. Cell.
710
1996;87(2):227-39.
711
8.
Pino MS, Balsamo M, Di Modugno F, Mottolese M, Alessio M, Melucci E, et al. Human
712
Mena+11a isoform serves as a marker of epithelial phenotype and sensitivity to epidermal
713
growth factor receptor inhibition in human pancreatic cancer cell lines. Clin Cancer Res.
714
2008;14(15):4943-50.
715
9.
Roussos ET, Goswami S, Balsamo M, Wang Y, Stobezki R, Adler E, et al. Mena
716
invasive (Mena(INV)) and Mena11a isoforms play distinct roles in breast cancer cell cohesion
717
and association with TMEM. Clin Exp Metastasis. 2011;28(6):515-27.
718
10.
Pignatelli J, Bravo-Cordero JJ, Roh-Johnson M, Gandhi SJ, Wang Y, Chen X, et al.
719
Macrophage-dependent tumor cell transendothelial migration is mediated by Notch1/MenaINV-
720
initiated invadopodium formation. Sci Rep. 2016;6:37874.
721
11.
Eddy RJ, Weidmann MD, Sharma VP, Condeelis JS. Tumor Cell Invadopodia: Invasive
722
Protrusions that Orchestrate Metastasis. Trends Cell Biol. 2017;27(8):595-607.
723
12.
Patsialou A, Wyckoff J, Wang Y, Goswami S, Stanley ER, Condeelis JS. Invasion of
724
human breast cancer cells in vivo requires both paracrine and autocrine loops involving the
725
colony-stimulating factor-1 receptor. Cancer Res. 2009;69(24):9498-506.
726
13.
Wang W, Mouneimne G, Sidani M, Wyckoff J, Chen X, Makris A, et al. The activity
727
status of cofilin is directly related to invasion, intravasation, and metastasis of mammary tumors.
728
J Cell Biol. 2006;173(3):395-404.
729
14.
Wyckoff J, Wang W, Lin EY, Wang Y, Pixley F, Stanley ER, et al. A paracrine loop
730
between tumor cells and macrophages is required for tumor cell migration in mammary tumors.
731
Cancer Res. 2004;64(19):7022-9.
732
15.
Leung E, Xue A, Wang Y, Rougerie P, Sharma VP, Eddy R, et al. Blood vessel
733
endothelium-directed tumor cell streaming in breast tumors requires the HGF/C-Met signaling
734
pathway. Oncogene. 2017;36(19):2680-92.
735
29
16.
Borriello L, Coste A, Traub B, Sharma VP, Karagiannis GS, Lin Y, et al. Primary tumor
736
associated macrophages activate programs of invasion and dormancy in disseminating tumor
737
cells. Nat Commun. 2022;13(1):626.
738
17.
Roussos ET, Wang Y, Wyckoff JB, Sellers RS, Wang W, Li J, et al. Mena deficiency
739
delays tumor progression and decreases metastasis in polyoma middle-T transgenic mouse
740
mammary tumors. Breast Cancer Res. 2010;12(6):R101.
741
18.
Pignatelli J, Goswami S, Jones JG, Rohan TE, Pieri E, Chen X, et al. Invasive breast
742
carcinoma cells from patients exhibit MenaINV- and macrophage-dependent transendothelial
743
migration. Sci Signal. 2014;7(353):ra112.
744
19.
Karagiannis GS, Pastoriza JM, Wang Y, Harney AS, Entenberg D, Pignatelli J, et al.
745
Neoadjuvant chemotherapy induces breast cancer metastasis through a TMEM-mediated
746
mechanism. Science Translational Medicine. 2017;9(397).
747
20.
Cabrera RM, Mao SPH, Surve CR, Condeelis JS, Segall JE. A novel neuregulin - jagged1
748
paracrine loop in breast cancer transendothelial migration. Breast Cancer Res. 2018;20(1):24.
749
21.
Urbanelli L, Massini C, Emiliani C, Orlacchio A, Bernardi G, Orlacchio A.
750
Characterization of human Enah gene. Biochim Biophys Acta. 2006;1759(1-2):99-107.
751
22.
Osipo C, Golde TE, Osborne BA, Miele LA. Off the beaten pathway: the complex cross
752
talk between Notch and NF-kappaB. Lab Invest. 2008;88(1):11-7.
753
23.
Guan E, Wang J, Laborda J, Norcross M, Baeuerle PA, Hoffman T. T cell leukemia-
754
associated human Notch/translocation-associated Notch homologue has I kappa B-like activity
755
and physically interacts with nuclear factor-kappa B proteins in T cells. J Exp Med.
756
1996;183(5):2025-32.
757
24.
Shin HM, Minter LM, Cho OH, Gottipati S, Fauq AH, Golde TE, et al. Notch1 augments
758
NF-kappaB activity by facilitating its nuclear retention. EMBO J. 2006;25(1):129-38.
759
25.
DiDonato JA, Mercurio F, Karin M. NF-kappaB and the link between inflammation and
760
cancer. Immunol Rev. 2012;246(1):379-400.
761
26.
Greten FR, Eckmann L, Greten TF, Park JM, Li ZW, Egan LJ, et al. IKKbeta links
762
inflammation and tumorigenesis in a mouse model of colitis-associated cancer. Cell.
763
2004;118(3):285-96.
764
27.
Maeda S, Kamata H, Luo JL, Leffert H, Karin M. IKKbeta couples hepatocyte death to
765
cytokine-driven compensatory proliferation that promotes chemical hepatocarcinogenesis. Cell.
766
2005;121(7):977-90.
767
28.
Dajee M, Lazarov M, Zhang JY, Cai T, Green CL, Russell AJ, et al. NF-kappaB
768
blockade and oncogenic Ras trigger invasive human epidermal neoplasia. Nature.
769
2003;421(6923):639-43.
770
29.
van Hogerlinden M, Rozell BL, Ahrlund-Richter L, Toftgard R. Squamous cell
771
carcinomas and increased apoptosis in skin with inhibited Rel/nuclear factor-kappaB signaling.
772
Cancer Res. 1999;59(14):3299-303.
773
30.
Nelson DE, Ihekwaba AE, Elliott M, Johnson JR, Gibney CA, Foreman BE, et al.
774
Oscillations in NF-kappaB signaling control the dynamics of gene expression. Science.
775
2004;306(5696):704-8.
776
31.
Ashall L, Horton CA, Nelson DE, Paszek P, Harper CV, Sillitoe K, et al. Pulsatile
777
stimulation determines timing and specificity of NF-kappaB-dependent transcription. Science.
778
2009;324(5924):242-6.
779
30
32.
Sung MH, Salvatore L, De Lorenzi R, Indrawan A, Pasparakis M, Hager GL, et al.
780
Sustained oscillations of NF-kappaB produce distinct genome scanning and gene expression
781
profiles. PLoS One. 2009;4(9):e7163.
782
33.
Patsialou A, Bravo-Cordero JJ, Wang Y, Entenberg D, Liu H, Clarke M, et al. Intravital
783
multiphoton imaging reveals multicellular streaming as a crucial component of in vivo cell
784
migration in human breast tumors. Intravital. 2013;2(2):e25294.
785
34.
Guy CT CR, Muller WJ. Induction of mammary tumors by expression of polyomavirus
786
middle T oncogene: a transgenic mouse model for metastatic disease. Molecular and cellular
787
biology. 1992;12(3).
788
35.
Patsialou A, Wang Y, Lin J, Whitney K, Goswami S, Kenny PA, et al. Selective gene-
789
expression profiling of migratory tumor cells in vivo predicts clinical outcome in breast cancer
790
patients. Breast Cancer Res. 2012;14(5):R139.
791
36.
Sharma VP, Tang B, Wang Y, Duran CL, Karagiannis GS, Xue EA, et al. Live tumor
792
imaging shows macrophage induction and TMEM-mediated enrichment of cancer stem cells
793
during metastatic dissemination. Nat Commun. 2021;12(1):7300.
794
37.
Shah MM, Zerlin M, Li BY, Herzog TJ, Kitajewski JK, Wright JD. The role of Notch
795
and gamma-secretase inhibition in an ovarian cancer model. Anticancer Res. 2013;33(3):801-8.
796
38.
Milano J, McKay J, Dagenais C, Foster-Brown L, Pognan F, Gadient R, et al. Modulation
797
of notch processing by gamma-secretase inhibitors causes intestinal goblet cell metaplasia and
798
induction of genes known to specify gut secretory lineage differentiation. Toxicol Sci.
799
2004;82(1):341-58.
800
39.
Sjolund J, Johansson M, Manna S, Norin C, Pietras A, Beckman S, et al. Suppression of
801
renal cell carcinoma growth by inhibition of Notch signaling in vitro and in vivo. J Clin Invest.
802
2008;118(1):217-28.
803
40.
Zhang X, Chen Q, Song H, Jiang W, Xie S, Huang J, et al. MicroRNA‑375 prevents
804
TGF‑β‑dependent transdifferentiation of lung fibroblasts via the MAP2K6/P38 pathway. Mol
805
Med Rep. 2020;22(3):1803-10.
806
41.
Clément-Lacroix P, Ai M, Morvan F, Roman-Roman S, Vayssière B, Belleville C, et al.
807
Lrp5-independent activation of Wnt signaling by lithium chloride increases bone formation and
808
bone mass in mice. Proc Natl Acad Sci U S A. 2005;102(48):17406-11.
809
42.
Ma L, Gong H, Zhu H, Ji Q, Su P, Liu P, et al. A novel small-molecule tumor necrosis
810
factor alpha inhibitor attenuates inflammation in a hepatitis mouse model. J Biol Chem.
811
2014;289(18):12457-66.
812
43.
Moellering RE, Cornejo M, Davis TN, Del Bianco C, Aster JC, Blacklow SC, et al.
813
Direct inhibition of the NOTCH transcription factor complex. Nature. 2009;462(7270):182-8.
814
44.
Lin EY, Jones JG, Li P, Zhu L, Whitney KD, Muller WJ, et al. Progression to
815
malignancy in the polyoma middle T oncoprotein mouse breast cancer model provides a reliable
816
model for human diseases. Am J Pathol. 2003;163(5):2113-26.
817
45.
Roussos ET, Balsamo M, Alford SK, Wyckoff JB, Gligorijevic B, Wang Y, et al. Mena
818
invasive (MenaINV) promotes multicellular streaming motility and transendothelial migration in
819
a mouse model of breast cancer. J Cell Sci. 2011;124(Pt 13):2120-31.
820
46.
Borriello L, Condeelis J, Entenberg D, Oktay MH. Breast Cancer Cell Re-Dissemination
821
from Lung Metastases-A Mechanism for Enhancing Metastatic Burden. J Clin Med.
822
2021;10(11).
823
47.
Nguyen DX, Bos PD, Massague J. Metastasis: from dissemination to organ-specific
824
colonization. Nat Rev Cancer. 2009;9(4):274-84.
825
31
48.
Kim MY, Oskarsson T, Acharyya S, Nguyen DX, Zhang XH, Norton L, et al. Tumor
826
self-seeding by circulating cancer cells. Cell. 2009;139(7):1315-26.
827
49.
Zhang W, Bado IL, Hu J, Wan YW, Wu L, Wang H, et al. The bone microenvironment
828
invigorates metastatic seeds for further dissemination. Cell. 2021;184(9):2471-86.e20.
829
50.
Borggrefe T, Oswald F. The Notch signaling pathway: transcriptional regulation at Notch
830
target genes. Cell Mol Life Sci. 2009;66(10):1631-46.
831
51.
Field JT, Martens MD, Mughal W, Hai Y, Chapman D, Hatch GM, et al. Misoprostol
832
regulates Bnip3 repression and alternative splicing to control cellular calcium homeostasis
833
during hypoxic stress. Cell Death Discov. 2018;4:37.
834
52.
Vanden Berghe W, De Bosscher K, Boone E, Plaisance S, Haegeman G. The nuclear
835
factor-kappaB engages CBP/p300 and histone acetyltransferase activity for transcriptional
836
activation of the interleukin-6 gene promoter. J Biol Chem. 1999;274(45):32091-8.
837
53.
Hughes R, Qian BZ, Rowan C, Muthana M, Keklikoglou I, Olson OC, et al. Perivascular
838
M2 Macrophages Stimulate Tumor Relapse after Chemotherapy. Cancer Res. 2015;75(17):3479-
839
91.
840
54.
DeNardo DG, Brennan DJ, Rexhepaj E, Ruffell B, Shiao SL, Madden SF, et al.
841
Leukocyte complexity predicts breast cancer survival and functionally regulates response to
842
chemotherapy. Cancer Discov. 2011;1(1):54-67.
843
55.
Chang YS, Jalgaonkar SP, Middleton JD, Hai T. Stress-inducible gene Atf3 in the
844
noncancer host cells contributes to chemotherapy-exacerbated breast cancer metastasis. Proc
845
Natl Acad Sci U S A. 2017;114(34):E7159-E68.
846
56.
Volmer L, Koch A, Matovina S, Dannehl D, Weiss M, Welker G, et al. Neoadjuvant
847
Chemotherapy of Patients with Early Breast Cancer Is Associated with Increased Detection of
848
Disseminated Tumor Cells in the Bone Marrow. Cancers (Basel). 2022;14(3).
849
57.
Shaked Y. The pro-tumorigenic host response to cancer therapies. Nat Rev Cancer. 2019.
850
58.
Karagiannis GS, Condeelis JS, Oktay MH. Chemotherapy-Induced Metastasis: Molecular
851
Mechanisms, Clinical Manifestations, Therapeutic Interventions. Cancer Res. 2019;79(18):4567-
852
76.
853
59.
Daenen LG, Houthuijzen JM, Cirkel GA, Roodhart JM, Shaked Y, Voest EE. Treatment-
854
induced host-mediated mechanisms reducing the efficacy of antitumor therapies. Oncogene.
855
2014;33(11):1341-7.
856
60.
Shaked Y, Henke E, Roodhart JM, Mancuso P, Langenberg MH, Colleoni M, et al. Rapid
857
chemotherapy-induced acute endothelial progenitor cell mobilization: implications for
858
antiangiogenic drugs as chemosensitizing agents. Cancer Cell. 2008;14(3):263-73.
859
61.
Ran S. The Role of TLR4 in Chemotherapy-Driven Metastasis. Cancer Res.
860
2015;75(12):2405-10.
861
62.
Lee SJ, Dimtchev A, Lavin MF, Dritschilo A, Jung M. A novel ionizing radiation-
862
induced signaling pathway that activates the transcription factor NF-kappaB. Oncogene.
863
1998;17(14):1821-6.
864
63.
Li X, Chen F, Zhu Q, Ding B, Zhong Q, Huang K, et al. Gli-1/PI3K/AKT/NF-kB
865
pathway mediates resistance to radiation and is a target for reversion of responses in refractory
866
acute myeloid leukemia cells. Oncotarget. 2016;7(22):33004-15.
867
64.
Gilmore TD, Herscovitch M. Inhibitors of NF-kappaB signaling: 785 and counting.
868
Oncogene. 2006;25(51):6887-99.
869
65.
Ryeom SW. The cautionary tale of side effects of chronic Notch1 inhibition. J Clin
870
Invest. 2011;121(2):508-9.
871
32
66.
Begalli F, Bennett J, Capece D, Verzella D, D'Andrea D, Tornatore L, et al. Unlocking
872
the NF-κB Conundrum: Embracing Complexity to Achieve Specificity. Biomedicines.
873
2017;5(3).
874
67.
Paramore A, Frantz S. Bortezomib. Nature Reviews Drug Discovery. 2003;2(8):611-2.
875
876
33
Figure Legends
877
Figure 1. Macrophage-mediated induction of MenaINV expression via Notch and NF-κB
878
cooperation. (A) MenaINV mRNA expression in MDA-MB-231 (231) cells co-cultured with or
879
without BAC2.1F macrophages (Mac) and with or without 10ng/ml TNFα for 4 hours. (B)
880
MenaINV mRNA expression in 231 cells co-cultured with or without Macs, NF-κB inhibitor
881
(DHMEQ), or Notch/γ secretase inhibitor (DAPT) for 4 hours. (C) Model of potential Notch1 and
882
NF-κB signaling “crosstalk” leading to enhanced transcriptional activity at the ENAH (Mena)
883
promoter. The released Notch intracellular domain (NICD - shaded in gray) can bind to the
884
transcription factors in the NF-κB signaling pathway and prevent their nuclear export, allowing
885
for enhanced and sustained transcriptional activation of target genes and alternative splicing.
886
The bars in (A) and (B) represent average fold change MenaINV mRNA compared to control
887
(231 cells), +/-S.D. The data were analyzed using a one-way ANOVA with Tukey’s multiple
888
comparisons test.*p<0.05, **p<0.01, ***p<0.001, ****p<0.0001, n.s.=not significant.
889
890
Figure 2. Notch1 enhances NF-κB signaling by sustaining p65 nuclear localization. (A)
891
Stills from movies at 0, 17, and 240 minutes of MDA-MB-231/GFP-p65 cells treatment with
892
vehicle, or 10 ng/ml human TNFα, or 80 µm Jagged1, or 10 ng/ml TNFα and 80 µm Jagged1. In
893
all treatment groups with TNFα, the cells were treated for an initial 10 minutes, and then TNFα
894
was washed out and replaced with minimal media, or with Jagged1 supplemented media. Cells
895
were imaged live for 240 minutes using an EPI fluorescence microscope for the duration of the
896
treatment, with one image captured every 2.5 minutes. Scale bar = 10 μm. (B) Quantification of
897
normalized GFP-p65 nuclear localization over time from experiment in (A). (C) Western blot
898
showing the amount of p65 in the cytoplasmic and nuclear fractions of wild type MDA-MB-231
899
cells treated for 30 minutes (upper blots) or 4 hours (lower blots) with vehicle, or 10 ng/ml TNFα,
900
or 80 μm Jagged1, or 10 ng/ml TNFα and 80 μm Jagged1 (TNFα + Jagged1). In all treatment
901
groups with TNFα, the cells were treated for an initial 10 minutes, and then TNFα was washed
902
34
out and replaced with minimal media, or with Jagged1 supplemented media. (D) Quantification
903
of western blots in (C) where the nuclear p65 signal was normalized to the lamin A/C signal.
904
The graph shows the fold change in nuclear p65 signal for each treatment relative to the control
905
treatment at both time points. (E) MenaINV mRNA expression in wild type MDA-MB-231 (231)
906
cells treated as in (C) for 1 or 4 hours. Bars in (D) show average fold change MenaINV mRNA
907
expression compared to Control at 1 or 4 hours. Data in (D) were analyzed using a one-way
908
ANOVA with Tukey’s multiple comparisons test. *p<0.05, ****p<0.0001, n.s.=not significant.
909
910
Figure 3. MenaINV expression in tumor cells induced by macrophages depends partially
911
on TNFα mediated NF-κB signaling and Notch1 Jagged1 signaling. (A) MenaINV mRNA
912
expression in MDA-MB-231 (231) cells co-cultured with or without BAC2.1F macrophages (Mac)
913
and with or without C87 (TNFα inhibitor) or SAHM1 (MAML1 inhibitor) for 4 hours. (B) MenaINV
914
mRNA expression in 231 cells co-cultured with or without Macs, TNFα inhibitor (C87), or Jag 1
915
or Jag2 blocking antibodies for 4 hours. Bars in (A) and (B) represent average fold change of
916
MenaINV mRNA expression compared to control cells (231). (C) MenaINV mRNA expression in
917
231 cells co-cultured with wildtype (WT) or Jagged1 knockout BAC2.1F macrophages (Jag1 KO
918
Macs). Bars in (A-C) represent average fold change of MenaINV mRNA expression compared
919
to control cells (A and B: 231; C: 231 +WT Macs). Data were analyzed using a one-way ANOVA
920
with Tukey’s multiple comparisons test. *p<0.05, **p<0.01, n.s.=not significant.
921
922
Figure 4. Macrophage depletion decreases NF-κB signaling and MenaINV expression in a
923
PDX model in vivo. (A) Experimental design for macrophage depletion in patient derived
924
xenografts (PDX) HT17 in SCID mice. i.p. = intraperitoneal. Red arrows indicate treatment days.
925
(B) Immunofluorescence co-staining of HT17 xenografted in nude mice treated as outlined in
926
(A) for the macrophage marker, Iba1 (white), (C) p65 (red), MenaINV (green) and nuclei (blue-
927
DAPI). Blue and orange outlined sections demonstrate examples of what is quantified as
928
35
primarily cytoplasmic (blue) or nuclear (orange) localization of p65. Scale bars = 100μm. (D)
929
Quantification of average fold change in p65 expression from mice in (A). (E) Quantification of
930
average fold change in p65 nuclear localization in PDX HT17 from mice treated as outlined in
931
(A). Only p65 co-localized with the nuclear DAPI signal was quantified. (F) Quantification of
932
average fold change MenaINV expression from PDX HT17 tumors treated as outlined in (A).
933
Data in (D-F) were analyzed using a student’s t-test. **p<0.01.
934
935
Figure 5. Inhibition of Notch1 signaling in vivo decreases activation of NF-κB signaling in
936
MDA-MB-231 orthotropic injection model. (A) Schematic of DAPT treatment of SCID mice
937
bearing orthotopically injected MDA-MB-231 tumor cells. Seven weeks post tumor cell injection
938
mice were treated with 10 mg/kg DAPT or vehicle (corn oil) by i.p. every day for 14 days. Red
939
arrows represent treatment days. (B) Immunofluorescence staining of primary tumor tissues
940
sections for DAPI (nuclear stain, blue), p65 (red) and MenaINV (green). White dotted circles
941
indicate nuclei in the DAPI and p65 channels. Yellow arrow heads denote nuclei with p65
942
positive stain (active NF-κB signaling), white arrowheads indicate nuclei without p65 positive
943
staining (inactive NF-κB signaling). (C) Quantification of p65 localization (%cytoplasmic/nuclear)
944
in tumor tissue from (B). (D) Quantification of average fold change in MenaINV expression
945
compared to control mice from (B). Data in (C) and (D) were analyzed using a student’s t-test.
946
*p<0.05, **p<0.01.
947
948
Figure 6. Chemotherapy treatment enhances NF-κB activation and MenaINV expression
949
through macrophage recruitment in patient xenograft model. (A) Experimental design of
950
chemotherapy and clodronate treatments in patient derived xenografts (PDX) HT17 in SCID
951
mice. i.p. = intraperitoneal, i.v. = intravenous. (B) Immunofluorescence staining of primary
952
breast tumor tissues from mice treated as outlined in (A) with DAPI (nuclear stain, blue), and
953
antibodies recognizing p65 (red), and MenaINV (green). Blue and orange outlined sections are
954
36
expanded below and demonstrate examples of what is quantified as primarily cytoplasmic (blue)
955
or nuclear (orange) localization of p65 in HT17 tumor tissue. (C) Quantification of average fold
956
change in p65 nuclear localization in treated primary tumors from (A) stained for p65 and DAPI.
957
Only p65 which co-localized with the nuclear DAPI signal was quantified. (D) Quantification of
958
average fold change in MenaINV expression in treated primary tumors from (A). (E)
959
Quantification of the percentage of MenaINV-hi expressing tumor cells which also co-expressed
960
p65 (regardless of cellular compartment localization), in primary tumor cells from treatments in
961
(A). (F) Quantification of the localization (% cytoplasmic/nuclear) of p65 in MenaINV-hi
962
expressing tumor cells from primary tumor cells treated in (A). (G) Quantification of average fold
963
change MenaINV expression associated with nuclear p65 staining of primary tumors from (A)
964
stained for MenaINV. Data in (C, D, and G) were analyzed using a one-way ANOVA with
965
Tukey’s multiple comparisons test. *p<0.05, **p<0.01, ***p<0.001, n.s.=not significant.
966
967
Figure 7. MenaINV expression in cancer cells is induced by macrophage-mediated co-
968
operative NF-κB and Notch1 signaling. Juxtacrine and paracrine signaling between
969
macrophages and tumor cells activate Notch1 and NF-κB pathways which co-operate to induce
970
MenaINV expression in cancer cells. (A) Notch1 signaling alone does not induce MenaINV
971
expression in tumor cells. (B) NF-κB signaling, activated by TNFα binding to the TNFR1
972
receptor, causes nuclear translocation of the transcription factor p65 and a 1.5-fold increase in
973
MenaINV expression. (C) Notch1 and NF-κB signaling crosstalk to increase MenaINV
974
expression further to 2.5-fold. Notch1 intracellular domain (NICD) enhances nuclear retention of
975
NF-κB transcription factor p65 leading to sustained NF-κB signaling and induction of MenaINV
976
expression. This mechanism of MenaINV induction is present in vivo and it explains previously
977
observed increase in MenaINV expression upon in chemotherapy treatment(19). This detailed
978
understanding of MenaINV induction in clinically relevant scenarios is needed for future
979
37
development of combination therapies to improve survival of patients with breast cancer. Figure
980
created with BioRender.com.
981
231
231
+ M ac
231
+ TN F α
231
+ M ac
+ TN F α
0
2
4
6
F o ld C h a n g e M e n a IN V
m R N A E x p re s s io n
A
****
****
n.s.
*
****
****
Macrophage
Tumor Cell
Notch
Jag1
NICD
RBPJ
NRE
NICD
κB
P
p50
p65
TNFα
TNFR
NF-κB
Nucleus
NICD
γ-secretase
**** ****
****
n.s.****
n.s.
2 3 1
2 3 1
+ M a c
2 3 1
+ M a c
+ D H M E Q
2 3 1
+ M a c
+ D A P T
2 3 1
+ M a c
+ D H M E Q
+ D A P T
0
2
4
6
F o ld C h a n g e M e n a IN V
m R N A E x p re s s io n
**
***
n.s.
B
C
Figure 1.
A
t=0
t=17’
t=240’
Jagged1
Control
TNFα + Jagged1
TNFα
MDA-MB-231/GFP-p65
B
C
75
37
GAPDH
50
Nuclear
Jagged1
TNFα
Control
Cytoplasmic
TNFα + Jagged1
Jagged1
TNFα
Control
TNFα + Jagged1
75
37
GAPDH
30 min
4h
75
MDA-MB-231 cells
Lamin A/C
75
Lamin A/C
ladder
D
0
0.5
1
1.5
2
2.5
3
Control
TNF
Jag
Jag+TNF
Fold Change MenaINV
mRNA Expression
1hr
4hr
*
****
α
α
n.s.
MW
(kDa)
p65
p65
Figure 2.
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Control
TNFα
Jag1
TNFα+
Jag1
Fold Change Nuclear p65
Localization Compared to Control
30min
4hr
TNFα +
Jag1
Jag1
TNFα
E
*
****
****
****
A
2 3 1
2 3 1
+ M a c
2 3 1
+ M a c
+ C 8 7
2 3 1
+ M a c
+ S A H M 1
2 3 1
+ M a c
+ C 8 7
+ S A H M 1
0
1
2
3
4
5
F o ld C h a n g e M e n a IN V
m R N A E x p re s s io n
***
n.s.
**
n.s.
***
*
***
**
*
***
B
2 3 1
2 3 1 +
M a c
2 3 1 +
M a c +
C 8 7
2 3 1 +
M a c +
J a g 1 A B
2 3 1 +
M a c +
J a g 2 A B
2 3 1 +
M a c +
J a g 1 + 2 A B
0
1
2
3
4
5
F o ld C h a n g e M e n a IN V
m R N A E x p re s s io n
*
n.s.
n.s.
***
**
****
*
C
Figure 3.
231
+ W T M ac s
231
+ J ag 1 K O M ac s
0 .0
0 .5
1 .0
1 .5
F o ld C h a n g e M e n a IN V
m R N A E x p re s s io n
SCID
HT17
6
7
8
Post-transplantation (Weeks)
0
Clodronate or control (PBS) liposomes 200 µl i.p.
Sacrifice
1
HT17 xenograft
A
B
Clodronate
Control
Iba1/DAPI
Clodronate
Control
MENAINVDAPI/p65
HT17
C
D
***
E
***
F
**
C trl
C lo d
0 .0
0 .5
1 .0
1 .5
A ve ra g e F o ld C h a n g e
p 6 5 E x p re s s io n
C trl
C lo d
0 .0
0 .5
1 .0
1 .5
2 .0
A ve ra g e F o ld C h a n g e
N u cle a r p 6 5 L o ca liza tio n
C trl
C lo d
0 .0
0 .5
1 .0
1 .5
2 .0
A ve ra g e F o ld C h a n g e
M e n a IN V E x p re ss io n
Figure 4.
HT17
SCID
MDA-MB-231
9
7
8
Post-injection (Weeks)
Collect tumors and
metastatic sites
DAPT (10 mg/kg) or corn oil i.p.
A
p65
DAPI
Control
DAPT
Merge
MenaINV
B
231-SCID
**
*
C
D
0%
20%
40%
60%
80%
100%
Control
DAPT
p65 Localization
Cytoplasmic
Nuclear
C o ntro l
D A P T
0 .0
0 .5
1 .0
1 .5
2 .0
F o ld C h a n g e
M e n a IN V E xp re ss io n
Figure 5.
Paclitaxel or vehicle 10 mg/kg i.v.
SCID
HT17
6
7
8
Post-transplantation (Weeks)
0
Clodronate or control (PBS) liposomes 200 µl i.p.
Sacrifice
1
HT17 xenograft
A
B
C trl
P tx
C trl
P tx
0
1
2
3
4
5
A ve ra g e F o ld C h a n g e
N u cle a r p 6 5 L o ca liza tio n
L ip o so m e s
C lo dronate
C
***
***
n.s.
*
C trl
P tx
C trl
P tx
0
1
2
3
4
A ve ra g e F o ld C h a n g e
M e n a IN V E x p re ss io n
L ip o so m e s
C lo d ro n a te
D
****
****
*
*
C trl
P tx
C trl
P tx
0
1
2
3
4
A v e ra g e F o ld C h a n g e M e n a IN V
E x p re s s io n A s so c ia te d w ith N u c le a r p 6 5
L ip o so m e s
C lo d ro n a te
C trl
P tx
C trl
P tx
0
2 0
4 0
6 0
8 0
1 0 0
% M e n a IN V -h i C e lls
E x p re s s in g p 6 5
L ip o so m e s
C lo d ro n a te
E
G
F
0%
20%
40%
60%
80%
100%
120%
Ctrl
Ptx
Ctrl
Ptx
% p65 Localization
Cytoplasmic
Nuclear
**
*
n.s.
Figure 6.
Clodronate Liposomes
Paclitaxel
Paclitaxel
PBS Liposomes
Vehicle
Vehicle
MenaINV DAPI/p65
DAPI p65
HT17
Figure 7.
A
B
C
| 2023 | Cooperative NF-κB and Notch1 signaling promotes macrophage-mediated MenaINV expression in breast cancer | 10.1101/2023.01.03.522642 | [
"Duran Camille L.",
"Karagiannis George S.",
"Chen Xiaoming",
"Sharma Ved P.",
"Entenberg David",
"Condeelis John S.",
"Oktay Maja H."
] | null |
1
Title: FK506-binding protein FklB is involved in biofilm formation through its peptidyl-prolyl
isomerase activity
Authors and affiliations: Chrysoula Zografou, Maria Dimou, Panagiotis Katinakis
Laboratory of General and Agricultural Microbiology, Faculty of Crop Science, Agricultural University of Athens, Iera Odos 75,
11855 Athens, Greece
Abstract
FklB is a member of the FK506-binding proteins (FKBPs), a family that consists of five genes in Escherichia
coli. Little is known about the physiological and functional role of FklB in bacterial movement. In the present
study, FklB knock-out mutant ΔfklB presented an increased swarming and swimming motility and biofilm
formation phenotype, suggesting that FklB is a negative regulator of these cellular processes.
Complementation with Peptidyl-prolyl isomerase (PPIase)-deficient fklB gene (Y181A) revealed that the
defects in biofilm formation were not restored by Y181A, indicating that PPIase activity of FklB is
modulating biofilm formation in E. coli. The mean cell length of ΔfklB swarming cells was significantly
smaller as compared to the wild-type BW25113. Furthermore, the mean cell length of swarming and
swimming wild-type and ΔfklB cells overexpressing fklB or Y181A was considerably larger, suggesting that
PPIase activity of FklB plays a role in cell elongation and/or cell division. A multi-copy suppression assay
demonstrated that defects in motility and biofilm phenotype were compensated by overexpressing sets of
PPIase-encoding genes. Taken together, our data represent the first report demonstrating the involvement
of FklB in cellular functions of E. coli.
Introduction
The previously prevalent view of bacteria development was considered to be the planktonic form of life.
That is, unicellular organisms that grow as individual entities. This view has changed as it has been found
that bacteria, under given conditions, behave as multicellular groups that grow on nutrient-rich surfaces,
secrete a polysaccharide material, through a process called swarming motility. Swarming motility is mainly
driven by rotating flagella, and swarming bacteria generally appear to be elongated as a result of cell
division suppression [1][2]. In contrast to swarming, swimming describes a mode in which cells move within
aqueous environments, not in groups, but independently, by operating their rotating flagella [3]. A
comparable type of multicellular behavior is the biofilm formation where bacteria form sessile communities
and disperse by secreting proteins and surfactants extracellularly [4]. Biofilms are complex systems and
can be composed of multiple species [5]. Environmental conditions and coordinated life cycles can affect
or set off heterogeneity and include, among other, expression of genes and proteins, as well as post-
translational protein modifications (PTMs), that could alter environmental sensing and signal transduction
[6][7]. Various PTMs such as glycosylation, N-terminal modifications and phosphorylation are few of the
2
functional properties of Peptidyl-prolyl cis/trans isomerases (PPIases). PPIases, being ubiquitous among
all organisms, are key regulators of numerous highly important biological processes; they accelerate the
rate of in vitro protein folding and they have the ability to bind proteins and act as chaperones.
Additionally, PPIases catalyze the folding of newly synthesized protein targets, particularly those that have
peptide bonds in the trans conformation. They are also able to alter the structure and conformation of
mature proteins thus affecting their intermolecular interactions [8]. In bacteria and other organisms, there
are three characterized PPIase subfamilies; the Cyclophilins, the FK506-binding proteins (FKBPs), and
the Parvulins. E. coli FKBP family consists of five genes; fkpA, fkpB, fklB, slyD and tig, none of which is
essential for growth [9]. FKBPs are found to be involved in a diverse series of cellular processes such as
cell division [10], stress response regulation and development [11], gene regulation through transcription
and translation [12], and most importantly, virulence and pathogenicity [13].
E. coli FklB (or FKPB22) possesses PPIase activity, exists in solution as a homodimer and shares a
significant homology with the protein Mip (macrophage infectivity potentiator) that is identified in a number
of human pathogenic bacteria, such as Legionella pneumophila, Neisseria gonorrheae, and Chlamydia
trachomatis in the psychrotrophic bacterium Shewanella sp. SIB1 [14] but also in the plant pathogen
Xanthomonas campestris [15]. Shewanella SIB1 FKBP22 is composed of two monomers that are
connected at their N-termini, bearing a V-shaped structure. Within the monomer, a 40-residue long a-helix
separates the N- and C-terminal domain [16]. An almost identical tertiary structure appears to be assumed
by the E. coli FklB [17]. Data suggest that the probable PPIase binding site of SIB1 FKBP22 for a protein
substrate is located at its C-terminal domain. Abrogation of SIB1 FKBP22 PPIase activity did not
significantly affect its chaperone function [18].
This paper describes a new approach to investigate the physiological role and assess the PPIase and
chaperone function of FklB through a series of phenotypic methods. We focused on genetic and
biochemical approaches to assess the swarming and swimming motility, biofilm and cell length phenotypes
in E. coli, caused either by the loss of fklB or by the overexpression of fklB and of PPIase-deficient fklB
(Y181A) gene. We found that deletion of fklB resulted in an enhancement of motility and biofilm, as well
as a decrease of swarming cells’ length. Complementation with fklB gene, in the mutant strain ΔFklB,
suppressed the fklB deletion motility and biofilm phenotype, while overexpression of Y181A suppressed
only the motility phenotype. Overexpression of fklB or Y181A gene exhibited opposite effects on the mean
cell length of swarming and swimming cells. We also used a multi-copy suppression approach to assess
if overexpression of other PPIase-encoding genes may suppress the ΔfklB strain motility and biofilm
phenotypes.
3
Results & Discussion
PPIase and chaperone activity of FklB
Initially, we examined the PPIase activity of FklB with a standard PPIase assay of isomer-specific
proteolysis by chymotrypsin, described by Kofron [19]. Based on the protein alignment of E. coli FklB
(Ec_4207) with the fully characterized human FkpB12 (hfkpB12), we located the FklB’s putative active
sites. We then constructed an active site mutated form of FklB, Y181A, which we used in the PPIase assay
in comparison to the wild-type FklB. N-succinyl-Ala-Ala-Pro-Phe-pnitroanilide was used as a substrate
known to mimic the internal peptidyl-prolyl moiety of proteins containing proline.
We found that Y181A had no measurable isomerase function, whereas the catalytic efficiency of FklB was
1.50 ± 0.0013 (Kcat/Km), suggesting that substitution of Y181 had a significant effect on its activity (Fig.
1A). This result suggests that Y181A, located on its C-terminal domain, is involved in the catalytic function
of the enzyme. Previous findings have demonstrated that the catalytic efficiencies of other mutant forms
of FklB indicate that W157 and F197 are also critically important for the isomerase activity of SIB1 FKBP22
[20].
Next, we sought to determine the chaperone activity of FklB by measuring its ability to suppress the thermal
aggregation of citrate synthase (CS) [21], while also evaluating any effects on it, caused by the induced
active site mutation. CS tends to aggregate at high temperatures because of hydrophobic interactions
between unfolding intermediates and results in the formation of high molecular weight particles. Proteins
with a chaperone function are able to recognize and bind to these unfolding intermediates and therefore
keep their concentration low in solution. We observed that wild-type FklB presents a chaperone function,
whereas the inhibition of the CS aggregation from Y181A appears to increase in proportion to its
concentration (Fig. 1B).
However, we noticed an increased ability of Y181A to inhibit the formation of CS aggregates, in comparison
to the wild-type FklB. This may be an indication that the loss of FklB’s PPIase activity improves its function
as a chaperone.
It has been previously shown that the PPIase and chaperone activity of SIB1 FKBP22 reside in two
structurally unrelated domains, but not necessarily functionally independent domains. Mutations at its
PPIase active site do not critically affect its chaperone function, an indication that SIB1 FKBP22 does not
require PPIase activity for protein folding. However, the authors highlight the importance and requirement
of the chaperone domain for the PPIase activity, as a way of enabling the formation of folding intermediates
[20]. Several other studies have also indicated that the presence of chaperone activity improves PPIase
activity [22] [23]. The substrates are bound to the chaperone site and are subsequently transferred to the
PPIase site, where the peptidyl-prolyl bonds of the proteins are being isomerized. Protein molecules, that
exited the PPIase site with an incorrect peptide-prolyl bond, are re-attached to the chaperone region and
the procedure is repeated [23]. Therefore, it could be assumed that the chaperone activity of FklB may
offset a high PPIase activity and concurrently the loss of PPIase activity might allow structural changes
4
that increase the chaperone activity. Elucidating the structural relationship and association of the two
functions is very important in order to uncover the role of the PPIase family in major cellular processes.
Role of FklB in swarming and swimming motility
The role of FklB was examined under swarming and swimming conditions, by inoculating the center of
swarming (LB-glucose, 0.5% agar) and swimming plates (LB, 0.3% agar) with liquid cultures expressing
and/or lacking the fklB gene (Fig. 2). We found that the mutant ΔFklB strain formed considerably larger
swarming and swimming colonies in comparison to the control strain, BW25113, indicating that the loss of
FklB is responsible for the observed phenotype (Fig. 2A-B). In order to validate that FklB functions as a
swarming and swimming motility repressor we examined the phenotypes of the ΔFklB strain and of the
control strain overexpressing the fklB gene (strain BW25113(FklB)). We found that the ΔFklB(FklB)
reverted the hyper-swarmer or hyper-swimmer phenotype to wild-type, while BW25113(FklB) further
suppressed the phenotype beyond the wild-type levels (Fig. 2A-D). This observation supports our initial
hypothesis that the lack of FklB was the causative factor of the increased swarming and swimming motility.
Growth rates of ΔFklB or BW25113 (FklB) strain liquid cultures were comparable to the control’s,
suggesting that the increased motility phenotype was not attributed to an increased growth.
Subsequently, we checked the involvement of the PPIase activity of FklB protein in swarming and
swimming motility by following the same conditions. Strain ΔFklB(Y181A) did not seem to differ from the
corresponding strain ΔfklB(FklB), even in highest IPTG concentrations, as both displayed no motility on
swarm or swim plates, suggesting that the PPIase activity of FklB is not likely to be involved in the
mechanism. Overall, we found that the expression of FklB and its mutant form, Y181A, is able to restore
the wild-type phenotype at all IPTG concentrations (0.1-0.5 mM). This confirms that the presence of FklB
is indispensable for maintaining a normal phenotype, perhaps through pre- and post-translational
modifications or indirect target protein interactions that control a wide range of cellular processes, including
motility [24] [25]. The negative regulation of swarming and swimming motility in E. coli by certain PPIase
family members was previously shown [26][27]. FkpB proteins are found to be involved in bacterial motility,
for example, an increase in the transcript levels of fklB gene was observed in P. mirabilis swarming cells
[28]. Another example is the GldI protein of the microorganism F. johnsoniae, a lipoprotein homologous to
FKBPs, essential for gliding mobility [29]. Although the structure of almost all FKBP proteins has been
extensively studied, our knowledge about their biological role still remains limited. We already know that
FKBPs catalyze the refolding of peptides preceding proline at polypeptide chains, as well as that all exhibit
some PPIase activity, but there are still several unanswered questions about their physiological role.
FklB is suppressing E. coli’s biofilm formation ability
Swarming motility and biofilm formation relationship seems to be complex and although both conditions
share some common constituents, they greatly differ. Specifically, the use of flagella is necessary for
5
biofilm initiation, but motility is also required for its initiation, as well as dispersion and release of bacteria
[30]. However, it is not clear whether there is an inverse regulation of swarming motility and biofilm
formation, as conflicting data have been published. For example, an increased EPS production suppressed
swarming motility, but enhanced biofilm formation among laboratory isolates [31]. Biofilm formation, as
well as swarming and swimming motility, was suppressed by overexpressing the cyclophilin PpiB.
However, this involvement of PpiB in the biofilm formation phenotype does not involve its prolyl isomerase
activity [32]. Biofilm formation was explored for FklB in order to elucidate the role of its PPIase function in
this multicellular behavior, but also to investigate into its relation to swarming. To this means, we initially
compared the biofilm formed by the control BW25113 and the mutant strain ΔFklB and we noticed that the
ΔFklB strain was capable of a greatly increased biofilm formation (Fig. 3). We hypothesized that the
increased biofilm formation by ΔFklB was attributed to the absence of the FklB protein and in order to
clarify this we tested the biofilm formation under the same conditions of the strain ΔFklB(FklB) as well as
the strain BW25113(FklB). Indeed, we noticed the restoration of the wild-type phenotype, when FklB was
expressed at intermediate IPTG concentrations (0.1-0.25 mM), in the mutant strain (ΔFklB(FklB)) and in
the wild-type strain (BW25113(FklB)). Interestingly, we further detected a biofilm repression phenotype
when FklB was overexpressed (0.5 mM IPTG), either in BW25113(FklB) or in ΔFklB(FklB) strain,
suggesting that FklB bears a key role in biofilm formation (Fig. 3A-B).
Similarly, we tested the strains that overexpress the mutant protein Y181A, ΔFklB(Y181A) and
BW25113(Y181A). The results showed that the biofilm of strain ΔFklB(Y181A), did not differ from the
mutant strain ΔFklB, even in the presence of high levels of IPTG (0.25 and 0.5 mM). The overexpression
of the mutant Y181A did not cause a restoration of the wild-type phenotype. Based on these results, we
can conclude that FklB’s PPIase activity is involved in this multicellular behavior (Fig. 3A).
Additionally, the strain BW25113(Y181A), in the absence or presence of low levels of IPTG (0.1 mM),
showed similar biofilm formation ability to the control BW25113. However, we noticed that even though the
strain BW25113(Y181A) at 0.25 and 0.5 mM IPTG, showed an important decrease in biofilm formation,
that decrease was slightly lower than BW25113 (FklB) ((Fig. 3B). These observations seem to suggest
that FklB has an important role in suppressing the biofilm formation phenotype of E. coli and that its PPIase
activity is indispensable for this involvement.
PPIase family members can functionally replace FklB in swarming, swimming and biofilm cells
Previous research has demonstrated that two members of the PPIase family are functionally linked in yeast
cells. It was found that although these proteins do not bind or catalyze the same peptides, they can
generate conformational changes to substrates [33]. Another study has showed that a parvulin and a FKBP
protein catalyze the cis/trans isomerization of peptide bonds in proteins with great homology [34]. Based
on the above studies, we questioned whether the previously observed phenotypes of ΔFklB mutant strain
could reverse upon expression of members of the PPIase family. To this end, we separately introduced
6
and expressed plasmids that contained each gene belonging to the PPIase family; fkpA, slyD, fkpB, tig,
ppiA, ppiB, surA, ppiC and ppiD into the ΔFklB mutant strain and we compared the ability of each one of
swarming, swimming and biofilm formation (Fig. 4).
Interestingly, we found that the multiple copies of all gene members of the PPIase family (0.5 mM IPTG),
but not the member tig of the FKBP family, could rescue the hyper-swarming phenotype of the ΔFklB
mutant (Fig 4A). The hyper-swarmer phenotype of ΔFklB was restored to wild-type levels even at single
copies of genes fkpA, ppiD, and of course fklB (0 mM IPTG). This observation could be evidential of a
functional overlap between FKBPs and parvulins, in swarming bacteria, perhaps hinting at the cis/trans
isomerization of some common substrates.
The hyper-swimmer phenotype of ΔFklB was rescued upon expression of the majority of PPIases,
excluding the cyclophilins ppiA and ppiB. Every member of the FKBP family was able to complement FklB’s
function in swimming cells even in single copies (0 mM IPTG). There was a significant increase detected
at high expression levels of tig (0.5 mM IPTG), which indicates a unique involvement of the trigger factor
protein in swimming motility (Fig. 4B).
Lastly, we checked the biofilm formation phenotype of the ΔFklB could be abrogated by members of the
PPIase family. We found that the expression of a great number of PPIases was not able to functionally
replace FklB. The members PpiA, PpiB, FkpB, Tig, PpiC, and PpiD did not rescue the increased biofilm of
the mutant strain ΔFklB, at low levels of expression (0 mM and 0.1 mM IPTG). Wild-type biofilm levels
were recovered in the high-copy presence of PpiB, PpiC, and SurA (0.5 mM IPTG) and in the intermediate-
copy presence of FklB and FkpB (0.25 mM IPTG). Interestingly, we identified a biofilm suppression
phenotype after expressing high-levels of FKBP encoding genes, fkpA, fkpB, slyD, and fklB (0.5 mM IPTG)
(Fig. 4C).
The evidence from the above experiments point towards the idea that members of the PPIase family can
compensate for the absence of FklB, in swarmer, swimmer and biofilm E. coli cells. This functional
replacement is even possible at very low copy numbers of PPIases, suggesting that there might be a
substrate regulation pathway shared within the PPIase family. The data also indicate that there is a
stronger physiological function commonality among members of the FKBP family through the regulation of
cellular processes, post-translationally [26].
FklB expression causes cell morphology alterations
Swarmer cells are described as elongated and hyperflagellated cells, that are able to migrate towards the
edge of a swarming plate or a nutrient-rich surface, away from the initial colony [35] [36] [37]. We have
previously examined E. coli cells in a planktonic phase that lack or overexpress the cyclophilin PpiB and
found that in both cases the present an impaired cell division [38].
In this study, we examined the cell morphology of the control BW25113 and of the mutant ΔFklB, as well
as of the strains that overexpress FklB; BW25113(FklB), BW25113(Y181A), ΔFklB(FklB) and
7
ΔFklB(Y181A), during swarming and swimming motility. The expression of plasmids that carried the fklB
gene and its mutant, Y181A, was performed in the presence of 0.1, 0.25, and 0.5 mM IPTG. We
microscopically observed all the above strains after Gram staining and after DAPI staining, using an optical
and a fluorescence microscope, respectively, showing the expression of FklB and Y181A at 0.25 mM IPTG
(Fig. 5, 6).
In swarming cells, we observed that the absence of the fklB gene (strain ΔFklB) did not result in a
differentiated cell phenotype when compared to the control strain (Fig. 5A). However, overexpression of
the fklB gene, in both the wild-type and mutant strains, BW25113(FklB) and ΔFklB(FklB), resulted into a
phenotype characterized by elongated cells that have stopped dividing (Fig. 5A, 6A).
Additionally, a mixed population of normal size cells and cells that were not dividing was noted in the
swarmer cells of strains BW25113(FklB) and ΔFklB(FklB) overexpressing the fklB gene (Fig. 5A, 6A). For
these cells, we noticed an abnormality in the septa formation, that did not allow the separation of the
cellular membrane for cell division.
Regarding the swimming motility, we noted that the phenotype of the mutant strain ΔFklB did not differ
profoundly from the wild-type strain, BW25113. However, overexpression of the fklB gene in both strains,
BW25113(FklB) and ΔFklB(FklB), caused a pronounced cell elongation, in which it appeared that the cell
division had been inhibited. Figure 5 shows that the cells of strain ΔFklB(FklB) formed multiple nucleoids,
therefore we concluded that the replication of the genetic material was being done normally, while the cell
wall and plasma membrane separation were not permitted (Fig. 5B, 6B).
The phenotype in both swarming and swimming cells seemed to reverse after the expression of the mutant
gene, Y181A. Strains BW25113(Y181A) and ΔFklB(Y181A) had a normal cell appearance under swarming
and swimming conditions, that phenotypically corresponded to the wild-type strain. This observation led
us to the conclusion that the cellular elongation of the strains BW25113(FklB) and ΔFklB(FklB) was due to
the increased levels of the FklB protein. Summing up the results, it was concluded that the accumulation
of the FklB protein, in the swarmer and swimmer cells, caused phenotype alterations that were specific to
the increased PPIase activity (Fig. 5, 6).
8
Materials and Methods
E. coli strains, growth conditions and growth rate assay
The bacterial strains and plasmids that were used in this study are described in Table 1. E. coli K-12
BW25113 and single-gene knockout mutants [39] were obtained from the E. coli Genetic Stock Center.
Plasmid pCA24N, as well as plasmids pCA24N containing the PPIase encoding genes were obtained from
the ASKA library of the NARA Institute [40]. Unless stated otherwise, bacteria were cultivated routinely in
LB (Luria–Bertani) agar or broth at 37oC with aeration. When necessary, media were supplemented with
chloramphenicol (25 μg/ml) or kanamycin (25 μg/ml) or ampicillin (100 μg/ml). The specific growth rates of
the E. coli wild-type and mutant strains were determined by measuring the turbidity at 600 nm and C.F.U/ml
for two independent cultures of each strain as a function of time with turbidity values less than 0.9.
Plasmids
The coding sequence of EcFklB (NC_000913.3) was amplified using PCR and E. coli genomic DNA as a
template. The primers used are Ec.b4207.H.F: 5’- CCAGGATCCGACCACCCCAACTTTTGACACC -3’
and Ec.b3349.H.R: 5’- CGCAAGCTTTTAGAGGATTTCCAGCAGTTC -3’. The fragment excised from
amplified EcFklB sequence was cloned between BamHI and HindIII sites of pPROEX-HTa, resulting in
pPROEX-HTa FklB. Y181A point mutation in FklB was engineered using the gene-SOE method described
by
Horton
[41].
The
mutagenic
primers
used
are
Ec.b4207.Y181A.F:
5’-
CCGCAGGAACTGGCAGCTGGCGAGCGCGGCGCA-3’
and
Ec.b4207.Y181A.R:
5’-
TGCGCCGCGCTCGCCATATGCCAGTTCCTGCGG-3’. The primary PCR products were purified and
then used as templates for the second round of PCR. The final product was introduced into pPROEX-HTa,
resulting in pPROEX-HTa Y181A. The nucleotide sequence of the gene encoding the mutant protein was
confirmed by Sanger sequencing.
Heterologous expression of FklB in E. coli and purification of recombinant protein
E. coli BL21(DE3)[ F– ompT gal dcm lon hsdSB(rB- mB-) λ(DE3 [lacI lacUV5-T7 gene 1 ind1 sam7 nin5]
(Novagen, Madison, WI, USA) was used as a host strain for overproduction of a His-tagged form of FklB
or Y181A. Synthesis of recombinant proteins in E. coli BL21 (DE3) cells was initiated by addition of 0.25
mM isopropyl 1-thio-β-D-galactopyranoside (IPTG) when the culture reached OD600 of 0.6 and continued
cultivation for additional 4h at 30°C. Recombinant proteins were purified with Ni-NTA chromatography
(Ni2+-nitrilotriacetate, Qiagen) according to the manufacturer’s instructions. To remove any imidazole and
salts in the collected fractions, fractions were pooled accordingly and dialyzed against 35 mM Hepes buffer
pH8.0 and 70 mM NaCl, for 12 h. Production levels and purity of the recombinant proteins were analyzed
by 15% SDS-PAGE electrophoresis.
9
Motility assays
Overnight cultures of the different strains were grown, standardized to an OD600 of 1.2, and 3 μl used to
stab or spotted at the center of swimming and swarming plates, respectively. The swimming plates were
prepared with 0.3% Fluka agar, 1% Bacto-tryptone, 0.5% Yeast Extract and 1% NaCl. The swarming
motility plates were prepared with 0.5% Fluka agar, 1% Bacto-tryptone, 0.5% Yeast Extract, 1% NaCl and
0.5% glucose. When necessary, media were supplemented with chloramphenicol (25 μg/ml) or kanamycin
(25 μg/ml) and the appropriate amounts of IPTG. The plates were dried for 1-2 h at room temperature
before being inoculated and were scanned after 20 h incubation at 30 °C. Petri dishes were scanned and
the swarming and swimming areas were measured with the imaging software ImageJ. The experiments
were carried out in three replicates.
Biofilm formation assay
The crystal violet biofilm assays were performed as previously described [4]. Briefly, BW25113 and the
fklB mutant strains containing either pPROEX-HTa FklB or pPROEX-HTa Y181A were grown overnight in
LB medium. The overnight cultures were 1:10 diluted in 100 μl of LB medium supplemented, when
necessary, with appropriate concentrations of antibiotics and IPTG, and the biofilm was formed in covered
96-well microtiter dish for 20 h without shaking at 30oC. The cell suspensions were removed and turbidity
was measured at OD600. The plates were washed once with sterile distilled H2O to remove unbound
bacteria and stained with 200 μl crystal violet (0.1% solution) for 20 min. Quantification was conducted by
suspending the crystal violet stained cells in 200 μl of 20% acetone (in ethanol). Total biofilm formation
was normalized by cell growth (turbidity at 600 nm) to avoid overestimating changes due to growth effects.
As controls, BW25113 fklB mutants with empty or pPROEX-HTa or pCA24N were used.
Peptidyl-prolyl cis/trans isomerase enzymatic assay
PPIase activity was tested using a chymotrypsin-coupled PPIase assay [19]. In this assay we measured
the ability of FklB or Y181A to convert the cis isomer of the synthetic oligopeptide substrate N-Suc-Ala-
Leu-Pro-Phe-p-nitroanilide into the trans form. The assay reaction contained 50 mM Hepes buffer pH 8.0
and 100 mM NaCl, 50 μg α-chymotrypsin (dissolved in 1 mM HCl) (Fluka), 25 μM Suc-AAPF-pNA (5 mM
stock dissolved in trifluoroethanol supplemented with 0.45 M LiCl) and the appropriate amount of
recombinant FklB or Y181A. The reaction was monitored at 4°C by the increase in absorbance at 390 nm
(corresponding to the release of p-nitroanilide) using a HITACHI U-2800 spectrophotometer.
Citrate synthase thermal aggregation assay
Thermal denaturation of citrate synthase (0.25 μΜ final concentration, Sigma) was achieved by incubation
at 45°C, in 40 mM Hepes pH: 7.5, for 15-20 min, in the absence or in the presence of additional proteins,
10
as previously described [21]. Aggregation of citrate synthase was measured by monitoring the increase in
turbidity at 500 nm in a HITACHI U-2800 spectrophotometer equipped with a thermostatic cell holder. The
absorbance change recorded is due to the increase in light scattering upon aggregation of citrate synthase.
Protein disulfide isomerase (Sigma) was used in positive control reactions and albumin (Research
Organics) was used in negative control reactions.
11
References
[1] N. Verstraeten et al., “Living on a surface: swarming and biofilm formation,” Trends in Microbiology,
vol. 16, no. 10, pp. 496–506, Oct. 2008.
[2] D. B. Kearns, “A field guide to bacterial swarming motility,” Nat. Rev. Microbiol., vol. 8, no. 9, pp. 634–
644, Sep. 2010.
[3] G. L. Hazelbauer, J. J. Falke, and J. S. Parkinson, “Bacterial chemoreceptors: high-performance
signaling in networked arrays,” Trends Biochem. Sci., vol. 33, no. 1, pp. 9–19, Jan. 2008.
[4] G. A. O’Toole and R. Kolter, “Flagellar and twitching motility are necessary for Pseudomonas
aeruginosa biofilm development,” Mol. Microbiol., vol. 30, no. 2, pp. 295–304, Oct. 1998.
[5] M. E. Davey and G. A. O’toole, “Microbial biofilms: from ecology to molecular genetics,” Microbiol. Mol.
Biol. Rev., vol. 64, no. 4, pp. 847–867, Dec. 2000.
[6] S. W. Singer et al., “Posttranslational modification and sequence variation of redox-active proteins
correlate with biofilm life cycle in natural microbial communities,” ISME J, vol. 4, no. 11, pp. 1398–
1409, Nov. 2010.
[7] Z. Li et al., “Diverse and divergent protein post-translational modifications in two growth stages of a
natural microbial community,” Nat Commun, vol. 5, p. 4405, Jul. 2014.
[8] S. L. Rutherford and C. S. Zuker, “Protein folding and the regulation of signaling pathways,” Cell, vol.
79, no. 7, pp. 1129–1132, Dec. 1994.
[9] C. B. Kang, Y. Hong, S. Dhe-Paganon, and H. S. Yoon, “FKBP family proteins: immunophilins with
versatile biological functions,” Neurosignals, vol. 16, no. 4, pp. 318–325, 2008.
[10] P. Vittorioso, R. Cowling, J. D. Faure, M. Caboche, and C. Bellini, “Mutation in the Arabidopsis
PASTICCINO1 gene, which encodes a new FK506-binding protein-like protein, has a dramatic effect
on plant development,” Mol. Cell. Biol., vol. 18, no. 5, pp. 3034–3043, May 1998.
[11] A. Breiman and I. Camus, “The involvement of mammalian and plant FK506-binding proteins (FKBPs)
in development,” Transgenic Res., vol. 11, no. 4, pp. 321–335, Aug. 2002.
[12] M. Geisler and A. Bailly, “Tête-à-tête: the function of FKBPs in plant development,” Trends Plant Sci.,
vol. 12, no. 10, pp. 465–473, Oct. 2007.
[13] M. Gaestel, “Molecular chaperones in signal transduction,” Handb Exp Pharmacol, no. 172, pp. 93–
109, 2006.
[14] J. U. Rahfeld et al., “Isolation and amino acid sequence of a new 22-kDa FKBP-like peptidyl-prolyl
cis/trans-isomerase of Escherichia coli. Similarity to Mip-like proteins of pathogenic bacteria,” J. Biol.
Chem., vol. 271, no. 36, pp. 22130–22138, Sep. 1996.
[15] N. Zang et al., “Requirement of a mip-like gene for virulence in the phytopathogenic bacterium
Xanthomonas campestris pv. campestris,” Mol. Plant Microbe Interact., vol. 20, no. 1, pp. 21–30, Jan.
2007.
[16] C. Budiman, K. Bando, C. Angkawidjaja, Y. Koga, K. Takano, and S. Kanaya, “Engineering of
monomeric FK506-binding protein 22 with peptidyl prolyl cis-trans isomerase. Importance of a V-
shaped dimeric structure for binding to protein substrate,” FEBS J., vol. 276, no. 15, pp. 4091–4101,
Aug. 2009.
[17] B. Jana, A. Bandhu, R. Mondal, A. Biswas, K. Sau, and S. Sau, “Domain structure and denaturation
of a dimeric Mip-like peptidyl-prolyl cis-trans isomerase from Escherichia coli,” Biochemistry, vol. 51,
no. 6, pp. 1223–1237, Feb. 2012.
[18] C. Budiman, C. Angkawidjaja, H. Motoike, Y. Koga, K. Takano, and S. Kanaya, “Crystal structure of
N-domain of FKBP22 from Shewanella sp. SIB1: dimer dissociation by disruption of Val-Leu knot,”
Protein Sci., vol. 20, no. 10, pp. 1755–1764, Oct. 2011.
[19] J. L. Kofron, P. Kuzmic, V. Kishore, E. Colón-Bonilla, and D. H. Rich, “Determination of kinetic
constants for peptidyl prolyl cis-trans isomerases by an improved spectrophotometric assay,”
Biochemistry, vol. 30, no. 25, pp. 6127–6134, Jun. 1991.
[20] C. Budiman, T. Tadokoro, C. Angkawidjaja, Y. Koga, and S. Kanaya, “Role of polar and nonpolar
residues at the active site for PPIase activity of FKBP22 from Shewanella sp. SIB1,” FEBS J., vol. 279,
no. 6, pp. 976–986, Mar. 2012.
[21] J. Buchner, H. Grallert, and U. Jakob, “Analysis of chaperone function using citrate synthase as
nonnative substrate protein,” Meth. Enzymol., vol. 290, pp. 323–338, 1998.
12
[22] R. P. Jakob, G. Zoldák, T. Aumüller, and F. X. Schmid, “Chaperone domains convert prolyl isomerases
into generic catalysts of protein folding,” Proc. Natl. Acad. Sci. U.S.A., vol. 106, no. 48, pp. 20282–
20287, Dec. 2009.
[23] G. Zoldák and F. X. Schmid, “Cooperation of the prolyl isomerase and chaperone activities of the
protein folding catalyst SlyD,” J. Mol. Biol., vol. 406, no. 1, pp. 176–194, Feb. 2011.
[24] A. Galat, “Peptidylprolyl cis/trans isomerases (immunophilins): biological diversity--targets--functions,”
Curr Top Med Chem, vol. 3, no. 12, pp. 1315–1347, 2003.
[25] T. Inoue, R. Shingaki, S. Hirose, K. Waki, H. Mori, and K. Fukui, “Genome-wide screening of genes
required for swarming motility in Escherichia coli K-12,” J. Bacteriol., vol. 189, no. 3, pp. 950–957, Feb.
2007.
[26] A. Skagia, C. Zografou, E. Vezyri, A. Venieraki, P. Katinakis, and M. Dimou, “Cyclophilin PpiB is
involved in motility and biofilm formation via its functional association with certain proteins,” Genes
Cells, vol. 21, no. 8, pp. 833–851, Aug. 2016.
[27] A. Skagia et al., “Structural and functional analysis of cyclophilin PpiB mutants supports an in vivo
function not limited to prolyl isomerization activity,” Genes Cells, vol. 22, no. 1, pp. 32–44, Jan. 2017.
[28] M. M. Pearson, D. A. Rasko, S. N. Smith, and H. L. T. Mobley, “Transcriptome of swarming Proteus
mirabilis,” Infect. Immun., vol. 78, no. 6, pp. 2834–2845, Jun. 2010.
[29] M. J. McBride and T. F. Braun, “GldI is a lipoprotein that is required for Flavobacterium johnsoniae
gliding motility and chitin utilization,” J. Bacteriol., vol. 186, no. 8, pp. 2295–2302, Apr. 2004.
[30] M. Klausen et al., “Biofilm formation by Pseudomonas aeruginosa wild type, flagella and type IV pili
mutants,” Mol. Microbiol., vol. 48, no. 6, pp. 1511–1524, Jun. 2003.
[31] N. C. Caiazza, R. M. Q. Shanks, and G. A. O’Toole, “Rhamnolipids modulate swarming motility
patterns of Pseudomonas aeruginosa,” J. Bacteriol., vol. 187, no. 21, pp. 7351–7361, Nov. 2005.
[32] A. Skagia et al., “Structure-Function Analysis of the Periplasmic Escherichia coli Cyclophilin PpiA in
Relation to Biofilm Formation,” J. Mol. Microbiol. Biotechnol., vol. 27, no. 4, pp. 228–236, 2017.
[33] F. Fujimori et al., “Crosstalk of prolyl isomerases, Pin1/Ess1, and cyclophilin A,” Biochem. Biophys.
Res. Commun., vol. 289, no. 1, pp. 181–190, Nov. 2001.
[34] Y. Mamane, S. Sharma, L. Petropoulos, R. Lin, and J. Hiscott, “Posttranslational regulation of IRF-4
activity by the immunophilin FKBP52,” Immunity, vol. 12, no. 2, pp. 129–140, Feb. 2000.
[35] C. Allison and C. Hughes, “Bacterial swarming: an example of prokaryotic differentiation and
multicellular behaviour,” Sci Prog, vol. 75, no. 298 Pt 3–4, pp. 403–422, 1991.
[36] R. M. Harshey, “Bees aren’t the only ones: swarming in gram-negative bacteria,” Mol. Microbiol., vol.
13, no. 3, pp. 389–394, Aug. 1994.
[37] L. McCarter and M. Silverman, “Surface-induced swarmer cell differentiation of Vibrio
parahaemolyticus,” Mol. Microbiol., vol. 4, no. 7, pp. 1057–1062, Jul. 1990.
[38] A. Skagia, C. Zografou, A. Venieraki, C. Fasseas, P. Katinakis, and M. Dimou, “Functional analysis of
the cyclophilin PpiB role in bacterial cell division,” Genes Cells, vol. 22, no. 9, pp. 810–824, Sep. 2017.
[39] T. Baba et al., “Construction of Escherichia coli K-12 in-frame, single-gene knockout mutants: the Keio
collection,” Mol. Syst. Biol., vol. 2, p. 2006.0008, 2006.
[40] M. Kitagawa et al., “Complete set of ORF clones of Escherichia coli ASKA library (a complete set of E.
coli K-12 ORF archive): unique resources for biological research,” DNA Res., vol. 12, no. 5, pp. 291–
299, 2005.
[41] R. M. Horton, H. D. Hunt, S. N. Ho, J. K. Pullen, and L. R. Pease, “Engineering hybrid genes without
the use of restriction enzymes: gene splicing by overlap extension,” Gene, vol. 77, no. 1, pp. 61–68,
Apr. 1989.
13
Figures and figure legends
Figure 1. Prolyl isomerase and chaperone activity is lost in mutant Y181A compared to the wild-type protein FklB.
A. PPIase activity of 0.25 uM FklB and Y181A. Mean
values were obtained from three independent replicates
and error bars represent standard errors.
B. Thermal aggregation of citrate synthase in the absence (●) and
presence of (n) 0.25 uM FklB or 0.25 uM Y181A (▲). The results are
representative of three series of measurements carried out with
different preparations of enzymes.
FklB
Y181A
-0.1
0.0
0.1
0.2
0.3
Δrate (min-1)
**
0
200
400
600
800
1000
0.00
0.02
0.04
0.06
0.08
0.10
0.25 µΜ CS
0.25 µΜ plus 2.5 µΜ FklB
0.25 µΜ plus 2.5 µΜ Y181A
Time (sec)
abs (500 nm)
14
Figure 2. Prolyl isomerase activity of FklB causes a suppression of the swarming and swimming phenotype in E.
coli. Swarming (A) and swimming (C) area of the ΔFklB mutant strain that overexpresses FklB and Y181A and
swarming (B) and swimming area (D) of BW25113 that overexpresses FklB and Y181A, compared to the wild-type
BW25113. Mean values were obtained from four independent replicates, and error bars represent standard errors.
Statistical comparisons were made using ANOVA followed by Dunnett’s multiple-comparison test. Asterisks
indicate statistically significant differences (P < 0.05).
BW25113
ΔFklB
ΔFklB (FklB)
ΔFklB (Y181A)
ΔFklB (FklB)
ΔFklB (Y181A)
ΔFklB (FklB)
ΔFklB (Y181A)
ΔFklB (FklB)
ΔFklB (Y181A)
0
5
10
15
20
25
swarming area (cm2)
0 mM IPTG
0.1 mM IPTG
0.25 mM IPTG
0.5 mM IPTG
A
**
BW25113
BW25113 (FklB)
BW25113 (Y181A)
BW25113 (FklB)
BW25113 (Y181A)
BW25113 (FklB)
BW25113 (Y181A)
BW25113 (FklB)
BW25113 (Y181A)
0.0
0.5
1.0
1.5
2.0
2.5
swarming area (cm2)
0 mM IPTG
0.1mM IPTG
0.25 mM IPTG
0.5 mM IPTG
B
***
*
***
BW25113
ΔfklB
ΔfklB (FklB)
ΔfklB (Y181A)
ΔfklB (FklB)
ΔfklB (Y181A)
ΔfklB (FklB)
ΔfklB (Y181A)
ΔfklB (FklB)
ΔfklB (Y181A)
0
10
20
30
40
50
swimming area (cm2)
0 mM IPTG
0.1 mM IPTG
0.25 mM IPTG
0.5 mM IPTG
****
****
*
*
****
C
BW25113
BW25113 (FklB)
BW25113 (Y181A)
BW25113 (FklB)
BW25113 (Y181A)
BW25113 (FklB)
BW25113 (Y181A)
BW25113 (FklB)
BW25113 (Y181A)
0
1
2
3
4
swimming area (cm2)
0mM IPTG
0.1 mM IPTG
0.25 mM IPTG
0.5 mM IPTG
**
**** ****
**** ****
**** ****
D
15
Figure 3. Prolyl isomerase activity of FklB causes a suppression of the biofilm phenotype in E. coli. Biofilm
formation of the ΔFklB mutant strain that overexpresses FklB and Y181A (A) and of the BW25113 that
overexpresses FklB and Y181A (B), compared to the wild-type BW25113. Mean values were obtained from four
independent replicates, and error bars represent standard errors. Statistical comparisons were made using ANOVA
followed by Dunnett’s multiple-comparison test. Asterisks indicate statistically significant differences (P < 0.05).
BW25113
ΔfklB
ΔfklB (FklB)
ΔfklB (Y181A)
ΔfklB (FklB)
ΔfklB (Y181A)
ΔfklB (FklB)
ΔfklB (Y181A)
ΔfklB (FklB)
ΔfklB (Y181A)
0.0
0.1
0.2
0.3
0.4
OD 550nm
0 mM IPTG
0.1 mM IPTG
0.25 mM IPTG
0.5 mM IPTG
****
****
****
**
****
****
A
BW25113
BW25113 (FklB)
BW25113 (Y181A)
BW25113 (FklB)
BW25113 (Y181A)
BW25113 (FklB)
BW25113 (Y181A)
BW25113 (FklB)
BW25113 (Y181A)
0.00
0.05
0.10
0.15
OD 550nm
0 mM IPTG
0.1 mM IPTG
0.25 mM IPTG
0.5 mM IPTG
****
****
****
****
****
****
B
16
Figure 4. Members of the prolyl isomerase family restore the FklB mutant strain phenotypes. Swarming area (A),
swimming area (B), and biofilm formation (C) of ΔFklB and ΔFklB overexpressing each PPIase family member, PpiA,
PpiB, PpiC, FkpA, FkpB, FklB, SlyD, Tig, PpiC, PpiD, SurA in pCA24N vector.
BW25113
ΔFklB
ΔFklB (PpiA)
ΔFklB (PpiB)
ΔFklB (FkpA)
ΔFklB (FkpB)
ΔFklB (FklB)
ΔFklB (SlyD)
ΔFklB (Tig)
ΔFklB (PpiC)
ΔFklB (PpiD)
ΔFklB (SurA)
0
10
20
30
swarming area (cm2)
0 mM IPTG
0.1 mM IPTG
0.5 mM IPTG
****
****
****
***
****
****
****
A
BW25113
ΔFklB
ΔFklB (PpiA)
ΔFklB (PpiB)
ΔFklB (FkpA)
ΔFklB (FkpB)
ΔFklB (FklB)
ΔFklB (SlyD)
ΔFklB (Tig)
ΔFklB (PpiC)
ΔFklB (PpiD)
ΔFklB (surA)
0
5
10
15
20
25
swimming area (cm2)
0 mM IPTG
0.1 mM IPTG
0.5 mM IPTG
****
** **
**
*
*
*
B
BW25113
ΔFklB
ΔFklB (PpiA)
ΔFklB (PpiB)
ΔFklB (FkpA)
ΔFklB (FkpB)
ΔFklB (FklB)
ΔFklB (SlyD)
ΔFklB (Tig)
ΔFklB (PpiC)
ΔFklB (PpiD)
ΔFklB (SurA)
0.00
0.05
0.10
0.15
0.20
OD 550nm
0 mM IPTG
0.1 mM IPTG
0.25 mM IPTG
0.5 mM IPTG
****
****
****
****
****
****
****
****
****
****
****
****
****
********
****
****
****
****
****
********
****
****
****
C
17
Figure 5. Overexpression of FklB, but not Y181A, in ΔFklB causes a cell elongation phenotype in swarming and
swimming E. coli. ΔFklB and ΔFklB overexpressing FklB or Y181A in pCA24N vector taken from swarming (A)
and swimming (B) cells were examined after DAPI (upper row) or Gram (bottom row) staining by light and
fluorescent microscopy and compared to the control, BW25113. Bars represent 10 um.
18
Figure 6. Overexpression of FklB, but not Y181A, in BW25113 causes a cell elongation phenotype in swarming
and swimming E. coli. BW25113 overexpressing FklB or Y181A in pCA24N vector taken from swarming (A)
and swimming (B) cells were examined after DAPI (upper row) or Gram (bottom row) staining by light and
fluorescent microscopy and compared to the control, BW25113. Bars represent 10 um.
| 2020 | FK506-binding protein FklB is involved in biofilm formation through its peptidyl-prolyl isomerase activity | 10.1101/2020.02.01.930347 | [
"Zografou Chrysoula",
"Dimou Maria",
"Katinakis Panagiotis"
] | creative-commons |
1
MicroRNA775 Promotes Intrinsic Leaf Size and Reduces Cell Wall Pectin Level via a
1
Target Galactosyltransferase in Arabidopsis
2
3
He Zhang1, Zhonglong Guo1, Yan Zhuang2, Yuanzhen Suo3, Jianmei Du2, Zhaoxu Gao2, Jiawei
4
Pan1, Li Li1, Tianxin Wang1, Liang Xiao4, Genji Qin1, Yuling Jiao5, Huaqing Cai6, Lei Li1,2,*
5
6
1State Key Laboratory of Protein and Plant Gene Research, School of Life Sciences and School
7
of Advanced Agricultural Sciences, Peking University, Beijing 100871, China
8
2Peking-Tsinghua Center for Life Sciences, Academy for Advanced Interdisciplinary Studies,
9
Peking University, Beijing 100871, China
10
3Biomedical Pioneering Innovation Center, School of Life Sciences and Beijing Advanced
11
Innovation Center for Genomics, Peking University, Beijing 100871, China
12
4College of Biological Sciences and Technology, Beijing Forestry University, Beijing 100083,
13
China
14
5State Key Laboratory of Plant Genomics, Institute of Genetics and Developmental Biology,
15
Chinese Academy of Sciences, and National Center for Plant Gene Research, 100101 Beijing,
16
China
17
6National Laboratory of Biomacromolecules, CAS Center for Excellence in Biomacromolecules,
18
Institute of Biophysics, Chinese Academy of Sciences, Beijing 100101, China
19
20
*Correspondence should be addressed to lei.li@pku.edu.cn
21
2
Abstract
22
Plants possess unique primary cell walls made of complex polysaccharides that play critical roles
23
in determining intrinsic cell and organ size. How genes responsible for synthesizing and
24
modifying the polysaccharides are regulated by microRNAs (miRNAs) to control plant size
25
remains largely unexplored. Here we identified 23 putative cell wall related miRNAs, termed
26
CW-miRNAs, in Arabidopsis thaliana and characterized miR775 as an example. We showed
27
that miR775 post-transcriptionally silences GALT9, which encodes an endomembrane-located
28
galactosyltransferase belonging to the glycosyltransferase 31 family. Over-expression of miR775
29
and deletion of GALT9 significantly enlarged leaf-related organs, primarily owing to increases in
30
cell size. Monosaccharide quantification, confocal Raman imaging, and immunolabelling
31
combined with atomic force microscopy (AFM) revealed that the MIR775A-GALT9 circuit
32
modulates pectin level and elastic modulus of the cell wall. We further showed that MIR775A is
33
directly repressed by the transcription factor ELONGATED HYPOCOTYL 5 (HY5). Genetic
34
analysis confirmed that HY5 is a negative regulator of leaf size and acts through the HY5-
35
MIR775A-GALT9 repression cascade to control pectin level. These results demonstrate that
36
miR775-regulated cell wall remodeling is an integral determinant for intrinsic leaf size in A.
37
thaliana and highlight the need to study other CW-miRNAs for more insights into cell wall
38
biology.
39
40
3
Introduction
41
Precise control of organ size is a fundamental feature of living organisms that results in distinct,
42
species-specific organ sizes and shapes (Bogre et al., 2008; Johnson and Lenhard, 2011; Hong et
43
al., 2018). Genetic analyses in both animals and plants have established that intrinsic organ size
44
is determined by the combinatory effects of cell proliferation and cell expansion (Bogre et al.,
45
2008; Johnson and Lenhard, 2011; Gonzalez et al., 2012; Tumaneng et al., 2012; Hepworth and
46
Lenhard, 2014; Hong et al., 2018). Over the past two decades, an increasingly detailed picture is
47
emerging on cell proliferation control in plants, which involves transcriptional regulators
48
(Mizukami and Fischer, 2000; Powell and Lenhard, 2012; Du et al., 2014), miRNAs (Rodriguez
49
et al., 2010; Schommer et al., 2014; Yang et al., 2018), and the ubiquitin-proteasome pathway
50
(Du et al., 2014). By comparison, our understanding of cell size control in plants is relatively
51
sparse (Ferjani et al., 2007; Hong et al., 2018).
52
Different from metazoan cells, plant cells are enclosed in the cell walls, which locate
53
between the middle lamella and the plasma membrane. To reach the desired size, plant cells rely
54
on the balance between the inner turgor pressure and the extensibility of the cell walls (Cosgrove,
55
2005; Palin and Geitmann, 2012; Hong et al., 2018). During growth and development, cell walls
56
need to be loosened in a highly controlled way to allow nondestructive cell expansion, which
57
might increase cell size by several orders of magnitude (Velasquez et al., 2011; Palin and
58
Geitmann, 2012; Hong et al., 2018). Moreover, being sessile organisms, plants are extremely
59
sensitive to the environment and exhibit a number of plastic responses, which allow them to
60
reliably tune size and shape according to the prevailing environmental conditions (Hepworth and
61
Lenhard, 2014; Hong et al., 2018). For example, in response to shading from neighbors, many
62
plants undergo increased stem and petiole elongation in the well-characterized shade avoidance
63
responses. Therefore, the plant cell wall is critical for determining both the intrinsic organ size
64
and how it is shaped by the environment.
65
Primary plant cell wall is a highly complex and dynamic structure mainly composed of
66
cellulose, hemicelluloses, and pectin (Somerville et al., 2004; Cosgrove, 2005; Somerville, 2006;
67
Palin and Geitmann, 2012). These polysaccharide constituents have different structural and
68
biological roles. Pectin is defined as a group of polysaccharides containing galacturonic acid that
69
acts as gel-forming polymers to cross-link the hemicellulose and cellulose microfibrils
70
(Somerville, 2006; Palin and Geitmann, 2012; Atmodjo et al., 2013). Studies using solid-state
71
4
nuclear magnetic resonance spectroscopy presented compelling evidence for extensive cellulose-
72
pectin contacts but less cellulose-hemicellulose interactions in the cell walls than previously
73
envisaged (Wang et al., 2015), suggesting that pectin plays an underappreciated role in cell wall
74
remodeling.
75
Three major classes of pectin polymers have been identified in the cell wall matrix.
76
These include homogalacturonan (HG), which possesses a backbone of 1,4-linked α-D-
77
galacturonosyluronic acid residues, rhamnogalacturonan I (RG-I), which consists of interspersed
78
α-D-galacturonosyl and rhamnosyl residues with galactosyl and arabinosyl side-chains, and the
79
lesser abundant rhamnogalacturonan II (RG-II) (Harholt et al., 2010; Palin and Geitmann, 2012;
80
Atmodjo et al., 2013). Structural data indicate that these pectic constitutes interconnect with each
81
other in the wall via covalent linkages of their backbones (Atmodjo et al., 2013). Recently,
82
nanoimaging studies have showed that HG in pavement cell walls may assemble into discrete
83
nanofilaments rather than an interlinked network (Haas et al., 2020). It was suggested that local
84
and polarized expansion of the HG nanofilaments could lead to cell enlargement without turgor-
85
driven growth (Haas et al., 2020). However, biosynthesis and modifications of the pectin
86
polysaccharides are highly complicated processes and their roles in cell wall remodeling remain
87
to be fully elucidated. Given that the involved enzymes are likely integral membrane proteins in
88
their active forms and the lack of robust in vivo assays, functional details of the pectin-related
89
genes in regulating intrinsic organ size remain largely unknown (Qu et al., 2008; Harholt et al.,
90
2010; Palin and Geitmann, 2012; Parsons et al., 2012; Atmodjo et al., 2013; Tan et al., 2013; Qin
91
et al., 2017).
92
MiRNAs are an endogenous class of sequence-specific, trans-acting small regulatory
93
RNAs that modulate gene expression mainly at the post-transcriptional level (Voinnet, 2009; Ma
94
et al., 2010; Yang et al., 2012; Rogers and Chen, 2013). In plants, miRNAs are recognized to
95
regulate an enormous collection of target genes that are implicated in numerous biological
96
processes (Voinnet, 2009; Rogers and Chen, 2013; Rodriguez et al., 2016; Guo et al., 2020).
97
Genetic analysis has uncovered that several miRNAs (e.g. miR319, miR396, and miR408)
98
participate in regulating cell proliferation and organ growth (Palatnik et al., 2003; Ori et al.,
99
2007; Rodriguez et al., 2010; Schommer et al., 2014; Zhang et al., 2014; Rodriguez et al., 2016;
100
Pan et al., 2018; Yang et al., 2018). However, no systematic efforts have been reported to identify
101
and functionally study miRNAs pertinent to the regulation of primary cell wall, even though
102
5
hundreds of genes are involved in wall biosynthesis and modifications. We reasoned that
103
elucidation of the regulatory roles of cell wall related miRNAs, termed CW-miRNAs, should
104
help expanding our understanding of how cell wall remodeling contributes to intrinsic organ size
105
adjustment in plants.
106
In the current study, we identified a group of 23 putative CW-miRNAs in A. thaliana.
107
We focused on functional characterization of miR775 as an exemplar CW-miRNA and
108
delineated the HY5-MIR775A-GALT9 repression pathway for modulating cell size and leaf size.
109
Cellular analyses combining monosaccharide quantification, confocal Raman imaging,
110
immunolabelling, and atomic force microscopy (AFM) revealed that this pathway regulates
111
pectin level and elastic modulus of the cell wall. Collectively, these results demonstrated the
112
importance of miRNA-based regulation of cell wall genes in controlling intrinsic organ size.
113
6
Result
114
Identification and Analysis of Putative CW-miRNAs in Arabidopsis
115
To identify CW-miRNAs in A. thaliana, we collected 572 genes annotated as cell wall
116
biosynthesis related and 491 genes encoding proteins enriched in the Golgi apparatus (Parsons et
117
al., 2012). Searching against the 427 annotated miRNAs in A. thaliana, coupling computational
118
prediction with degradome sequencing analysis, we identified 23 putative CW-miRNAs that are
119
predicted to target 78 genes pertinent to primary wall biosynthesis (Figure 1; Supplemental Table
120
1). Using 34 sequenced small RNA populations derived from six different organ types, we found
121
that most of these miRNAs did not show strong organ specific expression pattern (Figure 1B).
122
Together the CW-miRNAs account for 5.4% of all miRNAs annotated in A. thaliana. However,
123
except miR156h that represses a gene encoding a pectin methylesterase inhibitor (Stief et al.,
124
2014), miR773 that negatively regulates callose deposition in response to fungal infection
125
(Salvador-Guirao et al., 2018), and miR827 that involves in phosphate homeostasis (Kant et al.,
126
2011), functions of this cohort of miRNAs have not been investigated.
127
Sequence comparison in representative A. thaliana ecotypes and 13 Brassicaceae
128
species revealed that most (17 or 73.9%) CW-miRNAs are only found in A. thaliana (Figure 1A).
129
For example, miR775 was among the first batch of non-conserved miRNAs annotated in A.
130
thaliana (Rajagopalan et al., 2006). We found that miR775 is highly conserved in A. thaliana
131
ecotypes but absent in A. lyrata and A. halleri (Supplemental Figure 1). Consistent with previous
132
reports (Felippes et al., 2008), we found that the closest pre-miR775a homolog in A. lyrata
133
misses the mature miR775 sequence (Supplemental Figure 1A) and could not fold into the stem-
134
loop secondary structure typical for miRNA precursors (Supplemental Figure 1B). These results
135
suggest that miR775 has evolved specifically in A. thaliana after its divergence from the
136
common ancestor of the Arabidopsis genus.
137
On the other hand, 75 of the 78 (96.2%) predicted target genes for the CW-miRNAs
138
have apparent orthologs in the Brassicaceae. GALT9, the predicted target gene for miR775,
139
encodes a galactosyltransferase belonging to the carbohydrate-active glycosyltransferase 31
140
(Supplemental Figure 2). Sequence alignment revealed that the predicted miR775 binding site in
141
GALT9 contains five heterogeneous nucleotides across the examined Brassicaceae species
142
(Figure 1C), more frequent than the surrounding sequences (Figure 1D). The five variable
143
nucleotides have formed eight polymorphic combinations in the examined Brassicaceae species
144
7
(Figure 1E). Among these and possible paralogs in A. thaliana, the miR775 binding site in
145
GALT9 exhibited the highest MFE/MED ratio (Supplemental Figure 2B), which is the ratio
146
between the minimum free energy (MFE) of a predicted miRNA:target duplex to the minimum
147
duplex free energy (MED) of the miRNA bound to a fully complementary sequence, an
148
quantitative indicator for likelihood of miRNA targeting (Alves et al., 2009). These results
149
indicate that complementarity of GALT9 to miR775 was selected in A. thaliana.
150
151
Molecular Validation of GALT9 as a MiR775 Target
152
To validate GALT9 as a miR775 target, we first performed the 5’ RNA ligation mediated-rapid
153
amplification of cDNA ends (5’ RLM-RACE) assay (Llave et al., 2002). The detected 5’ ends of
154
truncated GALT9 transcript locate preferentially at the 14th and 15th nucleotides within the region
155
complementary to miR775, counting from the 5’ end of miR775 (Figure 2). While this result
156
supports miR775-guided GALT9 cleavage, the detected transcript ends deviated by about four
157
nucleotides from the conventional cleavage site between the 10th and 11th nucleotides of
158
complementarity (Llave et al., 2002; German et al., 2009). We therefore performed degradome
159
sequencing for further analysis. For comparison with the wild type, we generated miR775-
160
overexpressing plants (MIR775A-OX) in which the enhanced Cauliflower Mosaic Virus 35S
161
promoter was used to drive pre-miR775a expression (Supplemental Figure 3). From the
162
degradome sequencing data, we retrieved reads mapped to the predicted miR775 binding site in
163
GALT9, which were enriched in MIR775A-OX relative to the wild type (Figure 2B). Closer
164
inspection revealed that the enriched reads were not confined to a single nucleotide but
165
concentrated in a region several nucleotides downstream of the 10th position relative to the 5’ end
166
of miR775 (Figure 2C). These results are consistent with the 5’ RLM-RACE data (Figure 2A) to
167
support miR775-dependent cleavage of the GALT9 transcript at unconventional sites.
168
Next, we tested whether miR775 is sufficient for repressing GALT9 using the dual
169
firefly luciferase (LUC) and Renilla luciferase (REN) reporter system (Liu et al., 2014). We
170
generated a GALT9-LUC reporter construct in which the GALT9 coding region was fused with
171
that of LUC (Figure 2D). We also generated GALT9m-LUC by substituting the nucleotides of the
172
miR775 binding site in GALT9-LUC but not the encoded amino acids (Figure 2A and 2D).
173
Transient expression of these constructs in tobacco protoplasts showed that the LUC/REN
174
chemiluminescence ratio was significantly lowered in the presence of miR775 (Figure 2D).
175
8
Attenuation of the LUC/REN ratio was abolished in the GALT9m-LUC plus miR775 combination
176
(Figure 2D), indicating that miR775 represses GALT9-LUC expression in a site-specific manner.
177
Finally, we examined how endogenous GALT9 level is affected by genetic manipulation
178
of miR775. In addition to the MIR775A-OX lines, we employed the CRISPR/Cas9 system to
179
delete a 123 bp genomic region in MIR775A (the only locus in A. thaliana) encompassing
180
miR775 (Supplemental Figure 4). Homozygous lines with no detectable expression of miR775
181
were selected and named mir775 (Supplemental Figure 4B-4D). By quantitative reverse
182
transcription coupled PCR (RT-qPCR) analysis, we found that the level of miR775 was
183
significantly increased and decreased in MIR775A-OX and mir775 in comparison to the wild
184
type, respectively (Figure 2E). GALT9 transcript level was significantly decreased in MIR775A-
185
OX but increased in mir775 relative to the wild type (Figure 2E). These results indicate that
186
altering miR775 level is sufficient to reciprocally module GALT9 transcript abundance.
187
188
The MIR775A-GALT9 Circuit Controls Organ and Cell Sizes
189
To elucidate the biological role of miR775, we monitored morphology of the mir775 plants
190
throughout the life cycle. In comparison to the wild type, a size reduction of leaf-related organs,
191
including the cotyledon, the fifth rosette leaf, and the petal, was observed for mir775 (Figure 3;
192
Supplemental Figure 4E-4H). Quantification confirmed that mir775 has significantly smaller
193
phyllome organs than the wild type (Figure 3D-3F). In contrast, mature organs of MIR775A-OX
194
were significantly larger than those of the wild type (Figure 3). To confirm the mir775 phenotype,
195
we generated the MIR775A-OX mir775 double mutant through genetic crossing (Supplemental
196
Figure 5). We found that the 35S:pre-miR775a transgene in the used MIR775A-OX line was able
197
to restore miR775 transcript accumulation and rescue the organ reduction phenotype in the
198
mir775 background (Figure 3; Supplemental Figure 5).
199
To test the role of GALT9 in phyllome organs, we employed the CRISPR/Cas9 system
200
to delete the entire coding region of GALT9 (Supplemental Figure 6). In the homozygous
201
deletion lines (galt9-1), GALT9 expression was drastically compromised in comparison with the
202
wild type (Supplemental Figure 6A-6C). We also identified an Arabidopsis T-DNA line (galt9-2)
203
carrying insertion in the start codon of GALT9 (Supplemental Figure 6A). Both galt9 mutants
204
exhibited significantly enlarged phyllome organs than the wild type (Figure 3), phenotypes
205
similar to MIR775A-OX. We also generated transgenic plants over-expressing GALT9 (GALT9-
206
9
OX) and GALT9m (GALT9m-OX; see Figure 2A) driven by the 35S promoter (Supplemental
207
Figure 7). Both GALT9-OX and GALT9m-OX plants displayed significantly reduced sizes of leaf-
208
related organs than the wild type (Figure 3; Supplemental Figure 7), phenotypes similar to those
209
of mir775.
210
In contrast to the phyllome, there are organs in A. thaliana that rely on heterotropic
211
growth to reach the intrinsic sizes, such as the hypocotyl, the silique, and the inflorescence stem
212
(Geitmann and Ortega, 2009; Peaucelle et al., 2015; Andres-Robin et al., 2018). In comparison to
213
the wild type, we found that hypocotyl length, silique length, and inflorescence height of the
214
mir775 plants were not statistically different from those of the wild type (Figure 4). By contrast,
215
sizes of these organs of the MIR775A-OX, galt9, GALT9-OX, and GALT9m-OX plants were
216
significantly altered compared to the wild type with the exception of hypocotyl length of GALT9-
217
OX (Figure 4). Collectively, these results indicate that endogenous miR775 primarily promotes
218
phyllome organ growth by repressing GALT9 in A. thaliana.
219
In addition to GALT9, we have previously reported three other computationally
220
predicted target genes for miR775 including DICER-LIKE1 (DCL1) (Zhang et al., 2011).
221
Inspection of the degradome sequencing data from both the wild type and MIR775A-OX
222
backgrounds revealed no evidence for miR775-directed cleavage for these genes (Supplemental
223
Figure 8). Furthermore, consistent with previous characterizations of the dcl1 mutants (e.g.
224
Mallory and Vaucheret, 2006), an examined dcl1 T-DNA insertion mutant exhibited
225
significantly reduced organ sizes in comparison to the wild type (Supplemental Figure 9),
226
phenotype opposite to that of galt9 or MIR775A-OX. Thus, GALT9 is a bona fide miR775 target
227
that plays an opposite role to miR775 in determining intrinsic organ size.
228
A change in organ size can be attributed to altered cell size and/or cell number. To
229
assess the effects of the MIR775A-GALT9 circuit, we selected four cell types from three organs
230
for examination by cryo-scanning electron microscopy (cryo-SEM). Observed and quantified
231
sizes of MIR775A-OX and galt9 epidermal cells on the cotyledon, the petal, and the hypocotyl as
232
well as the guard cells on the cotyledon were significantly larger than those of the wild type
233
(Figure 5). Opposite phenotypes were observed for mir775 and GALT9-OX cells (Figure 5A-5E).
234
Moreover, a highly linear relationship with a virtually 1:1 slope between the cell size and the
235
organ size was observed for the three examined organ types across the five genotypes (Figure
236
5F). These results indicate that changes in cell size are primarily responsible for changes in organ
237
10
size caused by manipulating the MIR775A-GALT9 circuit.
238
239
MIR775A-GALT9 Modulates Pectin Level and Cell Wall Elasticity
240
Members of the GALT family have been extensively implicated in cell wall remodeling
241
(Supplemental Figure 2A) (Bouton et al., 2002; Qu et al., 2008; Qin et al., 2017). As most
242
proteins involved in cell wall remodeling locate on the endomembrane (Parsons et al., 2012), we
243
determined the subcellular localization of GALT9. RESPONSIVE TO ANTAGONIST1 (RAN1)
244
is a copper transporter reported to reside on the endomembrane (Hirayama et al., 1999). Using
245
GALT9 fused with the green fluorescent protein (GFP), we found that GALT9-GFP colocalized
246
with mCherry-tagged RAN1 transiently co-expressed in the same tobacco leaf epidermal cells
247
(Figure 6). This observation indicates that transiently expressed GALT9 is located on the
248
endomembrane.
249
To infer the molecular function of GALT9, we carried out a co-expression analysis and
250
identified 174 genes that are co-expressed with GALT9 in A. thaliana (Supplemental Dataset 1).
251
Gene Ontology (GO) analysis revealed that these genes were most significantly enriched with
252
GO terms related to cell wall biology and pectin metabolism in particular (Figure 6B). Manual
253
review revealed that 20 of these genes are linked to pectin metabolism and related processes,
254
including eight genes of the pectin lyase-like superfamily, four genes of the TRICHOME
255
BIREFRINGENCE-LIKE family, and eight other genes in pectin synthesis and modifications
256
based on experimental evidence in the literature (Figure 6C). As examples, co-expression
257
patterns between GALT9 and TRICHOME BIREFRINGENCE (TBR), which was shown to
258
regulate pectin composition in the trichome and stem (Bischoff et al., 2010), and between
259
GALT9 and POWDERY MILDEW RESISTANT6 (PMR6), a member of the pectin lyase-like
260
superfamily and whose mutation caused smaller rosette leaves with altered pectin composition
261
(Vogel et al., 2002), are shown in Figure 6D.
262
To confirm the involvement of GALT9 in pectin metabolism, we performed
263
monosaccharide composition analysis of the cell walls. We found that the relative amount of
264
glucose, the primary monosaccharide of cellulose, was not significantly different in the de-
265
starched fifth rosette leaves from the mir775, MIR775A-OX, galt9, and GALT9-OX plants in
266
comparison to the wild type (Figure 7). In contrast, the relative amount of galacturonic acid, the
267
representative derivative of pectin polysaccharides, was significantly lower in the MIR775A-OX
268
11
and galt9 plants but higher in the mir775 and GALT9-OX plants than the wild type (Figure 7A).
269
Moreover, an inverse linear relationship between the relative amount of galacturonic acid and the
270
relative cell size was observed among the five genotypes (Figure 7B). This linear relationship
271
was not found for the relative glucose level (Figure 7B). These results indicate that MIR775A-
272
GALT9 specifically influences pectin level in the leaf cell walls.
273
Raman imaging is a technique for obtaining high-resolution, chemically specific, and
274
non-destructive information of plant cell walls (Gierlinger et al., 2012; Zeng et al., 2016). Using
275
a home-built coherent Raman microscope, we mapped in situ pectin distribution in a mutant
276
defective in QUARTET2 (QRT2). Stronger than wild type signals encircling cotyledon epidermal
277
cells were observed in qrt2 (Supplemental Figure 10), consistent with previous reports that
278
QRT2 is required for pectin degradation (Rhee and Somerville, 1998). Similar to qrt2, we
279
detected stronger than wild type pectin signals in both mir775 and GALT9-OX plants
280
(Supplemental Figure 10A). The MIR775A-OX and galt9 plants, in contrast, exhibited the
281
opposite phenotype with weaker pectin signals than the wild type (Figure 8). This effect was
282
specific for pectin, as no difference in cellulose deposition among MIR775A-OX, galt9, and the
283
wild type was observed (Figure 8A and 8C). Quantification of the signal intensity confirmed that
284
pectin content was significantly reduced in MIR775A-OX and galt9 (Figure 8D).
285
As HG accounts for more than 60% of plant cell wall pectin (Caffall and Mohnen,
286
2009), we performed immunohistochemical analysis of cotyledons using a fluorescence-labeled
287
monoclonal antibody (LM19) specific for HG (Verhertbruggen et al., 2009). Fluorescence
288
microscopy revealed that LM19 signals in the MIR775A-OX and galt9 seedlings were drastically
289
reduced in comparison to the wild type (Figure 8E). By contrast, Fluorescent Brightener 28
290
(FB28), which mainly stains cellulose, generated signals with no obvious difference among the
291
genotypes (Figure 8E). These results confirmed that miR775 and GALT9 reduces and promotes
292
pectin deposition in the cell walls, respectively.
293
AFM is useful for determining the surface structures and mechanical characters of
294
biological samples at the nanometer scale (Yakubov et al., 2016). To investigate the link between
295
pectin content and mechanical property of the cell wall, we employed AFM to directly measure
296
the elastic properties of the epidermal cells. This analysis showed that the qrt2 mutant has higher
297
elastic modulus than the wild type (Supplemental Figure 10B), consistent with the notion that
298
higher pectin level leads to increased stiffness of the wall. We then applied AFM to measure the
299
12
elastic properties of the MIR775A-OX and galt9 cotyledon cells and petal cells (Figure 9). In
300
accordance with the cryo-SEM results (Figure 5), the 3D contour mapped by AFM revealed that
301
the MIR775A-OX and galt9 cells are larger than the wild type (Figure 9A and 9D). The
302
MIR775A-OX and galt9 cell walls, however, have elastic moduli significantly lower than the
303
wild type (Figure 9C and 9F), indicating that the enlarged cells have reduced wall rigidity. Taken
304
together, our results demonstrate that MIR775A-GALT9 modulates pectin abundance in the cell
305
wall and affects resistance to micro-indentation.
306
307
MIR775A Is Negatively Regulated by HY5 in Aerial Organs
308
A full-length cDNA BX81802 matches the MIR775A locus, allowing the transcription start site
309
and proximal promoter region (pMIR775A) to be determined (Figure 10). To find out how
310
MIR775A is transcriptionally regulated, we examined available whole genome chromatin
311
immunoprecipitation (ChIP) data and identified an ELONGATED HYPOCOTYL5 (HY5)
312
binding peak in pMIR775A (Figure 10A) (Zhang et al., 2011). As a key transcription factor for
313
photomorphogenesis, HY5 is known to bind to G-box-like motifs (Oyama et al., 1997; Yadav et
314
al., 2002; Song et al., 2008). Indeed, we located a typical G-box like motif in pMIR775A that
315
coincides with the HY5 binding peak (Figure 10A). Using ChIP-qPCR, significant enrichment of
316
HY5 occupancy at pMIR775A was confirmed (Figure 10B). These results reveal HY5 as a
317
plausible upstream regulator for the MIR775A-GALT9 circuit.
318
To examine the effect of HY5 on pMIR775A in vivo, we generated the 35S:GFP and
319
35S:HY5-GFP effector constructs. As reporters, we used pMIR775A to drive LUC and pMIR408,
320
which was previously shown to be activated by HY5 (Zhang et al., 2014), as a positive control.
321
We tested four effector-reporter combinations through co-infiltration of tobacco leaf epidermal
322
cells. Attesting to validity of the assay, co-expression of HY5 with pMIR408:LUC robustly
323
enhanced LUC activity (Figure 10C). However, in the presence of HY5, the pMIR775A activity
324
was markedly decreased (Figure 10C), indicating that HY5 negatively regulates MIR775A. To
325
corroborate this regulatory relationship in A. thaliana, we fused the β-glucuronidase (GUS) gene
326
with pMIR775A and expressed the reporter in either the wild type (pMIR775A:GUS) or the hy5-
327
215 (pMIR775A:GUS/hy5-215) genetic background (Figure 10D). In both seedlings and adult
328
plants, we found that GUS activity in the shoot was higher in hy5-215 than in the wild type
329
(Figure 10D; Supplemental Figure 11), confirming HY5-mediated MIR775A repression.
330
13
Finally, we performed RT-qPCR analysis to monitor the influence of HY5 on miR775
331
and GALT9 transcript accumulation. For this purpose, we also employed a HY5-OX line in which
332
expression of the HY5 coding region was driven by the 35S promoter (Gao et al., 2020). This
333
analysis revealed that miR775 abundance increased in the hy5-215 shoots but decreased in HY5-
334
OX with reference to the wild type (Figure 10E). Conversely, GALT9 transcript level was
335
significantly lower in hy5-215 but higher in HY5-OX shoots compared to the wild type (Figure
336
10E). Collectively these results indicate that HY5 binds to the MIR775A promoter to repress
337
miR775 accumulation and derepress GALT9 in aerial organs, thereby forming the HY5-
338
MIR775A-GALT9 repression cascade.
339
Previously, we reported that HY5 positively regulates MIR775A based on analysis of
340
miR775 abundance in whole young seedlings (Zhang et al., 2011). To ascertain whether HY5
341
positively or negatively regulates MIR775A, we compared GUS activities in different organs of
342
pMIR775A:GUS and pMIR775A:GUS/hy5-215 plants. This analysis revealed that, in contrast to
343
the aerial organs, GUS activity in pMIR775A:GUS/hy5-215 root was consistently lower than that
344
in the wild type background at different developmental stages (Supplemental Figure 11B-11D).
345
In separately sampled shoots and roots, miR775 level determined by RT-qPCR was higher and
346
lower in hy5-215 compared to the wild type, respectively (Supplemental Figure 11E). These
347
results indicate that HY5 differentially regulates MIR775A in the aerial and underground organs.
348
349
The HY5-MIR775A-GALT9 Pathway Regulates Leaf Size
350
The above findings prompted us to examine the role of HY5 in leaf size determination. We
351
generated a null hy5-ko allele by deleting almost the entire coding region using the
352
CRISPR/Cas9 system (Supplemental Figure 12). Similar to the well-characterized hy5-215 allele,
353
which carries a point mutation that abolishes proper splicing of the first intron (Oyama et al.,
354
1997), the hy5-ko seedlings exhibited larger cotyledons and longer hypocotyls than the wild type
355
(Supplemental Figure 12B-12D). In the adult stage, the hy5 mutants have larger rosette leaves
356
and longer petioles than the wild type (Supplemental 12E). On the contrary, HY5-OX plants
357
exhibited the opposite phenotypes in both the seedling and adult stages (Supplemental Figure
358
12C-12E). These results extended previous works documenting the organ enlargement
359
phenotypes of the hy5 mutants (Sibout et al., 2006; Brown and Jenkins, 2008; Burko et al., 2020).
360
Using cryo-SEM, we analyzed and quantitated size of epidermal cells from both the
361
14
cotyledons (Supplemental Figure 12F and 12G) and the fifth rosette leaves of adult plants
362
(Figure 11). In both cases, we confirmed that the hy5 mutants have significantly enlarged
363
epidermal cells compared to the wild type. To test whether these effects were related to the
364
pectin level, we performed Raman microscopy on the fifth rosette leaves and found that the hy5-
365
ko cells have significantly less pectin than the wild type (Figure 11C and 11D). This finding was
366
corroborated by quantifying the galacturonic acid content in the cell wall of the hy5-ko and wild
367
type leaves (Figure 11E). AFM analysis showed that the hy5-ko cell walls have significantly
368
reduced elastic modulus than the wild type (Figure 11F and 11G). These results indicate that
369
HY5 is a negative regulator for leaf size by increasing the pectin level and limiting cell expansion.
370
To genetically analyze whether HY5 and MIR775A-GALT9 act in the same pathway to
371
regulate leaf growth, we generated the hy5 mir775 and hy5 GALT9-OX double mutants through
372
genetic crossing using hy5-ko. Quantification of the size of the fifth rosette leaves revealed that
373
the leaf enlargement phenotype of hy5-ko was suppressed in both hy5 mir775 and hy5 GALT9-
374
OX (Figure 12). By cryo-SEM analysis and chemical quantification, we confirmed that the two
375
double mutants mitigated the cell enlargement and pectin reduction phenotypes of hy5-ko (Figure
376
12B). Moreover, a linear correlation between the cell size and leaf size was observed for the wild
377
type, hy5-ko, mir775, GALT9-OX, hy5 mir775 and hy5 GALT9-OX genotypes (Figure 12C).
378
Conversely, a reverse correlation between cell size and pectin level was observed across the
379
same genotypes (Figure 12D). Taken together, these results indicate that MIR775A and GALT9
380
act downstream of HY5 in the same genetic pathway to control pectin content and intrinsic leaf
381
size (Figure 13).
382
15
383
16
Discussion
384
Organ size is one of the dominating traits for plant development and architecture. Molecular
385
genetics studies in the past three decades have identified numerous genes in organ size control
386
(Bogre et al., 2008; Johnson and Lenhard, 2011; Gonzalez et al., 2012; Hepworth and Lenhard,
387
2014; Hong et al., 2018). Characterization of these genes has led to the conclusion that organ
388
size control is primarily exerted by cell number regulation and cell size control is also integral to
389
the intricate regulatory network governing organ size (Ferjani et al., 2007; Hong et al., 2018).
390
Because the presence of a rigid plant cell wall, increasing of cell volume must be accompanied
391
by mechanisms that allow timely wall relaxation. In this study, we identified 23 putative CW-
392
miRNAs in A. thaliana that are potentially pertinent to the regulation of primary wall
393
biosynthesis (Figure 1A). We selected miR775 as an example for functional characterization and
394
provided new insights into how miRNAs may regulate organ size by modulating cell wall
395
biosynthesis and/or modification.
396
We found that GALT9 is the bona fide target for miR775 specifically in A. thaliana
397
(Figures 1-3; Supplemental Figures 1 and 2). GALT9 is a member of the glycosyltransferase 31
398
family (Supplemental Figure 2A) and locates to the endomembrane (Figure 6A). It has been
399
shown that several members of this family are capable of adding galactose to various glycans
400
(Velasquez et al., 2011; Qin et al., 2017). The closest homolog to GALT9 in cotton is GhGALT1
401
(Supplemental Figure 2A). It was reported that GhGALT1 overexpression in cotton resulted in
402
smaller leaves, reduced boll size, and shorter fibers (Qin et al., 2017). In vitro purified
403
GhGALT1 exhibited galactosyltransferase enzyme activity in galactan backbone biosynthesis
404
(Qin et al., 2017). In this study, we provided a coherent body of evidence, including co-
405
expression pattern with pectin related genes (Figure 6B-6D), monosaccharide quantification
406
(Figure 7), confocal Raman microcopy and pectin immunolabelling (Figure 8; Supplemental
407
Figure 10), that support an indisputable role of GALT9 in modulating the level of cell wall
408
pectin in A. thaliana.
409
Moreover, reduction in pectin content in galt9 is associated with alteration to cell wall
410
mechanical property. Using AFM, we analyzed both the cotyledon and petal epidermal cells and
411
observed that the galt9 and MIR775A-OX cell walls displayed significantly lower elastic
412
modulus than that of the wild type (Figure 9; Supplemental Figure 10). This observation is
413
consistent with previous AFM analysis of epidermal cells that linked variation in the pectin
414
17
network to changes in cell wall elasticity (Peaucelle et al., 2015; Xi et al., 2015). Together with
415
studies on pectin biochemistry (Wolf et al., 2012; Xiao et al., 2014; Peaucelle et al., 2015;
416
Andres-Robin et al., 2018), these findings suggest that attenuation of the pectin constitute in
417
galt9 and MIR775A-OX cell walls might compromise cross-link with cellulose, which in turn
418
reduces elastic resistance to internal turgor pressure. This property of the cell wall would allow
419
more expandability that translates into enlarged cell sizes, which we observed by cryo-SEM and
420
AFM (Figures 5 and 9). Consistent with previous suggestions (e.g. Xiao et al., 2014), these
421
results imply that the capacity for cell expansion is not maximized in the wild type organs due to
422
rigidification of the pectin cross-linked cell walls. We hypothesize that by tuning pectin content,
423
GALT9 might act as a downstream component of the regulatory networks that control cell
424
expansion and present this idea in a conceptual model shown in Figure 13.
425
Regarding phyllome organs, we found that MIR775A-OX and galt9 plants have
426
significantly larger organs while mir775 and GALT9-OX plants have smaller organs than the
427
wild type (Figure 3; Supplemental Figures 3-7). Importantly, we did not observe substantial
428
changes in the number of epidermal cells in any the examined organs (Figure 5). Across multiple
429
organs of the mir775, MIR775A-OX, galt9, and GALT9-OX genotypes, a strong linear correlation
430
between organ size and cell size was observed (Figure 5F). These changes in cell size resulted in
431
essentially one-to-one changes in organ size across the examined genotypes (Figure 5F),
432
suggesting that altered cell proliferation is not the cause for the observed changes in organ size.
433
These findings thus indicate that the MIR775A-GALT9 circuit is part of the cellular machinery
434
that controls intrinsic organ size independent of cell proliferation (Ferjani et al., 2007; Hong et
435
al., 2018).
436
Organogenesis requires coordinated cellular responses to developmental and
437
environmental cues to realize the genetically determined growth potential. Through molecular
438
and genetic analyses, we showed that in aerial organs MIR775A is under negative transcriptional
439
control by HY5 (Figure 10; Supplemental Figure 11). Extending previous studies (Sibout et al.,
440
2006; Brown and Jenkins, 2008; Burko et al., 2020), we confirmed that HY5 is a negative
441
regulator for leaf size by modulating cell size (Figures 11 and 12; Supplemental Figure 12).
442
Importantly, we found that the effect of HY5 on cell size stems from alteration of pectin level and
443
elasticity of the cell walls (Figures 11 and 12). HY5-MIR775A-GALT9 is therefore a repression
444
cascade operating in A. thaliana that imposes restriction on cell wall flexibility via GALT9-
445
18
mediated pectin deposition and helps the plant to reach the desired intrinsic leaf size (Figure 13).
446
HY5 is a key gene regulator for light signaling and photomorphogenesis (Oyama et al., 1997;
447
Burko et al., 2020). Thus, whether the HY5-MIR775A-GALT9 pathway is a mechanism for
448
modulating pectin in the establishment of photomorphogenesis warrants investigation.
449
As HY5 is a negative regulator of MIR775A (Figure 10), there should exist positive
450
regulators for the spatiotemporal accumulation of miR775. Our preliminary results suggest that
451
members of the class II TCP (TEOSINTE BRANCHED1, CYCLOIDEA, PCF) transcription
452
factor family, which regulate the transition from cell division to cell expansion in dicot leaves
453
(Palatnik et al., 2003; Ori et al., 2007; Efroni et al., 2008; Schommer et al., 2014), are candidates
454
that activate MIR775A. It would be interesting to characterize these organogenesis-related factors
455
that regulate miR775 to further elucidate how this miRNA contributes to pectin dynamics during
456
leaf development. These efforts should be instrumental to reveal how other CW-miRNAs relay
457
developmental or environmental cues to regulate cell wall remodeling and prepare the cells
458
transitioning into expansion-driven growth with proper resistance to turgor pressure to reach the
459
intrinsic size.
460
As an important class of endogenous regulatory RNAs, miRNAs are known to regulate
461
leaf organogenesis (Palatnik et al., 2003; Ori et al., 2007; Rodriguez et al., 2010; Schommer et
462
al., 2014; Rodriguez et al., 2016; Yang et al., 2018). Several conserved miRNA families,
463
including miR156, miR319, and miR396, have been shown to regulate diverse aspects of leaf
464
organogenesis involving leaf initiation, phase transition, polarity establishment, and morphology
465
(Braybrook and Kuhlemeier, 2010; Efroni et al., 2010; Yang et al., 2018). For instance, over
466
activation of miR319 promotes cell proliferation and results in larger leaves made up of smaller
467
cells in comparison to the wild type (Palatnik et al., 2003; Efroni et al., 2008). These phenotypes
468
are in line with the “compensation phenomenon” whereby mutants defective in cell proliferation
469
may alter cell size to reach relatively the same final organ size (Ferjani et al., 2007; Kawade et
470
al., 2010; Czesnick and Lenhard, 2015). Our finding on the role of miR775 in regulating leaf size
471
through cell wall remodeling adds one more node to the miRNA networks governing leaf
472
development and morphogenesis in A. thaliana.
473
The miRNA families with known roles in leaf organogenesis, such as miR156, miR319,
474
and miR396, are deeply conserved in angiosperm (Yang et al., 2018; Guo et al., 2020). In
475
contrast, while the target gene GALT9 is conserved in angiosperm (Figure 1D; Supplemental
476
19
Figure 2A), miR775 is an evolutionarily young miRNA unique to A. thaliana (Figure 1A;
477
Supplemental Figure 1). Delineation of the HY5-MIR775A-GALT9 pathway and documentation
478
of the mir775 phenotype (Figures 3-5, 10, and 12) demonstrated that MIR775A has been
479
successfully integrated into the A. thaliana leaf developmental program. This finding suggests
480
that the miRNA networks governing leaf development in different plant species may contain
481
conserved “old” miRNAs interlaced with diverse species-specific “young” miRNAs. To confirm
482
miRNA diversity in contributing to differential organ size control mechanisms, it would be
483
interesting to test whether introducing species-specific CW-miRNAs such as miR775 or custom-
484
designed artificial miRNAs into diverse plant species is sufficient to repress the GALT9
485
orthologs and to modify organ size.
486
In summary, the evidence presented in this work highlights the function of a species-
487
specific CW-miRNA in regulating cell and organ size in A. thaliana. Future investigation of
488
other CW-miRNAs should provide additional insights into how plants orchestrate a complex
489
sequence of molecular behaviors to modify the cell walls during development and in response to
490
environmental cues. In addition to further elucidating the regulatory programs, these efforts
491
would serve as a proof-of-concept to employ CW-miRNAs to sculpture plant size and
492
architecture, which determine many agronomic traits in crops (Tang and Chu, 2017).
493
494
20
495
21
Methods
496
Plant Materials and Growth Conditions
497
The wild type plant used in this study was A. thaliana ecotype Col-0. To produce the
498
35S:MIR775A and 35S:GALT9 constructs, the genomic regions containing pre-miR775a and the
499
GALT9 coding region were PCR amplified using the Pfusion DNA polymerase (New England
500
Biolabs) and primers listed in Supplemental Table 2. The PCR products were cloned into the
501
35S-pKANNIBAL vector (Li et al., 2010). The 35S:GALT9m construct was generated by
502
substituting the nucleotides of the miR775 binding site within the GALT9 coding region but not
503
the encoded amino acids using primers listed in Supplemental Table 2. Following transformation
504
and selection with BASTA (20 mg L-1) (bioWORLD), transformants were allowed to propagate
505
to the T2 generation for analysis. The HY5-OX plants were as previously described (Gao et al.,
506
2020). The pMIR775A:GUS line was generated by cloning the 1,064 bp genomic fragment
507
upstream of the full-length cDNA BX81802 into the pCAMBIA-1381Xa vector (CAMBIA). The
508
construct was used to transform wild type plants following the standard floral dipping method
509
and selected with Hygromycin (20 mg L-1). T2 generation plants were screened for GUS activity
510
and a designated line was used for crossing into the hy5-215 background.
511
A CRISPR/Cas9 system specific for plants was used to delete MIR775A, GALT9, and
512
HY5 as described (Mao et al., 2013). In the modified pCAMBIA1300 vector, the 35S and the
513
AtU6-26 promoter respectively drive Cas9 and pairs of sgRNA designed to target both ends of
514
the target genes. The resulting constructs were introduced into wild type plants via
515
transformation. T1 generation plants were individually genotyped by PCR and sequencing to
516
identify deletion events. Approximately 200 individual T2 generation plants were further
517
genotyped to identify Cas9-free homozygous mutant lines.
518
To grow Arabidopsis plants, surface sterilized seeds were plated on agar-solidified MS
519
media including 1% (w/v) sucrose and incubated at 4°C for three days in the dark. Germinated
520
seedlings were either allowed to grow on the plate for three weeks (16 h light/8 h dark at
521
22°C/20°C) or transferred commercial soil and maintained in a growth chamber (16 h light/8 h
522
dark at 22°C/20°C, 50% relative humidity). The light intensity was approximately 120 μmol m-2
523
s-1. Tobacco seedlings used for transient assay were Nicotiana benthamiana, which were grown
524
under settings of 16 h light/8 h dark, 25°C/21°C, 50% relative humidity, and light intensity of
525
150-200 μmol m-2 s-1.
526
22
527
Identification of CW-miRNAs
528
The 572 cell wall biosynthesis related genes were collected by GO term search. The 491 genes
529
encoding Golgi-enriched proteins were obtained from previous studies (Parsons et al., 2012).
530
Full-length cDNA sequences for a nonredundant combination of these genes were obtained from
531
TAIR (www.arabidopsis.org). Searching against the 427 annotated miRNAs in A. thaliana
532
(miRBase, version 22) (Kozomara et al., 2018) was done using the computational tools
533
psRNATarget (Dai and Zhao, 2011) and psRobot (Wu et al., 2012). This process produced two
534
separate outputs, which were further searched against degradome sequencing data generated by
535
the CleaveLand4 or StarScan pipeline (Addo-Quaye et al., 2009; Liu et al., 2015). Possible
536
miRNA-target pairs predicted by both tools or by either one but compatible with degradome data
537
were combined into a nonredundant dataset, which contained 23 miRNAs and 78 target genes
538
listed in Supplemental Table 1. Conservation of CW-miRNAs was determined by searching
539
against miRNAs in miRBase (version 22) and PmiREN (Guo et al., 2020). Brassicaceae species
540
with genome sequences but no miRNA annotation were manually checked using BLASTN (E-
541
value < 1e-10) and RNAfold for evaluating the secondary structures as previously reported
542
(Gruber et al., 2008). The predicted target genes were searched against seven Brassicaceae
543
species with sequenced genomes for possible orthologs using BLASTP (E-value < 1e-10).
544
545
Degradome Sequencing and Analysis
546
Total RNA from MIR775A-OX leaves was isolated using Trizol reagent (Invitrogen). Degradome
547
library construction using biotinylated random primers was performed as previously described
548
(German et al., 2008; 2009). The library was subjected to single-end sequencing (50 bp) on the
549
Illumina Hiseq 2500 platform. A total of 63,558,618 clean reads were generated and 55,077,460
550
mapped to the TAIR10 A. thaliana genome using Bowtie2 (Langmead and Salzberg, 2012),
551
allowing no more than two mismatches. The sequencing data were deposited to the Sequence
552
Read Archive database (SRR10322040). Three sets degradome sequencing data from the wild
553
type seedlings (SRR3945024, SRR3945025, and SRR3945026) were used as control. Reads
554
mapped to the predicted target sites were used to extrapolate the positions of the 5’ transcript
555
ends and to calculate the RPM values using an in-house Perl script.
556
557
23
Quantitative RNA Analyses
558
Total RNA was isolated using the Quick RNA Isolation kit (Huayueyang). Each sample was
559
taken from the pooled tissues, such as leaves or roots. All experiments were repeated on at least
560
three sets of independently prepared RNA. mRNA and miRNA were reverse transcribed into
561
cDNA using the SuperScript III reverse transcriptase (Invitrogen) and the miRcute Plus miRNA
562
First-Stand cDNA Synthesis kit (Tiangen), respectively. Quantitative PCR was performed with
563
the ABI PRISM Fast 7500 Real-Time PCR engine using the TB Green Premix Ex TaqII (TIi
564
RNaseH Plus) (TaKaRa) and the miRcute Plus miRNA qPCR kit (SYBR Green) (Tiangen) with
565
three technical replicates, respectively. Actin7 and 5S RNA were used as internal controls.
566
Relative amounts of mRNA and miRNA were calculated using the comparative threshold cycle
567
method.
568
569
5’ RLM-RACE
570
The assay was performed using the 5’-Full RACE kit (TaKaRa) according to the manufacturer’s
571
instructions with modifications. Total RNA was isolated from seedlings and ligated to the 5’
572
RNA adaptor by T4 RNA ligase (TaKaRa). Reverse transcription was performed with 9-nt
573
random primers and the cDNA amplified by PCR with an adaptor primer and a gene-specific
574
primer. This was followed by a nested PCR and cloning of the products using the Mighty TA-
575
cloning kit (TaKaRa). Twenty independent clones were randomly picked and sequenced.
576
577
REN/LUC Dual Luciferase Assays
578
The REN/LUC construct was modified from the previous version (Liu et al., 2014) by using the
579
Actin2 promoter to drive the LUC fusion proteins. The GALT9m-LUC reporter construct was
580
generated by substituting the nucleotides in the miR775 binding site within GALT9 by PCR
581
using primers listed in Supplemental Table 2.Three combinations of the two effectors and/or
582
reporter constructs were used to transiently co-transform tobacco protoplasts as previously
583
described (Liu et al., 2014). Chemiluminescence was detected using the NightSHADE LB 985
584
system (Berthold) in the presence of 20 mg mL-1 potassium luciferin (Gold Biotech). The
585
LUC/REN ratio was calculated to infer effectiveness of miR775 targeting.
586
587
Protein Localization
588
24
The GALT9 and RAN1 coding sequences were respectively cloned into the pJIM19-
589
GFP/mCherry/ vectors. Agrobacterium GV3101 cells harboring the 35S:GALT9-GFP and
590
35S:RAN1-mCherry constructs were mixed and co-infiltrated into tobacco leaf epidermal cells
591
with a syringe. The cells were observed three days thereafter using an LSM 710 laser scanning
592
confocal microscope (Zeiss). Colocalization was analyzed using the Coloc 2 module in ImageJ.
593
594
Co-expression Analysis
595
The GALT9 co-expressed genes in A. thaliana were obtained from the ATTED-II database
596
(version 9) (Obayashi et al., 2018). The 174 co-expressed genes were identified based on the
597
mutual rank index as a co-expression measure using a cutoff value of 400. The co-expressed
598
genes were visualized using the built-in tools in ATTED-II.
599
600
Cryo-SEM
601
The method for cryo-SEM was as previously described (Esch et al., 2004) with minor
602
modifications. The scanning electron microscope FEI Helios NanoLab G3 UC (Thermo
603
Scientific) and the Quorum PP3010T workstation (Quorum Technologies), which has a cryo
604
preparation chamber connected directly to the microscope, were used as a unit. Plant samples
605
were frozen in subcooled liquid nitrogen (-210°C) and then transferred in vacuum cabin to the
606
cold stage of the chamber for sublimation (-90°C, 5 min) and sputter coating (10 mA, 30 sec)
607
with platinum. Images were taken using the electron beam at 2 kV and 0.2 nA with a working
608
distance of 4 mm. Projective cell area of indicated samples was measured using ImageJ. Average
609
cell size was determined by measuring 100 cells from at least three samples.
610
611
Chemical Analysis of Cell Wall Components
612
Cell wall cellulose level was determined using the Cellulose Extraction and Determination kit
613
(Comin Biotechnology, www.cominbio.com). Approximately 300 mg tissues per sample were
614
homogenized in 1 mL 80% ethanol, heated at 90°C for 20 min, cooled to room temperature, and
615
centrifuged at 6000g for 10 min. The insoluble pellets were washed once in 1 mL 80% ethanol
616
and once in 1 mL acetone by vertexing and centrifugation at 6000g for 10 min. The pellets were
617
resuspended in 1 mL solution I provided in the kit, de-starched for 15 h at room temperature, and
618
collected by centrifugation at 6000g for 10 min, and dried. Five milligrams of the resulting cell
619
25
wall materials were homogenized in 0.5 mL distilled water, mixed with 0.75 mL concentrated
620
sulfuric acid on ice, incubated for 30 min, and centrifuged at 8000g for 10 min at 4°C. Glucose
621
determination in the supernatants was based on the anthrone assay (Yuan et al., 2019; Huang et
622
al., 2020) using reagents provided in the kit and following the manufacturer’s protocol. The
623
glucose concentration from the blue-green samples was measured by absorbance at 630 nm using
624
a NanoPhotometer P-class USB spectrophotometer (Implen GmbH).
625
Pectin level was determined using the Pectin Extraction and Determination kit (Comin
626
Biotechnology). Briefly, approximately 50 mg tissues per sample were homogenized in 1 mL
627
extraction buffer I provided in the kit, heated at 90°C for 30 min, cooled to room temperature,
628
and centrifuged at 5000g for 10 min. The insoluble pellets were washed in 1 mL extraction
629
buffer I by vertexing and centrifugation at 5000g for 10 min. The pellets were resuspended in 1
630
mL extraction buffer II provided in the kit, heated at 90°C for 1 h, and centrifuged at 8000g for
631
15 min. Galacturonic acid in the supernatants was determined by colorimetry as previously
632
described (Taylor, 1993) using reagents provided in the kit. Absorbance of the pink- to red-
633
colored samples at 530 nm was read on the NanoPhotometer P-class USB spectrophotometer.
634
635
GUS Staining
636
Care was taken to make sure whole plants or seedlings were submerged and evenly incubated at
637
room temperature for 6 h in a GUS staining solution (1 mM 5-bromo-4-chloro-3-indolyl-b-D-
638
glucuronic acid, 100 mM Na3PO4 buffer, 3 mM each K3Fe(CN)6/K4Fe(CN)6, 10 mM EDTA,
639
and 0.1% Nonidet P-40). After staining, chlorophyll was removed using 70% ethanol for 4 h,
640
which was repeated three times.
641
642
Confocal Raman Imaging
643
Freshly detached Arabidopsis cotyledons and young leaves were washed sequentially with 70%,
644
100%, and 70% ethanol for 10 min each to remove chlorophyll. After that, the samples were kept
645
in water. Label-free imaging of cellulose and pectin was performed with a home-built coherent
646
Raman microscope, fitted with a picoEmerald (Applied Physics & Electronics) picosecond laser
647
as light source, which supplies tunable pump beam and fixed Stokes beam. As previously
648
described (Gierlinger et al., 2012), 1100 cm-1 (asymmetric stretching vibration of the glycoside
649
bond C-O-C) and 854 cm-1 (C-O-C skeletal mode of α-anomers) were used for specific in situ
650
26
mapping of cellulose and pectin, respectively. The pump beams were respectively tuned to 952.5
651
nm and 975.5 nm, synchronized, and visualized with an inverted microscope (Olympus)
652
equipped with a 25× objective lens and a coherent Raman detection module. Each image was
653
acquired with 512 by 512 pixels and averaged by 5 frames. A background image was acquired
654
for each sample by only illuminating with the pump laser beam. For normalization, difference of
655
the signal intensity between each image and the corresponding background image was divided
656
by the background image using ImageJ.
657
658
Pectin Immunolabelling
659
This procedure was performed as previously described (Qi et al., 2017). Briefly, seven-day-old
660
seedlings were fixed in absolute methanol under vacuum and embedded in Steedman’s wax
661
(Sigma-Aldrich). After rehydration, 8 μm sections were prepared and pre-treated for 1 h with 2%
662
(w/v) BSA in PBS, and then incubated overnight with the primary antibody LM19 (PlantProbes)
663
diluted 1:500 in 0.1% BSA. After three washes in BST buffer (0.1% BSA and 0.1% (v/v) Tween
664
20), sections were incubated for 1 h with the secondary antibody Alexa Fluor 546 goat anti-rat
665
IgG (Life Technologies) diluted 1:1,000 in 0.1% BSA. Sections were mounted in ProLong
666
Antifade (Life Technologies) with cover slips and the Fluorescent Brightener 28 dye solution
667
(Sigma-Aldrich) added. Fluorescence imaging was performed with an LSM 710 laser scanning
668
confocal microscope (Zeiss).
669
670
AFM Analysis
671
Freshly detached cotyledons and petals were subject to AFM analysis as described with
672
modifications (Peaucelle et al., 2015; Xi et al., 2015). Briefly, the samples were attached to glass
673
slide using transparent nail polish and submerged under water at room temperature to prevent
674
plasmolysis. The topographical images of epidermal cells were scanned with a BioScope
675
Resolve atomic force microscope equipped with a ScanAsyst-Fluid cantilever (Bruker) of 20 nm
676
tip radius and 0.7 N m-1 spring constant. For topography, peak force error and DMT modulus
677
images, Peak Force QNM mode of the acquisition software were used, with peak force frequency
678
at 2 kHz and peak force set-point at 3 nN. The topology image size was 10 × 10 μm2 or 20 × 20
679
μm2 with a resolution of 256 × 256 pixels recorded at a scan rate of 0.2 Hz. To map apparent
680
Young’s modulus, 1 to 2 mm-deep indentations were performed along the topological skeletons
681
27
of epidermal cells to ensure relative normal contact between the probe and sample surface. At
682
least three indentation positions were chosen for each cell, with each position consecutively
683
indented three times, making at least nine indentation force curves per cell. Data were analyzed
684
with Nanoscope Analysis version 1.8.
685
28
Supplemental Data
686
Supplemental Figure 1. Comparison of Pre-miR775a Homologs in A. thaliana and A. lyrata.
687
Supplemental Figure 2. MiR775 Specifically Targets GALT9 in A. thaliana.
688
Supplemental Figure 3. Characterization of MIR775A-OX Lines.
689
Supplemental Figure 4. Generation and Characterization of the mir775 Mutant Lines.
690
Supplemental Figure 5. Characterization of the MIR775A-OX mir775 Line.
691
Supplemental Figure 6. Generation and Characterization of the galt9 Mutant Lines.
692
Supplemental Figure 7. Characterization of the GALT9-OX Lines.
693
Supplemental Figure 8. Degradome Sequencing Profiles of Predicted MiR775 Targets.
694
Supplemental Figure 9. Phenotypic Comparison of the galt9 and dcl1 Mutants.
695
Supplemental Figure 10. Analysis of the qrt2 Mutant Defective in Pectin Turnover.
696
Supplemental Figure 11. HY5 Differentially Regulates MIR775A in the Shoot and the Root.
697
Supplemental Figure 12. Generation and Characterization of Mutants for HY5.
698
Supplemental Table 1. Putative CW-miRNAs and Predicted Target Genes in A. thaliana.
699
Supplemental Table 2. Oligonucleotide Sequences of the Primers Used in This Study.
700
Supplemental Dataset 1. GALT9 Co-expressed Genes in A. thaliana.
701
29
Accession Number
702
Sequence data from this article can be found in the Arabidopsis Genome Initiative or
703
GenBank/EMBL databases under the following accession numbers: MIR775A (At1g78206), HY5
704
(At5g11206), GALT9 (At1g53290), DCL1 (At1g01040), and QRT2 (At3g07970). T-DNA
705
insertion mutants used are galt9 (SALK_015338), dcl1 (SALK_056243C), and qurt2
706
(SALK_031337).
707
708
Author Contributions
709
L.L. designed and supervised the research. H.Z., Y.Z., J.D., J.P., L.L, T.W., and H.C. performed
710
the research. H.Z., Y.S., Z.G. (Guo), Z.G. (Gao), L.X., G.Q., and Y.J. analyzed the data. H.Z.
711
and L.L. wrote the paper.
712
713
Acknowledgements
714
We thank Drs. Dong Liu and Chan Li at the National Center for Protein Science at Peking
715
University for technical assistance in AFM operation and image analysis, Dr. Yiqun Liu and Ms.
716
Yifeng Jiang at the Core Facilities of School of Life Sciences at Peking University for assistance
717
with SEM. This work was supported by grants from the National Key Research and
718
Development Program of China (2017YFA0503800) and the National Natural Science
719
Foundation of China (31621001).
720
30
721
Arabidopsis thaliana
Arabidopsis halleri
Arabidopsis lyrata
Capsella rubella
Leavenworthia alabamica
Camelina sativa
Brassica napus
Aethionema arabicum
Brassica rapa
Brassica oleracea
Sisymbrium irio
Schrenkiella parvula
Eutrema salsugineum
A
miR156h
miR827
miR775
miR156j
miR838
miR861
miR837
miR1886
miR5630b
miR2936
miR156i
miR414
miR417
miR773b
miR776
miR854e
miR4227
miR4239
miR5015
miR5021
miR5628
miR5658
miR5662
Figure 1. Identification and Analysis of Putative CW-miRNAs in A. thaliana.
C
D
0.5
0.7
0.9
0.80
0.79
0.78
0.77
0.67
0.62
0.57
0.55
a
b
c
d
e
f
g
h
Identity (%)
50
75
100
AUCGCAGAGUAUGAUGACUUUGUACUGCUAGAUAUCGAAGAGGAGUACAGUAAGCUCCCU
MFE/MDE
Brassicaceae
miR775 ACCGUGACGAUCUGUAGCUU
: :.:::::::::.::::::
UAUGAUGACUUCGUACUGCUAGAUAUCGAAGAGGAGUAC
UAUGAUGACUUUGUUCAGCUAGAUAUCGAAGAGGAGUAU
UAUGAUGACUUUGUACUGCUAGAUAUCGAAGAGGAGUAC
UAUGAUGACUUUGUACUGCUAGAUAUCGAAGAGGAGUAC
UAUGAUGACUUUGUACUGCUAGAUAUCGAAGAGGAGUAU
UAUGAUGAUUUUAUACUGUUAGAUAUCGAAGAGGAGUAC
UAUGAUGACUUCGUACAGCUAGAUAUAGAAGAGGAGUAC
UAUGAUGACUUCGUACAGCUAGAUAUAGAAGAGGAGUAC
UAUGAUGACUUCGUACAGCUAGAUAUAGAAGAGGAGUAC
UACGAUGACUUUAUACUGCUCGAUAUCGAGGAGGAGUAU
UAUGAUGACUUUGUACUGCUGGAUAUCGAAGAGGAGUAC
UAUGAUGACUUUGUACUGCUAGAUAUCGAAGAGGAGUAC
UAUGAUGAUUUUAUACUGCUUGAUAUUGAGGAGGAGUAC
UAUGAUGAUUUUGUACUAUUGGACAUUGAAGAAGAGUAC
UACGAUGAUUUCGUGCUUUUAGAUCUGGAGGAGGAGUAU
UAUGAUGAUUUCUUGCUAUUGGAUGUUGAGGAGGAAUAU
UACAAGGAUUUUAUACGCAUCGAUAUCGAAGAAGAAUAU
CACAAGGACUUCAUGCUCAUUGACAUCGACGAGAAGUAC
Y D D F V L L D I E E E Y
Arabidopsis thaliana
Arabidopsis lyrata
Arabidopsis halleri
Capsella rubella
Camelina sativa
Leavenworthia alabamica
Brassica napus
Brassica rapa
Brassica oleracea
Sisymbrium irio
Schrenkiella parvula
Eutrema salsugineum
Aethionema arabicum
Solanum lycopersicum
Oryza sativa
Amborella trichopoda
Selaginella moellendorffii
Physcomitrella patens
Position
Polymorphism
E
Root
Seedling
Leaf
Flower
Inflorescence
Silique
Present
Absent
RPM > 50
5 < RPM ≤ 50
0 < RPM ≤ 5
RPM = 0
B
Figure 1. Identification and Analysis of Putative CW-miRNAs in A. thaliana.
(A) Conservation of the 23 putative CW-miRNAs in Brassicaceae. Circles in blue indicate
presence of a given CW-miRNA in the corresponding species. (B) Expression profile of the
CW-miRNAs in A. thaliana. RPM (reads per million) values in 34 small RNA sequencing
datasets, which are grouped into six organ types based on similarity of the sampled plant
materials, are used to profile the miRNAs. (C) Comparison of the complementarity between
miR775 and its possible binding site in GALT9 homologs. On the left is a phylogenetic tree
reconstructed with closest GALT9 homologs from 18 species. Species in Brassicaceae are
shaded in blue. On the right is an alignment of sequences flanking the miR775 binding site
(in bold). The five polymorphic nucleotides within the miR775 binding site are shaded in
green. (D) Quantification of nucleotide conservation in GALT9 at the miR775 binding site
across the 18 examined species. Red stars indicate the high-diversity nucleotides. The
consensus sequence is shown below. (E) Calculated MFE/MED ratios for predicted
miR775:target duplexes. Lower case letters represent observed combinations of the five
polymorphic nucleotides. a, CGUAC; b, UGAAC; c, UGUAC; d, UGUGC; e, UAUAC; f,
UAUCC; g, UAUUU; h, CGAAA.
LUC/REN ratio
D
0
1
2
3
4
5
6
7
Relative transcript level
E
GALT9
miR775
WT
MIR775A-OX mir775
a
b
a
0
0.5
1.0
2.0
1.5
2.5
Reporter
construct
Effector
construct
GALT9
a.a.
GALT9m
miR775
GALT9
9/20
5/20
ACCGUGACGAUCUGUAGCUU-5’
: :.:::::::::.::::::
5’-UUCGUACUGCUAGAUAUCGAA
5’-UUCGUCCUACUGGACAUUGAG
F V L L D I E
35S
Pre-miR775a
35S
REN
pACT2
35S
REN
pACT2
35S
REN
pACT2
35S
Pre-miR775a
+
+
GALT9-LUC
GALT9m-LUC
GALT9-LUC
A
B
0.1
0.2
0.3
0
RPM
160
360
560
760
Position
Position
160
360
560
760
5’-UUCGUACUGCUAGAUAUCGAA
: :.:::::::::.::::::
ACCGUGACGAUCUGUAGCUU-5’
0.1
0.2
0.3
0
RPM
C
WT
MIR775A-OX
WT
MIR775A-OX
miR775:
a
b
a
b
c
c
Figure 2. Validation of GALT9 as an Authentic MiR775 Target.
(A) 5’ RLM-RACE analysis of GALT9. Gene structure of GALT9 is shown on top. Base pairing
between miR775 and GALT9 is shown on bottom. Arrows mark detected cleavage sites along
with frequency of the corresponding clones. Substituted nucleotides for making GALT9m are
colored in blue. (B) Comparison of degradome sequencing data obtained from the wild type
(left) and MIR775A-OX (right) plants. Frequency of the sequenced 5’ ends is plotted against
the position in the GALT9 transcript. Red dots indicate position of reads with the highest
frequency mapped to the miR775 binding site. (C) Sliding window analysis of degradome
sequencing data at the miR775 binding site. Step of 4 nucleotides was used. Dashed line
marks the position between the 10th and 11th nucleotides from the 5’ end of miR775. Arrows
indicate positions of the cleavage sites mapped by 5’ RLM-RACE in A. (D) REN/LUC dual
luciferase assay validating GALT9 repression by miR775. The Actin2 promoter was used to
drive expression of GALT9-LUC or GALT9m-LUC. The 35S:pre-miR775a effector and the
reporters were used to transiently co-transform tobacco protoplasts. The LUC/REN ratio of
chemiluminescence is shown on the right. Data are means ± SD from four independent
transformation events. Different letters denote combinations with significant difference
(Student’s t-test, p < 0.05). (E) Quantitative analysis of the miR775 and GALT9 transcript
levels in seedlings of the three indicated genotypes. Data are means ± SD from three technical
replicates. Different letters denote groups with significant difference (Student’s t-test, p < 0.01).
Figure 3. MIR775A and GALT9 Oppositely Regulate Size of Leaf-related Organs.
(A-C) Morphological comparison of three representative organ types across the indicated
genotypes. (A) Cotyledon of seven-day-old seedlings; (B) The fifth rosette leaf of three-
week-old plants; (C) petal of open flowers. Bars, 2 mm. (D-F) Quantitative size
measurement of cotyledons (D), the fifth rosette leaves (E), and the petals (F). Data are
mean ± SD from individual organs normalized against the wild type. Different letters
denote genotypes with significant difference (Student’s t-test, n = 30, p < 0.001 for D, n =
20, p < 0.01 for E, n = 30, p < 0.001 for F).
0
0.5
1
1.5
2
Relative fifth leaf area
a
b
c
d
e
a
f
0
0.5
1
1.5
2
Relative cotyledon area
a
b
c
d
e
a
f
0
0.5
1
1.5
Relative petal area
a
a
b
b
c
cd cd
A
B
C
D
E
F
MIR775A-OX
mir775
WT
MIR775A-OX mir775
galt9
GALT9-OX
GALT9m-OX
0
0.5
1
1.5
2
Figure 4. MIR775A and GALT9 Play Different Roles in Regulating Size of Heterotrophic
Organs.
(A-C) Morphological comparison of three representative organs with heterotrophic growth
across the indicated genotypes. (A) Hypocotyl of seven-day-old seedlings, bar, 2 mm; (B)
Mature silique, bar, 2 mm; (C) Mature inflorescence, bar, 2 cm. (D-F) Quantitative
measurement of hypocotyl length (D), silique length (E), and inflorescence height (F). Values
are mean ± SD from individual organs normalized to the wild type. Different letters denote
genotypes with significant difference (Student’s t-test, n = 15, p < 0.01 for D, n = 30, p <
0.001 for E, n = 26, p < 0.001 for F).
0
0.5
1
1.5
2
a
a
ad
a
b
c
d
0
0.4
0.8
1.2
1.6
a
a
b
b
a
c
d
C
A
B
D
E
F
a
a
a
b
c
d
d
Relative hypocotyl length
Relative silique length
Relative inflorescence height
Figure 5. The MIR775A-GALT9 Circuit Controls Cell Size.
(A-D) cryo-SEM analysis of epidermal cells of the five indicated genotypes. Shown are
representative images for cotyledon (A), bar, 50 μm; stoma including guard cells (B), bar, 20
μm; petal (C), bar, 20 μm; and hypocotyl (D), bar, 50 μm. (E) Quantification of epidermal cell
size from cotyledon, petal, and hypocotyl and stoma area. Data are mean ± SD relative to the
wild type from 30 individual cells of several individual plants. Different letters denote
genotypes with significant difference (Student’s t-test, p < 0.01 for A, C and D, p < 0.05 for B).
(F) Correlation between cell size and organ size. Relative organ and cell sizes of three organs
(cotyledon, petal, and hypocotyl) across the wild type, MIR775A-OX, mir775, galt9, and
GALT9-OX genotypes were used for a linear regression analysis.
A
0.5
0.7
0.9
1.1
1.3
1.5
1.7
0.6
0.8
1
1.2
1.4
Relative organ size
Relative cell size
E
F
Y = 1.04X
R = 0.85
p = 0.00049
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
leaf
stoma
petal
hypocotyl
Cell size relative to wild type
Organ type
galt9
GALT9-OX
WT
MIR775A-OX
mir775
a
b
b
a
c
a
b
b
c
c
a
b
b
c
c
a
b
b
d
c
B
C
D
MIR775A-OX
mir775
WT
galt9
GALT9-OX
B
0
1
2
3
4
5
6
7
8
9
10
-Log(p-value)
PMR6
-1.5
-1.0
-0.5
0
0.5
-1.0
0
1.0
-2.0
-1.0
0
1.0
Correlation of expression
C
D
GALT9
TBR
RAN1-mCherry
GALT9-GFP
A
Merge
R = 0.72
cell wall organization or biogenesis (22)
carbohydrate metabolic process (27)
cell wall organization (17)
external encapsulating structure organization (17)
lipid biosynthetic process (16)
pectin metabolic process (9)
galacturonan metabolic process (9)
polysaccharide metabolic process (14)
lipid metabolic process (21)
single-organism process (89)
Figure 6. GALT9 Has a Deduced Role in Pectin Metabolism.
(A) Colocalization of GALT9-GFP with RAN1-mCherry in tobacco leaf epidermal cells. Scatter
plot on the right shows correlation of GFP and mCherry fluorescence intensity. R, Pearson
correlation coefficient. Bar, 50 μm. (B) Top ten most significantly enriched GO terms in the
biological process category associated with the 174 GALT9 co-expressed genes. Numbers in
parentheses are co-expression genes associated with each term. (C) Concentric display of
GALT9 co-expression genes with the 20 pectin-related genes shown on the periphery.
Narrow lines representing mutual rank value above 200, medium lines representing 50-200,
and wide lines representing 0-50. (D) Correlation pattern between GALT9 and the pectin-
related genes PMR6 and TBR. Axes are Log2-transformed expression levels against the
averaged level of each gene.
Figure 7. The MIR775A-GALT9 Circuit Regulates Cell Wall Pectin Level.
(A) Quantification of the relative glucose and galacturonic acid levels in the cell walls.
Hydroxylated cell wall materials extracted from the de-starched fifth rosette leaf of
the indicated genotypes were used for monosaccharide measurement by colorimetry.
Data are mean ± SD from three technical replicates performed on pooled leaves.
Within a set of measurements, different letters denote genotypes with significant
difference (Student’s t-test, p < 0.01). (B) Correlation between relative cell size and
the two quantified cell wall monosaccharides across the five genotypes by a linear
regression analysis.
0
0.5
1
1.5
2
0.7
0.9
1.1
1.3
Relative cell size
0
0.5
1
1.5
2
2.5
R = -0.95, p = 0.0068
a
b
b
c
c
A
B
a
a
a
a
a
R = 0.67, p = 0.107
Galacturonic acid
Glucose
Relative monosaccharide level
Galacturonic acid
Glucose
Relative monosaccharide level
0
0.1
0.2
0.3
0.4
Relative cellulose level
0
0.1
0.2
0.3
0.4
0.5
Relative pectin level
a
a
a
a
b
b
Pectin
Cellulose
MIR775A-OX
WT
galt9
LM19 + FB28
MIR775A-OX
WT
galt9
A
E
C
B
D
Figure 8. MIR775A-OX and galt9 Seedlings Have Reduced Cell Wall Pectin.
(A-B) Examination of cell wall constituents by confocal Raman microscopy. Cotyledon
mesophyll cells of seven-day-old wild type, MIR775A-OX, and galt9 seedlings were imaged for
cellulose (A) at 1100 cm−1 and pectin (B) at 854 cm−1. Bars, 50 μm. (C-D) Relative cellulose
and pectin levels deduced from Raman images. Average intensity in a 25 μm by 25 μm area at
the cell corner was used to represent the level of the wall components. Data are mean ± SD of
15 areas from five cotyledons. Different letters denote genotypes with significant difference
(Student’s t-test, p < 0.01). (E) Immunohistochemical localization of pectin. The LM19 antibody
(green) and the FB28 dye (red) were used to stain seven-day-old seedlings and examined by
fluorescence microscopy. Bar, 100 μm.
0
20
40
60
80
100
120
Elastic modulus (MPa)
10
0 μm
100
a
0 MPa
Figure 9. MIR775A-OX and galt9 Epidermal Cells Have Reduced Elastic Modulus.
(A) AFM mapping of three-dimensional topography of epidermal cells. Individual cells of seven-
day-old cotyledons were analyzed. Colors represent distance from the base, which is the deepest
point the probe reaches. (B) Cell topography overlaid with elastic modulus. Colors indicate
elasticity. (C) Quantification of apparent Young’s modulus using the Peak Force QNM mode. Each
measurement was the average in a 5 μm by 5 μm area of a cell with the highest modulus. Data are
mean ± SD of 10 cells from three cotyledons. Different letters denote genotypes with significant
difference (Student’s t-test, p < 0.001). (D-F) Cell topography (D), topography overlaid with
elasticity (E), and apparent Young’s modulus (F) of the petal epidermal cells. Individual cells of
petals of open flowers were analyzed. Each measurement was the average in a 10 μm by 10 μm
area with the highest modulus. Data are mean ± SD of 10 cells from three petals. Different letters
denote significant difference (Student’s t-test, p < 0.001).
200
10
0
40
80
120
160
200
240
280
Elastic modulus (MPa)
b
c
a
b
c
C
F
0 μm
0 MPa
A
D
WT
galt9
MIR775A-OX
B
E
AT1G78200
AT1G78210
BX818024
29421k
29422k
29423k
29424k
MIR775A
G-box like
HY5 binding profile
A
B
IgG
α-HY5
0.12
0.08
0.04
0
ChIP siganl (% of input)
hy5
WT
0
1
2
3
Relative transcript level
Figure 10. HY5 Represses MIR775A Expression by Directly Binding to Its Promoter.
(A) HY5 occupancy profile at the MIR775A locus. HY5 binding profile is based on global ChIP
data mapped onto the Arabidopsis genome coordinates. Loci are represented by block arrows.
Position of MIR775A, defined by the full-length cDNA BX818024, is depicted as a black arrow.
The triangle marks the G-box like motif. (B) Confirmation of HY5 binding to pMIR775A by ChIP-
qPCR. ChIP was performed in light-grown wild type and hy5 seedlings with or without the anti-
HY5 antibody. Values are normalized to the respective DNA inputs. Data are ± SD from three
technical replicates. Different letters denote significant difference (Student’s t-test, p < 0.001). (C)
Transient expression assay for testing the effect of HY5 on pMIR775A activity. Either the
pMIR775A:LUC or pMIR408:LUC construct was co-infiltrated with the 35S:HY5-GFP (+HY5) or
the vector alone (-HY5) in tobacco epidermal cells and imaged for LUC activity. (D) GUS staining
for HY5-dependent pMIR775A activity in A. thaliana. The same pMIR775A:GUS reporter gene
was expressed in either the wild type or the hy5-215 background. Bar, 1 mm. (E) RT-qPCR
analysis of the relative miR775 and GALT9 transcript abundance in the wild type, hy5-215, and
HY5-OX seedlings. Data are means ± SD from three technical replicates. Different letters denote
groups with significant difference (Student’s t-test, p < 0.01).
WT
hy5-215 HY5-OX
miR775
GALT9
pMIR408:LUC
pMIR775A:LUC
- HY5
+ HY5
- HY5
+ HY5
pMIR775A:GUS/hy5-215
pMIR775A:GUS
C
D
E
High
Low
a
a
a
b
a
a
b
b
c
c
0
0.5
1
1.5
Figure 11. HY5 Is a Negative Regulator of Leaf Size.
(A) Enlargement of the hy5-ko epidermal cells in comparison to the wild type. The upper
side of the fifth leaf from three-week-old plants was used for cryo-SEM analysis. Bar, 50
μm. (B) Quantification of epidermal cell size. Data are mean ± SD of 100 individual cells
from five rosette leaves. Different letters denote significant difference (Student’s t-test, p <
0.001). (C) Imaging pectin in mesophyll cells by confocal Raman microscopy. Bar, 50 μm.
(D) Average intensity of Raman images was used to deduce relative pectin levels. Data
are mean ± SD of 15 areas from five leaves. Different letters denote significant difference
(Student’s t-test, p < 0.01). (E) Quantification of the relative galacturonic acid level in the
wild type and hy5-ko cell walls. Data are mean ± SD from three technical replicates
performed on pooled leaves. Different letters denote significant difference (Student’s t-test,
p < 0.01). (F) Topography of the wild type and hy5-ko cotyledon epidermal cells mapped
by AFM (top) and cell topography overlaid with elasticity (bottom). (G) Quantification of
apparent Young’s modulus. Each measurement was the average in a 5 μm by 5 μm area
of a cell with the highest modulus. Data are mean ± SD of 10 cells from three cotyledons.
Different letters denote significant difference (Student’s t-test, p < 0.001).
0
250
500
750
1000
Cell area (μm2)
0
40
80
120
Elastic modulus (MPa)
100
10
a
b
WT
WT
B
a
b
hy5-ko
0 μm
0 MPa
0
0.1
0.2
0.3
0.4
0.5
Relative pectin level
WT
a
b
WT
WT hy5-ko
D
E
G
hy5-ko
WT
C
F
hy5-ko
hy5-ko
hy5-ko
a
b
Relative galacturonic
acid level
hy5-ko
F
WT
A
Figure 12. The HY5-MIR775A-GALT9 Pathway Regulates Leaf Size.
(A) Morphology of the fifth rosette leaves of three-week-old plants from the indicated
genotypes. Bar, 5 mm. (B) Quantification of the leaf size, epidermal cell size, and pectin
level relative to the wild type. Data are mean ± SD from 10 individual plants for leaf
size, from 100 individual cells of several plants for cell size, and from three technical
replicates performed on pooled leaves for galacturonic acid level. Within each set of
measurements, different letters denote genotypes with significant difference (Student’s
t-test, p < 0.05 for leaf size; p < 0.01 for cell size and galacturonic acid level). (C-D)
Linear regression between cell sizes and organ sizes (C) and between cell sizes and
galacturonic acid levels (D) across the six genotypes.
0
0.5
1
1.5
2
0.4
0.8
1.2
1.6
F
Y = 1.2X
R = 0.89
p = 0.009
0
0.5
1
1.5
2
2.5
0.4
0.8
1.2
1.6
Y = -1.63X
R = -0.88
p = 0.0096
0
0.5
1
1.5
2
2.5
a
b
ad
a
ad
c
c
a
ad
ade
de
a
a
adc
b
c
cd
Pectin level
Leaf size
Cell size
Relative values
b
A
B
Relative cell size
Relative leaf size
Relative pectin level
C
D
Figure 13. Model for the HY5-MIR775A-GALT9 Pathway in Controlling Intrinsic Leaf Size.
HY5-MIR775A-GALT9 is a delineated double repression cascade for regulating GALT9
accumulation for leaf size determination. GALT9 participates in cell wall remodeling by
promoting the pectin constituent and reducing cell wall elasticity, which may prepare the cells
with proper resistance to turgor pressure for reaching the intrinsic size during leaf
development.
miR775
GALT9
MIR775A
cellulose
hemicellulose
pectin
Cell expansion
Intrinsic leaf size
Pectin level & wall stiffness
HY5
Supplemental Table 1. Putative CW-miRNAs and Predicted Target Genes in A. thaliana.
MiRNA
Target
Description
miR156h
AT5G38610
PECTIN METHYLESTERASE INHIBITOR
miR156i
AT1G13560
AMINOALCOHOLPHOSPHOTRANSFERASE1
AT3G01390
VACUOLAR MEMBRANE ATPASE10
AT5G38610
PECTIN METHYLESTERASE INHIBITOR SUPERFAMILY PROTEIN
miR156j
AT2G33040
GAMMA SUBUNIT OF MITOCHONDRIAL ATP SYNTHASE
AT5G38610
PECTIN METHYLESTERASE INHIBITOR SUPERFAMILY PROTEIN
miR1886
AT1G02800
GLYCOSIDE HYDROLASE FAMILY9
AT2G36870
XYLOGLUCAN ENDOTRANSGLYCOSYLASE/HYDROLASE
miR2936
AT1G15690
INORGANIC H PYROPHOSPHATASE FAMILY PROTEIN
AT1G15690
PYROPHOSPHATE-ENERGIZED INORGANIC PYROPHOSPHATASE
miR414
AT1G09210
CALRETICULIN 1B
AT1G56340
CALRETICULIN 1A
AT2G16600
ROTAMASE CYP3
AT3G25520
RIBOSOMAL PROTEIN L5
AT4G33740
MYB-LIKE PROTEIN X
AT5G12110
ELONGATION FACTOR 1-BETA 1
AT5G13850
NASCENT POLYPEPTIDE-ASSOCIATED COMPLEX SUBUNIT ALPHA-LIKE PROTEIN3
AT5G61790
CALNEXIN1
AT4G33330
GLUCURONYLTRANSFERASE
AT2G31210
BHLH TRANSCRIPTION FACTOR
AT3G50240
KINESIN-RELATED PROTEIN
miR417
AT5G66460
ENDO-BETA-MANNANASE
miR4227
AT4G12650
ENDOMEMBRANE PROTEIN 70 FAMILY
miR4239
AT3G57330
AUTOINHIBITED Ca2+-ATPASE11
miR5015
AT1G71040
LOW PHOSPHATE ROOT2
miR5021
AT1G09330
ECHIDNA GOLGI APPARATUS MEMBRANE PROTEIN-LIKE PROTEIN
AT1G10950
TRANSMEMBRANE NINE1
AT1G11310
SEVEN TRANSMEMBRANE MLO FAMILY PROTEIN
AT1G11680
CYTOCHROME P450 51G1
AT1G71940
SNARE ASSOCIATED GOLGI PROTEIN FAMILY
AT2G18840
INTEGRAL MEMBRANE YIP1 FAMILY PROTEIN
AT2G20120
CONTINUOUS VASCULAR RING
AT2G26680
FKBM FAMILY METHYLTRANSFERASE
AT3G08550
ELONGATION DEFECTIVE1
AT3G09440
HEAT SHOCK PROTEIN 70 FAMILY PROTEIN
AT3G21160
ALPHA-MANNOSIDASE2
AT3G26370
O-FUCOSYLTRANSFERASE FAMILY PROTEIN
AT3G49310
MAJOR FACILITATOR SUPERFAMILY PROTEIN
AT3G52300
ATP SYNTHASE D CHAIN
AT4G30190
H(+)-ATPASE2
AT4G30440
UDP-D-GLUCURONATE 4-EPIMERASE1
AT4G34180
CYCLASE FAMILY PROTEIN
AT5G20350
TIP GROWTH DEFECTIVE1
AT5G51570
SPFH/BAND 7/PHB DOMAIN-CONTAINING MEMBRANE-ASSOCIATED PROTEIN
AT5G23870
PECTIN ACETYLESTERASE FAMILY PROTEIN
AT5G26670
PECTIN ACETYLESTERASE FAMILY PROTEIN
AT3G26370
O-FUCOSYLTRANSFERASE FAMILY PROTEIN
AT1G24170
GALACTURONOSYLTRANSFERASE
AT4G36160
NAC-DOMAIN TRANSCRIPTION FACTOR
AT5G33290
XYLOGALACTURONAN XYLOSYLTRANSFERASE
AT4G02130
GALACTURONOSYLTRANSFERASE
AT5G61130
CALLOSE BINDING
AT1G53000
NUCLEOTIDE-DIPHOSPHO-SUGAR TRANSFERASES SUPERFAMILY PROTEIN
miR5628
AT2G02860
SUCROSE TRANSPORTER2
miR5630b
AT1G33120
RIBOSOMAL PROTEIN L6 FAMILY
miR5658
AT1G14670
ENDOMEMBRANE PROTEIN 70 FAMILY
AT1G32090
EARLY-RESPONSIVE TO DEHYDRATION4
AT3G27220
GALACTOSE OXIDASE/KELCH REPEAT SUPERFAMILY PROTEIN
AT3G47670
PLANT INVERTASE/PECTIN METHYLESTERASE INHIBITOR SUPERFAMILY PROTEIN
AT4G11220
VIRB2-INTERACTING PROTEIN2
AT5G55500
BETA-1,2-XYLOSYLTRANSFERASE
AT5G57655
XYLOSE ISOMERASE FAMILY PROTEIN
AT1G05310
PECTIN LYASE-LIKE SUPERFAMILY PROTEIN
AT2G06850
ENDOXYLOGLUCAN TRANSFERASE EXGT-A1
AT3G54920
PECTATE LYASE-LIKE PROTEIN
AT4G29230
NAC DOMAIN CONTAINING PROTEIN75
AT1G62760
PECTIN METHYLESTERASE INHIBITOR
AT1G20190
ALPHA-EXPANSIN FAMILY PROTEIN
AT3G06260
GALACTURONOSYLTRANSFERASE
AT5G62380
NAC-DOMAIN TRANSCRIPTION FACTOR
AT3G62660
GALACTURONOSYLTRANSFERASE
miR5662
AT3G49010
BREAST BASIC CONSERVED1
miR773b
AT2G26890
GRAVITROPISM DEFECTIVE2
miR775
AT1G53290
GALACTOSYLTRANSFERASE
miR776
AT2G32530
CELLULOSE SYNTHASE
miR827
AT1G63010
VACUOLAR PHOSPHATE TRANSPORTER1
miR837
AT5G24810
ABC1 FAMILY PROTEIN
miR838
AT1G43170
RIBOSOMAL PROTEIN1
AT1G51630
O-FUCOSYLTRANSFERASE FAMILY PROTEIN
AT1G51630
PRENYLATED RAB ACCEPTOR 1.B1
miR854e
AT3G56110
PRENYLATED RAB ACCEPTOR1
miR861
AT3G58730
VACUOLAR ATP SYNTHASE SUBUNIT D
AT1G71990
LEWIS-TYPE ALPHA 1,4-FUCOSYLTRANSFERASE
Supplemental Table 2. Oligonucleotide Sequences of the Primers Used in This Study.
No.
For plasmid construction
Sequence (5 to 3)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
MIR775-OX-F
MIR775-OX -R
MIR775-sgRNA1-F
MIR775-sgRNA1-R
MIR775-sgRNA2-F
MIR775-sgRNA2-R
pMIR775-F
pMIR775-R
pre-miR775-F
pre-miR775-R
GALT9-OX-F
GALT9-OX-R
GALT9-sgRNA1-exon1-F
GALT9-sgRNA1-exon1-R
GALT9-sgRNA2-3UTR-F
GALT9-sgRNA2-3UTR-R
GALT9-fLuc-N
GALT9m-fLuc-N
GALT9-GFP-F
GALT9-GFP-R
GALT9-CDS-F
GALT9-CDS-R
GALT9-Bsite-mutation-F
GALT9-Bsite-mutation-R
GALT9-mutation-F
GALT9-mutation-R
C-GFP-1305.1-F
C-GFP-1305.1-R
pGALT9-F
pGALT9-R
sgHY5-2-1F
sgHY5-2-1R
sgHY5-3-F
sgHY5-3-R
GCTCTAGAGCGTTGTTCTTCCTTCTTTGCTGAT
GGGGTACCCCTCATTTTCACATTACCACTTCGT
GATTGGCGGTTGGCGACTGAATAAG
AAACCTTATTCAGTCGCCAACCGCC
GATTGTATCAGTTGATTTTAAACAT
AAACATGTTTAAAATCAACTGATAC
GGGAAAGCTTTGTGGATAG
CATCAAGAACACGATTATG
GCTCTAGACGTTGCACTACGTGACATTGA
CATGCCATGGTGGCACTGCTAGACATCGAAA
GGGTCTAGAATGCATTCTCCTCGTAAGCT
AAAGGTACCTTCATCATCTGATGGCAAAG
GATTGACTCGCCCGCGCCGATCAA
AAACTTGATCGGCGCGTGGCGAGTC
GATTGCTTTATAAACCTCTTCTCAG
TCGACCTGCAGGCATGCAAGCTTGTCACGATTCTTACGCCT
CATGCCATGGTCGTACTGCTAGATATCGAAGACGCCAAAAACATAAAGAAAGGCC
CATGCCATGGGCGTACTCGATCATATGGAAGACGCCAAAAACATAAAGAAAGGCC
TGAACTAGTATGCATTCTCCTCGTAAGC
GCCACGCGTTCATCATCTGATGGCA
ATGCATTCTCCTCGTAAGCTA
TCATTCATCATCTGATGGCAA
TTCGTTCTCCTCGACATAGAGGAGGAGTAC
CTCTATGTCGAGGAGAACGAAGTCATCATA
TTTGTCCTACTGGACATAGAG
GAAATCGCAGAGTATGATGAC
TGAACTAGTATGCATTCTCCTCGTAAGC
GCCACGCGTTCATCATCTGATGGCA
GCATGCAAGCTTACATTTTGAGTCCGAT
GCCGCCGCCACGCGTGTGTGTGCCTAC
ATTGTGTTGTCTTAGTAGCGAAGC
AAACGCTTCGCTACTAAGACAACA
ATTGAAGACTACAATAAGAGAACT
AAACAGTTCTCTTATTGTAGTCTT
For RT-qPCR
35
36
37
38
39
40
41
42
43
5sRNA-F
5sRNA-R
miR775_qPCR_F
Actin7-F
Actin7-R
GALT9-qPCR_F
GALT9-qPCR_R
HY5-qRT-F
HY5-qRT-R
GATGCGATCATACCAGCACTAA
GATGCAACACGAGGACTTCCC
GCTTCGATGTCTAGCAGTGCCA
GGTGTCATGGTTGGTATGGGTC
CCTCTGTGAGTAGAACTGGGTGC
TATCGAAGAGGAGTACAGTAAG
TAGCAGAGAGAGTCGATCTG
CCATCAAGCAGCGAGAGGTCATCAA
CGCCGATCCAGATTCTCTACCGGAA
For genotyping
44
45
46
47
48
49
50
51
5 RACE -RPM-F
GALT9-GSP-R
LBb1.3
SALK015338-LP
SALK015338-RP
MIR775-KO-F
MIR775-KO-R
GALT9-KO-F
CTAATACGACTCACTATAGGGCAAGCAGTGGTATCAACGCAGAGT
GATTACGCCAAGCTTATTCATTGCCAGCATCCACGCACCT
ATTTTGCCGATTTCGGAAC
GATGGCTAACCCCGTAGATTC
TGCGATAGCTGGTAGACAACAC
TGACTCTCATGGCTGTGTCAG
AGCTTGTAGGGGAAAGGGAGATAG
TCGAGCTTCCTTGACACCAC
52
53
54
55
56
GALT9-KO-R
hy5_215-F
hy5_215-R
HY5-CRISPR-F
HY5-CRISPR-R
TGCAGGTTCGCTCGAAGAAA
GTCATCAAGCTCTGCTCCACAT
AAGACACCTCTTCAGCCGCTTG
CAGAGATCTGACGGCGGTA
CCTTTCTACTACAGTGTCAC
A
B
Supplemental Figure 1. Comparison of Pre-miR775a Homologs in A. thaliana and A. lyrata.
(A) Alignment of pre-miR775a sequences from five representative A. thaliana ecotypes with the
closest homolog in A. lyrata. Sequences are 29,422,419-29,422,603 on A. thaliana (Col-0)
chromosome 1 and 18,060,424-18,060,639 on A. lyrate chromosome 2. Region corresponding to
mature miR775 is underlined in red. (B) Predicted secondary structures from sequences in A.
Red lines indicate the region corresponding to miR775 in A. thaliana. Supports Figure 1 in the
main manuscript.
A. lyrata
A. thaliana (Col-0)
A. thaliana (Cvi-0)
A. thaliana (Bur-0)
A. thaliana (Ler-0)
A. thaliana (Ws-0)
AATATAA-----GATGGTGACGAACGACTGAATAAAATGACTTAAAC--TGCGGTTACGTGGTCATTTGAGAACTGTGATGAGT
AACATCATGGCGGTTGG-------CGACTGAATAAGAGGATTTAAACGTTGC-ACTACGT-GACA-TTGA-AACTGT-------
AACATCNTGGCGGTTGG-------CGACTGAATAAGAGGATTTAAACGTTGC-ACTACGT-GACA-TTGA-AACTGT-------
AACATCNTGGCGGTTGG-------CGACTGAATAAGAGGATTTAAACGTTGC-ACTACGT-GACA-TTGA-AACTGT-------
AACATCNTGGCGGTTGG-------CGACTGAATAAGANNNNTTAAACGTTGC-ACTACGT-GACA-TTGA-AACTGT-------
AACATCNTGGCGGTTGG-------CGACTGAATAAGAGGATTTAAACGTTGC-ACTACGT-GACA-TTGA-AACTGT-------
ATACAATGGTTTTTATGCTCACGACAATTTTCAAAGCATCTCTATGTTTATGCTCATCACAGTTCTTGATTACCCACTAAACCG
---------------------------CTTTCAA--CATTCCAATATTT----------CAACTTTCGAATACCCAATATTTGG
---------------------------CTTTCAA--CATTCCAATATTT----------CAACTTTCGAATACCCAATATTTGG
---------------------------CTTTCAA--CATTCCAATATTT----------CAACTTTCGAATACCCAATATTTGG
---------------------------CTTTCAA--CATTCCAATATTT----------CAACTTTCGAATACCCAATATTTGG
---------------------------CTTTCAA--CATTCCAATATTT----------CAACTTTCGAATACCCAATATTTGG
ATGTTTAAAAAACCTTT-------------------ATGTTT-AAACCAA---ATTATTTGTCTCCCAT---ATT-ATCCGT
TTTGTTCAAAGACATTTTCGATGTCTAGCAGTGCCAATGTTTAAAATCAACTGATAATTT--------TGGAATTAATGTGT
TTTGTTNAAAGACATTTTCGATGTCTAGCAGTGCCAATGTTNAAAATCANCTGATAATTT--------TGGAATTAATGTGT
TTTGTTNAAAGACATTTTCGATGTCTAGCAGTGCCAATGTTNAAAATCANCTGATAATTT--------TGGAATTAATGTGT
TTTGTTNAAAGACATTTTCGATGTCTAGCAGTGCCAATGTTNAAAATCANCTGATAATTT--------TGGAATTAATGTGT
TTTGTTNGAAGACATTTTCGATGTCTAGCAGTGCCAATGTTNAAAATCANCTGATAATTT--------TGGAATTAATGTGT
A. lyrata
A. thaliana (Col-0)
A. thaliana (Cvi-0)
A. thaliana (Bur-0)
A. thaliana (Ler-0)
A. thaliana (Ws-0)
A. lyrata
A. thaliana (Col-0)
A. thaliana (Cvi-0)
A. thaliana (Bur-0)
A. thaliana (Ler-0)
A. thaliana (Ws-0)
A. lyrata
Col-0
Cvi-0
Bur-0
Ler-0
Ws-0
5’
3’
5’
3’
5’ 3’
5’ 3’
5’ 3’
5’ 3’
A. thaliana
A
0.1
Supplemental Figure 2. MiR775 Specifically Targets GALT9 in A. thaliana.
(A) Phylogeny of representative members of the glycosyltransferase 31 family. Shown is an
unrooted neighbor joining tree built with the JTT model. Bootstrap values are from 1,000
iterations. Circles indicate branches with a bootstrap value > 60. The clade containing GATL9 is
shade in green. Genes known for involvement in primary cell wall biosynthesis are highlighted
in red. (B) Sequence alignment at the miR775 binding site, shown in bold, between GALT9 and
two closest homologs in A. thaliana. Nucleotides undermining complementarity with miR775 are
shown in red. The MFE/MED ratios are shown on the right, which indicate that only GALT9 is a
potential target for miR775. Supports Figure 1 in the main manuscript.
ACCGUGACGAUCUGUAGCUU
: :.:::::::::.::::::
UAUGAUGACUUCGUACUGCUAGAUAUCGAAGAGGAGUAC
UACGAUGACUUUAUACUGCUCGAUAUCGAGGAGGAGUAC
UACAGAGAUUUUGUGCUUCUUGAUACCGAGGAAGAAUAU
miR775
GALT9
AT3G14960
AT2G26100
0.80
0.62
0.44
MFE/MED
B
Relative miR775 level
0
10
20
30
40
A
B
C
D
E
Supplemental Figure 3. Characterization of MIR775A-OX Lines.
(A-D) Morphological comparison of the indicated lines. Shown are cotyledon of eight-day-old
seedlings (A), seedling showcasing the hypocotyl (B), the fifth rosette leaf of three-week-old
plants (C), and petal of open flowers (D). Bars, 2 mm. MIR775A-OX was generated by
expressing the 35S:pre-miR775a transgene (pre-miR775a under control of the enhanced
35S promoter) in A. thaliana. Seventeen independent T1 lines were obtained and four further
analyzed at the T2 generation. Line #8 was selected for subsequent analyses. (E) RT-qPCR
analysis of relative miR775 abundance in the selected lines. Data are means ± SD from
three technical replicates. Supports Figures 2-4 in the main manuscript.
1
337
460
Supplemental Figure 4. Generation and Characterization of the mir775 Mutant Lines.
(A) Diagram showing the CRISPR/Cas9 vector for simultaneously introducing Cas9 with paired
sgRNAs. (B) Scheme for generating mir775 deletion using the CRISPR/Cas9 system. Numbers
mark positions according to the full length cDNA BX818024. The paired sgRNAs are designed
to delete a 123 bp region encompassing pre-miR775a. Sequence comparison for a typical
deletion allele with reference to the wild type allele is shown on the bottom. (C) Genotyping
result for 10 independent homozygous mir775 lines. Genomic DNA from individual deletion lines
was PCR-amplified and gel-separated. Size polymorphisms according to the wild type and
deletion alleles are indicated. Lines #18 and #45 were selected for subsequent analyses. (D)
RT-qPCR analysis of relative miR775 abundance in the two selected lines. Data are means ±
SD from three technical replicates. (E-H) Morphological comparison of the indicated lines. From
left to right: eight-day-old seedlings showcasing the cotyledon (E), seedlings showcasing the
hypocotyl (F), the fifth rosette leaves of three-week-old plants (G), and petals of open flowers
(H). Bars, 2 mm. Supports Figures 3 and 4 in the main manuscript.
pre-miR775a
atcatggcggttggcgactgaat---------------------------tttaaaatcaactgataattttggaatt
Wild type:
mir775:
atcatggcggttggcgactgaataagaggatt
gtgccaatgtttaaaatcaactgataattttggaatt
AtU6-26
AtU6-26
sgRNA1
2×35S
NLS
3×Flag
hSpCas9n
NLS
Nos Ter
sgRNA2
mir775
MIR775A
0
0.5
1
1.5
Relative miR775 level
A
C
E
B
F
D
G
H
0
1
2
3
4
A
B
Relative miR775 level
a
b
c
a
C
Supplemental Figure 5. Characterization of the MIR775A-OX mir775 Line.
(A-B) Morphological comparison of the indicated genotypes. Eight-day-old seedlings were
photographed to showcase the cotyledon (A) and the hypocotyl (B). MIR775A-OX mir775
was created by crossing T3 generation MIR775A-OX line #8 to mir775. F2 progenies
homozygous for mir775 and resistant to BASTA (MIR775A-OX positive) were selected for
analyses. Bars, 2 mm. (C) RT-qPCR analysis of relative miR775 abundance in the indicated
genotypes. Data are means ± SD from three technical replicates. Different letters denote
groups with significant difference (Student’s t-test, p < 0.001). Supports Figures 3 and 4 in
the main manuscript.
GALT9
A
galt9-2 (SALK_015338)
Supplemental Figure 6. Generation and Characterization of the galt9 Mutant Lines.
(A) Scheme for generating galt9 deletion mutants using the CRISPR/Cas9 system. Exons of
GALT9 are shown as horizontal boxes. Two sgRNAs are designed to create paired cleavage
sites positioned at 110 and 1,991, resulting in a 1,882 bp deletion. The corresponding mutant
was named galt9-1. A T-DNA insertion line (SALK_015338) with the T-DNA inserted into the
start codon was named galt9-2. (B) Genotyping result for the deletion lines. A total of seven
independent homozygous lines were identified. PCR product corresponding to the wild type
allele is marked. Lines #4 and #9 were selected for subsequent analyses. (C) RT-qPCR
analysis of relative GALT9 transcript levels in the indicated lines in comparison to the wild
type. Data are means ± SD from three technical replicates. (D-E) Morphology of the fifth
rosette leaf (D) and petal (E) of the indicated genotypes. Bars, 2 mm. Supports Figures 3 and
4 in the main manuscript.
110
1,991
tcatcactcgccacgcgccgatcaacgg
tgtctttataaacctcttctcagtggtcgaagctctatca
tcatcactcgccacgcgccgatca---------------------------------gtggtcgaagctctatca
Wild type:
galt9-1:
galt9-1
B
C
D
E
Relative GALT9 level
0
0.5
1
1.5
Supplemental Figure 7. Characterization of the GALT9-OX Lines.
(A-B) Morphological comparison of the fifth rosette leaf from three-week-old plants (A)
and petal from open flowers (B). Bars, 2 mm. GALT9-OX was generated by expressing
the GALT9 coding region under control of the enhanced 35S promoter in A. thaliana.
Twelve independent T1 lines were obtained and four further analyzed at the T2 generation.
GALT9m-OX was generated by substituting the nucleotides of the miR775 binding site in
GALT9 but not the encoded amino acids. Six independent T1 lines were obtained and two
further analyzed at the T2 generation. (C) RT-qPCR analysis of relative GALT9 transcript
levels in the indicated lines in comparison to the wild type. Data are means ± SD from
three technical replicates. Supports Figures 3 and 4 in the main manuscript.
A
B
0
4
8
12
16
20
Relative GALT9 level
0
1
2
3
4
5
C
Supplemental Figure 8. Degradome Sequencing Profiles of Predicted MiR775 Targets.
Degradome sequencing data were obtained from the wild type and MIR775A-OX plants.
Shown on top are normalized frequencies of reads with unique 5’ ends mapped to the four
potential miR775 target genes. Enlarged views at the predicted miR775-binding sites are
shown on the bottom along with base pairing pattern to miR775. Supports Figure 2 in the
main manuscript.
0.1
0.2
0.3
0.04
0.08
0.12
0.16
RPM
RPM
160
360
560
760
200
5000
1800
3400
0.1
0.2
0.3
0.04
0.08
0.12
0.16
RPM
RPM
GALT9
DCL1
5’UCGUACUGCUAGAUAUCGAA3’
: :.:::::::::.::::::
3’ACCGUGACGAUCUGUAGCUU5’
5’UGGAACUGCUAGACAUAGAG3’
::: :::::::::::: ::.
3’ACCGUGACGAUCUGUAGCUU5’
WT
MIR775A-OX
1
2
3
4
100 200 300 400 500 600
0.5
1
0.05
0.1
0.15
0.2
0.25
1000
2000
3000
4000
AT1G23390
AT4G12020
5’UGGAGCUGUUCGACAUCGAA3’
::: .:::.: :::::::::
3’ACCGUGACGAUCUGUAGCUU5’
5’UGUCACUGCUAUGCAUUGAG3’
:: :::::::: .:::.::.
3’ACCGUGACGAUCUGUAGCUU5’
RPM
RPM
RPM
RPM
Supplemental Figure 9. Phenotypic Comparison of the galt9 and dcl1 Mutants.
(A-C) Morphological comparison of the indicated genotypes. Photographs of eight-day-old
cotyledons (A), petals of open flowers (B), and mature siliques (C) are shown on the left.
Bars, 2 mm. Quantifications of the relative cotyledon area, petal area, and silique length
are shown on the right. Data are means ± SD from 30 individual organs normalized to
the wild type. Different letters denote genotypes with significant difference (Student’s t-
test, p < 0.01). Supports Figures 2-4 in the main manuscript.
0
0.5
1
1.5
0
0.4
0.8
1.2
1.6
0
0.5
1
1.5
2
a
c
d
b
c
a
b
c
b
Relative silique length
Relative petal area
Relative cotyledon area
a
b
b
C
A
B
100 MPa
10 μm
Supplemental Figure 10. Analysis of the qrt2 Mutant Defective in Pectin Turnover.
(A) Examination of cell wall pectin by Raman microscopy. Cotyledon cells of seven-day-
old wild type, mir775, GALT9-OX, and qrt2 seedlings were imaged for pectin. Bar, 20 μm.
(B) Topography of the wild type and qrt2 cotyledon epidermal cells mapped by AFM (left)
and topography overlaid with elasticity (right). Bar, 5 μm. Supports Figures 8 and 9 in the
main manuscript.
WT
qrt2
WT
mir775
GALT9-OX
qrt2
B
A
Topography
Elasticity
0
0
pMIR775A:GUS/hy5
pMIR775A:GUS
pMIR775A:GUS
pMIR775A:GUS/hy5
0
0.5
1
1.5
2
2.5
E
D
B
Shoot
Root
WT
hy5
Relative miR775 level
Supplemental Figure 11. HY5 Differentially Regulates MIR775A in the Shoot and the Root.
(A) GUS staining for pMIR775A activities in the wild type and hy5-215 backgrounds. Ten- (left) and
12-day-old (right) pMIR775A:GUS and pMIR775A:GUS/hy5-215 seedlings (right) were stained for
GUS activity. Bars, 1 mm. (B) The pMIR775A:GUS and pMIR775A:GUS/hy5 adult plants with
approximately ten true leaves were stained for GUS activity. Bar, 2 cm. (C-D) Root tips of
pMIR775A:GUS and pMIR775A:GUS/hy5 at the seedling (C) and adult (D) stages were compared
for GUS activity. Bars, 50 μm. (E) Quantitative analysis of relative miR775 levels separately in the
shoot and the root of wild type and hy5-215 seedlings by RT-qPCR. Data are means ± SD from three
technical replicates. Supports Figure 10 in the main manuscript.
pMIR775A:GUS
A
pMIR775A:GUS
pMIR775A:GUS/hy5
C
pMIR775A:GUS/hy5
0
1
2
3
Cotyledon size (mm²)
0
2000
4000
6000
8000
Cell size (μm²)
A
hy5-215 (GA)
225
1,611
Wild type:
hy5-ko:
hy5-ko
HY5
B
D
G
Supplemental Figure 12. Generation and Characterization of Mutants for HY5.
(A) Scheme for generating the hy5-ko allele using CRISPR/Cas9. Two sgRNAs are designed to
create paired cleavage sites resulting in a 1,386 bp deletion. The hy5-215 allele harbors a point
mutation near the end of the first intron that interferes splicing. (B) Genotyping result with PCR
products according to the wild type and deletion alleles indicated. Lines #4 and #5 were selected
for subsequent analyses. (C-D) Morphological comparison and quantification of cotyledon size.
Data are mean ± SD from 10 individual seedlings. Different letters denote genotypes with
significant difference (Student’s t-test, p < 0.05). Bar, 2 mm. (E) Morphological comparison of
adult plants. Bar, 5 mm. (F-G) SEM analysis of the cotyledon epidermal cells. Bar, 100 μm. (G)
Quantification of the cotyledon epidermal cell size. Data are mean ± SD from 100 individual cells
from three seedlings. Different letters denote genotypes with significant difference (Student’s t-
test, p < 0.05). Supports Figures 11 and 12 in the main manuscript.
ttttcaccagc-------------------------------ttctcttattgtagtcttagatttctcttaattgaa
ttttcaccagcttcgctactaagacaacaaat agttctcttattgtagtcttagatttctcttaattgaa
a
b
b
c
a
b
c
a
WT
hy5-215
hy5-ko
HY5-OX
WT
hy5-215
hy5-ko
HY5-OX
C
E
F
WT
hy5-215
hy5-ko
HY5-OX
Parsed Citations
Addo-Quaye, C., Miller, W., and Axtell, M.J. (2009). CleaveLand: a pipeline for using degradome data to find cleaved small RNA targets.
Bioinformatics 25, 130-131.
Google Scholar: Author Only Title Only Author and Title
Alves, L., Jr., Niemeier, S., Hauenschild, A., Rehmsmeier, M., and Merkle, T. (2009). Comprehensive prediction of novel microRNA
targets in Arabidopsis thaliana. Nucleic Acids Res 37, 4010-4021.
Google Scholar: Author Only Title Only Author and Title
Andres-Robin, A., Reymond, M.C., Dupire, A., Battu, V., Dubrulle, N., Mouille, G., Lefebvre, V., Pelloux, J., Boudaoud, A., Traas, J., Scutt,
C.P., and Monéger, F. (2018). Evidence for the regulation of gynoecium morphogenesis by ETTIN via cell wall dynamics. Plant Physiol
178, 1222-1232.
Google Scholar: Author Only Title Only Author and Title
Atmodjo, M.A., Hao, Z.Y., and Mohnen, D. (2013). Evolving views of pectin biosynthesis. Annu Rev Plant Biol, 64, 747-779.
Google Scholar: Author Only Title Only Author and Title
Bogre, L., Magyar, Z., and Lopez-Juez, E. (2008). New clues to organ size control in plants. Genome Biol 9, 226.
Google Scholar: Author Only Title Only Author and Title
Bouton, S., Leboeuf, E., Mouille, G., Leydecker, M.T., Talbotec, J., Granier, F., Lahaye, M., Hofte, H., and Truong, H.N. (2002).
Quasimodo1 encodes a putative membrane-bound glycosyltransferase required for normal pectin synthesis and cell adhesion in
Arabidopsis. Plant Cell 14, 2577-2590.
Google Scholar: Author Only Title Only Author and Title
Braybrook, S.A., and Kuhlemeier, C. (2010). How a plant builds leaves. Plant Cell 22, 1006-1018.
Google Scholar: Author Only Title Only Author and Title
Brown, B.A., and Jenkins, G.I. (2008). UV-B signaling pathways with different fluence-rate response profiles are distinguished in mature
Arabidopsis leaf tissue by requirement for UVR8, HY5, and HYH. Plant Physiol 146, 576-588.
Google Scholar: Author Only Title Only Author and Title
Burko, Y., Seluzicki, A., Zander, M., Pedmale, U.V., Ecker, J.R., and Chory, J. (2020). Chimeric activators and repressors define HY5
activity and reveal a light-regulated feedback mechanism. Plant Cell 32, 967-983.
Google Scholar: Author Only Title Only Author and Title
Caffall, K.H., and Mohnen, D. (2009). The structure, function, and biosynthesis of plant cell wall pectic polysaccharides. Carbohyd Res
344, 1879-1900.
Google Scholar: Author Only Title Only Author and Title
Cosgrove, D.J. (2005). Growth of the plant cell wall. Nat Rev Mol Cell Biol 6, 850-861.
Google Scholar: Author Only Title Only Author and Title
Czesnick, H., and Lenhard, M. (2015). Size control in plants - lessons from leaves and flowers. Cold Spring Harb Perspect Biol 7,
a019190.
Google Scholar: Author Only Title Only Author and Title
Dai, X.B., and Zhao, P.X. (2011). psRNATarget: a plant small RNA target analysis server. Nucleic Acids Res 39, W155-W159.
Google Scholar: Author Only Title Only Author and Title
Du, L., Li, N., Chen, L.L., Xu, Y.X., Li, Y., Zhang, Y.Y., Li, C.Y., and Li, Y.H. (2014). The ubiquitin receptor DA1 regulates seed and organ
size by modulating the stability of the ubiquitin-specific protease UBP15/SOD2 in Arabidopsis. Plant Cell 26, 665-677.
Google Scholar: Author Only Title Only Author and Title
Efroni, I., Blum, E., Goldshmidt, A., and Eshed, Y. (2008). A protracted and dynamic maturation schedule underlies arabidopsis leaf
development. Plant Cell 20, 2293-2306.
Google Scholar: Author Only Title Only Author and Title
Efroni, I., Eshed, Y., and Lifschitz, E. (2010). Morphogenesis of simple and compound leaves: a critical review. Plant Cell 22, 1019-1032.
Google Scholar: Author Only Title Only Author and Title
Esch, J.J., Chen, M.A., Hillestad, M., and Marks, M.D. (2004). Comparison of TRY and the closely related At1g01380 gene in controlling
Arabidopsis trichome patterning. Plant J 40, 860-869.
Google Scholar: Author Only Title Only Author and Title
Felippes, F.F., Schneeberger, K., Dezulian, T., Huson, D.H., and Weigel, D. (2008). Evolution of Arabidopsis thaliana microRNAs from
random sequences. RNA 14, 2455-2459.
Google Scholar: Author Only Title Only Author and Title
Ferjani, A., Horiguchi, G., Yano, S., and Tsukaya, H. (2007). Analysis of leaf development in fugu mutants of Arabidopsis reveals three
compensation modes that modulate cell expansion in determinate organs. Plant Physiol 144, 988-999.
Google Scholar: Author Only Title Only Author and Title
Gao, Z., Li, J., Li, L., Yang, Y., Li, J., Fu, C., Zhu, D., He, H., Cai, H., and Li, L. (2020). Structural and functional analyses of hub microRNAs
in an integrated gene regulatory network of Arabidopsis. Genomics Proteomics Bioinformatics. In press.
Google Scholar: Author Only Title Only Author and Title
Geitmann, A., and Ortega, J.K. (2009). Mechanics and modeling of plant cell growth. Trends Plant Sci 14, 467-478.
Google Scholar: Author Only Title Only Author and Title
German, M.A., Luo, S.J., Schroth, G., Meyers, B.C., and Green, P.J. (2009). Construction of Parallel Analysis of RNA Ends (PARE)
libraries for the study of cleaved miRNA targets and the RNA degradome. Nat Protoc 4, 356-362.
Google Scholar: Author Only Title Only Author and Title
German, M.A., Pillay, M., Jeong, D.H., Hetawal, A., Luo, S.J., Janardhanan, P., Kannan, V., Rymarquis, L.A., Nobuta, K., German, R., De
Paoli, E., Lu, C., Schroth, G., Meyers, B.C., and Green, P.J. (2008). Global identification of microRNA-target RNA pairs by parallel
analysis of RNA ends. Nat Biotechnol 26, 941-946.
Google Scholar: Author Only Title Only Author and Title
Gierlinger, N., Keplinger, T., and Harrington, M. (2012). Imaging of plant cell walls by confocal Raman microscopy. Nat Protoc 7, 1694-
1708.
Google Scholar: Author Only Title Only Author and Title
Gonzalez, N., Vanhaeren, H., and Inze, D. (2012). Leaf size control: complex coordination of cell division and expansion. Trends Plant
Sci 17, 332-340.
Google Scholar: Author Only Title Only Author and Title
Gruber, A.R., Lorenz, R., Bernhart, S.H., Neuböck, R., and Hofacker, I.L. (2008). The Vienna RNA websuite. Nucleic Acids Res 36, W70-
W74.
Google Scholar: Author Only Title Only Author and Title
Guo, Z., Kuang, Z., Wang, Y., Zhao, Y., Tao, Y., Cheng, C., Yang, J., Lu, X., Hao, C., Wang, T., Cao, X., Wei, J., Li, L., and Yang, X. (2020).
PmiREN: a comprehensive encyclopedia of plant miRNAs. Nucleic Acids Res 48, D1114-D1121.
Google Scholar: Author Only Title Only Author and Title
Haas, K.T., Wightman, R., Meyerowitz, E.M., and Peaucelle, A. (2020). Pectin homogalacturonan nanofilament expansion drives
morphogenesis in plant epidermal cells. Science 367, 1003-1007.
Google Scholar: Author Only Title Only Author and Title
Harholt, J., Suttangkakul, A., and Scheller, H.V. (2010). Biosynthesis of pectin. Plant Physiol 153, 384-395.
Google Scholar: Author Only Title Only Author and Title
Hepworth, J., and Lenhard, M. (2014). Regulation of plant lateral-organ growth by modulating cell number and size. Curr Opin Plant Biol
17, 36-42.
Google Scholar: Author Only Title Only Author and Title
Hirayama, T., Kieber, J.J., Hirayama, N., Kogan, M., Guzman, P., Nourizadeh, S., Alonso, J.M., Dailey, W.P., Dancis, A., and Ecker, J.R.
(1999). RESPONSIVE-TO-ANTAGONIST1, a Menkes/Wilson disease-related copper transporter, is required for ethylene signaling in
Arabidopsis. Cell 97, 383-393.
Google Scholar: Author Only Title Only Author and Title
Hong, L.L., Dumond, M., Zhu, M.Y., Tsugawa, S., Li, C.B., Boudaoud, A., Hamant, O., and Roeder, A.H.K. (2018). Heterogeneity and
robustness in plant morphogenesis: from cells to organs. Annu Rev Plant Biol, 69, 469-495.
Google Scholar: Author Only Title Only Author and Title
Huang, L., Li, X.H., Zhang, W.W., Ung, N., Liu, N.N., Yin, X.L., Li, Y., Mcewan, R.E., Dilkes, B., Dai, M.J., Hicks, G.R., Raikhel, N.V., Staiger,
C.J., and Zhang, C.H. (2020). Endosidin20 targets the cellulose synthase catalytic domain to inhibit cellulose biosynthesis. Plant Cell 32,
2141-2157.
Google Scholar: Author Only Title Only Author and Title
Johnson, K., and Lenhard, M. (2011). Genetic control of plant organ growth. New Phytol 191, 319-333.
Google Scholar: Author Only Title Only Author and Title
Kant, S., Peng, M.S., and Rothstein, S.J. (2011). Genetic regulation by NLA and microRNA827 for maintaining nitrate-dependent
phosphate homeostasis in Arabidopsis. PLoS Genet 7, e1002021.
Google Scholar: Author Only Title Only Author and Title
Kawade, K., Horiguchi, G., and Tsukaya, H. (2010). Non-cell-autonomously coordinated organ size regulation in leaf development.
Development 137, 4221-4227.
Google Scholar: Author Only Title Only Author and Title
Kozomara, A., Birgaoanu, M., and Griffiths-Jones, S. (2018). miRBase: from microRNA sequences to function. Nucleic Acids Res 47,
D155-D162.
Google Scholar: Author Only Title Only Author and Title
Langmead, B., and Salzberg, S.L. (2012). Fast gapped-read alignment with Bowtie 2. Nat Methods 9, 357-359.
Google Scholar: Author Only Title Only Author and Title
Li, J., Yu, M.A., Geng, L.L., and Zhao, J. (2010). The fasciclin-like arabinogalactan protein gene, FLA3, is involved in microspore
development of Arabidopsis. Plant J. 64, 482-497.
Google Scholar: Author Only Title Only Author and Title
Liu, Q.K., Wang, F., and Axtell, M.J. (2014). Analysis of complementarity requirements for plant microRNA targeting using a Nicotiana
benthamiana quantitative transient assay. Plant Cell 26, 741-753.
Google Scholar: Author Only Title Only Author and Title
Liu, S., Li, J.H., Wu, J., Zhou, K.R., Zhou, H., Yang, J.H., and Qu, L.H. (2015). StarScan: a web server for scanning small RNA targets from
degradome sequencing data. Nucleic Acids Res 43, W480-W486.
Google Scholar: Author Only Title Only Author and Title
Llave, C., Xie, Z.X., Kasschau, K.D., and Carrington, J.C. (2002). Cleavage of Scarecrow-like mRNA targets directed by a class of
Arabidopsis miRNA. Science 297, 2053-2056.
Google Scholar: Author Only Title Only Author and Title
Ma, Z.R., Coruh, C., and Axtell, M.J. (2010). Arabidopsis lyrata small RNAs: transient MIRNA and small interfering RNA loci within the
Arabidopsis genus. Plant Cell 22, 1090-1103.
Google Scholar: Author Only Title Only Author and Title
Mallory, A.C., and Vaucheret, H. (2006). Functions of microRNAs and related small RNAs in plants. Nat Genet 38, S31-36.
Google Scholar: Author Only Title Only Author and Title
Mao, Y., Zhang, H., Xu, N., Zhang, B., Gou, F., and Zhu, J.K. (2013). Application of the CRISPR-Cas system for efficient genome
engineering in plants. Mol Plant 6, 2008-2011.
Google Scholar: Author Only Title Only Author and Title
Mizukami, Y., and Fischer, R.L. (2000). Plant organ size control: AINTEGUMENTA regulates growth and cell numbers during
organogenesis. Proc Natl Acad Sci USA 97, 942-947.
Google Scholar: Author Only Title Only Author and Title
Obayashi, T., Aoki, Y., Tadaka, S., Kagaya, Y., and Kinoshita, K. (2018). ATTED-II in 2018: a plant coexpression database based on
investigation of the statistical property of the mutual rank index. Plant Cell Physiol 59, e3.
Google Scholar: Author Only Title Only Author and Title
Ori, N., Cohen, A.R., Etzioni, A., Brand, A., Yanai, O., Shleizer, S., Menda, N., Amsellem, Z., Efroni, I., Pekker, I., Alvarez, J.P., Blum, E.,
Zamir, D., and Eshed, Y. (2007). Regulation of LANCEOLATE by miR319 is required for compound-leaf development in tomato. Nat
Genet 39, 787-791.
Google Scholar: Author Only Title Only Author and Title
Oyama, T., Shimura, Y., and Okada, K. (1997). The Arabidopsis HY5 gene encodes a bZIP protein that regulates stimulus-induced
development of root and hypocotyl. Gene Dev 11, 2983-2995.
Google Scholar: Author Only Title Only Author and Title
Palatnik, J.F., Allen, E., Wu, X.L., Schommer, C., Schwab, R., Carrington, J.C., and Weigel, D. (2003). Control of leaf morphogenesis by
microRNAs. Nature 425, 257-263.
Google Scholar: Author Only Title Only Author and Title
Palin, R., and Geitmann, A. (2012). The role of pectin in plant morphogenesis. Biosystems 109, 397-402.
Google Scholar: Author Only Title Only Author and Title
Pan, J.W., Huang, D.H., Guo, Z.L., Kuang, Z., Zhang, H., Xie, X.Y., Ma, Z.F., Gao, S.P., Lerdau, M.T., Chu, C.C., and Li, L. (2018).
Overexpression of microRNA408 enhances photosynthesis, growth, and seed yield in diverse plants. J Integr Plant Biol 60, 323-340.
Google Scholar: Author Only Title Only Author and Title
Parsons, H.T., Christiansen, K., Knierim, B., Carroll, A., Ito, J., Batth, T.S., Smith-Moritz, A.M., Morrison, S., McInerney, P., Hadi, M.Z.,
Auer, M., Mukhopadhyay, A., Petzold, C.J., Scheller, H.V., Loque, D., and Heazlewood, J.L. (2012). Isolation and proteomic
characterization of the Arabidopsis Golgi defines functional and novel components involved in plant cell wall biosynthesis. Plant
Physiol 159, 12-26.
Google Scholar: Author Only Title Only Author and Title
Peaucelle, A., Wightman, R., and Hofte, H. (2015). The control of growth symmetry breaking in the Arabidopsis hypocotyl. Cur Biol 25,
1746-1752.
Google Scholar: Author Only Title Only Author and Title
Powell, A.E., and Lenhard, M. (2012). Control of organ size in plants. Curr Biol 22, R360-367.
Google Scholar: Author Only Title Only Author and Title
Qi, J.Y., Wu, B.B., Feng, S.L., Lu, S.Q., Guan, C.M., Zhang, X., Qiu, D.L., Hu, Y.C., Zhou, Y.H., Li, C.Y., Long, M., and Jiao, Y.L. (2017).
Mechanical regulation of organ asymmetry in leaves. Nat Plants 3, 724-733.
Google Scholar: Author Only Title Only Author and Title
Qin, L.X., Chen, Y., Zeng, W., Li, Y., Gao, L., Li, D.D., Bacic, A., Xu, W.L., and Li, X.B. (2017). The cotton beta-galactosyltransferase 1
(GalT1) that galactosylates arabinogalactan proteins participates in controlling fiber development. Plant J 89, 957-971.
Google Scholar: Author Only Title Only Author and Title
Qu, Y., Egelund, J., Gilson, P.R., Houghton, F., Gleeson, P.A., Schultz, C.J., and Bacic, A. (2008). Identification of a novel group of
putative Arabidopsis thaliana beta-(1,3)-galactosyltransferases. Plant Mol Biol 68, 43-59.
Google Scholar: Author Only Title Only Author and Title
Rajagopalan, R., Vaucheret, H., Trejo, J., and Bartel, D.P. (2006). A diverse and evolutionarily fluid set of microRNAs in Arabidopsis
thaliana. Gene Dev 20, 3407-3425.
Google Scholar: Author Only Title Only Author and Title
Rhee, S.Y., and Somerville, C.R. (1998). Tetrad pollen formation in quartet mutants of Arabidopsis thaliana is associated with
persistence of pectic polysaccharides of the pollen mother cell wall. Plant J 15, 79-88.
Google Scholar: Author Only Title Only Author and Title
Rodriguez, R.E., Mecchia, M.A., Debernardi, J.M., Schommer, C., Weigel, D., and Palatnik, J.F. (2010). Control of cell proliferation in
Arabidopsis thaliana by microRNA miR396. Development 137, 103-112.
Google Scholar: Author Only Title Only Author and Title
Rodriguez, R.E., Schommer, C., and Palatnik, J.F. (2016). Control of cell proliferation by microRNAs in plants. Curr Opin Plant Biol 34,
68-76.
Google Scholar: Author Only Title Only Author and Title
Rogers, K., and Chen, X.M. (2013). Biogenesis, turnover, and mode of action of plant micrornas. Plant Cell 25, 2383-2399.
Google Scholar: Author Only Title Only Author and Title
Salvador-Guirao, R., Baldrich, P., Weigel, D., Rubio-Somoza, I., and San Segundo, B. (2018). The microRNA miR773 is involved in the
Arabidopsis immune response to fungal pathogens. Mol Plant Microbe Interact 31, 249-259.
Google Scholar: Author Only Title Only Author and Title
Schommer, C., Debernardi, J.M., Bresso, E.G., Rodriguez, R.E., and Palatnik, J.F. (2014). Repression of cell proliferation by miR319-
regulated TCP4. Mol Plant 7, 1533-1544.
Google Scholar: Author Only Title Only Author and Title
Sibout, R., Sukumar, P., Hettiarachchi, C., Holm, M., Muday, G.K., and Hardtke, C.S. (2006). Opposite root growth phenotypes of hy5
versus hy5 hyh mutants correlate with increased constitutive auxin signaling. PLoS Genet 2, e202.
Google Scholar: Author Only Title Only Author and Title
Somerville, C. (2006). Cellulose synthesis in higher plants. Annu Rev Cell Dev Biol 22, 53-78.
Google Scholar: Author Only Title Only Author and Title
Somerville, C., Bauer, S., Brininstool, G., Facette, M., Hamann, T., Milne, J., Osborne, E., Paredez, A., Persson, S., Raab, T., Vorwerk,
S., and Youngs, H. (2004). Toward a systems approach to understanding plant cell walls. Science 306, 2206-2211.
Google Scholar: Author Only Title Only Author and Title
Song, Y.H., Yoo, C.M., Hong, A.P., Kim, S.H., Jeong, H., Shin, S.Y., Kim, H.J., Yun, D.J., Lim, C.O., Bahk, J.D., Lee, S.Y., Nagao, R.T., Key,
J.L., and Hong, J.C. (2008). DNA-binding study identifies C-box and hybrid C/G-box or C/A-box motifs as high-affinity binding sites for
STF1 and LONG HYPOCOTYL5 proteins. Plant Physiol 146, 1862-1877.
Google Scholar: Author Only Title Only Author and Title
Stief, A., Altmann, S., Hoffmann, K., Pant, B.D., Scheible, W.R., and Baurle, I. (2014). Arabidopsis miR156 regulates tolerance to
recurring environmental stress through SPL transcription factors. Plant Cell 26, 1792-1807.
Google Scholar: Author Only Title Only Author and Title
Tan, L., Eberhard, S., Pattathil, S., Warder, C., Glushka, J., Yuan, C.H., Hao, Z.Y., Zhu, X., Avci, U., Miller, J.S., Baldwin, D., Pham, C.,
Orlando, R., Darvill, A., Hahn, M.G., Kieliszewski, M.J., and Mohnen, D. (2013). An Arabidopsis cell wall proteoglycan consists of pectin
and arabinoxylan covalently linked to an arabinogalactan protein. Plant Cell 25, 270-287.
Google Scholar: Author Only Title Only Author and Title
Tang, J., and Chu, C. (2017). MicroRNAs in crop improvement: fine-tuners for complex traits. Nat Plants 3, 17077.
Google Scholar: Author Only Title Only Author and Title
Taylor, K.A.C.C. (1993). Protocols in biotechnology - a colorimetric method for the quantitation of galacturonic acid. Appl Biochem
Biotech 43, 51-54.
Google Scholar: Author Only Title Only Author and Title
Tumaneng, K., Russell, R.C., and Guan, K.L. (2012). Organ size control by Hippo and TOR pathways. Curr Biol 22, R368-R379.
Google Scholar: Author Only Title Only Author and Title
Velasquez, S.M., Ricardi, M.M., Dorosz, J.G., Fernandez, P.V., Nadra, A.D., Pol-Fachin, L., Egelund, J., Gille, S., Harholt, J., Ciancia, M.,
Verli, H., Pauly, M., Bacic, A., Olsen, C.E., Ulvskov, P., Petersen, B.L., Somerville, C., Iusem, N.D., and Estevez, J.M. (2011). O-
glycosylated cell wall proteins are essential in root hair growth. Science 332, 1401-1403.
Google Scholar: Author Only Title Only Author and Title
Google Scholar: Author Only Title Only Author and Title
Verhertbruggen, Y., Marcus, S.E., Haeger, A., Ordaz-Ortiz, J.J., and Knox, J.P. (2009). An extended set of monoclonal antibodies to
pectic homogalacturonan. Carbohyd Res 344, 1858-1862.
Google Scholar: Author Only Title Only Author and Title
Vogel, J.P., Raab, T.K., Schiff, C., and Somerville, S.C. (2002). PMR6, a pectate lyase-like gene required for powdery mildew
susceptibility in Arabidopsis. Plant Cell 14, 2095-2106.
Google Scholar: Author Only Title Only Author and Title
Voinnet, O. (2009). Origin, Biogenesis, and Activity of Plant MicroRNAs. Cell 136, 669-687.
Google Scholar: Author Only Title Only Author and Title
Wang, T., Park, Y.B., Cosgrove, D.J., and Hong, M. (2015). Cellulose-pectin spatial contacts are inherent to Never-Dried Arabidopsis
primary cell walls: evidence from solid-state nuclear magnetic resonance. Plant Physiol 168, 871-884.
Google Scholar: Author Only Title Only Author and Title
Wolf, S., Mravec, J., Greiner, S., Mouille, G., and Hofte, H. (2012). Plant cell wall homeostasis is mediated by Brassinosteroid feedback
signaling. Curr Biol 22, 1732-1737.
Google Scholar: Author Only Title Only Author and Title
Wu, H.J., Ma, Y.K., Chen, T., Wang, M., and Wang, X.J. (2012). PsRobot: a web-based plant small RNA meta-analysis toolbox. Nucleic
Acids Res 40, W22-W28.
Google Scholar: Author Only Title Only Author and Title
Xi, X.N., Kim, S.H., and Tittmann, B. (2015). Atomic force microscopy based nanoindentation study of onion abaxial epidermis walls in
aqueous environment. J Appl Phys 117, 024703.
Google Scholar: Author Only Title Only Author and Title
Xiao, C., Somerville, C., and Anderson, C.T. (2014). POLYGALACTURONASE INVOLVED IN EXPANSION1 functions in cell elongation
and flower development in Arabidopsis. Plant Cell 26, 1018-1035.
Google Scholar: Author Only Title Only Author and Title
Yadav, V., Kundu, S., Chattopadhyay, D., Negi, P., Wei, N., Deng, X.W., and Chattopadhyay, S. (2002). Light regulated modulation of Z-
box containing promoters by photoreceptors and downstream regulatory components, COP1 and HY5, in Arabidopsis. Plant J 31, 741-
753.
Google Scholar: Author Only Title Only Author and Title
Yakubov, G.E., Bonilla, M.R., Chen, H., Doblin, M.S., Bacic, A., Gidley, M.J., and Stokes, J.R. (2016). Mapping nano-scale mechanical
heterogeneity of primary plant cell walls. J Exp Bot 67, 2799-2816.
Google Scholar: Author Only Title Only Author and Title
Yang, T., Wang, Y., Teotia, S., Zhang, Z., and Tang, G. (2018). The making of leaves: how small rna networks modulate leaf
development. Front Plant Sci 9, 824.
Google Scholar: Author Only Title Only Author and Title
Yang, X., Zhang, H., and Li, L. (2012). Alternative mRNA processing increases the complexity of microRNA-based gene regulation in
Arabidopsis. Plant J 70, 421-431.
Google Scholar: Author Only Title Only Author and Title
Yuan, N., Balasubramanian, V.K., Chopra, R., and Mendu, V. (2019). The photoperiodic flowering time regulator FKF1 negatively
regulates cellulose biosynthesis. Plant Physiol 180, 2240-2253.
Google Scholar: Author Only Title Only Author and Title
Zeng, Y.N., Yarbrough, J.M., Mittal, A., Tucker, M.P., Vinzant, T.B., Decker, S.R., and Himmel, M.E. (2016). In situ label-free imaging of
hemicellulose in plant cell walls using stimulated Raman scattering microscopy. Biotechnol Biofuels 9, 256.
Google Scholar: Author Only Title Only Author and Title
Zhang, H., He, H., Wang, X., Wang, X., Yang, X., Li, L., and Deng, X.W. (2011). Genome-wide mapping of the HY5-mediated gene
networks in Arabidopsis that involve both transcriptional and post-transcriptional regulation. Plant J 65, 346-358.
Google Scholar: Author Only Title Only Author and Title
Zhang, H., Zhao, X., Li, J., Cai, H., Deng, X.W., and Li, L. (2014). MicroRNA408 is critical for the HY5-SPL7 gene network that mediates
the coordinated response to light and copper. Plant Cell 26, 4933-4953.
Google Scholar: Author Only Title Only Author and Title
| 2020 | MicroRNA775 Promotes Intrinsic Leaf Size and Reduces Cell Wall Pectin Level via a Target Galactosyltransferase in | 10.1101/2020.09.17.301705 | [
"Zhang He",
"Guo Zhonglong",
"Zhuang Yan",
"Suo Yuanzhen",
"Du Jianmei",
"Gao Zhaoxu",
"Pan Jiawei",
"Li Li",
"Wang Tianxin",
"Xiao Liang",
"Qin Genji",
"Jiao Yuling",
"Cai Huaqing",
"Li Lei"
] | null |
1
Leveraging omic features with F3UTER
1
enables identification of unannotated
2
3’UTRs for synaptic genes
3
Siddharth Sethi1,2, David Zhang2,3,4, Sebastian Guelfi2,5, Zhongbo Chen2,3,4, Sonia
4
Garcia-Ruiz2,3,4, Mina Ryten2,3,4*ᶲ, Harpreet Saini1*, Juan A. Botia2,6*
5
6
1. Astex Pharmaceuticals, 436 Cambridge Science Park, Cambridge, United Kingdom.
7
2. Department of Neurodegenerative Disease, Institute of Neurology, University College
8
London, London, UK.
9
3. NIHR Great Ormond Street Hospital Biomedical Research Centre, University College
10
London, London, UK.
11
4. Genetics and Genomic Medicine, Great Ormond Street Institute of Child Health,
12
University College London, London WC1E 6BT, UK.
13
5. Verge Genomics, South San Francisco, CA 94080, USA
14
6. Department of Information and Communications Engineering, University of Murcia,
15
Spain.
16
17
18
*These authors contributed equally to this manuscript.
19
Corresponding author: Professor Mina Ryten (mina.ryten@ucl.ac.uk)
20
21
22
23
Words: 3,826
24
Display items: 5
25
2
Abstract
26
27
There is growing evidence for the importance of 3’ untranslated region (3’UTR) dependent
28
regulatory processes. However, our current human 3’UTR catalogue is incomplete. Here, we
29
developed a machine learning-based framework, leveraging both genomic and tissue-specific
30
transcriptomic features to predict previously unannotated 3’UTRs. We identify unannotated
31
3’UTRs associated with 1,513 genes across 39 human tissues, with the greatest abundance
32
found in brain. These unannotated 3’UTRs were significantly enriched for RNA binding protein
33
(RBP) motifs and exhibited high human lineage-specificity. We found that brain-specific
34
unannotated 3’UTRs were enriched for the binding motifs of important neuronal RBPs such as
35
TARDBP and RBFOX1, and their associated genes were involved in synaptic function and brain-
36
related
disorders.
Our
data
is
shared
through
an
online
resource
F3UTER
37
(https://astx.shinyapps.io/F3UTER/). Overall, our data improves 3’UTR annotation and provides
38
novel insights into the mRNA-RBP interactome in the human brain, with implications for our
39
understanding of neurological and neurodevelopmental diseases.
40
3
Introduction
41
The 3’UTRs of protein-coding messenger RNAs (mRNAs) play a crucial role in regulating gene
42
expression at the post-transcriptional level. They do so by providing binding sites for trans factors
43
such as RBPs and microRNAs, which affect mRNA fate by modulating subcellular localisation,
44
stability and translation [1, 2]. There is evidence to suggest that these RNA-based regulatory
45
processes may be particularly important in large, polarised cells such as neurons. Recent studies
46
have shown that transcripts which are highly expressed in neurons have both significantly longer
47
3’UTRs and higher 3’UTR diversity [3, 4]. Furthermore, it has been shown that thousands of
48
mRNA transcripts localise within subcellular compartments of neurons and undergo regulated
49
local translation, allowing neurons to rapidly react to local extracellular stimuli [4-7]. Thus, there
50
has been growing interest in the impact of 3’UTR usage on neuronal function in health and
51
disease.
52
53
However, despite on-going efforts to identify and characterise 3’UTRs in the human genome [8-
54
11], there is evidence to suggest that our current catalogue is incomplete [3, 12-14]. Large-scale
55
3’end RNA-sequencing (RNA-seq) has identified a large number of novel polyadenylation
56
(poly(A)) sites, many of which are located outside of annotated exons [12, 13]. These insights are
57
complemented by an increasing recognition of the functional importance of transcriptional activity
58
outside of known exons, particularly in human brain tissues [15-17]. This raises the possibility of
59
developing new approaches for 3’UTR identification seeded from RNA-seq data analyses, an
60
area that has not been fully explored, in large part due to the limited availability of data and
61
appropriate tools.
62
63
In this study, we present a machine learning-based framework, named F3UTER, which leverages
64
both genomic and tissue-specific transcriptomic features. We apply F3UTER to RNA-seq data
65
from Genotype-Tissue Expression Consortium (GTEx) to predict hundreds of unannotated
66
3’UTRs across a wide range of human tissues, with the highest prevalence discovered in brain.
67
We provide evidence to suggest that these unannotated 3’UTR sequences are functionally
68
significant and have higher human lineage specificity than expected by chance. More specifically,
69
we found brain-specific unannotated 3’UTRs were enriched for genes involved in synaptic
70
function and interact with neuronal RBPs implicated in neurodegenerative and neuropsychiatric
71
disorders.
We
release
our
data
in
an
online
platform,
F3UTER
72
(https://astx.shinyapps.io/F3UTER/), which can be queried to visualise unannotated 3’UTR
73
predictions and the omic features used to predict them.
74
4
Results
75
Annotation-independent expression analysis suggests the existence of
76
unannotated 3’UTRs in the human brain
77
There is growing evidence to suggest that the annotation of the human brain transcriptome is
78
incomplete and disproportionately so when compared to other human tissues [15-17]. We
79
hypothesised that this difference may in part be attributed to an increased number of unannotated
80
3’UTRs in human brain. To investigate this possibility, we analysed unannotated expressed
81
regions of the genome (termed ERs) as previously reported by Zhang and colleagues [15]. These
82
ERs were identified through annotation-independent expression analysis of RNA-seq data
83
generated by GTEx with ER calling performed separately for 39 human tissues, including 11 non-
84
redundant human brain regions. We focused on the subset of ERs most likely to be 3’UTRs,
85
namely intergenic ERs which lie within 10 kb of a protein-coding gene (Methods). We found that
86
these intergenic ERs were significantly higher in number (𝑝 = 1.66 × 10−6, Wilcoxon Rank Sum
87
Test) and total genomic space (𝑝 = 2.39 × 10−9, Wilcoxon Rank Sum Test) in brain compared to
88
non-brain tissues (Figure 1a). Furthermore, we discovered that intergenic ERs were significantly
89
more likely to be located at 3’- rather than 5’-ends of their related protein-coding genes (𝑝 =
90
2.08 × 10−14, Wilcoxon Rank Sum Test) (Figure 1b), suggesting that a proportion of ERs
91
detected in human brain could represent unannotated 3’UTRs.
92
93
Differentiating 3’UTRs from other expressed genomic elements is
94
challenging
95
96
Given that existing studies indicate high levels of transcriptional noise and non-coding RNA
97
expression in intergenic regions [18-21], only some intergenic ERs are likely to be generated by
98
unannotated 3’UTRs. This prompted us to develop a method to distinguish 3’UTRs from other
99
transcribed genomic elements (non-3’UTRs) using short-read RNA-seq data. To achieve this aim,
100
we first constructed a training set of known 3’UTRs (positive examples) and non-3’UTRs (negative
101
examples) from Ensembl human genome annotation (v94). We obtained 17,719 3’UTRs and a
102
total of 162,249 non-3’UTRs, consisting of five genomic classes: 21,798 5’UTRs, 130,768 internal
103
coding exons (ICE), 3,718 long non-coding RNAs (lncRNAs), 3,819 non-coding RNAs (ncRNAs)
104
and 2,146 pseudogenes (Methods). For each of the positive and negative examples, we
105
5
constructed a set of 41 informative omic features, which were broadly categorised as either
106
genomic or transcriptomic in nature. Features calculated from genomic data included poly(A)
107
signal (PAS) occurrence, DNA sequence conservation, mono-/di-nucleotide frequency,
108
transposon occurrence and DNA structural properties. Features calculated from transcriptomic
109
data included entropy efficiency of the mapped reads (EE) and percentage difference between
110
the reads mapped at the boundaries (PD) (Methods). To gain a better understanding of these
111
features, we performed a univariate analysis to individually inspect the relationship between each
112
feature and the genomic classes in our training dataset (i.e. 3’UTRs and all types of non-3’UTRs).
113
Overall, while the genomic and transcriptomic features used had overlapping distributions
114
amongst some genomic classes, each feature was significantly different when compared across
115
all the genomic classes (𝑝 < 2.2 × 10−16, Kruskal-Wallis Test and proportion Z-Test,
116
Supplementary Figure S1). This suggested that the features selected could be used to
117
distinguish 3’UTRs from other genomic elements.
118
To further investigate this for all 41 features across all six genomic classes, we applied a uniform
119
manifold approximation and projection (UMAP) [22] for dimensionality reduction into a 2D
120
projection space. We found that while most 3’UTRs clustered separately from other classes within
121
that space, some of them highly overlapped with other genomic classes such, as lncRNAs, ICEs
122
and 5’UTRs (Figure 2a, Supplementary Figure S2). These findings suggested that many
123
unannotated 3’UTRs would be difficult to identify, and thus, may require an advanced
124
classification approach based on machine learning to accurately distinguish them from other
125
genomic elements.
126
127
F3UTER accurately distinguishes 3’UTRs from other genomic elements
128
Next, we measured the predictive value of the omic features we had identified to distinguish
129
between unannotated 3’UTRs and other expressed elements if used collectively. We trained an
130
elastic net multinomial logistic regression model and evaluated its performance using 5-fold cross
131
validation repeated 20 times (Methods). Taking all classes into account, the multinomial logistic
132
regression model achieved an accuracy of 74% and a kappa of 0.52 in distinguishing between
133
the different genomic classes. Consistent with the UMAP visualisation, we found that known
134
3’UTRs were most likely to be misclassified as lncRNAs (4.98%), followed by ICEs (2.46%) and
135
pseudogenes (0.88%) (Figure 2b). On the other hand, false-positive 3’UTR predictions, which
136
6
totalled 44%, were predominantly composed of known ICEs (17.23%) and 5’UTRs (16.06%)
137
(Figure 2b).
138
Since the high false-positive rate of our multinomial logistic regression model would be a
139
significant barrier to reliably predict unannotated 3’UTRs from intergenic ERs, we generated an
140
alternative machine-learning-based approach to address this problem. The resulting random
141
forest multinomial classifier was assessed for its performance using 5-fold cross validation
142
repeated 20 times (Methods). We found that the random forest multinomial classifier had a
143
significantly higher accuracy (76%; 𝑝 < 2.2 × 10−16, Wilcoxon Rank Sum Test) and kappa (0.56;
144
𝑝 < 2.2 × 10−16, Wilcoxon Rank Sum Test) in comparison to the multinomial logistic regression
145
model (Supplementary Figure S3). While the false-negative rate was higher (random forest
146
classifier rate of 22%; logistic regression rate of 9%, Figure 2c), importantly the random forest-
147
based classifier reduced false-positive calling of 3’UTRs to 10% (4.4% 5’UTR, 2.7% lncRNA,
148
1.5% ICE and 1.2% pseudogenes) compared to 44% using logistic regression. We also simplified
149
the classification problem to a binary one and generated a second random forest classifier, aiming
150
only to distinguish between 3’UTRs and non-3’UTRs. This resulted in the development of our final
151
random forest classifier, Finding 3’ Un-translated Expressed Regions (F3UTER, Figure 2d).
152
153
To assess F3UTER’s performance, we performed 5-fold cross validation (repeated 20 times) and
154
calculated metrics such as accuracy, sensitivity, specificity, kappa, area under the ROC curve
155
(AUC-ROC) and area under the precision-recall curve (AUC-PR). F3UTER achieved a mean
156
accuracy of 0.96, sensitivity of 0.92, specificity of 0.96, kappa of 0.78, AUC-ROC of 0.98 (Figure
157
2e) and AUC-PR of 0.91 (Figure 2f) on the validation datasets (hold out). We found that F3UTER
158
performed similarly on both the training and validation datasets in the cross validation (Figure
159
2g). In addition, increasing the sample size of training data reduced the variability in model
160
predictions and hence, made it more stable. Taken together, these findings suggested that we
161
were not overfitting the classifier. Finally, we investigated the contributions of individual features
162
towards the accuracy and node homogeneity (Gini coefficient, Methods) of 3’UTR classification.
163
Interestingly, we found that features derived directly from sequence data (e.g. conservation and
164
PAS) as well as from the transcriptomic data, namely mean-PD and mean-EE (Supplementary
165
Figure S4), most significantly contributed to the accuracy of F3UTER. This shows that F3UTER
166
leverages both genomic and transcriptomic features to classify 3’UTRs, which would be expected
167
to enable the identification of tissue-specific unannotated 3’UTRs.
168
7
Evaluation of F3UTER using 3’-end sequencing data validates unannotated
169
3’UTR predictions
170
171
We evaluated the performance of F3UTER using an independent dataset consisting of both RNA-
172
seq data and paired 3’-seq in B cells [23]. The latter, a form of 3’-end sequencing, was performed
173
to identify poly(A) sites experimentally. Since poly(A) sites are present at the very end of 3’UTRs,
174
unannotated 3’UTRs should overlap or be in the close vicinity of a poly(A) site. It should be noted
175
that unlike the GTEx RNA-seq dataset which we used for our previous analyses and which
176
consists of hundreds of samples for most tissues, this B cell dataset consisted of only two RNA-
177
seq samples. Since detecting unannotated ERs relies on averaging RNA-seq coverage across
178
many samples to reduce the contribution of transcriptional noise to ER definition, calling ERs from
179
only two samples would likely result in inaccuracies at ER boundaries. Although this would be
180
expected to significantly reduce the confidence in the detection of unannotated ERs and
181
potentially underestimate the performance of F3UTER, the paired RNA-seq and 3’-seq nature of
182
this B cell dataset enabled us to confidently validate 3’UTR predictions using gold standard
183
experimental data.
184
185
First, we identified 3’ unannotated intergenic ERs in B cells from the RNA-seq data following the
186
pipeline used by Zhang et al. [15]. Then we used F3UTER to predict unannotated 3’UTRs in this
187
B cell ER dataset, and compared these predictions to intergenic poly(A) clusters detected using
188
3’-seq (Figure 3a). We focused on confident 3’UTR predictions, defined as those with a prediction
189
probability of > 0.6. ERs predicted to be 3’UTRs which also overlapped with a poly(A) cluster were
190
considered to be validated, as exemplified by the intergenic ER predicted to be a novel 3’UTR of
191
the gene CYTIP (Figure 3b). As a reference, we noted that 87.9% of known 3’UTRs overlapped
192
with a poly(A) cluster in B cell. We found that on average, 38.5% of 3’UTR predictions were
193
validated. This was 17.5-fold higher than that for randomly selected intergenic regions (2.2%, 𝑝 <
194
0.0001, permutation test; Supplementary Figure S5) and 2.2-fold higher than the validation rate
195
of non-3’UTR predictions (17.4%, Figure 3c). Overall, these observations demonstrate the
196
accuracy of F3UTER and show that it can effectively distinguish unannotated 3’UTRs from other
197
functional genomic elements in the genome.
198
199
8
Applying F3UTER across 39 GTEx tissues identifies hundreds of
200
unannotated 3’UTRs with evidence of functional significance
201
202
We applied F3UTER to 3’ unannotated intergenic ERs identified by Zhang and colleagues [15] in
203
39 tissues using RNA-seq data provided by GTEx. Similar to the B cell ER dataset, we focused
204
on confident 3’UTR predictions with a prediction probability of > 0.6 (Supplementary File 1).
205
Across all tissues, we found that on average 7.9% of analysed ERs were predicted as
206
unannotated 3’UTRs, with 8.2% being called in brain (Supplementary Figure S6). This equated
207
to an average of 187 potentially unannotated 3’UTRs per tissue (ranging from 96 in adipose-
208
subcutaneous to 348 in hypothalamus, Figure 4a), covering 58 to 265 kb of genomic space (mean
209
across tissues = 138 kb, Figure 4b). By assigning predicted 3’UTRs to protein-coding genes
210
either through the existence of junction reads or by proximity, we estimated that 1,513 distinct
211
genes in total had unannotated 3’UTRs with an average of 167 genes per tissue (Figure 4c). As
212
expected, the number of predicted unannotated 3’UTRs was significantly higher in the brain
213
relative to non-brain tissues (median values of 295 and 142 in brain and non-brain tissues
214
respectively; 𝑝 = 1.65 × 10−6, Wilcoxon Rank Sum Test). This was associated with a significantly
215
higher total genomic space (median values of 232 kb and 104 kb in brain and non-brain tissues
216
respectively; 𝑝 = 1.43 × 10−8, Wilcoxon Rank Sum Test) and higher number of implicated genes
217
(median values of 270 and 127 in brain and non-brain tissues respectively; 𝑝 = 1.65 × 10−6,
218
Wilcoxon Rank Sum Test). This data suggests that incomplete annotation of 3’UTRs is present in
219
all human tissues but is most prevalent in the brain.
220
221
Next, we investigated the functional significance of unannotated 3’UTRs by analysing their
222
potential interaction with RBPs. This in silico analysis was performed because selective RBP
223
binding at 3’UTRs is thought to be key in explaining the selection of alternate PASs and its impact
224
on mRNA stability and localisation [24]. Using the catalogue of known RNA binding motifs from
225
the ATtRACT database [25], we examined the binding density of 84 RBPs across all unannotated
226
3’UTRs (Methods). Consistent with previous reports demonstrating higher RBP binding densities
227
in known 3’UTRs relative to other genomic regions [26], we found that 3’UTR predictions were
228
enriched for RBP binding motifs compared to non-3’UTR predictions (𝑝 < 2.2 × 10−16, effect size
229
(es) = 0.17, Wilcoxon Rank Sum Test, Figure 4d). Surprisingly, we noted that unannotated
230
3’UTRs were also enriched for RBP binding motifs compared to known 3’UTRs (𝑝 < 2.2 × 10−16,
231
es = 0.28, Wilcoxon Rank Sum Test, Figure 4d) suggesting that these regions may be of
232
particular functional significance. To investigate this further, we leveraged constrained, non-
233
9
conserved (CNC) scores [27], a measure of human-lineage-specificity, to determine whether the
234
unannotated 3’UTRs identified were of specific importance in humans. CNC score, a metric
235
combining cross-species conservation and genetic constraint in humans, was used to identify and
236
score genomic regions which are amongst the 12.5% most constrained within humans but yet are
237
not conserved. We found that unannotated 3’UTRs exhibited higher CNC scores compared to
238
known 3’UTRs (𝑝 = 0.012, es = 0.016, Wilcoxon Rank Sum Test, Figure 4e). Thus, together our
239
analyses suggested that unannotated 3’UTRs are not only functionally important but may be
240
particularly crucial in human-specific biological processes.
241
242
F3UTER identifies unannotated 3’UTRs of genes associated with synaptic
243
function
244
245
Given the evidence for the functional importance of unannotated 3’UTRs predicted by F3UTER,
246
we wanted to explore their biological relevance. To do this, we began by categorising all
247
unannotated 3’UTRs into four sets based on their tissue-specificity: absolute tissue-specific,
248
highly brain-specific, shared and ambiguous (Methods and Supplementary Figure S7). Using
249
this non-redundant set of 3’UTRs, we found that on average, we extended the current annotation
250
per gene by 681 bp in highly brain-specific (1.4x the known maximal 3’UTR length), 633.6 bp in
251
tissue-specific (0.95x the known maximal 3’UTR length), and 496.63 bp in shared predictions
252
(0.88x the known maximal 3’UTR length) respectively. Next, we repeated the RBP and CNC
253
analysis for each category finding that all unannotated 3’UTR sets showed significant enrichment
254
of RBP binding motifs when compared not only to non-3’UTR predictions (𝑝 ≤ 2.5 × 10−5,
255
Wilcoxon Rank Sum Test), but also to known 3’UTRs (𝑝 ≤ 3.9 × 10−7, Wilcoxon Rank Sum Test),
256
with the brain-specific set having the largest effect size (es ≥ 0.17) (Figure 5a). Focussing on
257
CNC scores, we found that while none of the unannotated 3’UTR sets showed significant
258
differences in score compared to known 3’UTRs (𝑝 ≥ 0.121, Wilcoxon Rank Sum Test), brain-
259
specific unannotated 3’UTRs trended to significance with the largest effect size relative to other
260
sets (Figure 5a). Together, these observations lead us to conclude that highly brain-specific
261
3’UTR predictions were likely to be of most biological interest.
262
263
These observations raised the question of what types of genes are associated with highly brain-
264
specific 3’UTR predictions. Interestingly, we found that while genes linked to brain-specific non-
265
3’UTR predictions had no GO term enrichments, those linked to an unannotated brain-specific
266
10
3’UTR were significantly enriched for synaptic function (“synaptic signalling”, “synapse
267
organisation” and “protein localization to postsynaptic specialization membrane”; 𝑞 =
268
4.97 × 10−3) (Figure 5b, Supplementary File 2). Using SynGO (the synaptic GO database [28])
269
to obtain more granular information, we found that genes associated with unannotated 3’UTRs
270
were more significantly enriched for terms relating to post-synaptic (“protein localisation in
271
postsynaptic density”, 𝑞 = 2.87 × 10−4; postsynaptic function, 𝑞 = 4.1 × 10−3), as compared to
272
presynaptic structures (“localisation in presynapse”, 𝑞 = 0.03; presynaptic function, 𝑞 = 0.1)
273
(Figure 5c, Supplementary File 2). Furthermore, we found that genes linked to unannotated
274
brain-specific 3’UTRs were significantly enriched for those already associated with rare
275
neurogenetic disorders (𝑝 = 0.01, hypergeometric test) and more specifically adult-onset
276
neurodegenerative disorders (𝑝 = 0.03, hypergeometric test). For example, we detected an
277
unannotated 3’UTR in the brain linked to the gene, APP, a membrane protein which when mutated
278
gives rise to autosomal dominant Alzheimer’s disease and encodes for amyloid precursor protein,
279
the main constituent of amyloid plaques [29]. We detected a 920 bp long brain-specific
280
unannotated 3’UTR located 1.8 kb downstream of APP (Figure 5d) and only 51 bp from an
281
intergenic poly(A) site on the same strand as APP gene as reported by the poly(A) atlas. Other
282
similar examples included the genes, C19orf12, RTN2, SCN2A and OPA1 (Supplementary
283
Figures S8 & S9).
284
285
Brain-specific unannotated 3’UTRs interact with RBPs implicated in
286
neurological disorders
287
288
Next, we investigated the information content of brain-specific unannotated 3’UTRs by comparing
289
RBP binding enrichments between brain-specific and shared 3’UTR predictions (Methods). By
290
using shared 3’UTR predictions as the negative control, we removed RBPs associated with non-
291
brain tissues and so identified RBP binding of greatest relevance to human brain function. This
292
analysis identified 22 RBPs with significantly enriched binding in the brain-specific unannotated
293
3’UTRs (𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑 𝑝 < 10−5) (Supplementary Table 1). We found that nine of these RBPs were
294
previously known to be associated with “mRNA 3’UTR binding” (𝑞 = 2.23 × 10−14,
295
Supplementary File 3), including TARDBP, an RNA binding protein implicated in both
296
frontotemporal dementia and amyotrophic lateral sclerosis [30]. Of the 75 gene targets that we
297
identified for TARDBP through unannotated 3’UTRs, up to 50 were known to be TARDBP targets
298
11
based on computational scanning of existing 3’UTR annotations for TARDBP motif (47%, 𝑝 =
299
0.008, hypergeometric test) and iCLIP experiments (44%, 𝑝 = 1.47 × 10−6, hypergeometric test).
300
However, this implied that 25 gene targets were not previously known to harbour TARDBP binding
301
motifs based on current annotation. Another RBP which was identified to be significantly enriched
302
in brain-specific unannotated 3’UTRs was RBFOX1 (𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑 𝑝 = 1.78 × 10−18), a neuronal
303
splicing factor implicated in the regulation of synaptic transmission [31] and whose mRNA targets
304
have been implicated in autism spectrum disorders [32]. We identified 89 gene targets with a
305
predicted RBFOX1 binding motif within their associated unannotated 3’UTRs. Of these 89 genes,
306
only 31 (35%) had a predicted RBFOX1 binding motif within their existing 3’UTRs, again implying
307
that unannotated 3’UTRs provide valuable novel binding sites. Furthermore, GO and SynGO
308
enrichment analyses (Supplementary File 3) demonstrated that the target genes of RBFOX1
309
were significantly enriched for synaptic (“synaptic membrane adhesion”, 𝑞 = 1.58 × 10−2) and
310
postsynaptic terms (“postsynapse”, 𝑞 = 0.01), consistent with the previously known functions of
311
RBFOX1 [31]. These results show that the identification of brain-specific unannotated 3’UTRs can
312
recognise new genes within known regulatory networks, which can provide novel, disease-
313
relevant insights.
314
315
Discussion
316
317
In this study we generate a machine learning-based classifier, F3UTER, which leverages
318
transcriptomic as well as genomic data to predict unannotated 3’UTRs. F3UTER outperforms a
319
state-of-the-art statistical learning approach, elastic net logistic regression, whilst retaining its
320
interpretability capabilities. We apply F3UTER to transcriptomic data covering 39 human tissues
321
studied within GTEx, enabling the identification of tissue-specific unannotated 3’UTRs. Using this
322
large, public, short-read RNA-seq data set, we predict unannotated 3’UTRs for 1,513 genes,
323
(equating to 5.4 Mb of genomic space in total across 39 tissues) and demonstrate that F3UTER
324
can be successfully applied to human genomic regions from any tissue with existing bulk RNA-
325
seq data. In fact, even though intergenic ERs in B cells were generated using only two samples,
326
we were able to validate 38.5% of the unannotated 3’UTR predictions using 3’-end sequencing
327
data, showing that F3UTER can be a useful tool even for small RNA-seq datasets. Furthermore,
328
it should be noted that F3UTER does not depend on ER datasets as input, but instead any set of
329
interesting human genomic regions can be used. Given the continued popularity and high
330
12
availability of short-read RNA-seq data across tissues, cell types and disease states, we believe
331
that (1) F3UTER could be applied more broadly to improve our understanding of 3’UTR diversity
332
and usage, and (2) the set of omic features devised within this study could form the basis for other
333
predictive models aimed at increasing the accuracy of human transcriptomic annotation.
334
335
We focus on F3UTER-predicted 3’UTRs in human brain, which we find to be most prevalent when
336
comparing predictions across all 39 human tissues. We believe that the higher frequency of
337
incomplete 3’UTR annotation in human brain could be attributed to several factors including: (1)
338
higher transcript diversity with many rare isoforms expressed in this tissue; (2) high cellular
339
heterogeneity complicating detection of tissue- /cell-type specific transcripts; (3) historically lower
340
availability of human brain samples; and (4) reliance on post-mortem tissues, which suffer from
341
RNA degradation resulting in decreased accuracy of transcript identification.
342
343
While we find that collectively the unannotated 3’UTRs predicted by F3UTER were significantly
344
enriched for RBP binding and exhibited high human lineage-specificity, the latter was primarily
345
driven by brain-specific 3’UTR predictions. Overall, these findings suggest that predicted 3’UTRs
346
are likely to be functionally important in the human genome. Moreover, these findings provide
347
some explanation for the difficulties of identifying 3’UTRs through cross-species analyses
348
particularly when considering brain-specific transcripts. Interestingly, we find that brain-specific
349
unannotated 3’UTRs were enriched for binding of RBPs already implicated in neurological
350
disorders, such as TARDBP and RBFOX1. Furthermore, genes linked to unannotated brain-
351
specific 3’UTRs were significantly enriched for those already associated with rare neurogenetic
352
and adult-onset neurodegenerative disorders, and for genes involved in synaptic function.
353
354
Taken together, our results demonstrate that F3UTER not only improved 3’UTR annotation, but
355
also identified unannotated 3’UTRs in the human brain which provided novel insights into the
356
mRNA-RBP interactome with implications for our understanding of neurological and
357
neurodevelopmental diseases. With this in mind, we note the growing interest in the role of 3’UTR-
358
based mechanisms in translational regulation within complex, large, polarised cell types such as
359
neurons [4, 5, 33, 34]. Although increasing use of single-nuclei RNA-seq, together with long-read
360
RNA-seq will provide further insights into alternative 3’UTR usage and will impact the field
361
considerably, these technologies still have significant limitations for the identification of rare
362
transcripts. Therefore, we believe that F3UTER, which can effectively utilise existing short-read
363
RNA-seq data sets, will be of interest to a wide range of researchers. Furthermore, we release
364
13
our results through an online resource (F3UTER: https://astx.shinyapps.io/F3UTER/) which
365
allows users to both easily query unannotated 3’UTRs and inspect the omic features driving the
366
classifier’s prediction for an ER of interest.
367
14
Figures
368
369
370
371
372
373
Figure 1.
374
Enrichment of intergenic ERs across 39 GTEx tissues. (a) Scatter plot showing the number
375
of intergenic ERs and their total genomic space covered in 39 human tissues. (b) Enrichment of
376
intergenic ERs grouped by location with respect to their associated protein-coding gene. Each
377
data point in the box plot represents the proportion of total intergenic ERs in a tissue. p: p-value
378
calculated using Wilcoxon Rank Sum Test.
379
15
380
Figure 2.
381
Classification of 3’UTRs from other transcribed elements in the genome. (a) UMAP
382
representation of features, with elements labelled by genomic classes. (b) Classification of
383
3’UTRs using an elastic net multinomial logistic regression. (c) Classification of 3’UTRs using a
384
multinomial random forest classifier. (d) General framework of F3UTER: the core of the
385
framework is a random forest classifier trained on omic features derived from known 3’UTRs and
386
non-3’UTRs. The omic features are based on either genomic (DNA sequence) or transcriptomic
387
16
(RNA-seq from GTEx) properties. To make predictions, genomic coordinates of ERs are given as
388
input, from which a feature matrix is constructed. The output of the framework is ERs categorised
389
into potential 3’UTRs and non-3’UTRs with their associated prediction probability scores. (e, f)
390
ROC and precision recall curves of F3UTER evaluated using 5-fold cross validation. (g) Bias-
391
variance trade-off plot demonstrating the performance of F3UTER on training and validation
392
datasets grouped by the sample size of the training data.
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
17
432
433
434
435
436
Figure 3.
437
Evaluation of F3UTER predictions on an independent ER dataset. (a) Schematic describing
438
the framework of the process implemented to evaluate the performance of F3UTER on ERs in B
439
cells. (b) Genome browser view of the CYTIP locus, showing intergenic ERs detected
440
downstream of CYTIP and poly(A) sites in B cells. (c) Bar plots showing the overlap between
441
predictions made by F3UTER and intergenic poly(A) sites from 3‘-end sequencing in B cells. The
442
bar for random predictions represents the mean overlap (from 10,000 permutations) between
443
randomly selected intergenic ERs and intergenic poly(A) sites.
444
18
445
Figure 4.
446
Unannotated 3’UTR predictions across 39 GTEx tissues. (a) Number of unannotated 3’UTRs
447
predicted by F3UTER. (b) Total genomic space of unannotated 3’UTRs. (c) Number of genes
448
associated with unannotated 3’UTRs. In each bar plot, tissues are sorted in descending order of
449
19
the values plotted on y-axis. The square boxes below the bars are color-coded to group the
450
tissues according to their physiology. The predictions are grouped and color-coded based on their
451
prediction probability scores from F3UTER. (d) Density distributions comparing the RBP binding
452
density across known 3’UTRs, predicted 3’UTRs and predicted non-3’UTRs. p: p-value of
453
comparison calculated using Wilcoxon Rank Sum Test; es: effect size; x: predicted 3’UTRs vs.
454
known 3’UTRs; y: predicted 3’UTRs vs. predicted non-3’UTRs. (e) Density distributions
455
comparing the “constrained non-conserved” (CNC) scores between known and predicted 3’UTRs.
456
p: p-value of comparison calculated using Wilcoxon Rank Sum Test; es: effect size.
457
458
459
460
461
462
20
463
Figure 5.
464
Functional significance of highly brain-specific unannotated 3’UTRs. (a) Density
465
distributions comparing RBP binding and “constrained non-conserved” (CNC) scores between
466
known, predicted 3’UTRs and predicted non-3’UTRs, categorised according to their tissue-
467
specificity. p: p-value of comparison calculated using Wilcoxon Rank Sum Test; es: effect size; x:
468
predicted 3’UTRs vs. known 3’UTRs; y: predicted 3’UTRs vs. predicted non-3’UTRs. (b) GO terms
469
enriched amongst the list of genes associated with highly brain-specific unannotated 3’UTRs. MF:
470
21
molecular function; CC: cellular component; BP: biological process. (c) Sunburst plot showing the
471
cellular component SynGO terms over-represented in genes associated with highly brain-specific
472
3’UTRs. The inner rings of the plot represent parent terms, while outer rings represent their more
473
specific child terms. Rings are colour coded based on the enrichment q-value of the terms. (d)
474
Genome browser view of the APP locus, showing intergenic ERs detected downstream of APP in
475
the hypothalamus, and poly(A) sites from the poly(A) atlas data.
476
22
Methods
477
ER data
478
We collected the set of intergenic ERs identified by Zhang and colleagues [15] in 39 GTEx tissues,
479
comprising of 11 non-redundant brain tissues and 28 non-brain tissues (total intergenic ERs =
480
9,339,770). Each ER was associated to a protein-coding gene by extracting genes which
481
connected to the ER via a junction read. In cases where no junction read was present, the nearest
482
protein-coding gene was assigned to the ER. From this dataset, we selected intergenic ERs
483
located within 10 kb of their associated gene, resulting in 237,540 ERs. In this dataset, 4% of the
484
ERs were associated to a gene via a junction read. Based on the location of intergenic ERs with
485
respect to their associated genes, i.e. whether upstream or downstream, we annotated their
486
orientation as 5’ (92,148 ERs) or 3’ (145,392 ERs) respectively. The total genomic space of these
487
intergenic ERs was calculated by adding the length of all ERs in each tissue. To further remove
488
ERs which were unlikely to be 3’UTRs, we selected 3’ intergenic ERs with a length ≤ 2 kb – which
489
is the third quartile limit of known 3’UTR exon lengths. We also removed small ERs with length ≤
490
40 nucleotides (nt) for which feature calculation can be problematic. This resulted in a set of
491
93,934 ERs across all 39 tissues, and this set was used as input to F3UTER.
492
Assembling positive and negative 3’UTR learning datasets
493
For positive examples, we used known 3’UTRs, while for negative examples, we used regions in
494
the genome which are known to be non-3’UTRs, namely 5’UTRs, internal coding exons (ICEs),
495
lncRNAs, ncRNAs and pseudogenes. Ensembl human genome annotation (v94 GTF) was used
496
to extract the genomic coordinates of these different genomic classes. For all classes in our
497
training dataset, firstly, we selected high confidence annotations at the transcript level with
498
transcript support level (TSL) = 1. Secondly, we collapsed and combined multiple transcripts
499
associated with a single gene to make a consensus “meta-transcript” per gene. This merged all
500
the overlapping regions emerging from the same gene. Finally, we extracted exons with width >=
501
40 (nt) from these meta-transcripts to serve as learning examples.
502
503
To capture regions of 3’UTR exons, 5’UTR exons and ICEs, transcripts from protein-coding genes
504
were selected. For ICE examples, transcripts with at least three coding exons were further
505
selected (as transcripts with less than three exons would not contain an internal exon) and their
506
23
first and last coding exons were removed to capture ICEs. To capture lncRNA, ncRNA and
507
pseudogene exons, we selected annotations from the GTF file with the following gene biotypes:
508
-
lncRNA: "non_coding", "3prime_overlapping_ncRNA", "antisense", "lincRNA",
509
"sense_intronic", "sense_overlapping", "macro_lncRNA"
510
-
ncRNA: "miRNA", "misc_RNA", "rRNA", "snRNA", "snoRNA", "vaultRNA"
511
-
pseudogene: "pseudogene", "processed_pseudogene", "unprocessed_pseudogene",
512
"transcribed_processed_pseudogene", "transcribed_unitary_pseudogene",
513
"transcribed_unprocessed_pseudogene", "translated_processed_pseudogene",
514
"unitary_pseudogene", "unprocessed_pseudogene", "TR_V_pseudogene",
515
"TR_J_pseudogene", "rRNA_pseudogene", "polymorphic_pseudogene",
516
"IG_V_pseudogene", "IG_pseudogene", "IG_J_pseudogene", "IG_C_pseudogene"
517
518
Calculating omic features
519
For each region in the training dataset, we calculated several genomic and transcriptomic based
520
features. Transcriptomic features were used to account for tissue-specific properties of
521
transcribed elements in the genome.
522
523
Genomic (sequence) based features:
524
525
● Poly(A) signals (number of features, n=1): Previous studies have shown that 3’UTR
526
sequences of most mammalian genes contain the consensus AAUAAA motif (or a close
527
variant) 10-30 nt upstream of the poly(A) site [8]. These motif sites are recognised and
528
bound by the cleavage and polyadenylation specificity factor (CPSF), and are referred to
529
as polyadenylation signals (PASs). PASs are an important characteristic of 3’UTRs and
530
are involved in the regulation of the polyadenylation process [8]. We used 12 commonly
531
occurring PASs (AAUAAA, AUUAAA, AGUAAA, UAUAAA, AAUAUA, AAUACA,
532
CAUAAA, GAUAAA, ACUAAA, AAUAGA, AAUGAA, AAGAAA) [9, 12, 35, 36] to construct
533
a consensus position weight matrix (PWM). Each region was scanned for potential PWM
534
matches and a binary outcome was reported i.e. whether the region contains a potential
535
PAS or not. The “searchSeq'' function (with min.score= “95%”) from the R package
536
“TFBSTools” [37] was used to detect PWM matches.
537
538
24
● Mono- and di-nucleotide frequency (n=20): The sequence composition in 3’UTRs,
539
especially near the poly(A) sites has been shown to be important for polyadenylation [8,
540
9, 35]. The frequency probability of each mono-nucleotide (i.e. A, T G, C; n=4) and di-
541
nucleotide pair (n=16; e.g. AA, AT, GC, GG) was calculated as the number of nucleotide
542
occurrences divided by the length of the region.
543
544
● DNA sequence conservation (n=1): Sequences of non-protein coding transcripts and
545
un-translated regions are poorly conserved compared to protein-coding sequences [38,
546
39]. For every genomic position, we extracted the phastCons score of the human genome
547
(hg38) across 7 species pre-computed by the UCSC genome browser, and averaged it
548
across the region to calculate mean sequence conservation score for each region.
549
550
● Transposons (n=1): Previous studies have revealed that transposons are highly enriched
551
within lncRNAs compared to protein-coding genes and other non-coding elements [40,
552
41]. These transposable elements are considered to be the functional domains of
553
lncRNAs. We calculated the total fraction of region covered with transposons – LINEs,
554
SINEs, LTRs, DNA and RC transposons. The hg38 genomic coordinates of the
555
transposable
elements
(Dfam
v2.0)
were
downloaded
from
556
http://www.repeatmasker.org/species/hg.html.
557
558
● DNA structural properties (n=16): The underlying sequence composition of a DNA
559
molecule plays an important role in determining its structure. As a result, similar DNA
560
sequences have a tendency to have similar DNA structures [42]. We calculated 16
561
properties of DNA structures which can be predicted from a nucleotide sequence based
562
on previous experiments. To quantitatively measure a structural property from a nucleotide
563
sequence,
we
used
pre-compiled
conversion
tables
downloaded
from
564
http://bioinformatics.psb.ugent.be/webtools/ep3/?conversion [43]. Depending on the
565
structural property, we extracted scores for each di-nucleotide or tri-nucleotide occurrence
566
in the sequence from the conversion tables, and averaged the scores across the region.
567
568
Transcriptomic based features:
569
570
● Entropy efficiency (n=1): We measured the uniformity of read coverage across a region
571
using entropy efficiency, as described in Gruber et al. [44]. The entropy efficiency (EE) of
572
25
a region (x) was calculated as,
𝐸𝐸(𝑥) = −
∑
𝑝(𝑥𝑖)×log (𝑝(𝑥𝑖
𝑛
𝑖=1
))
𝑙𝑜𝑔(𝑛)
;
𝑝(𝑥𝑖) =
𝑥𝑖
∑
𝑥𝑗
𝑛
𝑗=1
,
573
where 𝑛 represents the length of the region and 𝑝(𝑥𝑖) is the read count at position 𝑖 divided
574
by the total read count of the region. For each region, we calculated EE in 39 GTEx tissues
575
and averaged it across all the tissues to obtain a baseline distribution of EE scores.
576
577
● Percentage difference (n=1): We calculated the percentage difference (PD) between the
578
read counts at the boundaries of a region. For read counts 𝑟1 and 𝑟2 measured at the
579
boundaries of a region 𝑥, PD was calculated as: 𝑃𝐷(𝑥) =
|𝑟1− 𝑟2|
𝑚𝑒𝑎𝑛(𝑟1,𝑟2) × 100. For each
580
region, we calculated PD in 39 GTEx tissues and averaged it across all the tissues to
581
obtain a baseline distribution of PD scores.
582
583
Univariate and multivariate analysis
584
For univariate analysis, we performed non-parametric Kruskal–Wallis test and proportion Z-test
585
for continuous and categorical variables, respectively, to identify features with significant
586
differences across all the genomic classes. We used UMAP [22] to visualise all the features in
587
two-dimensional space. The UMAP analysis was performed using the R package “umap” with
588
default parameters. The clusters were visualised as a 2D density and a scatter plot. Each data
589
point was labelled and coloured according to its genomic class.
590
591
To perform multivariate analysis, a feature matrix was generated where rows represented regions
592
from the training dataset (n=179,968), and columns represented the quantified features (n=41).
593
The features were scaled and centred in R using the preProcess function of R “Caret” package
594
[45]. The elastic net multinomial logistic regression model was trained using the “glmnet” R
595
package [46] with the following parameters: family = "multinomial", alpha=0.5, nlambda=25 and
596
maxit=10,000. The random forest multinomial classifier was trained within Caret using the
597
“randomForest” package [47] with default parameters (ntree = 500, nodesize = 1). We performed
598
a 5-fold cross validation (repeated 20 times) to evaluate the performance of these multinomial
599
classifiers, where the model was trained on 80% of the data (training dataset) and tested on 20%
600
of the remaining data (validation dataset). Downsampling of the data was employed to correct for
601
imbalance in the sample size of the classes. For each cross validation run, we produced a
602
confusion matrix for each prediction class using the Caret’s confusionMatrix function and
603
computed the false- positive and negative rates. Additionally, we calculated Cohen’s kappa, which
604
26
reports the accuracy of a model compared to the expected accuracy and is a much accurate
605
measure of performance for imbalanced datasets. These metrics were averaged across all the
606
cross validation runs for reporting purposes.
607
F3UTER construction and evaluation
608
We designed F3UTER as a binary classifier to categorise an ER into a 3’UTR (positive) or a non-
609
3’UTR (negative). This random forest classifier was implemented in R using Caret as the machine
610
learning framework and “randomForest” as the machine learning algorithm within Caret. The
611
random forest classifier was trained using the default parameters (ntree = 500, nodesize = 1). We
612
performed a 5-fold cross validation (repeated 20 times) to evaluate the performance of the
613
F3UTER. For each cross validation run, we calculated the performance metrics such as accuracy,
614
kappa, sensitivity, specificity, ROC curve and precision-recall curve, using the caret’s
615
confusionMatrix function. Variable importance was measured using mean decrease in accuracy
616
and Gini coefficient, as natively reported by random forest. The Gini coefficient measures the
617
contribution of variables towards homogeneity of nodes in the random forest tree. These metrics
618
were averaged across all the cross validation runs for reporting purposes. For bias-variance trade-
619
off analysis, we trained F3UTER on sequentially increasing sample size of training data (0.1%,
620
0.5%, 1%, 5%, 10%, 30%, 50%, 80% and 100%), hence sequentially increasing the complexity
621
of the model. For each sample size value, a fraction of the training data was randomly selected,
622
and a 5-fold cross validation was performed which captured all the performance metrics for both
623
the training and validation datasets. This process was repeated 20 times for each sample size.
624
To make 3’UTR predictions on ER datasets, the classifier with the highest kappa statistic was
625
selected from the cross validation process.
626
627
Validation of 3’UTR predictions in B cells
628
Previously published RNA-seq and its corresponding 3’-end seq data in B cells [23] (two replicates
629
each) was used for validating 3’UTR predictions (GEO repository: GSE111310; samples:
630
GSM3028281, GSM3028282, GSM3028302 and GSM3028304). We processed each RNA-seq
631
replicate individually and detected 3’ intergenic ERs using the pipeline detailed in Zhang et al.
632
[15]. Analysed poly(A) site clusters associated with these RNA-seq samples were downloaded
633
from poly(A) atlas [13]. These poly(A) site clusters were compared to Ensembl human genome
634
annotation (v92) to identify sites which occur within the intergenic regions. F3UTER was applied
635
27
to 3’ intergenic ERs in B cells and the resulting predictions (with prediction probability > 0.6) were
636
compared to intergenic poly(A) site clusters to calculate their overlap. Predictions with at least a
637
1 bp overlap with a poly(A) site were considered to be overlapping. A permutation test was
638
performed to inspect if the observed overlap between 3’UTR predictions and intergenic poly(A)
639
sites is more than what we would expect by random chance. Using BEDTOOLS [48], the locations
640
of 3’UTR predictions were shuffled in the intergenic genomic space on the same chromosome,
641
hence generating random intergenic ERs with length, size and chromosome distribution similar
642
to 3’UTR predictions in B cells. To shuffle the locations within the intergenic space, we excluded
643
the genomic space covered by genes (all Ensembl bio-types) and intergenic ERs in B cells (both
644
3’ and 5’). The overlap between these randomly generated intergenic ERs and poly(A) sites was
645
then calculated, and this process was repeated 10,000 times to produce a distribution of expected
646
overlap. The p-value was calculated as
𝑥
𝑁, where 𝑥 is the number of expected overlap greater than
647
the observed overlap, and 𝑁 is the total number of permutations. The z-score was calculated as
648
𝑂𝑜𝑏𝑠 − 𝑂𝑝𝑒𝑟𝑚
𝑆𝐷𝑝𝑒𝑟𝑚
, where 𝑂𝑜𝑏𝑠represents the observed overlap, 𝑂𝑝𝑒𝑟𝑚is the median of the permuted
649
overlap, and 𝑆𝐷𝑝𝑒𝑟𝑚is the standard deviation of the permuted distribution.
650
651
3’UTR predictions in 39 GTEx tissues
652
A feature matrix of 3’ intergenic ERs was generated in each tissue. F3UTER was applied to each
653
matrix to categorise intergenic ERs into 3’UTR (𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 > 0.60) and non-3’UTR
654
(𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 ≤ 0.60) predictions. For each tissue, the lengths of the 3’UTR
655
predictions were added to calculate their total genomic space (in kb). To compare brain and non-
656
brain tissues, a two-sided Wilcoxon Rank Sum Test was applied to statistically compare the
657
associated numbers between the two groups. To explore the biological relevance of 3’UTR
658
predictions, they were categorised into four groups based on their tissue-specificity: absolute
659
tissue-specific, highly brain-specific, shared and ambiguous. To do such categorisation, the
660
genomic coordinates of ER predictions were compared across the 39 tissues. An ER which did
661
not overlap any other ER across the tissues was labelled as “absolute tissue-specific” or present
662
in only one tissue. On the other hand, for an ER which overlapped (≥ 1 bp) ERs from other tissues,
663
we calculated the proportion of brain tissues in which the ER was detected. If more than 75% of
664
the tissues were brain related, the ER was labelled as “highly brain-specific”. From the remaining
665
data, ERs detected in at least five tissues, with their start and end coordinates within a 10 bp
666
28
window, were labelled as “shared”. All the remaining ERs which did not fall in any of the above
667
categories were labelled as “ambiguous”.
668
669
RBP and CNCR analysis
670
The position weight matrices (PWMs) of RBP binding motifs in humans were collected from the
671
ATtRACT database [25]. Motifs with less than 7 nt in length and with a confidence score of less
672
than one, were removed to reduce false-positives in the motif matches. To remove redundancy
673
between multiple motifs of a RBP, we further selected the longest available motif. This resulted in
674
84 unique PWMs, which were then used for identifying potential RBP binding using tools from the
675
MEME suite [49]. We used FIMO [50] with a uniform background to scan query regions for
676
potential RBP motif matches. For each RBP motif and query sequence pair, we calculated
677
normalised counts as the number of motif matches (with 𝑝 < 10−4) per 100 nt of query sequence.
678
To summarise this analysis, we then calculated an overall RBP binding score for each query
679
sequence by adding the normalised counts across all the RBPs. We used AME [51] with default
680
parameters to compare binding enrichment of RBPs between highly brain-specific (query) and
681
shared 3’UTR predictions (control). RBP motifs with an enrichment 𝑎𝑑𝑗𝑢𝑒𝑠𝑡𝑒𝑑 𝑝 − 𝑣𝑎𝑙𝑢𝑒 < 10−5
682
were considered to be significantly over-represented in highly brain-specific 3’UTR predictions
683
compared to shared 3’UTR predictions. Previously reported gene targets of TARDBP identified
684
using iCLIP technology were extracted from the POSTAR2 database [52].
685
686
The CNC scores, as reported by Chen et al. [27], were used to quantify the occurrence of CNCRs
687
within unannotated 3’UTRs. We extracted the CNC score for each 10 bp window and averaged it
688
across the query region to calculate a mean CNC score for each query region.
689
690
Calculating gene enrichment
691
To investigate molecular functions and biological processes significantly associated with a gene
692
list, we performed GO enrichment analysis using the ToppFun tool in the ToppGene suite [53].
693
GO terms attaining an enrichment q-value (false-discovery rate computed using Benjamini-
694
Hochberg method) < 0.05 were considered significant. Similarly, SynGO [28] was used to identify
695
enriched GO terms (q-value < 0.05) associated with synaptic function. To calculate enrichment of
696
genes associated with rare neurogenetic disorders, OMIM [54] genes related to neurological
697
29
disorders were used (1,948 genes). The list of genes associated with adult-onset
698
neurodegenerative disorders was extracted from Genomic England Panel App (254 green
699
labelled genes) [55]. A hypergeometric test was used to calculate the enrichment using the total
700
number of protein-coding genes (22,686) as the ‘gene universe’.
701
702
Data availability
703
Code used to perform analyses in this study is publicly available at https://github.com/sid-
704
sethi/F3UTER. Accession numbers of all data used in this study are listed in methods.
705
706
Acknowledgements
707
We thank Matthew Davis, Greg O'Sullivan and Carla Bento for their thoughtful feedback on this
708
study. This work was funded by a postdoctoral fellowship awarded to S.S. under the “Sustaining
709
Innovation Postdoctoral Training Program” at Astex Pharmaceuticals (S.S., H.S.). D.Z., S.G-R.,
710
and M.R. were supported by the award of a Tenure Track Clinician Scientist Fellowship to M.R.
711
(MR/N008324/1). Z.C. was supported through the award of a Leonard Wolfson Doctoral Training
712
Fellowship in Neurodegeneration. S.G. was supported through the award of an Alzheimer’s
713
Research UK PhD fellowship. J.A.B. was supported by the Science and Technology Agency,
714
Séneca Foundation, Comunidad Autónoma Región de Murcia, Spain, through the research
715
project 00007/COVI/20. We thank colleagues at the University College London, University of
716
Murcia and Astex Pharmaceuticals for helpful comments.
717
718
Author contributions
719
S.S., H.S., J.A.B., M.R. conceived and designed the study. S.S. conducted all the research and
720
data analysis. M.R., J.A.B., H.S. jointly supervised this study. D.Z., S.G. provided ER datasets in
721
GTEx tissues and helped with the analysis of ERs. S.S. developed the F3UTER online platform.
722
Z.C. provided help and data for the CNC analysis. S.S. wrote the manuscript with help from M.R.,
723
J.A.B., and H.S. All authors contributed, read and approved the manuscript.
724
725
30
Competing interests
726
The authors declare no competing interests.
727
728
Corresponding authors
729
Correspondence to Juan A. Botia, Harpreet Saini and Mina Ryten.
730
31
Supplementary Figures
731
732
Supplementary Figure S1.
733
Univariate comparisons of features and genomic classes. Plots show the relationship
734
between quantified features and genomic classes in the training dataset. A Kruskal-Wallis Test
735
was used to compare continuous values of features across the classes, while a proportion Z-test
736
was used for proportions. For each feature, the comparison across the classes was statistically
737
significant with a 𝑝 − 𝑣𝑎𝑙𝑢𝑒 < 2.2 × 10−16.
738
739
740
741
742
743
32
744
745
746
747
748
749
Supplementary Figure S2.
750
UMAP visualisation of genomic features. UMAP representation of all 41 omic features, with
751
elements labelled by genomic classes.
752
753
754
755
756
757
758
33
759
760
761
762
763
764
765
766
767
768
Supplementary Figure S3.
769
Performance of multinomial classification models measured using 5-fold cross validation
770
repeated 20 times. Boxplots comparing the accuracy and kappa of random forest multinomial
771
classifier and elastic net multinomial logistic regression model, to classify different genomic
772
classes. p: p-value calculated using Wilcoxon Rank Sum Test.
773
774
775
776
34
777
778
Supplementary Figure S4.
779
Contribution of features towards 3’UTR classification. Variable importance chart showing the
780
importance of features in classifying 3’UTRs from other transcribed elements in the genome, as
781
measured by mean decrease in accuracy and Gini. The features are ordered in decreasing order
782
of their relative importance and grouped based on their type.
783
35
784
785
786
Supplementary Figure S5.
787
Overlap between randomly selected intergenic ERs and poly(A) sites. Distribution of overlap
788
between randomly selected intergenic ERs and poly(A) sites from 10,000 permutations. Operm:
789
mean overlap of the permuted distribution; Oobs: observed overlap of 3’UTR predictions.
790
791
792
793
794
36
795
796
797
798
799
800
801
802
Supplementary Figure S6.
803
F3UTER predictions across 39 GTEx tissues. Bar plot showing the number of predictions in
804
each tissue, grouped and color-coded according to their prediction probability scores. Tissues are
805
sorted in descending order of the total number of predictions in each tissue. The square boxes
806
below the bars are color-coded to group the tissues according to their physiology.
807
808
809
810
811
812
813
814
815
816
817
37
818
819
820
821
822
823
824
825
826
Supplementary Figure S7.
827
Categorisation of F3UTER predictions based on tissue-specificity. Bar plots showing the
828
number of predictions grouped according to their tissue specificity across 39 tissues. Tissues are
829
sorted in descending order of the number of predictions. The square boxes below the bars are
830
color-coded to group the tissues according to their physiology.
831
832
833
834
835
836
837
838
839
840
38
841
842
843
Supplementary Figure S8.
844
Unannotated 3’UTR associated with C19orf12 in brain. Genomic view of the C19orf12 locus
845
displaying intergenic ERs and poly(A) sites from poly(A) atlas in the region. Two tracks are
846
displayed for each tissue - the top track shows coloured boxes which represent the intergenic
847
ERs, while the bottom track shows black lines which represent RNA-seq junction reads.
848
849
850
851
852
853
854
855
856
39
857
Supplementary Figure S9.
858
Examples of highly brain-specific unannotated 3’UTRs. Genomic view of genes (top: SCN2A;
859
middle: RTN2; bottom: OPA1) associated with an unannotated 3’UTR in brain, displaying
860
intergenic ERs and poly(A) sites from poly(A) atlas in the region.
861
862
863
864
865
866
867
868
40
869
870
871
872
Supplementary Table 1.
873
List of RBPs with significantly enriched binding in the brain-specific unannotated 3’UTRs
874
compared to shared unannotated 3’UTRs (𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑 𝑝 < 10−5). The enrichment p-value of the
875
motifs were adjusted for multiple tests using a Bonferroni correction.
876
877
878
Rank
RBP Name
RBP Ensembl id
p-value
Adjusted p-
value
1
RBFOX1
ENSG00000078328
5.83E-21
1.78E-18
2
KHSRP
ENSG00000088247
4.12E-17
8.21E-15
3
ERI1
ENSG00000104626
1.07E-15
3.85E-13
4
TIAL1
ENSG00000151923
5.24E-15
2.41E-12
5
ELAVL3
ENSG00000196361
6.95E-15
3.60E-12
6
CELF1
ENSG00000149187
1.18E-12
1.16E-10
7
SSB
ENSG00000138385
3.96E-13
1.77E-10
8
TARDBP
ENSG00000120948
2.72E-12
5.09E-10
9
PUM2
ENSG00000055917
1.55E-11
3.49E-09
10
ZFP36L2
ENSG00000152518
7.52E-12
4.30E-09
11
ZFP36
ENSG00000128016
7.52E-12
4.30E-09
12
HNRNPDL
ENSG00000152795
7.37E-11
9.14E-09
13
AGO2
ENSG00000123908
6.50E-10
1.11E-07
14
SRSF10
ENSG00000188529
6.02E-10
1.58E-07
15
HNRNPAB
ENSG00000197451
6.02E-10
1.58E-07
16
RBM5
ENSG00000003756
3.52E-10
1.60E-07
17
HNRNPA2B1
ENSG00000122566
4.09E-09
3.68E-07
18
ZRANB2
ENSG00000132485
2.72E-09
5.80E-07
19
SRSF3
ENSG00000112081
8.68E-09
1.49E-06
20
TRA2B
ENSG00000136527
8.68E-09
1.75E-06
21
HNRNPD
ENSG00000138668
8.12E-08
3.51E-05
22
AKAP1
ENSG00000121057
5.94E-07
9.51E-05
879
880
881
882
883
884
885
41
References
886
887
888
1.
Glisovic, T., et al., RNA-binding proteins and post-transcriptional gene regulation. FEBS
889
Lett, 2008. 582(14): p. 1977-86.
890
2.
Mayr, C., What Are 3' UTRs Doing? Cold Spring Harb Perspect Biol, 2019. 11(10).
891
3.
Miura, P., et al., Widespread and extensive lengthening of 3' UTRs in the mammalian
892
brain. Genome Res, 2013. 23(5): p. 812-25.
893
4.
Tushev, G., et al., Alternative 3' UTRs Modify the Localization, Regulatory Potential,
894
Stability, and Plasticity of mRNAs in Neuronal Compartments. Neuron, 2018. 98(3): p.
895
495-511.e6.
896
5.
Glock, C., M. Heumüller, and E.M. Schuman, mRNA transport & local translation in
897
neurons. Curr Opin Neurobiol, 2017. 45: p. 169-177.
898
6.
Cajigas, I.J., et al., The local transcriptome in the synaptic neuropil revealed by deep
899
sequencing and high-resolution imaging. Neuron, 2012. 74(3): p. 453-66.
900
7.
Gumy, L.F., et al., Transcriptome analysis of embryonic and adult sensory axons reveals
901
changes in mRNA repertoire localization. Rna, 2011. 17(1): p. 85-98.
902
8.
Tian, B., et al., A large-scale analysis of mRNA polyadenylation of human and mouse
903
genes. Nucleic Acids Res, 2005. 33(1): p. 201-12.
904
9.
Derti, A., et al., A quantitative atlas of polyadenylation in five mammals. Genome Res,
905
2012. 22(6): p. 1173-83.
906
10.
Xiang, Y., et al., Comprehensive Characterization of Alternative Polyadenylation in
907
Human Cancer. J Natl Cancer Inst, 2018. 110(4): p. 379-389.
908
11.
Marini, F., D. Scherzinger, and S. Danckwardt, TREND-DB-a transcriptome-wide atlas of
909
the dynamic landscape of alternative polyadenylation. Nucleic Acids Res, 2020.
910
12.
Gruber, A.J., et al., A comprehensive analysis of 3' end sequencing data sets reveals
911
novel polyadenylation signals and the repressive role of heterogeneous
912
ribonucleoprotein C on cleavage and polyadenylation. Genome Res, 2016. 26(8): p.
913
1145-59.
914
13.
Herrmann, C.J., et al., PolyASite 2.0: a consolidated atlas of polyadenylation sites from
915
3' end sequencing. Nucleic Acids Res, 2020. 48(D1): p. D174-d179.
916
14.
Je, G., et al., A novel extended form of alpha-synuclein 3'UTR in the human brain. Mol
917
Brain, 2018. 11(1): p. 29.
918
42
15.
Zhang, D., et al., Incomplete annotation has a disproportionate impact on our
919
understanding of Mendelian and complex neurogenetic disorders. Sci Adv, 2020. 6(24).
920
16.
Zhang, Y.E., et al., New genes expressed in human brains: implications for annotating
921
evolving genomes. Bioessays, 2012. 34(11): p. 982-91.
922
17.
Jaffe, A.E., et al., Developmental regulation of human cortex transcription and its clinical
923
relevance at single base resolution. Nat Neurosci, 2015. 18(1): p. 154-161.
924
18.
Khalil, A.M., et al., Many human large intergenic noncoding RNAs associate with
925
chromatin-modifying complexes and affect gene expression. Proc Natl Acad Sci U S A,
926
2009. 106(28): p. 11667-72.
927
19.
Guttman, M., et al., Chromatin signature reveals over a thousand highly conserved large
928
non-coding RNAs in mammals. Nature, 2009. 458(7235): p. 223-7.
929
20.
Pertea, M., et al., CHESS: a new human gene catalog curated from thousands of large-
930
scale RNA sequencing experiments reveals extensive transcriptional noise. Genome
931
Biol, 2018. 19(1): p. 208.
932
21.
Palazzo, A.F. and E.S. Lee, Non-coding RNA: what is functional and what is junk? Front
933
Genet, 2015. 6: p. 2.
934
22.
McInnes, L., et al., UMAP: Uniform Manifold Approximation and Projection. Journal of
935
Open Source Software, 2018. 3(29): p. 861.
936
23.
Singh, I., et al., Widespread intronic polyadenylation diversifies immune cell
937
transcriptomes. Nat Commun, 2018. 9(1): p. 1716.
938
24.
Erson-Bensan, A.E., Alternative polyadenylation and RNA-binding proteins. J Mol
939
Endocrinol, 2016. 57(2): p. F29-34.
940
25.
Giudice, G., et al., ATtRACT-a database of RNA-binding proteins and associated motifs.
941
Database (Oxford), 2016. 2016.
942
26.
Plass, M., S.H. Rasmussen, and A. Krogh, Highly accessible AU-rich regions in 3'
943
untranslated regions are hotspots for binding of regulatory factors. PLoS Comput Biol,
944
2017. 13(4): p. e1005460.
945
27.
Chen, Z., et al., Human-lineage-specific genomic elements: relevance to
946
neurodegenerative disease and <em>APOE</em> transcript usage. bioRxiv, 2020: p.
947
2020.04.17.046441.
948
28.
Koopmans, F., et al., SynGO: An Evidence-Based, Expert-Curated Knowledge Base for
949
the Synapse. Neuron, 2019. 103(2): p. 217-234.e4.
950
29.
Tcw, J. and A.M. Goate, Genetics of β-Amyloid Precursor Protein in Alzheimer's
951
Disease. Cold Spring Harb Perspect Med, 2017. 7(6).
952
43
30.
Liscic, R.M., et al., ALS and FTLD: two faces of TDP-43 proteinopathy. Eur J Neurol,
953
2008. 15(8): p. 772-80.
954
31.
Vuong, C.K., et al., Rbfox1 Regulates Synaptic Transmission through the Inhibitory
955
Neuron-Specific vSNARE Vamp1. Neuron, 2018. 98(1): p. 127-141.e7.
956
32.
Lee, J.A., et al., Cytoplasmic Rbfox1 Regulates the Expression of Synaptic and Autism-
957
Related Genes. Neuron, 2016. 89(1): p. 113-28.
958
33.
Bae, B. and P. Miura, Emerging Roles for 3' UTRs in Neurons. Int J Mol Sci, 2020.
959
21(10).
960
34.
Holt, C.E. and E.M. Schuman, The central dogma decentralized: new perspectives on
961
RNA function and local translation in neurons. Neuron, 2013. 80(3): p. 648-57.
962
35.
Wang, L., R.D. Dowell, and R. Yi, Genome-wide maps of polyadenylation reveal
963
dynamic mRNA 3'-end formation in mammalian cell lineages. Rna, 2013. 19(3): p. 413-
964
25.
965
36.
Beaudoing, E., et al., Patterns of variant polyadenylation signal usage in human genes.
966
Genome Res, 2000. 10(7): p. 1001-10.
967
37.
Tan, G. and B. Lenhard, TFBSTools: an R/bioconductor package for transcription factor
968
binding site analysis. Bioinformatics, 2016. 32(10): p. 1555-6.
969
38.
Pang, K.C., M.C. Frith, and J.S. Mattick, Rapid evolution of noncoding RNAs: lack of
970
conservation does not mean lack of function. Trends Genet, 2006. 22(1): p. 1-5.
971
39.
Wang, J., et al., Mouse transcriptome: neutral evolution of 'non-coding' complementary
972
DNAs. Nature, 2004. 431(7010): p. 1 p following 757; discussion following 757.
973
40.
Kelley, D. and J. Rinn, Transposable elements reveal a stem cell-specific class of long
974
noncoding RNAs. Genome Biol, 2012. 13(11): p. R107.
975
41.
Kannan, S., et al., Transposable Element Insertions in Long Intergenic Non-Coding RNA
976
Genes. Front Bioeng Biotechnol, 2015. 3: p. 71.
977
42.
Meysman, P., K. Marchal, and K. Engelen, DNA structural properties in the classification
978
of genomic transcription regulation elements. Bioinform Biol Insights, 2012. 6: p. 155-68.
979
43.
Abeel, T., et al., Generic eukaryotic core promoter prediction using structural features of
980
DNA. Genome Res, 2008. 18(2): p. 310-23.
981
44.
Gruber, A.J., et al., Terminal exon characterization with TECtool reveals an abundance
982
of cell-specific isoforms. Nat Methods, 2018. 15(10): p. 832-836.
983
45.
Kuhn, M., Building Predictive Models in R Using the caret Package. 2008, 2008. 28(5):
984
p. 26.
985
44
46.
Friedman, J., T. Hastie, and R. Tibshirani, Regularization Paths for Generalized Linear
986
Models via Coordinate Descent. J Stat Softw, 2010. 33(1): p. 1-22.
987
47.
Breiman, L., Random Forests. Machine Learning, 2001. 45(1): p. 5-32.
988
48.
Quinlan, A.R. and I.M. Hall, BEDTools: a flexible suite of utilities for comparing genomic
989
features. Bioinformatics, 2010. 26(6): p. 841-2.
990
49.
Bailey, T.L., et al., MEME SUITE: tools for motif discovery and searching. Nucleic Acids
991
Res, 2009. 37(Web Server issue): p. W202-8.
992
50.
Grant, C.E., T.L. Bailey, and W.S. Noble, FIMO: scanning for occurrences of a given
993
motif. Bioinformatics, 2011. 27(7): p. 1017-8.
994
51.
McLeay, R.C. and T.L. Bailey, Motif Enrichment Analysis: a unified framework and an
995
evaluation on ChIP data. BMC Bioinformatics, 2010. 11: p. 165.
996
52.
Zhu, Y., et al., POSTAR2: deciphering the post-transcriptional regulatory logics. Nucleic
997
Acids Res, 2019. 47(D1): p. D203-d211.
998
53.
Chen, J., et al., ToppGene Suite for gene list enrichment analysis and candidate gene
999
prioritization. Nucleic Acids Res, 2009. 37(Web Server issue): p. W305-11.
1000
54.
Amberger, J.S., et al., OMIM.org: Online Mendelian Inheritance in Man (OMIM®), an
1001
online catalog of human genes and genetic disorders. Nucleic Acids Res, 2015.
1002
43(Database issue): p. D789-98.
1003
55.
Martin, A.R., et al., PanelApp crowdsources expert knowledge to establish consensus
1004
diagnostic gene panels. Nat Genet, 2019. 51(11): p. 1560-1565.
1005
1006
| 2021 | Leveraging omic features with F3UTER enables identification of unannotated 3’UTRs for synaptic genes | 10.1101/2021.03.08.434412 | [
"Sethi Siddharth",
"Zhang David",
"Guelfi Sebastian",
"Chen Zhongbo",
"Garcia-Ruiz Sonia",
"Ryten Mina",
"Saini Harpreet",
"Botia Juan A."
] | creative-commons |
Classification: Biological Sciences: Environmental Sciences
Title: Exposure to environmental level pesticides stimulates and diversifies evolution in
Escherichia coli towards greater antibiotic resistance
Yue Xinga, Shuaiqi Wua, Yujie Mena,b,1
aDepartment of Civil and Environmental Engineering, University of Illinois at Urbana-
Champaign, Urbana, IL, United States
bInstitute for Genomic Biology, University of Illinois at Urbana-Champaign, Urbana, IL, United
States
1To whom correspondence may be addressed. Email: ymen2@illinois.edu
2
Abstract
1
Antibiotic resistance is one of the most challenging issues in public health. Antibiotic resistance
2
can be selected by antibiotics at sub-inhibitory concentrations, the concentrations typically
3
occurring in natural and engineered environments. Meanwhile, many other emerging organic
4
contaminants such as pesticides are frequently co-occurring with antibiotics in agriculture-related
5
environments and municipal wastewater treatment plants. To investigate the effects of the co-
6
existing, non-antibiotic pesticides on the development of antibiotic resistance, we conducted
7
long-term exposure experiments using a model Escherichia coli strain. The results revealed that
8
1) the exposure to a high level (in mg/L) of pesticides alone led to the emergence of mutants with
9
significantly higher resistance to streptomycin; 2) the exposure to an environmental level (in
10
µg/L) of pesticides together with a sub-inhibitory level (in sub mg/L) of ampicillin
11
synergistically stimulated the selection of ampicillin resistance and the cross-selection of
12
resistance to three other antibiotics (i.e., ciprofloxacin, chloramphenicol, and tetracycline).
13
Resistance levels of mutants selected from co-exposure were significantly higher than those of
14
mutants selected from ampicillin exposure only. The comparative genomic and transcriptomic
15
analyses indicate that distinct and diversified genetic mutations in ampicillin- and ciprofloxacin-
16
resistant mutants were selected from co-exposure, which likely caused holistic transcriptional
17
regulation and the increased antibiotic resistance. Together, the findings provide valuable
18
fundamental insights into the development of antibiotic resistance under environmentally
19
relevant conditions, as well as the underlying molecular mechanisms of the elevated antibiotic
20
resistance induced by the exposure to pesticides.
21
Keywords: Antibiotic resistance; Emerging organic contaminants; Pesticides; Mutation;
22
Escherichia coli
23
3
Significance statement
24
Antibiotic resistance is a major threat to public health globally. Besides clinically relevant
25
environments, the emergence and spread of resistant bacteria in non-clinical environments can
26
also potentially pose risks of therapy failures. This study showed that the long-term,
27
environment-level exposure to pesticides with and without antibiotics significantly stimulated the
28
development of greater antibiotic resistance. The resistant strains selected from the exposure to
29
pesticides are genetically and metabolically distinct from the ones selected by the antibiotic only.
30
Although it is still being debated regarding whether or not a large use of antibiotics in plant
31
agriculture is harmful, our findings provide the first fundamental evidence that greater concerns
32
of antibiotic resistance may result if antibiotics are applied together with non-antibiotic
33
pesticides.
34
4
Introduction
35
Antibiotic resistance has been one of the most challenging environmental and public health
36
issues. The de novo mutation is one important route for bacteria to acquire antibiotic resistance,
37
under both clinical and environmental conditions (1-4). Antibiotics at both inhibitory (typically
38
in the mg/L range) and sub-inhibitory concentrations (below the minimal inhibitory
39
concentration (MIC); in the high µg/L or ng/L range) can lead to increased resistance emergence
40
(1, 3, 5). The latter is of more concern due to the ubiquitous occurrence of antibiotic residues at
41
low (i.e., sub-MIC) levels in the environment. This may explain the emergence of antibiotic
42
resistance in many non-clinically relevant environments such as domestic sewage, water bodies
43
receiving treated sewage from municipal wastewater treatment plants (WWTPs), as well as farm
44
run-off where antibiotics occur from tens of ng/L to several hundreds of g/L (6-9). Although
45
there is limited direct evidence in terms of resistant phenotypes, the wide surveillance of
46
antibiotic resistance genes (ARG) revealed elevated ARG levels in treated wastewater and the
47
receiving environment along with the occurrence of low-level antibiotics (10-13). In addition,
48
cross-selection occurs in mutants exposed to a single antibiotic at sub-MIC levels, which
49
developed resistance not only to the exposed antibiotic but also to other non-exposed antibiotics,
50
indicating the complexity of antibiotic resistance development (14, 15). With similar cytotoxicity
51
to antibiotics, disinfectants and disinfection byproducts were found to promote the development
52
of antibiotic resistance at high levels (several to a few thousand mg/L) (16-19).
53
Moreover, in natural and built environments a variety of other emerging organic
54
contaminants such as pesticides, non-antibiotic drugs, and personal care products are usually co-
55
occurring with antibiotics at low levels (20-24). It is still unclear how these non-antibiotic
56
emerging organic contaminants at environmentally relevant concentrations would affect the
57
5
selection of antibiotic resistance by antibiotics. If synergistic effects were present, antibiotic
58
resistance levels in those environments would be underestimated by only considering the
59
antibiotic occurrence. Thus, it is crucial to obtain a better understanding of the emergence of
60
antibiotic resistance with exposure to both antibiotics and non-antibiotic organic contaminants at
61
environmental levels.
62
Pesticides are one important group among those contaminants co-existing with
63
antibiotics. They are typically found in agricultural soils, run-offs, and the receiving water bodies
64
(25-28), which are also potentially antibiotic-impacted environments. For example, antibiotics
65
used in farms to treat sick animals and boost livestock growth can be released into those
66
environments (8, 29), leading to the co-occurrence of pesticides and antibiotics. Moreover,
67
antibiotics have also been applied to fight against plant bacterial diseases in plant farms (30),
68
where pesticides may also be used. For instance, currently the US Environmental Protection
69
Agency is in the process of allowing heavy usage of two antibiotics, streptomycin and
70
oxytetracycline, to combat citrus greening, a bacterial disease killing citrus trees (31). In
71
addition, pesticides also occurred in other non-clinical antibiotic-impacted environments such as
72
irrigation water and municipal wastewater (32-37). The environmental levels of each individual
73
pesticide range from less than 1 ng/L to tens of g/L (25, 26), which brings an overall
74
environmental occurrence to a high µg/L range due to the presence of many pesticide species
75
used for different purposes in farms and households. Some pesticides like biocides share similar
76
inhibitory mechanisms with antibiotics, such as membrane disruption (38-40) and inhibition of
77
cell wall synthesis (41, 42) , which might favor mutations towards the co-selection of antibiotic
78
resistance.
79
6
The goal of this study was to fill the knowledge gap regarding the emergence of
80
antibiotic resistance in environments with the occurrence of both antibiotics and pesticides. We
81
aimed to 1) investigate the effects of environmental level exposure to pesticides alone and the
82
interactive effects with sub-MIC antibiotics (synergistic, antagonistic, or neutral) on the
83
development of antibiotic resistance, and 2) identify the underlying mechanisms of the emerged
84
antibiotic resistance from 1). To accomplish our goals, we designed evolutionary experiments
85
with a susceptible Escherichia coli strain being exposed to pesticides only and co-exposed to
86
ampicillin and pesticides at environmental levels for 500 generations. The change in antibiotic
87
resistance of de no mutants was determined. Genetic mutations of the de no mutants selected
88
from different exposure conditions were identified by whole-genome sequencing (WGS).
89
Transcriptional regulation in resistant mutants from co-exposure compared to those from
90
antibiotic exposure alone was examined using RNA sequencing (RNA-seq) and reverse
91
transcription quantitative PCR (RT-qPCR). The genomic and transcriptomic analyses provide
92
insights into different mechanisms of antibiotic resistance developed under those exposure
93
conditions.
94
Results
95
Effects of the Exposure to Pesticides on the Development of Streptomycin Resistance. We
96
initiated evolutionary experiments with a susceptible strain E. coli K12 C3000. Starting from one
97
ancestor strain (G0), parallel population passages were exposed to constant concentrations of
98
pesticide mixture. The mixture consisted of 23 frequently detected pesticides in various natural
99
and engineered environments, including biocides, fungicides, herbicides, and insecticides (Table
100
S1). The concentrations of pesticides used in this study were based on their environmental
101
concentrations (EC) (0.1 – 4.8 µg/L, each; ~ 20 µg/L in total) (Table S1). The total pesticide
102
7
concentrations used for the exposure experiments range from 1/125EC (0.16 µg/L) to 125EC
103
(2.5 mg/L). In the control experiments, E. coli passages were obtained from the same ancestor
104
cells in the same way but without pesticide exposure.
105
Among all pesticide exposure levels, an increased mutation frequency (1 – 2 orders of
106
magnitude higher) towards streptomycin resistance was observed in E. coli populations being
107
exposed to 125EC pesticides for 500 generations (G500) (Figure S1), although with no statistical
108
significance due to the large variations among triplicated evolution passages. This indicates that
109
exposure to high-level (2.5 mg/L) pesticides stimulated the emergence of streptomycin (Strep)
110
resistance. To further characterize the resistance levels of Strep-resistant mutants, 36 resistant
111
mutants were randomly picked up on selective solid media, and their MICs were determined.
112
Mutants selected from 125EC exposure acquired significantly higher Strep-MICs compared to
113
the mutants from the no-exposure control (p-value = 0.0013, N = 36, Mann-Whitney U test)
114
(Figure 1). It is worth noting that the exposure to high-level pesticides favored the emergence of
115
de novo mutants more resistant to Strep only, but not to the other four tested antibiotics, i.e.,
116
ampicillin (Amp), tetracycline (Tet), ciprofloxacin (Cip), and chloramphenicol (Chl).
117
Effects of Pesticides + Ampicillin Co-exposure at Environmental Levels on the
118
Development of Antibiotic Resistance. We next tested the combined effects of environmental
119
level pesticides and Amp (1/125 – 1/5 MIC0) on the development of antibiotic resistance in E.
120
coli. According to the MIC distribution of the 36 Amp-resistant mutants, those selected from
121
populations co-exposed to 1/5MIC0 Amp and 1EC pesticides exhibited a shift to higher MICs
122
(Figure 2B), compared to those selected from populations exposed to 1/5MIC0 Amp (Figure 2A).
123
The shift of MIC distribution was statistically significant (p-value = 0.039, Mann-Whitney U
124
test). To explore the development of cross-resistance, we determined the MIC distributions of E.
125
8
coli mutants from co-exposed and Amp-exposed populations (G500) resistant to four other
126
antibiotics: Strep, Chl, Cip, and Tet. Except for Strep with similar resistance developed under co-
127
and Amp-exposure conditions, the resistance (i.e., MICs) to the other three antibiotics were
128
significantly higher (1.5 – 3.5 times) for resistant mutants from co-exposure than from Amp-
129
exposure (p-values of 1.1 × 10-10, 0.044, and 1.3 × 10-7 for Chl, Cip, and Tet, respectively)
130
(Figure 2 D, F and H). The co-exposure also accelerated the development of resistance to the
131
four antibiotics with mutation frequencies 1 – 4 orders of magnitude higher than those under
132
Amp-exposure only, although without statistical significance due to variations among the
133
triplicated evolution passages (Figure S2). Collectively, the co-exposure to environmental level
134
pesticides and 1/5MIC0 ampicillin exhibited synergistic effects on the emergence of mutants
135
resistant to not only the exposed antibiotic but also other non-exposed antibiotics (cross-
136
selection). The co-exposure condition selected mutants more resistant than those selected under
137
sub-MIC antibiotic selection pressure only. One should note that no accelerated development of
138
higher resistance was observed for G500 E. coli cells exposed to the same concentration of
139
pesticides but with lower ampicillin concentrations (1/125MIC0 and 1/25MIC0), as well as the
140
G500 non-exposure control (data not shown).
141
Genetic Mutations in Strep-resistant Mutants Selected by High-level Pesticide Exposure.
142
To unravel the mechanisms leading to the elevated mutation frequency towards Strep
143
resistance and the higher Strep resistance of E. coli mutants after being exposed to high-level
144
pesticides (125EC), we identified valid genetic mutations including non-synonymous single
145
nucleotide polymorphisms (SNPs), insertions and deletions (INDEL) in the resistant mutants
146
compared to the susceptible strains from G500 population without pesticide exposure, which
147
have the same MIC as the ancestor strain (G0).
148
9
The genomes of Strep-resistant mutants isolated from G500 E. coli with pesticide
149
exposure revealed genetic mutations in four genes including two SNPs, and two deletions against
150
the genome of the susceptible isolate from G500 E. coli without pesticide exposure (Table 1, see
151
Table S3 in the SI for a complete list of genetic mutations). These four mutated genes encode
152
proteins for: (i) target modification; (ii) DNA replication and repair; (iii) regulation. It is
153
noteworthy that all three sequenced Strep-resistant mutants from the pesticide-exposed cultures
154
shared the same genetic mutation of the rpsG gene (SNP: A → T), resulting in gaining a stop
155
codon (*) replacing Leu157 in the amino acid sequence (Table 1). The rpsG gene encodes the
156
30S ribosomal protein S7, which is essential for cell growth. The stop codon gained at a later
157
position (157 of 179 residues) of the amino acid sequence did not affect the function of this
158
protein, as no growth inhibition of the resistant mutants was observed (data not shown).
159
Genetic Mutations in Amp- and Cip-resistant Mutants Selected by the Pesticides +
160
Ampicillin Co-exposure. To explore mechanisms leading to the higher Amp-resistance of
161
mutants isolated in E. coli populations exposed to pesticides + Amp, we identified and compared
162
the genetic mutations in Amp-resistant mutants from E. coli under co-exposure and Amp-
163
exposure. We also did the same comparative genomic analysis to study the underlying
164
mechanisms of the increase of cross-resistance to antibiotics other than the exposed Amp. We
165
focused on Cip-resistant mutants, as they showed the highest MIC increase after the co-exposure.
166
For the three Amp-resistant mutants isolated from the co-exposed culture, the same
167
mutation occurred in gene ftsI (SNP: A → T; amino acid change: Gln536 → Leu) (Figure 3 &
168
Table 1). It encodes an Amp-binding protein, and this genetic mutation likely altered the protein
169
structure, hence lowering the binding affinity of Amp to this protein. In addition, multiple
170
mutations (non-synonymous SNPs and insertions) occurred in a prophage-related gene yagJ.
171
10
Besides, mutations also occurred in genes encoding membrane and flagellar structure proteins
172
(Table 1). The structural alteration of these proteins could potentially limit or avoid the entry of
173
the antibiotic into the cells (41, 42), thus resulting in antibiotic resistance.
174
Interestingly, the mutations identified in Amp-resistant mutants isolated from co-exposed
175
E. coli were completely different from those isolated from Amp-exposed E. coli. Fewer genetic
176
mutations were detected in the Amp-resistant mutants from Amp-exposure, none of which was
177
shared among the three sequenced mutants. One mutant had an SNP mutation in acrR involved
178
in multidrug transport (43, 44) (Figure 3 & Table S3). Another mutant had an SNP mutation in
179
the proline transport gene proV, which occurred in a multi-drug-resistant Salmonella strain (45).
180
The third mutant had an SNP mutation in an isocitrate dehydrogenase encoding gene icd, the
181
mutation of which has been observed in E. coli mutants resistant to nalidixic acid (46). Together,
182
many of the identified mutated genes in the Amp-resistant strains isolated from both co-exposure
183
and Amp-exposure conditions have resistance-related functions, which likely led to the
184
development of Amp-resistance. Moreover, the co-exposure selected Amp-resistant mutants with
185
distinct genetic mutations, which likely contributed to their higher MIC levels than those selected
186
by Amp-exposure only.
187
For Cip-resistant mutants isolated from co-exposed E. coli, mutations in the gyrA gene
188
occurred in all three sequenced mutants: two had Ser83 → Leu and one had Asp87 → Gly (Table
189
1). The DNA gyrase encoded by gyrA is the target of Cip, and the mutations in gyrA might lead
190
to the resistance to Cip (47). Along with gyrA mutations, more diverse genetic mutations were
191
detected in Cip-resistant mutants from co-exposure than from Amp-exposure, including genes
192
with various functions: (i) DNA replication and repair; (ii) drug transporter and degrader, and
193
efflux pumps; (iii) membrane structure and transporter; (iv) regulator; (v) prophage; and (vi)
194
11
energy metabolism. Most of these mutations were not directly involved in known Cip-resistant
195
mechanisms. It seems that the co-exposure not only accelerated but also diversified the
196
evolution, resulting in the selection of Cip-resistant mutants with higher resistance.
197
There were fewer genetic mutations detected in the Cip-resistant mutants from Amp-
198
exposure, which occurred in genes encoding proteins for target modification, transporters, and
199
regulators (Figure 3 & Table S3). One of the three sequenced mutants had the same gyrA
200
mutation (Asp87 → Gly) as the one that occurred to the Cip-resistant mutant from co-exposure
201
condition. The other two mutants had an SNP mutation (T → C, Thr120 → Ala) in the envZ gene
202
that encodes a membrane-associated protein kinase in the two-component regulatory system,
203
which might reduce the production of membrane porin and lead to antibiotic resistance (48). The
204
same genetic mutations of proV and acrR genes as those in the Amp-resistant mutants were
205
found in Cip-R mutants from Amp-exposure, suggesting a more general resistance mechanism
206
not only to ampicillin but also to other types of antibiotics.
207
As it is not financially applicable and practically feasible to sequence all antibiotic-
208
resistant mutants isolated from Amp- and co-exposure for genomic comparison, complementary
209
SNP genotyping assays were conducted to examine the prevalence of the identified genetic
210
mutations from three biological replicates by WGS in the entire resistant population of G500 E.
211
coli under co- and Amp-exposure conditions. As a representative, the ftsI gene, which showed
212
the same SNP mutation among all three sequenced mutants from the co-exposure condition was
213
targeted by the SNP genotyping assay. We treated the co-exposed and Amp-exposed G500 E.
214
coli with 4 mg/L ampicillin (i.e., MIC0, Amp) to select resistant E. coli populations in the liquid
215
media and then detected the genotyping patterns in the resistant populations. The mutated ftsI
216
genotype was only detected in the resistant populations selected from co-exposed G500 E. coli
217
12
(Figure S3). Despite the varied fractions of ftsI mutants in the three biological replicates (1.2%,
218
30.5%, and 99.8%), the presence/absence of mutated ftsI determined by SNP genotyping assay is
219
consistent with the WGS results. Thus, the detection and frequency of genetic mutations from
220
three selected mutant genomes can qualitatively represent the presence and dominance of the
221
genotypes in the resistant population. In line with the SNP genotyping results, the replicate from
222
co-exposure containing 99.8% mutated ftsI showed more than one order of magnitude higher
223
mutation frequency than the other two replicates (Figure S2). This suggests that the mutated ftsI
224
contributed to the accelerated development of Amp-resistance under the co-exposure condition,
225
and perhaps resulted in the higher Amp-resistance than the resistant mutants from Amp-exposure
226
that developed different genetic mutations and resistance mechanisms.
227
Differential Gene Expression of Resistant Mutants Isolated from Amp-exposed and Co-
228
exposed E. coli Cultures. To further investigate the resistance mechanisms developed under co-
229
and Amp-exposure conditions, differential gene expression analysis at the transcriptional level
230
was conducted using RNA-seq. Principal component analysis indicates a clear difference
231
between resistant mutants from co-exposure and those from Amp-exposure (Figure 4 A and B).
232
A total of 92 and 107 genes exhibited significantly higher/lower expression (FDR < 0.05, 2-
233
fold change) in Amp-R mutants and Cip-R mutants from co-exposure, respectively, compared to
234
those from Amp-exposure. Hierarchical clustering revealed six distinct clusters of the
235
differentially expressed genes under 8 functional categories. (Figure 4 C, D and details in Table
236
S4).
237
Some genes in cluster I and almost all genes in cluster III showed significantly lower
238
expression in Amp-R mutants from co-exposure than Amp-exposure, such as genes involved in
239
flagellar structure formation (e.g., fliC), arginine synthesis (e.g., argA), carbohydrate transport
240
13
(e.g., argA, mglA), cold shock defense (e.g., cspH), prophage (e.g., nmpC, yjhQ), and fatty acid
241
β-oxidation (e.g., fadB, fadH). Moreover, the expression of genes in cluster IV was completely
242
shut down in Amp-R mutants from co-exposure, including CP4-6 prophage genes (e.g., yagE,
243
the mutated yagJ, and mmuM (also involving methionine synthesis)) and arginine synthesis
244
genes (e.g., argF) (Figure S4A). In contrast, genes in cluster VI showed higher expression in
245
Amp-R mutants from co-exposure, including heat shock and acid stress defense genes, such as
246
ibpA and hdeA; genes involved in glutamate decarboxylation (gadA, gadB, and gadC), putrescine
247
degradation (e.g., puuB) and histidine synthesis (hisA and hisF); and a membrane structure gene
248
(yhiM) (Figure S4A). Two fimbriae-associated genes, fimB and fimE, in cluster II also showed
249
higher expression levels in Amp-R mutants from co-exposure.
250
The Cip-R mutants from co-exposure exhibited higher expression of most genes in
251
cluster I and II (Figure 4 C and Figure S4 B), including genes related to polymycin resistance
252
(pmrD and arnF), heat shock defense (patZ and ygcP), oxidative stress defense (bsmA), histidine
253
synthesis (e.g., hisJ and hisM), glyoxylate cycle (e.g., aceA), and N-acetylneuraminate
254
degradation (e.g., nanA). In contrast, the expression of genes in cluster V, such as those
255
associated with nitrate reduction (e.g., narV) and lipid degradation (pagP and hdhA) (Figure 4C
256
and Figure S4B), was substantially lower in Cip-R mutants from co-exposed culture. The above
257
gene expression patterns were quite different from those for Amp-R mutants, suggesting
258
different resistance mechanisms. Interestingly, exceptions are found for two genes (i.e., fimB and
259
fimE) encoding fimbrial structures, whose expression was stimulated in mutants resistant to both
260
Amp and Cip from co-exposure. Besides, genes in cluster IV (e.g., prophage genes yagJ, mmuM)
261
exhibited similar expression in both Amp-R and Cip-R mutants, which were turned off in
262
mutants from co-exposed E. coli (Figure 4 C). These shared responses of the mutants from the
263
14
co-exposure condition suggest the involvement of those genes in the resistance both to Amp and
264
Cip. In addition, among all mutated genes identified in Amp-R and Cip-R mutants, yagJ was the
265
only one that exhibited a significantly differential expression, in which there were several shared
266
site mutations between the Amp- and Cip-R mutants from the co-exposure condition.
267
Discussion
268
This work provides evidence that long-term exposure to pesticides alone or together with
269
sub-MIC level antibiotics can stimulate and diversify de novo mutations towards resistance of
270
certain antibiotics. The findings are of high relevance to the emergence of antibiotic resistance in
271
some natural and built environments. High pesticide levels (mg/L) triggering evolution towards
272
resistance may occur in biosolids and aquatic organisms where pesticides can be accumulated
273
(49-51). In aquatic environments receiving WWTP effluent and agricultural runoff, antibiotics at
274
sub-MIC levels are occurring together with pesticides at ng - µg/L (20-24). Such co-occurrence
275
may synergistically select for de novo mutants resistant to antibiotics from a susceptible
276
population, with even higher resistance than those that could have been selected by antibiotic
277
exposure alone.
278
Mutation in genes encoding antibiotic target proteins is one of the mechanisms leading to
279
higher resistance of mutants from pesticide-exposed and co-exposed E. coli. The higher
280
resistance to Strep for mutants from pesticide-exposure was attributed to the stop-gain mutation
281
in rpsG at a later amino acid position. The rpsG gene encodes a component (protein S7) of the
282
30S subunit of ribosome, and Strep binds to the 30S subunit to inhibit protein synthesis. This is
283
different from previous findings that several site mutations in rpsL, another gene in the same
284
operon encoding 30S subunit ribosomal protein S12, can lead to the structure alteration of 30S
285
subunit, thus Strep-resistance in E. coli strains (52-54). The genetic change of rpsG uniquely
286
15
selected under pesticide exposure may alter the structure of S7 and the entire 30S subunit,
287
resulting in lower affinity, hence less sensitivity to Strep in the de novo mutants. Mutations in the
288
antibiotic target genes, ftsI and gyrA for Amp and Cip, respectively, occurred exclusively (for
289
ftsI) or more frequently (for gyrA) in resistant mutants from co-exposure than those from Amp-
290
exposure. Direct alteration of the target proteins can be more effective to overcome the inhibitory
291
effect of antibiotics than mutations in other resistance-related genes in the resistant mutants from
292
Amp-exposure, leading to higher antibiotic resistance (MIC levels).
293
Moreover, the co-exposure to pesticides and Amp stimulated and diversified genome-
294
wide mutations, and mutants with diverse mutations were selected under Cip stress, thus likely
295
contributing to the higher Cip-resistance. Common mutations in a prophage gene yagJ were
296
shared in both Amp- and Cip-resistant mutants from co-exposure, but not from Amp-exposure.
297
The mutated gene yagJ exhibited differential expression (i.e., a complete shutdown) in both
298
Amp- and Cip-resistant mutants from co-exposure compared to Amp-exposure. This differs from
299
the previous findings that the removal of prophage CP4-6 genes including yagJ decreased the
300
resistance to nalidixic acid (55), which is a quinolone antibiotic, as Cip is.
301
Previous studies (14, 15, 47) about the resistance mechanisms mostly focused on genetic
302
mutations and the expression of antibiotic resistance genes. The global differential gene
303
expression has not been well understood. Compared to the resistant mutants from Amp-exposure
304
grown with antibiotic stress, the resistant mutants from co-exposure showed differential
305
expression of many genes involved in metabolic activities and cell structure formation. Such
306
different transcriptional responses to the same antibiotic stress may be related to the higher
307
antibiotic resistance observed for the mutants from co-exposure than those from Amp-exposure.
308
Amp- and Cip-R mutants from co-exposure shared several gene expression patterns, including (i)
309
16
the stimulated expression of fimbriae synthesis genes promoting cell adhesion, and (ii) the
310
deactivated expression of CP4-6 prophage-related genes, including yagJ, ykgS, and mmuM.
311
These features may promote bacterial survival under stress conditions, rendering multidrug
312
resistance.
313
In addition, we validated the differential gene expression results by RNA-seq using RT-
314
qPCR targeting selected genes (Figure S5). According to RT-qPCR results, we also found that
315
the differential expression of some genes in resistant mutants from co-exposure was independent
316
of whether they were grown with antibiotic stress or not. For example, fimB and fliC in resistant
317
mutants from co-exposure showed higher expression levels compared to the resistant mutants
318
from Amp-exposure even when growing without antibiotic stress (Figure S6). This suggests that
319
the distinct genetic mutations found in resistant mutants from co-exposure directly led to some
320
transcriptional regulation without an antibiotic stimulus.
321
Taken together, this study unravels an overlooked role of pesticides in promoting the
322
emergence of resistance to some antibiotics and selecting more resistant mutants with and/or
323
without the presence sub-MIC level antibiotics. It gives a better understanding of the molecular
324
mechanisms leading to the higher antibiotic resistance in E. coli after being exposed to multiple
325
selection pressures rather than to antibiotics alone. This provides important insights into
326
antibiotic resistance developed under more environmentally relevant exposure conditions.
327
328
Materials and Methods
329
Bacterial Strains, Growth and Selection Conditions. The antibiotic susceptible bacterium used
330
in this study was the gram-negative Escherichia coli K-12 C3000 (E. coli). The growth medium
331
for all selection experiments was Luria-Bertani (LB) broth. First, the stock E. coli cells from -80
332
17
ºC freezer were revived and then streaked on an LB agar plate and allowed to grow for 20 hours.
333
One single colony was picked and inoculated into a tube containing 3 mL of LB broth for 24-
334
hour incubation at 30 ºC. The cell culture was considered as the ancestor strain and used for
335
subsequent exposure experiments.
336
Twenty-three pesticides that have been frequently detected in environmental samples
337
were selected. Their environmental concentrations (EC) range from 0.1 to 4.8 μg/L. Detailed
338
information of the selected pesticides is in Table S1. Two exposure experiments were conducted:
339
(1) Exposure to pesticide mixture of 1/125EC, 1/25EC, 1/5EC, 1EC, 5EC, 25EC, 125EC,
340
mimicking a wide range of the pesticide occurrence in various environments with degradation or
341
accumulation of pesticides. A no chemical exposure was also set up as the control. (2) Exposure
342
to a combination of ampicillin of 1/125MIC0, 1/25MIC0, or 1/5MIC0 (MIC0, MIC of antibiotics
343
for the G0 E. coli strain in LB medium, MIC0, Amp = 4 mg/L) and pesticides of 1EC was applied.
344
The corresponding control was exposure only to ampicillin.
345
The pesticide stock mixture was dissolved in methanol. Appropriate volumes of the
346
mixture were added to the 96-well plate, which was air-dried until all the methanol was gone.
347
195 L LB medium and 5 L ampicillin stock solution were subsequently added to the wells.
348
The negative control group was added with the same volume of nanopure water as the ampicillin
349
solution. The cultures were incubated at 30 C and aerated by shaking. The culture was serially
350
passaged by 500-fold dilution every 24 hours for 500 generations (9 generations of growth per
351
serial passage). All exposure conditions and controls were performed with triplications.
352
Isolation of Resistant Strains and Determination of Minimum Inhibitory Concentrations.
353
The MIC0 of the ancestor strain was determined by the MIC test applied to 5 different types of
354
antibiotics: ampicillin, tetracycline, ciprofloxacin, streptomycin, and chloramphenicol. Briefly,
355
18
95 μL of LB medium and 5 μL of the antibiotic stock solution were added, respectively. An
356
overnight culture was prepared and diluted with 0.9% NaCl solution to optical density at 600 nm
357
(OD600) of around 0.1 as the standard solution. Then 0.5 μL of the standard solution was added
358
into fresh LB medium containing antibiotics at a series of concentrations. For the growth control
359
group, 5 μL of nanopore water was used instead of the antibiotic solution. For the negative
360
control group, 5 μL of nanopore water was used instead of the antibiotic solution and no E. coli
361
was inoculated. Cell culture was incubated at 35 C for 20 hours and OD600 were measured. The
362
MIC was determined as the concentration that inhibited 90% of growth based on OD600.
363
After 500 generations, 5-time diluted cell cultures were spread on LB agar plates
364
containing antibiotics at MIC0. Twelve resistant mutants were randomly picked up, and in total
365
there were 36 resistant mutants from each exposure condition. The MICs of these resistant
366
mutants were further determined. The Mann-Whitney U test was used to statistically analyze the
367
difference of MICs among resistant mutants under different exposure conditions (p-value <
368
0.05).
369
DNA Extraction, Whole-Genome Sequencing (WGS), and SNP Calling. Different E. coli
370
isolates were cultured overnight in LB media and cell pellets were collected by centrifugation.
371
DNA was extracted from each isolate using the DNeasy Blood and Tissue Kit (Qiagen)
372
according to the manufacturer’s instructions. The DNA concentration and quality were
373
determined on a Qubit 4 Fluorometer (Thermo Fisher Scientific, Wilmington, DE).
374
The obtained DNA obtained was then subjected to Illumina MiSeq 250-bp paired-end
375
sequencing carried out by Roy J. Carver Biotechnology Center at the University of Illinois. An
376
average coverage of 961,648 reads per isolate was obtained. A dynamic sequence trimming was
377
done by SolexaQA software (56) with a minimum quality score of 24 and a minimum sequence
378
19
length of 50 bp. The trimmed reads of the ancestor isolate (G0) were aligned against the E. coli
379
K12 MG1655 genome available at NCBI GenBank (NC_000913.3) using the Bowtie 2 toolkit
380
(57) to assemble the genome of E. coli at G0. All reads from isolates after G500 were then
381
aligned against the assembled G0 genome. SAMtools and Picard Tools were used to format and
382
reformat the intermediate-alignment files (58). SNPs and INDELs were identified with the
383
Genome Analysis Toolkit UnifiedGenotyper (59), with the calling criteria of > 5-read coverage
384
and > 50% mutation frequency at the mutation position.
385
RNA Extraction, RNA-Seq, and Differential Gene Expression Analyses. One Amp-resistant
386
mutant strain and one Cip-resistant mutant strain from different exposure conditions were
387
selected from sequenced mutants. Mutants were grown in a shaking incubator at 35 C in 8 mL
388
LB broth for 5 hours to OD600 = 0.75. Each condition had 3 biological replicates. The cultures
389
then were divided into two, one aliquot with the stress of 0.8×MIC antibiotic, one aliquot
390
without antibiotic treatment. Cultures were allowed to grow for an additional 30 minutes, then
391
cell pellets were collected by centrifugation.
392
Total RNA was isolated according to the acid phenol: chloroform extraction method, as
393
previously described (60) and treated with DNase to remove residual DNA using TURBO DNA-
394
free kit (Thermo Fisher Scientific). Ribosomal RNA was removed and sample libraries of
395
resistant mutants with antibiotic treatment were built using a Truseq mRNA-Seq Library
396
Preparation Kit (Illumina, USA), according to the manufacturer’s recommendations. Sequencing
397
was performed on a HiSeq 2500 system (Illumina, USA) and produced 100-base single-end
398
reads. The purified RNA samples of resistant mutants without antibiotic treatment, as well as
399
those of G500 susceptible strains were reverse-transcribed to cDNA and stored properly for RT-
400
qPCR measurement (See Supplementary Methods).
401
20
Low-quality RNA-seq reads (quality score < 30, sequence length < 25 bp) were removed
402
using SolexaQA software (56). The qualified sequences were subject to the alignment using
403
Bowtie 2 toolkit against the reference genome. Genes were counted using FeatureCounts
404
software (61), and the count data were then analyzed using R version 3.5.1 and Bioconductor
405
package DESeq2 version 3.8 (62). Genes were considered significantly differentially expressed
406
based on these three criteria: (a) TPM (Transcripts per million) > 5 in at least one of the samples;
407
(b) FDR (False Discovery Rate) adjusted p-value < 0.05; (c) > 2-fold difference in TPM values.
408
Principle component analysis was performed using normalized counts according to the
409
DESeq2 output. Hierarchical clustering by transforming the normalized count data was applied
410
based on the correlation distance and Ward aggregation criterion.
411
Accession numbers. All WGS and RNA sequencing data have been deposited in the NCBI SRA
412
database under accession no. PRJNA530028.
413
Acknowledgements
414
We would like to give thanks to Hernandez Alvaro Gonzalo and Chris L. Wright at the Roy J.
415
Carver Biotechnology Center, University of Illinois at Urbana-Champaign for whole genome
416
sequencing and RNA sequencing support.
417
References
418
1.
Andersson DI & Hughes D (2014) Microbiological effects of sublethal levels of
419
antibiotics. Nat Rev Microbiol 12(7):465-478.
420
2.
Davies J & Davies D (2010) Origins and evolution of antibiotic resistance. Microbiol Mol
421
Biol Rev 74(3):417-433.
422
21
3.
Gullberg E, Cao S, Berg OG, Ilbäck C, Sandegren L, Hughes D, & Andersson DI (2011)
423
Selection of resistant bacteria at very low antibiotic concentrations. PLoS Pathog
424
7:e1002158.
425
4.
Long H, Miller SF, Strauss C, Zhao C, Cheng L, Ye Z, Griffin K, Te R, Lee H, Chen C-
426
C, & Lynch M (2016) Antibiotic treatment enhances the genome-wide mutation rate of
427
target cells. Proc Natl Acad Sci USA 113(18):E2498-2505.
428
5.
Andersson DI & Hughes D (2011) Persistence of antibiotic resistance in bacterial
429
populations. FEMS Microbiol Rev 35(5):901-911.
430
6.
Brown KD, Kulis J, Thomson B, Chapman TH, & Mawhinney DB (2006) Occurrence of
431
antibiotics in hospital, residential, and dairy effluent, municipal wastewater, and the Rio
432
Grande in New Mexico. Sci Total Environ 366(2):772-783.
433
7.
McArdell CS, Molnar E, Suter MJF, & Giger W (2003) Occurrence and fate of macrolide
434
antibiotics in wastewater treatment plants and in the Glatt Valley watershed, Switzerland.
435
Environ Sci Technol 37(24):5479-5486.
436
8.
Watanabe N, Bergamaschi BA, Loftin KA, Meyer MT, & Harter T (2010) Use and
437
environmental occurrence of antibiotics in freestall dairy farms with manured forage
438
fields. Environ Sci Technol 44(17):6591-6600.
439
9.
Li Y, Wu X, Mo C, Tai Y, Huang X, & Xiang L (2011) Investigation of sulfonamide,
440
tetracycline, and quinolone antibiotics in vegetable farmland soil in the Pearl River Delta
441
area, southern China. J Agric Food Chem 59(13):7268-7276.
442
10.
Su J, Wei B, Ouyang W, Huang F, Zhao Y, Xu H, & Zhu Y (2015) Antibiotic resistome
443
and its association with bacterial communities during sewage sludge composting.
444
Environ Sci Technol 49(12):7356-7363.
445
22
11.
Karkman A, Do TT, Walsh F, & Virta MPJ (2018) Antibiotic-resistance genes in waste
446
water. Trends Microbiol 26(3):220-228.
447
12.
Qian X, Sun W, Gu J, Wang X-J, Sun J-J, Yin Y-N, & Duan M-L (2016) Variable effects
448
of oxytetracycline on antibiotic resistance gene abundance and the bacterial community
449
during aerobic composting of cow manure. J Hazard Mater 315:61-69.
450
13.
Zhu Y, Zhao Y, Li B, Huang C, Zhang S, Yu S, Chen Y, Zhang T, Gillings MR, & Su J
451
(2017) Continental-scale pollution of estuaries with antibiotic resistance genes. Nat
452
Microbiol 2:16270.
453
14.
Lazar V, Nagy I, Spohn R, Csorgo B, Gyorkei A, Nyerges A, Horvath B, Voros A, Busa-
454
Fekete R, Hrtyan M, Bogos B, Mehi O, Fekete G, Szappanos B, Kegl B, Papp B, & Pal C
455
(2014) Genome-wide analysis captures the determinants of the antibiotic cross-resistance
456
interaction network. Nat Commun 5:4352.
457
15.
Kohanski MA, DePristo MA, & Collins JJ (2010) Sublethal antibiotic treatment leads to
458
multidrug resistance via radical-induced mutagenesis. Mol Cell 37(3):311-320.
459
16.
Lv L, Jiang T, Zhang S, & Yu X (2014) Exposure to mutagenic disinfection byproducts
460
leads to increase of antibiotic resistance in Pseudomonas aeruginosa. Environ Sci
461
Technol 48(14):8188-8195.
462
17.
Li D, Zeng S, He M, & Gu AZ (2016) Water disinfection byproducts induce antibiotic
463
resistance-role of environmental pollutants in resistance phenomena. Environ Sci Technol
464
50(6):3193-3201.
465
18.
Baker-Austin C, Wright MS, Stepanauskas R, & McArthur JV (2006) Co-selection of
466
antibiotic and metal resistance. Trends Microbiol 14(4):176-182.
467
23
19.
Berg J, Thorsen MK, Holm PE, Jensen J, Nybroe O, & Brandt KK (2010) Cu exposure
468
under field conditions coselects for antibiotic resistance as determined by a novel
469
cultivation-independent bacterial community tolerance assay. Environ Sci Technol
470
44(22):8724-8728.
471
20.
Kasprzyk-Hordern B, Dinsdale RM, & Guwy AJ (2009) The removal of pharmaceuticals,
472
personal care products, endocrine disruptors and illicit drugs during wastewater treatment
473
and its impact on the quality of receiving waters. Water Res. 43(2):363-380.
474
21.
Kosma CI, Lambropoulou DA, & Albanis TA (2014) Investigation of PPCPs in
475
wastewater treatment plants in Greece: occurrence, removal and environmental risk
476
assessment. Sci Total Environ 466-467:421-438.
477
22.
Luo Y, Guo W, Ngo HH, Nghiem LD, Hai FI, Zhang J, Liang S, & Wang XC (2014) A
478
review on the occurrence of micropollutants in the aquatic environment and their fate and
479
removal during wastewater treatment. Sci Total Environ 473-474:619-641.
480
23.
Petrie B, Barden R, & Kasprzyk-Hordern B (2015) A review on emerging contaminants
481
in wastewaters and the environment: current knowledge, understudied areas and
482
recommendations for future monitoring. Water Res 72:3-27.
483
24.
Bradley PM, Journey CA, Romanok KM, Barber LB, Buxton HT, Foreman WT, Furlong
484
ET, Glassmeyer ST, Hladik ML, Iwanowicz LR, Jones DK, Kolpin DW, Kuivila KM,
485
Loftin KA, Mills MA, Meyer MT, Orlando JL, Reilly TJ, Smalling KL, & Villeneuve DL
486
(2017) Expanded target-chemical analysis reveals extensive mixed-organic-contaminant
487
exposure in U.S. streams. Environ Sci Technol 51(9):4792-4802.
488
25.
Harman-Fetcho JA, Hapeman CJ, McConnell LL, Potter TL, Rice CP, Sadeghi AM,
489
Smith RD, Bialek K, Sefton KA, Schaffer BA, & Curry R (2005) Pesticide occurrence in
490
24
selected South Florida canals and Biscayne Bay during high agricultural activity. J Agric
491
Food Chem 53(15):6040-6048.
492
26.
Bortoluzzi EC, Rheinheimer DS, Gonçalves CS, Pellegrini JBR, Maroneze AM, Kurz
493
MHS, Bacar NM, & Zanella R (2007) Investigation of the occurrence of pesticide
494
residues in rural wells and surface water following application to tobacco. Química Nova
495
30:1872-1876.
496
27.
Ccanccapa A, Masia A, Navarro-Ortega A, Pico Y, & Barcelo D (2016) Pesticides in the
497
Ebro River basin: occurrence and risk assessment. Environ Pollut 211:414-424.
498
28.
Robles-Molina J, Gilbert-López B, García-Reyes JF, & Molina-Díaz A (2014)
499
Monitoring of selected priority and emerging contaminants in the Guadalquivir River and
500
other related surface waters in the province of Jaén, South East Spain. Sci Total Environ
501
479:247-257.
502
29.
Wei R, Ge F, Huang S, Chen M, & Wang R (2011) Occurrence of veterinary antibiotics
503
in animal wastewater and surface water around farms in Jiangsu Province, China.
504
Chemosphere 82(10):1408-1414.
505
30.
McManus PS, Stockwell VO, Sundin GW, & Jones AL (2002) Antibiotic use in plant
506
agriculture. Annu Rev Phytopathol 40(1):443-465.
507
31.
McKenna M (2019) Antibiotics set to flood Florida's troubled orange orchards. Nature
508
567(7748):302.
509
32.
Heeb F, Singer H, Pernet-Coudrier B, Qi W, Liu H, Longrée P, Müller B, & Berg M
510
(2012) Organic micropollutants in rivers downstream of the megacity Beijing: sources
511
and mass fluxes in a large-scale wastewater irrigation system. Environ Sci Technol
512
46(16):8680-8688.
513
25
33.
Calderón-Preciado D, Matamoros V, & Bayona JM (2011) Occurrence and potential crop
514
uptake of emerging contaminants and related compounds in an agricultural irrigation
515
network. Sci Total Environ 412-413:14-19.
516
34.
Fairbairn DJ, Karpuzcu ME, Arnold WA, Barber BL, Kaufenberg EF, Koskinen WC,
517
Novak PJ, Rice PJ, & Swackhamer DL (2016) Sources and transport of contaminants of
518
emerging concern: a two-year study of occurrence and spatiotemporal variation in a
519
mixed land use watershed. Sci Total Environ 551:605-613.
520
35.
Schwarzenbach RP, Escher BI, Fenner K, Hofstetter TB, Johnson CA, von Gunten U, &
521
Wehrli B (2006) The challenge of micropollutants in aquatic systems. Science
522
313(5790):1072-1077.
523
36.
Xing Y, Yu Y, & Men Y (2018) Emerging investigators series: occurrence and fate of
524
emerging organic contaminants in wastewater treatment plants with an enhanced
525
nitrification step. Environ Sci Water Res Technol 4(10):1412-1426.
526
37.
Chang D-E, Smalley DJ, & Conway T (2002) Gene expression profiling of Escherichia
527
coli growth transitions: an expanded stringent response model. Mol Microbiol 45(2):289-
528
306.
529
38.
Denyer SP (1990) Mechanisms of action of biocides. Int Biodeterior 26(2):89-100.
530
39.
Heath RJ, Rubin JR, Holland DR, Zhang E, Snow ME, & Rock CO (1999) Mechanism of
531
triclosan inhibition of bacterial fatty acid synthesis. J Biol Chem 274(16):11110-11114.
532
40.
Newton BA (1956) The properties and mode of action of the polymyxins. Bacteriol Rev
533
20(1):14-27.
534
26
41.
Attmannspacher U, Scharf BE, & Harshey RM (2008) FliL is essential for swarming:
535
motor rotation in absence of FliL fractures the flagellar rod in swarmer cells of
536
Salmonella enterica. Mol Microbiol 68(2):328-341.
537
42.
Miryala SK & Ramaiah S (2018) Exploring the multi-drug resistance in Escherichia coli
538
O157:H7 by gene interaction network: a systems biology approach. Genomics:pii:
539
S0888-7543.
540
43.
Ma D, Alberti M, Lynch C, Nikaido H, & Hearst JE (1996) The local repressor AcrR
541
plays a modulating role in the regulation of acrAB genes of Escherichia coli by global
542
stress signals. Mol Microbiol 19(1):101-112.
543
44.
Okusu H, Ma D, & Nikaido H (1996) AcrAB efflux pump plays a major role in the
544
antibiotic resistance phenotype of Escherichia coli multiple-antibiotic-resistance (Mar)
545
mutants. J Bacteriol 178(1):306-308.
546
45.
Parkhill J, Dougan G, James KD, Thomson NR, Pickard D, Wain J, Churcher C, Mungall
547
KL, Bentley SD, Holden MTG, Sebaihia M, Baker S, Basham D, Brooks K,
548
Chillingworth T, Connerton P, Cronin A, Davis P, Davies RM, Dowd L, White N, Farrar
549
J, Feltwell T, Hamlin N, Haque A, Hien TT, Holroyd S, Jagels K, Krogh A, Larsen TS,
550
Leather S, Moule S, Ó'Gaora P, Parry C, Quail M, Rutherford K, Simmonds M, Skelton
551
J, Stevens K, Whitehead S, & Barrell BG (2001) Complete genome sequence of a
552
multiple drug resistant Salmonella enterica serovar Typhi CT18. Nature 413:848.
553
46.
Helling RB & Kukora JS (1971) Nalidixic acid-resistant mutants of Escherichia coli
554
deficient in isocitrate dehydrogenase. J Bacteriol 105(3):1224-1226.
555
47.
Oz T, Guvenek A, Yildiz S, Karaboga E, Tamer YT, Mumcuyan N, Ozan VB, Senturk
556
GH, Cokol M, Yeh P, & Toprak E (2014) Strength of selection pressure is an important
557
27
parameter contributing to the complexity of antibiotic resistance evolution. Mol Biol Evol
558
31(9):2387-2401.
559
48.
Jaffé A, Chabbert YA, & Derlot E (1983) Selection and characterization of beta-lactam-
560
resistant Escherichia coli K-12 mutants. Antimicrob Agents Chemother 23(4):622-625.
561
49.
Chopra AK, Sharma MK, & Chamoli S (2011) Bioaccumulation of organochlorine
562
pesticides in aquatic system—an overview. Environ Monit Assess 173(1):905-916.
563
50.
Kannan K, Kajiwara N, Watanabe M, Nakata H, Thomas NJ, Stephenson M, Jessup DA,
564
& Tanabe S (2004) Profiles of polychlorinated biphenyl congeners, organochlorine
565
pesticides, and butyltins in southern sea otters and their prey. Environ Toxicol Chem
566
23(1):49-56.
567
51.
Mohapatra DP, Cledon M, Brar SK, & Surampalli RY (2016) Application of wastewater
568
and biosolids in soil: occurrence and fate of emerging contaminants. Water Air Soil
569
Pollut 227(3):77.
570
52.
Wistrand-Yuen E, Knopp M, Hjort K, Koskiniemi S, Berg OG, & Andersson DI (2018)
571
Evolution of high-level resistance during low-level antibiotic exposure. Nat Commun
572
9(1):1599.
573
53.
Ozaki M, Mizushima S, & Nomura M (1969) Identification and functional
574
characterization of the protein controlled by the streptomycin-resistant locus in E. coli.
575
Nature 222(5191):333-339.
576
54.
Funatsu G & Wittmann HG (1972) Ribosomal proteins: XXXIII. Location of amino-acid
577
replacements in protein S12 isolated from Escherichia coli mutants resistant to
578
streptomycin. J Mol Biol 68(3):547-550.
579
28
55.
Wang X, Kim Y, Ma Q, Hong SH, Pokusaeva K, Sturino JM, & Wood TK (2010)
580
Cryptic prophages help bacteria cope with adverse environments. Nat Commun 1:147.
581
56.
Cox MP, Peterson DA, & Biggs PJ (2010) SolexaQA: at-a-glance quality assessment of
582
Illumina second-generation sequencing data. BMC Bioinformatics 11(1):485.
583
57.
Langmead B & Salzberg SL (2012) Fast gapped-read alignment with Bowtie 2. Nat
584
Methods 9:357.
585
58.
Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, Marth G, Abecasis G,
586
Durbin R, & 1000 Genome Project Data Processing Subgroup (2009) The sequence
587
alignment/map format and SAMtools. Bioinformatics 25(16):2078-2079.
588
59.
McKenna A, Hanna M, Banks E, Sivachenko A, Cibulskis K, Kernytsky A, Garimella K,
589
Altshuler D, Gabriel S, Daly M, & DePristo MA (2010) The genome analysis toolkit: a
590
mapreduce framework for analyzing next-generation DNA sequencing data. Genome Res
591
20(9):1297-1303.
592
60.
Men Y, Yu K, Bælum J, Gao Y, Tremblay J, Prestat E, Stenuit B, Tringe SG, Jansson J,
593
Zhang T, & Alvarez-Cohen L (2017) Metagenomic and metatranscriptomic analyses
594
reveal the structure and dynamics of a dechlorinating community containing
595
Dehalococcoides mccartyi and corrinoid-providing microorganisms under cobalamin-
596
limited conditions. Appl Environ Microbiol 83(8):e03508-03516.
597
61.
Liao Y, Smyth GK, & Shi W (2014) FeatureCounts: an efficient general purpose program
598
for assigning sequence reads to genomic features. Bioinformatics 30(7):923-930.
599
62.
Love MI, Huber W, & Anders S (2014) Moderated estimation of fold change and
600
dispersion for RNA-seq data with DESeq2. Genome Biol 15(12):550.
601
602
29
Figure legends:
603
Fig. 1. MICs of non-exposed mutants from G500 (A) and mutants from high-level pesticide
604
exposure (B). MIC0, Strep is the MIC (9 mg/L) of the ancestor strain. The p value of Mann-
605
Whitney U test is indicated (n = 36).
606
Fig. 2. MICs of resistant mutants selected from G500 E. coli under co-exposure (1EC pesticides
607
+ 1/5MIC0, Amp) (bottom-row; A: Amp-R mutants, C: Chl-R mutants, E: Cip-R mutants, and G:
608
Tet-R mutants) and single-exposure (1/5MIC0, Amp) (top-row; B: Amp-R mutants, D: Chl-R
609
mutants, F: Cip-R mutants, and H: Tet-R mutants). MIC0, Amp = 4 mg/L, MIC0, Chl = 8 mg/L,
610
MIC0, Cip = 0.016 mg/L, and MIC0, Tet = 1 mg/L. The p value of the Mann-Whitney U test
611
between Amp-exposure and co-exposure conditions is indicated (n = 36).
612
Fig. 3. Genetic mutations identified in resistant mutants (grey: reference genome, i.e., the
613
genome of G500 E. coli with no exposure; purple: genome of Strep-resistant mutants from high-
614
level pesticide exposure; blue: genome of Amp-resistant mutant from Amp-exposure; green:
615
genome of Amp-resistant mutant from co-exposure; yellow: genome of Cip-resistant mutants
616
from Amp-exposure; light pink: genome of Cip-resistant mutants from co-exposure. The links
617
represent genetic mutations (i.e., non-synonymous SNPs, insertions, or deletions). The capsules
618
stand for mutated genes and their positions on the genome. The darker the colors and the thicker
619
the links, the higher frequencies of the genetic mutations detected among the three resistant
620
mutants. Rings with dashed boarders contain the genes involved in a specific function).
621
Fig. 4. Differential gene expression analysis results, including principle component analysis of
622
gene transcripts in Amp-R (A) and Cip-R (B) mutants from Amp-exposure and co-exposure after
623
DESeq2 normalization, the heatmap of the relative abundance of differentially expressed genes
624
30
(C), and the bubble plot of the number of differentially expressed genes in terms of gene clusters
625
and gene functions (D).
626
31
Fig. 1. MICs of non-exposed mutants from G500 (A) and mutants from high-level pesticide
exposure (B). MIC0, Strep is the MIC (9 mg/L) of the ancestor strain. The p value of Mann-
Whitney U test is indicated (n = 36).
32
Fig. 2. MICs of resistant mutants selected from G500 E. coli under co-exposure (1EC pesticides
+ 1/5MIC0, Amp) (bottom-row; A: Amp-R mutants, C: Chl-R mutants, E: Cip-R mutants, and G:
Tet-R mutants) and single-exposure (1/5MIC0, Amp) (top-row; B: Amp-R mutants, D: Chl-R
mutants, F: Cip-R mutants, and H: Tet-R mutants). MIC0, Amp = 4 mg/L, MIC0, Chl = 8 mg/L,
MIC0, Cip = 0.016 mg/L, and MIC0, Tet = 1 mg/L. The p value of the Mann-Whitney U test
between Amp-exposure and co-exposure conditions is indicated (n = 36).
33
Fig. 3. Genetic mutations identified in resistant mutants (grey: reference genome, i.e., the
genome of G500 E. coli with no exposure; purple: genome of Strep-resistant mutants from high-
level pesticide exposure; blue: genome of Amp-resistant mutant from Amp-exposure; green:
genome of Amp-resistant mutant from co-exposure; yellow: genome of Cip-resistant mutants
from Amp-exposure; light pink: genome of Cip-resistant mutants from co-exposure. The links
represent genetic mutations (i.e., non-synonymous SNPs, insertions, or deletions). The capsules
stand for mutated genes and their positions on the genome. The darker the colors and the thicker
the links, the higher frequencies of the genetic mutations detected among the three resistant
mutants. Rings with dashed boarders contain the genes involved in a specific function).
34
Fig. 4. Differential gene expression analysis results, including principle component analysis of
gene transcripts in Amp-R (A) and Cip-R (B) mutants from Amp-exposure and co-exposure after
DESeq2 normalization, the heatmap of the relative abundance of differentially expressed genes
(C), and the bubble plot of the number of differentially expressed genes in terms of gene clusters
and gene functions (D).
35
Table 1. Selected genetic mutations identified in the resistant mutants.
Gene
Site
position
Nucleotide change
(SNP/INDEL)
Amino acid change
Samples
Gene annotation
rpsG
3473665
A → T
Stop gained (Leu157 → *)
Pesticide-expa, Strep-Rb-
1, 2, 3
30S ribosomal subunit protein S7
ftsI
93019
A → T
Gln536 → Leu
Co-exp, Amp-R-1, 2, 3
Peptidoglycan DD-transpeptidase
yagJ
292171
G → A
Val224 → Ile
Co-exp, Amp-R-1;
Co-exp, Cip-R-3
CP4-6 prophage
292177
A → C
Thr226 → Pro
Co-exp, Amp-R-1, 2, 3;
Co-exp, Cip-R-2, 3
292181
C → A
Ala227 → Asp
Co-exp, Amp-R-1, 2, 3;
Co-exp, Cip-R-2, 3
292186
A → C
Asn229 → His
Co-exp, Amp-R-1, 2, 3;
Co-exp, Cip-R-2, 3
292189
G → GATCTCATAT
Disruptive inframe insertion
(Ala230 → AspLeuIleSer)
Co-exp, Amp-R-1, 2, 3;
Co-exp, Cip-R-2, 3
292192
G → GGGACTTGTTC
Frameshift (Glu231 → fs)
Co-exp, Amp-R-1, 2, 3;
Co-exp, Cip-R-2
292193
A → G
Glu231 → Gly
Co-exp, Amp-R-1, 2, 3;
Co-exp, Cip-R-2, 3
292194
A → C
Glu231 → Asp
Co-exp, Amp-R-1, 2, 3;
Co-exp, Cip-R-2, 3
292200
A → C
Leu233 → Phe
Co-exp, Amp-R-1, 2, 3;
Co-exp, Cip-R-2, 3
292201
T → C
Phe234 → Leu
Co-exp, Amp-R-1, 2, 3;
Co-exp, Cip-R-2, 3
36
envZ
3535564
T → C
Thr120 → Ala
Amp-exp, Cip-R-1, 2
Sensory histidine kinase
gyrA
2339197
T → C
Asp87 → Gly
Amp-exp, Cip-R-3;
Co-exp, Cip-R-2
DNA gyrase subunit A
2339209
G → A
Ser83 → Leu
Co-exp, Cip-R-1, 3
nlpD
2867753
A → AT
Frameshift (Ile346 → fs)
Co-exp, Cip-R-1, 2, 3
Murein hydrolase activator
epmB
4375501
A → G
Val76 → Ala
Co-exp, Cip-R-1, 2, 3
Lysine 2, 3-aminomutase
maa
479493
A → G
Val143 → Ala
Co-exp, Cip-R-2, 3
Maltose O-acetyltransferase
sfmF
563407
A → G
Thr26 → Ala
Co-exp, Cip-R-2, 3
Putative fimbrial protein
ssuB
993767
A → G
Trp94 → Arg
Co-exp, Cip-R-2, 3
Aliphatic sulfonate ABC transporter
ATP binding subunit
dgcT
1093354
T → TG
Frameshift (Pro162 → fs)
Co-exp, Cip-R-2, 3
Putative diguanylate cyclase
ydcI
1495068
AT → T
Frameshift (Asn4 → fs)
Co-exp, Cip-R-2, 3
Putative DNA-binding
transcriptional repressor
ynfM
1670200
A → G
Tyr165 → Cys
Co-exp, Cip-R-2, 3
Putative transporter
btsS
2214369
GC → G
Frameshift (Gly104 → fs)
Co-exp, Cip-R-2, 3
High-affinity pyruvate receptor
hcaD
2672504
T → C
Leu141 → Pro
Co-exp, Cip-R-2, 3
Dioxygenase ferredoxin reductase
subunit
uvrA
4273523
A → G
Val139 → Ala
Co-exp, Cip-R-2, 3
Excision nuclease subunit A
a-exp: exposure; bR: resistant
| 2019 | Exposure to environmental level pesticides stimulates and diversifies evolution in towards greater antibiotic resistance | 10.1101/665273 | [
"Xing Yue",
"Wu Shuaiqi",
"Men Yujie"
] | creative-commons |
Comparing Task-Relevant Information Across Different Methods of Extracting
Functional Connectivity
Sophie Benitez Stulza,d, Andrea Insabatoa,b, Gustavo Decoa,c, Matthieu Gilsona, Mario
Sendend,e
a Center for Brain and Cognition, Computational Neuroscience Group, Department of
Information and Communication Technologies, Universitat Pompeu Fabra, Carrer de Ramon
Trias Fargas, 25-27, Barcelona, 08005, Spain
b The Italian Academy, Center for Theoretical Neuroscience, Columbia University, 1161
Amsterdam Ave., New York NY 10027, USA
c Institució Catalana de la Recerca i Estudis Avançats (ICREA), Universitat Pompeu Fabra,
Passeig Lluís Companys 23, Barcelona 08010, Spain
d Department of Cognitive Neuroscience, Faculty of Psychology and Neuroscience,
Maastricht University, 6201BC Maastricht, The Netherlands
e Maastricht Brain Imaging Centre, Faculty of Psychology and Neuroscience, Maastricht
University, P.O. Box 616, 6200 MD Maastricht, The Netherlands
Abstract:
186 words
Main text:
4176 words
References: 36
Abstract
The concept of brain states, functionally relevant large-scale patterns, has become popular in
neuroimaging. Not all components of such patterns are equally characteristic for each brain
state, but machine learning provides a possibility of extracting the structure of brain states from
functional data. However, the characterization in terms of functional connectivity measures
varies widely, from cross-correlation to phase coherence, and the idea that different measures
will provide the similar information is a common assumption made in neuroimaging. Here, we
compare the performance of phase coherence, pairwise covariance, correlation, model-based
covariance and model-based precision for a dataset of subjects performing five different
cognitive tasks. We employ multinomial logistic regression for classification and consider two
types of cross-validation schemes, between- and within-subjects. Furthermore, we investigate
whether classification is robust for different temporal window lengths. We find that informative
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
1
links for the classification, meaning changes between tasks that are consistent across subjects,
are entirely uncorrelated between correlation and covariance. These results indicate that the
corresponding FC signature can strongly differ across FC methods used and that interpretation
is subject to caution in terms of subnetworks related to a task.
Keywords: machine learning, functional connectivity, fMRI, task information, brain states
1. Introduction
At a macroscopic level the brain may be conceived of as a complex system of regions
engaging in dynamic, interactive behaviour (Bullmore & Sporns, 2009). Neuroscience has
developed various quantitative approaches to define stereotypical brain states corresponding to
cognitive functions. Brain states may refer to purely spatial patterns, activity distribution across
voxels or brain regions (Cabral, Kringelbach, & Deco, 2017). Alternatively, they may refer to
spatio-temporal patterns and distributions functional interactions between regions (Vidaurre,
Smith, & Woolrich, 2017).
Whole-brain modelling has been widely used to characterise spatio-temporal brain states and
capture their multivariate distributions. This approach attempts to explain observed functional
interaction in terms of models of underlying region dynamics as well as structural connections
between regions. Modelling of the oscillatory behaviour in brain regions has, for instance,
shown that there are differences in this local parameter across task-dependent brain states
(Senden, Reuter, van den Heuvel, Goebel, & Deco, 2017). On the other hand, models estimating
directed connectivity based on the functional interactions between brain regions have also
revealed differences in network parameters across task-dependent brain states (Pallares et al.,
2018; Senden et al., 2018).
Recently, the application of machine learning to infer brain states has also gained popularity
(Naselaris, Kay, Nishimoto, & Gallant, 2011; Pallares et al., 2018; Rahim, Thirion, Bzdok,
Buvat, & Varoquaux, 2017; Varoquaux et al., 2017; Xie et al., 2017). Machine learning is useful
since it can extract the relevant feature patterns of brain states from multivariate data and assess
the generalization capabilities of these brain states to novel data. This approach has been highly
successful for inferring brain states from functional connectivity (FC). Conventionally,
functional connectivity (FC) is calculated across the entire duration of a session. Recently,
however, focus has shifted towards dynamic functional connectivity (dFC) which is calculated
at shorter time scales in the range of tens of seconds (Gonzalez-Castillo et al., 2015; Hutchison
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
2
et al., 2013; Preti, Bolton, & Van De Ville, 2017). For example dFC can be calculated with the
sliding-window approach (Cabral, Kringelbach, et al., 2017), where Pearson correlation or
covariance is computed between the signals of every pair of region with a small temporal
window moving along the time series. A studies using the sliding window concept of dFC could
successfully distinguish between the brain states during five different cognitive tasks
(Gonzalez-Castillo et al., 2015; Xie et al., 2017). At the opposite end of the spectrum of time-
scales, FC can be obtained instantaneously with phase coherence (Cabral, Vidaurre, et al., 2017;
Senden et al., 2017). Evidently, there is a multitude of studies using various FC metrics to
investigate brain states during different tasks (Cabral, Vidaurre, et al., 2017; Gonzalez-Castillo
& Bandettini, 2017; Senden et al., 2018, 2017). However, the interchangeable use of FC metrics
rests on the assumption that the results are comparable across metrics. This has not been
validated since varying methodologies make it impossible to compare them across studies.
Our aim is test this assumption and to systematically evaluate the task-relevant information
structure of the corresponding brain states across metrics and time-scale. The tasks include rest,
a n-Back task, the Flanker task, a mental rotation task, and an Odd-man-out task (Senden et al.,
2018, 2017). Specifically, we want to investigate whether choice in FC metric (Pearson
correlation, covariance, phase coherence) affects classification performance and whether task-
dependent information is similar across metrics. Secondly, we investigate metrics across
different time scales, because it is possible that certain time scales do not capture information
relevant to the classification, which would not be an issue of the metric itself, but of the
parameter choice for its temporal window. Also, including metrics that reach from
instantaneous FC (phase coherence) until static FC (global FC) provides a broad systematic
overview of the temporal spectrum.
We find that the choice of parameters and metrics for connectivity classification strongly impact
the task-relevant information retrieved and call for a more careful approach towards the
interpretation of such results.
2. Material and methods
2.1 Functional MRI Data
We use an fMRI resting and task state dataset acquired in 14 subjects (8 females, M = 28.76,
22 – 43 years old) as described in a previous paper (Senden et al., 2017). The dataset comprised
the blood-oxygen-level dependent (BOLD) signal of 68 Regions of Interest (ROIs) obtained in
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
3
five functional runs per subject with 192 data points each. During each run, the subjects were
either resting, or engaging in one of four tasks: the Eriksen flanker task (Eriksen & Eriksen,
1974), a n-Back task (Kirchner, 1958), a mental rotation task (Shepard & Metzler, 1971), or a
verbal Odd-man-out task (Flowers & Robertson, 1985). A detailed description of the stimuli
used in the task paradigm can be found in Senden et al. (2017). The dataset was acquired at the
Maastricht Brain Imaging Centre, (Maastricht University) on a 3 Tesla scanner (Tim
Trio/upgraded to Prisma Fit, Siemens Healthcare, Germany). The data was pre-processed with
BrainVoyager QX (v2.6; Brain Innovation, Maastricht, the Netherlands) using slice scan time
correction, motion correction, and a high-pass filter with a frequency cut-off of .01 Hz. After
subsequent wavelet de-spiking and regressing out global noise signals estimated from the
ventricles, the average BOLD signal for each region was computed by taking the mean of voxels
uniquely belonging to one of the 68 ROIs specified by the DK atlas (Desikan et al., 2006) with
Matlab (2013a, The MathWorks, Natick, MA).
2.2 Spatiotemporal functional connectivity
2.2.1 Phase Coherence. To obtain the analytical signal (Smith, 2007), a complex-
valued function that has no negative frequency components, from the BOLD signal the Hilbret
transformation was applied to the BOLD signal for each ROI. To calculate the instantaneous
Figure 1. Extracting FC from Bold signal. (A) Bold signal of 68 ROIs for 384 s of a fMRI session. Dots indicating omitted BOLD
timeseries for visibility purposes. (B) FC matrices extracted from the BOLD signal in window with window length (WL). To eliminate
identical values a mask is applied and (ROI*(ROI-1))/2 = 2278 features are obtained for each timepoint t. Subsequent timepoints are
shifted by time step (Δt). (C) Table of FC types calculated from the Bold signal.
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
4
functional connectivity (iFC) between region i and j for time t the cosine of the phase difference
of the analytical signal of the two regions, was calculated.
𝑖𝐹𝐶(𝑖, 𝑗, 𝑡) = cos(𝜃(𝑖, 𝑡) − 𝜃(𝑗, 𝑡))
2.2.1.1 Eigenvector. To obtain the connectivity among eigenvectors we calculated the
outer product of the strongest eigenvector of iFC as previously described in Cabral, Vidaurre,
et al. (2017).
𝑒𝑖𝑔𝐹𝐶(𝑖, 𝑗, 𝑡) = 𝑒𝑖𝑔(𝑖𝐹𝐶(𝑖, 𝑡)) ⊗ 𝑒𝑖𝑔(𝑖𝐹𝐶(𝑗, 𝑡))
where,
𝑖𝐹𝐶(𝑡) = instantaneous FC at timepoint t.
𝑒𝑖𝑔 = largest eigenvector.
2.2.2 Covariance. The dynamic covariance (dCov) was calculated across window
lengths of 20 s, 40 s, 60 s, 80 s, 100 s, 120s with a timestep of 4 s. We also computed pairwise
Cov over the whole session to obtain global functional connectivity (gCov). Dynamic
covariance between region n and p for time window t was calculated as follows:
𝐶𝑜𝑣(𝑖, 𝑗, 𝑤) =(𝑋(𝑖, 𝑤)−𝑋(𝑖)
̅̅̅̅̅̅)∗(𝑋(𝑗, 𝑤)−𝑋(𝑗)
̅̅̅̅̅̅)
where,
𝑋(𝑘, 𝑤) = BOLD in region k in time window w.
𝑋(𝑘)
̅̅̅̅̅̅ = Mean BOLD in region k.
2.2.3 Pearson’s Correlation. Dynamic pairwise Pearson correlation (dPC) was
calculated with windows of 20 s, 40 s, 60 s, 80 s, 100 s, 120 s, and with a timestep of 4 s as well
as within 6 s window with a timestep of 2 s to make the timescale of the PC as similar as
possible to the timescale of the Hilbert transform. We also computed pairwise PC over the
whole session to obtain global functional connectivity (gPC). Dynamic Pearson correlation
between region i and j for time window w was calculated as follows:
𝐶𝑜𝑟𝑟(𝑖, 𝑗, 𝑡) =
(𝑋(𝑖, 𝑤)−𝑋(𝑖)
̅̅̅̅̅̅)∗(𝑋(𝑗, 𝑤)−𝑋(𝑗)
̅̅̅̅̅̅)
√(𝑋(𝑖, 𝑤)−𝑋(𝑖)
̅̅̅̅̅̅)
2∗ (𝑋(𝑖, 𝑤)−𝑋(𝑗)
̅̅̅̅̅̅)
2
where,
𝑋(𝑘, 𝑤) = BOLD in region k in time window w.
𝑋(𝑘)
̅̅̅̅̅̅ = Mean BOLD in region k.
2.2.4 Model-based Precision and Covariance. The model-based precision and
covariance (Scikit-learn, GraphLassoCV) attempts to estimate the inverse of the covariance
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
5
matrix, the precision matrix, which is proportional to the partial correlation matrix. The
empirical precision matrix is not included as the covariance matrix is underdetermined,
meaning it has less timepoints than regions in short time windows, and could not be
calculated. The GraphLasso algorithm achieves this by enforcing sparsity on the estimation of
the precision matrix by using an L1 penalty which is automatically estimated with cross-
validation. More specifically, the GraphLasso algorithm (Friedman, Hastie, & Tibshirani,
2008) minimizes the following function to estimate the precision matrix K and the
corresponding covariance matrix S.
𝐾̂ = 𝑎𝑟𝑔𝐾𝑚𝑖𝑛(𝑡𝑟𝑆𝐾 − 𝑙𝑜𝑔𝑑𝑒𝑡𝐾 + 𝛼||𝐾||1)
where,
𝐾 =precision matrix to be estimated.
𝑆 = sample covariance matrix.
||𝐾||1 =sum of absolute values of off-diagonal coefficients of K.
𝛼 =L1 penalty parameter.
2.3 Classification
2.3.1 Multinomial logistic regression. We use multinomial logistic regression (MLR) with
a cross-entropy loss. We use an L2 penalization in combination with a limited-memory
Broyden-Fletcher-Goldfarb-Shannon algorithm solver (Bishop, 2006) and an L1 penalty with
a SAGA algorithm solver (Defazio, Bach, & Lacoste-Julien, 2014). The SAGA algorithm is an
incremental gradient method which supports non-strongly convex problems. The penalty
parameter is optimized with nested cross-validation meaning that the parameters are first
optimized using cross-validation within the training set before being applied to the entire
training set.
2.3.2 Cross-validation.
2.3.2.1 Within Subject. Due to temporal autocorrelation simple permutation does not give us
any indication of the stability of the signal within a subject over time. Therefore, we use blocked
cross-validation. For each task and subject, the samples are divided in 10 consecutive folds.
The number of samples contained in each fold depends on the metric chosen. Subsequently, the
decoder is trained on the first fold and tested on the second fold. Then the decoder is trained on
the first and second fold and tested on the third fold. This procedure is continued until the last
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
6
fold is reached. The accuracy of the validation procedure is obtained from the mean of the
testing accuracy over the 10 trained decoders.
The penalty parameter is optimized using nested cross-validation. More specifically,
parameters for each training set are optimized with 2-fold temporal cross-validation on the
training set (see figure 3).
2.3.2.2 Between Subject. The decoder is trained on 13 of the 14 subjects and tested on the
remaining subject. This procedure is repeated with each subject being left out once. The
accuracy of the validation procedure is the mean of the testing accuracy over the 14 trained
decoders.
The penalty parameter is optimized using nested cross-validation. More specifically,
parameters for each training set are optimized with 13-fold subject cross-validation on the
training set (see figure 3).
2.3.3 Recursive feature elimination. Recursive feature elimination (RFE) iteratively
removes the feature that is least important for classification. Features leading to a maximal
accuracy using temporal and subject cross-validation are then deemed the best features to use
for the classification. The ranking of all features obtained by the RFE is indicative of the
structure of the information obtained from each FC. The number of best features was also
selected within the nested cross-validation before optimizing the penalty parameter.
The classification pipeline was implemented in python using the Scikit-learn library
(Pedregosa et al., 2011).
2.4 Similarity Measures
Spearman Rank. The Spearman Rank correlation 𝑟𝑠 is a measure of non-linear
correlation with a value between -1, denoting perfect anti-correlation, and 1, denoting perfect
correlation (Lehman & Rourke, 2005). It quantifies how well the relationship between two
variables can be expressed with a monotonic function.
𝑟𝑠 =𝑐𝑜𝑣(𝑟𝑔𝑋, 𝑟𝑔𝑌)
𝜎𝑟𝑔𝑋 ∗ 𝜎𝑟𝑔𝑌
where,
𝑟𝑔𝑋, 𝑟𝑔𝑌 =Ranks of variables X, Y.
𝑐𝑜𝑣(𝑟𝑔𝑋, 𝑟𝑔𝑌) =Covariance of the rank variables.
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
7
𝜎𝑟𝑔𝑋, 𝜎𝑟𝑔𝑌 = Standard deviation of the rank variables.
3. Results
3.1 Performance of the FC metrics
3.1.1 Covariance
Within subject cross-validation accuracy of covariance follows a monotonically increasing
trend starting from a window length of 20 s and saturates after 80 s (figure 3B). The necessity
of within-subject CV to quantify the temporal stability of the classes becomes clear when
compared to cross-validation with permutation sets which disregard the temporal
autocorrelation (S2). While the permutation CV achieves maximal accuracy for all window
lengths, within-subject CV shows a break-down of temporal stability which has also been
Figure 3: Within- and between-subject cross-validation procedure. Covariance outperforms correlation. (A) Within-subject and between-
subject cross-validation. In within-subject cross-validation the data is split in sections along time. (B) Between-subject CV accuracy of
covariance and correlation. (C) Between-subject CV accuracy of covariance and correlation.
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
8
shown by other studies (Roberts et al., 2017). Adding variance to covariance only improves
accuracy for a window length of 80 s but decreases on average by approximately 5% (S1A).
Global covariance achieves a slightly higher accuracy with 0.8 (figure 3B).
Between subject cross-validation accuracy of covariance increases from a window length of
20s and saturates at 100s with a dip at 80 s (figure 3C). The performance seems to follow a
growing trend (excluding 80 s) reaching a maximum at a window length of 100 s with an
accuracy of 77 % and decreasing thereafter. Global covariance achieves a similar accuracy as
dynamic covariance with a window length of 60 s.
3.1.2 Pearson correlation
Within-subject performance of the Pearson correlation increases from a window length of 20 s
and saturates after 80 s (figure 3B). Global correlation performance is 0.8.
Between-subject cross-validation accuracy of correlation follows a monotonically increasing
trend from a window length of 20 s until 120 s. Global Pearson correlation accuracy is ~15%
higher than performance of dPC with a window length of 120 s. The different trends observed
in empirical covariance and correlation suggests, that they are affected by the noise in the data
differently. At windows until 120 s covariance generally performs better possibly because the
standardization in the Pearson correlation also removes information in the variance at shorter
time scales. At longer time-scales the variance likely contained more noise and the removal
increases performance.
3.1.3 GraphLasso Precision
Figure 4: Performance of model-based FC measures GraphLasso covariance and precision. (A) Within-subject cross-validation accuracy
using GraphLasso covariance (green) and GraphLasso precision (pink). (B) Between-subject cross-validation accuracy using
GraphLasso covariance (green) and GraphLasso precision (pink). Chance level is 0.2.
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
9
Within-subject performance of the GraphLasso Precision follows an asymptotic trend towards
the maximal accuracy increasing from a window length of 20 s and reaching maximal accuracy
at a window length of 100 s (figure 4A).
Between-subject cross-validation accuracy of GraphLasso precision shows a growing
monotonical trend continually increasing from a window length of 20 s without saturating.
Similar to empirical covariance, model-based covariance does not improve at longer window
lengths, suggesting, that it might be affected by noisy lower frequency fluctuations.
Interestingly, removing noisy fluctuation by estimating the underlying precision performs much
better than standardizing it with the variance like in the Pearson correlation.
3.1.4 GraphLasso Covariance
Within-subject performance of the GraphLasso covariance increases from a window length of
20 s and reaches approximately maximal accuracy at a window length of 100 s (figure 4A).
Model-based as well as empirical metrics follow a similar asymptotic trend towards maximal
accuracy, suggesting that they are affected by similar noisy temporal fluctuation at shorter time-
scales.
Between-subject cross-validation accuracy of GraphLasso covariance continually increases
from a window length of 20 s reaching a maximum at 60 s and decreasing again until 120 s.
3.1.5 Phase coherence.
Figure 2: CV Accuracy at short time-scales. (A) Within-subject CV accuracy of the BOLD timeseries, phase coherence (iFC), the largest
eigenvector of the phase coherence (iFC), Pearson Correlation with a window length of 6s and a time step of 2 s (dPC 6s) and Pearson
Correlation with a window length or 20s and a time step of 4 s (dPC 20s). (B) Between-subject CV accuracy of the BOLD timeseries, phase
coherence (iFC), the largest eigenvector of the phase coherence (iFC), Pearson Correlation with a window length of 6s and a time step of 2 s
(dPC 6s) and Pearson Correlation with a window length or 20s and a time step of 4 s (dPC 20s). Chance level (0.2) indicated with black line.
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
10
Phase coherence performed poorly for both within- and between-subject CV. The median of
the within-subject performance for phase coherence was 0.42 with chance level at 0.2 (figure
2A). The largest eigenvector of phase coherence only scored slightly above 0.32. The median
of the between-subject performance of phase coherence was 0.33 and for the largest eigenvector
was 0.27 (figure 2B).
The BOLD signal does not seem to carry any information to distinguish among tasks and using
the eigenvector of the phase coherence leads to a decrease in accuracy and likely does not select
relevant axes of the variability. Interestingly, the Pearson Correlation with a similar time-step
as phase coherence and window of only 6 s did not outperform phase coherence.
3.2 Regularisation methods
Regularization is a commonly used tool to prevent a classifier from overfitting the training set
leading to low testing accuracy (Bishop, 2006). However, L2 regularization did not reduce
overfitting adequately as training accuracy was up to 50% higher than testing accuracy (see
table S4).
Figure 5. Feature selection performance for covariance. (A) The recurrent feature ranking (RFE) is used to test how many best features give
the highest within-subject CV and between-subject CV accuracy. (B) Within-subject CV and (C) Between-subject CV accuracy of all features
versus best features with covariance.
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
11
Using L1 regularization instead of L2 regularization in our classification did not improve the
performance of the classifier. Rather it reduced accuracies by approximately 3% on average
(see S2). Another tool that can be used to reduce dimensionality additionally is feature selection.
However, this did not lead to a significant increase in within- or between-subject CV accuracy
(figure 5A – C).
3.3 Task and rest are highly dissimilar
The strong decrease of accuracy towards smaller time-scales may be predominantly due to the
difficulty of differentiating among tasks rather than discriminating task states from rest. Here
we test this possibility by plotting the silhouette scores of phase coherence and covariance of
the axes along which activity is most different between tasks, extracted with Linear
Discriminant Analysis. Silhouette scores quantify if an observation (black dot) is closer to the
distribution of its own class (black 2) or to the distribution of another class (green 1) as shown
in figure 6A. If the observations are strongly clustered the silhouette score is high, whereas it
decreases if the classes are more overlapping such as in the example given in figure 6A. Figure
6B shows that at smaller time-scales task samples have significantly lower silhouette scores,
meaning that they are more similar to other classes as opposed to their own, whereas rest is
more similar to itself than other classes. With increasing window length the silhouette scores
increase, but the difference between rest and tasks remains except for the time window of 80 s.
Figure 6. Silhouette scores of linear discriminant analysis (LDA) of rest and task. Features were reduced to four components with LDA and
the silhouette score was calculated. (A) The first LDA component of covariance with window length 40 s is plotted on the x-axis and the
second LDA component is plotted on the y-axis. Rest is plotted in black and task is plotted in green. (B) Violinplot of silhouette scores of
LDA of rest and task for various FC Metrics. The metrics used were phase coherence (-), and covariance corresponding to the window lengths
on the x axis. Rest is plotted in green and task is plotted in pink. Black bars indicate the inner 50 percentiles. The white dot indicated the
median. A y-score of 0 indicates no clustering whereas 1 indicates strong clustering. Significance level indicated with symbols, p < 0.0001
(****), p < 0.001 (***), p < 0.01 (**), p < 0.05 (*), and non-significance (-).
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
12
3.2 The structure of task-relevant information differs strongly across time-scale and method of
FC extraction.
To evaluate the distribution of information structure across various FC methods we perform
recursive feature elimination for
each method and compare the
resulting
rankings
using
Spearman
rank
correlation
(figure 7). The model-based
metrics
(precision
and
covariance) as well as the
empirical
metrics
(Pearson
correlation
and
covariance)
display a similar decrease in
similarity across time scale.
GraphLasso
precision
and
covariance also retain most
similarity at similar time scales.
This pattern is also present for
GraphLasso covariance and empirical covariance, but not for GraphLasso precision and
empirical covariance. Most importantly, the feature ranking of covariance (as well as
covariance-based metrics) and correlation are not correlated at any time-scale, suggesting that
the task-relevant information structure retrieved by these two methods is very dissimilar. With
covariance and Pearson correlation the task-relevant information structure becomes more
dissimilar with increasing difference in window length. At the shortest time-scales, feature
rankings obtained from iFC are slightly correlated with Pearson correlation metrics and eigFC
are slightly correlated with covariance metrics. Instantaneous FC and eigFC do not seem to be
correlated. The decreasing correlation with size of time window suggests task-relevant
information content also differs across time-scale. Although the concurrent decrease in
accuracy for shorter time-scales might also indicate that sufficiently long window lengths are
necessary for a stable estimate for covariance or correlation.
Figure 7. Task-dependent information structure can differ strongly across metric.
Spearman rank correlation of all FC metrics.
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
13
4. Discussion
The aim of this paper was to evaluate if brain states can be classified with FC in a
systematic manner and whether the extracted brain states are influenced by the choice of FC
metric (phase coherence, Pearson correlation, covariance, GraphLasso precision, and
GraphLasso covariance). Among empirical measure covariance outperformed correlation under
certain conditions, in this five-task classification. Adding variance to covariance did not further
increase accuracy. GraphLasso precision outperformed all empirical measures and was only
outperformed by GraphLasso covariance for a window length of 20 s. Within-subject cross-
validation accuracy was generally higher than between-subject cross-validation and can be
conceptualized as an upper limit on accuracy. Another possibility is that more subjects are
needed for between-subject cross-validation as suggested in a study by Abraham & al. (2017).
They also found that accuracy increased with higher parcellation. Within- and between-subject
cross-validation accuracy increased in proportion with the time-scale, which is likely due to
high-frequency noise in the signal which is more likely to affect short time-scales (Cabral,
Kringelbach, et al., 2017; Hutchison et al., 2013). An alternative explanation for the low
accuracy at shorter time-scales is low task performance (Gonzalez-Castillo et al., 2015).
Gonzalez-Castillo et al. (2015) showed that large deviations in task performance are correlated
with substantial errors in classification accuracy. These deviations are more likely to bias
connectivity measures at shorter time-scales. However, we did not control for this possibility.
A third explanation could be that stable classification depends on specific frequency bands
which would require window lengths long enough to capture these functional interactions. A
study investigating the dependence of community structure on window length has already
shown that different frequency bands can address distinct neuronal processes (Telesford et al.,
2016). Specific neuronal processes could be better captured by models aimed at specific
frequency bands such as dynamic causal modelling or the Kuramoto model (Cabral, Hugues,
Sporns, & Deco, 2011; Friston, Kahan, Biswal, & Razi, 2014).
Accuracy at shorter time-scales was low for testing data, it was high for training data.
This finding highlights that proper cross-validation is necessary to draw conclusions regarding
classification performance since the data tends to get overfit. This is critical, since a high
accuracy of the classifier on the training set is necessary, but not sufficient for high accuracy
on the novel testing set. For example, in the study by Xie et al. (2017) the performance of the
trained classifier was not validated with a novel dataset. Such validation would have been
informative of whether the learned parameters can distinguish the brain states due to true
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
14
differences that hold at a population level or due to noise (Varoquaux et al., 2017). The problem
of overfitting can generally be addressed by feature selection or regularization. Here neither
feature selection nor L1 penalty regularization lead to an increase in accuracy for between- or
within-subject CV. While feature selection eliminates features, the L1 penalty forces their
weights to 0 indicating that the task-relevant signatures may be more distributed, because the
classification improves of no features are discarded. Note that we did not perform an exhaustive
search of the parameter space for the optimal combinations of feature number and L2 penalty
parameter. Instead, we searched the parameter space serially: We optimized the feature number
and then optimized the penalty parameter.
The strong decrease of accuracy at shorter time-scales was primarily driven by the difficulty of
distinguishing tasks from each other rather than distinguish task states from rest. This suggests
that the brain at rest is very dissimilar to the brain engaging in a task. This is in line with
previous studies using whole-brain modelling (Ponce-Alvarez, He, Hagmann, & Deco, 2015;
Senden et al., 2018, 2017). However, it could be argued that this stems from the fact that the
stimuli used here were all visual, making the classification entirely reliant on non-sensory
processes. It is, therefore, quite possible for other classification problems to reach better
accuracies at smaller time-scales and with different FC methods. Another limiting factor could
be the context-dependence of the features used in the multinomial classification. A feature can
be crucial for distinguishing between task A and B, but not between task A and C. If the
classification problem only includes tasks A and C the task-relevant information structure that
is extracted by the classifier changes depending on the tasks included.
Task-relevant features that are crucially depends on which tasks are included in the
classification. le specific functional interactions might be relevant in a pairwise discrimination
between two tasks, they could become irrelevant in a multinomial discrimination depending on
the tasks among which the classifier is discriminating.
The most important finding, however, is that the task-relevant information structure differs
strongly not only across time-scale, but also across connectivity measures. The absence of any
similarity in information structure retrieved from correlation and covariance is a
counterintuitive and problematic result. Correlation is merely normalized covariance and
evidence that such closely related methods can provide very different information contradicts
the implicit assumption that similar methods should lead to similar conclusions. That this is not
the case is problematic for the interpretation of any results obtained for different measures and
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
15
time-scales since there is no ground-truth on task-relevant functional interactions. For example,
how would one interpret evidence from studies using network theory to detect communities
based on different FC methods (Fuertinger & Simonyan, 2016; Najafi, Mcmenamin, Simon, &
Pessoa, 2016; Sporns, 2013)? This underlines the need for alternative, better defined metrics
such as model-based FC, where the relationships between the various metrics are better defined
(Cabral et al., 2011; Friston et al., 2014; Pallares et al., 2018; Senden et al., 2018, 2017).
However, the optimal metric may still strongly depend on the classification problem itself.
Consequentially, this will impact the research design, for example when attempting to classify
switching trials. Here, the task intervals have to be long enough for windows to only contain a
single task.
In conclusion, the following suggestions can be made for classification in neuroscience.
(1) When one is interested in groups and wished to obtain results which generalize to new
subjects, accuracy model-based FC metrics should be used and precision should be preferred
except for window lengths around 20 s. (2) When one is interested in individual subjects,
empirical covariance should be preferred for classification. (3) Generally, larger window
lengths should be preferred. (4) For MLR classifiers, L2 regularization should be preferred.
The pipeline developed here can be applied to other neuroimaging tools as well such as
electroencephalography (EEG) or functional near-infrared spectroscopy (fNIRS). Quantifying
the performance of a classifier is furthermore especially important in clinical settings when
aiming to identify pathological brain states in new patients. Predictive decoders, for example in
the case of brain-computer interfaces, can be implemented with FC metrics, but should be tuned
within-subject as the performance is better and more stable. The main result of this study,
namely, the dissimilarity of information-structure across FC methods, calls for greater care in
the selection of FC method with respect to the aim of a study as well as a more careful
interpretation of results in neuroscience using different FC methods in the future.
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
16
References
Abraham, A., Milham, M. P., Di Martino, A., Craddock, R. C., Samaras, D., Thirion, B., &
Varoquaux, G. (2017). Deriving reproducible biomarkers from multi-site resting-state
data: An Autism-based example. NeuroImage, 147, 736–745.
https://doi.org/10.1016/J.NEUROIMAGE.2016.10.045
Bishop, C. M. (2006). Pattern Recognition And Machine Learning. Springer.
https://doi.org/10.1117/1.2819119
Bullmore, E., & Sporns, O. (2009). Complex brain networks: graph theoretical analysis of
structural and functional systems. Nature Reviews Neuroscience, 10(4), 312–312.
https://doi.org/10.1038/nrn2618
Cabral, J., Hugues, E., Sporns, O., & Deco, G. (2011). Role of local network oscillations in
resting-state functional connectivity. NeuroImage, 57(1), 130–139.
https://doi.org/10.1016/j.neuroimage.2011.04.010
Cabral, J., Kringelbach, M. L., & Deco, G. (2017). Functional connectivity dynamically
evolves on multiple time-scales over a static structural connectome: Models and
mechanisms. NeuroImage, (March). https://doi.org/10.1016/j.neuroimage.2017.03.045
Cabral, J., Vidaurre, D., Marques, P., Magalhães, R., Silva Moreira, P., Miguel Soares, J., …
Kringelbach, M. L. (2017). Cognitive performance in healthy older adults relates to
spontaneous switching between states of functional connectivity during rest. Scientific
Reports, 7(1), 5135. https://doi.org/10.1038/s41598-017-05425-7
Defazio, A., Bach, F., & Lacoste-Julien, S. (2014). SAGA: A Fast Incremental Gradient
Method With Support for Non-Strongly Convex Composite Objectives. Retrieved from
http://arxiv.org/abs/1407.0202
Desikan, R. S., Ségonne, F., Fischl, B., Quinn, B. T., Dickerson, B. C., Blacker, D., …
Killiany, R. J. (2006). An automated labeling system for subdividing the human cerebral
cortex on MRI scans into gyral based regions of interest.
https://doi.org/10.1016/j.neuroimage.2006.01.021
Eriksen, B. A., & Eriksen, C. W. (1974). Effects of noise letters upon the identification of a
target letter in a nonsearch task. Perception & Psychophysics, 16(1), 143–149.
https://doi.org/10.3758/BF03203267
Flowers, K., & Robertson, C. (1985). The effect of Parkinson’s disease on the ability to
maintain a mental set. Journal of Neurology Neurosurgery, and Psychiatry, 48, 517–529.
https://doi.org/10.1136/jnnp.48.6.517
Friedman, J., Hastie, T., & Tibshirani, R. (2008). Sparse inverse covariance estimation with
the graphical lasso. Biostatistics, 9(3), 432–441.
https://doi.org/10.1093/biostatistics/kxm045
Friston, K. J., Kahan, J., Biswal, B., & Razi, A. (2014). A DCM for resting state fMRI.
NeuroImage, 94(100), 396–407. https://doi.org/10.1016/j.neuroimage.2013.12.009
Fuertinger, S., & Simonyan, K. (2016). Stability of Network Communities as a Function of
Task Complexity. Journal of Cognitive Neuroscience, 28(12), 2030–2043.
https://doi.org/10.1162/jocn_a_01026
Gonzalez-Castillo, J., & Bandettini, P. A. (2017). Task-based dynamic functional
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
17
connectivity: Recent findings and open questions. NeuroImage.
https://doi.org/10.1016/j.neuroimage.2017.08.006
Gonzalez-Castillo, J., Hoy, C. W., Handwerker, D. A., Robinson, M. E., Buchanan, L. C.,
Saad, Z. S., & Bandettini, P. A. (2015). Tracking ongoing cognition in individuals using
brief, whole-brain functional connectivity patterns. Proceedings of the National
Academy of Sciences, 112(28), 8762–8767. https://doi.org/10.1073/pnas.1501242112
Hutchison, R. M., Womelsdorf, T., Allen, E. A., Bandettini, P. A., Calhoun, V. D., Corbetta,
M., … Chang, C. (2013). Dynamic functional connectivity: Promise, issues, and
interpretations. Neuroimage, 80, 5–79. https://doi.org/10.1016/j.neuroimage.2013.05.079
Kirchner, W. K. (1958). Age differences in short-term retention of rapidly changing
information. Journal of Experimental Psychology, 55(4), 352–358.
https://doi.org/10.1037/h0043688
Lehman, A., & Rourke, N. O. (2005). JMP for Basic Univariate and Multivariate Statistics A
Step-by-Step Guide. Analysis. Retrieved from
http://books.google.com/books?id=1nlApuloc0AC&pgis=1
Najafi, M., Mcmenamin, B. W., Simon, J. Z., & Pessoa, L. (2016). Overlapping communities
reveal rich structure in large-scale brain networks during rest and task conditions HHS
Public Access. Neuroimage, 135, 92–106.
https://doi.org/10.1016/j.neuroimage.2016.04.054
Naselaris, T., Kay, K. N., Nishimoto, S., & Gallant, J. L. (2011). Encoding and decoding in
fMRI. NeuroImage, 56(2), 400–410. https://doi.org/10.1016/j.neuroimage.2010.07.073
Pallares, V., Insabato, A., Sanjuan, A., Kuehn, S., Mantini, D., Deco, G., & Gilson, M.
(2018). Subject- and behavior-specific signatures extracted from fMRI data using whole-
brain effective connectivity. Doi.Org, 201624. https://doi.org/10.1101/201624
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., …
Duchesnay, É. (2011). Scikit-learn: Machine Learning in Python. Journal of Machine
Learning Research, 12, 2825–2830. Retrieved from http://scikit-
learn.org/stable/about.html#citing-scikit-learn
Ponce-Alvarez, A., He, B. J., Hagmann, P., & Deco, G. (2015). Task-Driven Activity
Reduces the Cortical Activity Space of the Brain: Experiment and Whole-Brain
Modeling. PLoS Computational Biology, 11(8), 1–26.
https://doi.org/10.1371/journal.pcbi.1004445
Preti, M. G., Bolton, T. A., & Van De Ville, D. (2017). The dynamic functional connectome:
State-of-the-art and perspectives. NeuroImage, 160, 41–54.
https://doi.org/10.1016/j.neuroimage.2016.12.061
Rahim, M., Thirion, B., Bzdok, D., Buvat, I., & Varoquaux, G. (2017). Joint prediction of
multiple scores captures better individual traits from brain images. NeuroImage, 158,
145–154. https://doi.org/10.1016/j.neuroimage.2017.06.072
Roberts, D. R., Bahn, V., Ciuti, S., Boyce, M. S., Elith, J., Guillera-Arroita, G., … Dormann,
C. F. (2017). Cross-validation strategies for data with temporal, spatial, hierarchical, or
phylogenetic structure. Ecography, 40(8), 913–929. https://doi.org/10.1111/ecog.02881
Senden, M., Reuter, N., van den Heuvel, M. P., Goebel, R., & Deco, G. (2017). Cortical rich
club regions can organize state-dependent functional network formation by engaging in
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
18
oscillatory behavior. NeuroImage, 146(October 2016), 561–574.
https://doi.org/10.1016/j.neuroimage.2016.10.044
Senden, M., Reuter, N., van den Heuvel, M. P., Goebel, R., Deco, G., & Gilson, M. (2018).
Task-related effective connectivity reveals that the cortical rich club gates cortex-wide
communication. Human Brain Mapping, 39(3), 1246–1262.
https://doi.org/10.1002/hbm.23913
Shepard, R. N., & Metzler, J. (1971). Mental Rotation of Three-Dimensional Objects.
Science, 171(3972), 701–703. https://doi.org/10.1126/science.171.3972.701
Smith, J. O. (Julius O. (2007). Mathematics of the discrete Fourier transform (DFT) : with
audio applications (2nd editio). W3K Publishing.
Sporns, O. (2013). Structure and function of complex brain networks. Dialogues in Clinical
Neuroscience, 15(3), 247–62. Retrieved from
http://www.ncbi.nlm.nih.gov/pubmed/24174898
Telesford, Q. K., Lynall, M.-E., Vettel, J., Miller, M. B., Grafton, S. T., & Bassett, D. S.
(2016). Detection of functional brain network reconfiguration during task-driven
cognitive states. NeuroImage, 142, 198–210.
https://doi.org/10.1016/j.neuroimage.2016.05.078
Varoquaux, G., Reddy Raamana, P., Engemann, D. A., Hoyos-Idrobo, A., Schwartz, Y., &
Thirion, B. (2017). Assessing and tuning brain decoders: Cross-validation, caveats, and
guidelines. https://doi.org/10.1016/j.neuroimage.2016.10.038
Vidaurre, D., Smith, S. M., & Woolrich, M. W. (2017). Brain network dynamics are
hierarchically organized in time. PNAS, 114(48), 12827–12832.
https://doi.org/10.1073/pnas.1705120114
Xie, H., Calhoun, V. D., Gonzalez-Castillo, J., Damaraju, E., Miller, R., Bandettini, P. A., &
Mitra, S. (2017). Whole-brain connectivity dynamics reflect both task-specific and
individual-specific modulation: A multitask study.
https://doi.org/10.1016/j.neuroimage.2017.05.050
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
19
Supplementary Material
Supplementary Figure 1: Adding variance to covariance does not outperform covariance
alone. (A) Within-subject cross-validation accuracy using only covariance (green) and
covariance + variance (blue). (B) Between-subject cross-validation accuracy using only
covariance (green) and covariance + variance (blue). Chance level is 0.2.
Supplementary Figure 2: Using L1 regularization instead of L2 regularization does not
improve accuracy for covariance. (A) Within-subject cross-validation accuracy using L2
penalty (green) or L1 penalty (orange). (B) Between-subject cross-validation accuracy using
L2 penalty (green) or L1 penalty (orange). Chance level is 0.2.
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
20
Supplementary Figure 3: Random split overestimates the cross-validation accuracy within a
timeseries. Within-subject cross-validation within a run using covariance with a time series
split (green) and a stratified shuffle split (orange). Chance level is 0.2.
Table S4
Results of parameters for metric and cross-validation using all features
Cross-
Validatio
n
Feature
s (All/
Best)
Metric
Penalty
parameter
C (Mean)
Penalty
parameter
C (Std)
Testing
Accurac
y
(Median
)
Testing
Accurac
y (Std)
Training
Accurac
y
(Median
)
Training
Accurac
y (Std)
T
AF
Cov20
1012.72
1907.481
0.6768
0.0286
1
0
S
AF
Cov20
0.0008
0.0008
0.4989
0.0724
0.8052
0.0956
T
AF
Cov40
106.1225
35.3507
0.9224
0.0291
1
0
S
AF
Cov40
592400.6
2079076
0.5965
0.087
1
0
T
AF
Cov60
59.8244
58.09
0.9888
0.0079
1
0
S
AF
Cov60
591719.8
2079269
0.6889
0.0973
1
0
T
AF
Cov80
94.332
47.148
0.9878
0.0137
1
0
S
AF
Cov80
1155225
2829641
0.6195
0.1389
1
0
T
AF
Cov100
70.7443
57.7611
1
0
1
0
S
AF
Cov100
577996
2082457
0.7704
0.1128
1
0
T
AF
Cov120
24.1652
46.8836
1
0
1
0
S
AF
Cov120
757.4603
1661.841
0.7667
0.1172
1
0
S
AF
gCov
14186.89
50856.45
0.7
0.229
1
0.0713
T
AF
Covvar20
530.6649
1432.906
0.6241
0.0393
1
0
S
AF
Covvar20
2324511
3644483
0.4413
0.0751
0.9879
0.0129
T
AF
Covvar40
577.2371
1416.75
0.8837
0.033
1
0
S
AF
Covvar40
591795.2
2079248
0.5322
0.0897
1
0.0008
T
AF
Covvar60
117.906
0
0.9735
0.0118
1
0
S
AF
Covvar60
2325932
3643578
0.6
0.1218
1
0
T
AF
Covvar80
94.332
47.148
0.9898
0.0106
1
0
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
21
S
AF
Covvar80
1155621
2829479
0.6195
0.1347
1
0
T
AF
Covvar100
59.8244
58.09
0.9988
0.0035
1
0
S
AF
Covvar100
578701.7
2082262
0.6417
0.1353
1
0
T
AF
Covvar120
71.0459
57.3969
1
0.0015
1
0
S
AF
Covvar120
578365.5
2082355
0.6299
0.1443
1
0
T
AF
PC6
828418.9
2420044
0.3765
0.0365
0.9205
0.0647
S
AF
PC6
0.0016
0.0004
0.2931
0.0365
0.8337
0.0692
T
AF
PC20
59.5502
58.3639
0.6589
0.034
1
0
S
AF
PC20
14455.62
50797.02
0.4297
0.0716
0.9877
0.0044
T
AF
PC40
117.906
0
0.901
0.0242
1
0
S
AF
PC40
14119.31
50875.18
0.5395
0.109
0.9991
0.0004
T
AF
PC60
83.1174
53.1455
0.9796
0.0094
1
0
S
AF
PC60
42.1202
56.4874
0.6012
0.1339
0.9999
0.0003
T
AF
PC80
106.4035
34.5077
0.9976
0.0042
1
0
S
AF
PC80
16.8893
41.2399
0.6145
0.1448
1
0
T
AF
PC100
82.8226
53.596
1
0.0024
1
0
S
AF
PC100
8.6683
30.3058
0.6394
0.1538
1
0
T
AF
PC120
48.0339
57.0611
1
0.0019
1
0
S
AF
PC120
14144.6
50868.18
0.6561
0.1779
1
0.0052
S
AF
gPC
362.2131
1238.769
0.8
0.1767
1
0
T
AF
iFC
0.8653
1.3192
0.4185
0.0301
0.8386
0.0943
S
AF
iFC
0.0015
0.0006
0.3318
0.0448
0.7501
0.0773
T
AF
eigFC
808640.7
2425920
0.3239
0.0181
0.7442
0.1186
S
AF
eigFC
577600.3
2082567
0.2724
0.0299
0.633
0.0026
T
AF
Bold
0.2888
0.8639
0.1895
0.0384
0.252
0.0267
S
AF
Bold
8.6388
30.3142
0.2391
0.0224
0.2575
0.004
T
BF
Cov20
59.5502
58.3639
0.6679
0.0259
1
0
S
BF
Cov20
14110.8827
50877.5085
0.4824
0.0732
0.8012
0.1035
T
BF
Cov40
83.1174
53.1455
0.9173
0.0315
1
0
S
BF
Cov40
1170345.44
5
2823915.00
9
0.6151
0.0889
1
0.0002
T
BF
Cov60
82.8295
53.5854
0.9867
0.0083
1
0
S
BF
Cov60
578315.028
2
2082369.07
4
0.6605
0.104
1
0
T
BF
Cov80
94.332
47.148
0.9867
0.0133
1
0
S
BF
Cov80
577970.316
3
2082464.37
8
0.6039
0.1353
1
0
T
BF
Cov100
70.7443
57.7611
1
0
1
0
S
BF
Cov100
412.5183
1225.3701
0.7718
0.1126
1
0
T
BF
Cov120
24.1652
46.8836
1
0
1
0
S
BF
Cov120
592442.954
3
2079064.26
5
0.75
0.1155
1
0
T
AF
Prec10_GL
39510.52
79020.91
0.876
0.0127
0.9963
0.0035
S
AF
Prec10_GL
0.0017
0
0.5275
0.1068
0.9867
0.0014
T
AF
Prec20_GL
0.6255
1.1277
0.9934
0.002
0.9992
0.0003
S
AF
Prec20_GL
0.0213
0.031
0.6581
0.1298
0.9986
0.0005
T
AF
Prec30_GL
0.0223
0.0315
1
0.0007
1
0
S
AF
Prec30_GL
101.4738
40.2506
0.7321
0.159
1
0
T
AF
Prec40_GL
0.0017
0
1
0
1
0
TASK INFORMATION ACROSS FUNCTIONAL CONNECTIVITY METHODS
22
S
AF
Prec40_GL
51.5701
57.458
0.7645
0.17
1
0
T
AF
Prec50_GL
0.0086
0.0206
1
0.0008
1
0
S
AF
Prec50_GL
59.9869
57.9263
0.8169
0.1697
1
0
T
AF
Prec60_GL
0.0017
0
1
0
1
0
S
AF
Prec60_GL
14498.76
50784.77
0.8591
0.1859
1
0
T
AF
Cov10_GL
0.9134
1.2877
0.9349
0.0081
0.9963
0.0002
S
AF
Cov10_GL
345.5998
1242.707
0.5527
0.1188
0.9965
0.0005
T
AF
Cov20_GL
519.7243
1436.427
0.9967
0.0022
0.9992
0.0001
S
AF
Cov20_GL
1413.447
2158.975
0.6337
0.1353
0.9991
0.0001
T
AF
Cov30_GL
507.0977
1440.45
1
0.0012
1
0
S
AF
Cov30_GL
117.906
0
0.6951
0.1365
1
0
T
AF
Cov40_GL
0.0566
0.0275
1
0.0012
1
0.0006
S
AF
Cov40_GL
413.1352
1225.163
0.6882
0.1519
1
0
T
AF
Cov50_GL
0.3102
0.8573
1
0
0.9998
0.0001
S
AF
Cov50_GL
577651.8
2082552
0.6831
0.1412
1
0
T
AF
Cov60_GL
0.2896
0.8636
1
0.0009
0.9995
0.0002
S
AF
Cov60_GL
43.961
55.1154
0.6667
0.1351
1
0
| 2018 | Comparing Task-Relevant Information Across Different Methods of Extracting Functional Connectivity | 10.1101/509059 | [
"Stulz Sophie Benitez",
"Insabato Andrea",
"Deco Gustavo",
"Gilson Matthieu",
"Senden Mario"
] | creative-commons |
Mechanisms of up-regulation of Ubiquitin-Proteasome activity in the absence of NatA
dependent N-terminal acetylation
Ilia Kats1,2, Marc Kschonsak1,3, Anton Khmelinskii4, Laura Armbruster1,5, Thomas Ruppert1 and
Michael Knop1,6,*
1 Zentrum für Molekulare Biologie der Universität Heidelberg (ZMBH), DKFZ-ZMBH Alliance,
Im Neuenheimer Feld 282, 69120 Heidelberg, Germany.
2 present address: German Cancer Research Center (DKFZ), Im Neuenheimer Feld 280,
69120 Heidelberg, Germany
3 present address: Department of Structural Biology, Genentech Inc., South San Francisco,
CA, USA.
4 Institute of Molecular Biology (IMB), Ackermannweg 4, 55128 Mainz, Germany.
5 present address: Centre for Organismal Studies (COS), Im Neuenheimer Feld 360, 69120
Heidelberg, Germany
6 Deutsches Krebsforschungszentrum (DKFZ), DKFZ-ZMBH Alliance, Im Neuenheimer Feld
280, 69120 Heidelberg, Germany.
* corresponding author: m.knop@zmbh.uni-heidelberg.de
Abstract
N-terminal acetylation is a prominent protein modification and inactivation of N-terminal
acetyltransferases (NATs) cause protein homeostasis stress. Using multiplexed protein
stability (MPS) profiling with linear ubiquitin fusions as reporters for the activity of the ubiquitin
proteasome system (UPS) we observed increased UPS activity in NatA, but not NatB or NatC
mutants. We find several mechanisms contributing to this behavior. First, NatA-mediated
acetylation of the N-terminal ubiquitin independent degron regulates the abundance of Rpn4,
the master regulator of the expression of proteasomal genes. Second, the abundance of
several E3 ligases involved in degradation of UFD substrates is increased in cells lacking NatA.
Finally, we identify the E3 ligase Tom1 as a novel chain elongating enzyme (E4) involved in
the degradation of linear ubiquitin fusions via the formation of branched K11 and K29 ubiquitin
chains, independently of the known E4 ligases involved in UFD, leading to enhanced
ubiquitination of the UFD substrates.
Introduction
Selective protein degradation is essential for proteome homeostasis, to remove unnecessary
or abnormal proteins as part of quality control pathways or in response to changes in the
environment. In eukaryotes the bulk of selective protein degradation is handled by the
ubiquitin-proteasome system (UPS). Substrates of the UPS are recognized through features
known as degradation signals or degrons (Ravid and Hochstrasser, 2008), ubiquitinated by E3
ubiquitin ligases typically on lysine side chains, and finally degraded by the proteasome (Finley
et al., 2012; Hershko and Ciechanover, 1998).
Global activity of the UPS is tightly regulated and responds to environmental challenges such
as heat stress, DNA damage or cytotoxic compounds, which can damage or induce misfolding
of proteins (Hahn et al., 2006). In the budding yeast Saccharomyces cerevisiae, the
transcription factor Rpn4 is a master regulator of proteasome capacity. It trans-activates
promoters of all proteasomal subunits and several other proteins of the UPS (Mannhaupt et
al., 1999; Shirozu et al., 2015). Expression of Rpn4 is in turn regulated by several stress-
induced transcription factors such as Hsf1 and Yap1 (Hahn et al., 2006).
In addition to global regulation of the UPS that affects the entire proteome, selective
degradation of specific proteins can be induced through post-translational modifications
creating or exposing degradation signals. N-degrons that target for degradation via an N-
terminal destabilizing residue can be formed by specific endoproteolytic cleavage. For
example, cohesin cleavage by separase at the metaphase-anaphase transition induces
degradation of the C-terminal fragment by the Arg/N-end rule pathway that recognizes the
newly exposed N-terminal residue as a degradation signal (Rao et al., 2001).
Nα-terminal acetylation of proteins (Nt-acetylation) is a co-translational modification catalyzed
by ribosome-associated Nα-terminal acetyltransferase (NAT) complexes. Three NATs, NatA,
NatB, and NatC, are responsible for the acetylation of 50-90% of all protein N-termini in yeast
and human cells (Aksnes et al., 2016; Starheim et al., 2012). These NATs differ in their
substrate specificity. NatA acetylates the small residues (S,A,V,C,G) after they have been
exposed at the N-terminus through cleavage of the initiator methionine (iMet) by methionine
aminopeptidases (MetAPs). NatB and NatC acetylate the iMet if it is followed by a polar residue
(one of (D,E,N,Q)) or a large hydrophobic residue (one of (F,L,I,W)), respectively. The identity
of the first two N-terminal residues is however not sufficient to trigger Nt-acetylation, and
numerous proteins lack this modification despite being potential NAT substrates according to
their primary sequence (Aksnes et al., 2016).
Nt-acetylation of nascent chains is a prevalent protein modification affecting the majority of all
proteins and it has been implicated in a multitude of cellular processes. Deletion of the major
N-acetyl transferase genes leads to pleiotropic effects with distinct influences on the
physiology and cellular proteostasis of S. cerevisiae. For NatA several individual targets are
known where Nt-acetylation functions in mediating protein-protein interactions, prevention of
incorrect protein secretion, protein folding and degradation. This includes key transcriptional
regulators as well as protein folding machinery or structural components of the cytoskeleton
(Aksnes et al., 2016; Friedrich et al.), therefore it is not surprising that the attribution of specific
functions to Nα-acetylated N-termini is not possible. Another important point is the question to
what extent Nt-acetylation is subject to specific regulation, e.g. via regulation of the activity of
the individual NATs. While there is some evidence from plants that NatA activity can be
regulated as a function of drought stress (Linster et al., 2015), in yeast no clear reports about
specific regulation of NATs exist. This is consistent with the observation that Nt-acetylation
appears to be irreversible and that is hardly affected by reduced Acetyl-CoA levels (Varland et
al., 2018).
Nt-acetylation was proposed to act as a degradation signal (Hwang et al., 2010b; Shemorry et
al., 2013) and Nt-acetylated N-termini are thought to be recognized and ubiquitinated by
specific E3 ligases of the Ac/N-end rule pathway. The universality of this pathway is debatable,
because acetylation is not a self-sufficient degron and the involved E3 ligases recognize a
broad palette of N-degrons independent on Nt-acetylation (Friedrich et al.; Gawron et al., 2016;
Kats et al., 2018; Zattas et al., 2013). Still, Nt-acetylation can be part of N-degrons that contain
adjacent sequence motifs (Hwang et al., 2010b; Shemorry et al., 2013). In recognition of the
fact that Nt-acetylation is not a general degron, it was finally proposed to refer to ‘N-terminal
degrons’, and to avoid the wording ‘N-end rule’ (Varshavsky, 2019), in favor of specific
terminology that refers to the individualistic nature of each N-terminal degron. This is even
more important given that accumulating evidence suggests that N-acetylation can fulfill the
exact opposite function: as a protein stabilizing modification. First, Nt-acetylation can prevent
direct ubiquitylation of the Nα amino group of proteins (Caron et al., 2005; Hershko et al., 1984;
Kuo et al., 2004). This may be the underlying mechanism how Nt-acetylation protects the
Derlin protein Der1 from degradation by the associated E3 ligase Hrd1 (Zattas et al., 2013).
Acetylation can also protect N-termini from non-canonical processing by aminopeptidases, i.e.
methionine aminopeptidases 1 and 2 (MAP1/2), which, in the absence of Nt-acetylation, can
remove the initiator methionine (iMet) form the nascent chain. This leads to the exposure of
the second residue, which in the case of NatB and NatC N-termini will lead to the exposure of
an Arg/N-degron that can targets the protein for Ubr1 dependent degradation (Kats et al., 2018;
Nguyen et al., 2018).
The yeast genome encodes several linear ubiquitin fusion proteins which serve as a source of
free ubiquitin, since the N-terminal ubiquitin moiety is usually co-translationally cleaved off by
endogenous deubiquitinating enzymes (DUBs) (Amerik and Hochstrasser, 2004). Linear
ubiquitin fusions that escape DUB cleavage or that are generated post-translationally by
ubiquitination of the Nα group of the first amino acid residue of a protein can be further
ubiquitinated by E3 ligases of the ubiquitin-fusion degradation (UFD) pathway using
conventional lysine-ε-amino-specific linkage on at least one of the seven lysine residues of the
N-terminal ubiquitin moiety and degraded by the proteasome. In yeast, Ufd4 is the major E3
ligase of the UFD pathway (Johnson et al., 1995), while the accessory E3 ligases Ufd2 and
Ubr1 promote degradation by acting as chain elongating enzymes (E4 ligases) (Hwang et al.,
2010a; Koegl et al., 1999). The UFD pathway is conserved in humans, where it is composed
of the Ufd4 ortholog TRIP12 and the Ufd2 orthologs UFD2a and UFD2b (Park et al., 2009).
The pathway was first identified in yeast using artificial substrates consisting of linear ubiquitin
fusions (UbiG76V) that are resistant to cleavage by DUBs (Johnson et al., 1995). Such UFD
substrates were subsequently used as a high-throughput-compatible readout of proteasome
activity (Dantuma et al., 2000; Stack et al., 2000). However, endogenous substrates of the
UFD pathway have proven difficult to identify, and only few are known to date. Nevertheless,
mammalian cells possess the E2 conjugating enzyme Ube2w that monoubiquitinates N-
terminal residues if they are followed by an intrinsically disordered sequence (Scaglione et al.,
2013; Tatham et al., 2013; Vittal et al., 2014) as well as the E3 ligase LUBAC that assembles
linear M1-linked ubiquitin chains and was implicated in immune signaling (Fiil et al., 2013;
Gerlach et al., 2011; Tokunaga et al., 2009). However, to the best of our knowledge, the origin
of the N-terminal ubiquitin moiety in known endogenous UFD substrates has not been
investigated, and all known instances of N-terminal ubiquitination by LUBAC or Ube2w do not
induce degradation of the substrate, but rather mediate protein-protein interactions or activate
signaling cascades (Rittinger and Ikeda, 2017). N-terminal ubiquitination has been suggested
to be regulated by Nt-acetylation, as both modifications involve the same amino group (Caron
et al., 2005; McDowell and Philpott, 2013).
We have developed multiplexed protein stability (MPS) profiling, a quantitative and high-
throughput compatible method that enables the degradation profiling of large peptide libraries
using fluorescence activated cell sorting (FACS) and analysis of enriched fractions by deep
sequencing (Kats et al., 2018). We used MPS profiling to explore the degron propensity of
native and non-native N-termini and a large fraction of the yeast N-termini (N-terminome) (Kats
et al., 2018). In this work we explore the influence of NatA on protein degradation in the
budding yeast S. cerevisiae starting with the observation that artificial UFD substrates are
degraded faster in NatA-deficient cells. Using screening and targeted we describe a role for
Nt-acetylation on regulation of UPS activity via Rpn4 and we investigate how the abundance
of several E3 and E4 ubiquitin ligases is influenced by NatA and how this contributes to UFD.
We furthermore identify Tom1 as a novel ubiquitin chain-elongating enzyme (E4) of the UFD
pathway and using in vivo and in vitro assays we investigate ubiquitination by Tom1.
Altogether, our data provide new insights into the molecular processes governing UPS activity
regulation in the absence of NatA activity, emphasizing the importance of NatA for cellular
protein homeostasis.
Results
NatA affects turnover of UFD substrates
We performed a systematic survey of degrons in protein N-termini using linear ubiquitin fusion
reporter constructs (Kats et al., 2018). These reporters consisted of an N-terminal ubiquitin
followed by two variable residues (X and Z), a linker sequence (eK) and a tandem fluorescent
protein timer (tFT). The tFT reports on protein stability independently of expression through
the intensity ratio of the slow maturing mCherry and the fast maturing sfGFP fluorescent
proteins, which increases as a function of protein half-life in steady state (Khmelinskii et al.,
2012; Khmelinskii et al., 2016). In the course of that study we observed that reporters with a
proline residue immediately following the ubiquitin moiety (Ubi-PZ-tFT reporters) exhibited
increased turnover in strains lacking the N-terminal acetyltransferase NatA (Fig 1A), whereas
no destabilization was observed in NatB and NatC mutants (see Supplementary Figure S3 in
(Kats et al., 2018)). The N-terminal ubiquitin moiety is usually co-translationally cleaved by
endogenous deubiquitinating enzymes (DUBs) (Bachmair et al., 1986), which enables the
exposure of non-native amino acid residues at the N-terminus of the reporter protein. However,
a proline residue located directly after ubiquitin impairs cleavage of the ubiquitin moiety by
DUBs. Such linear ubiquitin fusions are rapidly degraded by the ubiquitin fusion degradation
(UFD) pathway (Johnson et al., 1995), primarily through the action of the ubiquitin E3 ligases
Ufd4 and Ubr1 (Hwang et al., 2010a). In contrast, cleaved Ubi-PZ-tFT reporters with an
exposed N-terminal proline are stable (Bachmair et al., 1986; Bachmair and Varshavsky,
1989).
To understand how NatA affects turnover of Ubi-PZ-tFT reporters, we first confirmed that these
reporters are affected by deletion of the catalytic NatA subunit NAA10 using cycloheximide
chase experiments. These immunoblots indicated that abundance and/or degradation of an
uncleaved Ubi-PP-tFT reporter are influenced by deletion of NAA10, the catalytic subunit of
NatA (Fig 1B). These results can be explained either by accelerated degradation of uncleaved
Ubi-PZ-tFT reporters or by impaired DUB activity in the naa10Δ mutant. In DUB-impaired cells,
a larger fraction of Ubi-PZ-tFT reporters would remain uncleaved, and rapid degradation of
uncleaved Ubi-PZ-tFT reporters by the UFD pathway would account for their apparent
destabilization. To distinguish between these possibilities, we investigated turnover of a non-
cleavable UbiG76V-tFT reporter, in which the last glycine of ubiquitin is exchanged for valine
to completely prevent cleavage by DUBs (Johnson et al., 1992). Degradation of this reporter
was inferred from mCherry/sfGFP ratio as measured by flow cytometry. Stability of the
UbiG76V-tFT reporter in wild type yeast was at the lower end of the tFT dynamic range,
therefore no clear effect of NAA10 deletion could be detected by flow cytometry (Fig 1C, pos.
1&2, 5&6). As expected, this reporter was strongly stabilized in ufd4Δ and ubr1Δ ufd4Δ cells.
Surprisingly however, it was still degraded in these mutants and moreover, it was clearly
destabilized upon deletion of NAA10 (Figs 1C, pos. 3&4, 7&8, for a CHX chase, see Fig. S1).
This suggests that NatA-dependent acceleration of UFD substrate turnover is independent of
DUB activity. The results also suggest that accelerated degradation does not involve the
canonical E3 ubiquitin ligases implicated in the degradation of such linear ubiquitin fusions.
Fig 1. Accelerated degradation of linear ubiquitin fusion proteins in NatA-deficient
strains.
(A) Average stability of Ubi-PZ-tFT reporters in the indicated strains. The protein stability index
(PSI) is a measure of protein turnover resulting from high-throughput analysis of tFT-tagged
constructs and increases as a function of the mCherry/sfGFP ratio and is therefore
anticorrelated with degradation rate. Data from Kats et al. (Kats et al., 2018). Boxplots show
median, 1st and 3rd quartile, whiskers extend to ± 1.5x interquartile range (IQR) from the box.
p: two-sided paired t-test.
(B) Degradation of the Ubi-PP-tFT reporter after blocking translation with cycloheximide.
Whole-cell extracts were separated by SDS-PAGE followed by immunoblotting with antibodies
against GFP and Pgk1 as loading control. A product resulting from mCherry autohydrolysis
during cell extract preparation (Gross et al., 2000) is marked (∗).
(C) Flow cytometry analysis of strains expressing the UbiG76V-tFT reporter. For all flow
cytometry experiments, mCherry/sfGFP ratios were normalized to a stable control measured
in the same strain background. Mean mCherry/sfGFP ratios and 95% CI of six replicates are
plotted together with the median mCherry/sfGFP ratio of each replicate.
Nt-acetylation by NatA promotes ubiquitin-independent degradation of Rpn4
DUB-independent destabilization of the UbiG76V-tFT reporter in strains lacking the known E3s
of the UFD pathway suggested that at least one additional E3 ligase involved in degradation
of UFD substrates exists. While searching for this E3, we noticed that deletion of the Ubr2 E3
ligase in the ubr1Δ ufd4Δ background accelerated degradation of the UbiG76V-tFT reporter.
This destabilization was additive to the effect of NAA10 deletion on UbiG76V-tFT reporter
stability (Fig 2A, pos. 1 to 4). Ubr2 acts via the Rpn4 transcription factor to regulate expression
of UPS genes. More specifically, Rpn4 possesses two degrons, a ubiquitin-dependent degron
that is recognized by Ubr2, and an N-terminal ubiquitin-independent degron that is directly
recognized by the 26S proteasome (Ju et al., 2004; Ju and Xie, 2004; Ju and Xie, 2006; Wang
et al., 2004a) (Fig 2B). These degrons induce a negative feedback loop regulating UPS activity,
such that Rpn4 abundance and consequently proteasome biogenesis are balanced to meet
the proteolytic load (Xie and Varshavsky, 2001). Deletion of the Ubr2-dependent degron of
Rpn4 (Rpn4Δ(211-229) (Wang et al., 2010)) destabilized the UbiG76V-tFT reporter in the
ubr1Δ ufd4Δ background. No further destabilization of this reporter was observed upon
additional deletion of UBR2 (Fig 2A, pos. 5 to 8). This indicates that accelerated degradation
of the UbiG76V-tFT reporter upon ablation of Ubr2 is due to stabilization of Rpn4. Rpn4 is a
potential NatA substrate according to its primary sequence, which starts with MA. To explain
the additive effect of NatA deletion on degradation of the UbiG76V-tFT reporter, we
hypothesized that Nt-acetylation of Rpn4 affects its N-terminal ubiquitin-independent degron.
Consistent with this idea, abundance of C-terminally TAP-tagged Rpn4 was strongly increased
in the naa10Δ mutant (Fig 2C). To test this hypothesis directly, we exploited the portability of
the ubiquitin-independent degron of Rpn4 (Ha et al., 2012) and measured turnover of an
Rpn4(1-80)-tFT reporter containing the N-terminal Ubi-independent degron of Rpn4 fused to
the tFT. This reporter was stabilized upon deletion of NAA10 (Fig 2D, Pos. 1&2). Preventing
NatA-mediated Nt-acetylation by substituting the second residue for asparagine strongly
reduced stabilization of the reporter in the naa10Δ mutant (Fig 2D, Pos. 4&5). Instead, this
Rpn4A2N(1-80)-tFT reporter, a potential target of NatB, was stabilized in naa20Δ cells lacking
the catalytic subunit of NatB (Fig 2D, Pos. 4&6) to a similar extent as the Rpn4(1-80) reporter
in naa10Δ cells (Fig 2D, Pos. 1&2). These results indicate that Rpn4A2N(1-80)-tFT is
acetylated by NatB, and that Nt-acetylation, regardless of the NAT, promotes ubiquitin-
independent degradation of Rpn4. Label-free quantitative mass spectrometry of full-length
Rpn4 confirmed NatA-dependent acetylation of Rpn4 and NatB-dependent acetylation of
Rpn4A2N (Fig 2E).
Next we investigated the influence of NatA on ubiquitin-independent degradation of Rpn4 in a
physiological context. We performed cycloheximide chases of C-terminally TAP-tagged Rpn4
lacking its ubiquitin-dependent degron (Rpn4Δ(211-229)-TAP). Deletion of NAA10 doubled the
half-life of Rpn4Δ(211-229-TAP), but not of Rpn4A2N,Δ(211-229-TAP) (Fig 2F and Fig S2A,B).
Taken together, these results argue that NatA-mediated N-terminal acetylation of Rpn4
promotes its ubiquitin-independent degradation, thereby modulating its abundance.
To assess if Rpn4 mediates the accelerated degradation of the UbiG76V-tFT reporter in the
naa10Δ mutant, we measured turnover of this reporter in cells carrying the rpn4A2N allele
using flow cytometry. Destabilization of this reporter upon deletion of NAA10 was only
marginally reduced in the Rpn4A2N mutant compared to cells expressing wild type Rpn4 (Fig
2G). This suggests that elevated levels of Rpn4 could contribute to, but are not solely
responsible for, accelerated turnover of UFD substrates in the absence of NatA.
Fig 2: Regulation of the ubiquitin independent degron of Rpn4 by NatA and contribution
to degradation of UFD substrates.
(A) Flow cytometry analysis of strains expressing the UbiG76V-tFT reporter.
(B) Domain architecture of Rpn4.
(C) Degradation of C-terminally TAP-tagged Rpn4 after blocking translation with
cycloheximide. Whole-cell extracts were separated by SDS-PAGE followed by immunoblotting
with antibodies against protein A and Pgk1 as loading control.
(D) Flow cytometry analysis of strains expressing the indicated Rpn4 N-terminal sequences
fused to the tFT. mCherry/sfGFP ratios were normalized to the mean mCherry/sfGFP ratio of
the wild type strain.
(E) Extracted ion chromatograms of Nt-acetylated and unmodified N-terminal peptides derived
from full-length Rpn4 variants obtained by label-free mass spectrometry.
(F) Half-lives of C-terminally TAP-tagged Rpn4 variants estimated by cycloheximide chase.
Mean half-lives and 95% CI of six replicates are plotted together with the half-life of each
replicate. p: one-sided unpaired t-test.
(G) Flow cytometry analysis of strains expressing the UbiG76V-tFT reporter. mCherry/sfGFP
ratios were normalized to the mean mCherry/sfGFP ratio of NAA10 wild type strains. AN:
Rpn4A2N. p: one-sided unpaired t-test.
Tom1 is an E4 ligase of the UFD
Rpn4-independent destabilization of the UbiG76V-tFT reporter in ubr1Δ ufd4Δ cells upon
deletion of NatA (Fig 2G) is consistent with our initial hypothesis, the existence of an unknown
E3 ligase targeting this reporter for degradation. In human cells, the E3 ligase HUWE1 was
implicated in the UFD pathway (Poulsen et al., 2012). The yeast homolog Tom1 targets excess
histones (Singh et al., 2012), ribosomal subunits (Sung et al., 2016) and other proteins (Kim
and Koepp, 2012; Kim et al., 2012) for degradation, but has not yet been described to mediate
UFD. We used flow cytometry to test if Tom1 participates in degradation of UFD substrates
and observed only weak stabilization of the UbiG76V-tFT reporter in the tom1Δ mutant (Fig
3A, Pos. 1&2). This could explain why Tom1 was not previously identified as a component of
the UFD pathway. Nevertheless, we were able to co-immunoprecipitate C-terminally TAP-
tagged Tom1 with the UbiG76V-tFT reporter (Fig 3B), suggesting a direct role for Tom1 in
degradation of UFD substrates.
According to the current model of the UFD pathway, linear ubiquitin fusions are first
oligoubiquitinated by Ufd4 on the K29 residue of the N-terminal ubiquitin moiety (Johnson et
al., 1995; Tsuchiya et al., 2013). These short chains are then extended by the chain-elongating
E4 enzymes Ufd2 and Ubr1 to degradation-promoting length (Hwang et al., 2010a; Koegl et
al., 1999). While Ubr1 activity has not been investigated in detail, Ufd2 is known to require K48
of the N-terminal ubiquitin moiety (Johnson et al., 1995; Koegl et al., 1999; Liu et al., 2017).
The weak stabilization of the UbiG76V-tFT reporter in the tom1Δ mutant suggests that Tom1
is redundant with Ufd4 or one of the E4 ligases. UFD substrates lacking K29 are fully stable
(Johnson et al., 1995) and thus cannot be used to distinguish between these possibilities. To
more confidently place Tom1 in the UFD pathway, we therefore mutated K48 of the UbiG76V-
tFT reporter to arginine and measured turnover of the resulting UbiK48R,G76V-tFT reporter
using flow cytometry. In wild type yeast, the UbiK48R,G76V-tFT reporter was degraded slower
than the UbiG76V-tFT reporter and was not stabilized in a ufd2Δ mutant, consistent with the
current model. Strikingly, deletion of TOM1 almost completely abolished degradation of the
UbiK48R,G76V-tFT reporter (Fig 3A, Pos. 8&9) and the tom1Δ and ubr1Δ ufd4Δ mutants were
indistinguishable in terms of UbiK48R,G76V-tFT turnover (Fig 3A, Pos. 9&10). Interestingly,
the UbiK48R,G76V-tFT reporter was slightly more stable in a tom1Δ ubr1Δ ufd4Δ mutant
compared to either tom1Δ or ubr1Δ ufd4Δ cells (Fig 3A, Pos. 9 to 11). Altogether, these
observations argue that Tom1 can play a major role in degradation of UFD substrates.
However, Tom1 is not essential for degradation of UFD substrates, as other ubiquitin ligases
can use K48 to promote degradation of UFD substrates independently of Tom1. One such
ligase is Ufd2, but it is likely that additional ligases performing this function exist, as the
UbiG76V-tFT reporter was still degraded in a tom1Δ ufd2Δ mutant (Fig 3A).
Fig 3. Role of Tom1 in degradation of UFD substrates.
(A) Flow cytometry analysis of strains expressing UbiG76V-tFT or UbiK48R,G76V-tFT
reporters.
(B) Co-purification of Tom1 with the UbiG76V-tFT reporter. Proteins were separated by SDS-
PAGE followed by immunoblotting with antibodies against GFP, protein A, and Zwf1 as loading
control. Input: whole-cell extract prepared by glass bead lysis. IP: proteins eluted after
incubation of whole-cell extracts with GFP binding protein coupled to sepharose beads. The
EH reporter is not a UFD substrate. It is therefore thought to not be targeted by Tom1 and
served as negative control. (∗) marks a non-specific band.
We considered two mechanisms that could explain our results: (i) UFD substrates are
ubiquitinated sequentially by Ufd4 and Tom1 and ubiquitination by Tom1 depends on Ufd4; or
(ii) Tom1 ubiquitinates UFD substrates independently of Ufd4 on a lysine residue distinct from
K48. In the absence of E4 activity on K48, ubiquitination by either Ufd4 or Tom1 alone is not
sufficient to target the reporter for degradation and both E3 ligases are required. To distinguish
between these possibilities and to investigate the effect of Tom1 on ubiquitin chain formation,
we purified ubiquitin conjugates from whole-cell extracts. The abundance of high molecular
weight species originating from the UbiG76V-tFT reporter was reduced in the tom1Δ mutant
(Fig 4A, lanes 9&13). Moreover, only mono- and diubiquitinated species were seen in the
tom1Δ mutant, when using the UbiK48R,G76V-tFT reporter, despite strong polyubiquitination
of this reporter in wild type cells (Fig 4B, lanes 5&7). In a ubr1Δ ufd4Δ background, the
UbiK48R,G76V-tFT reporter was only weakly ubiquitinated (Fig 4B, lane 6). Altogether, these
results are consistent with the idea that Tom1 acts as a chain elongating enzyme (E4) in the
UFD pathway, which recognizes proteins that carry linear oligoubiquitin chains added by Ufd4
and extends these to a degradation-promoting length.
To test this hypothesis directly, we reconstituted ubiquitination of UFD substrates in vitro
(Hwang et al., 2010a; Koegl et al., 1999). We first investigated ubiquitination by Ufd4, Ufd2,
and Ubr1. Using Ubi-ProtA as a substrate, Ubr1 or Ufd4 alone generated short ubiquitin chains
of up to three or four ubiquitin monomers in length, respectively, while Ufd2 was inactive in the
absence of other E3 ligases (Fig 4C, lanes 1 to 5). On the other hand, Ufd4 combined with
Ufd2 and/or Ubr1 generated high molecular weight conjugates (Fig 4C, lanes 6 to 8). When
UbiK48R-ProtA was used as a substrate, the combination of Ufd4 and Ufd2 did not synthesize
appreciable amounts of polyubiquitin conjugates (Fig 4D, lanes 3&7). Altogether, these results
reproduce previous observations (Hwang et al., 2010a; Koegl et al., 1999) and hence confirm
the integrity of our in vitro system. Next, we used this assay to investigate the effect of Tom1
on ubiquitin chain formation. Tom1 alone was inactive towards both Ubi-ProtA and UbiK48R-
ProtA, but generated high molecular weight polyubiquitin chains in the presence of Ufd4
regardless of the model substrate (Fig 4C and D, lanes 1, 3, 9 & 11 each). This indicates that
Tom1 recognizes oligoubiquitinated UFD substrates and either extends pre-formed chains or
synthesizes new chains conjugated directly to the substrate, but using a residue distinct from
K48 of the N-terminal ubiquitin moiety for chain attachment. Since HUWE1, the mammalian
homologue, was shown to synthesize K6- and K11-linked chains (Michel et al., 2017; Yau et
al., 2017), it is possible that Tom1 can use those lysine residues of the N-terminal ubiquitin
moiety to initiate new chains. Moreover, detailed analysis of the banding pattern revealed that
in the presence of Tom1 tri-ubiquitinated species of different apparent molecular weight were
generated (Fig 4E), indicating that ubiquitin conjugates synthesized by Tom1 and Ufd4 are
clearly distinct.
We next used mass spectroscopy in order to identify the type of linkages formed in the in vitro
ubiquitination reactions. In reactions that included Ufd4 alone (Fig 4C, lane 3), only K29
linkages were observed (Fig 4F) as expected (Koegl et al., 1999; Liu et al., 2017). Upon
addition of Tom1 a strong signal for K48 linkages was observed (Fig 4F) indicating the
formation of elongated chains based on K48 linkages. When we tested the high molecular
weight products of full reactions (Fig. 4C, lane 15) that included Ufd4, Ufd2 and Tom1 we could
also detect K11 linkages, whereas these linkages were absent in this fraction when Tom1 was
omitted (Fig. 4C, lane 8). Together these results support the idea Tom1 functions as an E4
enzyme and that it is able to form different types of ubiquitin linkages.
Fig 4. Tom1 is an E4 ubiquitin ligase and catalyzes the formation of K48 and K11
ubiquitin linkages
(A and B) Ubiquitination of UbiG76V-tFT (A) or UbiK48R,G76V-tFT (B) in strains expressing
10xHis-tagged ubiquitin. Total cell extracts and ubiquitin conjugates purified by immobilized
metal affinity chromatography were analyzed by SDS-PAGE followed by immunoblotting
against GFP, Zwf1, and ubiquitin. A product of mCherry hydrolysis during cell extract
preparation (Gross et al., 2000) (∗) and a product resulting from inefficient proteasomal
degradation of sfGFP (Khmelinskii et al., 2016) (∗∗) are marked.
(C and D) In vitro reconstitution of ubiquitin chain formation with Ubi-ProtA (C) or UbiK48R-
ProtA (D) as substrate using immunoblotting against protein A.
(E) Comparison of the banding pattern of lanes 3 and 11 from (D). Length of ubiquitin chains
is indicated.
(F) Analysis of ubiquitin linkages by mass spec. Ubiquitinated proteins were isolated from SDS
PAGE gels prepared from samples in (C) and analyzed for the presence of branched chains
as described in methods. The abundance of characteristic fragments in the eluates is shown.
Traces were normalized to the non-modified K63 peptide.
Next, we tested if Tom1 contributes to the destabilization of UFD substrates in NatA-deficient
cells. In the ubr1Δ ufd4Δ background, cells lacking Tom1 showed a markedly reduced
acceleration of UbiG76V-tFT reporter degradation upon deletion of NAA10 compared to Tom1-
proficient cells, and this destabilization was further reduced, but not completely abolished, in
cells carrying the rpn4A2N allele (Fig 5A). These results indicate that accelerated turnover of
UFD substrates in the naa10Δ mutant is mediated partially by Tom1, partially by reduced
ubiquitin-independent degradation of Rpn4, and partially by other factors.
Increased abundance of Tom1 and/or other UFD-specific E3 ligases in the naa10Δ
background could explain accelerated turnover of UFD substrates in this mutant. Supporting
this notion, elevated levels of Ubr1 in NatA-deficient cells have been observed previously (Oh
et al., 2017). We therefore tested if NatA affects abundance of Ufd4 and Tom1 using
immunoblotting. Levels of both E3 ligases were elevated in naa10Δ cells compared to wild
type (Fig 5A and Fig S3).
To test if increased abundance of E3s participating in UFD can accelerate degradation of UFD
substrates, we measured degradation of UbiG76V-tFT and Ubi-PZ-tFT reporters in strains
overexpressing Ufd4, Tom1, or Ubr1 using flow cytometry. No clear changes in turnover of the
UbiG76V reporter were detected, most likely because it is at the lower limit of the tFT dynamic
range in wild type cells. The Ubi-PP-tFT reporter was more stable in the wild type background
but was only weakly destabilized in a strain overexpressing Ubr1 (Fig 5B), consistent with a
negligible contribution of Ubr1 to UFD in vivo (Figs 1C and 3C (Hwang et al., 2010a)). However,
overexpression of Ufd4 or Tom1 strongly destabilized the PP reporter (Fig 5B). Only a fraction
of this reporter is degraded by the UFD pathway, while the other fraction is stable due to
removal of the N-terminal ubiquitin moiety by DUBs. Increased turnover of this reporter upon
overexpression of Ufd4 and Tom1 therefore indicates that these E3 ligases can compete with
DUB activity. Moreover, these results suggest that increased abundance of UFD E3 ligases
could explain accelerated turnover of UFD substrates in the naa10Δ mutant.
Given that deletion of NAA10 in a ubr1Δ ufd4Δ tom1Δ rpn4A2N background still slightly
accelerated degradation of the UbiG76V-tFT reporter (Fig 5A), we hypothesize that this
destabilization is not due to the action of one single protein, but rather the result of a systemic
upregulation of the UPS, caused in part by reduced ubiquitin-independent degradation of
Rpn4, but also by other factors currently unknown. A reason for this could be unspecific, low-
efficiency ubiquitination of the N-terminal ubiquitin moiety by most, if not all, cellular E3 ligases,
in addition to the specific, high-efficiency ubiquitination by Ufd4 and Tom1. Upregulation of the
UPS would therefore lead to not only an increase in specific and unspecific ubiquitination of
UFD substrates but also accelerated proteasomal degradation.
Fig 5. Role of NatA in regulation of the UFD.
(A) Flow cytometry analysis of strains expressing the UbiG76V-tFT reporter. mCherry/sfGFP
ratios were normalized to the mean mCherry/sfGFP ratio of NAA10 wild type strains. AN:
Rpn4A2N. p: one-sided unpaired t-test.
(B) Abundance of C-terminally 3HA-tagged Ufd4 or TAP-tagged Tom1 in cells lacking NatA
compared to wild type yeast. Whole-cell extracts were separated by SDS-PAGE followed by
immunoblotting with antibodies against HA and Pgk1 (Ufd4) or with antibodies against protein
A and Fas (Tom1). Pgk1 and Fas served as loading controls. Mean fold-change and 95% CI
of six replicates are plotted together with the fold-change of each replicate. p: one-sample t-
test.
(C) Flow cytometry analysis of strains expressing UbiG76V-tFT or Ubi-PP-tFT reporters. OE:
overexpression from the GPD promoter.
Discussion
Our study sheds light on the impact of NatA Nt-acetylation on protein homeostasis. NatA
mutants exhibit specific phenotypes, some of which can be explained by impaired protein-
protein interactions in the absence of correctly acetylated N-termini, with various
consequences: transcriptional alterations caused by defective Sir3-dependent gene silencing
(Wang et al.), impaired function and stability of the Hsp90 chaperonin system and its client
proteins (Oh:2017hx), cellular sorting of secretory proteins, functions of the Golgi apparatus
and the actin cytoskeleton and targeting of specific proteins for degradation (summarized in
(Aksnes et al.)). It is easy to imagine that a multitude of individual effects can challenge
proteostasis regulation that then demands for a higher activity of the UPS in order to remove
damage: mistargeted proteins, misfolded proteins, mis-expressed proteins and subunits. This
higher UPS activity then could at least partially account for the increased degradation rate of
linear ubiquitin fusion proteins.
Interestingly, our observation that Rpn4 Nt-acetylation enhances the strength of its Nt-degron
provides a hint towards a more direct coupling of NatA and proteostasis regulation. Here we
demonstrate that Nt-acetylation can act independently of E3 ligases to promote ubiquitin-
independent degradation of Rpn4, thereby linking NatA activity to regulation of UPS activity.
Importantly, in this context Nt-acetylation is neither required nor sufficient to trigger degradation
of Rpn4, but rather accelerates degradation of this already unstable protein. Although
abundance and half-life of Rpn4 were increased in NatA-deficient cells, we did not observe
clearly increased activity of proteasomal subunit promoters (S4 Fig). This could be explained
by the relatively weak effect of NatA on Rpn4 degradation and abundance, and it is consistent
with the previous report that the abundance of proteasomal subunits was not significantly
increased even when the N-terminal degron of Rpn4 was completely removed (Wang et al.,
2004a), and that it showed only a modest increase in response to expression of a non-
degradable Rpn4 variant lacking both degrons (Wang et al., 2010). Since promoters of E3
ligases involved in UFD appear to lack obvious Rpn4 binding motifs (Shirozu et al., 2015) it is
unlikely that the increased abundance of Tom1, Ubr4 (Fig. 5b) and Ubr1 (Oh et al., 2017) E3
ligases in the naa10Δ mutant is mediated by Rpn4. It can be imagined that load-dependent
inhibition of autoubiquitination regulates E3 abundance, as shown for other E3 ligases (P de
Bie, 2011). Alternatively, Rpn4-independent NatA-mediated regulation of E3 expression is
possible.
We furthermore show that degradation of UFD substrates is accelerated in NatA-deficient cells
and subsequently identify the E3 ligase Tom1 as a novel E4 chain elongating enzyme of this
pathway. This function of Tom1 is clearly distinct from its previously recognized roles as an E3
ligase that is sufficient for ubiquitination of substrate proteins (Sung et al., 2016) and its E3-
independent function in ribosome-associated quality control (Defenouillère et al., 2013). While
no endogenous substrates of the UFD pathway are known in yeast, the pathway is conserved
in mammalian cells, where several functions have been identified. UBB+1, a mutant ubiquitin
variant with a short C-terminal extension caused by a frameshift mutation, is a substrate of the
mammalian UFD pathway (Park et al., 2009) and has been linked to neurodegenerative
disorders (van Leeuwen et al., 1998). The cell cycle regulator p21 (Bloom et al., 2003), the
ERK3 MAP kinase (Coulombe et al., 2004), and the Arf tumor suppressor (Kuo et al., 2004)
were shown to be degraded following N-terminal ubiquitination. It was recently demonstrated
that HUWE1, the mammalian ortholog of Tom1, can ubiquitinate MyoD, the first known UFD
substrate, on its N-terminal residue (Noy et al., 2012). Given the conserved nature of UFD and
its components, we speculate that Tom1 can generate endogenous UFD substrates in yeast.
Altogether, our results complement the knowledge about the role of NatA dependent Nt-
terminal acetylation and how this is coupled to the activity of the UPS. We believe that our
work will contribute to a better understanding of this protein modification and its functions.
Materials and Methods
Yeast genome manipulations
Yeast gene deletions and promoter duplications were performed by PCR targeting, as
described (Huber et al., 2014; Janke et al., 2004). Seamless genome editing was performed
using the 50:50 technique (Horecka and Davis, 2014). Yeast strains and plasmids used in this
study are listed in Tables S1 and S2, respectively.
tFT-based protein stability measurements with flow cytometry (tFT assay)
Yeast strains containing the desired plasmids were inoculated into 200 µl SC medium lacking
the appropriate amino acids for plasmid selection and grown to saturation in 96-well plates.
The cultures were then diluted into fresh medium by pinning to a new 96-well plate using a
RoToR pinning robot (Singer Instruments) and incubated at 23°C for 20-24 h to 1x106-8x106
cells/ml. Flow cytometry was performed on a FACSCanto RUO HTS flow cytometer (BD
Biosciences) equipped with a high-throughput sample loader, a 561 nm laser with 600 nm long
pass and 610/20 nm band pass filters for mCherry, and a 488 nm laser with 505 nm long pass
and 530/30 nm band pass filters for sfGFP. Data analysis was performed in R (R Core Team,
2016) with the flowCore and flowWorkspace packages using a custom script. Briefly, the
events were gated for mCherry- and sfGFP-positive cells, the median intensity of a negative
control was subtracted from each channel, and the mCherry/sfGFP ratio was calculated for
each cell. The median mCherry/sfGFP ratio of each sample was used for further analysis.
Unless otherwise stated, each experiment was performed using two biological replicates with
three technical replicates each. To account for growth rate differences, sample mCherry/sfGFP
ratios were normalized to the stable Ubi-TH-eK-tFT reporter (plasmid pAnB19-TH, Table S2),
which was measured in each strain background.
Flow cytometry of promoter duplications
Yeast cells were inoculated into 200 µl SC medium and grown to saturation in 96-well plates.
The cultures were then diluted into fresh medium by pinning to a new 96-well plate using a
RoToR pinning robot (Singer Instruments) and incubated at 23°C for 20-24 h to 1x106-8x106
cells/ml. Flow cytometry was performed on a FACSCanto RUO HTS flow cytometer (BD
Biosciences) equipped with a high-throughput sample loader, a 561 nm laser with 600 nm long
pass and 610/20 nm band pass filters for mCherry, and a 488 nm laser with 505 nm long pass
and 530/30 nm band pass filters for sfGFP. Data analysis was performed in R (R Core Team,
2016) with the flowCore and flowWorkspace packages using a custom script. Briefly, the
events were gated for single cells using forward and side scatter pulse width, followed by gating
for fluorescent cells. The median intensity of a negative control was subtracted from each cell.
The median sfGFP intensity of each sample was used for further analysis. Unless otherwise
stated, each experiment was performed using two biological replicates with three technical
replicates each.
Cycloheximide chases
Cells were grown at 23°C to 6x106-1x107 cells/ml in synthetic medium before addition of
cycloheximide (Sigma Aldrich, 100 mg/ml stock in 100% ethanol) to 100 µg/ml final
concentration. At each time point, 1 ml of the culture was removed, mixed with 150 µl 1.85 M
NaOH and 10 µl 2-mercaptoethanol and flash-frozen in liquid nitrogen. Protein extracts were
prepared as described (Knop et al., 1999), followed by SDS-PAGE and immunoblotting.
For Ubi-P-tFT constructs, membranes were probed with rabbit anti-GFP (ab6556, Abcam) and
mouse anti-Pgk1 (22C5D8, Molecular Probes) antibodies. A secondary donkey anti-mouse
antibody coupled to IRDye800 (610-732-002, biomol, Rockland) or donkey anti-rabbit coupled
to Alexa 680 (A10043, life technologies) were used for detection on an Odyssey infrared
imaging system (Li-Cor).
For Rpn4-TAP strains, membranes were probed with rabbit peroxidase-anti-peroxidase (PAP)
antibodies (Z0113, Dako) and imaged on an LAS-4000 system (Fuji), followed by probing with
mouse anti-Pgk1 (22C5D8, Molecular Probes) and goat anti-mouse HRP (115-035-003,
Dianova) antibodies and imaging. Quantification was performed in ImageJ (Schneider et al.,
2012).
For HA-tagged Ufd4, membranes were probed with mouse anti-HA (12CA5) and mouse anti-
Pgk1 (22C5D8, Molecular Probes), followed by probing with mouse anti-Pgk1 (22C5D8,
Molecular Probes) and imaging on a LAS-4000 system (Fuji).
Tom1 abundance
Cells expressing protein A-tagged Tom1 were grown at 23°C to 6x106-1x107 cells/ml in
synthetic medium and samples were taken and cell extracts were prepared as described (Knop
et al., 1999). Following SDS PAGE and Western blotting, membranes were probed with rabbit
peroxidase-anti-peroxidase (PAP) antibodies (Z0113, Dako) and imaged on an LAS-4000
system (Fuji), followed by probing with rabbit anti-Fas (Egner et al., 1993) and goat anti-rabbit
HRP (111-035-003, Dianova) antibodies and imaging. Quantification was performed in ImageJ
(Schneider et al., 2012).
Rpn4 mass spectrometry
pdr5Δ ubr2Δ yeast cells expressing transcriptionally inactive Rpn4C477A mutants (Wang et
al, 2004) C-terminally tagged with 10xHis-sfGFPcp8 (Khmelinskii et al., 2016)from a GPD
promoter were grown in SC-His to 7x106 – 8x106 cells/ml. Bortezomib was added to 50 µM
and cultures were incubated for 1 h. 2.5x109 cells were harvested by centrifugation, washed
with 20% (w/v) trichloroacetic acid, and stored at -80°C. Cell pellets were resuspended in 1600
µl 20% (w/v) trichloroacetic acid and lysed with 0.5 mm glass beads (Sigma) in a FastPrep
FP120 (Thermo) for 8x 40 s at 6.5 m/s. After precipitation, proteins were washed with cold
acetone, air-dried and resuspended in 3 ml purification buffer (6M guanidium chloride, 100 mM
Tris-HCl pH 9.0, 300 mM NaCl, 10 mM imidazole, 0.2 % (v/v) Triton X-100). DTT was added
to 10 mM and samples were incubated at 60 °C for 30 min, followed by quenching with 100
mM chloroacetamide at RT for 60 min. Lysates were clarified at 21 000 g, 4 °C for 45 min and
the supernatants incubated with TALON beads (Clontech) pre-equlibrated with purification
buffer at RT over night with overhead rotation followed by washing with wash buffer (8M urea,
100 mM sodium phosphate pH 7.0, 300 mM NaCl, 5 mM imidazole, 5 mM chloroacetamide,
0.2% (v/v) Triton X-100) without (twice) and with 0.2%(w/v) SDS (twice). Rpn4 was eluted in
30 µl elution buffer (8M urea, 100 mM sodium phosphate pH 7.0, 300 mM NaCl, 500 mM
imidazole, 5 mM chloroacetamide, 0.2% (v/v) Triton X-100). 7 µl of eluate were used for SDS-
PAGE followed by Coomassie Brilliant Blue staining. Bands of the expected size were excised,
digested with trypsin, and analyzed with ESI LC-MS/MS on a Q-Exactive HF (Thermo
Scientific) coupled with Dionex Ultimate 3000 RSLCnano (Thermo Scientific). Mass
spectrometry was performed at the ZMBH core facility for mass spectrometry and proteomics.
Ubiquitin pulldowns
Ubiquitinated proteins were purified from yeast cells expressing N-terminally 10xHis-tagged
ubiquitin using a protocol adapted from (Khmelinskii et al., 2014). Yeast were grown in SC-
His/Leu to 7x106 – 8x106 cells/ml. Approx. 1x109 cells were harvested by centrifugation,
washed with cold H¬2O, and stored at -80 °C. Cell pellets were resuspended in 800 µl 20%
(w/v) trichloroacetic acid and lysed with 0.5 mm glass beads (Sigma) in a FastPrep FP120
(Thermo) for 8x 40 s at 6.5 m/s. After precipitation, proteins were washed with cold acetone,
air-dried, resuspended in 1.5 ml purification buffer (6M guanidium chloride, 100 mM Tris-HCl
pH 9.0, 300 mM NaCl, 10 mM imidazole, 5 mM chloroacetamide, 0.2 % (v/v) Triton X-100),
and clarified at 21 000 g, 4 °C for 45 min. Protein concentration was determined with Bradford
assay (BioRad) in purification buffer diluted 1:10 with H2O. 1% of the amount of protein to be
used for purification was removed, precipitated with 150 µl 20% (w/v) trichloroacetic acid, and
resuspended in 100 µl HU buffer (8 M Urea, 5% (w/v) SDS, 200 mM sodium phosphate pH
7.0, 1 mM EDTA, 15 mg/ml DTT) to be used as total extract. Equal amounts of protein were
incubated with TALON beads (Clontech) pre-equilibrated with purification buffer at RT for 1.5
h with overhead rotation, followed by washing with wash buffer (8M urea, 100 mM sodium
phosphate pH 7.0, 300 mM NaCl, 5 mM imidazole, 5 mM chloroacetamide, 0.2% (v/v) Triton
X-100) without (twice) and with 0.2%(w/v) SDS (twice). Ubiquitin conjugates were eluted in 50
µl elution buffer (8M urea, 100 mM sodium phosphate pH 7.0, 300 mM NaCl, 500 mM
imidazole, 5 mM chloroacetamide, 0.2% (v/v) Triton X-100) and analyzed by SDS-PAGE on
4-12% NuPAGE Bis-Tris gradient gels (Invitrogen) followed by immunoblotting. After probing
with rabbit anti-GFP (ab6556, Abcam) and rabbit anti-Zwf1 (Miller et al., 2015) followed by goat
anti-rabbit IgG-HRP (#111-035-003, Dianova) and imaging on an LAS-4000 system (Fuji),
membranes were stripped (100 mM glycine, 2% (w/v) SDS, pH 2.0) and re-probed with mouse
anti-ubiquitin (P4G7) followed by goat anti-mouse IgG-HRP (#115-035-003, Dianova) and
imaging.
LC-MS Analysis of ubiquitin linkages
SDS-PAGE gels of in vitro ubiquitination products (Fig. 4C,D) were stained using Coomassie
and from each lane the regions corresponding to the polyubiquitinated species were cut out
and processed as described with minor modifications (Fecher-Trost et al., 2013). In brief, after
reduction with dithiothreitol and alkylation with iodoacetamide, trypsin digestion was done
overnight at 37°C. The reaction was quenched by addition of 20 µL of 0.1% trifluoroacetic acid
(TFA; Biosolve, Valkenswaard, The Netherlands) and the supernatant was dried in a vacuum
concentrator before LC-MS analysis. Nanoflow LC-MS2 analysis was performed with an
Ultimate 3000 liquid chromatography system coupled to an QExactive HF mass spectrometer
(Thermo-Fischer, Bremen, Germany). Samples were dissolved in 0.1% TFA and injected to a
self-packed analytical column (75um x 200mm; ReproSil Pur 120 C18-AQ; Dr Maisch GmbH)
and eluted with a flow rate of 300nl/min in an acetonitrile-gradient (3%-40%). The mass
spectrometer was operated in data-dependent acquisition mode, automatically switching
between MS and MS2. Collision induced dissociation MS2 spectra were generated for up to
20 precursors with normalized collision energy of 29 %.
Database search - Raw files were processed using Proteome Discoverer 2.3. (Thermo
Scientific) for peptide identification and quantification. MS2 spectra were searched with the
SEQUEST software (Thermo Scientific) against the Uniprot yeast database (6910 entries) and
the contaminants database (MaxQuant version 1.5.3.30 (Cox and Mann, 2008) with the
following parameters: Carbamidomethylation of cysteine residues as fixed modification and
Acetyl (Protein N-term), Oxidation (M), deamidation (NQ) and GG signature for ubiquitination
(K) as variable modifications, trypsin/P as the proteolytic enzyme with up to 2 missed
cleavages. Peptide identifications were accepted if they could be established at greater than
95,0% probability by the Peptide Prophet algorithm (Keller et al., 2002). Protein identifications
were accepted if they could be established at greater than 95,0% probability and contained at
least 2 identified peptides. Protein probabilities were assigned by the Protein Prophet
algorithm (Alexey et al., 2003). Scaffold (version Scaffold_4.8.4, Proteome Software Inc.,
Portland, Oregon) was used to validate and visualize MS/MS based peptide and protein
identifications. For graphic presentation of XICs retention times were aligned and exported as
.csv files using FreeStyle (Thermo Scientific).
Tom1 co-immunoprecipitation
Yeast strains expressing the desired construct were grown to 7x106 – 8x106 cells/ml. 1x109
were harvested by centrifugation, washed with cold H2O, and stored at -80°C. GFP fusions
were immunoprecipitated using lab-purified GFP binding protein (GBP) (Rothbauer et al.,
2008) coupled to NHS-activated Sepharose FastFlow beads (GE Healthcare) using a protocol
adapted from (Babiano et al., 2012). Cell pellets were resuspended in 200 µl cold lysis buffer
(50 mM Tris-HCl pH 7.4, 150 mM CH3COOK, 5 mM EDTA, 5 mM EGTA, 0.2 % Triton X-100)
with protease inhibitors (2x Roche Complete EDTAfree, 5 mM benzamidine, 5 mM Pefabloc
SC, 5 mM 1,10-phenanthroline, 25 mM N-ethylmaleimide) and lysed with 0.5 mm glass beads
(Sigma) in a FastPrep FP120 for 6x 20 s at 6.5 m/s. Lysates were clarified at 21 000 g for 30
min and the supernatants incubated for 2 h at 4 °C with overhead rotation together with 40 µl
GBP-beads previously equilibrated by washing 3 times with 1 ml lysis buffer. The beads were
washed 3 times with lysis buffer and eluted in 50 µl HU buffer (8 M Urea, 5% (w/v) SDS, 200
mM sodium phosphate pH 7.0, 1 mM EDTA, 15 mg/ml DTT). Samples were analyzed by SDS-
PAGE followed by immunoblotting with rabbit peroxidase-anti-peroxidase (PAP) antibodies
(Z0113, Dako) or rabbit anti-GFP (ab6556, Abcam) and rabbit anti-Zwf1 (Miller et al., 2015)
followed by goat anti-rabbit IgG-HRP (#111-035-003, Dianova) and imaging on an LAS-4000
system (Fuji).
In vitro ubiquitination assays
6xHis-Rad6, 6xHis-Ubc4, Ubi-ProtA-6xHis, and UbiK48R-ProtA-6xHis were expressed in
E.coli BL21(DE3) pRIL and purified over a pre-packed HisTrap FastFlow column (GE
Healthcare). FLAG-Ufd4, FLAG-Ubr1, FLAG-Ufd2, and FLAG-Tom1 were overexpressed in
yeast from a GPD promoter and purified as described (Hwang et al., 2009; Hwang and
Varshavsky, 2008). Purified yeast Uba1 and ubiquitin were purchased from BostonBiochem
(#E-300 and #U-100SC, respectively). Final protein concentrations were 100 nM (Uba1), 80
µM (ubiquitin), 1 µM (Rad6), 1 µM (Ubc4), 200 nM (Ubr1), 200 nM (Ufd4), 200 nM (Ufd2), 200
nM (Tom1), 125 nM (Ubi-ProtA), 125 nM (UbiK48R-ProtA), in 20 µl reactions containing 4 mM
ATP (#1191, Merck), 150 mM NaCl, 5 mM MgCl2, 1 mM DTT, 50 mM HEPES (pH 7.5). All
reactions contained Uba1, ubiquitin, Rad6, and Ubc4. Reactions were pipetted on ice,
incubated at 30 °C for 30 min, quenched by addition of 20 µl 5x SDS sample buffer (50% (v/v)
glycerol, 10% (w/v) SDS, 250 mM Tris-HCl pH 6.8, 62.5 mM EDTA, 5% (v/v) ß-
mercaptoethanol) and incubation at 95 °C for 5 min, and analyzed using 4-12% NuPAGE Bis-
Tris gradient gels (Invitrogen) followed by immunoblotting with rabbit peroxidase-anti-
peroxidase (PAP) antibodies (Z0113, Dako) and imaging on a LAS-4000 system (Fuji).
Fluorescence microscopy
Yeast strains were grown in SC medium at 23°C to ~8x106 cells/ml. Control strains not
expressing fluorescent proteins and Tom1-GFP strains additionally expressing mCherry from
a constitutive promoter were mixed 1:1 and attached to glass-bottom 96-well plates (MGB096-
1-2-LG-L, Matrical) as described (Khmelinskii and Knop, 2014). Image stacks were acquired
on a Nikon Ti-E wide field epifluorescence microscope with a 60x ApoTIRF oil immersion
objective (1.49 NA, Nikon), an LED light source (SpectraX, Lumencor), a Flash4 sCMOS
camera (Hamamatsu). Segmentation was performed in the bright-field channel using CellX
(Mayer et al., 2013). Flat-field correction was performed using a reference image derived from
a well containing recombinant mCherry-sfGFP fusion protein and average fluorescence across
the stack was calculated for each cell. Cells were classified as autofluorescence control or
sample separately for each field of view by fitting a 2-component Gaussian mixture model to
the mCherry intensity values and assigning each cell to the class with the higher posterior
probability. GFP intensity of all control cells within a field of view was averaged and subtracted
from the sample GFP intensities.
Acknowledgments
We acknowledge the support of Bernd Heßling (ZMBH Mass Spectrometry facility) and Monika
Langlotz (ZMBH Flow Cytometry and FACS facility). We thank Birgit Besenbeck for technical
support, Daniel Kirrmaier for support with fluorescence microscopy, Frauke Melchior and
Marius Lemberg for critically reading the manuscript. Work was supported by the Deutsche
Forschungsgemeinschaft (SFB1036) to MK and an MSc/PhD fellowship from the HBIGS
graduate school to IK.
( )
Author contributions
MK conceived the project. MK, IK, MKs, AK, and TR designed the experiments and discussed
the results. IK, MKs, LA, and TR performed the experiments. MK and IK wrote the manuscript.
Conflict of interest
The authors declare no conflict of interest.
Supplementary information
Table S1: Yeast strains
Strain
Background
Genotype
Reference
FY1679
S288c
MATa/α ura3-52/ura3-52 leu2Δ1/LEU2 his3Δ200/HIS3 trp1Δ63/TRP1
GAL2/GAL2
[71]
ESM356-1
FY1679
MATa ura3-52 leu2Δ1 his3Δ200 trp1Δ63
Elmar Schiebel
YCT1084
ESM356-1
ubr1Δ::hphNT1
[72]
YMaM632
ESM356-1
naa20Δ::hphNT1
[12]
YBB4
ESM356-1
ufd4Δ::hphNT1
[12]
YBB5
ESM356-1
naa10Δ::hphNT1
[12]
YBB9
ESM356-1
ufd4Δ::natNT2 ubr1Δ::hphNT1
[12]
YBB52
ESM356-1
UFD4-3HA-kanMX6
This study
YBB53
ESM356-1
naa10Δ::hphNT1 UFD4-3HA-kanMX6
This study
YEO2
ESM356-1
naa10Δ::kanMX6 ubr1Δ::hphNT1
This study
YEO3
ESM356-1
naa10Δ::kanMX6 ufd4Δ::natNT2 ubr1Δ::hphNT1
[12]
YIK35
ESM356-1
naa10Δ::kanMX6 ufd4Δ::hphNT1
This study
YIK55
ESM356-1
ufd2Δ::klURA3
This study
YIK56
ESM356-1
ufd2Δ::klURA3 ufd4Δ::natNT2 ubr1Δ::hphNT1
This study
YIK241
ESM356-1
ubr2Δ::klUra3 ufd4Δ::natNT2 ubr1Δ::hphNT1
This study
YIK242
ESM356-1
ubr2Δ::klUra3 ufd4Δ::natNT2 ubr1Δ::hphNT1 naa10Δ::kanMX6
This study
YIK278
ESM356-1
tom1Δ::klUra3
This study
YIK280
ESM356-1
tom1Δ::klURA3 ufd4Δ::natNT2 ubr1Δ::hphNT1
This study
YIK281
ESM356-1
tom1Δ::klURA3 naa10Δ::kanMX6 ufd4Δ::natNT2 ubr1Δ::hphNT1
This study
YIK292
ESM356-1
ufd2Δ::natNT2 tom1Δ::klURA3
This study
YIK300
ESM356-1
tom1Δ::natNT2 ubr1Δ::hphNT1
This study
YIK301
ESM356-1
tom1Δ::natNT2 ufd4Δ::hphNT1
This study
YIK305
ESM356-1
TOM1-TAP-kanMX4
This study
YIK309
ESM356-1
RPN4-TAP-kanMX4
This study
YIK311
ESM356-1
naa10Δ::natNT2 RPN4-TAP-kanMX4
This study
YIK330
ESM356-1
pep4Δ0
This study
YIK343
ESM356-1
natNT2-pGPD-FLAG-UFD4 pep4Δ0
This study
YIK344
ESM356-1
natNT2-pGPD-FLAG-TOM1 pep4Δ0
This study
YIK345
ESM356-1
natNT2-pGPD-FLAG-UBR1 pep4Δ0
This study
YIK346
ESM356-1
natNT2-pGPD-FLAG-UFD2 pep4Δ0
This study
YIK358
ESM356-1
pPRE4-sfGFP-KanMX-pPRE4-PRE4
This study
YIK359
ESM356-1
pPRE5-sfGFP-KanMX-pPRE5-PRE5
This study
YIK360
ESM356-1
pPRE6-sfGFP-KanMX-pPRE6-PRE6
This study
YIK361
ESM356-1
pRPT3-sfGFP-KanMX-pRPT3-RPT3
This study
YIK362
ESM356-1
pRPT5-sfGFP-KanMX-pRPT5-RPT5
This study
YIK363
ESM356-1
pPUP1-sfGFP-KanMX-pPUP1-PUP1
This study
YIK364
ESM356-1
pTUB1-sfGFP-KanMX-pTUB1-TUB1
This study
YIK366
ESM356-1
pRPB2-sfGFP-KanMX-pRPB2-RPB2
This study
YIK367
ESM356-1
naa10Δ::natNT2 pPRE4-sfGFP-KanMX-pPRE4-PRE4
This study
YIK368
ESM356-1
naa10Δ::natNT2 pPRE5-sfGFP-KanMX-pPRE5-PRE5
This study
YIK369
ESM356-1
naa10Δ::natNT2 pPRE6-sfGFP-KanMX-pPRE6-PRE6
This study
YIK370
ESM356-1
naa10Δ::natNT2 pRPT3-sfGFP-KanMX-pRPT3-RPT3
This study
YIK371
ESM356-1
naa10Δ::natNT2 pRPT5-sfGFP-KanMX-pRPT5-RPT5
This study
YIK372
ESM356-1
naa10Δ::natNT2 pPUP1-sfGFP-KanMX-pPUP1-PUP1
This study
YIK373
ESM356-1
naa10Δ::natNT2 pTUB1-sfGFP-KanMX-pTUB1-TUB1
This study
YIK375
ESM356-1
naa10Δ::natNT2 pRPB2-sfGFP-KanMX-pRPB2-RPB2
This study
YIK385
ESM356-1
pPRE4-sfGFP-KanMX-pPRE4-PRE4 Rpn4A2N
This study
YIK386
ESM356-1
pPRE5-sfGFP-KanMX-pPRE5-PRE5 Rpn4A2N
This study
YIK387
ESM356-1
pPRE6-sfGFP-KanMX-pPRE6-PRE6 Rpn4A2N
This study
YIK388
ESM356-1
pRPT3-sfGFP-KanMX-pRPT3-RPT3 Rpn4A2N
This study
YIK389
ESM356-1
pRPT5-sfGFP-KanMX-pRPT5-RPT5 Rpn4A2N
This study
YIK390
ESM356-1
pPUP1-sfGFP-KanMX-pPUP1-PUP1 Rpn4A2N
This study
YIK391
ESM356-1
pTUB1-sfGFP-KanMX-pTUB1-TUB1 Rpn4A2N
This study
Strain
Background
Genotype
Reference
YIK393
ESM356-1
pRPB2-sfGFP-KanMX-pRPB2-RPB2 Rpn4A2N
This study
YIK398
ESM356-1
naa10Δ::natNT2 pPRE4-sfGFP-KanMX-pPRE4-PRE4 Rpn4A2N
This study
YIK399
ESM356-1
naa10Δ::natNT2 pPRE5-sfGFP-KanMX-pPRE5-PRE5 Rpn4A2N
This study
YIK400
ESM356-1
naa10Δ::natNT2 pPRE6-sfGFP-KanMX-pPRE6-PRE6 Rpn4A2N
This study
YIK401
ESM356-1
naa10Δ::natNT2 pRPT3-sfGFP-KanMX-pRPT3-RPT3 Rpn4A2N
This study
YIK402
ESM356-1
naa10Δ::natNT2 pRPT5-sfGFP-KanMX-pRPT5-RPT5 Rpn4A2N
This study
YIK403
ESM356-1
naa10Δ::natNT2 pPUP1-sfGFP-KanMX-pPUP1-PUP1 Rpn4A2N
This study
YIK404
ESM356-1
naa10Δ::natNT2 pTUB1-sfGFP-KanMX-pTUB1-TUB1 Rpn4A2N
This study
YIK406
ESM356-1
naa10Δ::natNT2 pRPB2-sfGFP-KanMX-pRPB2-RPB2 Rpn4A2N
This study
YIK414
ESM356-1
ubr1Δ::klTrp1 ufd4Δ::hphNT1 Rpn4A2N
This study
YIK415
ESM356-1
ubr1Δ::klTrp1 ufd4Δ::hphNT1 naa10Δ::natNT2 Rpn4A2N
This study
YIK423
ESM356-1
tom1Δ::kanMX6 ufd4Δ::hphNT1 ubr1Δ::klTrp1 Rpn4A2N
This study
YIK424
ESM356-1
tom1Δ::kanMX6 naa10Δ::natNT2 ufd4Δ::hphNT1 ubr1Δ::klTrp1 Rpn4A2N
This study
YIK427
ESM356-1
Rpn4Δ(211-229)-TAP-kanMX4
This study
YIK428
ESM356-1
Rpn4A2NΔ(211-229)-TAP-kanMX4
This study
YIK429
ESM356-1
naa10Δ::hphNT1 Rpn4Δ(211-229)-TAP-kanMX4
This study
YIK430
ESM356-1
naa10Δ::hphNT1 Rpn4A2NΔ(211-229)-TAP-kanMX4
This study
YIK431
ESM356-1
naa20Δ::natNT2 Rpn4Δ(211-229)-TAP-kanMX4
This study
YIK432
ESM356-1
naa20Δ::natNT2 Rpn4A2NΔ(211-229)-TAP-kanMX4
This study
YIK460
ESM356-1
ubr1Δ::kanMX6 ufd4Δ::natNT2 Rpn4Δ(211-229)
This study
YIK462
ESM356-1
ubr1Δ::kanMX6 ufd4Δ::natNT2 naa10Δ::hphNT1 Rpn4Δ(211-229)
This study
YIK464
ESM356-1
ubr2Δ::klUra3 ubr1Δ::kanMX6 ufd4Δ::natNT2 Rpn4Δ(211-229)
This study
YIK466
ESM356-1
ubr2Δ::klUra3 ubr1Δ::kanMX6 ufd4Δ::natNT2 naa10Δ::hphNT1 Rpn4Δ(211-
229)
This study
YIK469
ESM356-1
pdr5Δ::kanMX6 ubr2Δ::natNT2
This study
YIK470
ESM356-1
pdr5Δ::kanMX6 ubr2Δ::natNT2 naa10Δ::hphNT1
This study
YIK471
ESM356-1
pdr5Δ::kanMX6 ubr2Δ::natNT2 naa20Δ::hphNT1
This study
YIK476
ESM356-1
pIK117 in YIK469
This study
YIK477
ESM356-1
pIK118 in YIK469
This study
YIK478
ESM356-1
pIK117 in YIK470
This study
YIK479
ESM356-1
pIK118 in YIK470
This study
YIK480
ESM356-1
pIK117 in YIK471
This study
YIK481
ESM356-1
pIK117 in YIK471
This study
YIK585
ESM356-1
natNT2-pGPD-UFD4
This study
YIK586
ESM356-1
natNT2-pGPD-TOM1
This study
YIK587
ESM356-1
natNT2-pGPD-UBR1
This study
YIK619
ESM356-1
naa10Δ::natNT2 TOM1-TAP-kanMX4
This study
YIK644
ESM356-1
leu2Δ::pGPD-mCherry-tCYC1-hphNT1 TOM1-sfGFP-kanMX
This study
YIK645
ESM356-1
leu2Δ::pGPD-mCherry-tCYC1-hphNT1 naa10Δ::natNT2 TOM1-sfGFP-kanMX
This study
Table S2: Plasmids
Plasmid
Description
Reference
pET28c
E. coli expression vector
Novagen
pFA6a-kanMX6
template for gene deletion by PCR targeting with kanMX6 selection
marker
(Wach et al., 1994)
pFA6a-hphNT1
template for gene deletion by PCR targeting with hphNT1 selection
marker
(Janke et al., 2004)
pFA6a-natNT2
template for gene deletion by PCR targeting with natNT2 selection marker
(Janke et al., 2004)
pYM13
Template for C-terminal tagging with TAP-tag by PCR targeting with
kanMX6 selection marker
(Janke et al., 2004)
pYM23
Template for C-terminal tagging with 3Myc-tag by PCR targeting with
klTrp1 selection marker
(Janke et al., 2004)
pRS413
CEN ARS HIS3
(Sikorski and Hieter)
p413-GPD
CEN ARS HIS3 pGPD-tCYC1
(Mumberg et al.)
pArd1
pRS416-NAA10
Ulrike Friedrich
pGR295
p415-TEF-10xHis-Ubi
(Khmelinskii et al.,
2014)
pAnB19
pRS413-pGPD-Ubi-EcoRV-STOP-eK-mCherry-sfGFP
(Kats et al., 2018)
pAnB19-PP
pRS413-pGPD-Ubi-PP-eK-mCherry-sfGFP
(Kats et al., 2018)
pAnB19-EH
pRS413-pGPD-Ubi-EH-eK-mCherry-sfGFP
(Kats et al., 2018)
pAnB19-UbiG76V
pRS413-pGPD-UbiG76V-eK-mCherry-sfGFP
(Kats et al., 2018)
pIK35
pFA6a-klUra3
(Kats et al., 2018)
pIK41
pRS413-pGPD-UbiG76V-eK-mCherry-sfGFPcp8
This study
pIK45
pRS413-pGPD-UbiK48R,G76V-eK-mCherry-sfGFP
This study
pIK57
pRS413-pGPD-Rpn4(1-80)-eK-mCherry-sfGFP
This study
pIK59
template for pGPD-driven overexpression and N-terminal tagging with
FLAG-tag by PCR targeting with natNT2 selection marker
This study
pIK66
pRS413-pGPD-Rpn4A2N(1-80) -eK-mCherry-sfGFP
This study
pIK78
6xHis-Ubc4 in pET28c
This study
pIK79
6xHis-Rad6 in pET28c
This study
pIK100
Ubi-ProtA-6xHis in pET28c
This study
pIK102
UbiK48R-ProtA-6xHis in pET28c
This study
pIK117
p413-GPD-Rpn4C477A-10xHis-sfGFPcp8
This study
pIK118
p413-GPD-Rpn4A2NC477A-10xHis-sfGFPcp8
This study
Fig. S1. Degradation of UbiG76V-tFT.
Degradation of UbiG76V-tFT after blocking translation with cycloheximide. Whole-cell extracts
were separated by SDS-PAGE followed by immunoblotting with antibodies against GFP and
Zwf1 as loading control. A product of mCherry hydrolysis during cell extract preparation (Gross
et al., 2000) is marked (∗).
S2 Fig. Degradation of Rpn4 variants.
Cyclohexiide chase analysis of the degradation of Rpn4 variants. Whole-cell extracts were
separated by SDS-PAGE followed by immunoblotting with antibodies against protein A and
Pgk1 as loading control.
(A) Rpn4Δ(211-229) lacking the ubiquitin-dependent degron. Representative immunoblot from
Fig 2F.
(B) Rpn4A2N,Δ(211-229) lacking the ubiquitin-dependent degron and acetylated by NatB
instead of NatA. Representative immunoblot from Fig 2F.
S3 Fig. E3 ligases in cells lacking NatA.
(A) Degradation of Ufd4 after blocking translation with cycloheximide. Whole-cell extracts were
separated by SDS-PAGE followed by immunoblotting with antibodies against HA and Pgk1 as
loading control. Representative immunoblot from Fig 5A. Time point 0 was used for
quantification.
(B) Abundance of Tom1. Whole-cell extracts were separated by SDS-PAGE followed by
immunoblotting with antibodies against protein A and Fas as loading control. Representative
immunoblot from Fig 5A.
(C) Abundance of Tom1. Live-cell imaging of strains carrying C-terminally GFP-tagged Tom1
was performed and the fluorescence intensity was quantified. p: Mann-Whitney U-test.
S4 Fig: Influence of NatA and Rpn4 on activity of proteasomal promoters.
Promoters of the indicated genes were duplicated while simultaneously inserting a sfGFP
coding sequence, such that expression of sfGFP is driven by the second copy of the promoter.
Fluorescence intensity was measured by flow cytometry and normalized to wild type cells.
Mean intensity and 95% CI of six replicates are plotted together with the intensity of each
replicate. The TUB1 and RPB2 promoters served as Rpn4-independent controls.
References
Aksnes, H., A. Drazic, M. Marie, and T. Arnesen. 2016. First Things First: Vital Protein Marks
by N-Terminal Acetyltransferases. Trends Biochem. Sci. 41:746-760.
Alexey, I.N., K. Andrew, a. Eugene Kolker, and R. Aebersold. 2003. A Statistical Model for
Identifying Proteins by Tandem Mass Spectrometry. Anal. Chem. 75:4646-4658.
Amerik, A.Y., and M. Hochstrasser. 2004. Mechanism and function of deubiquitinating
enzymes. Biochimica et Biophysica Acta (BBA) - Molecular Cell Research. 1695:189-
207.
Babiano, R., M. Gamalinda, J.L. Woolford, and J. de la Cruz. 2012. Saccharomyces cerevisiae
ribosomal protein L26 is not essential for ribosome assembly and function. Mol. Cell.
Biol. 32:3228-3241.
Bachmair, A., D. Finley, and A. Varshavsky. 1986. In vivo half-life of a protein is a function of
its amino-terminal residue. Science. 234:179-186.
Bachmair, A., and A. Varshavsky. 1989. The degradation signal in a short-lived protein. Cell.
56:1019-1032.
Bloom, J., V. Amador, F. Bartolini, G. DeMartino, and M. Pagano. 2003. Proteasome-mediated
degradation of p21 via N-terminal ubiquitinylation. Cell. 115:71-82.
Caron, C., C. Boyault, and S. Khochbin. 2005. Regulatory cross-talk between lysine acetylation
and ubiquitination: role in the control of protein stability. Bioessays. 27:408-415.
Coulombe, P., G. Rodier, E. Bonneil, P. Thibault, and S. Meloche. 2004. N-Terminal
ubiquitination of extracellular signal-regulated kinase 3 and p21 directs their
degradation by the proteasome. Mol. Cell. Biol. 24:6140-6150.
Cox, J., and M. Mann. 2008. MaxQuant enables high peptide identification rates, individualized
p.p.b.-range mass accuracies and proteome-wide protein quantification. Nat.
Biotechnol. 26:1367-1372.
Dantuma, N.P., K. Lindsten, R. Glas, M. Jellne, and M.G. Masucci. 2000. Short-lived green
fluorescent proteins for quantifying ubiquitin/proteasome-dependent proteolysis in
living cells. Nat. Biotechnol. 18:538-543.
Defenouillère, Q., Y. Yao, J. Mouaikel, A. Namane, A. Galopier, L. Decourty, A. Doyen, C.
Malabat, C. Saveanu, A. Jacquier, and M. Fromont-Racine. 2013. Cdc48-associated
complex bound to 60S particles is required for the clearance of aberrant translation
products. Proc. Natl. Acad. Sci. U.S.A. 110:5046-5051.
Egner, R., M. Thumm, M. Straub, A. Simeon, H.J. Schüller, and D.H. Wolf. 1993. Tracing
intracellular proteolytic pathways. Proteolysis of fatty acid synthase and other
cytoplasmic proteins in the yeast Saccharomyces cerevisiae. J. Biol. Chem.
268:27269-27276.
Fecher-Trost, C., U. Wissenbach, A. Beck, P. Schalkowsky, C. Stoerger, J. Doerr, A. Dembek,
M. Simon-Thomas, A. Weber, P. Wollenberg, T. Ruppert, R. Middendorff, H.H. Maurer,
and V. Flockerzi. 2013. The in vivo TRPV6 protein starts at a non-AUG triplet, decoded
as methionine, upstream of canonical initiation at AUG. Journal of Biological Chemistry.
288:16629-16644.
Fiil, B.K., R.B. Damgaard, S.A. Wagner, K. Keusekotten, M. Fritsch, S. Bekker-Jensen, N.
Mailand, C. Choudhary, D. Komander, and M. Gyrd-Hansen. 2013. OTULIN Restricts
Met1-Linked Ubiquitination to Control Innate Immune Signaling. Mol. Cell. 50:818-830.
Finley, D., H.D. Ulrich, T. Sommer, and P. Kaiser. 2012. The ubiquitin-proteasome system of
Saccharomyces cerevisiae. Genetics. 192:319-360.
Friedrich, U.A., M. Zedan, B. Hessling, K. Fenzl, L. Gillet, J.D. Barry, M. Knop, G. Kramer, and
B. Bukau. Nα-terminal acetylation of proteins by NatA and NatB serves distinct
physiological roles in Saccharomyces cerevisiae. biorxiv.org.
Gawron, D., E. Ndah, K. Gevaert, and P. Van Damme. 2016. Positional proteomics reveals
differences in N-terminal proteoform stability. Mol. Syst. Biol. 12:858.
Gerlach, B., S.M. Cordier, A.C. Schmukle, C.H. Emmerich, E. Rieser, T.L. Haas, A.I. Webb,
J.A. Rickard, H. Anderton, W.W.L. Wong, U. Nachbur, L. Gangoda, U. Warnken, A.W.
Purcell, J. Silke, and H. Walczak. 2011. Linear ubiquitination prevents inflammation
and regulates immune signalling. Nature. 471:591-596.
Gross, L.A., G.S. Baird, R.C. Hoffman, K.K. Baldridge, and R.Y. Tsien. 2000. The structure of
the chromophore within DsRed, a red fluorescent protein from coral. Proc. Natl. Acad.
Sci. U.S.A. 97:11990-11995.
Ha, S.-W., D. Ju, and Y. Xie. 2012. The N-terminal domain of Rpn4 serves as a portable
ubiquitin-independent degron and is recognized by specific 19S RP subunits. Biochem.
Biophys. Res. Commun. 419:226-231.
Hahn, J.S., D.W. Neef, and D.J. Thiele. 2006. A stress regulatory network for co-ordinated
activation of proteasome expression mediated by yeast heat shock transcription factor.
Mol. Microbiol. 60:240-251.
Hershko, A., and A. Ciechanover. 1998. The ubiquitin system. Annu. Rev. Biochem. 67:425-
479.
Hershko, A., H. Heller, E. Eytan, G. Kaklij, and I.A. Rose. 1984. Role of the alpha-amino group
of protein in ubiquitin-mediated protein breakdown. Proc. Natl. Acad. Sci. U.S.A.
81:7021-7025.
Horecka, J., and R.W. Davis. 2014. The 50:50 method for PCR-based seamless genome
editing in yeast. Yeast. 31:103-112.
Huber, F., M. Meurer, D. Bunina, I. Kats, C.I. Maeder, M. Stefl, C. Mongis, and M. Knop. 2014.
PCR Duplication: A One-Step Cloning-Free Method to Generate Duplicated
Chromosomal Loci and Interference-Free Expression Reporters in Yeast. 9:e114590.
Hwang, C.-S., A. Shemorry, D. Auerbach, and A. Varshavsky. 2010a. The N-end rule pathway
is mediated by a complex of the RING-type Ubr1 and HECT-type Ufd4 ubiquitin ligases.
Nat. Cell Biol. 12:1177-1185.
Hwang, C.-S., A. Shemorry, and A. Varshavsky. 2009. Two proteolytic pathways regulate DNA
repair by cotargeting the Mgt1 alkylguanine transferase. Proceedings of the National
Academy of Sciences. 106:2142-2147.
Hwang, C.-S., A. Shemorry, and A. Varshavsky. 2010b. N-terminal acetylation of cellular
proteins creates specific degradation signals. Science. 327:973-977.
Hwang, C.-S., and A. Varshavsky. 2008. Regulation of peptide import through phosphorylation
of Ubr1, the ubiquitin ligase of the N-end rule pathway. Proc. Natl. Acad. Sci. U.S.A.
105:19188-19193.
Janke, C., M.M. Magiera, N. Rathfelder, C. Taxis, S. Reber, H. Maekawa, A. Moreno-Borchart,
G. Doenges, E. Schwob, E. Schiebel, and M. Knop. 2004. A versatile toolbox for PCR-
based tagging of yeast genes: new fluorescent proteins, more markers and promoter
substitution cassettes. Yeast. 21:947-962.
Johnson, E.S., B. Bartel, W. Seufert, and A. Varshavsky. 1992. Ubiquitin as a degradation
signal. EMBO J. 11:497-505.
Johnson, E.S., P.C. Ma, I.M. Ota, and A. Varshavsky. 1995. A proteolytic pathway that
recognizes ubiquitin as a degradation signal. J. Biol. Chem. 270:17442-17456.
Ju, D., L. Wang, X. Mao, and Y. Xie. 2004. Homeostatic regulation of the proteasome via an
Rpn4-dependent feedback circuit. Biochem. Biophys. Res. Commun. 321:51-57.
Ju, D., and Y. Xie. 2004. Proteasomal degradation of RPN4 via two distinct mechanisms,
ubiquitin-dependent and-independent. Journal of Biological Chemistry. 279:23851-
23854.
Ju, D., and Y. Xie. 2006. Identification of the preferential ubiquitination site and ubiquitin-
dependent degradation signal of Rpn4. J. Biol. Chem. 281:10657-10662.
Kats, I., A. Khmelinskii, M. Kschonsak, F. Huber, R.A. Knieß, A. Bartosik, and M. Knop. 2018.
Mapping Degradation Signals and Pathways in a Eukaryotic N-terminome. Mol. Cell.
70:488-501.e485.
Keller, A., A.I. Nesvizhskii, E. Kolker, and R. Aebersold. 2002. Empirical Statistical Model To
Estimate the Accuracy of Peptide Identifications Made by MS/MS and Database
Search. Anal. Chem. 74:5383-5392.
Khmelinskii, A., E. Blaszczak, M. Pantazopoulou, B. Fischer, D.J. Omnus, G. Le Dez, A.
Brossard, A. Gunnarsson, J.D. Barry, M. Meurer, D. Kirrmaier, C. Boone, W. Huber, G.
Rabut, P.O. Ljungdahl, and M. Knop. 2014. Protein quality control at the inner nuclear
membrane. Nature. 516:410-413.
Khmelinskii, A., P.J. Keller, A. Bartosik, M. Meurer, J.D. Barry, B.R. Mardin, A. Kaufmann, S.
Trautmann, M. Wachsmuth, G. Pereira, W. Huber, E. Schiebel, and M. Knop. 2012.
Tandem fluorescent protein timers for in vivo analysis of protein dynamics. Nat.
Biotechnol. 30:708-714.
Khmelinskii, A., and M. Knop. 2014. Analysis of protein dynamics with tandem fluorescent
protein timers. Methods Mol. Biol. 1174:195-210.
Khmelinskii, A., M. Meurer, C.-T. Ho, B. Besenbeck, J. Füller, M.K. Lemberg, B. Bukau, A.
Mogk, and M. Knop. 2016. Incomplete proteasomal degradation of green fluorescent
proteins in the context of tandem fluorescent protein timers. Mol. Biol. Cell. 27:360-370.
Kim, D.-H., and D.M. Koepp. 2012. Hect E3 ubiquitin ligase Tom1 controls Dia2 degradation
during the cell cycle. Mol. Biol. Cell. 23:4203-4211.
Kim, D.-H., W. Zhang, and D.M. Koepp. 2012. The Hect domain E3 ligase Tom1 and the F-
box protein Dia2 control Cdc6 degradation in G1 phase. Journal of Biological
Chemistry. 287:44212-44220.
Knop, M., K. Siegers, G. Pereira, W. Zachariae, B. Winsor, K. Nasmyth, and E. Schiebel. 1999.
Epitope tagging of yeast genes using a PCR-based strategy: more tags and improved
practical routines. Yeast. 15:963-972.
Koegl, M., T. Hoppe, S. Schlenker, H.D. Ulrich, T.U. Mayer, and S. Jentsch. 1999. A novel
ubiquitination factor, E4, is involved in multiubiquitin chain assembly. Cell. 96:635-644.
Kuo, M.-L., W. den Besten, D. Bertwistle, M.F. Roussel, and C.J. Sherr. 2004. N-terminal
polyubiquitination and degradation of the Arf tumor suppressor. Genes Dev. 18:1862-
1874.
Linster, E., I. Stephan, W.V. Bienvenut, J. Maple-Grødem, L.M. Myklebust, M. Huber, M.
Reichelt, C. Sticht, S.G. Møller, T. Meinnel, T. Arnesen, C. Giglione, R. Hell, and M.
Wirtz. 2015. Downregulation of N-terminal acetylation triggers ABA-mediated drought
responses in Arabidopsis. Nat Commun. 6:7640.
Liu, C., W. Liu, Y. Ye, and W. Li. 2017. Ufd2p synthesizes branched ubiquitin chains to promote
the degradation of substrates modified with atypical chains. Nat Commun. 8:14274.
Mannhaupt, G., R. Schnall, V. Karpov, I. Vetter, and H. Feldmann. 1999. Rpn4p acts as a
transcription factor by binding to PACE, a nonamer box found upstream of 26S
proteasomal and other genes in yeast. FEBS Lett. 450:27-34.
Mayer, C., S. Dimopoulos, F. Rudolf, and J. Stelling. 2013. Using CellX to quantify intracellular
events. Curr Protoc Mol Biol. Chapter 14:Unit 14.22.
McDowell, G.S., and A. Philpott. 2013. Non-canonical ubiquitylation: mechanisms and
consequences. Int. J. Biochem. Cell Biol. 45:1833-1842.
Michel, M.A., K.N. Swatek, M.K. Hospenthal, and D. Komander. 2017. Ubiquitin Linkage-
Specific Affimers Reveal Insights into K6-Linked Ubiquitin Signaling. Mol. Cell. 68:233-
246.e235.
Miller, S.B., C.T. Ho, J. Winkler, M. Khokhrina, A. Neuner, M.Y. Mohamed, D.L. Guilbride, K.
Richter, M. Lisby, E. Schiebel, A. Mogk, and B. Bukau. 2015. Compartment-specific
aggregases direct distinct nuclear and cytoplasmic aggregate deposition. EMBO J.
34:778-797.
Mumberg, D., R. Müller, and M. Funk. 1995. Yeast vectors for the controlled expression of
heterologous proteins in different genetic backgrounds. Gene. 156:119-122.
Nguyen, K.T., S.-H. Mun, C.-S. Lee, and C.-S. Hwang. 2018. Control of protein degradation
by N-terminal acetylation and the N-end rule pathway. Exp. Mol. Med. 50:91.
Noy, T., O. Suad, D. Taglicht, and A. Ciechanover. 2012. HUWE1 ubiquitinates MyoD and
targets it for proteasomal degradation. Biochem. Biophys. Res. Commun. 418:408-
413.
Oh, J.-H., J.-Y. Hyun, and A. Varshavsky. 2017. Control of Hsp90 chaperone and its clients by
N-terminal acetylation and the N-end rule pathway. Proceedings of the National
Academy of Sciences. 157:201705898.
P de Bie, A.C. 2011. Ubiquitination of E3 ligases: self-regulation of the ubiquitin system via
proteolytic and non-proteolytic mechanisms. Cell Death Differ. 18:1393-1402.
Park, Y., S.K. Yoon, and J.-B. Yoon. 2009. The HECT domain of TRIP12 ubiquitinates
substrates of the ubiquitin fusion degradation pathway. J. Biol. Chem. 284:1540-1549.
Poulsen, E.G., C. Steinhauer, M. Lees, A.-M. Lauridsen, L. Ellgaard, and R. Hartmann-
Petersen. 2012. HUWE1 and TRIP12 collaborate in degradation of ubiquitin-fusion
proteins and misframed ubiquitin. PLoS ONE. 7:e50548.
Rao, H., F. Uhlmann, K. Nasmyth, and A. Varshavsky. 2001. Degradation of a cohesin subunit
by the N-end rule pathway is essential for chromosome stability. Nature. 410:955-959.
Ravid, T., and M. Hochstrasser. 2008. Diversity of degradation signals in the ubiquitin-
proteasome system. Nat. Rev. Mol. Cell Biol. 9:679-690.
Rittinger, K., and F. Ikeda. 2017. Linear ubiquitin chains: enzymes, mechanisms and biology.
Open Biology. 7:170026.
Rothbauer, U., K. Zolghadr, S. Muyldermans, A. Schepers, M.C. Cardoso, and H. Leonhardt.
2008. A Versatile Nanotrap for Biochemical and Functional Studies with Fluorescent
Fusion Proteins. Molecular \& Cellular Proteomics. 7:282-289.
Scaglione, K.M., V. Basrur, N.S. Ashraf, J.R. Konen, K.S.J. Elenitoba-Johnson, S.V. Todi, and
H.L. Paulson. 2013. The Ubiquitin-conjugating Enzyme (E2) Ube2w Ubiquitinates the
N Terminus of Substrates. J. Biol. Chem. 288:18784-18788.
Schneider, C.A., W.S. Rasband, and K.W. Eliceiri. 2012. NIH Image to ImageJ: 25 years of
image analysis. Nat Methods. 9:671-675.
Shemorry, A., C.-S. Hwang, and A. Varshavsky. 2013. Control of protein quality and
stoichiometries by N-terminal acetylation and the N-end rule pathway. Mol. Cell.
50:540-551.
Shirozu, R., H. Yashiroda, and S. Murata. 2015. Identification of minimum Rpn4-responsive
elements in genes related to proteasome functions. FEBS Lett. 589:933-940.
Sikorski, R.S., and P. Hieter. 1989. A system of shuttle vectors and yeast host strains designed
for efficient manipulation of DNA in Saccharomyces cerevisiae. Genetics. 122:19-27.
Singh, R.K., M. Gonzalez, M.H. Kabbaj, and A. Gunjan. 2012. Novel E3 ubiquitin ligases that
regulate histone protein levels in the budding yeast Saccharomyces cerevisiae. PLoS
ONE. 7:e36295.
Stack, J.H., M. Whitney, S.M. Rodems, and B.A. Pollok. 2000. A ubiquitin-based tagging
system for controlled modulation of protein stability. Nat. Biotechnol. 18:1298-1302.
Starheim, K.K., K. Gevaert, and T. Arnesen. 2012. Protein N-terminal acetyltransferases: when
the start matters. Trends Biochem. Sci. 37:152-161.
Sung, M.-K., T.R. Porras-Yakushi, J.M. Reitsma, F.M. Huber, M.J. Sweredoski, A. Hoelz, S.
Hess, and R.J. Deshaies. 2016. A conserved quality-control pathway that mediates
degradation of unassembled ribosomal proteins. eLife. 5:3429.
Tatham, M.H., A. Plechanovová, E.G. Jaffray, H. Salmen, and R.T. Hay. 2013. Ube2W
conjugates ubiquitin to α-amino groups of protein N-termini. Biochemical Journal.
453:137-145.
Tokunaga, F., S.-i. Sakata, Y. Saeki, Y. Satomi, T. Kirisako, K. Kamei, T. Nakagawa, M. Kato,
S. Murata, S. Yamaoka, M. Yamamoto, S. Akira, T. Takao, K. Tanaka, and K. Iwai.
2009. Involvement of linear polyubiquitylation of NEMO in NF-κB activation. Nat. Cell
Biol. 11:123-132.
Tsuchiya, H., K. Tanaka, and Y. Saeki. 2013. The parallel reaction monitoring method
contributes to a highly sensitive polyubiquitin chain quantification. Biochem. Biophys.
Res. Commun. 436:223-229.
van Leeuwen, F.W., D.P.V. de Kleijn, H.H. van den Hurk, A. Neubauer, M.A.F. Sonnemans,
J.A. Sluijs, S. Köycü, R.D.J. Ramdjielal, A. Salehi, G.J.M. Martens, F.G. Grosveld,
J.P.H. Burbach, and E.M. Hol. 1998. Frameshift Mutants of β Amyloid Precursor
Protein and Ubiquitin-B in Alzheimer's and Down Patients. Science. 279:242-247.
Varland, S., H. Aksnes, F. Kryuchkov, F. Impens, D. Van Haver, V. Jonckheere, M. Ziegler, K.
Gevaert, P. Van Damme, and T. Arnesen. 2018. N-terminal Acetylation Levels Are
Maintained During Acetyl-CoA Deficiency in Saccharomyces cerevisiae. Mol. Cell
Proteomics. 17:2309-2323.
Varshavsky, A. 2019. N-degron and C-degron pathways of protein degradation. Proceedings
of the National Academy of Sciences. 116:358-366.
Vittal, V., L. Shi, D.M. Wenzel, K.M. Scaglione, E.D. Duncan, V. Basrur, K.S.J. Elenitoba-
Johnson, D. Baker, H.L. Paulson, P.S. Brzovic, and R.E. Klevit. 2014. Intrinsic disorder
drives N-terminal ubiquitination by Ube2w. Nat. Chem. Biol. 11:83-89.
Wach, A., A. Brachat, R. Pöhlmann, and P. Philippsen. 1994. New heterologous modules for
classical or PCR-based gene disruptions in Saccharomyces cerevisiae. Yeast.
10:1793-1808.
Wang, L., X. Mao, D. Ju, and Y. Xie. 2004a. Rpn4 is a physiological substrate of the Ubr2
ubiquitin ligase. J. Biol. Chem. 279:55218-55223.
Wang, X., J.J. Connelly, C.-L. Wang, and R. Sternglanz. 2004b. Importance of the Sir3 N
Terminus and Its Acetylation for Yeast Transcriptional Silencing. Genetics. 168:547-
551.
Wang, X., H. Xu, S.W. Ha, D. Ju, and Y. Xie. 2010. Proteasomal Degradation of Rpn4 in
Saccharomyces cerevisiae Is Critical for Cell Viability Under Stressed Conditions.
Genetics. 184:335-342.
Xie, Y., and A. Varshavsky. 2001. RPN4 is a ligand, substrate, and transcriptional regulator of
the 26S proteasome: a negative feedback circuit. Proc. Natl. Acad. Sci. U.S.A.
98:3056-3061.
Yau, R.G., K. Doerner, E.R. Castellanos, D.L. Haakonsen, A. Werner, N. Wang, X.W. Yang,
N. Martinez-Martin, M.L. Matsumoto, V.M. Dixit, and M. Rape. 2017. Assembly and
Function of Heterotypic Ubiquitin Chains in Cell-Cycle and Protein Quality Control. Cell.
171:918-933.e920.
Zattas, D., D.J. Adle, E.M. Rubenstein, and M. Hochstrasser. 2013. N-terminal acetylation of
the yeast Derlin Der1 is essential for Hrd1 ubiquitin-ligase activity toward luminal ER
substrates. Mol. Biol. Cell. 24:890-900.
| 2020 | Mechanisms of up-regulation of Ubiquitin-Proteasome activity in the absence of NatA dependent N-terminal acetylation | 10.1101/2020.03.23.003053 | [
"Kats Ilia",
"Kschonsak Marc",
"Khmelinskii Anton",
"Armbruster Laura",
"Ruppert Thomas",
"Knop Michael"
] | null |
1
Myomegalin regulates Hedgehog pathway by controlling PDE4D at the
centrosome
Hualing Peng1*, Jingyi Zhang1*, Amanda Ya1,3, Winston Ma1, Sammy Villa1, Shahar
Sukenik2, Xuecai Ge1**
1Department of Molecular and Cell Biology, University of California, Merced, Merced,
CA 95340
2Department of Chemistry and Chemical Biology, University of California, Merced,
Merced, CA 95340
3Current address: Molecular and Cell Biology Graduate Program at Dartmouth College,
Hanover, NH 03755
* These authors contribute equally to this work
** Correspondence: xge2@ucmerced.edu
Running Head: Control Hedgehog pathway at the centrosome
Total number of characters: 19,135
2
Abstract
Mutations in the Hedgehog (Hh) signaling are implicated in birth defects and cancers,
including medulloblastoma, one of the most malignant pediatric brain tumors. Current Hh
inhibitors face the challenge of drug resistance and tumor relapse, urging new insights in the Hh
pathway regulation. Our previous study revealed how PDE4D controls global levels of cAMP in
the cytoplasm to positively regulate Hh signaling; in the present study we found that a specific
isoform PDE4D3 is tethered to the centrosome by myomegalin, a centrosome/Golgi associated
protein. Myomegalin loss dislocates PDE4D3 from the centrosome, leading to local PKA over-
activation and inhibition of the Hh signaling, leaving other PKA-related pathways unaffected.
Myomegalin loss suppresses the proliferation of granule neuron precursors, and blocks the
growth of medulloblastoma in mouse model. Our findings specify a new regulatory mechanism of
the Hh pathway, and highlight an exciting therapeutic avenue for Hh-related cancers with reduced
side effects.
Introduction
The Hedgehog (Hh) pathway is widely implicated in birth defects and human
tumors(Briscoe and Therond, 2013). One of the Hh-related tumors is medulloblastoma, a
malignant pediatric brain tumor(Goodrich et al., 1997). Current treatment of medulloblastoma,
surgery removal followed by chemo- or radiotherapy, brings devastating side effects to the young
patients(Fouladi et al., 2005); while the available Hh-pathway inhibitor targeting Smoothened
(Smo) is challenged by drug resistance and tumor relapse(Yauch et al., 2009). Therefore, new
approaches to inhibit Hh signaling are needed. The Hh signal transduction involves a series of
protein transport into and out of the primary cilium, and eventually converges on the regulation of
Gli transcription factors(Hui and Angers, 2011). Without the ligand Sonic Hedgehog (Shh), the
receptor Patched (Ptch) resides in the cilium and prevents the cilium translocation and activation
of Smo. Upon Shh stimulation, Ptch exits the cilium, followed by Smo’s accumulation and
activation in the cilium. The signaling cascade ultimately activates the transcription activator Gli2,
and eliminates the transcription suppressor Gli3R, a proteolytic product from the Gli3 full length
(FL) protein(Wang and Li, 2006; Han and Alvarez-Buylla, 2010). The activated Hh signaling
quickly induces the transcription of Gli1, an amplifier of Hh signaling, forming a positive feedback
loop.
PKA plays a central role in Hh signaling activation and Gli regulations. PKA
phosphorylates Gli3, which primes its further phosphorylation by GSK and CK1. The
phosphorylated Gli3 was recognized by the ubiquitin proteosome system that cleaves Gli3FL into
Gli3R(Wang and Li, 2006). In addition, PKA also controls Gli2 activation. Within the cell, PKA
concentrates at the centrosome (cilium base) where it controls the cilium translocation of Gli2, a
step required for Gli2 activation(Tuson, He and Anderson, 2011). Genetic removal of PKA leads
to full activation of Hh pathway in the developing neural tube(Epstein et al., 1996; Huang, Roelink
and McKnight, 2002; Tuson, He and Anderson, 2011), further substantiating the strong inhibitory
effect of PKA on Hh signaling. Recent studies from Mukhopadhyay lab suggest that inhibiting the
cAMP-PKA levels in the cilium markedly activates Hh signaling in a manner independent of Smo
activation (Somatilaka et al., 2020). Conversely, pharmacological activation of PKA inhibits Hh
signaling and suppresses Hh-related tumor growth(Yamanaka et al., 2010, 2011). However, PKA
is widely involved in many signaling and metabolic pathways; ubiquitous activation of PKA
inevitably impacts all signaling pathways. Hence, treatments directly targeting PKA are not
practical due to their severe side effects. To avoid these side effects, one feasible strategy is to
selectively control PKA activities at the specific subcellular sites where it regulates Hh signaling,
leaving other pathways unaffected.
3
It is known that PKA activity in the cell is compartmentalized by forming complexes that
include cAMP-specific phosphodiesterase (PDE) (Zaccolo and Pozzan, 2002; Houslay, 2010;
McCormick and Baillie, 2014). In specific compartments, PKA activity is precisely regulated by
PDE. In our previous studies, we found that PDE4D, recruited to the cytoplasmic membrane by
sema3-Neuropilin signaling, governs cAMP levels in the entire cell to regulate Hh signaling(Tyler
Hillman et al., 2011; Ge et al., 2015). Our results were corroborated by Williams et al. who
independently discovered PDE4D as a positive regulator of the Hh pathway in a chemical
screen(Williams et al., 2015). Since PKA at the centrosome directly participate in Hh
signaling(Barzi et al., 2010; Tuson, He and Anderson, 2011), can we selectively manipulate
PDE4D activity at the centrosome to control local PKA activity? In the current study, we found an
approach to dislocate PDE4D3 from the centrosome; the subsequent elevation in local PKA
activity suppresses Hh signal transduction and Hh-related tumor growth. Our results highlight an
exciting avenue to treat Hh-related cancers with reduced side effects.
Results
Myomegalin (Mmg) interacts with PDE4D3 at the centrosome
To identify an effective approach of selectively modulating cAMP levels at the centrosome,
we did a literature search on the subcellular localization of all cAMP-specific phosphodiesterases.
We found that one PDE4D isoform, PDE4D3 was reported to interact with Mmg, a protein
associated with the centrosome/Golgi(Verde et al., 2001). But it remains unclear whether
PDE4D3 localizes to the centrosome and whether it is involved in the regulation of the Hh
signaling. To answer these questions, we first validated the Mmg-PDE4D3 interaction. Mmg is a
large protein of 270KD, and the full-length protein is not effectively expressed in cells. But in the
previous study the C-terminus of Mmg was identified to mediate its interaction with
PDE4D3(Verde et al., 2001) (Fig. 1A). We thus fused this domain (Mmg-C) with Flag and
expressed it together with HA-PDE4D3 in the cell. We then performed co-immunoprecipitation
assay with Flag- and HA- conjugated magnetic beads, and found that two proteins co-
immunoprecipitated each other (Fig. 1B). It is noteworthy that although Flag antibody pulled down
significant amount of Flag-Mmg (red triangle in Fig. 1B), it only coimmunoprecipitated small
amount of HA-PDE4D3 (red star in Fig. 1B), presumably because only a small fraction of HA-
PDE4D3 in the cell is interacting with Mmg. This is consistent with what we observed in
immunostaining results in Fig. 1E.
To validate the subcellular localization of Mmg, we stained NIH3T3 cells with the Mmg
antibody. As reported before(Roubin et al., 2013), Mmg immunofluorescence significantly
overlaps with pericentrin, a marker of the centrioles and pericentriolar material (Fig. 1C). The non-
centrosomal Mmg signal may represent its localization to Golgi. Mmg-C exhibits similar
localization pattern when expressed in NIH3T3 cells (Fig. 1D). We then expressed both HA-
PDE4D3 and Flag-Mmg-C in the cell. HA-PDE4D3 overlaps with Flag-Mmg at the
centrosome/Golgi area, although a significant fraction of HA-PDE4D3 also diffusively distributes
to the cytosol (Fig. 1E). Taken together, these results suggest that Mmg may recruit a small
fraction of PDE4D3 from the cytosol to the centrosome.
Mmg loss impairs Hh signal transduction and dislocates PDE4D3 from the centrosome
Next, we tested whether eliminating PDE4D3 from the centrosome impacts the Hh
pathway. We silenced Mmg expression with shRNA in NIH3T3 cells, a cell line that contains all
components of the Hh pathways and is commonly used to study Hh signaling transduction. Two
of the five tested shRNAs significantly reduced the transcript and the protein levels of Mmg (Fig.
S1A-B). We then treated cells with SAG, a small molecule agonist of the Hh pathway, and
assessed the Hh pathway activation with qPCR measuring the transcript level of the Hh target
4
gene Gli1. Mmg shRNA significantly reduced SAG-induced Gli1 expression, indicating that Hh
signal transduction was impaired (Fig. S1C).
To thoroughly eliminate Mmg protein expression, we employed CRISPR/Cas9 to knockout
Mmg in mouse embryonic fibroblasts (MEFs). We choose MEF because it transduces Hh
signaling but has lower ploidy level than NIH3T3 cells. We used two gRNAs targeting the 1st exon
of Mmg, and transfected the plasmid containing the two gRNAs and Cas9 into MEF cells, together
with EGFP. Single cell clones were isolated via flow cytometry and expanded (Fig. 2A). We
obtained two cell clones (#7, #10) of Mmg knockout (KO). Both clones appear normal in cell
morphology and cell proliferate (data not shown), and the Mmg mRNA and protein levels are
undetectable (Fig. 2B-C). Interestingly, among the 4 alternative splicing isoforms of mouse Mmg
(https://www.ncbi.nlm.nih.gov/gene/83679), CRISPR/Cas9 abolished the expression of the longer
isoforms (~270KD) and spared the shorter isoform (~130KD) (Fig. 2C), presumably because the
shorter isoform uses an alternative transcription starting point. The shorter isoform, however, does
not interact with PDE4D3 as it lacks the C-terminus. To identify the INDEL mutations induced by
CRISPR/Cas9, We amplified exon 1 and its flanking region with PCR from Mmg KO cells, and
sequenced individual PCR products. The sequencing results show that 3 type of mutations were
generated in each clone, resulting in frameshift that eventually leads to nonsense-mediated
mRNA decay (Fig. S2A-C).
To determine the Hh signaling in Mmg KO clones, we stimulated cells with SAG, and
detected Gli1 expression with qPCR and western blot. SAG-induced Gli1 expression was
dramatically reduced at the transcript and protein levels in both Mmg KO cell clones (Fig. 2D-E).
These results suggest a blockage of Hh transduction after Mmg loss.
Next, we determined the impact of Mmg loss on PDE4D3 localization at the centrosome.
Due to the high similarity between PDE4D isoforms, the antibody specific to PDE4D3 is
unavailable. Therefore, we expressed very low levels of EGFP-PDE4D3 in Mmg KO cells to mimic
the endogenous protein. As expected, in wild type cells PDE4D3 shows significant overlap with
pericentrin, in addition to its diffusive localization to other subcellular sites (Fig. 2F). However, in
Mmg KO cells, the intensity of PDE4D3 at the centrosome is significantly reduced (Fig. 2F-G).
Therefore, without Mmg, PDE4D3 is dislocated from the centrosome.
In summary, our data suggests that loss of Mmg markedly suppresses Hh signal
transduction, and dislocates PDE4D3 from the centrosome.
Mmg loss selectively increases local PKA activity at the centrosome, and blocked further
PKA activation by PDE4D inhibitors
Dislocation of PDE4D3 from the centrosome increases local cAMP levels, which may
eventually lead to PKA over-activation. To confirm this, we employed two method to evaluate local
PKA activity at the centrosome. First, to measure the basal levels of active PKA, we stained cells
with an antibody that recognizes active PKA (phosphoPKA T197). This antibody has been used
to evaluate PKA activity in previous studies(Barzi et al., 2010; Tuson, He and Anderson, 2011;
Ge et al., 2015). We highlighted the centrosome and pericentriolar area with pericentrin staining,
and measured the phosphoPKA levels in this area in ImageJ. As expected, the centrosomal active
PKA levels are much higher in Mmg knockout cells compared to that in wild type MEF (Fig. 3A-
B). Further, expressing exogenous Mmg in Mmg KO cells restored active PKA levels to normal
(Fig. 3B). This change of PKA activity, however, is limited locally to the centrosome, since the
overall phosphoPKA levels remain the same in Mmg KO cells (Fig. 3C). In addition, we detected
the phosphorylation levels of CREB, the cytosolic substrate of PKA(Shaywitz and Greenberg,
1999). The phospho-CREB (S133) levels show no difference between wild type and Mmg
5
knockout cells (Fig. 3C). Thus, Mmg loss increased the basal PKA activity selectively at the
centrosome.
Second, we monitored the dynamic PKA activity in live cells with A kinase-activity reporter
(AKAR4), a fluorescence resonance energy transfer (FRET)-based PKA probe developed in Jin
Zhang’s lab(Zhang et al., 2001; Herbst, Allen and Zhang, 2011). In this probe, a FRET pair (CFP
and YFP) is connected by a linker sequence that contains PKA phosphorylation sites and a
phosphoamino acid binding domain (PAABD). PKA phosphorylation induces conformational
changes in the linker, which brings the FRET pair in close proximity to efficiently produce FRET
(Fig. 3D). To target AKAR4 to the centrosome, we fused it with the regulatory subunit of PKA
(PKARIIɑ), a protein predominantly localized to the centrosome(Zhang et al., 2001). As expected,
when RIIɑ-AKAR4 was expressed in MEF, the probe is concentrated at the centrosomal area (Fig.
3E). The centrosome is marked by co-expression of RFP-PACT(Gillingham and Munro, 2000).
The PDE inhibitor IBMX is commonly used to enhance the PKA activity, because it
effectively elevates cAMP levels in the cell (Zhang et al., 2001; Herbst, Allen and Zhang, 2011).
Since PDE4D3 is dislocated from the centrosome in Mmg KO cells and the local PKA activity is
constitutively high, we hypothesize that IBMX’s effect will be masked at the centrosome in Mmg
KO cells (Fig. 3F). To test this hypothesis, we treated cells with IBMX and analyzed FRET locally
at the centrosome. The FRET efficiency was analyzed as the ratio of YFP/CFP, and this ratio at
each time point was normalized to time 0 (Fig. 3G & Fig. S3). In WT cells, IBMX gradually
increased FRET efficiency, peaking at 12 min. In contrast, Mmg knockout significantly dampened
FRET efficiency at all time points (Fig. 3G). Taken together, Mmg loss dislocates PDE4D3 from
the centrosome, thereby promoting the local basal PKA activity that cannot be further elevated by
the PDE inhibitor.
Mmg loss promotes Gli3R production and blocks Gli2 transportation to the cilium tip
The transcription factor Gli2 and Gli3 are PKA substrates in the Hh pathway. After PKA
phosphorylation, Gli3 is proteolytically processed into Gli3R, a transcription repressor (Fig. 4A).
Upon Hh signaling activation, the Gli3 processing ceases and Gli3R levels markedly reduce
(Wang and Li, 2006; Humke et al., 2010; Tukachinsky, Lopez and Salic, 2010; Hui and Angers,
2011). We examined Gli3 processing by western blot. In wild type cells, SAG treatment
significantly reduced Gli3R levels in cell lysates. In contrast, in Mmg CRISPR clones, the Gli3R
levels remain the same after SAG treatment (Fig. 4B). It is likely that without Mmg, the hyperactive
PKA at the centrosome continues to phosphorylate Gli3 to promote Gli3R production even after
SAG treatment. PKA affects the proteolysis of Gli2 only very slightly but more dramatically controls
its accumulation at cilia tips, a step required for Gli2 activation(Barzi et al., 2010; Tuson, He and
Anderson, 2011). We therefore examined the levels of Gli2 at the cilia tips after SAG stimulation.
The Gli2 intensity at the cilium tips in Mmg KO cells was significantly lower, compared to that in
wild type cells (Fig. 4C). When exogenous Mmg was expressed in Knockout clones, the Gli2
levels at the cilium tips was restored (Fig. 4C-D). Therefore, Mmg loss overactivates PKA at the
centrosome, which blocks Gli2 transport and activation in the cilium tip, leading to inhibition of the
Hh signaling.
In summary, our data suggest a model of how Mmg and PDE4D3 at the centrosome
control local PKA activity to regulate the Hh pathway. Without Shh, PKA activity at the centrosome
is high due to high local cAMP levels. The centrosomal cAMP may be produced in the cilium by
proteins such as GPR16(Mukhopadhyay et al., 2013), or diffuses to the centrosome from other
cytosolic areas. After Shh stimulation, GPR161 exits the cilium and stops the cAMP production;
the cAMP from nearby cytosolic areas is degraded by PDE4D3. The inactive PKA at the
centrosome allows Gli2 to be translocated and activated in the cilium tips, and stops Gli3R
production, leading to Hh pathway activation (Fig. 4E). In Mmg KO cells, since PDE4D3 is
6
dislocated from the centrosome, the cAMP diffused from the nearby areas is not effectively
degraded and the PKA activity remains high. This suppresses Gli2 activation and keeps Gli3R
levels high, and subsequently blocks the activation of Hh signaling (Fig. 4F).
Mmg loss blocks cell proliferation in primary cultured granule neuron precursors (GNPs)
In
the
developing
cerebellum,
Shh
is
the
mitogen
that
stimulates
GNP
proliferation(Dahmane and Ruiz i Altaba, 1999; Wallace, 1999; Wechsler-Reya and Scott, 1999).
Overactive Hh signaling leads to GNP over proliferation that eventually results as
medulloblastoma (MB), one of the most malignant pediatric brain tumor(Kool et al., 2012). In situ
hybridization
results
show
that
both
PDE4D3
and
Mmg
(www.informatics.jax.org/image/MGI:5332354,
www.informatics.jax.org/image/MGI:5333985)
are highly expressed in the developing cerebellum(Richter, Jin and Conti, 2005); it is likely that
the mechanism of Mmg-PDE4D3 regulation on Hh pathway applies to the control of GNP
proliferation. To test this hypothesis, we cultured GNPs from P7 mouse neonates in dishes, and
infected GNPs with lentiviral particles expressing shRNA against Mmg (shRNA #99) (Fig. 5A).
GNP proliferation was induced by SAG. After 3 days of primary culture, GNP proliferation was
assessed by BrdU incorporation assay. As expected, Mmg shRNA significantly reduced Mmg
transcript levels, and reduced the rate of BrdU incorporation after pulse labeling (Fig. 5B-D). Hh
signal activity was significantly reduced in Mmg knockdown cells, demonstrated by decreased
transcript levels of Gli1, a Hh target gene (Fig. 5E). In summary, our results suggest that the same
mechanism of Hh signaling regulation by Mmg-PDE4D3 may control GNP proliferation in the
developing cerebellum.
Mmg loss reduced the growth rate of medulloblastoma in mouse model
Next, we assess the effect of Mmg loss on the Hh-related tumor growth in the mouse
model of MB subcutaneous xenograft used in our previous studies(Ge et al., 2015). We employed
MB56, MB tumor cells directly taken from Ptch+/- mouse, the first and well-established MB mouse
model(Goodrich et al., 1997; Purzner et al., 2018). We infected MB56 with lentiviral particles that
express Mmg (shRNA #99) or control shRNA. 2 days after infection, tumor cells were injected
subcutaneously in the hind flank of nude mice. 6 days after injection, tumors size was measured
daily for two weeks (Fig. 5F). We found that Mmg loss significantly slowed tumor growth starting
from day 4 of measurement (Fig. 5G). At the end of the experiment, we evaluated Gli1 levels in
randomly sampled tumors and found that Mmg loss significantly reduced Hh signal activity (Fig.
5H). Thus, knockdown of Mmg suppressed the growth of Hh-related tumors.
Discussion
Genetic removal of PKA leads to full activation of the Hh pathway in the developing neural
tube(Epstein et al., 1996; Tuson, He and Anderson, 2011), suggesting PKA as a strong inhibitor
of the Hh signaling. However, as a multifaceted enzyme, PKA is widely involved in many signaling
and metabolic pathways. Therefore, global inhibition of PKA is not a feasible strategy for treatment.
Our previous study pointed PDE4D as a potential target to inhibit Hh signaling(Ge et al., 2015).
Our current study highlights an effective approach to selectively inhibit PDE4D at one specific
subcellular site. We provide evidence that dislocating PDE4D3 from the centrosome overactivates
PKA locally at the centrosome to inhibits the Hh pathway, while sparing other PKA-related cellular
events.
Cells have evolved two mechanisms to accurately govern local levels of cAMP and PKA
activity: 1) controlling its production by adenylyl cyclase, and 2) managing its degradation by
cAMP-specific PDEs. We believe that activation of Hh signaling involves both mechanisms. The
first mechanism has been shown to be mediates by GPR161 that resides at the cilium. When the
Hh pathway is off, GPR161 activates the Gαs-adenylyl cyclase pathway and keeps the local
7
cAMP levels high(Mukhopadhyay et al., 2013). Upon Shh stimulation, GPR161 exits the cilium
and stops cAMP production(Mukhopadhyay et al., 2013). Synergistic to this mechanism, PDE4D
at the centrosome degrades the cAMP that is diffused from the nearby subcellular areas. The
combined effects of these two mechanisms keep local PKA activities in check to allow the ensuing
Hh signaling events to occur. When PDE4D activity is absent from the centrosome, the local
cAMP concentration fails to reduce to the subthreshold level, even though the GPR161-Gαs-
adenylyl cyclase pathway stops to produce cAMP. As a result, the high PKA levels at the
centrosome suppresses the Hh signal transduction. Our study, for the first time, unmasked the
roles of centrosomal PDE4D in the Hh pathway.
PDE4D is a large protein family comprising more than 12 alternative splicing isoforms in
mammalian cells(Maurice et al., 2014). PDE4D3 was originally identified to bind to Mmg by Verde
et al (2001). Besides PDE4D3, other isoforms could be tethered to the centrosome as well, and
Mmg loss might dislocate all these isoforms from the centrosome. It is also noteworthy that after
adding IBMX, FRET was not completely abolished in Mmg knockout cells, indicating the existence
of other PDE isoforms at the centrosome (Fig. 3G). It will be intriguing for future studies to
delineate the identify of these PDE isoforms and their targeting mechanism to the centrosome.
Our study pointed to an effective method to suppress Hh signaling in cancers. Current Hh
inhibitors target Smo, and these inhibitors are facing challenges of drug resistance and tumor
relapse(Yauch et al., 2009). Since the Mmg-PDE4D-PKA axis acts directly at Gli transcription
factors, downstream of Smo, targeting this axis will be effective for cancers that have developed
resistance to Smo inhibitors. Further, it is known that the basal activity of PDE4D is high, and
most PDE4D small molecular inhibitors act by blocking the catalytic domain of PDE4D(Gavaldà
and Roberts, 2013). These inhibitors block all PDE4D isoforms and are associated with severe
side effects. Our results suggest that we may eliminate PDE4D activity specifically from the
centrosome without blocking its catalytic domain. It pinpointed an effective therapeutic avenue to
treat Hh-related cancers with reduced side effects.
Materials and Methods
Plasmids and generation of Myomegalin knockout CRISPR cell clones
Human PDE4D3 is generously provided by Marco Conti lab at UCSF, and was subcloned to
include HA and EGFP tag. Myomegalin-C is cloned by RT-PCR with a mouse total mRNA library,
and subcloned to included Flag and EGFP tag. The FRET probe AKAR4 was generously provided
by Jin Zhang lab (available in Addgene). pcDNA3-mPKA-RIIα-AKAR4-NES was constructed by
linking mPKA-RIIα with AKAR4-NES. The linker sequence is GGGGSGS. The two gRNA were
designed via the Guide Design Resources of Feng Zhang lab at MIT(https://zlab.bio/guide-design-
resources). The two gRNAs were cloned into the backbone of pX330-U6-Chimeric_BB-CBh-
hSpCas9 (Addgene 42230), and transfected into MEF cells together with EGFP via lipofectamine
2000. 48hr after transfection, EGFP-positive cells were sorted by flow cytometer and plated into
individual wells in 96-well plate. Individual cell clones were cultured for 2-4 weeks, and transferred
to 24-well plate for further expansion.
Time-lapse image with AKAR4
MEF cells were co-transfected with RIIα-AKAR4 and RFP-PACT via electroportation, and
cultured in DMEM supplemented with 10% FBS at 37˚C. 24hr later, cells were plated onto 8-
chambered lab-Tex II coverglass (Thermo Fisher) at a density of 3.5 x 104/well, and then grown
for approximately 24h before imaging.
For imaging, cells were washed once with Extracellular Imaging Buffer (ECB, 5mM KCl, 125mM
NaCl, 1.5mM CaCl2, 1.5mM MgCl2, 10mM Glucose, 20mM HEPES) and kept in ECB in the dark
8
at room temperature. Images were collected with an epiflourescence microscope (Zeiss
Observer3) with a 40X dry 0.9 NA objective lens connected to two linked Hamammatsu Flash
v3 sCMOS cameras to facilitate real-time FRET imaging. The CFP fluorophore was excited
using a 430 nm LED (Colibi7, Zeiss), and emission collected using a triple-bandpass emission
filter, 467/24 + 555/25 + 687/145 (set 91 HE from Zeiss). Downstream, the collected emission
was further split onto the two cameras using a 520 nm dichroic. Exposure time was set for
200ms. Images were acquired every 2min. IBMX was added to the cell as indicated in the
experiment.
FRET analysis
Results were analyzed in ImageJ. The centrosome was identified in red channel via RFP-PACT
and selected as region of interest (ROI). An automated, stack-based thresholding was built on
Renyi entropy method to identify strong fluorescence in the RFP channel throughout the time
course. Intensities of the CFP and YFP at each time point in the ROI were measured. To control
for different expression levels of AKAR, intensity at each time point was normalized to time zero.
Western blot
Cells were lysed on ice in RIPA buffer containing 25mM Tris-HCl (pH7.6), 150mM Nacl, 1% NP-
40, 1% sodium deoxycholate, 0.1% SDS, 1mM PMSF, 10mM sodium fluoride, 2mM sodium
pyrophosphate, 1mM sodium orthovanadate, Roche protease inhibitor cocktail and Roche
PhosSTOP inhibitor cocktail for 30 min. Lysates were cleared with centrifugation at 13,000 rpm
for 30 min at 4℃. Protein concentrations of the supernatants were determined with BCA protein
assay kit (Pierce). Protein samples were boiled in 6x SDS sample buffer for 10 min, and resolved
in SDS-PAGE. Protein binds were transferred to PVDF membrane (88520, Thermofisher), which
were blocked in Tris buffer (PH7.0) containing 0.1% Tween-20 and 5% BSA. The membrane was
incubated in primary antibodies (diluted in blocking buffer) overnight at 4 ℃, and washed 3 times
before incubation with HRP-conjugated secondary antibodies. Protein bands were visualized with
ECL Western Blot substrate (Pierce, 32109).
Primary antibodies used: mouse anti-GAPDH (ab9484, Abcam), rabbit anti-phosphoPKA-T197
(5661S, Cell Siganling), rabbit anti-phosphoCREB-S133 (9198S, Cell Signaling), mouse anti-PKA
(610625, BD Biosciences), rabbit anti-Gli1 (V812, Cell Signaling), rabbit anti-Myomegalin (PA5-
30324, Invitrogen).
Co-Immunoprecipitation
Plasmids were transfection into HEK293Tcells with lipofectamine 2000 reagent (Invitrogen)
according to manufacturer’s instruction. Plasmids used : HA-PDE4D3, 3xFlag-Mmg-C560 and
3xFlag-vector (E4026, Sigma, MO). 24hr after transfection, cells were lysed in ELB buffer
(150mM NaCl, 1% TritonX100, 50mM Tris pH8.0, 5mM EDTA, 5mM NaF, 2mM Na3VO4)
supplemented with Protease inhibitor cocktail (Roche 11836170001) for 30min at 4˚C. Lysates
were cleared by centrifugation at 14,000rpm for 15min. Protein concentration of supernatants
were determined using Pierce BCA Protein Assay Kit (Thermo Scientific). Equal amount of
protein was loaded to the anti-Flag M2 Magnetic beads (M8823, Sigma) and anti-HA Magnetic
beads (88836, Thermo Scientific) and incubated for 1h at room temperature. Beads were
washed according to manufacturer’s instruction and incubated with 2x Laemmli sample buffer at
95˚C for 5min. Samples were loaded to 10% SDS-PAGE gel and western blot was performed.
Antibodies used: HA-tag rabbit antibody (3724S, Cell Signaling), anti-Flag M2 antibody (F1804,
Sigma).
Immunofluorescence staining
NIH3T3 or MEF cells grown on Poly-D-Lysine (A003E, Sigma) coated coverslips were fixed with
4% paraformaldehyde for 10min at room temperature. Cells were then blocked with 2% donkey
9
serum and 0.1% triton in PBS for 1h. Primary and secondary antibodies are incubated with cells
in the blocking buffer. Images were taken with LEICA DMi8 microscopy, or Zeiss LSM880
confocal microscope, with 60x oil lens.
Primary antibodies used: anti-Flag M2 antibody (F1804, Sigma), anti-mouse Pericentrin
(611814, BD Biosciences), anti-rabbit myomegalin antibody (PA552969, Invitrogen), anti-rabbit
GFP antibody (A11122, Thermo Fisher), mouse anti-acetylated tubulin (T6793, SIGMA), goat
anti-Gli2 (AF3635, R&D SYSTEMS), rabbit anti-pPKA (ab59218, Abcam).
Quantification of PhosphoPKA, PDE4D3 and Gli2
The levels of phosphoPKA and EGFP-PDE4D3 in centrosome were measured using ImageJ
software as follows. First, an area of interest (AOI) was delineated based on the signal intensity
of pericentrin staining; second, the mean gray value in AOI was measured in the phosphoPKA or
EGFP-PD4D3 channel (F1); third, the contour of AOI was manually dragged to a nearby region
within the cell, and the mean gray value of the enclosed area was measured as background (F2).
The final values of phosphoPKA and PDE4D3 were calculated as F = F1-F2.
To quantify Gli2 levels at the cilium tips, the contour of the cilium tips was outlines in red channel
(acetylated tubulin staining). The mean gray density in the enclosed area was measure in green
channel (Gli2 staining). The background gray density was measured and subtracted to obtain the
final Gli2 intensity at the cilium tips.
For each condition, 35-60 cells were measured. In myomegalin rescue experiment, only cells that
were transfected with EGFP-Mmg are measured. Data analysis were done with Graphpad Prism
8.0 Software. Kruskal–Wallis non-parametric One-Way ANOVA was used for statistical analysis.
Quantitative PCR
Cells were plated in 6-well plates at 0.5 x 106 cells per well and cultured overnight. For Hh
induction, cells were stimulated with 100nM SAG in starvation medium (0.5% FBS in DMEM) for
20-24hr. Total RNAs were isolated with Trizol reagent. The concentration of total RNA was
normalized, and the same amount of RNA was mixed with qScript XLT-1 Step, RT-qPCR
ToughMix (Quantabio 66149433,) together with specific TaqMan expression assays. The real-
time PCR is performed in QuantStudio 3 (ThermoFisher).
The following TaqMan gene expression probes used: Mm00494654_g1 (Gli1), Mm00626240_m1
and Mm01257004_m1 (Mmg), Mm99999915_g1 (GAPDH).
Primary culture of GNPs and Brdu incorporation assay
Neonate CD1 mouse were sacrificed at P7. The cerebellum was taken out and cut into small
pieces with razor blades, incubated at 37˚C for 15min in digestion buffer (HBSS with 20mM
HEPES, PH 7.3, supplemented with trypsin and DNase I). At the end of incubation, digestion
buffer was aspirated and replaced with Neurobasal medium with 250U/ml DNase I. Tissues were
then triturated with pipet tips and polished Pasteur pipettes. After seated for 2min, dissociated
cells were collected from the upper layer and centrifuged at 1000rpm for 5 min. Cells were
washed one time, and resuspended in Neurobasal, supplemented with B27 (17504044, Gibco),
Glutamax, and 1% Penicillin/streptomycin. Cells were then plated on coverslips coated with Poly-
D-Lysine and Laminin. Lentivirus expressing Mmg shRNA were added 5hr after plating, and
incubated overnight. 48hr after plating, 10nM SAG was added to the cells and incubated overnight.
Brdu (20μM) was added to the culture and pulse labeled for 4hr, after which cells are fixed with
4% paraformaldehyde. Brdu immunostaining were performed with mouse-anti-Brdu antibody
(662411Ig, Proteintech).
10
Medulloblastoma xenograft mouse model
All the in vivo surgery steps and treatments were performed in accordance with the animal
protocols approved by UC Merced’s Institutional Biosafety Committee (IACUC). MB56 tumor
cells were cultured in neurobasal medium supplemented with B-27 (21103-049, Thermofisher).
Cells were infected with lentiviral particles of shRNA against Myomegalin or control shRNA.
48hr later, cells were then collected, centrifuged and resuspended in PBS at 2 x 107cells per
50ul. 50ul cells were mixed with Matrigel (354234, ThermoFisher) at 1:1 volume ration. The
100ul mixture was slowly injected into hind flanks of 8-10 nude mice of 7-week old (002019,
Jackson Laboratory) under isoflurane anesthesia. 6-days after injection, the tumor volume was
measured daily with digital caliper for two weeks. At the end of the experiments, mice were
euthanized and 5-6 tumors were harvested randomly. Tumor tissues were proceeded to RNA
extraction and qPCR. To generate a tumor growth curve, the relative tumor size is calculated as
the ration of tumor size on each day over the size of the same tumor on day 1 of measurement.
Quantification and Statistical Analysis
Statistical analysis was performed using Graphpad Prism 8.0 Software (Graphpad Software; La
Jolla, CA, USA). Statistical significance was determined by Student’s t-Test or Kruskal–Wallis
non-parametric One-Way ANOVA as mentioned in the figure legends.
Acknowledgements
We thank Dr. Marco Conti for their generous gifts of PDE4D constructs, and helpful discussions.
We thank Lavpreet Jammu, Anh Diep and Christi Waer for initial characterization of Myomegalin
shRNA and for generating Myomegalin and PDE4D3 constructs. We thank Dr. Lin Gan for
helpful discussions on the manuscript. The research was supported by N.I.H. grant R15
CA235749 to X.G.
Author contributions
Jingyi Zhang: Acquisition, analysis and interpretation of data on Mmg-PDE4D interaction, GNP
culture, and FRET assay
Hualing Peng: Acquisition, analysis and interpretation of data on characterization of Mmg
CRISPR Knockout clones and mouse tumor model
Winston Ma: Establishment, expanding, initial selection and maintenance of Mmg CRISPR
Knockout clones
Amanda Ya: Initial selection and maintenance of Mmg CRISPR Knockout clones,
Characterization of INDEL mutations in Mmg CRISPR clones
Sammy Villa: Cloning and transfection of Mmg CRISPR construct into MEF cells
Shahar Sukenik: Supervision of the AKAR4 experiment and FRET data analysis
Xuecai Ge: Conception and design of all experiments, Acquisition of a portion of data, Analysis
and interpretation of data, Drafting the article
References
Barzi, M. et al. (2010) ‘Sonic-hedgehog-mediated proliferation requires the localization of PKA
to the cilium base’, Journal of Cell Science, 123(1), pp. 62–69. doi: 10.1242/jcs.060020.
Briscoe, J. and Therond, P. P. (2013) ‘The mechanisms of Hedgehog signalling and its rolesin
developement and disease’, Nature Reviews Molecuar Cell Biology, 14(7), p. 416. doi:
http://dx.doi.org.ezproxy.lib.ucalgary.ca/10.1038/nrm3598.
Dahmane, N. and Ruiz i Altaba, a (1999) ‘Sonic hedgehog regulates the growth and patterning
11
of the cerebellum.’, Development (Cambridge, England), 126(14), pp. 3089–3100.
Epstein, D. J. et al. (1996) ‘Antagonizing cAMP-dependent protein kinase A in the dorsal CNS
activates a conserved Sonic hedgehog signaling pathway.’, Development (Cambridge,
England), 122, pp. 2885–2894.
Fouladi, M. et al. (2005) ‘Intellectual and functional outcome of children 3 years old or younger
who have CNS malignancies’, Journal of Clinical Oncology, pp. 7152–7160. doi:
10.1200/JCO.2005.01.214.
Gavaldà, A. and Roberts, R. S. (2013) ‘Phosphodiesterase-4 inhibitors: a review of current
developments (2010 - 2012).’, Expert opinion on therapeutic patents, 23(8), pp. 997–1016. doi:
10.1517/13543776.2013.794789.
Ge, X. et al. (2015) ‘Phosphodiesterase 4D acts downstream of Neuropilin to control Hedgehog
signal transduction and the growth of medulloblastoma’, eLife. eLife Sciences Publications Ltd,
4(September2015).
Gillingham, A. K. and Munro, S. (2000) ‘The PACT domain, a conserved centrosomal targeting
motif in the coiled-coil proteins AKAP450 and pericentrin’, EMBO Reports, 1(6), pp. 524–529.
doi: 10.1093/embo-reports/kvd105.
Goodrich, L. V. et al. (1997) ‘Altered neural cell fates and medulloblastoma in mouse patched
mutants’, Science. doi: 10.1126/science.277.5329.1109.
Han, Y.-G. and Alvarez-Buylla, A. (2010) ‘Role of Primary Cilia in Brain Development and
Cancer.’, Current opinion in Neuroviology, 20(1), pp. 58–67. doi:
10.1016/j.conb.2009.12.002.Role.
Herbst, K. J., Allen, M. D. and Zhang, J. (2011) ‘Spatiotemporally Regulated Protein Kinase A
Activity Is a Critical Regulator of Growth Factor-Stimulated Extracellular Signal-Regulated
Kinase Signaling in PC12 Cells’, Molecular and Cellular Biology, 31(19), pp. 4063–4075. doi:
10.1128/mcb.05459-11.
Houslay, M. D. (2010) ‘Underpinning compartmentalised cAMP signalling through targeted
cAMP breakdown’, Trends in Biochemical Sciences, pp. 91–100. doi:
10.1016/j.tibs.2009.09.007.
Huang, Y., Roelink, H. and McKnight, G. S. (2002) ‘Protein kinase A deficiency causes axially
localized neural tube defects in mice’, Journal of Biological Chemistry, 277(22), pp. 19889–
19896. doi: 10.1074/jbc.M111412200.
Hui, C. and Angers, S. (2011) ‘Gli Proteins in Development and Disease’, Annual Review of Cell
and Developmental Biology, 27(1), pp. 513–537. doi: 10.1146/annurev-cellbio-092910-154048.
Humke, E. W. et al. (2010) ‘The output of Hedgehog signaling is controlled by the dynamic
association between Suppressor of Fused and the Gli proteins’, Genes and Development. doi:
10.1101/gad.1902910.
Kool, M. et al. (2012) ‘Molecular subgroups of medulloblastoma: An international meta-analysis
of transcriptome, genetic aberrations, and clinical data of WNT, SHH, Group 3, and Group 4
medulloblastomas’, Acta Neuropathologica, 123(4), pp. 473–484. doi: 10.1007/s00401-012-
0958-8.
Maurice, D. H. et al. (2014) ‘Advances in targeting cyclic nucleotide phosphodiesterases.’,
Nature reviews. Drug discovery. Nature Publishing Group, 13(4), pp. 290–314. doi:
10.1038/nrd4228.
McCormick, K. and Baillie, G. S. (2014) ‘Compartmentalisation of second messenger signalling
12
pathways’, Current Opinion in Genetics and Development. Elsevier Ltd, 27, pp. 20–25. doi:
10.1016/j.gde.2014.02.001.
Mukhopadhyay, S. et al. (2013) ‘The ciliary G-protein-coupled receptor Gpr161 negatively
regulates the sonic hedgehog pathway via cAMP signaling’, Cell, 152(1–2), pp. 210–223. doi:
10.1016/j.cell.2012.12.026.
Purzner, T. et al. (2018) ‘Developmental phosphoproteomics identifies the kinase CK2 as a
driver of Hedgehog signaling and a therapeutic target in medulloblastoma’, Science Signaling.
doi: 10.1126/scisignal.aau5147.
Richter, W., Jin, S.-L. C. and Conti, M. (2005) ‘Splice variants of the cyclic nucleotide
phosphodiesterase PDE4D are differentially expressed and regulated in rat tissue.’, The
Biochemical journal, 388(Pt 3), pp. 803–811. doi: 10.1042/BJ20050030.
Roubin, R. et al. (2013) ‘Myomegalin is necessary for the formation of centrosomal and Golgi-
derived microtubules’, Biol Open, 2(2), pp. 238–250. doi: 10.1242/bio.20123392.
Shaywitz, A. J. and Greenberg, M. E. (1999) ‘CREB: A Stimulus-Induced Transcription Factor
Activated by A Diverse Array of Extracellular Signals’, Annual Review of Biochemistry. doi:
10.1146/annurev.biochem.68.1.821.
Somatilaka, B. N. et al. (2020) ‘Ankmy2 Prevents Smoothened-Independent Hyperactivation of
the Hedgehog Pathway via Cilia-Regulated Adenylyl Cyclase Signaling’, Developmental Cell.
Elsevier Inc., 54(6), pp. 710-726.e8. doi: 10.1016/j.devcel.2020.06.034.
Tukachinsky, H., Lopez, L. V. and Salic, A. (2010) ‘A mechanism for vertebrate Hedgehog
signaling: Recruitment to cilia and dissociation of SuFu-Gli protein complexes’, Journal of Cell
Biology. doi: 10.1083/jcb.201004108.
Tuson, M., He, M. and Anderson, K. V. (2011) ‘Protein kinase A acts at the basal body of the
primary cilium to prevent Gli2 activation and ventralization of the mouse neural tube’,
Development, 138(22), pp. 4921–4930. doi: 10.1242/dev.070805.
Tyler Hillman, R. et al. (2011) ‘Neuropilins are positive regulators of Hedgehog signal
transduction’, Genes and Development. doi: 10.1101/gad.173054.111.
Verde, I. et al. (2001) ‘Myomegalin is a Novel Protein of the Golgi/Centrosome That Interacts
with a Cyclic Nucleotide Phosphodiesterase’, Journal of Biological Chemistry, 276(14), pp.
11189–11198. doi: 10.1074/jbc.M006546200.
Wallace, V. A. (1999) ‘Purkinje-cell-derived Sonic hedgehog regulates granule neuron precursor
cell proliferation in the developing mouse cerebellum’, Current Biology, 9(8), pp. 445–448. doi:
10.1016/S0960-9822(99)80195-X.
Wang, B. and Li, Y. (2006) ‘Evidence for the direct involvement of βTrCP in Gli3 protein
processing’, Proceedings of the National Academy of Sciences of the United States of America,
103(1), pp. 33–38. doi: 10.1073/pnas.0509927103.
Wechsler-Reya, R. J. and Scott, M. P. (1999) ‘Control of neuronal precursor proliferation in the
cerebellum by sonic hedgehog’, Neuron, 22(1), pp. 103–114. doi: 10.1016/S0896-
6273(00)80682-0.
Williams, C. H. et al. (2015) ‘An In Vivo Chemical Genetic Screen Identifies Phosphodiesterase
4 as a Pharmacological Target for Hedgehog Signaling Inhibition’, Cell Reports, 11(1), pp. 43–
50. doi: 10.1016/j.celrep.2015.03.001.
Yamanaka, H. et al. (2010) ‘Forskolin, a Hedgehog signal inhibitor, inhibits cell proliferation and
induces apoptosis in pediatric tumor cell lines’, Molecular Medicine Reports. doi: 10.3892/mmr-
13
00000230.
Yamanaka, H. et al. (2011) ‘Hedgehog signal inhibitor forskolin suppresses cell proliferation and
tumor growth of human rhabdomyosarcoma xenograft’, Journal of Pediatric Surgery. doi:
10.1016/j.jpedsurg.2010.11.010.
Yauch, R. L. et al. (2009) ‘Smoothened mutation confers resistance to a Hedgehog pathway
inhibitor in medulloblastoma.’, Science (New York, N.Y.), 326(5952), pp. 572–4. doi:
10.1126/science.1179386.
Zaccolo, M. and Pozzan, T. (2002) ‘Discrete microdomains with high concentration of cAMP in
stimulated rat neonatal cardiac myocytes’, Science, 295(5560), pp. 1711–1715. doi:
10.1126/science.1069982.
Zhang, J. et al. (2001) ‘Genetically encoded reporters of protein kinase A activity reveal impact
of substrate tethering’, Proceedings of the National Academy of Sciences, 98(26), pp. 14997–
15002. doi: 10.1073/pnas.211566798.
14
Figure 1. PDE4D3 interacts with Mmg at the centrosome.
(A) Protein structure of Mmg. Cnn_1N, Centrosomin N-terminal motif 1; NBPF (DUF1220), domain
of Neuroblastoma breakpoint family; Mmg-C, the domain previously shown to interact with PDE4D3.
(B) When overexpressed in HEK293T cells, HA-PDE4D3 and Flag-Mmg C-terminus co-
immunoprecipitated with each other, suggesting the interaction of the two proteins. (C)
Immunostaining of endogenous Mmg shows that Mmg overlaps with pericentrin, a marker for the
centrosome and pericentriolar materials (white arrows). (D) When expressed in NIH3T3 cells, Mmg-
C colocalizes to the centrosome and pericentriolar material (white arrows). (E) When expressed in
NIH3T3 cells, HA-PDE4D3 diffusively distribute to the cytoplasm, but a significant fraction of PDE4D3
is recruited by Mmg to the centrosome and pericentriolar material (white arrows).
15
S1 (related to Figure 2). Mmg knockdown impairs Hh signal transduction.
(A) NIH3T3 cells were transfected with shRNA against myomegalin. 72hr after transfection,
myomegalin transcript levels were assessed by qPCR. Two of the shRNAs significantly reduced
Mmg transcription. (B) Western blot showing that among all the shRNAs tested, #98 and #99
decreased the Mmg protein expression levels. (C) 72hr after shRNA transfection, NIH3T3 cells were
treated with SAG overnight. Hh signaling activity was evaluated by qPCR to detect the transcript
levels of Gli1, a Hh pathway target gene. All error bars represent SD; statistics: Student’s t-Test.
**p<0.01, ***p<0.001.
16
S2 (related to Fig 2). Design of Mmg CRISPR and genomic mutations in individual cell clones.
(A) Design of Mmg CRISPR. Two guide RNAs were designed and both target exon 1 of mouse Mmg.
Brown arrows underlie the sequence of gRNAs. (B-C) The gRNA targeting region in mouse genomic
was amplified by genomic PCR, ligated into TOPO vector, and transfected into chemically competent
cells. 20 bacterial colonies of each cell clones were randomly picked and sequenced. 3 type of
mutations were found in cell line #7 (B) and #10 (C). No wild type sequences were identified in the
20 colonies, suggesting that all alleles of Mmg were mutated. Mutations lead to frame shift (B1, B2,
C1, C2, C3), or alter the 5’ UTR (B3) that prevents the initiation of translation, all eventually leading
to nonsense-mediated decay (NMD) of the mutant mRNA.
17
Figure 2. Mmg knockout dislocates PDE4D from the centrosome and impairs Hh signal
transduction.
(A) Procedure of generating Mmg CRISPR cell clones. 30 cell clones were established and tested for
Mmg transcript levels and protein levels. (B) In two of the Mmg CRISPR cell clones, the transcript of
Mmg was hardly detectable by qPCR. (C) Western blot shows that in cell clone #7 and #10, the
CRISPR abolished the expression of the longer isoforms of Mmg, but the shorter isoform remains
unaffected. (D-E) Mmg CRISPR knockout clones were stimulated with SAG for 24hr, and Hh signaling
activity was evaluated by Gli1 transcript levels and protein levels. The Hh activity was dramatically
suppressed in Mmg knockout cells. (F) Representative images of EGFP-PDE4D3 expressed in wild
type or Mmg knockout cell clones. PDE4D3 concentrates at the centrosome and pericentriolar
material in wild type cells (white arrow); however, in CRISPR cell clones, it only exhibits diffusive
distribution to the cytoplasm and lacks the significant overlap with pericentrin (white arrows). (G)
Quantification of PDE4D intensity at the centrosome. All error bars represent SD; statistics in B and
D: Student’s t-Test. **p<0.01, ***p<0.001. Statistics in G: Kruskal–Wallis non-parametric One-Way
ANOVA, followed by Dunn’s multiple comparison. ***p<0.001, ****p<0.0001. A.U.: arbitrary unit.
18
Figure 3. Mmg loss increases basal PKA activity at the centrosome, and abolishes further PKA activation
by IBMX
(A) Representative images of active PKA immunostaining demonstrate that the basal active PKA levels at the
centrosome increased in Mmg knockout cells (white arrows). Dotted lines circled the areas where pPKA intensity
were measured based on the staining of pericentrin. (B) Quantification of basal active PKA levels at the centrosome.
Expression of exogenous Mmg restored active PKA levels at the centrosome. Data are shown as mean ± SD.
Statistics: Kruskal–Wallis non-parametric One-Way ANOVA, followed by Dunn’s multiple comparison. ***p<0.001,
****p<0.0001. A.U.: arbitrary unit. (C) Western blot shows that the global levels of active PKA do not change in
Mmg knockout cells. Forskolin treatment serves as a positive control of PKA overactivation in the entire cell. (D)
Schematic view of AKAR4, a FRET based probe for PKA activity. The CFP (Cerulean) and YFP (cpVE172) is linked
by a linker sequence that contain a PKA phosphorylation site and a phosphoanimo acid binding domain (PAABD).
PKA phosphorylation induces conformational change in the linker, which brings CFP and YFP in close proximity to
produce FRET. (E) By fusing AKAR4 to the RII subunit of PKA, we targeted the probe to the centrosome.
Immunostaining results confirmed the centrosome localization of RIIɑ-AKAR4. (F) Diagram showing that inhibiting
PDE4D increases PKA activity in wild type cells, but has little effect on PKA activity in Mmg knockout cells. (G)
Normalized emission of FRET acceptor over donor before and after IBMX 0.1mM. The ratio of YFP/CFP at each
time point was normalized to time zero. n = 8-13 cells. Data are shown as mean ± SEM. Statistics: t-Test, between
the wild type cells and Mmg knockout cells at the same time point. *p<0.05, **p<0.01.
19
S3 (related to Fig 3). Mmg loss blocked IBMX’s effect on PKA activity at the centrosome.
(A) Fluorescence images in live cells con-expressing RFP-PACT and RIIɑ-AKAP4. The centrosomes
(white arrow) was identified by RFP-PACT in the red channel, selected as the region of interest (ROI),
and FRET signaling were analyzed in the ROI. (B) Ratiometric view of FRET efficiency before and
after IBMX treatment.
20
Figure 4. Mmg loss impacts Gli3 processing and Gli2 transportation to the cilium tips.
(A) Diagram showing the proteolytic processing of Gli3 after PKA phosphorylation. (B) 24hr SAG
treatment reduced Gli3R levels in wild type cells, but not in Mmg knockout cell clones. (C)
Representative images of Gli2 immunostaining after cells are stimulated with SAG. White arrows point
to Gli2 at the cilium tips. (D) Quantification of Gli2 levels at the cilium tips. Mmg knockout reduced Gli2
levels at the cilium tips, while Mmg overexpression restored Gli2 intensity. Data are shown as mean ±
SD. Statistics: Kruskal–Wallis non-parametric One-Way ANOVA, followed by Dunn’s multiple
comparison. ****p<0.0001. A.U.: arbitrary unit. (E-F) Diagram showing PED4D3 specifically controls
PKA activities at the centrosome to regulate the Hh signaling transduction. Under normal conditions,
Upon Shh stimulation, SMO is translocated and activated in the cilium, which then triggers a signaling
cascade that reduces cAMP levels at the cilium base. The subsequent inhibition of PKA allows Gli2 to
be translocated and activated in the cilium tips (E). Without myomegalin, PDE4D3 is dislocated from
the centrosome and fails to degrade the local cAMP. Thus, PKA levels remain high at the centrosome
even after Shh stimulations. Hyperactive PKA suppresses Gli2 activation and promotes Gli3R
production. As a result, the Hh pathway cannot be activated (F).
21
Figure 5. Mmg knockdown blocked cell proliferation in primary cultured GNPs and
suppressed the growth rate of medulloblastoma in mouse model.
(A) Schematic of BrdU incorporation assay in primary cultured GNPs. Lentivirus expressing shRNA
against Mmg or control shRNA were added to the cell 5-6 hr after GNPs are plated in dishes. SAG
was added to the culture 24 hr before cells were fixed. BrdU pulse labeling lasted for 4 hr right before
cells were fixed. (B) Representative images of BrdU immunostaining in GNPs. (C) Mmg transcript
levels at the end of the experiments, measured by qPCR. (D) BrdU incorporation rate in GNPs. (E)
Levels of Hh signaling activity evaluated by Gli1 transcript levels. (F) Schematic diagram of the MB56
tumor allograft experiment in mouse. Measurement started 6 day after tumor allograft when the size
of tumors could be accurately measured. (G) The relative tumor size is defined as the tumor volume
on the indicated day divided by that on day 0. For each treatment 8–9 mice were used, and each
mouse was transplanted with two tumors on their hind flank. Results shown are from one of the two
independent experiments. (H) At the end of the experiment, the Gli1 transcript levels in 6 of randomly
sampled tumors were assessed by qPCR. Hh signaling activity was reduced by Mmg RNAi. Data
are presented as Mean ± SEM. Statistics: Student’s t-Test. *p<0.05, **p < 0.01, ***p < 0.001.
| 2021 | Myomegalin regulates Hedgehog pathway by controlling PDE4D at the centrosome | 10.1101/2020.04.24.059923 | [
"Peng Hualing",
"Zhang Jingyi",
"Ya Amanda",
"Ma Winston",
"Villa Sammy",
"Sukenik Shahar",
"Ge Xuecai"
] | creative-commons |
1
Extracellular adenosine induces hypersecretion of IL-17A by T-helper 17 cells through the
1
adenosine A2a receptor to promote neutrophilic inflammation
2
3
Mieko Tokano1,2, Sho Matsushita1,3, Rie Takagi1, Toshimasa Yamamoto2, and Masaaki
4
Kawano1,*
5
6
1Department of Allergy and Immunology, Faculty of Medicine, Saitama Medical University, 38
7
Morohongo, Moroyama, Saitama 350-0495, Japan
8
2Department of Neurology, Saitama Medical University, 38 Morohongo, Moroyama, Saitama,
9
350-0495, Japan
10
3Allergy Center, Saitama Medical University, 38 Morohongo, Moroyama, Saitama 350-0495,
11
Japan
12
13
*Correspondence: Department of Allergy and Immunology, Faculty of Medicine, Saitama
14
Medical University, 38 Morohongo, Moroyama, Saitama 350-0495, Japan
15
Tel.: +81 49 276 1173; Fax: +81 49 294 2274
16
E-mail address: mkawano@saitama-med.ac.jp
17
18
Running title: Adenosine modulates neutrophilic inflammation
19
20
Key words: Adenosine, Adenosine A2a receptor, CD4+ T cells, IL-17A, Th17 cells, EAE
21
2
Abstract
22
23
Extracellular adenosine, produced from ATP secreted by neuronal or immune cells, may play a
24
role in endogenous regulation of inflammatory responses. However, the underlying molecular
25
mechanisms are largely unknown. Here, we show that adenosine primes hypersecretion of
26
interleukin (IL)-17A by CD4+ T cells via T cell receptor activation. This hypersecretion was also
27
induced by an adenosine A2a receptor (A2aR) agonist, PSB0777. In addition, an A2aR
28
antagonist, Istradefylline, and inhibitors of adenylcyclase and protein kinase A (both of which
29
are signaling molecules downstream of the Gs protein coupled with the A2aR), suppressed
30
IL-17A production, suggesting that activation of A2aR induces IL-17A production by CD4+ T
31
cells. Furthermore, immune subset studies revealed that adenosine induced hypersecretion of
32
IL-17A by T-helper (Th)17 cells. These results indicate that adenosine is an endogenous
33
modulator of neutrophilic inflammation. Administration of an A2aR antagonist to mice with
34
experimental autoimmune encephalomyelitis led to marked amelioration of symptoms,
35
suggesting that suppression of adenosine-mediated IL-17A production is an effective treatment
36
for Th17-related autoimmune diseases.
37
3
Introduction
38
39
T-helper (Th)17 cells are a subset of T-helper cells induced by stimulation of naïve CD4+ T cells
40
with both tumor growth factor (TGF)-β and interleukin (IL)-6 in the presence of T cell receptor
41
signaling. IL-17A production by Th17 cells increases neutrophilic inflammation (1-3); however,
42
not all neutrophilic inflammatory diseases are explained by known Th17 responses (4). Indeed, it
43
is likely that as-yet-unknown Th17 or neutrophilic inflammatory responses occur. Here, we show
44
that adenosine induces IL-17A production by CD4+ T cells directly. Extracellular adenosine is
45
one of the first “signals” identified during regulation of a large number of physiological and
46
pathological processes, including bulging of an artery (5), sleep promotion (6), and regulation of
47
nerve action (7,8). Extracellular adenosine is produced from secreted ATP that undergoes rapid
48
stepwise dephosphorylation by ectonucleotidases such as the E-NTPDase CD39, which converts
49
ATP or ADP to ADP or AMP, respectively, and the 5’-nucleotidase CD73, which
50
dephosphorylates AMP to adenosine (9); both CD39 and CD73 are expressed by activated CD4+
51
T cells and antigen presenting cells (APCs) (10,11). Extracellular adenosine stimulates adenosine
52
receptors (A1R, A2aR, A2bR, and A3R) belonging to a superfamily of membrane proteins called
53
the G protein-coupled receptor family of class A seven-transmembrane domain receptors. A2aR
54
and A2bR signal the Gs protein to trigger cAMP synthesis, which in turn activates adenyl
55
cyclase and protein kinase A. By contrast, A1R and A3R signal the Gi protein to trigger cAMP
56
degradation. In addition, A2bR also signals the Gq protein, which in turn activates phospholipase
57
C. In an immunological context, adenosine receptors are expressed by various immune cells,
58
including T cells and APCs (12). It is also suggested that adenosine stimulates neutrophil
59
chemotaxis and phagocytosis via A1R and A3R (13). In addition, adenosine induces Th17
60
4
differentiation by activating A2bR on CD4+ T cells (14). By contrast, several reports suggest that
61
adenosine suppresses Th17 differentiation via activation of A2aR on CD4+ T cells (15-17).
62
Considering that the G protein downstream of A2aR is Gs, and those of A2bR are Gs and Gq (5),
63
it is assumed that induction of A2bR-mediated Th17 differentiation is induced through
64
simultaneous activation of Gq and suppression of Gs, or via other unknown mechanisms.
65
Therefore, the precise molecular mechanism(s) underlying the effect of adenosine on Th cells is
66
unclear. Here, we show that adenosine promotes IL-17A production in a two-way mixed
67
lymphocyte reaction (MLR). In addition, an A2aR agonist (PSB0777) induced IL-17A
68
production, and an A2aR antagonist (Istradefylline) inhibited production, induced by adenosine,
69
suggesting that activation of A2aR plays a role in adenosine-mediated IL-17A production. This
70
notion was further supported by the observation that inhibitors of adenylcyclase and protein
71
kinase A, both of which are signaling molecules downstream of the Gs protein (18), also
72
suppressed adenosine-mediated IL-17A production. Immune subset studies suggested that Th17
73
cells play a role in adenosine-mediated hypersecretion of IL-17A. Administration of an A2R
74
antagonist to mice with experimental autoimmune encephalomyelitis (EAE) (19,20) markedly
75
ameliorated symptoms. Taken together, the data indicate that adenosine-dependent
76
hypersecretion of IL-17A by Th17 cells contributes not only to antibacterial defense but also to
77
neutrophilic autoimmune diseases, and that suppressing this process may be an effective therapy
78
for the latter.
79
5
Methods
80
81
Mice
82
Balb/c mice were obtained from Japan SLC, Inc. SJL/J mice were obtained from Charles River
83
Laboratories Japan, Inc. Mice were housed in appropriate animal care facilities at Saitama
84
Medical University and handled according to international guidelines for experiments with
85
animals. All experiments were approved by the Animal Research Committee of Saitama Medical
86
University.
87
88
Two-way MLR
89
Splenic lymphocytes were collected by lyzing tissue in a Dounce homogenizer, followed by
90
layering over Ficoll Paque (GE health care, Chicago, IL, USA), as described previously (21).
91
Balb/c splenic lymphocytes (3 × 106) were mixed with SJL/J splenic lymphocytes (3 × 106) in
92
500 µL of DMEM medium containing 10% FCS, 100 U/ml penicillin, 100 µg/mL streptomycin,
93
2 mM L-glutamine, 1 mM sodium pyruvate, and 50 µM 2-mercaptoethanol (D10 medium) in 24
94
well plates in the presence of adenosine (0–1 mM) (Sigma, St. Louis, MO, USA); in the presence
95
of each adenosine receptor agonist (0–10 µM) (A1R: 2-Chloro-N6-cyclopentyladenosine
96
(CCPA), Tocris, Bristol, UK; A2aR: PSB0777, Tocris; A2bR: BAY 60-6583, Tocris; A3R:
97
HEMADO, Tocris); in the presence of A2aR antagonist (0–1 nM) (Istradefylline, Sigma) plus
98
adenosine (100 µM); in the presence of an adenyl cyclase inhibitor (0–1 µM) (MDL-12330A,
99
Enzo Life Sciences, Farmingdale, NY, USA) or a protein kinase A inhibitor (0–1 µM) (H-89,
100
Tocris) plus adenosine (100 µM) to inhibit A2aR signaling; or in the presence of a CD39
101
inhibitor (0–1 µM) (ARL67156, Tocris) or a CD73 inhibitor (0–1 µM) (adenosine 5'-(α,
102
6
β-methylene) diphosphate (AMP-CP; Tocris) plus ATP (100 µM) (GE healthcare). After mixing,
103
the plates were incubated for 7 days at 37°C. The supernatants were collected for use in cytokine
104
ELISAs.
105
106
Flow cytometry analysis
107
The MLR was performed for 7 days in the presence or absence of adenosine (100 µM) (as
108
described above). After 7 days, cells were blocked with anti-mouse CD16/CD32 antibodies
109
(BioLegend, San Diego, CA, USA) and then stained for 30 min at 4°C with a fluorescein
110
isothiocyanate (FITC)-conjugated anti-mouse CD4 antibody (BioLegend). The cells were then
111
fixed, permeabilized, and stained with a phycoerythrin (PE)-conjugated anti-mouse IL-17A
112
antibody (BioLegend). Finally, the cells were washed and analyzed on a FACSCanto II flow
113
cytometer (BD Biosciences, Franklin Lakes, NJ) using FACSDiva acquisition software (BD
114
Biosciences).
115
116
Isolation of CD4+, CD4+CD62L+ T cells, and B cells
117
CD4+ T cells within the Balb/c and SJL/J splenocyte populations were isolated from the mixture
118
prepared previously after rupturing red blood cells (22). Cells were isolated by positive selection
119
of CD4+ T cells using magnetic-activated cell sorting (MACS) (Miltenyi Biotec, Bergisch
120
Gladbach, Germany), according to the manufacturer’s instructions. CD4+CD62L+ T cells were
121
isolated from Balb/c splenocytes by a combination of negative and positive selection by MACS.
122
During positive selection of CD62L+ T cells, negatively isolated CD4+ T cells were collected as
123
the flow through fraction. B cells were isolated from Balb/c splenocytes by positive selection
124
using MACS. Cells were resuspended in 1 mL of D10 medium and counted. The purity and
125
7
viability of CD4+ T cells, CD4+CD62L+ T cells, and B cells were >90% (Sup. Fig. 1). The purity
126
and viability of cells in the flow through fraction collected during isolation of CD4+CD62L+ T
127
cells are shown in Supplementary Figure 1.
128
129
Cell sorting
130
Balb/c CD4+ T cells (prepared as described above) were labeled for 30 min at 4°C with
131
PE-conjugated anti-mouse CCR3, CCR5, CCD6, CD25, or CD62L antibodies (BioLegend). The
132
cells were then washed and sorted using a FACS Aria II flow cytometer (BD Biosciences). The
133
purity and viability of the sorted cells are Supplementary Figure 1.
134
135
CD3/CD28 stimulation
136
CD4+ T cells (1 × 106) were stimulated for 7 days at 37°C with anti-mouse CD3 (BioLegend) (1
137
µg/mL) and CD28 (BioLegend) (0.5 µg/mL) antibodies (CD3/CD28) in the presence of
138
adenosine
(0–1
mM),
each
adenosine
receptor
agonist
(0–10
µM)
(A1R:
139
2-Chloro-N6-cyclopentyladenosine (CCPA), Tocris, Bristol, UK; A2aR: PSB0777, Tocris;
140
A2bR: BAY 60-6583, Tocris; A3R: HEMADO, Tocris), an A2aR antagonist (Istradefylline; 0–1
141
nM) plus adenosine (600 µM), or an adenyl cyclase inhibitor (0–1 µM) (MDL-12330A) or a
142
protein kinase A inhibitor (0–1 µM) (H-89) plus adenosine (600 µM) in 500 µL of D10 medium.
143
After cell sorting, cells (3 × 105) were plated in 24 well plates and stimulated for 7 days at 37°C
144
with anti-mouse CD3 (1 µg/mL) and CD28 (0.5 µg/mL) antibodies in the presence of adenosine
145
(600 µM) in 100 µL of D10 medium. After stimulation, the supernatants were collected for use
146
in cytokine ELISAs.
147
148
8
Adenosine or ATP ELISAs
149
MLR was performed by mixing Balb/c lymphocytes (6 × 106) with SJL/J lymphocytes (6 × 106)
150
in a 15 mL tube for 0–24 h at 37°C in the presence of a CD39 inhibitor (ARL67156) (0–1 µM)
151
and a CD73 inhibitor (AMP-CP) (0–1 µM) in 200 µL of D10 medium. CD3/CD28 stimulation
152
was performed for 24 h at 37°C in a 15 mL tube by incubating Balb/c CD4+ T cells (1 × 107)
153
with anti-mouse CD3 (1 µg/mL) and CD28 (0.5 µg/mL) antibodies plus ARL67156 (1 µM) or
154
AMP-CP (1 µM) in 200 µL of D10 medium. LPS stimulation was performed for 24 h at 37°C in
155
a 15 ml tube by incubating Balb/c B cells (1 × 107) or Balb/c bone marrow (BM)
156
derived-dendritic cells (BM-DCs) (1 × 107) generated from mouse bone marrow cells as
157
described previously (23) with LPS (Sigma) (0.5 µg/ml) plus ARL67156 (1 µM) or AMP-CP (1
158
µM) in 200 µL of D10 medium. The purity and viability of BM-DCs were >90% (Sup. Fig. 1).
159
After incubation, the supernatants were collected and tested in adenosine or ATP ELISAs
160
(Biovision, Milpitas, CA, USA).
161
162
Differentiation of naïve CD4+ T cells
163
Naïve CD4+ T cells (3 × 105) in 500 µL of D10 medium in 24 well plates were stimulated for 7
164
days at 37°C with anti-mouse CD3 (1 µg/mL) and CD28 (0.5 mg/mL) antibodies plus mouse
165
IL-6 (20 ng/mL) (Peprotech, Rocky Hill, NJ), and human TGF-β1 (2 ng/mL) (Peprotech) in the
166
presence of an A2aR antagonist (Istradefylline) (0–1 nM). After incubation, cells in 500 µL of
167
D10 medium were stimulated for another 7 days at 37°C with anti-mouse CD3 (1 µg/mL) and
168
CD28 (0.5 µg/mL) antibodies in the presence of Istradefylline (0–1 nM). Supernatants were
169
collected for use in cytokine ELISAs.
170
171
9
EAE model
172
EAE was induced as described previously (19). Briefly, SJL/J mice received a subcutaneous
173
inguinal injection (100 µg/mouse) of the proteolipid protein (PLP) peptide (PLP139-151, Tocris)
174
emulsified in complete Freund’s adjuvant (CFA) containing mycobacterium tuberculosis H37Ra
175
(100 µg/mouse; Difco, Detroit, MI, USA). Mice also received oral PBS(-) or an A2aR antagonist
176
(Istradefylline) (6 µg/mouse) once every 2 days from Day –7 to Day +18 after immunization
177
with PLP peptide (Day 0). Mice were examined daily for signs of EAE, which were graded as
178
described (24).
179
180
Peptide pulse assay
181
At 7 days post-subcutaneous immunization with PLP peptide emulsified in CFA, splenocytes (2
182
× 106 in 200 µL of D10 medium) were seeded in 96-well round plates and pulsed for 3 days at
183
37°C with PLP peptide (10 µM) in the presence of adenosine (600 µM) and an A2aR antagonist
184
(Istradefylline; 0–1 nM). Supernatants were collected for use in cytokine ELISAs.
185
186
Cytokine ELISAs
187
The concentrations of IFN-γ, IL-5, IL-17A, IL-17F, and IL-22 in cell supernatants were
188
measured using specific ELISA kits (DuoSet Kit, R&D, Minneapolis, MN, USA). Any value
189
below the lower limit of detection (15.6 pg/mL) was set to 0. No cytokine cross-reactivity was
190
observed within the detection ranges of the kits. If necessary, samples were diluted appropriately
191
so that the measurements fell within the appropriate detection range for each cytokine.
192
193
Statistical analysis
194
10
Differences between two groups were analyzed using an unpaired Student’s t-tests. Differences
195
between three or more groups were analyzed using one-way ANOVA with Tukey’s post-hoc test.
196
Clinical scores were analyzed using a non-parametric Mann-Whitney U-test. All calculations
197
were performed using KaleidaGraph software (Synergy software, Reading, PA, USA). A P value
198
< 0.05 was considered statistically significant.
199
11
Results
200
201
Adenosine promotes IL-17A production by CD4+ T cells in an MLR
202
First, we analyzed the effect of adenosine on CD4+ T cells during T cell-APC interactions in an
203
MLR (25). We found that CD4+ T cells exposed to adenosine secreted IL-17A in a
204
dose-dependent manner (Fig. 1A–C). Since both agonist-mediated IL-17A production and
205
antagonist-mediated suppression of adenosine-mediated IL-17A production were observed in the
206
presence of an adenosine A2aR agonist (PSB0777) and an antagonist (Istradefylline),
207
respectively (Fig. 1D and Sup. Fig. 2), we hypothesized that IL-17A in the MLR was produced
208
by CD4+ T cells stimulated via the A2aR. This notion was supported by the finding that
209
inhibitors of signaling molecules downstream of the A2aR also suppressed IL-17A production
210
(Fig. 1E). Furthermore, ATP induced IL-17A production in the MLR, which was suppressed by
211
the A2aR antagonist and by inhibitors of CD39 and CD73 (Fig. 1F), suggesting that adenosine
212
plays a role in IL-17A production. Since the A2aR antagonist and inhibitors of CD39 and CD73
213
also inhibited basal production of IL-17A in the MLR (Fig. 1G), we postulate that de novo
214
adenosine production is induced in the MLR.
215
216
Adenosine production in the MLR
217
CD39 and CD73 expressed on the surface of endothelial cells (26,27) and immune cells (10,11),
218
including T cells and DCs, are critical for production of adenosine from ATP. Since we found
219
that inhibitors of CD39 and CD73 inhibited basal IL-17A production in the MLR (Fig. 1G), we
220
next addressed the source of adenosine production in the MLR. As shown in Figure 2A,
221
production of adenosine and ATP was time-dependent. In accordance with suppression of
222
12
IL-17A production, inhibitors of CD39 and CD73 suppressed adenosine production at 24 h after
223
the start of the MLR (Fig. 2B and C). Since the MLR induces activation of CD4+ T cells by
224
APCs (25), we also addressed adenosine production by activated CD4+ T cells, B cells, and
225
BM-DCs (Fig. 2D–F). Production of both adenosine and ATP by activated CD4+ T cells, B cells,
226
and BM-DCs was observed; inhibitors of CD39 and CD73 suppressed production by activated
227
CD4+ T cells (Fig. 2D). This suggests that adenosine produced by CD4+ T cells may induce
228
spontaneous IL-17A secretion by CD4+ T cells in the MLR.
229
230
Th17 cells hypersecrete IL-17A in the presence of anti-CD3/CD28 antibodies and adenosine
231
Next, we tried to identify the Th subset that generated IL-17A in the presence of adenosine. First,
232
we confirmed that CD4+ T cells expressed the A2aR and secreted IL-17A by stimulating them
233
with agonistic anti-CD3/CD28 antibodies in the presence of adenosine (Fig. 3A). Time course
234
studies showed that adenosine-mediated IL-17A production was detected from 3 days
235
post-CD3/CD28 stimulation (Fig. 3B and Sup. Fig. 3). Administration of adenosine within 6 h of
236
antibody stimulation triggered IL-17A production; however, administration at 24 h
237
post-stimulation did not (Fig. 3C). As in the MLR, CD4+ T cells also produced IL-17A after
238
stimulation with anti-CD3/CD28 antibodies in the presence of an A2aR agonist (Fig. 3D and Sup.
239
Fig. 4). Adenosine-mediated IL-17A production by CD3/CD28-stimulated CD4+ T cells was
240
suppressed by an A2aR antagonist (Fig. 3D and Sup. Fig. 4). This was supported by data
241
showing that inhibitors of signaling molecules downstream of the A2R also suppressed IL-17A
242
production (Fig. 3E). Production of other Th17-related cytokines was also induced by adenosine
243
through the A2aR (Fig. 4). This again suggests that activated CD4+ T cells produce IL-17A upon
244
A2aR activation. Furthermore, we noticed that CD4+CD62L-, but not CD4+CD62L+, cells
245
13
produced IL-17A after CD3/CD28 stimulation in the presence of adenosine, suggesting that
246
adenosine induces IL-17A production by effector Th cells (Fig. 3F). Therefore, we performed
247
immune subset studies after separating CD4+ T cells using anti-chemokine receptor (CCR)
248
antibodies. As shown in Figure 3G, CD4+CCR6hi T cells produced IL-17A upon stimulation of
249
CD3/CD28, and production was strongly up-regulated in the presence of adenosine. Since CCR6
250
is a typical marker of Th17 cells (28,29), this suggests that activated Th17 cells hypersecrete
251
IL-17A in the presence of adenosine.
252
253
An adenosine A2aR antagonist ameliorates IL-17A-related autoimmune EAE responses
254
The above results raise the possibility that adenosine-mediated hypersecretion of IL-17A by
255
Th17 cells contributes to Th17-related autoimmune diseases. This hypothesis is supported by a
256
report showing that CD73 knockout mice are resistant to EAE (30), a Th17-mediated
257
autoimmune disease (20). We expected, therefore, that A2aR antagonist-mediated suppression of
258
Th17 responses should improve EAE. To address this, we examined the efficacy of an A2aR
259
antagonist in EAE model SJL/J mice (19). EAE was induced by immunization of mice with an
260
I-As restricted helper peptide derived from a myelin PLP peptide comprising amino acids
261
139–151 (HSLGKWLGHPDKF). The peptide was emulsified in CFA. First, we confirmed that
262
the
A2aR
antagonist
suppressed
adenosine-mediated
IL-17A
production
by
263
CD3/CD28-stimulated CD4+ T cells from SJL/J strain mice (Fig. 5A). The A2aR antagonist also
264
significantly suppressed adenosine-mediated IL-17A production after, but not during,
265
differentiation of Th17 cells from naïve CD4+ T cells (Fig. 5B). By contrast, and in agreement
266
with Figure 3G, adenosine administration did not induce IL-17A production during and after
267
differentiation of Th1, Th2, and Treg cells from naïve CD4+ T cells (data not shown). Next, we
268
14
pulsed splenocytes with the PLP peptide after immunization to confirm that IL-17A production
269
was induced in a peptide-dependent manner, and that production was up-regulated by adenosine.
270
As expected, IL-17A production occurred in a peptide-dependent manner and was up-regulated
271
by adenosine (Fig. 5C). Furthermore, the A2aR antagonist suppressed production of IL-17A,
272
suggesting that the A2aR antagonist inhibits IL-17A production by CD4+ T cells induced by
273
immunization with the PLP peptide. Finally, the A2aR antagonist was administered orally to
274
mice before and during EAE induction (Fig. 5D and E). As shown, the clinical scores of mice
275
receiving the A2aR antagonist were markedly lower than those of control mice (receiving water)
276
at 18 days post-immunization with the PLP peptide (Fig. 5D). Accordingly, histological studies
277
showed that the numbers of central nervous system-infiltrating CD3+ cells in mice receiving the
278
A2aR antagonist were much lower than those in mice receiving water (Fig. 5E). This suggests
279
that the A2aR antagonist suppresses IL-17A-mediated autoimmune responses by suppressing
280
hypersecretion of IL-17A by Th17 cells.
281
15
Discussion
282
283
Here, we showed that adenosine induces hypersecretion of IL-17A by Th17 cells. Addition of
284
adenosine (1 mM) to a two-way MLR increased IL-17A production to > 25 times the basal level;
285
however, both basal production and increased production of IL-17A were suppressed by an
286
A2aR antagonist, and by CD39/CD73 inhibitors. This indicates that hypersecretion of IL-17A in
287
the presence of adenosine occurs by other mechanisms in addition to T cell-APC interactions.
288
Since endothelial cells and nervous system cells also express CD39/CD73 (26,27,31,32) and
289
produce adenosine (8,33), and activated CD4+ T cells in the present study hypersecreted IL-17A
290
at 6 h post-CD3/CD28 stimulation, it is possible that activated Th17 cells also receive adenosine
291
from endothelial and neuronal cells to induce hypersecretion of IL-17A.
292
It is suggested that physiological concentrations of adenosine are lower than 1 µM, but
293
can be increased by stimuli such as high K+ levels, electrical stimulation, glutamate receptor
294
agonists, hypoxia, hypoglycemia, and ischemia (34). To obtain sufficient adenosine (> 100 µM)
295
to trigger hypersecretion of IL-17A, activated Th17 cells may need to make contact with
296
non-immune cells such as adenosine-producing endothelial cells (35) and neuronal cells (36) to
297
form a microenvironment with a high adenosine concentration (as observed during T cell-APC
298
interactions at immunological synapses) (37). Thus, A2aR antagonists rather than CD39/CD73
299
inhibitors might be more effective at inhibiting de novo adenosine-mediated hypersecretion of
300
IL-17 by Th17 cells. A previous study suggests that intracellular adenosine is transported out of
301
cells by efficient equilibrative transporters (38); CD39/CD73 inhibitors would not suppress this
302
type of de novo adenosine production.
303
With regard to the effect of adenosine on other Th subsets, our observations were
304
16
different from those of previous reports (39,40); here, we observed that adenosine up-regulated
305
IFN-γ (a Th1-related cytokine) secretion at 5 and 7 days and had no significant effect on IL-5 (a
306
Th2-related cytokine) secretion by CD4+ T cells after CD3/CD28 stimulation with 600 µM of
307
adenosine, although IL-17A production was significant (Sup. Fig. 3). This suggests that
308
adenosine induces hypersecretion of IL-17A by Th17 cell but does not suppress Th1 and Th2
309
activity. However, previous studies report that adenosine-mediated suppression of IFN-γ and
310
IL-5 was observed 1 day after T cell receptor-mediated stimulation of CD4+ T cells (39,40). This
311
may indicate that in the short term adenosine prioritizes stimulation of Th17 cell activity rather
312
than that of Th1 and Th2 cells, and that it does not suppress effector Th activity in the long term.
313
It is also suggested that the A2aR agonist, CGS21680, suppresses Th17 differentiation
314
(15-17). This result is opposite to ours; one reason for this may be differences in the source of
315
the A2aR agonist. The A2aR agonist CGS 21680 is much less selective than the A2aR agonist
316
we used this study (PSB0777); this is because CGS21680 not only binds to the A2aR but also to
317
A1R and A3R, which are associated with the Gi protein (which has opposite effects to the Gs
318
protein) (41,42). Therefore, it is probable that CGS21680 may cancel out any agonist effects by
319
activating A1R and A3R. Also, it is suggested that the A2aR antagonist, SCH58261,
320
up-regulates Th17 differentiation in mice (17). We hypothesized that up-regulation of Th17
321
differentiation by SCH58261 may be induced through relative downregulation of A2aR activity
322
compared with that of A2bR; this relative increase in A2bR activity induces Th17 differentiation
323
in mice (43,44). Also, it is probable that SCH58261 induces relative increase in activity of G
324
protein-coupled receptors other than adenosine receptors (e.g., dopamine receptors) to induce
325
Th17 differentiation in mice (45). We also hypothesize that although SCH58261 may induce
326
Th17 differentiation in vivo, it may not stimulate Th17 activity; this is because our data show
327
17
that an A2a antagonist (Istradefylline) suppressed IL-17A secretion by differentiated Th17 cells
328
but did not suppress Th17 differentiation (Fig. 5B). This hypothesis is supported by previous
329
data showing that SCH58261 markedly suppresses symptoms of EAE, a typical Th17-mediated
330
disease (30).
331
Our data suggest that production of IL-17A is relatively higher after exposure to an
332
A2aR agonist, PSB0777, than after exposure to an A2bR agonist, BAY 60-6583. PSB0777 is a
333
potent adenosine A2aR agonist (Ki = 44.4 nM for rat brain striatal A2aR) (42), and BAY
334
60-6583 is a potent adenosine A2bR agonist (Ki = 100 nM for rat A2bR) (46). By assuming that
335
the Ki values of PSB0777 and BAY 60-6583 are comparable, we thought that production of
336
IL-17A mediated by activation of the A2aR might be higher than that mediated by activation of
337
the A2bR. This hypothesis is supported by the notion that the A2aR is a high affinity receptor
338
with activity in the low to mid-nanomolar range, whereas the A2bR has a much lower affinity for
339
adenosine (micromolar) (5); this suggests that adenosine activates the A2aR rather than the
340
A2bR.
341
A2aR antagonists have been developed for treatment of Parkinsonism (47) and
342
malignancies (48). In addition, inhibitors of CD39/CD73 have been developed as anti-tumor
343
drugs (49,50). Regarding the effects of adenosine on tumor immunity, a previous study suggests
344
that adenosine suppresses effector T cell function since tumor cells express both CD39 and
345
CD73 and secrete adenosine (51). By contrast, several reports suggest that IL-17A promotes
346
emergence of pro-tumorigenic neutrophil phenotypes (52,53). Neutrophils in mouse tumor
347
models promote tumor metastasis (54-56), and observations in cancer patients have linked
348
elevated neutrophil counts in blood with increased risk of metastasis (57). Therefore, it is
349
probable that tumor-produced adenosine induces IL-17A secretion by CD4+CCR6hi T cells
350
18
followed by neutrophilic inflammation, which promotes tumor metastasis.
351
The results presented herein indicate that these drugs may also be effective treatments
352
for Th17-mediated diseases (4) such as psoriasis, neutrophilic bronchial asthma, severe atopic
353
dermatitis, and autoimmune diseases by suppressing hypersecretion of IL-17A by Th17 cells.
354
Moreover, these drugs might be effective treatments for diseases caused by neutrophilic
355
inflammation of unknown cause in the dermis; such diseases include Behcet uveitis (58) and
356
vasculitis of adenosine deaminase 2 deficiency (59). This is because endothelial cells express
357
CD39/CD73 and produce adenosine, which could induce hypersecretion of IL-17A by Th17 cells,
358
thereby contributing to inflammation.
359
19
References
360
361
1
Matsuzaki, G. and Umemura, M. 2007. Interleukin-17 as an effector molecule of innate
362
and acquired immunity against infections. Microbiol Immunol 51:1139.
363
2
Manel, N., Unutmaz, D., and Littman, D. R. 2008. The differentiation of human T(H)-17
364
cells requires transforming growth factor-beta and induction of the nuclear receptor
365
RORgammat. Nat Immunol 9:641.
366
3
Ivanov, II, Zhou, L., and Littman, D. R. 2007. Transcriptional regulation of Th17 cell
367
differentiation. Semin Immunol 19:409.
368
4
Tesmer, L. A., Lundy, S. K., Sarkar, S., and Fox, D. A. 2008. Th17 cells in human
369
disease. Immunol Rev 223:87.
370
5
Cronstein, B. N. and Sitkovsky, M. 2017. Adenosine and adenosine receptors in the
371
pathogenesis and treatment of rheumatic diseases. Nat Rev Rheumatol 13:41.
372
6
Jones, B. E. 2009. Glia, adenosine, and sleep. Neuron 61:156.
373
7
Boison, D., Chen, J. F., and Fredholm, B. B. 2010. Adenosine signaling and function in
374
glial cells. Cell Death Differ 17:1071.
375
8
Sebastiao, A. M. and Ribeiro, J. A. 2009. Tuning and fine-tuning of synapses with
376
adenosine. Curr Neuropharmacol 7:180.
377
9
Yegutkin, G. G. 2008. Nucleotide- and nucleoside-converting ectoenzymes: Important
378
modulators of purinergic signalling cascade. Biochim Biophys Acta 1783:673.
379
10
Whiteside, T. L., Mandapathil, M., and Schuler, P. 2011. The role of the adenosinergic
380
pathway in immunosuppression mediated by human regulatory T cells (Treg). Curr Med
381
Chem 18:5217.
382
20
11
Allard, B., Longhi, M. S., Robson, S. C., and Stagg, J. 2017. The ectonucleotidases CD39
383
and CD73: Novel checkpoint inhibitor targets. Immunol Rev 276:121.
384
12
Junger, W. G. 2011. Immune cell regulation by autocrine purinergic signalling. Nat Rev
385
Immunol 11:201.
386
13
Barletta, K. E., Ley, K., and Mehrad, B. 2012. Regulation of neutrophil function by
387
adenosine. Arteriosclerosis, thrombosis, and vascular biology 32:856.
388
14
Wilson, J. M., Kurtz, C. C., Black, S. G., Ross, W. G., Alam, M. S., Linden, J., and Ernst,
389
P. B. 2011. The A2B adenosine receptor promotes Th17 differentiation via stimulation of
390
dendritic cell IL-6. Journal of immunology 186:6746.
391
15
Alchera, E., Rolla, S., Imarisio, C., Bardina, V., Valente, G., Novelli, F., and Carini, R.
392
2017. Adenosine A2a receptor stimulation blocks development of nonalcoholic
393
steatohepatitis in mice by multilevel inhibition of signals that cause immunolipotoxicity.
394
Transl Res 182:75.
395
16
Wang, L., Wan, H., Tang, W., Ni, Y., Hou, X., Pan, L., Song, Y., and Shi, G. 2018.
396
Critical roles of adenosine A2A receptor in regulating the balance of Treg/Th17 cells in
397
allergic asthma. Clin Respir J 12:149.
398
17
Ansari, M. A., Nadeem, A., Attia, S. M., Bakheet, S. A., Raish, M., and Ahmad, S. F.
399
2017.
Adenosine
A2A
receptor
modulates
neuroimmune
function
through
400
Th17/retinoid-related orphan receptor gamma t (RORgammat) signaling in a BTBR T(+)
401
Itpr3(tf)/J mouse model of autism. Cell Signal 36:14.
402
18
Hasko, G., Linden, J., Cronstein, B., and Pacher, P. 2008. Adenosine receptors:
403
therapeutic aspects for inflammatory and immune diseases. Nat Rev Drug Discov 7:759.
404
19
Nakano, K., Higashi, T., Hashimoto, K., Takagi, R., Tanaka, Y., and Matsushita, S. 2008.
405
21
Antagonizing dopamine D1-like receptor inhibits Th17 cell differentiation: preventive
406
and therapeutic effects on experimental autoimmune encephalomyelitis. Biochem
407
Biophys Res Commun 373:286.
408
20
Kleinewietfeld, M., Manzel, A., Titze, J., Kvakan, H., Yosef, N., Linker, R. A., Muller, D.
409
N., and Hafler, D. A. 2013. Sodium chloride drives autoimmune disease by the induction
410
of pathogenic TH17 cells. Nature 496:518.
411
21
Higashi, T., Shimojo, N., Suzuki, S., Nakaya, M., Takagi, R., Hashimoto, K., Nakagome,
412
K., Nakamura, K., Kohno, Y., and Matsushita, S. 2011. Coenzyme A contained in
413
mothers' milk is associated with the potential to induce atopic dermatitis. Int Immunol
414
23:741.
415
22
Kawano, M., Takagi, R., Saika, K., Matsui, M., and Matsushita, S. 2018. Dopamine
416
regulates cytokine secretion during innate and adaptive immune responses. Int Immunol
417
30:591.
418
23
Lutz, M. B., Kukutsch, N., Ogilvie, A. L., Rossner, S., Koch, F., Romani, N., and Schuler,
419
G. 1999. An advanced culture method for generating large quantities of highly pure
420
dendritic cells from mouse bone marrow. Journal of immunological methods 223:77.
421
24
Podojil, J. R., Padval, M. V., and Miller, S. D. 2011. Combination treatment of mice with
422
CRx-153 (nortriptyline and desloratadine) decreases the severity of experimental
423
autoimmune encephalomyelitis. Cell Immunol 270:237.
424
25
McDevitt, H. O. 2000. Discovering the role of the major histocompatibility complex in
425
the immune response. Annu Rev Immunol 18:1.
426
26
Kaczmarek, E., Koziak, K., Sevigny, J., Siegel, J. B., Anrather, J., Beaudoin, A. R., Bach,
427
F. H., and Robson, S. C. 1996. Identification and characterization of CD39/vascular ATP
428
22
diphosphohydrolase. J Biol Chem 271:33116.
429
27
Jalkanen, S. and Salmi, M. 2008. VAP-1 and CD73, endothelial cell surface enzymes in
430
leukocyte extravasation. Arteriosclerosis, thrombosis, and vascular biology 28:18.
431
28
Acosta-Rodriguez, E. V., Rivino, L., Geginat, J., Jarrossay, D., Gattorno, M.,
432
Lanzavecchia, A., Sallusto, F., and Napolitani, G. 2007. Surface phenotype and antigenic
433
specificity of human interleukin 17-producing T helper memory cells. Nat Immunol
434
8:639.
435
29
Annunziato, F., Cosmi, L., Santarlasci, V., Maggi, L., Liotta, F., Mazzinghi, B., Parente,
436
E., Fili, L., Ferri, S., Frosali, F., Giudici, F., Romagnani, P., Parronchi, P., Tonelli, F.,
437
Maggi, E., and Romagnani, S. 2007. Phenotypic and functional features of human Th17
438
cells. J Exp Med 204:1849.
439
30
Mills, J. H., Thompson, L. F., Mueller, C., Waickman, A. T., Jalkanen, S., Niemela, J.,
440
Airas, L., and Bynoe, M. S. 2008. CD73 is required for efficient entry of lymphocytes
441
into the central nervous system during experimental autoimmune encephalomyelitis.
442
Proceedings of the National Academy of Sciences of the United States of America
443
105:9325.
444
31
Braun, N., Sevigny, J., Robson, S. C., Enjyoji, K., Guckelberger, O., Hammer, K., Di
445
Virgilio, F., and Zimmermann, H. 2000. Assignment of ecto-nucleoside triphosphate
446
diphosphohydrolase-1/cd39 expression to microglia and vasculature of the brain. Eur J
447
Neurosci 12:4357.
448
32
Kulesskaya, N., Voikar, V., Peltola, M., Yegutkin, G. G., Salmi, M., Jalkanen, S., and
449
Rauvala, H. 2013. CD73 is a major regulator of adenosinergic signalling in mouse brain.
450
PLoS One 8:e66896.
451
23
33
Gunther, G. R. and Herring, M. B. 1991. Inhibition of neutrophil superoxide production
452
by adenosine released from vascular endothelial cells. Ann Vasc Surg 5:325.
453
34
Latini, S. and Pedata, F. 2001. Adenosine in the central nervous system: release
454
mechanisms and extracellular concentrations. J Neurochem 79:463.
455
35
Carman, C. V. and Martinelli, R. 2015. T Lymphocyte-Endothelial Interactions:
456
Emerging Understanding of Trafficking and Antigen-Specific Immunity. Frontiers in
457
immunology 6:603.
458
36
Siffrin, V., Radbruch, H., Glumm, R., Niesner, R., Paterka, M., Herz, J., Leuenberger, T.,
459
Lehmann, S. M., Luenstedt, S., Rinnenthal, J. L., Laube, G., Luche, H., Lehnardt, S.,
460
Fehling, H. J., Griesbeck, O., and Zipp, F. 2010. In vivo imaging of partially reversible
461
th17 cell-induced neuronal dysfunction in the course of encephalomyelitis. Immunity
462
33:424.
463
37
Monks, C. R., Kupfer, H., Tamir, I., Barlow, A., and Kupfer, A. 1997. Selective
464
modulation of protein kinase C-theta during T-cell activation. Nature 385:83.
465
38
Fredholm, B. B. 2007. Adenosine, an endogenous distress signal, modulates tissue
466
damage and repair. Cell Death Differ 14:1315.
467
39
Csoka, B., Himer, L., Selmeczy, Z., Vizi, E. S., Pacher, P., Ledent, C., Deitch, E. A.,
468
Spolarics, Z., Nemeth, Z. H., and Hasko, G. 2008. Adenosine A2A receptor activation
469
inhibits T helper 1 and T helper 2 cell development and effector function. FASEB J
470
22:3491.
471
40
Lappas, C. M., Rieger, J. M., and Linden, J. 2005. A2A adenosine receptor induction
472
inhibits IFN-gamma production in murine CD4+ T cells. Journal of immunology
473
174:1073.
474
24
41
Klotz, K. N. 2000. Adenosine receptors and their ligands. Naunyn Schmiedebergs Arch
475
Pharmacol 362:382.
476
42
El-Tayeb, A., Michael, S., Abdelrahman, A., Behrenswerth, A., Gollos, S., Nieber, K.,
477
and Muller, C. E. 2011. Development of Polar Adenosine A2A Receptor Agonists for
478
Inflammatory Bowel Disease: Synergism with A2B Antagonists. ACS Med Chem Lett
479
2:890.
480
43
Kenakin, T. 2001. Inverse, protean, and ligand-selective agonism: matters of receptor
481
conformation. FASEB J 15:598.
482
44
Prather, P. L. 2004. Inverse agonists: tools to reveal ligand-specific conformations of G
483
protein-coupled receptors. Sci STKE 2004:pe1.
484
45
Nakano, K., Yamaoka, K., Hanami, K., Saito, K., Sasaguri, Y., Yanagihara, N., Tanaka,
485
S., Katsuki, I., Matsushita, S., and Tanaka, Y. 2011. Dopamine induces IL-6-dependent
486
IL-17 production via D1-like receptor on CD4 naive T cells and D1-like receptor
487
antagonist SCH-23390 inhibits cartilage destruction in a human rheumatoid
488
arthritis/SCID mouse chimera model. Journal of immunology 186:3745.
489
46
Alnouri, M. W., Jepards, S., Casari, A., Schiedel, A. C., Hinz, S., and Muller, C. E. 2015.
490
Selectivity is species-dependent: Characterization of standard agonists and antagonists at
491
human, rat, and mouse adenosine receptors. Purinergic Signal 11:389.
492
47
Kanda, T., Jackson, M. J., Smith, L. A., Pearce, R. K., Nakamura, J., Kase, H., Kuwana,
493
Y., and Jenner, P. 1998. Adenosine A2A antagonist: a novel antiparkinsonian agent that
494
does not provoke dyskinesia in parkinsonian monkeys. Ann Neurol 43:507.
495
48
Ohta, A., Gorelik, E., Prasad, S. J., Ronchese, F., Lukashev, D., Wong, M. K., Huang, X.,
496
Caldwell, S., Liu, K., Smith, P., Chen, J. F., Jackson, E. K., Apasov, S., Abrams, S., and
497
25
Sitkovsky, M. 2006. A2A adenosine receptor protects tumors from antitumor T cells.
498
Proceedings of the National Academy of Sciences of the United States of America
499
103:13132.
500
49
Kobie, J. J., Shah, P. R., Yang, L., Rebhahn, J. A., Fowell, D. J., and Mosmann, T. R.
501
2006. T regulatory and primed uncommitted CD4 T cells express CD73, which
502
suppresses effector CD4 T cells by converting 5'-adenosine monophosphate to adenosine.
503
Journal of immunology 177:6780.
504
50
Deaglio, S., Dwyer, K. M., Gao, W., Friedman, D., Usheva, A., Erat, A., Chen, J. F.,
505
Enjyoji, K., Linden, J., Oukka, M., Kuchroo, V. K., Strom, T. B., and Robson, S. C. 2007.
506
Adenosine generation catalyzed by CD39 and CD73 expressed on regulatory T cells
507
mediates immune suppression. J Exp Med 204:1257.
508
51
Moesta, A. K., Li, X. Y., and Smyth, M. J. 2020. Targeting CD39 in cancer. Nat Rev
509
Immunol.
510
52
Tuting, T. and de Visser, K. E. 2016. CANCER. How neutrophils promote metastasis.
511
Science 352:145.
512
53
Coffelt, S. B., Kersten, K., Doornebal, C. W., Weiden, J., Vrijland, K., Hau, C. S.,
513
Verstegen, N. J. M., Ciampricotti, M., Hawinkels, L., Jonkers, J., and de Visser, K. E.
514
2015. IL-17-producing gammadelta T cells and neutrophils conspire to promote breast
515
cancer metastasis. Nature 522:345.
516
54
Bald, T., Quast, T., Landsberg, J., Rogava, M., Glodde, N., Lopez-Ramos, D., Kohlmeyer,
517
J., Riesenberg, S., van den Boorn-Konijnenberg, D., Homig-Holzel, C., Reuten, R.,
518
Schadow, B., Weighardt, H., Wenzel, D., Helfrich, I., Schadendorf, D., Bloch, W.,
519
Bianchi, M. E., Lugassy, C., Barnhill, R. L., Koch, M., Fleischmann, B. K., Forster, I.,
520
26
Kastenmuller, W., Kolanus, W., Holzel, M., Gaffal, E., and Tuting, T. 2014.
521
Ultraviolet-radiation-induced inflammation promotes angiotropism and metastasis in
522
melanoma. Nature 507:109.
523
55
El Rayes, T., Catena, R., Lee, S., Stawowczyk, M., Joshi, N., Fischbach, C., Powell, C.
524
A., Dannenberg, A. J., Altorki, N. K., Gao, D., and Mittal, V. 2015. Lung inflammation
525
promotes metastasis through neutrophil protease-mediated degradation of Tsp-1.
526
Proceedings of the National Academy of Sciences of the United States of America
527
112:16000.
528
56
Wculek, S. K. and Malanchi, I. 2015. Neutrophils support lung colonization of
529
metastasis-initiating breast cancer cells. Nature 528:413.
530
57
Templeton, A. J., McNamara, M. G., Seruga, B., Vera-Badillo, F. E., Aneja, P., Ocana,
531
A., Leibowitz-Amit, R., Sonpavde, G., Knox, J. J., Tran, B., Tannock, I. F., and Amir, E.
532
2014. Prognostic role of neutrophil-to-lymphocyte ratio in solid tumors: a systematic
533
review and meta-analysis. J Natl Cancer Inst 106:dju124.
534
58
Yazici, H., Seyahi, E., Hatemi, G., and Yazici, Y. 2018. Behcet syndrome: a
535
contemporary view. Nat Rev Rheumatol 14:107.
536
59
Meyts, I. and Aksentijevich, I. 2018. Deficiency of Adenosine Deaminase 2 (DADA2):
537
Updates on the Phenotype, Genetics, Pathogenesis, and Treatment. J Clin Immunol
538
38:569.
539
540
541
27
Abbreviations: Th, T-helper; TGF, tumor growth factor; IL, interleukin; APCs, antigen
542
presenting cells; MLR, two-way mixed lymphocyte reaction; EAE, experimental autoimmune
543
encephalomyelitis; BM, bone marrow; BM-DC, bone-marrow-derived dendritic cell; CD3/CD28,
544
agonistic anti-CD3/CD28 antibodies; PLP, myelin proteolipid protein; PLP peptide, I-As
545
restricted helper peptide derived from the PLP; CFA, complete Freund’s adjuvant; CCR,
546
chemokine
receptor;
FITC,
fluorescein
isothiocyanate;
PE,
phycoerythrin;
MACS,
547
magnetic-activated cell sorting; Ab, antibody; n, number of repeat experiments; SD, standard
548
deviation.
549
550
Acknowledgments
551
This work was supported by a Grant-in-Aid for Scientific Research (C) (no. 19K07201),
552
awarded to M.K., a Grant-in-Aid for Young Scientists (B) (no. 18K15327) to R.T., and a
553
Grant-in-Aid for Scientific Research (C) (no. 19K08887) awarded to S.M. by the Japanese
554
Society for the Promotion of Science. This work was also supported by the 44th and 45th
555
Science Research Promotion Fund, awarded to M.K. by the Promotion and Mutual Aid
556
Corporation for Private Schools of Japan.
557
558
Author contributions
559
M.T., R.T., S.M., and M.K., performed the experiments. M.T., S.M., and M.K., conceived and
560
designed the experiments. M.T., S.M., T.Y., and M.K., wrote the manuscript. All authors
561
discussed the results and commented on the manuscript.
562
563
Conflicts of interest
564
28
Sho Matsushita is an employee of iMmno, Inc.
565
The other authors have no conflicts of interest to declare.
566
29
Figure legends
567
568
Fig. 1. Adenosine induces hypersecretion of IL-17A by CD4+ T cells in an MLR. A–C, An
569
MLR was performed for 7 days in the presence of adenosine (0–1 mM). After 7 days, cells were
570
stained with anti-CD4 (x-axis) and IL-17A (y-axis) antibodies (Abs), followed by flow
571
cytometry analysis (A, n (number of repeat experiments) = 4). The percentage of
572
IL-17A-producing CD4+ T cells within the total CD4+ T cell population is shown (B, n = 6–9).
573
Cells supernatants were analyzed in an IL-17A ELISA (C, n = 6–9). D, The effects of the A2aR
574
on IL-17A production in the presence of PSB0777 (an A2aR agonist) (left, n = 6–9),
575
Istradefylline (an A2aR antagonist) plus adenosine (100 µM) (right, n = 6–9). E, The effects of
576
A2aR signaling on IL-17A production in the presence of MDL-12330A (an adenyl cyclase
577
inhibitor) plus adenosine (100 µM) (left, n = 4–6), or H-89 (a protein kinase A inhibitor) plus
578
adenosine (100 µM) (right, n = 4–6) were analyzed in an IL-17A ELISA. F, The effects of
579
CD39/CD73 inhibitors on IL-17A production in the presence of ARL67156 (a CD39 inhibitor)
580
plus ATP (100 µM) (left, n = 4 – 6) or AMP-CP (a CD73 inhibitor) plus ATP (100 µM) (right, n
581
= 4–6) were analyzed in an IL-17A ELISA. G, The effects of an A2aR antagonist and
582
CD39/CD73 inhibitors on basal IL-17A production in the presence of Istradefylline (left, n =
583
4–7), ARL67156 (center, n = 4–7), or AMP-CP (right, n = 4–7) were analyzed in an IL-17A
584
ELISA. Data are expressed as the mean ± standard deviation (SD) and were compared using an
585
unpaired Student’s t-test (B) or one-way ANOVA with Tukey’s post-hoc test (C–G). *P < 0.05
586
and **P < 0.01, compared with medium.
587
588
Fig. 2. Adenosine production in the MLR. A, Concentrations of adenosine or ATP in MLR
589
30
supernatants were measured in an adenosine (left, n = 4–6) or ATP (right, n = 4–6) ELISA (0–24
590
h). B and C, The effects of CD39 (B, ARL67156, n = 4–6) and CD73 (C, AMP-CP, n = 4–6)
591
inhibitors on production of adenosine or ATP in the MLR were analyzed in an adenosine (left) or
592
ATP (right) ELISA. D, E and F, Levels of adenosine and ATP in the supernatants of
593
CD3/CD28-stimulated CD4+ T cells (D, n = 4), LPS-stimulated B cells (E, n = 4), and
594
LPS-stimulated BM-DCs (F, n = 4) were analyzed in adenosine and ATP ELISAs at 24 h
595
post-stimulation in the presence of CD39 or CD73 inhibitors. Data are expressed as the mean ±
596
SD and were compared using one-way ANOVA with Tukey’s post-hoc test. *P < 0.05 and **P <
597
0.01, compared with medium (B and C), CD3/CD28 stimulation (D), or LPS stimulation (E and
598
F).
599
600
Fig. 3. Adenosine induces hypersecretion of IL-17A by Th17 cells. A and B, CD4+ T cells
601
were stimulated for 1–7 days with anti-CD3/CD28 antibodies in the presence of adenosine (0–1
602
mM). After stimulation, the supernatants were analyzed in an IL-17A ELISA (n = 4–6). C,
603
Adenosine (600 µM) was added at 0–3 days after CD3/CD28 stimulation. At 7 days
604
post-CD3/CD28 stimulation, supernatants were analyzed in an IL-17A ELISA (n = 4). D, Effects
605
of the A2aR on IL-17A production in the presence of PSB0777 (an A2aR agonist) (left, n = 4–6),
606
Istradefylline (an A2aR antagonist) plus adenosine (600 µM) (right, n = 4–6). E, Effects of
607
A2aR signaling on IL-17A production in the presence of MDL-12330A (an adenyl cyclase
608
inhibitor) plus adenosine (600 µM) (left, n = 4–6), or H-89 (a protein kinase A inhibitor) plus
609
adenosine (600 µM) (right, n = 4–6) were analyzed in an IL-17A ELISA. F, CD4+CD62L+ and
610
CD4+CD62L+FT cells were stimulated for 7 days with anti-CD3/CD28 antibodies in the
611
presence of adenosine (0–1 mM). After 7 days, supernatants were analyzed in an IL-17A ELISA
612
31
(n = 4). G, Subsets of CD4+ T cells were stimulated for 7 days by anti-CD3/CD28 antibodies in
613
the presence of adenosine (600 µM) after isolation of each CCR cell type (high (hi) and low (lo)
614
expression) (n = 4). Data are expressed as the mean ± SD and were compared using an unpaired
615
Student’s t-test (g) or one-way ANOVA with Tukey’s post-hoc test (A-F). *P < 0.05 and **P <
616
0.01, compared with CD3/CD28 stimulation (A–C, D, left, and F) or CD3/28 stimulation plus
617
adenosine (600 µM) (D, right and E).
618
619
Fig. 4. Effect of adenosine and an A2aR antagonist on production of Th17-related cytokines.
620
A, An MLR was performed for 7 days in the presence of adenosine (0–1 mM) (top row),
621
PSB0777 (an A2aR agonist) (second row), or Istradefylline (an A2aR antagonist) plus adenosine
622
(100 µM) (bottom row). At 7 days post-incubation, the supernatants were analyzed in an
623
IL-17A (left), IL-17F (center), or IL-22 (right) ELISA. All experiments were repeated six to nine
624
times. B, CD4+ T cells were stimulated with an anti-CD3/CD28 antibody for 7 days in the
625
presence of adenosine (600 µM) (top row), PSB0777 (an A2aR agonist) (second row), or
626
Istradefylline (an A2aR antagonist) plus adenosine (600 µM) (bottom row). After stimulation,
627
supernatants were analyzed in IL-17A (left), IL-17F (center), and IL-22 (right) ELISAs. All
628
experiments were repeated four to six times. Data are expressed as the mean ± SD and were
629
compared using one-way ANOVA with Tukey’s post-hoc test. *P < 0.05 and **P < 0.01,
630
compared with CD3/CD28 stimulation.
631
632
Fig. 5. Suppression of adenosine-mediated hypersecretion of IL-17A ameliorates EAE. A
633
and B, SJL/J CD4+ T cells were stimulated for 7 days by anti-CD3/CD28 antibodies and an
634
A2aR antagonist (A, n = 4). Naïve CD4+ T cells were stimulated for 7 days by anti-CD3/CD28
635
32
antibodies in the presence of IL-6, TGF-β1, and Istradefylline (an A2aR antagonist) (0–1 nM) (B,
636
left, n = 4). Alternatively, naïve CD4+ T cells were stimulated for 7 days by anti-CD3/CD28
637
antibodies in the presence of IL-6 and TGF-β1, followed by another 7 day incubation with
638
anti-CD3/CD28 antibodies and Istradefylline (0–1 nM) (B, right, n = 4). After the stimulation,
639
supernatants were analyzed in an IL-17A ELISA. C, Mice were immunized subcutaneously with
640
PLP peptide emulsified in CFA (PLP peptide/CFA). At 7 days post-immunization, splenocytes
641
were incubated for 3 days with PLP peptide in the presence of Istradefylline and adenosine (600
642
µM). After the incubation, supernatants were analyzed in an IL-17A ELISA (c, n = 4). D, To
643
induce EAE, SJL/J mice were immunized with PLP peptide/CFA. Before and post-immunization
644
with the PLP peptide (Days 0 to 18), mice received oral Istradefylline (6 µg/mouse) or water
645
once every 2 days. Clinical scores were recorded every day during EAE induction (D, n = 15).
646
Data were obtained from five independent experiments (n = 3 mice/group). (E) At 18 days
647
post-immunization, spinal cord sections from mice administered oral Istradefylline (right) or
648
water (left) were stained with hematoxylin and eosin (upper panels) or with an anti-mouse CD3
649
antibody (lower panels) (Scale bar, 100 µm). Data are representative of five independent
650
experiments (all with similar results) and are expressed as the mean ± SD. Data were compared
651
using one-way ANOVA with Tukey’s post-hoc test (A–C), or using a non-parametric
652
Mann-Whitney U-test (D). *P < 0.05 and **P < 0.01, compared with CD3/CD28 stimulation
653
plus adenosine (600 µM) (A), CD3/CD28 stimulation plus IL-6 and TGF-β1 (B), or PLP peptide
654
pulse (C).
655
33
Legends for the supplementary figures
656
657
Sup. Fig. 1. Purity and viability of each immune cell subset. A, First row (left): Splenocytes
658
or isolated CD4+ T cells were stained with a FITC-conjugated anti-mouse CD4 antibody (Ab)
659
(BioLegend) (x-axis) and a PE-conjugated anti-mouse CD3 Ab (BioLegend) (y-axis). Second
660
row (left): Splenocytes or CD4+CD62L+ T cells isolated by MACS were stained with a
661
FITC-conjugated anti-mouse CD4 Ab (BioLegend) (x-axis) and a PE-conjugated anti-mouse
662
CD62L Ab (BioLegend) (y-axis). First row (right): Splenocytes and B cells were stained with a
663
FITC-conjugated anti-mouse CD19 Ab (BioLegend) (x-axis) and cell numbers were counted
664
(y-axis). Second row (right): Splenocytes, BM leukocytes, and BM-DCs were stained with a
665
FITC-conjugated anti-mouse CD11c Ab (BioLegend) (x-axis) and cell numbers were counted
666
(y-axis). The number in each panel represents the percentage of each immune subset within the
667
total cell population. B, First row: Isolated CD4+ (by MACS)-, CD4+CD62L- (by cell sorting)-,
668
or CD4+CD62L+ (by cell sorting) T cells were analyzed with a PE-conjugated anti-mouse
669
CD62L Ab (BioLegend) (x-axis) and cell numbers were counted (y-axis). The number in each
670
panel represents the percentage of each immune subset (- or +) within the total cell population (-
671
plus +). Data are representative of at least three repeat experiments. Second row: Isolated CD4+
672
(by MACS), CD4+CCR3low (lo) (by cell sorting), or CD4+CCR3high (hi) (by cell sorting) T cells
673
were analyzed with a PE-conjugated anti-mouse CCR3 Ab (BioLegend) (x-axis) and cell
674
numbers were counted (y-axis). Third row: Isolated CD4+ (by MACS), CD4+CCR5lo (by cell
675
sorting), or CD4+CCR5hi (by cell sorting) T cells were analyzed with a PE-conjugated
676
anti-mouse CCR5 Ab (BioLegend) (x-axis) and cell numbers were counted (y-axis). Fourth row:
677
Isolated CD4+ (by MACS), CD4+CCR6lo (by cell sorting), or CD4+CCR6hi (by cell sorting) T
678
34
cells were analyzed with a PE-conjugated anti-mouse CCR6 Ab (BioLegend) (x-axis) and cell
679
numbers were counted (y-axis). Fifth row: Isolated CD4+ (by MACS), CD4+CD25lo (by cell
680
sorting), or CD4+CD25hi (by cell sorting) T cells were analyzed with a PE-conjugated anti-mouse
681
CD25 Ab (BioLegend) (x-axis) and cell numbers were counted (y-axis). The number in each
682
panel represents the mean percentage ± SD of each immune subset (hi or lo) within the total cell
683
population (hi plus lo). Data are representative of at least three repeat experiments. C, After cell
684
isolation, each immune cell type was mixed with Trypan blue. Viability was calculated as the
685
number of unstained cells/(stained cells + unstained cells) × 100. Each measurement was
686
performed at least three times. The percentage represents the mean percentage ± SD.
687
688
Sup. Fig. 2. Effects of an adenosine receptor agonist and antagonist on production of
689
IL-17A in an MLR. MLR was performed for 7 days in the presence of each adenosine receptor
690
agonist or each adenosine receptor antagonist plus adenosine (100 µM). After 7 days, the
691
supernatants were analyzed in an IL-17A ELISA. A: CCPA (an A1R agonist). B: PSB0777 (an
692
A2aR agonist, left) and Istradefylline (an A2aR antagonist, right). C: BAY 60-653 (an A2bR
693
agonist). D: HEMADO (an A3R agonist). All experiments were repeated six to nine times. Data
694
are expressed as the mean ± SD and were compared using one-way ANOVA with Tukey’s
695
post-hoc test. *P < 0.05 and **P < 0.01, compared with medium.
696
697
Sup. Fig. 3. Effect of adenosine and an A2aR antagonist on production of IFN-γ, IL-5, and
698
IL-17A by CD3/CD28-stimulated CD4+ T cells. CD4+ T cells were stimulated for 1–7 days
699
with anti-CD3/CD28 antibodies in the presence or absence of adenosine (600 µM). After
700
stimulation, supernatants were analyzed in IFN-γ (top row), IL-5 (second row), and IL-17A
701
35
(bottom row) ELISAs. Data are expressed as the mean ± SD and were compared using one-way
702
ANOVA with Tukey’s post-hoc test. *P < 0.05 and **P < 0.01, compared with CD3/CD28
703
stimulation. All experiments were repeated four to six times.
704
705
Sup. Fig. 4. Effects of an adenosine receptor agonist and antagonist on production of
706
IL-17A by CD3/CD28-stimulated CD4+ T cells. CD4+ T cells were stimulated with an
707
anti-CD3/CD28 antibody for 7 days in the presence of each adenosine receptor agonist, or in the
708
presence of each adenosine receptor antagonist plus adenosine (600 µM). After 7 days, the
709
supernatants were analyzed in an IL-17A ELISA. A: CCPA (an A1R agonist). B: PSB0777 (an
710
A2aR agonist, left) and Istradefylline (an A2aR antagonist, right). C: BAY 60-653 (an A2bR
711
agonist). D: HEMADO (an A3R agonist). All experiments were repeated six to nine times. Data
712
are expressed as the mean ± SD and were compared using one-way ANOVA with Tukey’s
713
post-hoc test. *P < 0.05 and **P < 0.01, compared with medium.
714
Adenosine100 μM
Fig. 1
A
B
Medium
0.01
0.1
1
10
IL-17A (pg/mL)
D
**
*
PSB0777 (μM)
0
100
200
300
400
500
Medium
Medium
0.01
0.1
1
Adenosine 100 μM
**
IL-17A (pg/mL)
0
500
1000
1500
2000
**
**
Istradefylline (nM)
PE-anti-mouse IL-17A Ab
FITC-anti-mouse CD4 Ab
medium
0.66
0.16
IL-17A (pg/mL)
0
50
100
150
200
250
Medium
0.01
0.1
1
Istradefylline (nM)
**
**
**
Medium
Medium
0.01
0.1
1
H89 (μM)
Adenosine 100 μM
** ** **
Medium
Medium
0.01
0.1
1
MDL-12330A (μM)
Adenosine 100 μM
0
50
100
150
200
250
300
*
** **
IL-17A (pg/mL)
0
200
400
600
800
1000
IL-17A (pg/mL)
IL-17A (pg/mL)
Medium
0.01
0.1
1
AMP-CP (μM)
**
**
0
50
100
150
200
250
300
350
ARL67156 (μM)
**
IL-17A (pg/mL)
0
50
100
150
200
250
300
Medium
0.01
0.1
1
ARL67156
(μM)
Medium
ATP 100 μM
**
**
IL-17A (pg/mL)
0
100
200
300
400
500
Medium
0.01
0.1
1
AMP-CP
(μM)
IL-17A (pg/mL)
0
100
200
300
400
500
600
Medium
ATP 100 μM
Medium
0.01
0.1
1
**
IL-17A (pg/mL)
IL-17A (%)
**
Medium
100
Adenosine (μM)
C
0
0.4
0.8
1.2
IL-17A (pg/mL)
**
0
500
1000
1500
2000
Medium
100
Adenosine (μM)
E
F
G
Medium
1
10
300
600
1000
100
Adenosine (μM)
**
**
**
**
2000
0
4000
6000
**
5
10
15
20
25
30
0
CD4+ T cells
Fig. 2
Adenosine (μM)
0
5
10
15
20
25
30
3hr
6hr
12hr
24hr
30min
1hr
15min
0min
ATP(μM)
0
1
2
3
4
5
6
3hr
6hr
12hr
24hr
30min
1hr
15min
0min
A
B
Adenosine (μM)
ARL67156 (μM)
Medium
0.01
0.1
1
ATP (μM)
0
1
2
3
4
5
ARL67156 (μM)
Medium
0.01
0.1
1
C
Adenosine (μM)
0
5
10
15
20
25
30
Medium
0.01
0.1
1
AMP-CP (μM)
ATP (μM)
0
1
2
3
4
5
Medium
0.01
0.1
1
AMP-CP (μM)
MLR
MLR
MLR
ATP (μM)
CD3/CD28
0
0.5
1
1.5
2
1.5
3
Medium
AMP-CP 1 μM
ARL67156 1 μM
Medium
D
E
F
Medium
ARL67156 1 μM
AMP-CP 1 μM
Medium
8
0
2
4
6
10
12
14
Adenosine (μM)
ATP (μM)
0
2
4
6
8
10
12
Medium
ARL67156 1 μM
AMP-CP 1 μM
Medium
0
0.5
1
1.5
2
2.5
3
ATP (μM)
0
10
20
30
40
50
60
Adenosine (μM)
B cells
BM-DCs
**
**
Adenosine (μM)
0
2
4
6
8
10
Medium
AMP-CP 1 μM
ARL67156 1 μM
Medium
CD3/CD28
**
****
****
****
0
LPS
LPS
Medium
AMP-CP 1 μM
ARL67156 1 μM
Medium
Medium
AMP-CP 1 μM
ARL67156 1 μM
Medium
LPS
LPS
IL-17A (pg/mL)
Medium
Medium
0.01
0.1
1
Istradefylline (nM)
CD3/CD28
**
** **
0
100
200
300
400
500
Adenosine 600 μM
Medium
300
600
1000
100
Adenosine (μM)
**
**
**
0
200
400
600
800
CD3/CD28
IL-17A (pg/mL)
A
B
day1
day3
day5
day7
Adenosine 600 μM
CD3/CD28
Medium
0
100
200
300
400
500
IL-17A (pg/mL)
C
0 hr
6 hr
24 hr
Medium
72 hr
0
100
200
300
400
500
Adenosine 600 μM
CD3/CD28
IL-17A (pg/mL)
Medium
Medium
0.01
0.1
1
H89 (μM)
**
** **
0
100
200
300
400
500
IL-17A (pg/mL)
CD3/CD28
Adenosine 600 μM
Medium
Medium
0.01
0.1
1
**
**
**
0
100
200
300
400
500
MDL-12330A (μM)
IL-17A (pg/mL)
CD3/CD28
Adenosine 600 μM
Medium
Medium
1
10
100
1000
1
10
100
1000
Adenosine (μM)
Adenosine (μM)
CD4+CD62L+
CD3/CD28
400
800
1200
IL-17A (pg/mL)
CD3/CD28
Flow Through
F
G
Fig. 3
D
E
**
0
20
40
60
1000
2000
3000
Adenosine 600 μM
CD62L
CCR3
CCR5
CCR6
CD25
IL-17A (pg/mL)
**
*
**
**
**
**
**
0
- +
- +
-
+
lo hi lo hi lo hi
lo hi
- +
- +
- +
- +
- +
- +
- +
- +
CD3/CD28
Medium
0.01
0.1
1
10
CD3/CD28
IL-17A (pg/mL)
PSB0777 (μM)
*
0
50
100
150
200
250
Fig. 4
**
**
**
0
200
400
600
800
CD3/CD28
0
200
400
600
800
1000
IL-17F (pg/mL)
**
**
0
200
400
600
800
**
**
**
IL-22 (pg/mL)
IL-17A (pg/mL)
IL-17A (pg/mL)
IL-17F (pg/mL)
IL-22 (pg/mL)
IL-17A (pg/mL)
IL-17F (pg/mL)
IL-22 (pg/mL)
0
200
400
600
800
**
**
*
0
100
200
300
400
500
0
200
400
600
800
**
** **
**
** **
0
100
200
300
400
500
0
200
400
600
** ** **
IL-17F
IL-17A
IL-22
Medium
100
300
600
1000
Adenosine (μM)
CD3/CD28
CD3/CD28
Medium
100
300
600
1000
Adenosine (μM)
CD3/CD28
Medium
100
300
600
1000
Adenosine (μM)
PSB0777 (μM)
Medium
0.01
0.1
1
10
CD3/CD28
PSB0777 (μM)
Medium
0.01
0.1
1
10
CD3/CD28
PSB0777 (μM)
Medium
0.01
0.1
1
10
Medium
0.01
0.1
1
Adenosine 600 μM
Istradefylline (nM)
Medium
CD3/CD28
Medium
0.01
0.1
1
Adenosine 600 μM
Istradefylline (nM)
Medium
CD3/CD28
Medium
0.01
0.1
1
Adenosine 600 μM
Istradefylline (nM)
Medium
CD3/CD28
*
**
*
0
50
100
150
200
250
CD3/CD28-stimulated CD4+ T cells
IL-17A (pg/mL)
0
600
1200
1800
IL-17F (pg/mL)
Adenosine (μM)
Medium
100
300
600
1000
**
**
**
**
** ** **
0
2000
4000
6000
IL-17F
IL-17A
IL-22
0
1000
2000
3000
4000
5000
IL-22 (pg/mL)
**
**
Adenosine (μM)
Medium
100
300
600
1000
Adenosine (μM)
Medium
100
300
600
1000
**
IL-17F (pg/mL)
IL-17A (pg/mL)
**
*
0
100
200
300
400
500
0
200
600
1000
1400
IL-22 (pg/mL)
**
*
PSB0777 (μM)
Medium
0.01
0.1
1
10
0
600
1200
1800
** **
**
IL-22 (pg/mL)
0
400
800
1200
** ** **
IL-17F (pg/mL)
Medium
0.01
0.1
1
Adenosine 100 μM
**
IL-17A (pg/mL)
0
500
1000
1500
2000
**
**
Istradefylline (nM)
Medium
Medium
0.01
0.1
1
Adenosine 100 μM
Istradefylline (nM)
Medium
Medium
0.01
0.1
1
Adenosine 100 μM
Istradefylline (nM)
Medium
PSB0777 (μM)
Medium
0.01
0.1
1
10
PSB0777 (μM)
Medium
0.01
0.1
1
10
** **
0
200
400
600
800
1000
MLR
A
B
CD4+ T cells
E
C
A
B
IL-17A (pg/mL)
IL-17A (pg/mL)
0
500
1000
1500
2000
Medium
0.01
0.1
1
Istradefylline (nM)
IL-17A (pg/mL)
IL-17A (pg/mL)
Fig. 5
water (orally)
Isradefylline (orally)
PLP peptide/CFA
0
1000
2000
3000
4000
5000
Medium
Medium
0.01
0.1
1
CD3/CD28
Adenosine 600 μM
**
*
**
0
1000
2000
3000
4000
5000
Medium
Medium
0.1
1
Istradefylline (nM)
CD3/CD28
Adenosine 600 μM
0
1000
2000
3000
4000
5000
Medium
0.1
1
Istradefylline
(nM)
Istradefylline
(nM)
Medium
PLP peptide
Medium
Adenosine 600 μM
Medium
0.1
1
PLP peptide
**
**
*
**
*
* *
**
**
Days post induction
Mean clinical score
0
1
2
3
4
5
0
5
10
15
18
D
PLP peptide/CFA + Istradefylline (orally) (6 μg/day)
PLPpeptide/CFA + water (orally)
Istradefylline (nM)
CD3/CD28
IL-6 + TGF-β1
Naïve CD4+ T cells
Th17 induction
After Th17 induction
PLP peptide/CFA
HE
CD3
**
**
BM-DCs
94.5 ± 1.67
Sup. Fig. 1
B
C
Cell type
Viability
CD4+ T cells
CD4+CD62L- T cells
CD4+CD62L+ T cells
CD4+CCR3lo T cells
CD4+CCR3hi T cells
Splenocytes
100 ± 0
B cells
100 ± 0
Naive CD4+ T cells
100 ± 0
Flow through of
CD4+CD62L+ T cells
100 ± 0
CD4+ T cells
100 ± 0
Cell type
Viability
CD4+CD62Llo
95.6 ± 2.90
CD4+CD62Lhi
98.3 ± 2.50
CD4+CCR3lo
97.1 ± 1.91
CD4+CCR3hi
91.1 ± 6.37
CD4+CCR5lo
98.8 ± 0.89
CD4+CCR5hi
91.1 ± 6.40
CD4+CCR6lo
98.8 ± 0.80
CD4+CCR6hi
94.0 ± 4.24
CD4+CD25lo
97.1 ± 1.77
CD4+CD25hi
98.4 ± 1.62
Splenocytes
49.7
± 0.18
A
Isolated CD4+ T cells
Flow through fraction in an isolation
of CD4+CD62L+ T cells
Isolated CD4+CD62L+ T cells
5.40
± 0.21
CD4+CCR5lo T cells
CD4+CCR5hi T cells
CD4+CCR6lo T cells
CD4+CCR6hi T cells
CD4+CD25lo T cells
CD4+CD25hi T cells
78.7
± 3.81
21.0
± 4.01
BM leukocytes
Splenocytes
29.3
± 0.48
BM-DCs
Splenocytes
CD4+ T cells
CD4+ T cells
CD4+ T cells
CD4+ T cells
Bcells
Splenocytes
PE-anti-mouse CD3 Ab
FITC-anti-mouse CD4 Ab
PE-anti-mouse CD62L Ab
FITC-anti-mouse CD4 Ab
PE-anti-mouse CD62L Ab
Cell count
2.85
± 0.16
97.1
± 0.15
Cell count
FITC-anti-mouse CD19 Ab
Cell count
FITC-anti-mouse CD11c Ab
PE-anti-mouse CCR3 Ab
PE-anti-mouse CCR5 Ab
PE-anti-mouse CCR6 Ab
PE-anti-mouse CD25 Ab
43.5
± 0.69
4.48
± 0.12
94.6
± 0.20
96.9
± 0.15
50.9
± 0.48
45.4
± 0.43
Cell count
4.30
± 0.08
95.7
± 0.08
Cell count
14.6
± 1.91
85.4
± 1.86
Cell count
21.1
± 1.65
78.9
± 1.65
Cell count
98.7
± 0.24
69.0
± 0.85
99.9
± 0.08
80.5
± 0.96
99.9
± 0.04
84.9
± 0.54
99.9
± 0.05
99.9
± 0.05
99.8
± 0.26
95.7
± 0.24
4.00
± 0.37
95.9
± 0.21
10.8
± 0.79
93.4
± 0.37
Medium
0.01
0.1
1
10
HEMADO (μM)
IL-17A (pg/mL)
IL-17A (pg/mL)
Medium
0.01
0.1
1
10
IL-17A (pg/mL)
**
*
PSB0777 (μM)
100
200
300
400
500
0
100
200
300
400
500
Medium
0.01
0.1
1
10
CCPA (μM)
Medium
0.01
0.1
1
10
BAY 60-6583 (μM)
0
200
400
600
0
100
200
300
400
500
IL-17A (pg/mL)
IL-17A (pg/mL)
Medium
Medium
0.01
0.1
1
**
0
500
1000
1500
2000
**
**
Istradefylline (nM)
Sup. Fig. 2
Adenosine 100 μM
0
A
B
C
D
Sup. Fig. 3
IL-17A(pg/mL)
*
*
**
**
**
Adenosine 600 μM
CD3/CD28
(-) (+)
1
7
5
3 days
(+)
(+)
(+)
(-)
(-)
(-)
Adenosine 600 μM
CD3/CD28
(-) (+)
1
7
5
3 days
(+)
(+)
(+)
(-)
(-)
(-)
0
1000
800
600
400
200
IL-5 (pg/mL)
Adenosine 600 μM
CD3/CD28
(-) (+)
1
7
5
3 days
(+)
(+)
(+)
(-)
(-)
(-)
0
0
500
400
300
200
100
IFN-γ (pg/mL)
400
800
1200
1600
Sup. Fig. 4
A
B
C
D
Medium
0.01
0.1
1
10
CD3/CD28
CCPA (µM)
PSB0777 (µM)
Medium
0.01
0.1
1
10
CD3/CD28
*
*
Medium
0.01
0.1
1
10
CD3/CD28
BAY60-6853 (µM)
Medium
Medium
0.01
0.1
1
Istradefylline (nM)
CD3/CD28
**
**
Adenosine 600 µM
Medium
0.01
0.1
1
10
CD3/CD28
HEMADO (µM)
IL-17A (pg/mL)
*
0
50
100
150
200
250
IL-17A (pg/mL)
0
100
200
300
IL-17A (pg/mL)
0
100
200
300
400
500
IL-17A (pg/mL)
0
100
200
300
IL-17A (pg/mL)
0
100
200
300
| 2021 | Extracellular adenosine induces hypersecretion of IL-17A by T-helper 17 cells through the adenosine A2a receptor to promote neutrophilic inflammation | 10.1101/2021.04.29.441713 | [
"Tokano Mieko",
"Matsushita Sho",
"Takagi Rie",
"Yamamoto Toshimasa",
"Kawano Masaaki"
] | creative-commons |
1
Neurotrophin signaling is modulated by specific transmembrane domain
interactions
María L. Franco1,4, Kirill D. Nadezhdin2,4, Taylor P. Light3, Sergey A. Goncharuk2,
Andrea Soler-Lopez1, Fozia Ahmed3, Konstantin S. Mineev2, Kalina Hristova3,
Alexander S. Arseniev2* and Marçal Vilar1*
1Unit of Molecular Basis of Neurodegeneration, Institute of Biomedicine CSIC.
46010 València, SPAIN
2Shemyakin-Ovchinnikov Institute of Bioorganic Chemistry of the Russian Academy
of Sciences, Moscow 117997, Russian Federation.
3 Department of Materials Science and Engineering, Johns Hopkins University,
Baltimore, MD, USA
4 Contributed equally
*Corresponding authors; mvilar@ibv.csic.es and aars@nmr.ru
Running Title: TrkA and p75 complex formation
Abstract
The neurotrophin receptors p75 and TrkA play an important role in the development
and survival of the nervous system. Biochemical data suggest that p75 and TrkA
regulate the activities of each other. For instance, p75 is able to regulate the response
of TrkA to lower concentrations of NGF and TrkA promotes p75 shedding by α-
secretases in a ligand-dependent manner. The current model is that p75 and TrkA are
regulated by means of a physical direct interaction, however the nature of such
interaction has been elusive so far. Here using NMR in micelles, multiscale molecular
dynamics (MD), FRET and functional studies we identified and characterized the
direct interaction between TrkA and p75 through the transmembrane domains
(TMDs). MD of p75-TMD mutants suggests that although the interaction between
TrkA and p75 TMDs is maintained, a specific protein interface is required to facilitate
TrkA active homodimerization in the presence of NGF. The same mutations in the
TMD protein interface of p75 reduced the activation of TrkA by NGF and cell
differentiation. In summary we provide a structural model of the p75/TrkA receptor
complex stabilized by transmembrane domain interactions.
2
Introduction
Nerve growth factor (NGF) is a member of the mammalian neurotrophin (NT) protein
family, which also includes BDNF, NT3, and NT4/5 (1). NTs are implicated in the
maintenance and survival of the peripheral and central nervous systems and mediate
several forms of synaptic plasticity (2–5). NTs interact with two distinct receptors, a
cognate member of the Trk receptor tyrosine kinase family and the common p75
neurotrophin receptor, which belongs to the tumor necrosis factor receptor (TNFR)
superfamily of death receptors (6, 7). While Trk receptor signaling is involved in
survival and differentiation (8, 9), p75 participates in several signaling pathways
(reviewed in (10)). p75-mediated signaling is governed by the cell context and the
formation of complexes with different co-receptors and ligands, such as sortilin/pro-
NGF in cell death (11), Nogo/Lingo-1/NgR in axonal growth (12, 13), and TrkA/NGF
in survival and differentiation (14). p75 also undergoes shedding and receptor
intramembrane proteolysis (RIP), resulting in the release of its intracellular domain
(ICD), which itself possesses signaling capabilities (15–17).
Several lines of evidence implicate functional interactions between TrkA and
p75NTR in NGF-triggered signal transduction (3, 18–20). TrkA and p75 receptors
have nanomolar affinities for NGF and cooperate in transducing NGF signals (7, 21).
The expression patterns of these two receptors overlap extensively (22) and in some
instances, such as in the neurons of the dorsal root ganglion (DRG), TrkA is
exclusively expressed in conjunction with p75 (23).
p75 has been experimentally demonstrated to enhance the response of TrkA to NGF
(14, 24–26). In sympathetic neurons and oligodendrocytes, TrkA signaling inhibits
the pro-apoptotic signaling of p75 (27–29). Primary DRG and sympathetic neurons
derived from p75-null animals show attenuated survival responses to NGF (25, 26,
30), confirming the physiological role of p75/TrkA interactions. As the interaction
between the two receptors seems to not engage the ligand binding domains of the
extracellular region (31), the structural basis of such direct interaction is still
unknown.
Here we demonstrate that the interaction between TrkA and p75 is mediated, at least
in part, by the transmembrane domains. We validate these findings using functional
studies in cells expressing the full-length receptors.
3
RESULTS
p75 and TrkA form a constitutive complex at the plasma membrane
We performed Förster Resonance Energy Transfer (FRET) experiments to determine
if TrkA and p75 interact directly at the plasma membrane of live cells. HEK 293T
cells transiently co-transfected with full-length TrkA tagged with mTurquoise (the
donor fluorescent protein) and full-length p75 tagged with eYFP (the acceptor
fluorescent protein) were imaged, and small regions of the plasma membrane were
selected and analyzed. Illustrations of the TrkA-mTurq and p75-eYFP constructs used
in FRET experiments are shown in Figure 1A. In each region of the cell membrane,
we determined the FRET efficiency, the concentration of TrkA-mTurq and the
concentration of p75-eYFP using the FSI-FRET software (Figure 1B) (32). These
experiments were designed such that FRET can only occur between TrkA and p75,
not between TrkA-TrkA or p75-p75. We also performed control FSI-FRET
experiments using two unrelated proteins, LAT (Linker for the Activation of T-cells)
and FGFR3 (Fibroblast Growth Factor Receptor 3), which are not expected to interact
specifically and thus should give zero hetero-FRET. In addition, the proteins were
designed such that the fluorescent tags are positioned differently with respect to the
plasma membrane—the mTurq fluorophore is attached to the C-terminus of full-
length LAT while the eYFP fluorophore is attached to the C-terminus of an FGFR3
construct lacking the intracellular region, “ECTM” (Figure 1A). Therefore, these two
proteins will also not give rise to a non-specific FRET signal (random or “proximity”
FRET) (33). As expected, due to the large separation between the fluorescent tags, the
FRET efficiencies measured between these two control proteins are localized around
zero at all concentrations measured (Figure 1B). Therefore, this control dataset
demonstrates the scenario where there is no FRET between the proteins.
In the absence of ligand, full-length TrkA and p75 exhibit positive (greater than zero)
FRET efficiency values over all TrkA and p75 concentrations measured (Figure 1B).
Therefore, this data suggests that TrkA and p75 interact directly at the plasma
membrane. With this data alone, we cannot determine an accurate stoichiometry of
the TrkA-p75 heterocomplex. Given that TrkA and p75 exist in monomer-dimer
equilibrium in the absence of ligand, it is possible that TrkA and p75 associate as
heterodimers, or oligomers of higher order (Figure 1C). Next, we sought to determine
4
if NGF ligand binding influences the TrkA-p75 heterocomplex, and we performed
similar FSI-FRET experiments for TrkA-p75 in the presence of 100 ng/µL NGF
(Figure 1D). The FRET efficiencies measured for TrkA-p75 in the presence of NGF
are noticeably lower compared to the data in the absence of ligand. Furthermore,
comparison of the liganded TrkA-p75 FRET data to the LAT-FGFR3 control
experiment data revealed no significant differences (Figure 1E), which suggests that
the fluorophores attached to the C-termini of TrkA and p75 are too far away from one
another to observe a FRET signal in the ligand-bound state. The expression levels of
the TrkA and p75 at the cell surface are similar in both sets of experiments (+/- NGF)
so these differences are not a reflection of altered gene expression (Figure 1F). The
decrease in FRET may mean that the heterointeractions are abolished, for instance,
due to ligand-induced homodimer stabilization, or it may be due to conformational
changes in the heterocomplex which leads to decreased FRET.
The FRET data for TrkA-p75 in the absence and presence of NGF and the control
dataset were binned and compared in order to visualize the average FRET efficiency
as a function of receptor concentration (Figure 1G). For the control dataset and the
TrkA-p75 data in the presence of NGF, the average FRET efficiencies remain around
zero as expected from the raw data (Figure 1G). For the TrkA-p75 data in the absence
of ligand, we observe average FRET efficiencies greater than zero over all
concentrations (Figure 1G). Furthermore, at the low receptor concentration regime,
the average FRET efficiencies increase as a function of receptor concentration,
suggesting increasing TrkA-p75 interactions (Figure 1G).
Direct interaction between p75 and TrkA transmembrane domains
Previous findings have suggested that TrkA can form a complex with p75-CTF (a
membrane-anchored C-terminal fragment) by means of transmembrane domain
interaction (17). In addition, the TM domain of p75 is involved in the formation of the
high-affinity NGF binding sites (34), suggesting that the TM domain may mediate the
direct interaction between p75 and TrkA. Therefore, we were interested in
investigating the interaction between the p75 and TrkA, taking into account the
recently reported NMR structures of p75 and TrkA TM domains (35, 36). We
examined the interaction of p75-TM-wt with the TrkA-TM domain in lipid micelles
using NMR spectroscopy. Increasing amounts of TrkA-TM were added to the 15N-
labeled p75-TM in DPC micelles and the chemical shifts were monitored in a 1H-15N
5
HSQC spectrum (Figure 2A). Chemical shifts are very sensitive to the electronic
environment of a nucleus, and serve as an ideal instrument to probe the protein-
protein interaction. Previous work in our laboratory found that p75-TM-wt forms
spontaneous disulfide dimers (35). We titrated the 15N labeled p75-TM-wt disulfide
dimer with increasing concentrations of TrkA-TM-wt solubilized in DPC micelles,
retaining the constant lipid-to-protein ratio (LPR). The titration revealed no chemical
shift changes. We used several LPRs and at least two independent preparations of
p75-TM-wt. As the dimerization of p75-TMD-wt is irreversible (35) we performed
the experiments with the mutant p75-C257A, which forms non-covalent homodimers
(35) and allows the possibility to obtain the monomeric p75-TMD. According to the
previous work (35), the C257A mutation does not induce any substantial changes to
the structure of p75 TMD. Several chemical shift changes were observed in the
HSQC-NMR spectrum of p75-TM-C257A upon titration with TrkA-TM-wt,
suggesting the formation of specific p75/TrkA heterocomplexes (Figure 2B).
To identify the oligomer size of the complex, we measured the cross-correlated
relaxation rates of p75-TM-C257A signals (Figure S1). According to the recent work,
the NMR-derived hydrodynamic radii of TM domains in DPC micelles can be used to
distinguish the various oligomeric forms of the proteins (37). Here we observed the
rotational correlation time (and hydrodynamic radius) of a p75-TM-C257A monomer
at 45 oC to be 10.2±0.4 ns (2.61 nm), a p75-TM-C257A homodimer to be 13.1±0.6 ns
(2.85 nm) and the heterocomplex to be 12.7±0.8 ns (2.82 nm). In other words, the
observed new complex formed by TrkA-TM and p75-TM-C257A is a heterodimer as
the rotational correlation time of the heterocomplex was similar to that of the
homodimer.
With the increase of TrkA concentration, the percentage of p75 homodimer decreased
while that of p75/TrkA heterodimer increased (Figure 2C). This implies that homo-
and heterodimerization of p75-TM are the competing processes. The titration curve
revealed homodimerization and heterodimerization constants of comparable
magnitudes. Similar effects were observed when 15N-labeled TrkA-TMD sample was
titrated with the unlabeled p75-TM-C257A (Figure 2E, F). Addition of p75-TM-
C257A decreased the concentration of TrkA-TM homodimer, while the novel
heterodimeric state had emerged, which is indicative of the competition. Thus, we can
state that TrkA interacts with the monomeric form of p75 TM domain but does not
bind the disulfide-crosslinked dimer of the protein. Most likely, the covalent
6
dimerization shields some of the p75 residues necessary to interact with the TrkA TM
domain, or the interaction requires a rearrangement of the dimer that cannot be
achieved due to the restraints imposed by the disulfide bonds.
Chemical shift (CS) changes were detected along the p75 TMD sequence (Figure 2D),
which is expected as the TrkA interaction breaks the p75-TM-C257A dimerization.
The residues with the highest chemical shift changes are shown in the Figure 2D. To
find the residues undergoing chemical shifts changes in the TrkA-TMD, we
performed the titration on labeled 15N-TrkA-TMD homodimer with unlabeled p75-
TM-C257A (Figures 2E). With increasing p75-TM-C257A concentration, the
percentage of TrkA homodimer decreased while the heterodimer increased (Figure
2F). The NMR chemical shifts indicated that the region of higher CS changes (Figure
2G, Δδ>0.1) upon interaction of p75-TMD is located mainly at the N-terminus of
TrkA TMD.
These results support a direct interaction between p75-TMD and TrkA-TMD and
suggest that the formation of a heterodimer outcompetes the homodimerization of
each TM domain. Although the NMR shows that the interaction is direct, we cannot
use the CS changes to identify the protein-protein interface between the TM domains
in a membrane. Recently it has been shown that, by contrast to soluble proteins, CS
changes have almost zero predictive power to map protein interfaces in
transmembrane regions (38). CS changes primarily report hydrogen bonding and are
insensitive to van-der-Waals contacts between the protein side chains, which are the
main driving force for dimerization of membrane proteins (38).
Multiscale Molecular Dynamics
The crowding of the NMR spectra with several TrkA and p75 species (monomer,
homo- and heterodimers) precludes the complete CS assignment and the structure
calculation of the heterocomplex. To explore further the interaction between TrkA
and p75 TMDs we used molecular dynamics (MD) (Table 1 and Figure 3). MD
simulations provide a useful approach for modeling the transmembrane domain
interactions (39). Both full-atom (FA) and coarse-grained (CG) modeling has been
previously used to optimize the dynamics and interactions between different
transmembrane domains (39). To model the heterodimerization of p75-TMD and
TrkA-TMD, two CG helices were inserted in a parallel orientation relative to one
7
another separated 6 nm in a preformed POPC bilayer and 24 simulations of 5 µs were
run (total time 120 µs) (Table S1 and Figure 3A). In all but one of the 12 simulations,
the TrkA/p75 heterodimer was formed within the first 2 µs (except for the simulation
number #5 that formed the heterodimer at 5 µs) and did not dissociate during the
remainder of the simulation (Figure 3B). The POPC model membrane was well
equilibrated with average values for the area per lipid and hydrophobic thickness
(between glycerol groups) of 63.2 Å2 and 34.8 Å respectively that are in good
agreement with the experimental values (40) (Table S2). From each of the
heterocomplexes (Figure 3C), we compute the root mean squared deviation (RMSD)
between them and found a cluster of 7 models with an average RMSD of 2.12 Å
(Figure 3D). The final model was converted from coarse-grained to full-atom to
further study the packing of the interaction in a POPC lipid bilayer during 100 ns of
FA-MD, done in triplicate. The final POPC model membrane was well equilibrated
with average values for the area per lipid and hydrophobic thickness (between
phosphate groups) of 63.3 Å2 and 38.4 Å respectively that are in good agreement with
the experimental values (40). The membrane electron density was calculated and
shown in the Figure S2. The interhelix distance between residues at the C-terminus of
the helix (p75-W276 and TrkA-K441) was calculated along the total simulation time
(Figure 3E), indicating the equilibration of a stable complex.
The protein interface of p75-TMD participating in the interaction with TrkA-TMD is
C257S258xxA261A262xxV265G266xxA269xx (Figures 4A and 4B). This interface contains
the motif A262xxxG266xxA269 that was previously identified in the homodimerization
of p75-TMD-C257A (35) and is supported by the NMR experiments shown above,
indicating that heterodimerization with TrkA-TMD competes with the p75-TMD non-
covalent homodimerization. In addition, the residue C257 forms a part of the
heterodimer interface supporting our observations that disulfide dimers do not
significantly bind to the TrkA-TMD. The TrkA-TMD heterodimer interface is
formed by the motif V418xxxV422xxxV426F427xxL430 (Figure 4B) where the central
valine residues make the closest contact with the p75-TMD. Interestingly, several of
these residues are conserved in TrkB and TrkC (Figure 4C) suggesting that these
receptors interact with p75 in a similar manner as TrkA.
Altogether, the NMR and FRET data support the direct interaction between TrkA and
p75 and the MD provides insight into a possible heterodimer model.
8
The transmembrane heterodimer interface modulates TrkA activation and
sensitization to lower concentrations of NGF.
In vivo data suggest that in sensory neurons p75 helps TrkA to respond to the lower
concentrations of NGF (26) and enhances the response of TrkA to NGF (14, 24). One
current hypothesis is that the binding of p75 to TrkA induces a conformational change
in TrkA that facilitates both the binding of NGF to TrkA (24) and the activation of
TrkA (26). To test if the protein interface found above has any physiological role we
sought to determine if mutations on the p75 transmembrane protein interface
influences TrkA activation to lower concentrations of NGF (Figure 5A). We co-
expressed p75 with TrkA full-length receptors in Hela cells and stimulated with
increasing concentrations of NGF (0, 0.1, 1, 10 and 100 ng/mL). Western blot of cell
lysates were probed with specific antibodies against the activation loop of the TrkA
kinase domain (Tyr675 and Tyr676) (Figure 5B). Quantification of the protein bands
corresponding to the phosphorylation of TrkA was plotted against NGF concentration.
Fitting the data to a dose (NGF)–response (phosphorylation) curve allows an
estimation of the EC50 of NGF, the concentration of NGF that provokes a response
half way between the basal response and the maximal response (Figure 5D). Hela
cells transfected with TrkA present a LogEC50 of -9.219 ±�0.087 (an EC50 = 6.03 x10-
10 M). In cells co-expressing TrkA and p75 an LogEC50 of -9.524 ± 0.176 (an EC50 =
2.99x10-10 M) was found, showing a small, but significant effect of p75 on the
activation of TrkA by NGF. The parallel curve suggested an agonist effect of p75 and
NGF on the activation of TrkA. To analyze the effect of p75-TMD we used a
construct of p75 with its transmembrane domain swapped with the one from the
tumor necrosis factor receptor (TNFR), mutant p75-TNFR. A decrease in the NGF
sensitivity was observed in comparison to p75-wt (LogEC50 -8.56 ±�0.54, EC50=
2.7x10-9 M), indicating that the effect of p75-wt is lost in the p75-TNFR construct. As
the protein heterodimer interface contains the motif A261A262xxxG266xxA269 we made
a construct with a triple mutation A262,G266,A269 to Ile (p75-AGA mutation). The
rationale behind this is that the introduction of a hydrophobic bulky residue, Ile,
would impair the proper interaction with the TrkA-TMD. Fitting of the values
obtained from the lysates transfected with TrkA and p75-AGA showed a LogEC50 of
-8.776 ± 0.037, that corresponds to an EC50 = 1.7x10-9 M (Figure 5D), that accounts
for more than one order of magnitude higher than in the presence of p75-wt
supporting that this interface plays a key role in TrkA activity modulation by p75.
9
p75 needs a specific interface in the transmembrane domain to interact to TrkA
The finding that the activation of TrkA in the presence of the p75-AGA mutant is
lower than in the absence of p75 suggested an antagonist or inhibitor behavior for this
mutant. To further study the effect of this mutation on the heterodimer complex, we
introduced the triple mutation AGA/III into the p75-TMD and performed a CG-MD
followed by FA-MD simulation similar to the p75-TMD-wt constructs shown above
(Figure S3). MD analysis showed that although p75-TMD-AGA mutant still interacts
and binds to the TrkA-TMD with similar kinetics as the p75-TMD-wt, the
heterodimer arrangement is changed significantly. It has been previously shown that
TrkA-TMD contains two homodimer interfaces; an active dimer formed upon NGF
binding and an inactive dimer formed in the absence of NGF. The 12 independent
simulations of p75-TMD-wt showed a restricted binding interface localized close to
the inactive homodimer interface, leaving the active homodimer interface of TrkA
free and accessible (Figure 6A). However, after 12 independent simulations the end-
point of p75-TMD-AGA is almost equally distributed in all the possible TrkA-TMD
interfaces (Figure 6B), where the active homodimer interface is hidden by p75-TMD.
This result indicates that p75-TMD-AGA could impair TrkA active homodimerization
and may explain the weaker activation of TrkA in the presence of p75-TMD-AGA.
p75-AGA/III reduces NGF-induced differentiation of PC12 cells
To further support our finding that p75 needs a specific heterodimer interface to fully
activate TrkA, we overexpressed p75-wt and p75-AGA/III in PC12 cells that
endogenously express TrkA and quantified the neurite length upon stimulation with
NGF. As shown in Figure 7A, the PC12 cells transfected with p75-AGA/III had
shorter neurite lengths at 24h than cells transfected with p75-wt (16.02 µm ±�0.98,
n=181 vs 22.63 µm ±�1.69, n=89) and similar length as PC12 cells transfected with
the empty vector (14.09 µm�±�1.25, n=92) (control in Figure 7A). These experiments
suggest a reduction in the activation of TrkA by NGF of p75-AGA/III in comparison
to p75-wt.
Discussion
The present study provides, to the best of our knowledge, the first structural evidence
of a direct interaction between p75 and TrkA. While data from in vitro and in vivo
10
experiments has suggested the existence of a complex formed by p75 and TrkA (41–
43), repeated attempts to observe the direct interaction between both receptors using
different biochemical and structural approaches have been unsuccessful. Experimental
evidence of the existence of a TrkA/p75 complex were based on co-
immunoprecipitation studies (17, 21, 44) and by biophysical methods such as co-
patching (45) and fluorescence recovery after photobleaching (46). In addition, a
handful of studies have suggested that the transmembrane and intracellular domains
of p75 could be responsible for its interaction with TrkA (21, 34, 47, 48).
Here we demonstrated that the complex formed by p75 and TrkA is mediated by the
TM domains, supporting the findings by previous reports (21, 34). The results of our
NMR titration experiments point to a relatively weak affinity constant, similar to that
calculated for p75 non-covalent dimerization. This is around 10 times weaker than the
affinity constant calculated for glychopohrin-A homodimerization, and explains why
these complexes have been difficult to detect by co-immunoprecipitation in the
presence of detergents (i.e, glycophorin A TM domain dimers are resistant to SDS-
PAGE). Hetero-crosslinking experiments similarly failed to detect p75/TrkA
complexes, although probably for different reasons, as crosslinking requires the
specific residues (i.e Lys) to be close to each other and oriented in a specific manner,
not always possible even in a heterocomplex. Our results are in agreement with those
of fluorescence recovery after photobleaching (FRAP) experiments, which show that
p75 is fully mobile at the cell membrane but becomes restricted in mobility upon
TrkA co-expression (46), and with biochemical evidence suggesting that the TM
domain of p75 is necessary for the formation of high-affinity NGF binding sites (34).
Although the TMD interaction is weak, in vivo the levels of p75 and Trk normally
exist at a ratio of approximately 10:l (49, 50) favoring their heterointeractions
interaction over TrkA homointeractions.
Recently it has been shown that TrkA has two homodimer interfaces in the TMD; one
active and one inactive (36). The active interface corresponds to the TrkA bound to its
ligand NGF. And the inactive dimer interface corresponds to the pre-formed dimer of
TrkA in the absence of NGF. The observed binding of p75-TMD to TrkA takes place
mainly through an interface that is opposite to the active interface and partially
covering part of the inactive dimer interface, suggesting that binding of p75 to TrkA
may favor the formation or stabilization of TrkA active homodimers. In addition,
stabilization of a pre-formed dimer would be compatible to an increase in the affinity
11
of TrkA for NGF in the presence of p75 (18), suggesting that the heterodimer
p75/TrkA described here forms the basic unit of the NGF high-affinity sites. The
finding that mutations in the p75 protein interface, as shown here with the p75-AGA
mutant, impact the TrkA activation and supports the requirement of specific TMD
interactions in the neurotrophin receptors. As it has been shown recently, NGF
binding can induce the rotation of the TrkA TM dimer form the inactive to the active
interface (36, 51). This conformational change is supported by our FRET analysis,
which reveals that NGF binding alters the TrkA-p75 heterocomplex that we observed
in the absence of ligand. There are some possible explanations for this result, which
are both illustrated in Figure 1H. The first option is that NGF binding could cause the
dissociation of the TrkA-p75 heterocomplex, stabilizing the respective homodimers
instead. Another explanation is that NGF binding induces a conformational change of
the TrkA-p75 heterocomplex that alters the positioning of the fluorescent proteins,
increasing their separation and thus decreasing the FRET signal. While this data
cannot distinguish between these two possible effects, the FRET data clearly
demonstrate that TrkA and p75 interact directly in the absence of ligand and that NGF
binding alters the heterocomplex.
Our MD analysis of p75-AGA/TrkA interactions showed that the inactive dimer
interface is accessible suggesting that p75-AGA interaction may displace the
equilibrium towards the inactive homodimer of TrkA in the absence of NGF. This
would affect the activation of TrkA and lead to lower cell differentiation capabilities
of PC12 cells overexpressing the p75-AGA mutant. Alternatively, the binding of the
p75-AGA mutant may affect the conformational change induced by NGF binding
resulting in a less activation of TrkA.
Altogether, we show that a specific transmembrane interaction is required for the
positive role of p75 in TrkA activation by NGF. In conclusion, we provide a new
structural insight on the highly dynamic p75/TrkA heterocomplex, paving the way to
new investigations about the biological relevance of such interactions.
EXPERIMENTAL PROCEDURES
p75-TM and TrkA-TM constructs for cell-free expression
12
The gene encoding transmembrane and juxtamembrane residues 245-284
(MT245RGTTDNLIPVYCSILAAVVVGLVAYIAFKRWNSSKQNKQ284)
of
human p75 receptor (p75-TM-wt) was amplified by PCR from six chemically
synthesized oligonucleotides (Evrogen, Russia) partially overlapped along its
sequence. The C257A point mutant form of p75TM (p75-TM-C257A) was obtained
by site-directed mutagenesis by PCR. The PCR products were cloned into a
pGEMEX-1 vector by three-component ligation using the NdeI, AatII and BamHI
restriction
sites.
Expression
constructs
for
human
TrkA-TM
(MK410KDETPFGVSVAVGLAVFACLFLSTLLLVLNKAGRRNK447)
were
similarly prepared by PCR.
Fully Quantified Spectral Imaging (FSI)-FRET experiments
Human embryonic kidney (HEK) 293T cells used in the FRET experiments were
purchased from American Type Culture Collection (Manassas, VA; CRL-3216). The
cells were cultured at 37 °C and 5% CO2 in Dulbecco’s Modified Eagle Medium
(DMEM; Thermo Scientific; 31600-034) containing 3.5 g/L D-glucose, 1.5 g/L
sodium bicarbonate, and 10% fetal bovine serum (FBS; Sigma-Aldrich; F4135).
HEK293T cells were seeded in 35 mm glass bottom collagen-coated petri dishes
(MatTek Corporation, MA) at a density of 2 x 105 cells/dish and cultured for 24
hours. The cells were co-transfected with pcDNA constructs encoding for TrkA
tagged with mTurquoise (mTurq, the donor) and p75 tagged with enhanced yellow
fluorescent protein (eYFP, the acceptor). The TrkA-mTurq plasmid was generated as
described (32, 52). The p75-eYFP construct was cloned by overlapping PCR into the
same pcDNA vector. The LAT and ECTM FGFR3 plasmids used for control
experiments were generated as described previously (53, 54). Transfection was
performed with Lipofectamine 3000 (Invitrogen, CA; L3000008) using 1-4 µg of total
DNA at a TrkA:p75 ratio of 2:1 or 1:1. In addition, cells singly transfected with either
TrkA-mTurq or p75-eYFP were used for calibration as described (32). After twelve
hours following transfection, the cells were washed twice with starvation media
(serum-free, phenol red-free media) and serum-starved in starvation media for 12
hours overnight. Prior to imaging, the starvation media was replaced with hypo-
osmotic media (10% starvation media, 90% diH2O, 25 mM HEPES) to ‘unwrinkle’
the highly ruffled cell membrane under reversible conditions as described (55). Cells
were incubated for 10 minutes and then imaged under these conditions for
approximately 1 hour. In some experiments, soluble human beta nerve growth factor
13
(hβ-NGF; Cell Signaling Technology; 5221SC) was diluted to a final concentration of
100 ng/µl with the hypo-osmotic media before adding to the cells.
Cell images were obtained following published protocols (32) with a spectrally
resolved two-photon microscope set up using a Zeiss Inverted Axio Observer and the
OptiMis True Line Spectral Imaging system (Aurora Spectral Technologies, WI) with
line-scanning capabilities (56, 57). Fluorophores were excited with a mode-locked
laser (MaiTai™, Spectra-Physics, Santa Clara, CA) that generates femtosecond pulses
between wavelengths 690 nm to 1040 nm. For each cell, two images were collected:
the first at 840 nm to excite the donor and the second at 960 nm to primarily excite
the acceptor. Solutions of purified soluble fluorescent proteins (mTurq and eYFP)
were produced at known concentrations following a published protocol (58) and
imaged at each of these excitation wavelengths. A linear fit generated from the pixel-
level intensities of the solution standards was used to calibrate the effective three-
dimensional protein concentration which can be converted into two-dimensional
membrane protein concentrations in the cell membrane as described (32). Small
micron sized regions of the cell membrane were selected and the FRET efficiency, the
concentration of TrkA-mTurq, and the concentration of p75-eYFP present in the cell
membrane were quantified using the FSI-FRET software (32).
Cell-free gene expression
Bacterial S30 cell-free extract was prepared from 10 L of cell culture of the E. coli
Rosetta(DE3)pLysS strain, using a previously described protocol (Aoki et al., 2009;
Kai et al., 2012; Schwarz et al., 2007). Preparative-scale reactions (2-3 mL of reaction
mixture) were carried out in 50-mL tubes.
Titration of TrkA and p75 transmembrane domains by NMR
All TrkA/p75 titration 15N-TROSY experiments were carried out at LPR 80, pH 5.9,
temperature 318K with 20 mM NaPi buffer. Two independent sets of experiments
were conducted: (1) unlabeled p75-TM-C257A was incrementally added to 0.5 mM
sample of 15N-labeled TrkA-TM, and (2) unlabeled TrkA-TM was incrementally
added to the 0.4 mM sample of 15N-labeled p75-C257A-TM sample to observe p75-
TrkA interactions. Intensities of corresponding peaks were measured at each point,
population of the p75-p75 dimer, TrkA-TrkA dimer and TrkA/p75 complex were
calculated and plotted against TrkA/p75 molar ratio.
14
Modulation of TrkA activity by p75
Hela cells were transfected with 1µg of TrkA and 1µg of p75 or p75-TNFR using PEI
(ratio 10:1). 24 hours after transfections cells were lifted and split in identical
numbers to a 6 well plate. 48 hours after transfection were starved for 2 hours with
DMEM without serum and stimulated with different concentrations of NGF (from 0
to 100 ng/mL) for 15 minutes. Cells were washed with PBS and lysed with TNE
buffer on ice for 15 minutes. Lysates were clarified by centrifugation and the cell
supernatants quantified and analyzed by SDS-PAGE western immunoblots. Phospho-
Tyrosine specific antibodies (anti P-Tyr674/675 from Cell Signalling 1:3000) and
anti-p75 intracellular antibody (Promega) were used. To quantify the effect of p75 on
TrkA we consider an allosteric interaction between p75 and TrkA and fit to a
dose/response curve. The protein band corresponding to the phospho-Tyr signal was
quantified and the ratio to the total TrkA was calculated. This is the response in the
Figure 4. We plot the log of the concentration of NGF versus the response and the
curve was fit to a log(agonist) vs response (three parameters) equation using the
GraphPad software. The equation is Y=Bottom + (Top-Bottom)/(1+10^((LogEC50-
X))), and the EC50 is the concentration of agonist, in this case NGF, that gives a
response half way between Bottom and Top. At least three independent experiments
were quantified.
Coarse Grained Molecular Simulation Methods
One monomer from the TrkA-TMD dimer structure (PDB:2n90) and one monomer
from the p75-TMD dimer structure (PDB:2mic) were converted to a coarse grained
CG model using the script martinize.py from the martini web page
(www.cgmartini.nl) and the tools from Gromacs 5.0.5. In CG models 4 heavy atoms
are grouped together in one coarse-grain bead. Each residue has one backbone bead
and zero to four side-chain beads depending on the residue type (Monticelli,
Kandasamy et al., 2008). For all helix dimerization simulations, two α-helices were
inserted into a preformed 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC)
bilayer (containing 260 lipids) such that they were separated by an interhelix distance
dHH ≈ 55 Å (Figure 6). Each system was solvated with 2975 CG water particles and
0.15 M NaCl counter ions. The energy of the system was minimized and followed by
12 MD simulations of 5 µs each simulation, total time 60 µs. CG simulations were
15
performed using GROMACS v 5.0.5 (www.gromacs.org) (Van Der Spoel, Lindahl et
al., 2005). All simulations were performed at constant temperature, pressure, and
number of particles. The temperatures of the protein, lipid, and solvent were each
coupled separately using the Berendsen algorithm at 305 K, with 774 a coupling
constant τT = 1 ps. The system pressure was semiisotropically using the Parrinello-
Rahman algorithm at 1 bar with a coupling constant τP = 12 ps and a compressibility
of 3 Å~ 10−4 bar−1. The time step for integration was 20 fs. Coordinates were saved
for subsequent analysis every 200 ps.
Atomic Molecular Dynamics
GROMACS 5.0.5 was also used for all full atom MD simulations. CG models were
converted to FA using the CHARMM-GUI portal (www.charmm-gui.org). FA was
calculated using the CHARMM36m force field. Long-range electrostatics was
calculated using the particle mesh Ewald method with a real-space cutoff of 10 Å. For
the van der Waals interactions, a cutoff of 10 Å was used. The simulations were
performed at a temperature of 303.15 K using a Nose-Hoover thermostat with τT = 1
ps. A constant pressure of 1 bar was maintained with a Parrinello-Rahman algorithm
with an semiisotropic coupling constant τP = 5.0 ps and compressibility = 4.5 Å~
10−5 bar−1. The integration time step was 2 fs. The LINCS method was used to
constrain bond lengths. Coordinates were saved every 5 ps for analysis. Analysis of
all simulations was performed using the GROMACS suite of programs. VMD
(Humphrey, Dalke et al.,1996) and Chimera UCSF (Pettersen, Goddard et al., 2004)
were used for visualization and graphics. Membrane equilibration was assed
measuring the area per lipid and the membrane thickness using the APLVoro
application (59). The electron density profiles were calculated using the gmx
density tool in Gromacs. A representation of the electron density of the POC model
membrane with TrkA and p75 TMDs is shown in the Figure S2.
Cell culture and transfection
Hela cells, which do not endogenously express neither TrkA nor p75, were cultured in
DMEM medium (Fisher) supplemented with 10% FBS (Fisher) at 37 °C in a
humidified atmosphere with 5% CO2. PC12 and PC12nnr5 cells were cultured in
DMEM with 10% FBS and 5% horse serum. Transfection in Hela cells was
performed using polyethyleminime (Sigma) at 1-2µg/µl. We found that by using
polyethylenimine (PEI) as the transfection reagent in Hela cells the transfection is
16
suboptimal (10-15% of cells transfected) that allow having a small amount of TrkA
expressed in the cells and with. As a comparison using the same PEI/DNA ratio in
Hek293 cells TrkA is expressed in higher amounts and ligand-independent activation
is seen at this quantities of TrkA DNA. 500-1000 ng of DNA per plate was used in
TrkA activation experiments. 24h after transfection cells were lifted and re-plated in
12-well plates with 100,000 cells per well. Using this procedure the percentage of
transfection is identical in all the wells. 48h after transfection the cells were starved
with serum free medium for 2h and stimulated with NGF (Alomone) at the indicated
concentrations and time intervals. Cells were lysed with TNE buffer (Tris-HCl pH
7.5, 150 mM NaCl, 1mM EDTA) supplemented with 1% triton X-100 (Sigma),
protease inhibitors (Roche), 1 mM PMSF (Sigma), 123 mM sodium orthovanadate
(Sigma), and 1 mM sodium fluoride 545 (Sigma). The lysates containing p75 were
supplemented with iodoacetamide (Sigma) to avoid post-lysate dimer disulfide
formation. Lysates were kept on ice for 10 minutes and centrifuged at 13,000 rpm for
15 minutes on a tabletop centrifuge. The lysates were quantified using a Bradford kit
(Pierce) and analyzed by SDS-PAGE or used in immunoprecipitation.
Western blot analysis
Cells were washed in PBS and lysed in cold lysis buffer (50 mM Tris-HCL [pH7.5],
150 mM NaCl, 1 mM EDTA, 0.1% SDS, 0.1% Triton X-100, 1 mM PMSF, 10 mM
NaF, 1 mM Na2VO3, 10 mM iodoacetamide and protease inhibitor cocktail) at 4ºC.
Cellular debris was removed by centrifugation at 13,000 g for 15 minutes and protein
quantification was performed by Bradford assay. Proteins were resolved in reducing
and non-reducing SDS-PAGE gels and membranes were incubated overnight at 4ºC
with the following antibodies: rabbit polyclonal anti-human p75 intracellular domain
(1:1000, Promega); mouse monoclonal anti-HA (1:2000, SIGMA); rabbit polyclonal
MBP-probe (1:1000, Santa Cruz); rabbit anti-phosphoTyr674/5 (1:1000, Cell
Signaling); rabbit anti-TrkA (1:1000, Millipore). Following incubation with the
appropriate secondary antibody, membranes were imaged using enhanced
chemiluminescence and autoradiography.
Electroporation of PC12 and differentiation experiments
The electroporation of the different plasmids was carried out with the Multiporator®
(Eppendorf). PC12 cells were gwon with DMEM supplemented with 10% FBS and
17
5% Horse Serum and antibiotics (gentamycin and penicillin). For elecrtoporation cells
were grown to 70-80% confluence on a 10 cm plate and washed with PBS. They were
then raised with 3 ml of DMEM medium and centrifuged for 2.5 minutes at 500 rpm.
The pellet obtained was resuspended in 3 ml of the hypoosmolar electroporation
buffer (KCl 25mM, KH2PO4 0.3 mM, K2HPO4 0.85 mH, pH 7.2) and a viable
counting with trypna blue was carried out. 1 x 105 cells, and a concentration of 5 µg /
ml of the plasmid of interest (control, wt or mutant) and a concentration of 5 µg / ml
of the plasmid with GFP (Green Fluorescent Protein) were transferred to an
electroporation cuvette (2 mm wide and 400 µl in volume (Eppendorf)). After
optimizing the transfection parameters, it was determined that the best results were
obtained with a pulse of 100 µs at 200V, therefore the electroporation was carried out
under these conditions. Finally, the cells were seeded on a 6-well plate with 2 ml of
DMEM medium supplemented with 5% horse serum (Gibco). At 24 hours after
transfection, the cells were treated with NGF (50 ng/mL) in order to induce the
differentitaion of neurites as a function of the plasmid. The length of each neurite was
quantified from fluorescence microscopy images uisng the ImageJ software. Three
independent electroporation experiments were analyzed and at least 100 neurites per
each condition was quantified.
DATA AVAILABILITY
All the data are contained within the manuscript. Chemical shifts from TrkA-TMD
and p75-TMD are deposited in the Biological Magnetic Resonance Data Bank BMRB
with accession number 25872 for TrkA-TMD and 19673 for p75-TMD.
SUPPORTING INFORMATION
This article contains supporting information.
ACKNOWLEDGMENTS.
We thank Dr. M.D. Paul for the acquisition of the LAT-FGFR3 control FRET data
set.
FUNDING INFORMATION
This study was supported by the Spanish Ministry of Economy and Competitiveness
(MINECO; project BFU2013-42746-P and SAF2017-84096-R), by the Generalitat
18
Valenciana Prometeo Grant 2018/055 to MV), and by NIH GM068619 (to KH).
NMR studies of TRKA-TM and p75-TM were supported by the Russian Science
Foundation (grant No# 19-74-30014 to A.S.A).
CONFLICT OF INTEREST
The authors declare that they have no conflicts of interest with the contents of this
article.
REFERENCES
1.
Bothwell, M. (2014) NGF, BDNF, NT3, and NT4. Handb Exp Pharmacol 220,
3–15
2.
Ceni, C., Unsain, N., Zeinieh, M. P., and Barker, P. A. (2014) Neurotrophins in
the regulation of cellular survival and death. Handb Exp Pharmacol 220, 193–
221
3.
Chao, M. V. (2003) Neurotrophins and their receptors: a convergence point for
many signalling pathways. Nat. Rev. Neurosci. 4, 299–309
4.
Hempstead, B. L. (2014) Deciphering proneurotrophin actions. Handb Exp
Pharmacol 220, 17–32
5.
Lu, B., Nagappan, G., and Lu, Y. (2014) BDNF and synaptic plasticity,
cognitive function, and dysfunction. Handb Exp Pharmacol 220, 223–250
6.
Friedman, W. J., and Greene, L. A. (1999) Neurotrophin signaling via Trks and
p75. Exp. Cell Res. 253, 131–142
7.
Huang, E. J., and Reichardt, L. F. (2003) Trk receptors: roles in neuronal signal
transduction. Annu. Rev. Biochem. 72, 609–642
8.
Deinhardt, K., and Chao, M. V. (2014) Trk receptors. Handb Exp Pharmacol
220, 103–119
9.
Reichardt, L. F. (2006) Neurotrophin-regulated signalling pathways. Philos.
Trans. R. Soc. Lond. B, Biol. Sci. 361, 1545–1564
10. Kraemer, B. R., Yoon, S. O., and Carter, B. D. (2014) The biological functions
and signaling mechanisms of the p75 neurotrophin receptor. Handb Exp
Pharmacol 220, 121–164
11. Nykjaer, A., Lee, R., Teng, K. K., Jansen, P., Madsen, P., Nielsen, M. S.,
Jacobsen, C., Kliemannel, M., Schwarz, E., Willnow, T. E., Hempstead, B. L.,
and Petersen, C. M. (2004) Sortilin is essential for proNGF-induced neuronal
cell death. Nature 427, 843–848
12. Mi, S., Lee, X., Shao, Z., Thill, G., Ji, B., Relton, J., Levesque, M., Allaire, N.,
Perrin, S., Sands, B., Crowell, T., Cate, R. L., McCoy, J. M., and Pepinsky, R.
B. (2004) LINGO-1 is a component of the Nogo-66 receptor/p75 signaling
complex. Nat. Neurosci. 7, 221–228
13. Wang, K. C., Kim, J. A., Sivasankaran, R., Segal, R., and He, Z. (2002) P75
interacts with the Nogo receptor as a co-receptor for Nogo, MAG and OMgp.
Nature 420, 74–78
19
14. Hempstead, B. L., Martin-Zanca, D., Kaplan, D. R., Parada, L. F., and Chao, M.
V. (1991) High-affinity NGF binding requires coexpression of the trk proto-
oncogene and the low-affinity NGF receptor. Nature 350, 678–683
15. Kanning, K. C., Hudson, M., Amieux, P. S., Wiley, J. C., Bothwell, M., and
Schecterson, L. C. (2003) Proteolytic processing of the p75 neurotrophin
receptor and two homologs generates C-terminal fragments with signaling
capability. J. Neurosci. 23, 5425–5436
16. Skeldal, S., Sykes, A. M., Glerup, S., Matusica, D., Palstra, N., Autio, H.,
Boskovic, Z., Madsen, P., Castrén, E., Nykjaer, A., and Coulson, E. J. (2012)
Mapping of the interaction site between sortilin and the p75 neurotrophin
receptor reveals a regulatory role for the sortilin intracellular domain in p75
neurotrophin receptor shedding and apoptosis. J. Biol. Chem. 287, 43798–43809
17. Jung, K.-M., Tan, S., Landman, N., Petrova, K., Murray, S., Lewis, R., Kim, P.
K., Kim, D. S., Ryu, S. H., Chao, M. V., and Kim, T.-W. (2003) Regulated
intramembrane proteolysis of the p75 neurotrophin receptor modulates its
association with the TrkA receptor. J. Biol. Chem. 278, 42161–42169
18. Bothwell, M. (1995) Functional interactions of neurotrophins and neurotrophin
receptors. Annu. Rev. Neurosci. 18, 223–253
19. Chao, M. V., and Hempstead, B. L. (1995) p75 and Trk: a two-receptor system.
Trends Neurosci. 18, 321–326
20. Yano, H., and Chao, M. V. (2000) Neurotrophin receptor structure and
interactions. Pharm. Acta Helv. 74, 253–260
21. Bibel, M., Hoppe, E., and Barde, Y. A. (1999) Biochemical and functional
interactions between the neurotrophin receptors trk and p75NTR. EMBO J. 18,
616–622
22. Rifkin, J. T., Todd, V. J., Anderson, L. W., and Lefcort, F. (2000) Dynamic
expression of neurotrophin receptors during sensory neuron genesis and
differentiation. Dev. Biol. 227, 465–480
23. White, F. A., Silos-Santiago, I., Molliver, D. C., Nishimura, M., Phillips, H.,
Barbacid, M., and Snider, W. D. (1996) Synchronous onset of NGF and TrkA
survival dependence in developing dorsal root ganglia. J. Neurosci. 16, 4662–
4672
24. Mahadeo, D., Kaplan, L., Chao, M. V., and Hempstead, B. L. (1994) High
affinity nerve growth factor binding displays a faster rate of association than
p140trk binding. Implications for multi-subunit polypeptide receptors. J. Biol.
Chem. 269, 6884–6891
25. Lee, K. F., Davies, A. M., and Jaenisch, R. (1994) p75-deficient embryonic
dorsal root sensory and neonatal sympathetic neurons display a decreased
sensitivity to NGF. Development 120, 1027–1033
26. Davies, A. M., Lee, K. F., and Jaenisch, R. (1993) p75-deficient trigeminal
sensory neurons have an altered response to NGF but not to other neurotrophins.
Neuron 11, 565–574
27. Majdan, M., and Miller, F. D. (1999) Neuronal life and death decisions
functional antagonism between the Trk and p75 neurotrophin receptors. Int. J.
Dev. Neurosci. 17, 153–161
28. Majdan, M., Walsh, G. S., Aloyz, R., and Miller, F. D. (2001) TrkA mediates
developmental sympathetic neuron survival in vivo by silencing an ongoing
p75NTR-mediated death signal. J. Cell Biol. 155, 1275–1285
29. Yoon, S. O., Casaccia-Bonnefil, P., Carter, B., and Chao, M. V. (1998)
Competitive signaling between TrkA and p75 nerve growth factor receptors
20
determines cell survival. J. Neurosci. 18, 3273–3281
30. Lee, K. F., Li, E., Huber, L. J., Landis, S. C., Sharpe, A. H., Chao, M. V., and
Jaenisch, R. (1992) Targeted mutation of the gene encoding the low affinity
NGF receptor p75 leads to deficits in the peripheral sensory nervous system.
Cell 69, 737–749
31. Wehrman, T., He, X., Raab, B., Dukipatti, A., Blau, H., and Garcia, K. C.
(2007) Structural and mechanistic insights into nerve growth factor interactions
with the TrkA and p75 receptors. Neuron 53, 25–38
32. King, C., Stoneman, M., Raicu, V., and Hristova, K. (2016) Fully quantified
spectral imaging reveals in vivo membrane protein interactions. Integr Biol
(Camb) 8, 216–229
33. King, C., Raicu, V., and Hristova, K. (2017) Understanding the FRET
signatures of interacting membrane proteins. J. Biol. Chem. 292, 5291–5310
34. Esposito, D., Patel, P., Stephens, R. M., Perez, P., Chao, M. V., Kaplan, D. R.,
and Hempstead, B. L. (2001) The cytoplasmic and transmembrane domains of
the p75 and Trk A receptors regulate high affinity binding to nerve growth
factor. J. Biol. Chem. 276, 32687–32695
35. Nadezhdin, K. D., García-Carpio, I., Goncharuk, S. A., Mineev, K. S., Arseniev,
A. S., and Vilar, M. (2016) Structural basis of p75 transmembrane domain
dimerization. J. Biol. Chem. 291, 12346–12357
36. Franco, M. L., Nadezhdin, K. D., Goncharuk, S. A., Mineev, K. S., Arseniev, A.
S., and Vilar, M. (2020) Structural basis of the transmembrane domain
dimerization and rotation in the activation mechanism of the TRKA receptor by
nerve growth factor. J. Biol. Chem. 295, 275–286
37. Kot, E. F., Wang, Y., Goncharuk, S. A., Zhang, B., Arseniev, A. S., Wang, X.,
and Mineev, K. S. (2020) Oligomerization analysis as a tool to elucidate the
mechanism of EBV latent membrane protein 1 inhibition by pentamidine.
Biochim. Biophys. Acta Biomembr. 1862, 183380
38. Lesovoy, D. M., Mineev, K. S., Bragin, P. E., Bocharova, O. V., Bocharov, E.
V., and Arseniev, A. S. (2017) NMR relaxation parameters of methyl groups as
a tool to map the interfaces of helix-helix interactions in membrane proteins. J.
Biomol. NMR 69, 165–179
39. Chavent, M., Chetwynd, A. P., Stansfeld, P. J., and Sansom, M. S. P. (2014)
Dimerization of the EphA1 receptor tyrosine kinase transmembrane domain:
Insights into the mechanism of receptor activation. Biochemistry 53, 6641–6652
40. Shahane, G., Ding, W., Palaiokostas, M., and Orsi, M. (2019) Physical
properties of model biological lipid bilayers: insights from all-atom molecular
dynamics simulations. J Mol Model 25, 76
41. Chang, M.-S., Arevalo, J. C., and Chao, M. V. (2004) Ternary complex with
Trk, p75, and an ankyrin-rich membrane spanning protein. J. Neurosci. Res. 78,
186–192
42. Zampieri, N., and Chao, M. V. (2006) Mechanisms of neurotrophin receptor
signalling. Biochem. Soc. Trans. 34, 607–611
43. Franco, M. L., García-Carpio, I., Comaposada-Baró, R., Escribano-Saiz, J. J.,
Chávez-Gutiérrez, L., and Vilar, M. (2021) TrkA-mediated endocytosis of p75-
CTF prevents cholinergic neuron death upon γ-secretase inhibition. Life Sci.
Alliance 4
44. Huber, L. J., and Chao, M. V. (1995) A potential interaction of p75 and trkA
NGF receptors revealed by affinity crosslinking and immunoprecipitation. J.
Neurosci. Res. 40, 557–563
21
45. Ross, A. H., Daou, M. C., McKinnon, C. A., Condon, P. J., Lachyankar, M. B.,
Stephens, R. M., Kaplan, D. R., and Wolf, D. E. (1996) The neurotrophin
receptor, gp75, forms a complex with the receptor tyrosine kinase TrkA. J. Cell
Biol. 132, 945–953
46. Wolf, D. E., McKinnon, C. A., Daou, M. C., Stephens, R. M., Kaplan, D. R.,
and Ross, A. H. (1995) Interaction with TrkA immobilizes gp75 in the high
affinity nerve growth factor receptor complex. J. Biol. Chem. 270, 2133–2138
47. Iacaruso, M. F., Galli, S., Martí, M., Villalta, J. I., Estrin, D. A., Jares-Erijman,
E. A., and Pietrasanta, L. I. (2011) Structural model for p75(NTR)-TrkA
intracellular domain interaction: a combined FRET and bioinformatics study. J.
Mol. Biol. 414, 681–698
48. Sykes, A. M., Palstra, N., Abankwa, D., Hill, J. M., Skeldal, S., Matusica, D.,
Venkatraman, P., Hancock, J. F., and Coulson, E. J. (2012) The effects of
transmembrane sequence and dimerization on cleavage of the p75 neurotrophin
receptor by γ-secretase. J. Biol. Chem. 287, 43810–43824
49. Holtzman, D. M., Kilbridge, J., Li, Y., Cunningham, E. T., Lenn, N. J., Clary,
D. O., Reichardt, L. F., and Mobley, W. C. (1995) TrkA expression in the CNS:
evidence for the existence of several novel NGF-responsive CNS neurons. J.
Neurosci. 15, 1567–1576
50. Verge, V. M., Merlio, J. P., Grondin, J., Ernfors, P., Persson, H., Riopelle, R. J.,
Hökfelt, T., and Richardson, P. M. (1992) Colocalization of NGF binding sites,
trk mRNA, and low-affinity NGF receptor mRNA in primary sensory neurons:
responses to injury and infusion of NGF. J. Neurosci. 12, 4011–4022
51. Ahmed, F., Paul, M. D., and Hristova, K. (2020) The biophysical basis of
receptor tyrosine kinase ligand functional selectivity: Trk-B case study.
Biochem. J. 477, 4515–4526
52. Ahmed, F., and Hristova, K. (2018) Dimerization of the Trk receptors in the
plasma membrane: effects of their cognate ligands. Biochem. J. 475, 3669–3685
53. Paul, M. D., Grubb, H. N., and Hristova, K. (2020) Quantifying the strength of
heterointeractions among receptor tyrosine kinases from different subfamilies:
Implications for cell signaling. J. Biol. Chem. 295, 9917–9933
54. Chen, L., Novicky, L., Merzlyakov, M., Hristov, T., and Hristova, K. (2010)
Measuring the energetics of membrane protein dimerization in mammalian
membranes. J. Am. Chem. Soc. 132, 3628–3635
55. Sinha, B., Köster, D., Ruez, R., Gonnord, P., Bastiani, M., Abankwa, D., Stan,
R. V., Butler-Browne, G., Vedie, B., Johannes, L., Morone, N., Parton, R. G.,
Raposo, G., Sens, P., Lamaze, C., and Nassoy, P. (2011) Cells respond to
mechanical stress by rapid disassembly of caveolae. Cell 144, 402–413
56. Raicu, V., Stoneman, M. R., Fung, R., Melnichuk, M., Jansma, D. B., Pisterzi,
L. F., Rath, S., Fox, M., Wells, J. W., and Saldin, D. K. (2009) Determination of
supramolecular structure and spatial distribution of protein complexes in living
cells. Nat. Photonics 3, 107–113
57. Biener, G., Stoneman, M. R., Acbas, G., Holz, J. D., Orlova, M., Komarova, L.,
Kuchin, S., and Raicu, V. (2013) Development and experimental testing of an
optical micro-spectroscopic technique incorporating true line-scan excitation.
Int. J. Mol. Sci. 15, 261–276
58. Sarabipour, S., King, C., and Hristova, K. (2014) Uninduced high-yield
bacterial expression of fluorescent proteins. Anal. Biochem. 449, 155–157
59. Lukat, G., Krüger, J., and Sommer, B. (2013) APL@Voro: a Voronoi-based
membrane analysis tool for GROMACS trajectories. J. Chem. Inf. Model. 53,
22
2908–2925
FIGURE LEGENDS
Figure 1. TrkA-p75 FSI-FRET experiments. (A) Illustrations of the TrkA-mTurq
and p75-eYFP proteins used in FRET experiments along with the LAT-mTurq and
ECTM FGFR3-eYFP proteins used in control experiments. (B) FRET efficiencies as
a function of total receptor concentration measured for TrkA-mTurq and p75-eYFP in
the absence of ligand compared to a zero FRET control dataset. (C) Illustrations of
some possible stoichiometries of the TrkA-p75 heterocomplex: (i) heterodimer, (ii)
heterotrimer of two TrkA and one p75, (iii) heterotrimer of one TrkA and two p75,
(iv) heterotetramer or two TrkA and two p75. (D) FRET data for TrkA-mTurq and
p75-eYFP in the presence of 100 ng/µL NGF compared to the data in the absence of
NGF. (E) The FRET data for TrkA and p75 in the presence of NGF compared to the
zero FRET control dataset. (F) Expression of TrkA-mTurq and p75-eYFP measured
on the cell surface for the experiments performed in the absence and presence of
NGF. (G) The FRET data for TrkA-p75 in the absence and presence of NGF and for
the control dataset were binned and compared. (H) Illustrations of the possible
consequences of NGF binding to the TrkA-p75 heterocomplex, which could be either
dissociation of the heterocomplex to stabilize the respective homodimers or an NGF-
induced conformational change.
Figure 2. p75/TrkA interactions as observed by NMR.
23
A) Overlay of two 15N-TROSY experiments: (black) 15N-labeled p75 without TrkA
and (red) 15N-labeled p75 after addition of unlabeled TrkA with p75:TrkA molar ratio
1:4. 1H-15N assignments of p75 backbone amid proton resonances are provided. B)
15N-labeled p75-TM-C257A titration with unlabeled TrkA TM. Left to right: p75
monomer (black), p75-p75 homodimer (blue) and p75-TrkA heterodimer (red) states
are observed in the G266 amide proton cross-peak in 1H/15N-HSQC spectra. G266
was chosen as representative as its cross-peak is situated away from other peaks and it
shows clear monomer-homodimer-heterodimer transitions. C) Chemical shift changes
observed upon interaction with TrkA are shown on top of p75-TM sequence. D)
Population of p75-p75 homodimers relative to that of p75-TrkA heterodimers (p75-
p75 peak intensity is divided by sum of p75-p75 and p75-TrkA peak intensities),
expressed as a function of the p75/TrkA molar ratio. The population of p75-p75 dimer
decreases while that of p75-TrkA dimer increases as more TrkA is added to the
sample. For all experiments the lipid to protein molar ratio (LPR) remains constant at
80. E) 15N-labeled TrkA-TMD titration with unlabeled p75-TM-C257A. Left to right:
TrkA monomer (black), TrkA-TrkA homodimer (blue) and p75-TrkA heterodimer
(red) states are observed in the amide proton cross-peak in 1H/15N-HSQC spectra. F)
Chemical shift changes observed upon interaction with p75 are shown on top of
TrkA-TMD sequence. G) Population of TrkA-TrkA homodimers relative to that of
TrkA-p75 heterodimers (TrkA-TrkA peak intensity is divided by sum of TrkA-TrkA
and TrkA-p75 peak intensities), expressed as a function of the TrkA/p75 molar ratio.
The population of TrkA-TrkA dimer decreases while that of TrkA-p75 dimer
increases as more p75 is added to the sample.
Figure 3. Multiscale Molecular dynamics of TrkA-TMD and p75-TMD
A) Coarse-grained TrkA-TMD and p75-TMD helix dimerization simulation. The
initial system configuration (0 µs) consists of two helices (red and blue) inserted in a
POPC bilayer in a parallel orientation with an interhelix separation of dHH ≈ 55 Å.
The choline, phosphate and glycerol (gray) backbone particles of the POPC molecules
are shown. The snapshot at 5 µs illustrates the stable TM helix heterodimer. B)
Distance between TrkA-TMD and p75-TMD during CG-MD simulation time. C)
Structural models of the final conformations from the 12 simulations. In blue p75 and
in red TrkA is shown. D) Superposition of the 7 conformations with lowest rmsd
24
found by CG-MD. E) Interhelical distance between p75-TMD-W276 and TrkA-
TMD-K441 in the FA-MD simulation done by triplicate.
Figure 4. Structural models of the p75/TrkA TMD heterodimer.
A-B) Schematic representation of the spatial structure of the heterodimer p75-TMD
(blue) and TrkA-TMD (orange) after 100 ns full-atom MD. The residues participating
in the dimer interface are shown by blue (p75) and red (TrkA). C) Protein sequence
alignement of TrkA, TrkB and TrkC TMDs. In bold the conserved residues.
Figure 5. TrkA activation is modulated by p75-TMD.
A) Protein sequences alignment of the different mutant constructs of p75-TMD. The
residue mutated is shown in bold. B) Western blots of lysates from Hela cells
transfected with the indicated constructs and stimulated with increasing
concentrations of NGF. Membranes were probed using a TrkA-P-Tyr675 specific
antibody. C-D) Normalized activation of TrkA using increasing concentrations of
NGF in the absence of the presence of p75 mutant constructs indicated. Bars represent
the standard error of at least three independent experiments. P values are reported in
the text.
Figure 6. Effect of the mutation of the p75 heterodimer interface.
A-B) Result of 12 simulations by CG-MD of p75-TMD-AGA mutant (A) or p75-
TMD-wt (B) and TrkA-TMD in POPC model membranes. The position of the p75-
TMD helix (gray) respect to the TrkA-TMD (red) after each simulation is shown. In
green and red are shown the residues that belong to the active and inactive
homodimer interface of TrkA described in Franco et al. C) Quantification of the
neurite length (µm) of PC12 cells electroporated with the indicated constructs and
GFP at 24h of addition of NGF (50 ng/mL). Bars represent the standard error of at
least three independent electroporation experiments. Statistical analysis was
performed with one-way Anova analysis was used and the P values are reported
above each bar.
B) Representative fluorescence microscopy of PC12 cells electroporated with the
indicated constructs stimulated with NGF (50 ng/mL) for 24 hours post-
electroporation. Bar represents 50 µm.
! thd An
1 P76, n0 THA
Wh pTSTHA 14
MIE ae
end
distance (nm)
hTrkA TPFGVSVAVGLAVFACLFLSTLLLV
hTrkB HLSVYAVVVIASVVGFCLLVMLFLL
hTfrkC DTFGVSIAVGLAAFACVLLVVLFVM
MLTPVYCSILAAVYVGLVAYIATERIE
MLIPVYCBIEAIVYVILVEYIATER
PIS-TMER-TMD VLLPLVIFFOLCLLGLLFIGLNYRYGRM
BTA TRA THA Tika
° prt pISTNFR — p7S-AGAAH
2011 1010 0.011 seiee 9.011 10140 0.81 1 10100 NOH fapal)
Paymsreers)
‘TAA mat
THA + prs
A sa
seve P=00001
sideview
control p7S-wt— pT7S-AGAMIIE
control —p7S-wt_—p?S-AGAVIII
‘
topview : J
a Y t
p7S-wt/TtkaA p?S-AGA/THKA NGF (SOngimL)2h
a
oman
2 w:
§ 40:
=)
o
contol ——-~piSwt_p?S-AGAI
D
control p7Swt——_pTSAGAI
one 20h ania} 24
| 2021 | Neurotrophin signaling is modulated by specific transmembrane domain interactions | 10.1101/2021.05.24.445441 | [
"Franco María L.",
"Nadezhdin Kirill D.",
"Light Taylor P.",
"Goncharuk Sergey A.",
"Soler-Lopez Andrea",
"Ahmed Fozia",
"Mineev Konstantin S.",
"Hristova Kalina",
"Arseniev Alexander S.",
"Vilar Marçal"
] | creative-commons |
beta-blocker reverses inhibition of beta-2 adrenergic receptor resensitization
by hypoxia
Yu Sun1, Manveen K. Gupta1, Kate Stenson1, Maradumane L. Mohan1, Nicholas Wanner2,
Kewal Asosingh2, Serpil Erzurum2 and Sathyamangla V. Naga Prasad1
Department of Cardiovascular and Metabolic Sciences1, and Inflammation and Immunity2,
Lerner Research Institute, Cleveland Clinic, Cleveland, OH, 44195.
Address correspondence to:
Sathyamangla V. Naga Prasad, PhD, FAHA
Professor of Molecular Medicine
Cleveland Clinic Lerner College of Medicine at Case Western Reserve University
Staff, Department of Cardiovascular and Metabolic Sciences
Lerner Research Institute
Cleveland Clinic
9500 Euclid Avenue
Cleveland, 44195
Abstract
Ischemia/hypoxia is major underlying cause for heart failure and stroke. Although beta-
adrenergic receptor (βAR) is phosphorylated in response to hypoxia, less is known about the
underlying mechanisms. Hypoxia results in robust GRK2-mediated β2AR phosphorylation but
does not cause receptor internalization. However, hypoxia leads to significant endosomal-β2AR
phosphorylation accompanied by inhibition of β2AR-associated protein phosphatase 2A (PP2A)
activity impairing resensitization. Phosphoinositide 3-kinase γ (PI3Kγ) impedes resensitization
by phosphorylating endogenous inhibitor of protein phosphatase 2A, I2PP2A that inhibits PP2A
activity. Hypoxia increased PI3Kγ activity leading to significant phosphorylation of I2PP2A
resulting in inhibition of PP2A and consequently resensitization. Surprisingly, β-blocker
abrogated hypoxia-mediated β2AR phosphorylation instead of phosphorylation in normoxia.
Subjecting mice to hypoxia leads to significant cardiac dysfunction and β2AR phosphorylation
showing conservation of non-canonical hypoxia-mediated pathway in vivo. These findings
provide mechanistic insights on hypoxia-mediated βAR dysfunction which is rescued by β-
blocker and will have significant implications in heart failure and stroke.
Introduction
Oxygen is a key currency driving the sustenance of the cells as it plays a central role in
metabolism and respiration [1]. The need of oxygen for such a fundamental and existential
physiology has led the eukaryotes to develop exquisite mechanisms to maintain and match the
ever changing needs of oxygen by the cells/tissues [1, 2]. The reduction in oxygen supply
classically occurs due to increased demand that exceeds the supply either locally or systemically
causing hypoxia leading to metabolic crisis with implications in cell survival. In recognition of
the critical role oxygen plays in functional homeostasis, eukaryotes have developed an efficient
and rapid oxygen sensing system, the hypoxia-inducible factors (HIFs) which are master
transcription factors [2-5]. HIF family is represented by members HIF-1, -2 and -3 off which
HIF-1α isoform is most well studied. HIF-1α is stabilized in hypoxia and dimerizes with HIF-1β
to form a potent transcription factor that drives the hypoxia response [2, 6-8]. However, our
previous work has shown that HIF-1α can be stabilized by beta-adrenergic receptor (βAR)
activation in normoxia [9]. While, it is also known that β2ARs are regulated by oxygen through
hydroxylation that alters β2AR stability and responses [9, 10].
βARs are prototypic G-protein coupled receptors (GPCRs) that play a key role in cardiac
function [11] wherein their sustained dysfunction is associated with deleterious cardiac
remodeling and heart failure [11, 12]. There are three sub-types of βARs (β1, β2, and β3AR) of
which β2AR is ubiquitously expressed while, β1AR is primarily expressed in the heart. Agonist
binding to βARs like endogenous ligands epinephrine and norepinephrine leads to G-protein
coupling resulting in dissociation of hetero-trimeric G-protein into Gαs and Gβγ subunits, and
cAMP generation [11, 13-15]. The dissociated Gβγ subunits recruit GPCR kinase 2 (GRK2) to
the receptor leading to βAR phosphorylation and desensitization ie inability to couple to G-
protein despite agonist [11, 16]. As adaptor scaffolding protein β-arrestin binds to
phosphorylated βAR and targets it for endocytosis [17, 18]. The endocytosed β2AR undergoes
dephosphorylation in the endosomes before being recycled back to the plasma membrane as
naïve receptors [17, 19]. Our previous studies have shown that phosphoinositide 3-kinase γ
(PI3Kγ) that is recruited to the receptor complex inhibits protein phosphatase 2A (PP2A) at
plasma membrane through phosphorylation of inhibitor of PP2A (I2PP2A) [20]. Thus, agonist
activation leads to increased phosphorylation of β2AR due to kinase activity of GRK2 and
simultaneous inhibition of PP2A by PI3Kγ. In contrast to this traditional agonist mediated
mechanisms, we have shown that hypoxia causes β2AR phosphorylation in the absence of
agonist which is associated with HIF-1α accumulation [9]. Furthermore, GRK2 inhibition
results in reduction of hypoxia induced-β2AR phosphorylation and HIF1-α accumulation [9].
Interestingly, despite agonist-independent βAR dysfunction, use of β-blocker surprisingly
reduces HIF-1α accumulation with hypoxia [9].
Traditionally, β-blockers are antagonists that block the activation of the βARs and as a
consequence there is no G-protein coupling and cAMP generation [21]. Accumulating evidence
has shown that β-blocker can mediate biased signaling wherein, they block G protein-dependent
signaling while simultaneously initiating G protein-independent β-arrestin dependent signaling
[22, 23]. Studies have shown that β-blockers mediate downstream G protein-independent
signaling through EGFR transactivation [21]. This suggests that β-blockers are able to confer a
unique βAR conformation that allows for G protein independent signaling. Consistent with this
idea that βARs can attain different conformations that allows receptors to activate unique
downstream signal, studies have shown that hypoxia leads to unique phosphorylation bar-code
on the receptor that regulates HIF-1α accumulation [9]. Given that the mechanistic
underpinnings of this regulation is not well understood, our current studies have focused on
identifying the determinants of the unique non-canonical agonist-independent hypoxia mediated
regulation of β2AR function. We show in our current study that in addition to selective
upregulation of GRK2, there also simultaneous inhibition of PP2A-mediated resensitization
accounting for accumulation of phosphorylated β2ARs. Consistent with inhibition of
resensitization, there is significant increase in endosomal PI3Kγ activity and concomitant
reduction in β2AR-associated phosphatase activity. Moreover, there was marked increase in
I2PP2A phosphorylation accounting for the loss in PP2A activity. Correspondingly, subjecting
mice to acute hypoxia resulted in deleterious cardiac remodeling associated with significant βAR
dysfunction showing conservation of these pathways in vivo. Surprisingly, β-blocker
propranolol in hypoxia reversed β2AR phosphorylation in contrast to normoxia wherein, it
consistently induced β2AR phosphorylation showing non-canonical regulation of β2ARs by
hypoxia.
Results
Selective increase in GRK2 mediates hypoxia-induced β2AR dysfunction:
To test whether hypoxia per se causes β2AR phosphorylation, HEK 293 cells stably expressing
FLAG-β2AR (β2AR-HEK 293 cells) were serum starved and subjected to hypoxia for 0, 3 and 6
hours. Immunoblotting of cell lysates with anti-phospho-β2AR antibody showed significant
phosphorylation of β2AR by 6 hours (n=5) [Fig. 1A]. The membranes were stripped and re-
blotted for FLAG as a loading control for the β2AR expression in the cells [Fig. 1A]. To further
determine whether the cells were subjected to hypoxia stress, the membranes were stripped and
re-immunoblotted with anti-HIF-1α antibody. Accumulation of HIF-1α occurred by 6 hours
with no appreciable difference at 3 hours of hypoxia treatment [Fig. 1A] showing increase in
β2AR phosphorylation associated with HIF-1α accumulation. Confocal microscopy showed that
in contrast to normoxia, hypoxia resulted in significant accumulation of phosphorylated β2AR as
visualized by anti-phospho-β2AR antibody (green) (n=4)[Fig. 1B (panels 2 & 6) and C].
Since GRKs mediate phosphorylation of β2ARs, cell lysates were assessed to determine which
GRKs are involved in mediating receptor phosphorylation in response to hypoxia.
Comprehensive immunoblotting for ubiquitously expressed GRKs (GRK 2, 3, 5 and 6 [24])
showed selective and significant increase only in GRK2 expression (n=5) [Fig. 2A & B] with no
appreciable changes in other GRKs [Fig. 2A]. Given the consistent correlation between GRK2
upregulation and β2AR phosphorylation at 6 hours, all the studies described from hereon used 6
hours of hypoxia treatment for assessing mechanistic underpinnings of βAR dysfunction. To test
whether GRK2 activity is sufficient to mediate β2AR phosphorylation in response to hypoxia,
β2AR-HEK 293 cells were pre-treated with GRK2 inhibitor paroxetine. Paroxetine treatment
abrogated the hypoxia-mediated phosphorylation of β2ARs (n=4) [Fig. 2C] showing the GRK2
is the key kinase that phosphorylates β2ARs following hypoxia. Given that hypoxia causes
β2AR phosphorylation indicating receptor dysfunction, immediate downstream signal was
assessed by measuring cAMP level in β2AR-HEK 293 cells. Consistent with the loss in β2AR
function, there was significant reduction in the amount of cAMP following hypoxia (n=5) [Fig.
2D]. To further test whether G protein coupling is altered following hypoxia, plasma membranes
and endosomal fractions were isolated from normoxia and hypoxia treated cells. These fractions
were subjected to in vitro isoproterenol (ISO) (βAR agonist) stimulation to assess for G protein
coupling. Increased adenylyl cyclase activity was observed following ISO stimulation in the
plasma membrane (n=5) [Fig. 2E] as well as in endosomal fractions (n=5) [Fig. 2F] in normoxia
treated cells. While significant reduction in adenylyl cyclase activity was observed in hypoxia
treated cells showing that hypoxia-induced β2AR phosphorylation impairs receptor activation
promoting β2AR dysfunction.
Hypoxia disengages β2AR phosphorylation-internalization axis:
Traditionally, β2AR phosphorylation by GRK2 mediates β-arrestin recruitment leading to
desensitization and internalization of the receptors into the endosomes. To test whether β-
arrestin plays a role in hypoxia-induced β2AR desensitization and dysfunction, β-arrestin 2 GFP-
β2AR double-stable HEK 293 cells underwent normoxia or hypoxia treatment (6 hours) or ISO
stimulation for 10 minutes (that was used as a positive control). Consistent with previous studies
[25], ISO stimulation resulted in significant recruitment of β-arrestin 2 GFP to the plasma
membrane (green) (n=3) [Fig. 3A, panels 5 & 8] as assessed by the clearance of the cytosolic β-
arrestin 2 GFP. This is associated with a subset of β2ARs that are phosphorylated (red) [Fig.
3A, panels 6 & 8] and internalized. While hypoxia results in marked phosphorylation of β2ARs
(red) [Fig. 3A, panels 10 & 12], no appreciable changes in distribution of β-arrestin 2 GFP was
observed [Fig. 3A, panels 9 & 12] suggesting non-canonical regulation of β2AR by hypoxia.
Since we observed marked accumulation of phosphorylated β2ARs in the cytosol following
hypoxia, we tested whether hypoxia mediates internalization of receptors following β2AR
phosphorylation by pretreating the cells with internalization blockers (sucrose and β-cyclodextrin
[20]). Consistent with previous studies [20], ISO treatment resulted in significant
phosphorylation of β2ARs (green) that decorates the plasma membrane as receptors do not
internalize (n=3) [Fig. 3B, panels 4 & 6]. In contrast, despite pre-treatment with internalization
blockers, accumulation of phosphorylated β2ARs were observed in the cytosol [Fig. 3B, panels
7 & 9] showing that internalization blockers do not alter hypoxia mediated phosphorylation
and/or internalization. To directly test whether hypoxia mediates internalization of
phosphorylated β2ARs, radio-ligand binding was performed on plasma membrane and
endosomal fractions from β2AR-HEK 293 cells subjected to normoxia or hypoxia. Surprisingly,
radio-ligand binding showed no appreciable differences in the β2AR distribution following
hypoxia (n=6) [Fig. 3C & D]. These observations suggest that hypoxia mediates
phosphorylation of endosomal β2ARs independent of internalization consistent with the findings
internalization blockers [Fig. 3B].
Endosomal accumulation of phosphorylated β2ARs with hypoxia is associated with
inhibition of resensitization:
Given the observation that hypoxia mediates β2AR phosphorylation independent of
internalization, β2AR phosphorylation was assessed by immunoblotting of the plasma membrane
and endosomal fractions following hypoxia. Significant β2AR phosphorylation was observed in
the endosomal fractions of cells subjected to hypoxia when compared to normoxia while no
appreciable differences were observed at the plasma membrane (n=5) [Fig. 4A left and right
panels]. Endosomal β2ARs traditionally undergo dephosphorylation/resensitization by PP2A
[26]. Since PP2A is acutely regulated by PI3Kγ activity [20], PI3Kγ was immunoprecipitated
from plasma membrane and endosomal fractions and the immunoprecipitates were subjected to
in vitro lipid kinase activity. Significant PI3Kγ activity was observed in the endosomal fractions
following hypoxia compared to normoxia (n=4) [Fig. 4B, right panel] while no appreciable
differences were observed in the plasma membrane fractions [Fig. 4B, left panel]. As
endosomal PI3Kγ activity is higher, we assessed β2AR-associated phosphatase activity by
immunoprecipitating β2ARs by using anti-FLAG antibody from plasma membrane and
endosomal fractions. While no appreciable difference was observed in β2AR-associated
phosphatase activity at the plasma membrane (n=6) [Fig. 4C, left panel], significant reduction in
β2AR-associated phosphatase activity was observed in the endosomal fraction (n=6) [Fig. 4C,
right panel]. Since we have previously shown that PI3Kγ inhibits PP2A activity by
phosphorylating the endogenous inhibitor of PP2A, I2PP2A, immunoblotting was performed to
assess I2PP2A phosphorylation using an in-house generated anti-phospho-I2PP2A antibody.
Although total I2PP2A levels did not change [Fig. 4D], significant increase in I2PP2A
phosphorylation was observed in hypoxia compared to normoxia (n=4) [Fig. 4D, left and right
panel]. Despite reduced PP2A activity, there was no appreciable difference in the expression of
PP2A following hypoxia [Fig. 4D]. These observations show that hypoxia mediates activation
of PI3Kγ inhibiting resensitization leading to accumulation of phosphorylated β2ARs in the
endosomal fractions accounting for receptor dysfunction.
Hypoxia causes adverse cardiac remodeling and is associated with β2AR dysfunction:
Since it is known that hypoxia/ischemia is one of the leading causes of heart failure and stroke,
studies were performed to assess whether acute hypoxia can cause deleterious cardiac
remodeling. C57Bl6 mice were placed in hypoxia chamber for 20 hours [27] and cardiac
function was assessed by echocardiography. Acute hypoxia resulted in deleterious cardiac
remodeling as observed by increased cardiac lumen post-hypoxia (n=12) [Fig. 5A, upper panel]
and as measured by functional parameters of % fractional shortening (%FS) and % ejection
fraction (%EF) [Fig. 5A, lower panel]. Consistently, significant increase in heart weight to
body ratio (HW/BW) was observed in mice subjected to hypoxia (n=12) [Fig. 5B] and H & E
staining showed increased ventricular lumen following hypoxia (n=4) [Fig. 5C]. Since βARs are
powerful regulators of cardiac function, we assessed whether acute hypoxia causes increase in
cardiac β2AR phosphorylation. Immunoblotting of cardiac lysates showed significant increase
in β2AR phosphorylation following hypoxia (n=6) [Fig. 5D, upper panel and 5E, left panel].
HIF-1α, the sentinel marker for hypoxia was also significantly stabilized in the hypoxia
compared to normoxia [Fig. 5D, middle panel and 5E, right panel]. To test whether increased
phosphorylation of β2AR is associated with receptor dysfunction, in vitro ISO-stimulated
adenylyl cyclase activity was performed on the cardiac plasma membranes. There was
significant reduction in adenylyl cyclase activity following hypoxia both at baseline and upon in
vitro ISO stimulation (n=6) [Fig. 5F] which was preserved in normoxia. Together these findings
show that acute hypoxia causes adverse cardiac remodeling and βAR dysfunction.
β-blocker reverses hypoxia-mediated β2AR phosphorylation:
As β-blocker treatment in hypoxia reduces HIF-1α accumulation [9], experiments were
conducted to test whether β-blocker pre-treatment alters the state of β2AR phosphorylation
despite hypoxia. β2AR-HEK 293 cells were pre-treated with β-blocker propranolol followed by
either hypoxia or normoxia and phosphorylation of β2AR was assessed by immunoblotting.
Consistent with previous studies [28], significant phosphorylation of β2ARs was observed in
normoxia in response to β-blocker (n=5) [Fig. 6A & B]. In contrast, β-blocker treatment in
hypoxia surprisingly resulted in abrogation of hypoxia-mediated β2AR phosphorylation [Fig. 6A
& B]. To further test whether β-blocker treatment results in loss of β2AR phosphorylation,
confocal microscopy was performed following hypoxia. β-blocker treatment in normoxia
resulted in marked increase of phosphorylated β2ARs as visualized by anti-phospho-β2AR
antibody (green) (n=4) [Fig. 6C & 6D (panels 5 and 6)]. Hypoxia resulted in significant
increase in phosphorylated β2ARs [Fig. 6C & 6D (panels 3 and 4)] consistent with our data
[Fig. 1]. In contrast, β-blocker pre-treatment significantly reduced β2AR phosphorylation [Fig.
6C & 6D (panels 7 and 8)] showing a unique role of β-blocker in hypoxia. To further test
whether unexpected reduction in phosphorylation by β-blocker in hypoxia is due to the ability of
β-blocker to engage the resensitization pathway in hypoxia. Since hypoxia decreases endosomal
β2AR-associated PP2A activity in hypoxia, FLAG-β2AR was immunoprecipitated from plasma
membrane and endosomal fractions following hypoxia and β-blocker treatment to assess receptor
associated activity. FLAG-β2AR associated PP2A activity was not appreciably different in the
plasma membranes following β-blocker pre-treatment [Fig. 6E, gray bar plasma membrane].
However, there was significant increase in the FLAG-β2AR associated PP2A activity in the
endosomes of cells pre-treated with β-blockers and subjected to hypoxia [Fig. 6E, gray bar
endosomes]. This observation suggests that β-blockers may act differently in hypoxia than in
normoxia wherein, they could mechanistically engage the resensitization pathway to reduce
β2AR phosphorylation and underlie the benefits provided by β-blockers in patients with heart
failure.
Discussion
Here we show that hypoxia leads to β2AR phosphorylation independent of its agonist
indicating unique non-canonical regulation of the receptor. Hypoxia-induced β2AR
phosphorylation is GRK2-dependent as GRK2 is selectively upregulated and GRK2 inhibition
reverses β2AR phosphorylation. Hypoxia also leads to reduced cAMP and decreased adenylyl
cyclase activity suggesting that GRK2 recruitment to the β2ARs may be independent of the Gβγ
subunits [29, 30]. Although GRK2-mediates β2AR phosphorylation, there was no β-arrestin
recruitment to plasma membrane or changes in dynamics of receptor internalization.
Interestingly, hypoxia leads to selective accumulation of phosphorylated β2ARs in the
endosomes with no changes at the plasma membrane. Importantly, hypoxia inhibits β2AR
resensitization as β2AR-associated phosphatase activity was significantly impaired in the
endosomes following hypoxia. Significant PI3Kγ activity was observed only in the endosomal
fractions upon hypoxia consistent with inhibition of PP2A through the PI3Kγ-I2PP2A axis [20].
This is supported by the findings of significant I2PP2A phosphorylation showing that hypoxia
mediates β2AR phosphorylation by activating GRK2 while simultaneously inhibiting PP2A
which accounts for accumulation of phosphorylated receptors. These observations are
strengthened by in vivo studies showing that acute hypoxia results in significant βAR
dysfunction and is associated with adverse cardiac remodeling. β-blocker pre-treatment
surprisingly reduced hypoxia-mediated β2AR phosphorylation and was associated with increased
β2AR-associated phosphatase activity in contrast to its known role in mediating β2AR
phosphorylation in normoxia.
Previous studies have shown that GRK2 phosphorylation of βARs is one of the key regulators
of HIF-1α stabilization [9]. Consistently, our data shows that GRK2 is the key mediator of
β2AR phosphorylation in hypoxia as inhibition of GRK2 results in loss of β2AR phosphorylation
suggesting that GRK2 plays a critical role in hypoxia-mediated β2AR regulation. Traditionally,
GRK2 is recruited to the βARs by the dissociated Gβγ subunits of the hetero-trimeric G protein
following agonist stimulation of the receptor [11, 24]. However, hypoxia-mediated βAR
phosphorylation is agonist independent suggesting that hypoxia may engage non-canonical
pathways to mediate β2AR phosphorylation. In contrast to the classical GRK2-mediated
phosphorylation that initiates receptor internalization through β-arrestin-dependent pathways [24,
31], there were no significant recruitment of β-arrestin to the β2ARs following hypoxia.
Similarly, there were no differences in β2AR density/distribution between plasma membrane or
endosomal fractions as assessed by radio-ligand binding studies. This suggests that hypoxia may
not engage trafficking/internalization machinery to alter β2AR distribution despite our consistent
observation of increased endosomal β2AR phosphorylation by confocal microscopy or western
immunoblotting studies [Figs. 1, 3 & 4]. These set of unexpected observations brings forth a
unique conceptual idea that hypoxia may directly initiate β2AR phosphorylation in the
endosomes by GRK2 thus, by-passing the need of Gβγ subunits for recruitment. Such an idea
would be consistent with phosphorylation and regulation of non-receptor substrates of GRK2
[30].
Classically following β2AR phosphorylation by GRKs, the receptor is endocytosed and
resensitization occurs by dephosphorylation mediated by PP2A. Our previous study has shown
that PI3Kγ regulates resensitization by inhibiting PP2A activity through phosphorylation of
I2PP2A [20]. Also, agonist stimulation leads to kinase activation while, simultaneously
inhibiting PP2A activity thus buttressing the kinase activation [11, 20]. Hypoxia leads to
accumulation of phosphorylated β2ARs in the endosomes and is associated with significantly
increased endosomal PI3Kγ activity and inhibition of PP2A activity. Furthermore, hypoxia leads
to marked increase in phosphorylation of I2PP2A, the endogenous inhibitor of PP2A showing
that mechanistically PP2A is inhibited by the PI3Kγ-I2PP2A axis. This shows that loss in PP2A
activity and inability to dephosphorylate the receptor in part, contributes to the accumulation of
phosphorylated β2ARs in the endosomes. PI3Kγ is also recruited to plasma membrane by the
dissociated Gβγ subunits [32] but selective increase only in the endosomal PI3Kγ activity under
hypoxia suggests non-canonical regulation of PI3Kγ. Hypoxia activation of PI3Kγ in the cytosol
now mediates phosphorylation of I2PP2A inhibiting PP2A activity and thereby,
dephosphorylation of the receptors. This consequently leads to impairment of resensitization
accounting for accumulation of phosphorylated β2ARs in the endosomes.
Given the recognition that acute hypoxia could underlie stroke due to changes in cardiac
function, mice were subjected to acute hypobaric hypoxia to assess cardiac remodeling.
Interestingly, acute hypoxia resulted in adverse cardiac remodeling with left ventricular cardiac
dysfunction associated with βAR dysfunction as assessed by adenylyl cyclase activity and β2AR
phosphorylation. These observations suggest that βAR dysfunction in acute hypoxia may
underlie the deleterious left ventricular cardiac remodeling compared to studies showing long
term effects of hypoxia that were associated with marked alterations in the right ventricles [33,
34]. These studies support the idea that under conditions of acute hypoxia, the heart may have
difficulty meeting the mechanical demands leading to stroke due to tissue hypoxia/ischemia. In
this regard, recent studies have shown that in vivo β-blocker treatment markedly reduces renal
HIF-1α stabilization and erythropoiesis [9]. While mechanisms underlying role of β-blockers in
hypoxia are not well understood, our data surprisingly shows that β-blocker pre-treatment in
hypoxia leads to loss in β2AR phosphorylation. This is in contrast to the observation that β-
blockers mediate β2AR phosphorylation that initiates G protein-independent β-arrestin signaling
in normoxia [35, 36]. Consistent with this paradigm, our data shows that β2ARs are
significantly phosphorylated with β-blocker in normoxia but this phosphorylation is blocked in
hypoxia due to the presence of the β-blocker. However, these unexpected observations may have
significant clinical implications given that hypoxia per se initiates β2AR dysfunction through
increased accumulation of phosphorylated receptors. Given that accumulation of phosphorylated
βAR leads to reduced cardiac function/output [37-39] pre-disposing to the stroke, β-blocker
treatment in these conditions may reverse the phosphorylation of βARs preserving cardiac
function. This is supported by recent clinical trial wherein, use of β-blocker in patients with
pulmonary hypertension (who are associated with HIF-1α glycolytic shift) showed increased
cAMP [29]. This suggests that β-blocker in hypoxia may resensitize the β2ARs leading to
increased cAMP that may potentially underlie the beneficial outcomes. Such an idea is
supported by the observation of increased β2AR-associated phosphatase activity in the
endosomes of cells pre-treated with β-blockers in hypoxia suggesting a yet to be understood role
of β-blockers in hypoxia. Thus, our study shows that hypoxia mediates β2AR phosphorylation
by selectively increasing GRK2-dependent kinase pathway and simultaneously inhibiting the
PP2A phosphatase pathway through the PI3Kγ-I2PP2A axis [Fig. 7]. Also, unexpectedly our
study identified that β-blocker reduces β2AR phosphorylation in hypoxia suggesting
mechanisms beyond the current understanding of β-blocker function in normoxia and is being
investigated.
Methods
Experimental Animal:
C57/BL6 wild type (WT) mice of either sex 3-6 months of age were subjected to hypoxia (10%
O2) [27] or normoxia for 20 hours. The studies were performed in accordance with institutional
and national guidelines and regulations, as approved by Cleveland Clinic Institutional Animal
Care and Use Committee.
Cell Culture:
HEK 293 cells were maintained in minimum essential media with 10% heat-inactivated fetal
bovine serum and 1% penicillin/streptomycin [30]. Cells were seeded at a standard density of
~1-3 x 105 cells /35 mm dish and the experiments were performed at 70-80% confluence. Cells
were serum starved for 3 or 6 hours when treated with normoxia or hypoxia and for positive
control cells were stimulated with 10µM isoproterenol (ISO) (Sigma-Aldrich) for 10 minutes in
normoxia. For hypoxia studies, cells were incubated in a sealed chamber at 37°C with 2% O2,
5% CO2, balanced with 93% N2. GRK2 inhibitor Paroxetine (30nM) (Sigma-Aldrich) was
added to β2AR-HEK 293 cells 45 minutes prior to normoxia or hypoxia for 6 hours. Similarly,
the cells were pre-treated with propranolol (10 μM) (Sigma-Aldrich) for 45 minutes prior to
incubation in hypoxia for 6 hours. β2AR//β-arrestin 2 GFP double stable cells were pre-treated
with endocytosis inhibitors 0.45M sucrose and 2% β-cyclodextrin for 1 hour and subjected to
normoxia or hypoxia treatment. For control ISO treatment, the cells were pre-treated with
endocytosis blocker for 1 hours and stimulated with 10 µM ISO for 10 minutes. HEK 293 cells
stably expression FLAG-β2AR (β2AR-HEK 293) cells was a generous gift from R. Lefkowtiz,
Duke University, Durham, NC. HEK293 cells stably overexpressing β2AR and β-arrestin 2 GFP
was a generous gift from Dr. Marc G. Caron, Duke University, Durham, NC.
Isolation of Plasma Membranes and Early Endosomes:
Plasma membranes and early endosomes were isolated as previously described [20]. Plasma
membranes were prepared by homogenizing of samples in non-detergent lysis buffer (5mmol/L
Tris-HCl pH 7.5, 5 mM EDTA, 1 mM PMSF, and 2 μg/mL Leupeptin and Aprotinin). Cell
debris/nuclei were removed by centrifugation at 1000 x g for 5 minutes and the supernatant was
centrifuged at 30,000 x g for 30 minutes. Pellet representing membrane fraction was suspended
in 75 mM Tris-HCl pH 7.5, 2 mM EDTA, and 12.5 mM MgCl2 while supernatant was
centrifuged for 1 hour at 100,000 rpm to obtain early endosomes. Endosomes as pellets were re-
suspended in the same buffer as used on plasma membranes.
Confocal microscopy:
β2AR-HEK 293 cells were plated onto glass coverslips pre-treated with 0.01% poly L-Lysine
(Sigma-Aldrich). Cells were serum starved while cultured in normoxia or hypoxia incubator for
6 hours or stimulated with ISO in normoxia as positive control along with endocytosis inhibitors.
The cells were fixed with 4% paraformaldehyde, permeabilized with 0.1% Triton X-100, and
incubated in 5% goat serum in PBS. Anti-phospho-β2AR 355/356 [40] antibody (1:500, Santa
cruz) diluted in 1% goat serum was used as primary antibody, and anti-rabbit Alexa Flour 488
(1:500) was used a secondary antibody (Molecular Probes). Similar treatments were performed
for HEK293 cells stably overexpressing β2AR and β-arrestin 2 GFP except that phospho-β2AR
were visualized by using anti-rabbit Alexa Flour 568 (1:500) as secondary antibody (Molecular
Probes). Samples were visualized using sequential line excitation at 488 and 568 nm for green
and red, respectively. 70 to 100 positive cells were analyzed in each experiment and quantitation
was performed using IMAGE PRO PLUS7 (Media Cybernetics, Inc).
Western immunoblotting:
Standard procedure for western immunoblotting were performed as described previously [30].
The proteins were resolved on SDS-PAGE and transferred to PVDF (BIO-RAD) and assessed
for protein using primary anti-bodies as described below. Antibodies for HIF-1α (1:500),
phosphorylated-β2AR (1:1000), PI3Kγ (1:200), I2PP2A (1:5000), GRK2, 3, 5, 6 diluted at
1:1000 were from Santa Cruz Biotechnology, Flag antibody (1:1000) was from Roche, PP2Ac
antibody (1:2000) was from Upstate Biotechnology (Millipore), β-actin antibody (1:20000) was
from Sigma. Antibody for phosphorylated I2PP2A (anti-phospho-I2PP2A) (1:1000) was
generated in house and described in our recent publication [41].
Phosphatase assay:
PP2A phosphatase activity was measured using phosphatase assay kit (Upstate Biotechnology,
Millipore) following manufacturer’s protocol. Immunoprecipitated samples were resuspended in
the phosphate free assay buffer and incubated in presence or absence of PP2A specific Serine-
Threonine phospho-peptide substrate for 10 minutes. The reaction mix was incubated with acidic
malachite green solution and absorbance was measured at 630 nm in a plate reader.
Lipid kinase assays:
Assays were performed as previously described [42]. Briefly, cells were solubilized in Triton X-
100 lysis buffer (0.8% Triton X-100, 20 mM Tris-HCl pH 7.4, 300 mM NaCl, 1 mM EDTA,
20% glycerol, 1 mM PMSF, 2 μg/ml each of Leupeptin and Aprotinin). PI3Kγ was
immunoprecipitated using anti-PI3Kγ antibody. The beads were washed with lysis buffer and
suspended in reaction buffer TNE (10 mM Tris-HCl, pH 7.4, 150 mM NaCl, 5 mM EDTA, and
100 μM sodium-orthovandate,). To the resuspended beads, 10 μl of 100 mM MgCl2, 10 μl of 2
mg/ml PtdIns (20 μg) sonicated in TE buffer (10 mM Tris–HCl pH 7.4 and 1 mM EDTA), 10 μl
of 440 μM ATP containing 10 μCi of 32P-γ-ATP were added. The assay was performed at 23°C
for 10 min with continuous agitation and stopped by 6N HCl. Lipids were extracted by
chloroform:methanol (1:1) and spotted on to 200 μm silica-coated TLC plates (Selecto-flexible;
Fischer Scientific, Pittsburgh, PA), and phosphorylation was assessed by autoradiography.
βAR Density, Adenylyl Cyclase Activity and cAMP assays:
βAR density was measured as described previously [32]. Briefly, 20 μg plasma membranes or
endosomes were incubated with 250 pmol of [125]I-Cyanopindolol alone or along with 40 μmol/L
ICI (to determine nonspecific binding) at 37°C for 1 hour. The non-specific counts in presence
of ICI were subtracted from the total [125]I-Cyanopindolol counts to calculate for the receptor
density. Adenylyl cyclase assays were determined by incubating 20 μg of membranes or
endosomes at 37°C for 15 min with vehicle, isoproterenol or NaF (G-protein activator) in 50 μL
of assay mixture containing 20 mM Tris-HCl, 0.8 mM MgCl2, 2 mM EDTA, 0.12mM ATP, 0.05
mM GTP, 0.1 mM cAMP, 2.7 mM phosphoenolpyruvate, 0.05 IU/ml myokinase, 0.01 IU/ml
pyruvate kinase and 32P-α-ATP and generated cAMP was quantified by scintillation counting
[30]. The cAMP content in the lysates was determined according to the manufacturer’s
instruction by catch point cAMP immunoassay kit (Molecular Devices) [20].
Immunohistochemistry:
Freshly harvested cardiac samples were placed in fresh 4% paraformaldehyde at room
temperature for 24 hours, followed by ethanol dehydration, xylene exchange, wax soaking and
embedding tissues into paraffin blocks. Paraffin slides (5 μm thickness) were subsequently
stained with H&E. Photographs were taken using a Slide Scanner-Aperio AT2 (Leica
Biosystems).
Echocardiography:
Echocardiography was performed on anesthetized 8 -12 weeks old mice using a VEVO 2100
(VisualSonics) pre- and post-hypoxia treatment as previously described [43]. The mice in
normoxia also were imaged at the same time. M-mode views were recorded including left
ventricular systolic and diastolic dimensions, septum, and posterior wall which were used to
calculate the functional parameters.
Statistical analysis:
Results are expressed as means ± SD. Data were analyzed by t test for two-group comparison
(for example, βAR density in plasma membrane or endosome in Fig. 3B). For comparison of
more than two groups, we used one-way analysis of variance (ANOVA) if there was one
independent variable (for example, p- β2AR densitometric analysis in Fig. 1A) and two-way
ANOVA if there were two independent variables (for example, p- β2AR densitometric analysis
in Fig. 2B and adenylyl cyclase assay in Fig. 5E). A probability value of <0.05 was considered
significant.
Acknowledgments
We would like to thank Dr Sadashiva Karnik and Dr. Sarah Schumacher-Bass for constant
feedback and insightful thoughts during the development and progression of the project.
Funding: These studies were supported by NIH R01 HL089473 and HL128382 (SVNP).
References
1.
Ratcliffe PJ. Oxygen sensing and hypoxia signalling pathways in animals: the implications of
physiology for cancer. The Journal of physiology. 2013;591(8):2027-42. doi:
10.1113/jphysiol.2013.251470. PubMed PMID: 23401619; PubMed Central PMCID: PMC3634517.
2.
Liu Y, Cox SR, Morita T, Kourembanas S. Hypoxia regulates vascular endothelial growth factor
gene expression in endothelial cells. Identification of a 5' enhancer. Circulation research.
1995;77(3):638-43. doi: 10.1161/01.res.77.3.638. PubMed PMID: 7641334.
3.
Beall CM, Cavalleri GL, Deng L, Elston RC, Gao Y, Knight J, et al. Natural selection on EPAS1
(HIF2alpha) associated with low hemoglobin concentration in Tibetan highlanders. Proceedings of the
National Academy of Sciences of the United States of America. 2010;107(25):11459-64. doi:
10.1073/pnas.1002443107. PubMed PMID: 20534544; PubMed Central PMCID: PMC2895075.
4.
Kapitsinou PP, Liu Q, Unger TL, Rha J, Davidoff O, Keith B, et al. Hepatic HIF-2 regulates
erythropoietic responses to hypoxia in renal anemia. Blood. 2010;116(16):3039-48. doi: 10.1182/blood-
2010-02-270322. PubMed PMID: 20628150; PubMed Central PMCID: PMC2974609.
5.
Semenza GL, Wang GL. A nuclear factor induced by hypoxia via de novo protein synthesis binds
to the human erythropoietin gene enhancer at a site required for transcriptional activation. Molecular
and cellular biology. 1992;12(12):5447-54. doi: 10.1128/mcb.12.12.5447. PubMed PMID: 1448077;
PubMed Central PMCID: PMC360482.
6.
Jaakkola P, Mole DR, Tian YM, Wilson MI, Gielbert J, Gaskell SJ, et al. Targeting of HIF-alpha to
the von Hippel-Lindau ubiquitylation complex by O2-regulated prolyl hydroxylation. Science.
2001;292(5516):468-72. doi: 10.1126/science.1059796. PubMed PMID: 11292861.
7.
Wallace DC. Colloquium paper: bioenergetics, the origins of complexity, and the ascent of man.
Proceedings of the National Academy of Sciences of the United States of America. 2010;107 Suppl
2:8947-53. doi: 10.1073/pnas.0914635107. PubMed PMID: 20445102; PubMed Central PMCID:
PMC3024017.
8.
Huang LE, Gu J, Schau M, Bunn HF. Regulation of hypoxia-inducible factor 1alpha is mediated by
an O2-dependent degradation domain via the ubiquitin-proteasome pathway. Proceedings of the
National Academy of Sciences of the United States of America. 1998;95(14):7987-92. doi:
10.1073/pnas.95.14.7987. PubMed PMID: 9653127; PubMed Central PMCID: PMC20916.
9.
Cheong HI, Asosingh K, Stephens OR, Queisser KA, Xu W, Willard B, et al. Hypoxia sensing
through beta-adrenergic receptors. JCI insight. 2016;1(21):e90240. doi: 10.1172/jci.insight.90240.
PubMed PMID: 28018974; PubMed Central PMCID: PMC5161214.
10.
Xie L, Xiao K, Whalen EJ, Forrester MT, Freeman RS, Fong G, et al. Oxygen-regulated beta(2)-
adrenergic receptor hydroxylation by EGLN3 and ubiquitylation by pVHL. Science signaling.
2009;2(78):ra33. doi: 10.1126/scisignal.2000444. PubMed PMID: 19584355; PubMed Central PMCID:
PMC2788937.
11.
Rockman HA, Koch WJ, Lefkowitz RJ. Seven-transmembrane-spanning receptors and heart
function. Nature. 2002;415(6868):206-12. doi: 10.1038/415206a. PubMed PMID: 11805844.
12.
Naga Prasad SV, Nienaber J, Rockman HA. Beta-adrenergic axis and heart disease. Trends in
genetics : TIG. 2001;17(10):S44-9. doi: 10.1016/s0168-9525(01)02487-8. PubMed PMID: 11585676.
13.
Daaka Y, Luttrell LM, Lefkowitz RJ. Switching of the coupling of the beta2-adrenergic receptor to
different G proteins by protein kinase A. Nature. 1997;390(6655):88-91. doi: 10.1038/36362. PubMed
PMID: 9363896.
14.
Ferguson SS. Evolving concepts in G protein-coupled receptor endocytosis: the role in receptor
desensitization and signaling. Pharmacological reviews. 2001;53(1):1-24. PubMed PMID: 11171937.
15.
Rajagopal S, Rajagopal K, Lefkowitz RJ. Teaching old receptors new tricks: biasing seven-
transmembrane receptors. Nature reviews Drug discovery. 2010;9(5):373-86. doi: 10.1038/nrd3024.
PubMed PMID: 20431569; PubMed Central PMCID: PMC2902265.
16.
Lefkowitz RJ, Pitcher J, Krueger K, Daaka Y. Mechanisms of beta-adrenergic receptor
desensitization and resensitization. Advances in pharmacology. 1998;42:416-20. doi: 10.1016/s1054-
3589(08)60777-2. PubMed PMID: 9327928.
17.
Drake MT, Shenoy SK, Lefkowitz RJ. Trafficking of G protein-coupled receptors. Circulation
research. 2006;99(6):570-82. doi: 10.1161/01.RES.0000242563.47507.ce. PubMed PMID: 16973913.
18.
Lefkowitz RJ. G protein-coupled receptors. III. New roles for receptor kinases and beta-arrestins
in receptor signaling and desensitization. The Journal of biological chemistry. 1998;273(30):18677-80.
doi: 10.1074/jbc.273.30.18677. PubMed PMID: 9668034.
19.
Magalhaes AC, Dunn H, Ferguson SS. Regulation of GPCR activity, trafficking and localization by
GPCR-interacting proteins. British journal of pharmacology. 2012;165(6):1717-36. doi: 10.1111/j.1476-
5381.2011.01552.x. PubMed PMID: 21699508; PubMed Central PMCID: PMC3372825.
20.
Vasudevan NT, Mohan ML, Gupta MK, Hussain AK, Naga Prasad SV. Inhibition of protein
phosphatase 2A activity by PI3Kgamma regulates beta-adrenergic receptor function. Molecular cell.
2011;41(6):636-48. doi: 10.1016/j.molcel.2011.02.025. PubMed PMID: 21419339; PubMed Central
PMCID: PMC3063893.
21.
Noma T, Lemaire A, Naga Prasad SV, Barki-Harrington L, Tilley DG, Chen J, et al. Beta-arrestin-
mediated beta1-adrenergic receptor transactivation of the EGFR confers cardioprotection. The Journal
of clinical investigation. 2007;117(9):2445-58. doi: 10.1172/JCI31901. PubMed PMID: 17786238;
PubMed Central PMCID: PMC1952636.
22.
Nobles KN, Xiao K, Ahn S, Shukla AK, Lam CM, Rajagopal S, et al. Distinct phosphorylation sites
on the beta(2)-adrenergic receptor establish a barcode that encodes differential functions of beta-
arrestin. Science signaling. 2011;4(185):ra51. doi: 10.1126/scisignal.2001707. PubMed PMID: 21868357;
PubMed Central PMCID: PMC3415961.
23.
Wisler JW, DeWire SM, Whalen EJ, Violin JD, Drake MT, Ahn S, et al. A unique mechanism of
beta-blocker action: carvedilol stimulates beta-arrestin signaling. Proceedings of the National Academy
of Sciences of the United States of America. 2007;104(42):16657-62. doi: 10.1073/pnas.0707936104.
PubMed PMID: 17925438; PubMed Central PMCID: PMC2034221.
24.
Vasudevan NT, Mohan ML, Goswami SK, Naga Prasad SV. Regulation of beta-adrenergic receptor
function: an emphasis on receptor resensitization. Cell cycle. 2011;10(21):3684-91. doi:
10.4161/cc.10.21.18042. PubMed PMID: 22041711; PubMed Central PMCID: PMC3266006.
25.
Naga Prasad SV, Laporte SA, Chamberlain D, Caron MG, Barak L, Rockman HA. Phosphoinositide
3-kinase regulates beta2-adrenergic receptor endocytosis by AP-2 recruitment to the receptor/beta-
arrestin complex. The Journal of cell biology. 2002;158(3):563-75. doi: 10.1083/jcb.200202113. PubMed
PMID: 12163475; PubMed Central PMCID: PMC2173831.
26.
Krueger KM, Daaka Y, Pitcher JA, Lefkowitz RJ. The role of sequestration in G protein-coupled
receptor resensitization. Regulation of beta2-adrenergic receptor dephosphorylation by vesicular
acidification. The Journal of biological chemistry. 1997;272(1):5-8. doi: 10.1074/jbc.272.1.5. PubMed
PMID: 8995214.
27.
Asosingh K, Wanner N, Weiss K, Queisser K, Gebreab L, Kassa B, et al. Bone marrow
transplantation prevents right ventricle disease in the caveolin-1-deficient mouse model of pulmonary
hypertension. Blood Adv. 2017;1(9):526-34. doi: 10.1182/bloodadvances.2016002691. PubMed PMID:
29296972; PubMed Central PMCID: PMCPMC5728598 interests.
28.
Shen A, Chen D, Kaur M, Bartels P, Xu B, Shi Q, et al. beta-blockers augment L-type Ca(2+)
channel activity by targeting spatially restricted beta2AR signaling in neurons. eLife. 2019;8. doi:
10.7554/eLife.49464. PubMed PMID: 31609201; PubMed Central PMCID: PMC6813027.
29.
Stephens OR, Weiss K, Frimel M, Rose JA, Sun Y, Asosingh K, et al. Interdependence of hypoxia
and beta-adrenergic receptor signaling in pulmonary arterial hypertension. American journal of
physiology Lung cellular and molecular physiology. 2019;317(3):L369-L80. doi:
10.1152/ajplung.00015.2019. PubMed PMID: 31242023; PubMed Central PMCID: PMC6766716.
30.
Vasudevan NT, Mohan ML, Gupta MK, Martelli EE, Hussain AK, Qin Y, et al. Gbetagamma-
independent recruitment of G-protein coupled receptor kinase 2 drives tumor necrosis factor alpha-
induced cardiac beta-adrenergic receptor dysfunction. Circulation. 2013;128(4):377-87. doi:
10.1161/CIRCULATIONAHA.113.003183. PubMed PMID: 23785004; PubMed Central PMCID:
PMC3874808.
31.
Shenoy SK, Lefkowitz RJ. beta-Arrestin-mediated receptor trafficking and signal transduction.
Trends in pharmacological sciences. 2011;32(9):521-33. doi: 10.1016/j.tips.2011.05.002. PubMed PMID:
21680031; PubMed Central PMCID: PMC3159699.
32.
Naga Prasad SV, Barak LS, Rapacciuolo A, Caron MG, Rockman HA. Agonist-dependent
recruitment of phosphoinositide 3-kinase to the membrane by beta-adrenergic receptor kinase 1. A role
in receptor sequestration. The Journal of biological chemistry. 2001;276(22):18953-9. doi:
10.1074/jbc.M102376200. PubMed PMID: 11259422.
33.
Farha S, Saygin D, Park MM, Cheong HI, Asosingh K, Comhair SA, et al. Pulmonary arterial
hypertension treatment with carvedilol for heart failure: a randomized controlled trial. JCI insight.
2017;2(16). doi: 10.1172/jci.insight.95240. PubMed PMID: 28814664; PubMed Central PMCID:
PMC5621927.
34.
Leon-Velarde F, Bourin MC, Germack R, Mohammadi K, Crozatier B, Richalet JP. Differential
alterations in cardiac adrenergic signaling in chronic hypoxia or norepinephrine infusion. American
journal of physiology Regulatory, integrative and comparative physiology. 2001;280(1):R274-81. doi:
10.1152/ajpregu.2001.280.1.R274. PubMed PMID: 11124161.
35.
Matkovich SJ, Diwan A, Klanke JL, Hammer DJ, Marreez Y, Odley AM, et al. Cardiac-specific
ablation of G-protein receptor kinase 2 redefines its roles in heart development and beta-adrenergic
signaling. Circulation research. 2006;99(9):996-1003. doi: 10.1161/01.RES.0000247932.71270.2c.
PubMed PMID: 17008600.
36.
Erickson CE, Gul R, Blessing CP, Nguyen J, Liu T, Pulakat L, et al. The beta-blocker Nebivolol Is a
GRK/beta-arrestin biased agonist. PloS one. 2013;8(8):e71980. doi: 10.1371/journal.pone.0071980.
PubMed PMID: 23977191; PubMed Central PMCID: PMC3748024.
37.
Adachi H, Strauss W, Ochi H, Wagner HN, Jr. The effect of hypoxia on the regional distribution of
cardiac output in the dog. Circulation research. 1976;39(3):314-9. doi: 10.1161/01.res.39.3.314. PubMed
PMID: 954160.
38.
Calbet JA. Chronic hypoxia increases blood pressure and noradrenaline spillover in healthy
humans. The Journal of physiology. 2003;551(Pt 1):379-86. doi: 10.1113/jphysiol.2003.045112. PubMed
PMID: 12844510; PubMed Central PMCID: PMC2343162.
39.
Hansen J, Sander M. Sympathetic neural overactivity in healthy humans after prolonged
exposure to hypobaric hypoxia. The Journal of physiology. 2003;546(Pt 3):921-9. doi:
10.1113/jphysiol.2002.031765. PubMed PMID: 12563015; PubMed Central PMCID: PMC2342582.
40.
Tran TM, Friedman J, Baameur F, Knoll BJ, Moore RH, Clark RB. Characterization of beta2-
adrenergic receptor dephosphorylation: Comparison with the rate of resensitization. Molecular
pharmacology. 2007;71(1):47-60. doi: 10.1124/mol.106.028456. PubMed PMID: 17012621.
41.
Davuluri G, Welch N, Sekar J, Gangadhariah M, Alsabbagh Alchirazi K, Mohan ML, et al. Activated
protein phosphatase 2A disrupts nutrient sensing balance between mTORC1 and AMPK causing
sarcopenia in alcoholic liver disease. Hepatology. 2020. doi: 10.1002/hep.31524. PubMed PMID:
32799332.
42.
Naga Prasad SV, Jayatilleke A, Madamanchi A, Rockman HA. Protein kinase activity of
phosphoinositide 3-kinase regulates beta-adrenergic receptor endocytosis. Nature cell biology.
2005;7(8):785-96. doi: 10.1038/ncb1278. PubMed PMID: 16094730.
43.
Mohan ML, Jha BK, Gupta MK, Vasudevan NT, Martelli EE, Mosinski JD, et al. Phosphoinositide 3-
kinase gamma inhibits cardiac GSK-3 independently of Akt. Science signaling. 2013;6(259):ra4. doi:
10.1126/scisignal.2003308. PubMed PMID: 23354687; PubMed Central PMCID: PMC3967506.
Figure 1 - Hypoxia leads to increased phosphorylation of β2AR: A, Total cell lysates (80 μg)
from serum starved HEK 293 cells stably expressing FLAG-β2AR (β2AR-HEK 293 cells)
following 0, 3 and 6 hours (h) of hypoxia (2% oxygen) or normoxia were immunoblotted with
anti-phospho-β2AR antibody to assess β2AR phosphorylation (upper panel). The immunoblot
was stripped and re-probed with anti-HIF-1α antibody to determine expression of HIF-1α as a
molecular surrogate for hypoxia. The immunoblots were probed with anti-FLAG and anti-actin
antibody as loading controls. Cumulative densitometric analysis of five independent
experiments (n=5), *p< 0.01 vs. 0 and 3 hour time point. B, Confocal images of cell stained with
anti-phospho-β2AR antibody (green) after 6-hour normoxia or hypoxia treatment. Nucleus was
visualized by DAPI (blue) staining. Scale, 20 μm. C, Fluorescent intensity of phosphorylated
β2ARs/cell. Over 70-100 cells/experiment were used for fluorescent assessment, (n=4), *p<
0.05.
Figure 2 - Hypoxia mediates β2AR dysfunction through GRK2: A, Total cell lysates (80 μg)
were immunoblotted for ubiquitously expressed GRKs, GRK2, 3, 5 or 6 from β2AR-HEK 293
cells following 6 hours of normoxia or hypoxia. B, Summary densitometric analysis of GRK2
(n=5), *p<0.05 vs. 0 hour. C, β2AR-HEK 293 cells were pre-treated with GRK2 inhibitor
(paroxetine, 30µM) 45 minutes prior to 6 hours of hypoxia or normoxia treatment and total cell
lysates (80 µg) were immunoblotted with anti-phospho-β2AR antibody. D, cAMP levels were
measured in β2AR-HEK 293 cells following 6 hours of hypoxia treatment compared to
normoxia (n=5). *p<0.001 vs normoxia. E & F, In vitro isoproterenol (ISO)-stimulated
adenylyl cyclase activity was measured in the plasma membrane (E) and endosomal fractions (F)
extracted from β2AR-HEK 293 cells after 6 hours of hypoxia or normoxia. The data is presented
as fold change following in vitro ISO-stimulation/baseline (n=5). Plasma membranes (E)
*p<0.05 vs normoxia and (F) endosomes, *p<0.01 vs normoxia.
Figure 3 - Hypoxia mediated non-canonical endosomal accumulation of phosphorylated β2ARs:
A, Confocal microscopy was performed on β2AR-β-arrestin-2 GFP double stable HEK 293 cells
following 6 hours of hypoxia in serum free media. β-arrestin-2 GFP was visualized by GFP
(green) and phospho-β2AR by using anti-phospho- β2AR antibody (red). ISO (100 μM)
stimulation for 10 minutes was used a positive control to show β-arrestin-2 GFP recruitment to
plasma membrane. Nucleus was visualized by DAPI (blue) staining (n=3). Scale, 20 μm. B,
β2AR-HEK 293 cells were pre-treated with internalization blockers (0.45 M sucrose and 2% β-
cyclodextrin) and subjected to 6 hours of hypoxia or normoxia. Confocal microscopy was
performed to visualize phospho-β2AR using anti-phospho-β2AR antibody (green) following
hypoxia or 10 minutes of ISO treatment (used as positive control following 6 hours pre-treatment
with internalization blockers) (n=3). Nucleus was visualized by DAPI (blue) staining. Scale, 20
μm. C & D, [125]I-Cyanopindalol radio-ligand binding was performed on plasma membrane (C)
or endosomal fraction (D) isolated from β2AR-HEK 293 cells following 6 hours of normoxia or
hypoxia.
Figure 4 - Hypoxia inhibits β2AR resensitization by impairing β2AR-associated PP2A activity:
A, Plasma membrane (50 μg) or endosomal fractions (50 μg) from β2AR-HEK 293 cells were
immunoblotted with anti-phospho-β2AR antibody following 6 hours of normoxia or hypoxia.
Cumulative densitometry is shown as bar graphs (n=5). *p<0.05 vs normoxia. B, PI3Kγ was
immunoprecipitated from plasma membrane or endosomal fractions (80 μg) from β2AR-HEK
293 cells following 6 hours of normoxia or hypoxia. The immunoprecipitated beads were
washed and subjected to in vitro lipid kinase assay by providing phosphatidylinositol (PI) to
generate phosphatidylinositol mono-phosphate (PIP) (upper panel) (n=4). Summary
densitormetric data on generation of labeled PIP (lower panel). *p<0.05 vs normoxia. C, FLAG-
β2AR was immunoprecipitated (IP) from plasma membrane (50 μg) or endosomal fractions (50
μg) and associated PP2A activity was measured in the FLAG immunoprecipitates (n=6). *p<
0.05 vs. normoxia. D, Western immunoblotting was performed on 80 µg total lysates from
β2AR-HEK 293 cells following 6 hours of normoxia or hypoxia to detect PP2A, phospho-
I2PP2A and I2PP2A. Actin was used as loading control (n=4). Cumulative densitometry is
shown as bar graphs. *p<0.05 vs normoxia.
Figure 5 - Acute hypobaric hypoxia in mice leads to adverse cardiac remodeling associated with
βAR dysfunction: A, C57Bl6 mice were subjected to acute (20 hours) of hypobaric hypoxia. M-
mode echocardiography was performed pre- and post-hypoxia or normoxia treatments. Acute
hypoxia leads to larger ventricular chamber as assessed by echocardiography (n=12) (upper 4
panels). Lower panel shows cardiac functional parameters of % ejection fraction (%EF) (left
panel, *p<0.05 vs. normoxia) and % fractional shortening (%FS) (right panel, *p<0.05 vs.
normoxia) as measures of cardiac function. B, Heart weight (HW) and body weight (BW) were
measured for the mice at the termination the experiment post-hypoxia or normoxia to assess
HW/BW ratio as a measure of adverse cardiac remodeling (n=12). C, Heart sections from mice
subjected to normoxia or hypoxia were stained with H & E to assess cardiac remodeling (n=4).
H & E staining shows large ventricular lumen in mice subjected to hypoxia. Scale bar (3 mm).
D, Total cardiac lysates (100 μg) were immunoblotted with anti-phospo-β2AR antibody (upper
panel). The blots were stripped and re-probed with anti-HIF-1α a sentinel marker for hypoxic
response (middle). Actin was used as a loading control. E, Cumulative densitometry data (n=6)
for phospho-β2AR and HIF-1α is shown in the bar-graphs (left panel, *p<0.05 vs. phospho-
β2AR normoxia; right panel, *p<0.01 vs. HIF-1α normoxia). F, In vitro isoproterenol (ISO)-
stimulated adenylyl cyclase activity was measured in the cardiac plasma membranes isolated
from the hearts of mice subjected to 20 hours of normoxia or hypoxia (n=6). *p<0.01 vs basal:
#p< 0.05 vs. basal normoxia and ISO (normoxia or hypoxia).
Figure 6 - β-blocker reverses hypoxia-mediated β2AR phosphorylation: A, β2AR-HEK 293 cells
were pretreated (45 minutes) with β-blocker propranolol (10 μM) and then subjected to normoxia
or hypoxia for 6 hours. Total cell lysates (80 µg) were immunoblotted with anti-phospho-β2AR
antibody to assess β2AR phosphorylation. The blot was stripped and re-probed with anti-GRK2
antibody and actin was used as loading control. B, Cumulative data (n=5) for phospho-β2AR is
shown in the bar graph. *p<0.05 vs. respective vehicle (-propranolol); #p<0.05 vs. vehicle (-
propranolol) normoxia. C, Bar graph showing average fluorescent intensity/cell (n=4) (>100
cells/experiment). *p< 0.05 vs. respective vehicle (-propranolol); #p<0.05 vs. vehicle (-
propranolol) normoxia. D, Confocal microscopy was performed to visualize phosphorylated
β2ARs by using anti-phospho-β2AR antibody (green) and nucleus (blue) by DAPI after pre-
treatment with β-blocker propranolol (45 minutes) followed by 6 hours of hypoxia or normoxia.
E, FLAG-β2AR was immunoprecipitated (IP) from plasma membrane (50 μg) or endosomal
fractions (50 μg) following hypoxia alone or along with propranolol and associated PP2A
activity was measured in the FLAG immunoprecipitates (n=8). *p< 0.05 vs. endosomal hypoxia.
Figure 7 - Schematic illustration: Proposed model showing that hypoxia non-canonically
mediates β2AR dysfunction by selective upregulation of GRK2 that mediates receptor
phosphorylation and endosomal accumulation of phosphorylated β2ARs. Simultaneously,
hypoxia also impairs resensitization by inhibiting protein phosphatase 2A (PP2A) activity.
Hypoxia inhibits PP2A by activating PI3Kγ that phosphorylates endogenous inhibitor of PP2A
(I2PP2A) [20]. Phosphorylated-I2PP2A (phospho-I2PP2A) robustly binds to PP2A inhibiting
PP2A activity [20]. Thus, the inability of PP2A to dephosphorylate β2ARs leads to impairment
of resensitization and accumulation of phosphorylated in the endosomes. Surprisingly, β-blocker
treatment reverses β2AR phosphorylation in hypoxia preserving receptor function by potentially
reducing GRK2 levels and decreasing PI3Kγ activity normalizing PP2A that may now mediate
β2AR dephosphorylation despite hypoxia. These studies bring-to-fore yet to be appreciated role
of β-blockers in providing beneficial effects in hypoxia contrary to its currently understood role
in normoxia.
a ave? lava ODIO
a“
<
a
+
a
Fy
IB: HIF-10
IB: FLAG
IB: Actin
=} o <
4
ry °o 1189 sed
- o SuUdzUl JUBISaIN
Ss HH unoe/yyzd-d a '
=
ae
MW _ Normoxia Hypoxia
* (KD) — + — + Paroxetine
0 3 6 (h)
Plasma membrane Endosomes
n
nd
o
Adenyly!l cyclase
(ISO/Basal acivity)
o
0
Normoxia Hypoxia Normoxia Hypoxia
MW oo 3 6 (h)
(KD)
D
% CAMP/normoxia
m
Adenylyl cyclase
GRK2/actin
o
(ISO/Basal acivity)
S
a
Vv
o
=
o
°
0 3. 6 (h)
Plasma membrane
(Kt
75
50
MW oo 3 6 (h)
(KD)
MW Normoxia Hypoxia
(KD) — + — __+ Paroxetine
IB: p-B2AR
“_—— —| ‘er Actin
3
3
* 2
>
=
s
%
E
S
2
3 2 °
ad J
BIXOWOU/d WY? %
Endosomes
Normoxia Hypoxia
3 3 °
N =
(Ayatoe jeseg/Osi)
asejoAo |AjAuepy
Normoxia Hypoxia
Endosome
3 =
(urajosd Bwysjows) (utopoud Bwysjowy)
wsuep Yd wsuep UY”
Plasma membrane
c
£
a
o
2
£
Vg
Qa
BIXOWJON (os) eixodAyH PIXOWJON (os) erxodAy
RIXOWJON eIXOULION
B-arrestin
Normoxia
s&s
so
ED
J
Ss
z
Hypoxia
Ss
*
°
E
=
S
z
" (ISO) -
Hypoxia
“_ —S —l—LEE eee a
0.
membrane
3 Plasma membrane Endosome
IP: PI3Ky IP:PI3Ky
Assay: Lipid kinase activity Assay: Lipid kinase activity
_Normoxia_ _Hypoxia _Normoxia_ _Hypoxia
laereee [roever
3 ri
SE 1.0 SE 15
s- s 2 1.0
S Fos oF
es es°°
<0 <0
~ Normoxia Hypoxia ~ Normoxia Hypoxia
° Plasma membrane Endosome
IP: FLAG; Assay: PP2A activity IP: FLAG; Assay: PP2A activity
51.5 S
5 510
5‘° é *
30.5 305
zo £9
Normoxia Hypoxia Normoxia Hypoxia
D
wo Normoxia Hypoxia
7 | IB: p-I2PP2A g 1.0
NX
a
o
50—e wmmme| IB: Actin
COS e oes a Swe ee
Normoxia Hypoxia
xk
)
o-=-4nN
ouounsd
Fold Change
( Arbitrary Units
Normoxia Hypoxia
ni
=
o o o
- os
unovivZddzrd
Normoxia Hypoxia
PHA Stle TC ae
IP: FLAG; Assay: PP2A activity
Normoxia Hypoxia
wo Normoxia Hypoxia
==) IB: p-lI2PP2A
——— —— | IB: 12PP2A
50— eam | 1B: Actin
Plasma membrane
IP: FLAG; Assay: PP2A activity
VST WIMHUUeS LUNUVLaluIuyiaviy
Pre-Normoxia Post-Normoxia
—S ou
Normoxia Hypoxia
Pre-Hypoxia
HW/BW (mg/g)
oN fF OD
50 C Normoxia (H & E)
40 *
u 2 30
= 20
0 0
Normoxia Hypoxia Normoxia Hypoxia
Hypoxia (H & E)
%E!
Ny Do
sooo
*
=
°
D Normoxia Hypoxia
Mw
KD)
IB: p-B2AR
75:
#240
E se
* 7 =§,30
£15 £2.0 ge
3 S £220
$4.0 g c=
g =1.0 5310
20.5 wo Ss
a = SE 9
0+— 0 <= ~ Basal ISO Basal ISO
Normoxia Hypoxia Normoxia Hvpoxia
D Normoxia Hypoxia
MW
"min oan
IB: Actin
E *
IB: p-B2AR
~
a
IB: HIF-1c
h
!
Adenylyl cylase activity
ne poem, fk
in
oS
HIF-1c/actin
So
Normoxia Hypoxia Normoxia Hypoxia
- — Ot
2.0 # >
c 2
315 10
Ss =
@4 .
8 ° * 8 5
co
50: peti 20
Normoxia Hypoxia Nor
D Normoxia Hypoxia E
DAPI-Merge DAPI-Merge
|
- Fropranorot
a <. «a
Fold over/hypoxia
S
o
c
S
a
2
.
+
FA Normoxia Hypoxia
pay -— + — + Prop
IB: p-B2AK
+ Prop
#
Hypoxia
*
= +
a o 0 °
= =
90/Aj1suezU! UeDSe10N|4
7
Normoxia
it
*
— +Prop
Hypoxia
Hypoxia
DAPI-Merge
=
a
=
o
a
Fluorescent intensity/cell
o
— + — +Prop
Normoxia Hypoxia
E
IP: FLAG;
Assay: PP2A activity
WB Hypoxia
1B Hypoxia + Propranol
*
o= =
5
0
5
0
Fold over/hypoxia
Plasma _Endosome
membrane
B-blocker rescues BAR function
Hypoxia mediates SAR dysfunction
p-DIocKker rescues PAN TUNCHON
Hypoxia>
I+--<Ep>
LP « P|
Phosphorylation Dephosphorylation
4
<a
Phosphorylation Dephosphorylatic
GRKD)|
ee ee eee ee ee ee
| 2020 | beta-blocker reverses inhibition of beta-2 adrenergic receptor resensitization by hypoxia | 10.1101/2020.09.17.301903 | [
"Sun Yu",
"Gupta Manveen K.",
"Stenson Kate",
"Mohan Maradumane L.",
"Wanner Nicholas",
"Asosingh Kewal",
"Erzurum Serpil",
"Naga Prasad Sathyamangla V."
] | null |
1
Small Molecule Inducers of Neuroprotective miR-132 Identified by HTS-HTS in Human
1
iPSC-derived Neurons
2
Lien D. Nguyen1,†, Zhiyun Wei1,2,†,*, M. Catarina Silva3, Sergio Barberán-Soler4, Rosalia
3
Rabinovsky1, Christina R. Muratore1, Jonathan M. S. Stricker1, Colin Hortman4, Tracy L. Young-
4
Pearse1, Stephen J. Haggarty3, and Anna M. Krichevsky1,*
5
6
Affiliations
7
1. Department of Neurology, Brigham and Women’s Hospital and Harvard Medical School,
8
Boston, MA 02115, USA
9
2. Shanghai Key Laboratory of Maternal Fetal Medicine, Shanghai Institute of Maternal-Fetal
10
Medicine and Gynecologic Oncology, Shanghai First Maternity and Infant Hospital, School of
11
Medicine, Tongji University, Shanghai 200092, China
12
3. Chemical Neurobiology Laboratory, Center for Genomic Medicine, Department of Neurology,
13
Massachusetts General Hospital and Harvard Medical School, Boston, MA 02114, USA
14
4. RealSeq Biosciences, Santa Cruz, California, USA
15
16
17
18
† Lien D. Nguyen and Zhiyun Wei contributed equally to this work.
19
*Correspondence: akrichevsky@bwh.harvard.edu, zhiyun_wei@163.com
20
2
SUMMARY
21
MicroRNAs (miRNAs) are short RNAs that regulate fundamental biological processes. miR-132,
22
a key miRNA with established functions in Tau homeostasis and neuroprotection, is consistently
23
downregulated in Alzheimer’s disease (AD) and other tauopathies. miR-132 overexpression
24
rescues neurodegenerative phenotypes in several AD models. To complement research on
25
miRNA-mimicking oligonucleotides targeting the central nervous system, we developed a high-
26
throughput-screen coupled high-throughput-sequencing (HTS-HTS) in human induced pluripotent
27
stem cell (iPSC)-derived neurons to identify small molecule inducers of miR-132. We discovered
28
that cardiac glycosides, which are canonical sodium-potassium ATPase inhibitors, selectively
29
upregulated miR-132 in the sub-μM range. Coordinately, cardiac glycoside treatment
30
downregulated total and phosphorylated Tau in rodent and human neurons and protected against
31
toxicity by glutamate, N-methyl-D-aspartate, rotenone, and Aβ oligomers. In conclusion, we
32
identified small-molecule drugs that upregulated the neuroprotective miR-132 and ameliorated
33
neurodegenerative phenotypes. Our dataset also represents a comprehensive resource for
34
discovering small molecules that modulate specific miRNAs for therapeutic purposes.
35
36
Keywords: miRNAs, drug screen, iPSC-derived neurons, miR-132, ADRD, cardiac glycosides
37
3
INTRODUCTION
38
Despite the enormous burden of Alzheimer’s disease and related dementias (ADRDs) on patients,
39
caregivers, and society, there is still a lack of effective, disease-modifying treatments. Traditional
40
drug discovery has focused on disease-relevant proteins and peptides such as Aβ, Tau, and β-
41
amyloid cleaving enzyme 1. However, RNAs have recently emerged as promising targets for broad
42
disease categories, with several approved RNA therapeutics in the last five years 2. Particularly,
43
>70% of the human genome is transcribed into noncoding RNAs (ncRNAs) that play essential yet
44
largely understudied roles in biological processes 3, 4. MicroRNAs (miRNAs) are short, single-
45
stranded ncRNAs of 18–25 nucleotides that facilitate the degradation and inhibit the translation of
46
mRNA targets 5. Specific miRNAs have been shown to be dysregulated in various diseases 6,
47
making them valuable targets for both diagnostic and therapeutic purposes.
48
Here, we focus on miR-132, one of the most consistently downregulated miRNAs in the cortex
49
and hippocampus of ADRD patients 7-11. miR-132 deficiency promotes Aβ plaque deposits 12, 13,
50
and Tau accumulation, phosphorylation, and aggregation 13-16. We recently showed that miR-132
51
mimics protected mouse and human primary neurons against Aβ oligomer and glutamate toxicity
52
16. miR-132 viral overexpression reduced Tau toxicity and neuronal loss in presymptomatic PS19
53
mice overexpressing an autosomal dominant mutation (P301S) in the human microtubule-
54
associated protein tau (MAPT) transgene, likely through directly targeting Tau modifiers, including
55
glycogen synthase kinase 3 β (GSK3β), E1A binding protein P300 (EP300), RNA binding Fox-1
56
homolog 1 (RBFOX1), and calpain-2 (CAPN2) 16. Furthermore, miR-132 level was decreased in
57
the hippocampus of several AD mouse models, and viral overexpression of miR-132 rescued adult
58
hippocampal neurogenesis and memory deficits in these models 11, 12, 16, 17. These findings
59
4
collectively support upregulating miR-132 in the central nervous system (CNS) as a promising
60
approach for preventing or treating AD and tauopathies.
61
Two common approaches to upregulating miRNAs, oligonucleotide mimics and gene delivery,
62
have serious limitations. miRNA mimics, which are synthetic oligonucleotides imitating mature
63
miRNA sequences and structures, often have poor intracellular delivery and on-target activity,
64
must be heavily modified to avoid rapid degradation, and can induce immunotoxicity 18-20.
65
Similarly, delivering genes coding for miRNAs through viral or non-viral vectors is generally
66
inefficient and can induce immunotoxicity or off-target integration 19. The CNS presents additional
67
challenges for drug delivery and efficacy due to the blood-brain barrier that blocks the entrance of
68
most compounds. We proposed small molecules as an alternative approach for upregulating
69
miRNAs 21. Compared to miRNA mimics and gene therapy, small molecules usually have better
70
brain and cell penetrance. Small molecules already approved for treating human diseases have
71
well-established safety profiles and pharmacokinetics. Repurposing or improving these
72
compounds would accelerate the development of miRNA therapeutics to enter clinical trials.
73
However, only a few small molecules affecting miRNA levels have been described 22, and no
74
systematic effort has been made to identify such modulators of miRNA expression and activity.
75
Therefore, we designed a pipeline for discovering small molecules that upregulate miR-132 in
76
human induced pluripotent stem cell (iPSC)-derived excitatory neurons. To our knowledge, no
77
miRNome-wide high-throughput screen of small molecule modulators of miRNA, and particularly
78
miRNA inducers, has been developed to date.
79
We performed high-throughput-screen coupled high-throughput-sequencing (HTS-HTS) of ~1900
80
bioactive compounds in iPSC-derived human neurons and validated that several members of the
81
cardiac glycoside family, which are sodium-potassium (Na+/K+) ATPase pump inhibitors,
82
5
upregulated miR-132 in the sub-μM range. Treating rodent and human neurons with sub-μM
83
cardiac glycosides protected neurons against various toxic insults and downregulated Tau and
84
other miR-132 targets. Overall, we identified small-molecule compounds that upregulated the
85
neuroprotective miR-132 in neurons and provided a pipeline for discovering small-molecule
86
compounds that regulate miRNAs for therapeutic purposes.
87
88
RESULTS
89
Optimization of the high-throughput screen on human iPSC-derived neurons
90
We used human neurogenin 2 (NGN2)-driven iPSC-derived neurons (NGN2-iNs) as a
91
physiologically relevant cell-based screening platform to discover miR-132 inducers. iPSC lines
92
generated from donors were utilized for direct differentiation through NGN2 overexpression into
93
excitatory neurons based on established protocols (Figure 1A) 23. These cells closely mimic the
94
transcriptome and function of human neurons ex vivo and can be scaled and reproducibly employed
95
in multiple assays 23. Among 36 NGN2-iN lines obtained from the Religious Orders
96
Study/Memory and Aging Project (ROS-MAP) cohort, 25 lines from donors without cognitive
97
impairment were considered (S1A). The transcriptomes of these lines were previously profiled 23.
98
The BR43 line was selected for the screen based on its median expression of major miR-132
99
targets, including GSK3β, EP300, RBFOX1, CAPN2, FOXO3, TMEM106B, and MAPT (S1B).
100
Importantly, BR43 NGN2-iNs had the lowest variation of baseline miR-132 expression among the
101
replicate cultures and exhibited miR-132 upregulation by the known inducers BDNF and forskolin,
102
thus providing a reliable platform for the high-throughput screen (S1C).
103
6
Several steps of NGN2-iN culture and RNA collection were optimized for the high-throughput
104
screen (HTS) to maximize neuronal health, lysing efficiency, and RNA yield (S1D, E). The
105
protocol was tested for its compatibility with small RNA-seq using the RealSeq ultra-low input
106
system, long RNA RT-qPCR using the PrimeScript system, and small RNA RT-qPCR using the
107
miRCURY system (S1D, E), further supporting its application in diverse quantitative RNA-based
108
assays.
109
Small molecules screen of miR-132 inducers
110
Day 4 NGN2-iNs were plated onto 25 Matrigel-coated 96-well plates and differentiated into
111
neurons, as verified by NeuN and Tau expression (Figure 1A). On day 19, the Selleckchem library
112
(N=1,902 compounds), a diverse library of bioactive molecules, was pin-transferred into plates to
113
achieve 10 μM final concentration. DMSO (0.1% final concentration) and forskolin (10 μM) were
114
used as the negative and positive controls, respectively. NGN2-iNs were imaged to monitor
115
neuronal health 24h later, followed by direct lysis to release RNA (Figure 1A). Among all wells
116
with test compounds, 324 (17.0%) were excluded because of cell death, neurite degeneration, loss
117
of cells during washes, or enrichment of astrocytes. RNA lysates of the remaining wells, including
118
positive and negative controls, were used for RealSeq small RNA library preparation designed for
119
ultra-low input without RNA purification 24. RealSeq libraries from each set of four 96-well culture
120
plates were indexed with 384 multiplex barcodes and pooled for deep sequencing (Figure 1A).
121
After miRNA annotation, wells with less than 1,000 total annotated read counts were excluded
122
from further analysis (N=169, 10.7%). On average, 55,529 miRNA reads were counted per sample,
123
and 455, 240, 182, and 64 miRNA species per sample were detected with minimal read counts of
124
1, 5, 10, and 100, respectively (Figure 1B). Numerous neuron-enriched miRNAs, such as miR-
125
124, miR-26a, miR-128, miR-9, and miR-191, were abundant in DMSO-treated control NGN2-
126
7
iNs (Figure 1C). As expected, miR-132 was consistently detected and ranked among the 30 most
127
abundant miRNAs (Figure 1C). We further determined the top housekeeping neuronal miRNAs
128
by calculating the coefficient of variation (COV) for each miRNA within each batch of RNA-seq
129
and identified the miRNAs with the smallest COVs, including miR-103a/b, miR-107, and miR-
130
191 (Figure 1D). As library preparation and sequencing for different 384-well plates were carried
131
out on different days, to minimize batch effects, compounds within each 384-well plate were
132
ranked for miR-132 expression. Figure 1E showed the miR-132 waterfall plot for 221 compounds
133
in a 384-well plate.
134
Selection of compounds for further validation
135
To select compounds for further validation, we used a matrix with miR-132 plate rank as the
136
primary criterion and adjusted with secondary criteria, including the US Food and Drug
137
Administration (FDA) approval, BBB penetrance, clinical trials, published data on neuroprotective
138
effects, and effects on other miRNAs (Table S1). We treated DIV14 primary rat cortical neurons
139
and DIV21 human NGN2-iNs with 10 μM of 44 reordered compounds, and monitored miR-132
140
expression by RT-qPCR. 12 and 10 compounds significantly upregulated miR-132 in rat neurons
141
after 24h and 72h, respectively, and 4 compounds significantly upregulated miR-132 in NGN2-
142
iNs after 24h (Figure 2A, Table S2). Notably, the cardiac glycosides, ouabain and digoxin,
143
upregulated miR-132 in all conditions. Several chemotherapeutics, including rigosertib, pelitinib,
144
letrozole, XL888, and etoposide, appeared to mildly upregulate miR-132 but also caused toxicity
145
after 72h.
146
Dose-response of hit compounds
147
To investigate dose response, we selected forskolin as the positive control, digoxin, ouabain,
148
BIX02188, nitazoxanide as the hits, and pelitinib as a representative chemotherapeutic. We also
149
8
included 6 additional cardiac glycosides (digitoxin, oleandrin, bufalin, bufotalin, cinobufagin, and
150
proscillaridin A) and BIX02189, an analog of BIX02188. These compounds represent diverse
151
chemical groups and mechanisms of action (Figure 2B and Table S3). DIV14 primary rat cortical
152
neurons were treated with drugs at doses ranging from 1 nM to 100 μM for 24h. Remarkably, all
153
8 cardiac glycosides upregulated miR-132 2.5-3-fold in the sub-μM range, with proscillaridin A
154
having the lowest EC50 of 3.2 nM (Table S3). Other compounds also dose-dependently upregulated
155
miR-132 but with higher EC50. For all compounds tested, miR-212, which is a miRNA co-
156
transcribed and co-functional with miR-132 25, was similarly upregulated at almost identical EC50,
157
suggesting that the mechanism was largely transcriptional (S2A and Table S3). The cardiac
158
glycosides proscillaridin A, oleandrin, digoxin, ouabain, and bufalin also upregulated miR-132
159
and miR-212 in a dose-dependent manner in human NGN2-iNs in the sub-μM range (S2B, C and
160
Table S3). However, BIX02188, which robustly upregulated miR-132 in primary rat neurons, had
161
no effect on miR-132 in NGN2-iNs (S2C and Table S3), suggesting potential differences between
162
the two cell models.
163
Upregulation of miR-132/212 was specific and transcriptional
164
To investigate the specificity of miR-132 upregulation, we measured the expression level of 10
165
other abundant neuronal miRNAs in rat primary cortical neurons after 24h of treatment with
166
oleandrin and BIX02188. When normalized to the geometric mean of all 12 miRNAs 26, only miR-
167
132 and miR-212 were upregulated (Figure 3A). The precursors pre-miR-132 and pre-miR-212
168
(Figure 3B) were also upregulated by forskolin, BIX02118, and the cardiac glycosides, suggesting
169
that these compounds activated the transcription of the miR-132/212 locus. Correspondingly, the
170
upregulation of miR-132 by forskolin and oleandrin was completely blocked by pretreatment with
171
the transcription inhibitor actinomycin D (Figure 3C, S3). As miR-132/212 locus is regulated by
172
9
the transcription factor CREB 27, cells were also pretreated with the maximum tolerated doses of
173
a CREB inhibitor (1 μM CREB-I). Co-treatment with the CREB inhibitor partially attenuated miR-
174
132 upregulation by forskolin and oleandrin (Figure 3C). As cardiac glycosides are conventional
175
inhibitors of Na+/K+ pumps, we also knocked down ATP1A1 and ATP1A3, the dominant isoforms
176
in neurons, with siRNAs. Knocking down either ATP1A1 or ATP1A3 also increased the
177
expression of products of the miR-132/212 locus (Figure 3D), suggesting that cardiac glycosides
178
upregulated miR-132 by inhibiting their conventional targets.
179
Kinetics of miR-132 upregulation and effects on known targets
180
To investigate the kinetics of miR-132 upregulation by cardiac glycosides, we treated primary rat
181
cortical neurons with 100 nM oleandrin and measured the expression of the precursor and the
182
mature forms of miR-132 and miR-212 overtime (Figure 4A, B). Both pre-miR-132 and pre-miR
183
-212 were rapidly upregulated following treatment, peaked at 8h, and rapidly declined back to
184
baseline after 72h (Figure 4A). As oleandrin was shown to upregulate BDNF 28, 29, a known
185
transcriptional regulator of the miR-132/212 locus 27, we also measured BDNF expression level.
186
Oleandrin upregulated BDNF as expected but at a slower kinetics than pre-miR-132/212,
187
suggesting that BDNF was not mediating the effects of oleandrin on miR-132/212 expression
188
(Figire 4A). Compared to their precursors, mature miR-132 and -212 were upregulated at slower
189
kinetics, peaked at 24h, then slowly declined but were still ~2-fold above baseline at 72h (Figure
190
4B).
191
We hypothesized that the increase in miR-132 expression would lead to the downregulation of its
192
targets. Indeed, we observed a time-dependent downregulation of MAPT, FOXO3a, and EP300
193
mRNAs that matched the upregulation of miR-132 (Figure 4C). mRNA targets were significantly
194
reduced to ~50% of baseline at 24h and to ~75% of baseline at 72h (Figure 4C), which was similar
195
10
to the observed effects for miR-132 mimics 72h after transfection (S4A-C). Tau, pTau S202/T305
196
(AT8), pTau S396, and FOXO3a proteins were also downregulated, though the ratio of pTau: total
197
Tau was unchanged (Figure 4D-H). In primary Tau wild-type (WT) and PS19 mouse neurons that
198
overexpress human mutant Tau-P301S 30, 100 nM oleandrin upregulated miR-132 and
199
downregulated both mouse MAPT and human MAPT after 72h treatment (S5D-G).
200
Cardiac glycosides protected mature neurons against glutamate and Aβ toxicity
201
We hypothesized that the upregulation of miR-132 by the cardiac glycosides would be
202
neuroprotective 16. As several studies have reported possible neurotoxic effects associated with
203
cardiac glycosides 31, 32, we first treated rat neurons at different ages in vitro (DIVs 7/14/21/28)
204
with digoxin, oleandrin, and proscillaridin A for 96h before measuring cellular viability (Figure
205
5A). Interestingly, DIV7 neurons were highly susceptible to cardiac glycoside toxicity, with
206
significant loss of viability observed at the miR-132 EC100 for all compounds tested (Figure 5B-
207
D). However, mature neurons were more resistant to cardiac glycoside toxicity, and no loss of
208
viability was observed at miR-132 EC100 for neurons treated at DIV14 or later (Figure 5B-D).
209
To investigate neuroprotective effects, we first treated DIV21 rat neurons with oleandrin and
210
proscillaridin A at EC100 for 24h, followed by 100 μM glutamate or 10 µM Aβ42 oligomers (Figure
211
5E). Pretreatment with proscillaridin A and oleandrin rescued neuronal viability loss due to toxic
212
insults without affecting viability at baseline (Figure 5F, G). As we previously showed that miR-
213
132 mimics rescued loss of viability in younger neurons treated with glutamate 16, we performed
214
similar experiments in DIV7 neurons (Figure 5H). We observed a small loss of viability due to
215
proscillaridin A at baseline (Figure 5I). However, both oleandrin and proscillaridin A rescued loss
216
of viability caused by glutamate excitotoxicity (Figure 5I). Oleandrin and proscillaridin A also
217
11
partially and dose-dependently rescued neurite loss induced by glutamate without affecting neurite
218
at baseline (Figure 5J, S5).
219
Cardiac glycosides significantly reduced Tau and pTau in human iPSC-neurons
220
To investigate the effects of cardiac glycosides in human neurons, we utilized two additional iPSC-
221
derived neural progenitor cell (NPC) lines: MGH-2046-RC1 derived from an individual with FTD
222
carrying the autosomal dominant mutation Tau-P301L (referred here as P301L), and MGH-2069-
223
RC1 derived from a healthy individual directly related to MGH-2046 (referred here as WT). These
224
NPCs, when differentiated into neurons (iPSC-neurons) for 6-8 weeks, represent well-established
225
models for studying tauopathy phenotypes in patient-specific neuronal cells relative to a WT
226
control 33-35.
227
Since Tau metabolism is regulated by miR-132 14, and Tau lowering is a promising therapeutic
228
strategy for ADRD 36, we first investigated the effects of cardiac glycosides on Tau protein levels.
229
All three tested cardiac glycosides strongly and dose-dependently downregulated Tau, as
230
exemplified by proscillaridin A. The treatment led to a clear reduction in total Tau (TAU5
231
antibody) and pTau S396 in WT neurons (Figure 6A) and in P301L mutant neurons after 24h and
232
72h (Figure 6D). For total Tau (TAU5), the upper band (>50 kDa, monomeric Tau + post-
233
translational modifications (PTMs)) was more intense at lower drug concentrations. With
234
increasing concentrations, the upper band disappeared, whereas the lower band (<50kDa, possibly
235
non-pTau) became slightly more intense. This downward band shift suggested that proscillaridin
236
A reduced both Tau accumulation and altered PTMs. Consistent with the latter, proscillaridin A
237
reduced the monomeric form of pTau S396 (~50 kDa) as well as the high molecular weight
238
oligomeric pTau (≥250 kDa, Figure 6D).
239
12
RT-qPCR was performed on a matched set of WT and P301L iPSC-neurons and showed a dose-
240
dependent reduction in MAPT mRNA, a large increase in pre-miR-132, and a more modest
241
increase in mature miR-132 (Figure 6 B, C, E, F). Similar results were obtained with digoxin and
242
oleandrin treatments (S6). Further immunoblot results showed that in P301L iPSC-neurons, 72 h
243
treatment with 1 µM proscillaridin A, digoxin, or oleandrin treatment reduced both soluble and
244
insoluble total Tau and pTau S396 (S7A-C). 72h treatment also resulted in a dose-dependent
245
reduction in miR-132 targets at the protein levels, including FOXO3a, EP300, GSK3β, and
246
RBFOX1 (S7D-O).
247
For all compounds, the concentration of 10 µM was associated with >70% reduction in Tau and
248
pTau with 24h and 72h treatments. However, this concentration also reduced neuronal synaptic
249
markers, including post-synaptic density protein 95 (PSD95), synapsin 1 (SYN1), and β-III-tubulin
250
representative of microtubules’ structural integrity. These results suggest that at high
251
concentrations and with prolonged exposure, cardiac glycosides can compromise neuronal
252
integrity. Nevertheless, for each drug, we observed a significant safety window in which Tau
253
lowering was not associated with reduced synaptic or microtubule markers (Figure 6G-R). In all
254
graphs, the yellow shade indicates the dose range where the loss of at least 2 synaptic markers was
255
30% or less (Figure 6G-R). Interestingly, WT neurons appeared more susceptible to loss of
256
synaptic markers upon treatment than P301L neurons, particularly at 72h. For example,
257
proscillaridin A was much less toxic to P301L neurons than WT neurons (Figure 6O, P, Q, R).
258
Cardiac glycosides were neuroprotective in NPC-derived neuronal models of tauopathy
259
To examine the effects of the cardiac glycosides on neuronal viability, WT and P301L iPSC-
260
neurons were treated with various doses of digoxin, oleandrin, and proscillaridin A for 24h or 72h.
261
A dose-dependent loss of viability was observed with all three compounds, particularly at 72h. In
262
13
Tau-WT neurons, there was up to 30% loss of viability after 72h treatment, particularly at the
263
highest dose of 10 μM (Figure 7A-C). Interestingly, in Tau-P301L neurons, the toxicity observed
264
was minimal, with <10% viability loss at the highest concentrations at 72h (Figure 7 D-F). These
265
results were consistent with the previous immunoblot data (Figure 6G-R), showing that P301L
266
neurons were more resistant to cardiac glycoside toxicity than WT neurons.
267
We next tested whether cardiac glycosides can protect human neurons from various cell stressors
268
that specifically affect human iPSC-neurons expressing mutant Tau 35. These include the
269
excitotoxic agonist of glutamatergic receptors NMDA, an inhibitor of the mitochondrial electron
270
transport chain complex I, rotenone, and the aggregation-prone Aβ (1–42) amyloid peptide. P301L
271
neurons differentiated for 8 weeks were pretreated with cardiac glycosides for 6h prior to the
272
addition of stressors for 18h, and viability was measured at the 24h time point (Fig.7g). Cardiac
273
glycosides were added at the concentrations of 1 μM and 5 μM, which did not affect cell viability
274
in P301L neurons at 24h (Figure 7 D-F). All cardiac glycosides significantly rescued neuronal
275
viability in the presence of stressors (Figure 7H-J). The rescue could also be observed with
276
immunofluorescent staining (Figure 7K). At baseline, 1 μM of digoxin, oleandrin, or proscillaridin
277
A reduced Tau staining in agreement with the immunoblot data (Figure 6D) without visibly
278
affecting neuronal health. Treatment with the stressors led to a significant loss of neurites and cell
279
bodies in neurons pretreated with vehicle alone, which was rescued by pretreatment with the
280
cardiac glycosides. Overall, these results demonstrate that low concentrations of cardiac glycosides
281
were neuroprotective in human tauopathy neurons.
282
Transcriptome analysis of human iPSC-neurons confirmed shared pathways affected by
283
cardiac glycosides
284
14
To uncover the molecular effects of cardiac glycosides on neuroprotection beyond miR-132 and
285
its canonical targets, we profiled transcriptomes of human iPSC-neurons after 72h of treatment
286
with increasing doses of digoxin, oleandrin, proscillaridin A or vehicle alone (0.1% DMSO) using
287
RNA sequencing (Figure 8A). Starting from low doses, cardiac glycosides remarkably changed
288
the global transcriptome of P301L neurons as seen in principal component analysis (PCA, Figure
289
8B), with single principal component (PC1) being able to clearly separate controls from treatments.
290
More importantly, three different cardiac glycosides regulated transcriptomes similarly and in a
291
prominent dose-dependent manner (Figure 8B). Differential expression analysis identified
292
thousands of genes significantly regulated with fold-change higher than 4, even though the
293
statistical power was weakened by the intrinsic variance of dosage gradient (Figure 8C). Many
294
genes were related to neuronal health and activity, including the strongly upregulated ARC which
295
encapsulates RNA to mediate various forms of synaptic plasticity 37, 38, and downregulated MAPT
296
and the SLITRK3/4/6 family which plays a role in suppressing neurite outgrowth 39. We further
297
focused on the biological pathways that were commonly regulated by all three cardiac glycoside
298
compounds. Notably, these treatments affected a substantial number of shared pathways (Figure
299
8D). Many downregulated genes belong to 74 pathways related to neuronal development,
300
morphology, health, or activity (Figure 8E). Upregulated genes were highly enriched in positive
301
regulators of transcription, negative regulators of programmed cell death, and regulators of stress
302
and the unfolded protein response (Figure 8F). Furthermore, dozens of transcription factors that
303
had binding sites on MIR132 promoter and may upregulate its expression, including CREB5, were
304
commonly upregulated by cardiac glycosides (Figure 8G). The neuroprotective BDNF signaling
305
pathway was significantly upregulated (S8), corroborating our previous observation that cardiac
306
glycosides upregulated BDNF in rat neurons (Figure 4A) . Therefore, while digoxin, oleandrin,
307
15
and proscillaridin A all induced miR-132 expression, they likely also regulated multiple other
308
pathways. Overall, shared transcriptomic alterations and regulated pathways further confirmed the
309
common molecular mechanisms of action of cardiac glycosides and their ability to activate stress-
310
protective programs in highly vulnerable Tau-mutant neurons (Figure 8H).
311
312
DISCUSSION
313
As miRNAs have been increasingly recognized as master regulators of many biological processes
314
and promising therapeutic targets, screens for miRNA modulators have recently emerged. Several
315
studies have reported successful screens for small molecules that inhibit the activity of oncogenic
316
or pathogenic miRNAs, including miR-21 40, 41, miR-122 42, and miR-96 43. Small-molecule
317
inhibitors of miRNAs can be chemically modified to improve pharmacological properties and
318
efficient CNS delivery, even though with potentially inferior target specificity relative to miRNA
319
antisense oligonucleotides. On the other hand, miRNA mimic oligonucleotides require chemical
320
modifications for stabilization and durable activity in vivo, which may reduce overall potency in
321
the simultaneous regulation of multiple downstream targets. Therefore, small-molecule inducers
322
of specific miRNAs could provide additional advantages as therapeutics. To date, no miRNA
323
inhibitor or mimic oligonucleotide therapeutics have been FDA-approved, very few reporter-based
324
screens have been published, and no systematic screens relying on broader miRNome-level
325
readouts have been performed for small-molecule miRNA modulators 22.
326
Most HTSs for modulators of gene expression employ cell lines as screening platforms and gene-
327
specific heterologous reporter systems as primary assays. However, proliferating, immortalized
328
cells have limited value for identifying neuroprotective agents, and neurons are known to be
329
technically difficult to transfect efficiently and uniformly, especially on a large scale 21. Here, we
330
16
applied HTS with miRNA-seq to directly quantify expression levels of hundreds of miRNAs in
331
human neurons treated with small molecule compounds. Notably, the present study is the first
332
HTS-HTS for small RNAs that was enabled by the low-input requirement of RealSeq technology
333
24, although HTS-HTS for mRNA has been conducted previously 44-46. Despite the relatively small
334
scale of ~1900 compounds, we successfully validated 4 different classes of drugs that upregulate
335
miR-132, most notably the cardiac glycosides family. As the first small molecule screen for
336
neuronal miRNA modulators, the obtained dataset can be reanalyzed to identify compounds that
337
regulate any of the ~450 miRNAs, providing a unique new resource (Table S4) and facilitating
338
further discoveries of miRNA-targeting drugs.
339
In this study, we focus on miR-132, a master neuroprotector. Several members of the cardiac
340
glycoside family, Na+/K+ ATPase pump inhibitors, were successfully validated to upregulate
341
miR-132/212 consistently. Of note, cardiac glycosides such as digoxin and digitoxin are widely
342
used for treating congestive heart failure and cardiac arrhythmias. However, they have a narrow
343
therapeutic index and can be toxic at high doses47. Recent studies have reported that cardiac
344
glycosides are neuroprotective in animal models at low (sub-μM to μM) concentrations in stroke
345
48, 49, traumatic brain injury 50, systemic inflammation 51, and AD and tauopathies 52, 53.
346
Furthermore, clinical studies suggest that treatment with digoxin might improve cognition in older
347
patients with or without heart failure 54. Our data supported that the cardiac glycosides reduced
348
Tau accumulation and rescued Tau-mediated toxicity. Further work remains to be done to
349
investigate if any member of the cardiac glycosides can be developed into effective and safe
350
therapeutics for long-term treatment against neurodegenerative diseases. Oleandrin, which was
351
previously shown to be neuroprotective with excellent brain penetrance and retention 49, 55, and
352
proscillaridin A, which exhibited the lowest EC50 in rat and human neurons, may be good starting
353
17
points. Furthermore, as the expression of ATP1A3 is restricted to neurons, whereas ATP1A1 and
354
ATP1A2 are more ubiquitously expressed 56, drugs that selectively target the ATP1A3 isoform
355
may alleviate the systemic impact of the cardiac glycosides such as on the cardiac system.
356
Several questions that emerge from our observations are worth investigating further. First, there is
357
a significant difference in the fold change of mature and pre-miR132. Pre-miR-132 was
358
upregulated by 10 to 30-fold, whereas mature miR-132 in the same treatment group was
359
upregulated by only 1.5-3-fold (Figures 4, 6). We hypothesize that there may be physiological
360
mechanisms that maintain the levels of mature miR-132 within a 2-fold difference, perhaps a
361
bottleneck in processing pre-miR-132 to mature miR-132. Interestingly, miR-132 is
362
downregulated by ~1.5-2.5-fold in various neurodegenerative diseases 7, suggesting that the
363
increase promoted by treatment with cardiac glycosides is sufficient to restore physiological miR-
364
132 levels. Second, the effect of cardiac glycosides on downregulating human MAPT mRNA and
365
Tau protein appears to be much stronger than for rodent Tau. After 72h treatment, 100 nM
366
oleandrin downregulated rat MAPT mRNA by 27% and rat Tau protein by 35% (Fig. 4). The same
367
treatment downregulated human MAPT mRNA by 87% and human Tau protein by 59% in mutant
368
P301L neurons (Figure 6 and S6). Some differences may be attributed to differences in sensitivity
369
to cardiac glycosides between rodents and humans, as mouse ATP1A1 is inhibited by ouabain and
370
digitoxin at >100-fold higher concentration than human ATP1A157. However, in PS19 mouse
371
neurons, which express ~8-fold more human MAPT than endogenous mouse MAPT mRNA,
372
oleandrin downregulated both mouse and human MAPT by ~40-50% (S4), suggesting that cardiac
373
glycosides and miR-132 may target human MAPT more effectively. Third, several studies have
374
proposed that cardiac glycosides downregulate MAPT and Tau and provide neuroprotection
375
through other pathways, including increased autophagy 52, alternative splicing of MAPT mRNA58,
376
18
increased BDNF 28, and inhibiting reactive astrocytes 53. Our transcriptomic results support that
377
many neuronal pathways are altered, suggesting that cardiac glycosides can modulate multiple
378
pathways that converge on the downregulation of Tau and increased neuroprotection. While
379
further investigation is needed to determine the contribution of miR-132 upregulation to Tau
380
downregulation and neuroprotection, cardiac glycosides emerge as promising therapeutics for
381
neurological disorders, if they can be improved to reduce systemic toxicity and enhance brain
382
penetrance and retention.
383
In summary, our pilot HTS-HTS of miRNA regulators on human neurons discovered the cardiac
384
glycoside family as novel miR-132 inducers. These compounds specifically upregulated miR-132
385
expression via transcription activation by inhibiting the Na+/K+ ATPases and could protect rat
386
primary neurons and a human iPSC-derived neuronal model of tauopathy against diverse insults,
387
including glutamate, Aβ oligomers, NMDA, and rotenone. Our pilot study not only highlights
388
cardiac glycosides as promising treatments for neurodegenerative diseases but also provides a key
389
omics resource for future neuronal miRNA regulator discoveries.
390
391
Acknowledgments
392
This work was supported by the R56 AG069127 and the Rainwater Foundation/ Tau Consortium
393
grants to A.M.K. S.J.H., and M.C.S. were supported by Rainwater Foundation/ Tau Consortium
394
funding. T.L.Y.P., C.R.M., and J.M.S.S. were supported by R01AG055909. The BWH iPSC
395
NeuroHub provided support for NGN2-iNs related work. The ICCB-Longwood Screening Facility
396
provided the compounds and instruments for performing the high-throughput drug treatment. The
397
NeuroTechnology Studio at Brigham and Women’s Hospital provided IncuCyte instrument access
398
and consultation on data acquisition and data analysis. Dr. Bradford Dickerson (MGH), Dr. James
399
19
Gusella (MGH), Diane Lucente (MGH), and Dr. Bruce Miller (UCSF) are thanked for the generous
400
sharing of patient cell lines. Dr. Michelle Arkin (UCSF), Dr. Erik Uhlmann (DFCI), and Dr.
401
Evgeny Deforzh (BWH) are thanked for their helpful discussion, comments, and edits. Ramil
402
Arora and Harini Saravanan are thanked for annotating the compounds, as shown in Table S1. Dr.
403
Rachid El Fatimy is thanked for the preparation of Ab oligomers. PubChem Sketcher was used to
404
prepare the chemical structures in Supp. Table BioRender was used in the preparation of Figures
405
1, 8, and S8.
406
407
Contributions
408
A.M.K., Z.W., and L.D.N. conceived and designed the study. Z.W. and L.D.N. equally contributed
409
as first authors. Z.W., L.D.N., M.C.S., S.B.S., R.R., C.R.M., J.M.S.S., and C.H. performed
410
experiments for this study. T.L.Y.P., S.J.H., and A.M.K. provided the resources needed for
411
experiments. L.D.N. wrote an original draft of the manuscript. A.M.K, Z.W., and L.D.N. reviewed
412
and edited the manuscript. All authors reviewed and commented on the manuscript.
413
414
Corresponding authors
415
Correspondence to Zhiyun Wei or Anna M. Krichevsky.
416
417
Competing interests
418
S.B.S. and C.H. are employees of RealSeq Biosciences, which performed the RealSeq miRNA-
419
seq. S.J.H. is a consultant/member of the scientific advisory board for Psy Therapeutics, Frequency
420
20
Therapeutics, Vesigen Therapeutics, 4M Therapeutics, Souvien Therapeutics, Proximity
421
Therapeutics, and Sensorium Therapeutics, none of which were involved in the present study.
422
Other authors have no competing interests to declare.
423
424
Availability of data and materials
425
miRNA-sequencing and mRNA-sequencing data that support the findings of this study will be
426
deposited into Sequence Read Archive with accession number to be determined. Contact
427
corresponding authors for requests of materials and cell lines used in the manscript.
428
429
21
FIGURE AND FIGURE LEGENDS
430
22
Figure 1. Experimental workflow and overview of screen results. a, NGN2-iN generation, drug
431
treatment, and miRNA-seq workflow (N=1 per drug). b, Average number of miRNA species
432
detected per sample by miRNA-seq at various count cut-offs. c, Expression levels of the 100 most
433
abundant miRNAs in vehicle-treated samples. miR-26a-5p was the most abundant miRNA
434
detected, and miR-132-3p was the 27th. d, Shared miRNAs with the lowest coefficient of variation
435
among the 7 plates tested. e, Waterfall plot for miR-132 expression in plate 2. Samples treated
436
with ouabain, digoxin, and the positive control forskolin showed the highest level of miR-132.
437
23
438
Figure 2. Validation of top candidates from HTS-HTS and dose curve experiment. a, Drugs
439
that showed significant upregulation of miR-132 in primary rat cortical neurons after 24 and 72h
440
treatment and human NGN2-iNs after 24h treatment (RT-qPCR analysis, N=4, unpaired two-tailed
441
Student’s t-test, p<0.05 compared to DMSO). b, Dose curve experiments were performed in
442
DIV14 rat neurons after 24h treatment. Solid lines were used for cardiac glycosides, and dotted
443
lines were used for other drugs. EC50 and max fold change were calculated using sigmoidal fit, 4
444
parameters. (N=4-6, error bars represent SD).
445
24
446
25
Figure 3. Cardiac glycosides transcriptionally upregulated miR-132/212 by inhibiting the
447
Na+/K+ ATPases. a, Forskolin, oleandrin, and BIX02188 specifically upregulated miR-132/212
448
without affecting other abundant miRNAs. Expression was normalized to the geometric mean of
449
all 12 miRNAs tested. b, Cardiac glycosides, forskolin, and BIX02188 also upregulated the
450
precursors of miR-132/212 24h after treatment (unpaired two-tailed Student’s t-test compared to
451
DMSO control, N=4). c, Upregulation of miR-132 by forskolin or oleandrin was completely
452
blocked by the transcription inhibitor actinomycin D and partially blocked by CREB inhibitor
453
(unpaired two-tailed Student’s t-test compared to DMSO control, N=8-19). d, Knocking down
454
ATP1A1 or ATP1A3, the predominant isoforms in neurons, also upregulated pre- and mature miR-
455
132 (unpaired two-tailed Student’s t-test compared to DMSO control, N=4-6).
456
457
26
458
Figure 4. Oleandrin upregulated miR-132 and downregulated its targets over time. a-c, 100
459
nM oleandrin upregulated pre- and mature miR-132/212 and downregulated their mRNA targets
460
over time (RT-qPCR analysis). d-h, Oleandrin downregulated total Tau, pTau (AT8 and S396),
461
and FOXO3a protein after 72h treatment (Western blot analysis, unpaired two-tailed Student’s t-
462
test, N= 4-8, error bars represent SD).
463
27
464
28
Figure 5. Cardiac glycosides rescued glutamate and Aβ oligomer-induced toxicity in primary
465
rat neurons. a-d, Younger neurons were more susceptible to cardiac glycoside toxicity, whereas
466
more mature neurons were resistant. Primary rat neurons were treated with various doses of
467
digoxin, oleandrin, and proscillaridin A for 96h before viability was measured using WST-1. Cells
468
treated at DIV7 showed a dose-dependent reduction in viability. In contrast, cells treated at DIV14,
469
21, or 28 showed little loss of viability, particularly at EC100 for miR-132 upregulation (unpaired
470
t-test comparing to DMSO condition for each dose, N=4-8 per dose, error bars represent SD.). e-
471
g, For DIV21 neurons, proscillaridin A and oleandrin were not toxic at baseline and fully rescued
472
viability loss due to glutamate or Aβ oligomer treatment (2-way ANOVA, followed by Šídák’s
473
multiple comparisons test, N=8-16 per condition, error bars represent SD). h-j, Proscillaridin A
474
was mildly toxic to DIV7 neurons at baseline. However, both proscillaridin A and oleandrin fully
475
rescued viability loss and partially rescued neurite loss due to glutamate treatment (2-way
476
ANOVA, followed by Šídák’s multiple comparisons test, N=8-16 per condition, error bars
477
represent SD).
478
479
29
480
30
Figure 6. Dose-dependent reduction in Tau in human iPSC-neurons treated with cardiac
481
glycosides. WT and P301L neurons were differentiated for 6 weeks, then treated with cardiac
482
glycosides for 24h or 72h. a, Representative western blot for WT neurons treated with
483
proscillaridin A (ProsA). A dose-dependent reduction in total Tau and p-Tau S396 was observed
484
at both 24h and 72h. b-c, In parallel, a reduction in MAPT mRNA and an increase in pre-miR-132
485
and miR-132 RNA were observed. d-f, Similar results were also observed in Tau P301L neurons
486
by western blot (d) and mRNA (e, f) analysis. g-r, Western blot densitometry quantification of
487
dose-dependent effects on Tau (TAU5), pTau S396, and the synaptic makers PSD95 and SYN1 in
488
WT and P301L neurons treated for 24h or 72h. The yellow shades indicate compound
489
concentrations leading to <30% loss of at least two synaptic/microtubule markers (N=1-2, error
490
bars represent SEM, the dotted lines indicated that separate Western blots were put together).
491
31
492
32
Figure 7. Cardiac glycosides rescued Tau-P301L neuronal vulnerability to stress. a-f,
493
Compounds concentration effect on neuronal viability after 24h or 72h treatment of WT (a-c) and
494
P301L (d-f) neurons. Data points indicate mean ±SD (N=2); unpaired two-tailed Student’s t-test.
495
g, Schematic of the assay used to measure neuroprotective effects by cardiac glycosides in
496
tauopathy neurons. h-j, Cardiac glycosides rescued the loss of viability in P301L neurons due to
497
NMDA, rotenone, or Aβ42 oligomer treatment. Graph bars and data points show mean values
498
±SEM (N=2); unpaired two-tailed Student’s t-test. k, Representative images for P301L neurons at
499
8 weeks of differentiation treated with cardiac glycosides and each stressor compound. Total Tau
500
(K9JA antibody) staining is shown in red, and MAP2 is shown in green. Scale bars are 200 μm.
501
33
502
34
Figure 8. Transcriptome analysis of human iPSC-neurons treated with cardiac glycosides. a,
503
Workflow of the experiment design. b, Principal component analysis (PCA) indicated the strong
504
and dose-dependent alteration of global transcriptomic profiles after treatments. c, Volcano plots
505
showed significant down- and up-regulated genes, labeled in blue and red dots, respectively. Stars
506
highlighted dysregulated genes involved in neuronal activity and health. d, Venn diagram
507
indicated the similarity of pathways affected by three cardiac glycosides. e, Selected neuronal
508
pathways highlighted in common pathways of down-regulated genes. f, Selected transcription- and
509
response-related pathways highlighted in common pathways of up-regulated genes. g, Effects of
510
cardiac glycosides on the expression of transcription factors (TFs) that have binding sites on
511
MIR132 promoter. h, Working model showing the effects of cardiac glycosides: cardiac
512
glycosides act through their conventional mechanism leading to the transcriptional upregulation
513
of miR-132. The increase in miR-132, together with other pathways altered by cardiac glycosides,
514
downregulated various forms of Tau and provided neuroprotection against toxic insults.
515
35
SUPPLEMENTAL DATA LEGEND
516
517
36
Supplemental Figure S1: Optimization of screen setup. a, ROS-MAP NGN2-iN lines available.
518
b, The BR43 line was selected among the NGN2-iNs established from donors without a clinical
519
diagnosis of AD for its median expression of previously validated miR-132 targets. This line came
520
from an 89-year-old female donor without clinical AD diagnosis but with pathological AD
521
diagnosis. c, miR-132 was upregulated by positive controls forskolin and BNDF in BR43 NGN-2
522
iNs. d-e, 40 to 50 μL of Takara direct lysis buffer was optimal for direct lysing. 1:10 and 1:60
523
dilutions were optimal for qPCR on long RNA RT and small RNA RT, respectively.
524
37
525
38
Supplemental Figure S2: Dose-dependent upregulation of miR-132 and miR-212 in primary
526
rat neurons and human NGN-2iNs. a, Dose curve experiments for miR-212 were performed in
527
DIV14 rat neurons after 24h treatment (N=4-6). b-c, Dose curve experiments for miR-132 and
528
miR-212 were performed in DIV21 human NGN2-iNs after 24h treatment (N=1-2). Solid lines
529
were used for cardiac glycosides, and dotted lines were used for non-cardiac glycosides. EC50 and
530
max fold change were calculated using sigmoidal fit, 4 parameters. Error bars represent SD.
531
39
532
Supplemental Figure S3: Actinomycin D inhibited the upregulation of miR-132 and miR-
533
212. Time-dependent upregulation of miR-132 and miR-212 was completely abolished by
534
pretreatment with 10 μM actinomycin-D before forskolin or oleandrin in DIV14 rat neurons (a-
535
b) and DIV28 rat neurons (c-d).
536
40
537
41
Supplemental Figure S4: Additional effects of miR-132 mimics and cardiac glycosides. a-c,
538
miR-132 target mRNAs were downregulated 72h after transfection with miR-132 mimics. a,
539
MAPT. b, FOXO3a. c, EP300 (N=12, unpaired two-tailed Student’s t-test, Error bars represent
540
SD). d-g, miR-132 and miR-212 were also upregulated and mouse and human MAPT mRNA were
541
downregulated by oleandrin in primary PS19 mouse neurons. Similar observations were also
542
observed in WT neurons but were not statistically significant. h, human MAPT mRNA was
543
expressed at ~8-fold higher than endogenous mouse MAPT (N=3, unpaired two-tailed Student’s
544
t-test, error bars represent SD).
545
42
546
43
Supplemental Figure S5: Oleandrin and proscillaridin A rescued viability loss from
547
glutamate toxicity but were also mildly toxic in younger primary neurons. a, Experimental
548
scheme. b, Proscillaridin A dose-dependently reduced baseline viability (solid line) but also dose-
549
dependently rescued loss of viability due to glutamate (dotted line). c, Similar results were
550
obtained for oleandrin. d-e, Proscillaridin A and oleandrin did not affect neurite length at baseline
551
and partially rescued loss of neurites due to glutamate. f, Representative images of neurons treated
552
with glutamate and proscillaridin A or oleandrin. Cell bodies were highlighted in yellow, and
553
neurites were traced in pink. N=6-8, error bars represent SD. Scale bars are 200 µm.
554
44
555
45
Supplemental Figure S6: Dose-dependent reduction of Tau in iPSC-neurons treated with
556
cardiac glycosides. Tau WT and P301L neurons were differentiated for 6 weeks, then treated with
557
cardiac glycosides for 24h or 72h. a, Representative western blot for WT neurons treated with
558
digoxin. A dose-dependent reduction in total Tau (TAU5) and p-Tau S396 was observed at both
559
24h and 72h. b-c, In parallel, a reduction in MAPT mRNA and an increase in pre-miR-132 RNA
560
were observed. d-f, Similar results were also observed in Tau P301L neurons. g-l, Similar results
561
were observed for WT and P301L neurons treated with oleandrin.
562
46
563
47
Supplemental Figure S7: Cardiac glycosides’ effect on Tau solubility and miR-132 targets in
564
P301L neurons. iPSC-neurons differentiated for 6 weeks were treated for 72h at concentrations
565
of oleandrin (Ole), proscillaridin A (Pros A) and digoxin (DGX) leading to maximum Tau
566
reduction without detectable toxicity. a, Representative western blot analysis of protein lysates
567
generated by detergent fractionation for detection of total Tau (TAU5) and pTau S396 in the
568
soluble (S) and insoluble-pellet (P) fractions. b-c, Densitometry analysis of the western blots (N
569
=2). Graph bars represent mean densitometry ± SD for soluble (b) and insoluble (c) Tau levels
570
relative to vehicle (DMSO). d-g, 24h and 72h treatment with digoxin resulted in a dose-dependent
571
reduction in miR-132 targets, including p300, FOXO3a, GSK3β, and RBFOX1 (N =2). Error bars
572
represent SEM. Similar results were obtained for oleandrin (h-k), and proscillaridin A (l-o).
573
48
574
Supplemental Figure S8: BDNF signaling pathway was enriched with genes upregulated by
575
cardiac glycosides in iPSC-neurons. Pathway plot was modified based on KEGG neurotrophin
576
signaling pathway. Genes in red, blue, and white boxes represented upregulated, downregulated,
577
and unaffected ones, respectively. Mean fold changes (FC) among three cardiac glycosides were
578
labeled, of which positive value represented upregulation and vice versa.
579
49
Supplemental Table 1: Selection of compounds for further validation.
580
Well
_ID
Compound
Pl
at
e#
Sum
mati
on
score
mi
R1
32
ran
k
in
pla
te
Sc
or
e
ra
nk
Dupl
icate
?
Scor
e_mi
R-
129_
Top1
0
Scor
e
miR-
26a
(Bott
om1
0)
Sc
or
e
mi
R-
21
2
(T
op
15
)
An
y
cli
nic
al
tri
als
(C
T)
?
CT
-
Ne
uro
rel
ate
d?
CT
-
AD
rel
ate
d?
CT
-
PD
rel
ate
d?
"Neuro
protecti
ve"
effects?
Cr
os
s
B
B
B?
FDA
appr
oved
?
0365
2:B1
0
Digoxin
2
11
2
4.
5
1
1
1
1
1
0.
5
1
0365
1:A1
2
Almotriptan Malate
1
11
4
4
4
1
1
1
0365
1:L0
3
Inosine
1 10.75
1
4.
75
1
1
1
1
1
1
0365
5:P2
1
Roflumilast
5 10.75
1
4.
75
1
1
1
1
1
1
0365
1:K0
9
Rigosertib (ON-
01910)
1R
e
10.5
2
4.
5
4
1
1
-1
1
0365
5:O0
7
Apixaban
5
9.5
2
4.
5
1
1
1
1
1
50
0365
2:D1
5
Azaperone
2
9.25
3
4.
25
1
1
1
1
1
0365
1:D0
9
Pioglitazone
1
9
12
2
1
1
1
1
1
1
1
0365
6:A0
9
Quercetin
Dihydrate
6
9
2
4.
5
1
1
1
0.
5
1
0365
4:L2
0
LY2811376
4
8.75
3
4.
25
1
1
1
0.
5
1
0365
1:H0
6
SB742457
1
8.75
17
0.
75
1
1
1
1
1
1
1
1
0365
2:E0
9
Betahistine 2HCl
2
8.75
1.0
0
4.
75
1
1
1
1
0365
4:P0
6
Rolipram
4
8.75
9
2.
75
1
1
1
1
1
1
0365
6:E1
6
Azelastine HCl
6
8.75
1
4.
75
1
1
1
1
0365
6:D1
2
Dabrafenib
(GSK2118436)
6
8.25
3
4.
25
1
1
1
1
0365
1:H1
8
Azilsartan
Medoxomil
1
8.25
3
4.
25
1
1
1
1
0365
5:M1
1
Nitazoxanide
5
8.25
3
4.
25
1
1
1
1
51
0365
5:O2
1
Estradiol
5
8.25
11
2.
25
1
1
1
1
1
1
0365
3:F1
7
Rocilinostat (ACY-
1215)
3
8
2
4.
5
1
1
0.
5
1
0365
1:B0
6
PJ34
1
8
2
4.
5
1
1
0.
5
1
0365
4:P2
1
Pelitinib (EKB-
569)
4
7.75
1
4.
75
1
1
1
0365
5:O1
7
Aminoglutethimide
5
7.5
6
3.
5
1
1
1
1
0365
3:D2
1
Empagliflozin (BI
10773)
3
7.5
4
4
1
1
0.
5
1
0365
6:B0
7
URB597
6
7.5
6
3.
5
1
1
1
1
0365
1:N1
4
Pravastatin sodium
1R
e
7.25
13
1.
75
1
1
1
1
0.
5
1
0365
6:E1
8
Amoxicillin
Sodium
6
7
4
4
1
1
1
-1
1
0365
4:P1
1
Budesonide
4
6.75
5
3.
75
1
1
1
0365
6:A0
3
Glycyrrhizic Acid
6
6.75
7
3.
25
1
1
0.
5
1
52
0365
1:M0
5
Entecavir Hydrate
1R
e
6.75
5
3.
75
1
1
1
0365
5:E2
1
Didanosine
5
6.75
5
3.
75
1
1
1
0365
6:A1
0
Quinine HCl
Dihydrate
6
6.75
9
2.
75
1
1
1
1
0365
4:P0
3
Etoposide
4
6.5
6
3.
5
1
1
1
-1
1
0365
1:K0
5
Letrozole
1R
e
6.5
10
2.
5
1
1
1
1
0365
4:P2
0
BIX 02188
4
6.5
2
4.
5
1
1
0365
1:A0
4
Deflazacort
1
6.5
6
3.
5
4
-1
0365
5:P1
0
Mubritinib (TAK
165)
5
6.25
7
3.
25
1
1
1
0365
1:B2
2
Ulipristal
1R
e
6.25
9
2.
75
1
1
0.
5
1
0365
2:E1
7
Ouabain
2
6.25
1
4.
75
1
-1
0.
5
1
0365
2:O1
5
Desloratadine
2
6.25
7
3.
25
1
1
1
53
0365
1:B2
1
ML130 (Nodinitib-
1)
1R
e
6.25
15
1.
25
4
1
0365
4:P0
5
Vincristine
4
6
4
4
1
1
-1
1
0365
5:A1
5
Nitrofural
5
6
4
4
1
1
0365
6:E1
0
Ribavirin
6
6
8
3
1
1
1
0365
1:E2
1
Cimetidine
1R
e
5.75
11
2.
25
1
1
0.
5
1
0365
1:D1
4
Rivaroxaban
1
5.75
13
1.
75
1
1
1
1
0365
3:G0
7
XL888
3
5.75
5
3.
75
1
1
0365
2:P1
0
Dicoumarol
2
5.75
5
3.
75
1
1
0365
3:D1
9
AG-18
3
5.75
1
4.
75
1
0365
1:B1
4
Pifithrin-?
1
5.5
16
1
1
1
1
0.
5
1
0365
1:D1
6
Purmorphamine
1
5.5
8
3
1
0.
5
1
54
0365
4:P1
3
FT-207 (NSC
148958)
4
5.25
7
3.
25
1
-1
1
1
0365
5:A2
1
Busulfan
5
5
12
2
1
1
1
0365
5:P0
6
MLN9708
5
5
8
3
1
1
0365
4:P0
9
Vemurafenib
(PLX4032,
RG7204)
4
5
8
3
1
1
-1
1
0365
1:P1
0
HC-030031
1R
e
4.5
4
4
0.
5
0365
5:B1
3
Phentolamine
Mesylate
5
4.5
10
2.
5
1
1
0365
2:N2
2
GSK2334470
2
4.5
6
3.
5
1
0365
1:H2
1
U-104
1R
e
4.25
3
4.
25
0365
3:C1
2
STF-118804
3
4.25
3
4.
25
0365
1:H1
6
Cinepazide maleate 1R
e
4
14
1.
5
1
1
0.
5
0365
3:C2
0
E-64
3
4
10
2.
5
1
0.
5
55
0365
2:E2
2
Amfenac Sodium
Monohydrate
2
4
4
4
0365
1:A2
0
Myricitrin
1
3.75
11
2.
25
1
0.
5
0365
1:O0
4
Triamcinolone
1R
e
3.75
17
0.
75
1
1
1
0365
5:A1
2
Vidarabine
5
3.75
9
2.
75
1
0365
6:D1
4
Carbazochrome
sodium sulfonate
(AC-17)
6
3.75
5
3.
75
0365
6:B2
1
PF-04691502
6
3.5
14
1.
5
1
1
0365
6:E1
9
Astragaloside A
6
3.5
12
2
1
0.
5
0365
6:B0
5
AZD5438
6
3.25
11
2.
25
1
0365
1:J10
Dapivirine
(TMC120)
1
3.25
7
3.
25
1
-1
0365
1:P1
7
BML-190
1R
e
3.25
7
3.
25
0365
3:D0
5
CRT0044876
3
3.25
7
3.
25
56
0365
6:E2
1
Benserazide HCl
6
3.25
15
1.
25
1
1
0365
1:M0
3
Docetaxel
1R
e
3
8
3
1
-1
0365
1:A1
6
Irinotecan HCl
Trihydrate
1
3
10
2.
5
0.
5
0365
1:O1
9
Fenoprofen
Calcium
1R
e
3
16
1
1
1
0365
3:F1
1
ML167
3
3
8
3
0365
1:H2
2
Moguisteine
1
2.75
5
3.
75
-1
0365
3:D1
7
Tenovin-1
3
2.75
9
2.
75
0365
3:E2
2
Puromycin 2HCl
3
2.5
6
3.
5
-1
0365
4:N1
4
HMN-214
4
2.5
10
2.
5
0365
6:F0
3
Clorsulon
6
2.5
10
2.
5
0365
1:B1
7
SAR131675
1
1.75
15
1.
25
0.
5
57
0365
6:G1
2
Methacycline HCl
6
0.75
13
1.
75
-1
581
582
Supplemental Table 2: Validation results for 44 selected compounds in primary rat neurons and human NGN2-iNs.
583
24h treatment, rat neurons
Drug
Average fold change
P value (two tailed)
P value summary
Significant (alpha<0.05)?
Forskolin
3.639 <0.0001
****
Yes
BIX 02188
3.175
0.011 *
Yes
Ouabain
3.06
0.0001 ***
Yes
Digoxin
2.707
0.0014 **
Yes
Nitazoxanide
1.567
0.0012 **
Yes
Rigosertib (ON-01910)
1.221
0.0173 *
Yes
Pelitinib (EKB-569)
1.199
0.0185 *
Yes
Letrozole
1.198
0.0302 *
Yes
Rocilinostat (ACY-1215)
1.163
0.0296 *
Yes
Quercetin Dihydrate
1.127
0.0492 *
Yes
Betahistine 2HCl
1.127
0.0134 *
Yes
Amoxicillin Sodium
1.105
0.039 *
Yes
Edaravone
1.081
0.0134 *
Yes
Almotriptan
0.9284
0.0344 *
Yes
Etoposide
1.206
0.1674 ns
No
XL888
1.192
0.0649 ns
No
Didanosine
1.084
0.2885 ns
No
U-104
1.078
0.5031 ns
No
58
FT-207 (NSC 148958)
1.069
0.5568 ns
No
Budesonide
1.065
0.5814 ns
No
AG-18
1.064
0.1908 ns
No
LY2811376
1.061
0.1923 ns
No
Aminoglutethimide
1.057
0.3665 ns
No
Apixaban
1.053
0.1083 ns
No
Azaperone
1.049
0.3419 ns
No
Entecavir Hydrate
1.046
0.3067 ns
No
Inosine
1.042
0.5278 ns
No
Roflumilast
1.03
0.4851 ns
No
Ulipristal
1.021
0.7969 ns
No
Azelastine HCl
1.02
0.6803 ns
No
Desloratadine
1.018
0.7573 ns
No
PJ34
1.017
0.8487 ns
No
Empagliflozin (BI 10773)
1.004
0.9457 ns
No
Deflazacort
1.003
0.9635 ns
No
Dabrafenib (GSK2118436)
0.9891
0.3583 ns
No
Nitrofural
0.9846
0.6657 ns
No
Glycyrrhizic Acid
0.9801
0.8447 ns
No
Ribavirin
0.9731
0.6684 ns
No
Rolipram
0.972
0.3145 ns
No
Azilsartan
0.9716
0.6438 ns
No
Dicoumarol
0.9577
0.5499 ns
No
ML130 (Nodinitib-1)
0.9557
0.4339 ns
No
HC-030031
0.9248
0.265 ns
No
URB597
0.9235
0.5509 ns
No
Mubritinib (TAK 165)
0.9158
0.1805 ns
No
DMSO
1 N/A
N/A
N/A
584
59
72h treatment, rat neurons
Drug
Average fold change P value (two tailed)
P value summary
Significant (alpha<0.05)?
Forskolin
5.172 <0.0001
****
Yes
BIX 02188
3.787
0.0002 ***
Yes
Ouabain
2.674
0.0015 **
Yes
Digoxin
2.204
0.0033 **
Yes
Pelitinib (EKB-569)
1.838
0.0097 **
Yes
XL888
1.715
0.0335 *
Yes
Etoposide
1.552
0.0077 **
Yes
Rocilinostat (ACY-1215)
1.411
0.0049 **
Yes
Nitazoxanide
1.386
0.0222 *
Yes
Budesonide
1.287
0.0014 **
Yes
LY2811376
1.129
0.0178 *
Yes
Aminoglutethimide
0.8203
0.0374 *
Yes
Rigosertib (ON-01910)
1.357
0.0652 ns
No
Letrozole
1.192
0.0645 ns
No
Quercetin Dihydrate
1.159
0.0635 ns
No
Mubritinib (TAK 165)
1.147
0.2364 ns
No
Empagliflozin (BI 10773)
1.141
0.0916 ns
No
Desloratadine
1.132
0.1516 ns
No
Azelastine HCl
1.113
0.1408 ns
No
Ulipristal
1.112
0.1648 ns
No
Rolipram
1.088
0.2542 ns
No
Apixaban
1.087
0.0924 ns
No
U-104
1.079
0.3396 ns
No
Betahistine 2HCl
1.073
0.2481 ns
No
Edaravone
1.061
0.2385 ns
No
Entecavir Hydrate
1.055
0.3392 ns
No
60
Glycyrrhizic Acid
1.037
0.6864 ns
No
Amoxicillin Sodium
1.032
0.7329 ns
No
Nitrofural
1.029
0.8807 ns
No
Deflazacort
1.018
0.8445 ns
No
Didanosine
1.017
0.8106 ns
No
FT-207 (NSC 148958)
1.017
0.6764 ns
No
Ribavirin
1.007
0.9078 ns
No
Azaperone
0.9961
0.9079 ns
No
Dicoumarol
0.9762
0.7672 ns
No
AG-18
0.9666
0.5969 ns
No
Azilsartan
0.9656
0.6346 ns
No
Dabrafenib (GSK2118436)
0.9353
0.2136 ns
No
Inosine
0.9109
0.1424 ns
No
URB597
0.9092
0.1946 ns
No
Almotriptan
0.8949
0.0799 ns
No
HC-030031
0.8796
0.1298 ns
No
ML130 (Nodinitib-1)
0.8795
0.1024 ns
No
Roflumilast
0.8752
0.1591 ns
No
PJ34
0.8582
0.0887 ns
No
DMSO
1 N/A
N/A
N/A
585
24h treatment, human NGN2-iNs
Drug
Average fold change
P value (two tailed)
P value summary
Significant (alpha<0.05)?
Digoxin
2.86
0.0251 *
Yes
Ouabain
2.425
0.0129 *
Yes
Rigosertib (ON-01910)
1.663
0.0206 *
Yes
Forskolin
1.391
0.0153 *
Yes
61
Budesonide
1.199
0.0355 *
Yes
Didanosine
1.578
0.1375 ns
No
Empagliflozin (BI 10773)
1.534
0.2076 ns
No
Quercetin Dihydrate
1.437
0.3354 ns
No
Azelastine HCl
1.378
0.1517 ns
No
Dabrafenib (GSK2118436)
1.378
0.381 ns
No
Deflazacort
1.338
0.2725 ns
No
Betahistine 2HCl
1.326
0.2401 ns
No
Nitazoxanide
1.324
0.1063 ns
No
Almotriptan
1.311
0.2891 ns
No
AG-18
1.266
0.0557 ns
No
Glycyrrhizic Acid
1.265
0.461 ns
No
Inosine
1.234
0.2321 ns
No
BIX 02188
1.228
0.2411 ns
No
Entecavir Hydrate
1.221
0.2515 ns
No
Letrozole
1.217
0.2271 ns
No
Rolipram
1.182
0.224 ns
No
Aminoglutethimide
1.177
0.1727 ns
No
Etoposide
1.173
0.3406 ns
No
URB597
1.168
0.1867 ns
No
U-104
1.163
0.4164 ns
No
Nitrofural
1.145
0.3707 ns
No
ML130 (Nodinitib-1)
1.143
0.2225 ns
No
FT-207 (NSC 148958)
1.133
0.4391 ns
No
Rocilinostat (ACY-1215)
1.111
0.4474 ns
No
Roflumilast
1.104
0.6772 ns
No
Amoxicillin Sodium
1.102
0.1956 ns
No
XL888
1.099
0.4062 ns
No
Azilsartan
1.093
0.5713 ns
No
62
Pelitinib (EKB-569)
1.089
0.5371 ns
No
Azaperone
1.079
0.4025 ns
No
PJ34
1.056
0.6615 ns
No
Edaravone
1.043
0.5514 ns
No
Dicoumarol
1.041
0.8104 ns
No
Ulipristal
1.039
0.7041 ns
No
Apixaban
1.023
0.764 ns
No
Desloratadine
1.023
0.8387 ns
No
HC-030031
1.023
0.8394 ns
No
LY2811376
1.007
0.9426 ns
No
Ribavirin
0.9954
0.9648 ns
No
Mubritinib (TAK 165)
0.9671
0.7106 ns
No
DMSO
1 N/A
N/A
N/A
586
587
63
Supplemental Table 3: Chemical structures and miR-132/212 EC50 and max fold change of various compounds.
588
Drugs
(M.W in
g/mol)
Chemical structure
Class
miR-132
miR-212
EC50
(nM)
Rat
(human)
Max F.C.
Rat
(Human)
EC50
(nM)
Rat
(Human)
Max F.C.
Rat
(Human)
Proscillaridin
A
(530.7)
Cardiac
glycoside,
bufadinolide
3.2
(6.86)
2.72
(4.0)
3.32
(10.8)
5.64
(6.70)
Bufalin
(386.5)
Cardiac
glycoside,
bufadinolide
46.54
(40.89)
2.69
(2.62)
35.87
(57)
6.34
(5.1)
64
Oleandrin
(576.7)
Cardiac
glycoside,
cardenolide
50.5
(31)
2.89
(3.96)
44.25
(258.5)
5.82
(7.8)
Digitoxin
(765)
Cardiac
glycoside,
cardenolide
59.83
2.42
60.38
4.66
Digoxin
(780.9)
Cardiac
glycoside,
cardenolide
97.22
(143)
2.42
(2.54)
N/A*
(117)
23.99
(6.51)
Ouabain
(584.7)
Cardiac
glycoside,
cardenolide
118.7
(150-300)
2.82
(2.79)
134.2
(124)
4.05
(5.15)
65
Bufotalin
(444.6)
Cardiac
glycoside,
bufadinolide
130.1
1.75
136.6
2.8
Pelitinib
(467.9)
EGFR inhibit
or
182.6
1.34
326.2
1.57
Cinobufagin
(442.6)
Cardiac
glycoside,
bufadinolide
196.6
2.49
294
6.05
66
Forskolin
(410.5)
cAMP
activator
331.7
4.09
339.2
8.45
BIX02189
(440.5)
MEK5
inhibitor
430.1
1.5
626.9
1.55
BIX02188
(412.5)
MEK5
inhibitor
3064
(N/A)
3
(N/A)
3118
(N/A)
5.57
(N/A)
67
589
590
591
592
593
594
595
* While robust upregulation of miR-212 with digoxin was observed, the results could not be fit into a sigmoidal curve, so that EC50
596
was not calculated.
597
598
Nitazoxanide
(307.3)
Antiprotozoal
agent
10010
1.5
12650
3.9
68
Supplemental Table 4: Candidate small molecule compounds that may regulate miRNAs implicated in neurological disease.
599
Target
Regulator
Potential application
References
(PMID)
Hits
miR-107
Inducer
Neurogenesis
25662174
Ammonium Glycyrrhizinate, Cysteamine HCl, Amonafide,
MGCD-265 analog
miR-34a-5p
Inducer
Synaptogenesis
22160687,
22160706
Vilazodone HCl, Z-FA-FMK, DMXAA, PF-4708671, Quinine
HCl Dihydrate, Quercetin
miR-9-5p
Inducer
Neuronal
differentiation, treat
Huntington’s disease
16357340,
21753754,
19118166
Doxycycline HCl, OSI-906, Osthole, Tenovin-1, Famotidine
miR-124-3p
Inducer
Neuronal differentiation
16357340,
21753754
Roxatidine Acetate HCl, Santacruzamate A, Hematoxylin,
BMS-777607
miR-29a-3p
Inducer
Treat Alzheimer's
disease
18434550,
21930776
Lomeguatrib, Quisinostat, Bendamustine HCl, Rosiglitazone
miR-101-3p
Inducer
Treat Alzheimer’s
disease
20395292
Meclofenoxate HCl, PF-4708671, Axitinib, Cetirizine DiHCl,
Vardenafil HCl Trihydrate
miR-26b-5p
Suppressor
Treat Alzheimer’s
disease
24027266
Olsalazine Sodium, Uracil, (+,-)-Octopamine HCl, EX 527, ZM
336372, Roxatidine Acetate HCl
miR-133b
Inducer
Treat Parkinson’s
Disease
17761882
EPZ004777, Vemurafenib, Dicoumarol, MK-2461,
Itraconazole, GW788388
miR-27a/b-
3p
Suppressor
Treat Parkinson’s
Disease
27456084
Mefenamic Acid, Brivanib, Saracatinib, Probucol
miR-128-3p
Inducer
Treat epilepsy
24311694,
29581509
Mycophenolate Mofetil, EPZ004777, Crenolanib, Milciclib,
Tinidazole
miR-134-5p
Suppressor
Treat epilepsy
22683779
GS-9973, Tilmicosin, Cilengitide, Methotrexate, CCT128930,
Ambroxol HCl, L-Adrenaline
miR-137-3p
Suppressor
Treat schizophrenia
26005852
Saxagliptin, LY294002, Telaprevir, ZM 336372, Roxatidine
Acetate HCl, PU-H71
600
601
69
Supplemental Table 5: List of commercial miRNA primers used for miRNA RT-qPCR.
602
No.
Name
GeneGlobe ID
Species
1 hsa-let-7a-5p
YP00205727
Human, mouse, rat
2 hsa-mir-22-3p
YP00204606
Human, mouse, rat
3 hsa-miR-26b-5p
YP00204172
Human, mouse, rat
4 hsa-miR-99a-5p
YP00204521
Human, mouse, rat
5 hsa-miR-103a-3p
YP00204063
Human, mouse, rat
6 mmu-miR-124-3p
YP02119832
Mouse, rat
7 hsa-miR-125a-5p
YP00204339
Human, mouse, rat
8 hsa-miR-128-3p
YP00205995
Human, mouse, rat
9 hsa-miR-132-3p
YP00206035
Human, mouse, rat
10 hsa-miR-138-1-3p
YP00205881
Human, mouse, rat
11 hsa-miR-191-5p
YP00204306
Human, mouse, rat
12 hsa-miR-212-3p
YP00204170
Human
13 mmu-miR-212-3p
YP00206022
Mouse, rat
14 hsa-miR-107
YP00204468
Human, mouse, rat
603
604
70
Supplemental Table 6: List of mRNA primer sequences used for mRNA RT-qPCR.
605
Primer
Sequence (5' - 3')
Amplicon size
Species
Accession number
18S_F
ACCACATCCAAGGAAGGCAG
243nt
Rat, mouse, human
NR_146119.1
18S_R
CCGCTCCCAAGATCCAACTA
243nt
Rat, mouse, human
455-697
FOXO3a_F
GGCAAAGCAGACCCTCAAAC
65nt
Rat, mouse, human
NM_001455.4
FOXO3a_R
TGAGAGCAGATTTGGCAAAGG
65nt
Rat, mouse, human
2339-2403
pre_miR132_F
CCTCCGGTTCCCACAGTAACAA
52 nt
Rat, mouse, human
NR_031878.1
pre_miR132_R CCGCGTCTCCAGGGCAAC
52nt
Rat, mouse, human
1073-1107
pre_miR212_F
GGCTCTAGACTGCTTACTGCC
70nt
Rat, mouse, human
NR_031925.1
pre_miR212_R GGCCAGGCGTCGGTG
70nt
Rat, mouse
37-106
r_MAPT_F
GAACCACCAAAATCCGGAGA
164nt
Rat
NM_017212.2
r_MAPT_R
CTCTTACTGGCAGACGGTGAC
164nt
Rat
502-655
r_EP300_F
AATGGACAAGGGATAATGCCCA
120nt
Rat, mouse
XM_039080287.1
r_EP300_R
CTCAGTCAATAAACTGCCAGCA
120nt
Rat, mouse
753-872
r_BDNF_F
CTACGAGACCAAGTGTAATCC
147nt
Rat
NM_001270638.1
r_BDNF_R
AACCGCCAGCCAATTCTCTTT
147nt
Rat
651-797
r_GAPDH_F
CAACTCCCTCAAGATTGTCAGCAA 118nt
Rat
NM_001394060.2
r_GAPDH_R
GGCATGGACTGTGGTCATGA
118nt
Rat
495-612
h_MAPT_F
GTCGAAGATTGGGTCCCT
147nt
Human
M_016835.5
h_MAPT_R
GACACCACTGGCGACTTGTA
147nt
Human
2154-2300
h_GAPDH_F
CATCACTGCCACCCAGAAGACTG
153nt
Human, mouse
NM_002046.7
h_GAPDH_R
ATGCCAGTGAGCTTCCCGTTCAG
153nt
Human, mouse
616-768
m_MAPT_R
GAACCACCAAAATCCGGAGA
164nt
Mouse
NM_001038609.3
m_MAPT_R
CTCTTACTAGCTGATGGTGAC
164nt
Mouse
668-831
606
607
71
Supplemental Table 7: List of primary and secondary antibodies used for Western blots.
608
No
Target
Species
Source
Catalog #
Note
1 FOXO3a
Rabbit
Abcam
ab70315
Ref. Fig. 4
2 Tau-5
Mouse
Invitrogen
AHB0042
Ref. Fig. 4 & 6
3 Tau AT8
Mouse
Invitrogen
MN1020
Ref. Fig. 4
4 Tau S396
Rabbit
Abcam
ab109390
Ref. Fig. 4
5 β-actin
Mouse
Abcam
ab3280
Ref. Fig. 4
6 Tau S396
Rabbit
Invitrogen
44752G
Ref. Fig. 6
7 β-actin
Mouse
Sigma-Aldrich
A1978
Ref. Fig. 6
8 β-III-Tubulin
Mouse
Sigma-Aldrich
T-8660
Ref. Fig. 6
9 PSD-95
Mouse
Neuro-Mab
K28/43
Ref. Fig. 6
10 SYN1
Rabbit
Synaptic Systems
106-103
Ref. Fig. 6
11 Tau K9JA
Rabbit
Agilent
A002401-2
Ref. Fig. 7
12 MAP2
Rabbit
Sigma-Aldrich
AB5543
Ref. Fig. 7
13 GSK3β
Rabbit
Cell Signaling
#9315
Ref. Ex. Fig. 7
14 EP300
Mouse
Novus Biologicals
E15NB100-507SS
Ref. Ex. Fig. 7
15 FOXO3a
Mouse
Protein Tech/Thermo Fisher
1F12D11
Ref. Ex. Fig. 7
16 RBFOX1
Mouse
Thermo Fisher
MA5-33104 (A2BP1) Ref. Ex. Fig. 7
17 Anti-rabbit IgG
Cell Signaling
#14708
Ref. Fig. 4
18 Anti-mouse IgG
Cell Signaling
#14709
Ref. Fig. 4
19 Anti-rabbit IgG
Cell Signaling
#7074
Ref. Fig. 6
20 Anti-mouse IgG
Cell Signaling
#7076
Ref. Fig. 6
21 Anti-rabbit IgG Alexa Fluor 594
Invitrogen
A11012
Ref. Fig. 7
22 Anti-mouse IgG Alexa Fluor 488
Invitrogen
A11029
Ref. Fig. 7
23 Anti-rabbit IgG Alexa Fluor 595
Invitrogen
A11032
Ref. Fig. 7
609
72
MATERIALS AND METHODS
610
Induced neuron differentiation from iPSC
611
Induced pluripotent stem cell (iPSC) lines were retrieved and differentiated into neurons with
612
NGN2 expression, as previously reported23. Briefly, iPSCs were plated in mTeSR1 media at a
613
density of 9.5x104 cells/cm2 on Matrigel (Corning #354234)-coated plates. Cells were then
614
transduced with the following virus: pTet-O-NGN2-puro (Addgene #52047): 0.1 μL per 5x104
615
cells; Tet-O-FUW-eGFP (Addgene #30130): 0.05μL per 5x104 cells; Fudelta GW-rtTA (Addgene
616
#19780): 0.11 μL per 5x104 cells. Transduced cells were dissociated with Accutase (StemCell
617
Technologies) and plated onto Matrigel-coated plates in mTeSR1 (StemCell Technologies) at
618
5x104/cm2 (Day 0). On day 1, media was changed to KSR media (Knockout DMEM, 15% KOSR,
619
1x MEM-NEAA, 55 μM beta-mercaptoethanol, 1x GlutaMAX; Gibco) with doxycycline (2 μg/ml,
620
Sigma-Aldrich). Doxycyline was maintained in the media for the remainder of the differentiation.
621
On day 2, the media was changed to 1:1 KSR: N2B media with puromycin (5 μg/ml, Gibco), where
622
N2B was composed of DMEM/F12, 1x GlutaMAX, 1x N2 supplement B (StemCell Technologies)
623
and 0.3% dextrose (Sigma-Aldrich). Puromycin was maintained in the media throughout the
624
differentiation. On day 3, the media was changed to N2B media + 1:100 B-27 supplement
625
(GIBCO) and puromycin (10 μg/ml). From day 4 on, cells were cultured in NBM media
626
(Neurobasal medium, 0.5x MEM-NEAA, 1x GlutaMAX, 0.3% dextrose) + 1:50 B-27 + BDNF,
627
GDNF, CNTF (10 ng/ml each, Peprotech). After day 4, half of the media was replaced by fresh
628
media twice per week. Cells were stocked on day 4 at 1~2x106 cells in 200 μL freezing media
629
(50% day 4 media + 40% FBS + 10% DMSO) per cryovial in -80oC overnight, followed by liquid
630
nitrogen storage. iPSC-derived neurons used for validation experiments were prepared similarly.
631
73
iPSC lines were generated following review and approval through Brigham and Women’s Hospital
632
Institutional Review Board (IBR#2015P001676).
633
Preparation of NGN2-iNs for high-throughput screen.
634
Twenty-five 96-well plates (Corning) were coated with Matrigel solution (0.2 mg/mL in
635
DMEM/F12) at 60 μL per well for 1.5 hours at 37oC. Then, the Matrigel solution was completely
636
removed, and 100 μL PBS (Gibco) was added per well using electronic 12-channel pipettes in
637
speed 3 (e12c-pip; Eppendorf). The plates were temporality incubated at 37oC. Frozen day 4 iPSC-
638
iNs were thawed in 500 μL pre-warmed resuspension media per vial, which was composed of
639
NBM, 1:100 B27, and 1:1000 ROCKi (StemCell Technologies), and were kept in a warm metal
640
bath to facilitate the thawing. Multiple vials were pooled into one 50 mL conical tube, then pre-
641
warmed resuspension media were added drop-wisely to reach the volume of 40 mL. After gently
642
mixing by reverting the tube, viable cell concentration was counted with trypan blue (Bio-Rad).
643
The cells were spun down (220 g, 5 min, room temperature) and resuspended in day 4 media at
644
1x105 cells/mL. Then, PBS was completely removed from the plates, and 100 μL cell suspension
645
was added per well, using e12c-pip at speed 3. The cells were incubated at 37oC after shaking the
646
plates for even distribution (day 4). To reduce evaporation during incubation, plates were kept in
647
plastic containers lined with sterile wet paper towels. On day 5, an additional 100 μL pre-warmed
648
day 4 medium was added per well using e12c-pip in speed 1. On days 7/10/14/18, 95 μL
649
conditioned medium was removed, and 100 μL pre-warmed day 4 medium was added per well,
650
both using e12c-pip.
651
High-throughput screen in NGN2-iNs
652
On day 19, half (three 384-well plates) of the Selleck bioactive compound library (N=1902
653
compounds) were pin transferred (V&P Scientific) to twelve NGN2-iN plates using the Seiko
654
74
Compound Transfer Robot at 200 nL per well (final concentration at 10 μM). Positive control
655
(Forskolin) and negative control (DMSO) were also pin transferred to the wells without library
656
compound. On day 20, four 10x photos were taken per well automatically using the ImageXpress
657
Micro Confocal microscope (Molecular Devices). Then, the media in the plates was removed with
658
approximately 20 μL media left, using a 24-channel stainless steel manifold (Drummond #3-000-
659
101) linked with a vacuum at a low speed. With the help of Multidrop™ Combi Reagent Dispenser
660
(Thermo Scientific) and the standard cassette (speed: low), 250 μL ice-cold DPBS (Wisent) was
661
added per well. The DPBS was removed with approximately 20 μL liquid left, using another 24-
662
channel stainless steel manifold linked with the vacuum at a low speed. The residual DPBS was
663
completely removed using a mechanic 12-channel pipette. Next, 45 μL lysing solution was added
664
per well using another Multidrop™ Combi Reagent Dispenser with the small cassette (speed: low),
665
where the lysing solution is composed of single-cell lysis buffer (Takara #635013): 1x RNase
666
Inhibitor Murine (NEB): nuclease-free water (Exiqon) = 19:1:190. Thorough lysis was achieved
667
by shaking the plates on a shaker for 5 min at room temperature. The lysis samples were transferred
668
from four 96-well plates to each 384-square-well plate using the 96-well module-coupled VPrep
669
liquid handler (Agilent) with 30 μL tips (twice without changing tips). After sealing, the 384-
670
square-well plates were spun down at 4000 rpm for 5 min, and 10 μL supernatant was aliquoted
671
to a 384-well plate (Eppendorf) using the 384-well module-coupled VPrep liquid handler. The
672
plates were finally sealed with the PlateLoc Heat Sealer (Agilent) and stored at -20oC. The other
673
half (three and a quarter 384-well plates) of the compound library were added to the remaining
674
thirteen 96-well plates after one day delay (on day 20), using the identical protocol due to the time
675
consumption. The high-throughput screen was conducted in the ICCB-Longwood Screening
676
Facility, Harvard Medical School.
677
75
Ultra-low input miRNA-seq using the RealSeq
678
To avoid RNA purification, we used RealSeq-T technology (RealSeq Biosciences) following
679
manufacturer recommendations. In summary, cell lysates were incubated at 70C for 5 minutes on
680
RealSeq hybridization buffer (100 mM NaCl, 50 mM Tris-HCl, 10 mM MgCl2, 1mM DTT, pH
681
7.9) with 1x RealSeq biotinylated DNA probes to target all miRNAs in miRbase 21. After 2 hours
682
of incubation at 37C, 10 μL of RealSeq Beads were added, and miRNAs were captured using a
683
384-well Magnet Plate (Alpaqua, MA). Following three washes with RealSeq Wash buffer,
684
miRNA was eluted from beads in 10uL of RNase-free water. All the miRNA elusion was input to
685
prepare sequencing libraries with RealSeq-Biofluids following manufacturer instructions (RealSeq
686
Biosciences). In summary, a single adapter and circularization approach was used 24. Libraries
687
were barcoded with dual indexes and sequenced with a NextSeq 550 (Illumina, CA). FastQ files
688
were trimmed of adapter sequences using Cutadapt with the following parameters: cutadapt -u 1 -
689
a TGGAATTCTCGGGTGCCAAGG -m 15. Trimmed reads were aligned to the corresponding
690
reference by using Bowtie 59. Counts of each miRNA were normalized among samples by total
691
miRNA read counts.
692
Animal Use
693
This study was carried out in accordance with the recommendations in the U.S. National Institutes
694
of Health Guide for the Care and Use of Laboratory Animals. The protocol was approved by the
695
Institutional Animal Care and Use Committee at Brigham and Women’s Hospital. Mice were
696
maintained on a 12:12-h light/dark cycle (7:00 am on/7:00 pm off) with food and water provided
697
ad libitum before experimental procedures
698
Rat and Mouse Primary Neuron Culture
699
76
Rat primary cortical neuron cultures were prepared from E18 SAS Sprague Dawley pups (Charles
700
River). Brain tissues were dissected, dissociated enzymatically by 0.25% Trypsin-EDTA (Thermo
701
Fisher Scientific), triturated with fire-polished glass Pasteur pipettes, and passed through a 40 μm
702
cell strainer (Sigma-Aldrich) to remove clumps. After counting, neurons were seeded onto poly-
703
D-lysine (Sigma-Aldrich) coated cell culture plates at 80,000 cells/cm2 in neurobasal medium
704
supplemented with 1X B27 and 0.25X GlutaMax. Half of the cell medium was changed every 4
705
days until use.
706
Mouse primary cortical neuron cultures were prepared from P1 or P2 postnatal pups from PS19
707
mouse breeding pairs. After dissection, mouse brain tissues were kept in Hibernate-A medium at
708
4oC in the dark for ~4h. After genotyping, brains from pups of the same genotype, either WT or
709
PS19, were pooled together and dissociated enzymatically with papain solution (Worthington).
710
After dissociation, mouse neurons were prepared and cultured similarly to rat neurons.
711
siRNA and miRNA Mimics Transfection
712
siRNAs and miRNA mimics were purchased from Dharmacon (Horizon Discovery) and were
713
dissolved in nuclease-free water to prepare 50 μM stock concentrations. Transfection was
714
performed using NeuroMag (OZ Biosciences). For siRNA knockdown, transfection was
715
performed with 50 nM siRNAs on DIV7 and DIV9, and RNA was collected for analysis at DIV11.
716
Transfection of DIV14 neurons with 50 nM miR-132 or CTRL mimics was performed similarly.
717
RNA was collected for analysis 72h later at DIV17.
718
RNA Extraction, cDNA Preparation, and RT-qPCR
719
Total RNA from cells was extracted using the Norgen Total RNA Purification Kit (Norgen Biotek)
720
following the manufacturer’s protocol. DNAse1 was applied during RNA extraction to remove
721
77
genomic DNA. RNA was eluted in nuclease-free water, and the concentration was measured using
722
Nanodrop (Thermo Fisher Scientific).
723
For miRNA analysis, 50ng of RNA was reverse transcribed into cDNA using the miRCURY LNA
724
RT kit (Qiagen). RT-qPCR mix was prepared using the miRCURY LNA SYBR Green PCR kit
725
(Qiagen). qPCR was performed using the QuantStudio 7 Flex System. The cycling conditions were
726
95°C for 10 min, 50 cycles of 95°C for 15 s, and 60°C for 1 min following dissociation analysis.
727
miRNA expression was normalized to the geometric mean of miR-103a and let7a unless stated
728
otherwise in figure legends. For mRNA analysis, 250-1000 ng of RNA was reverse transcribed
729
into cDNA using the High Capacity cDNA Reverse Transcription Kit (Thermo Fisher Scientific).
730
RT-qPCR mix was prepared using the PowerUp SYBR Green Master Mix (Thermo Fisher
731
Scientific). qPCR was performed using the QuantStudio 7 Flex System. The cycling conditions
732
were 95°C for 10 min, 50 cycles of 95°C for 15 s, and 60°C for 1 min following dissociation
733
analysis. mRNA expression was normalized to the geometric mean of 18S and GAPDH.
734
Quantification was performed using the delta-delta Ct method. miRNA and mRNA primers used
735
were listed in Supplemental Tables 5 and 6.
736
Transcriptome profile by RNA-seq
737
After quality control by Agilent 2100 Bioanalyzer, the total RNA was used as input for library
738
preparation by Novogene Co., Ltd, followed by high-throughput sequencing on Illumina HiSeq X
739
with PE150 mode to produce approximately 20 M reads per sample. The reads were quality
740
controlled with FastQC, trimmed with Trimmomatic, aligned with HiSat2 to hg38, and quantified
741
with HTSeq-count using the Galaxy platform. Read counts were processed for differential
742
expression analysis using the R package DEBrowser with DESeq2. Pathway analysis was
743
78
performed by Enrichr. Promoter binding sites were extracted from JASPAR 2022 TFBS via the
744
UCSC genome browser.
745
Western Blot Analysis
746
Total protein was extracted using RIPA buffer (Boston Bioproducts) supplemented with Complete,
747
Mini, EDTA-free Protease Inhibitor Cocktail (Millipore Sigma). Protein concentrations were
748
determined using the Micro BCA Protein Assay Kit (Thermo Fisher Scientific). Equal amounts of
749
protein were loaded, and electrophoresis was performed in NuPAGE 4 to 12% gradient Bis-Tris
750
polyacrylamide protein gels (Thermo Fisher Scientific). Proteins were transferred to Immun-Blot
751
PVDF membranes (Bio-Rad) and then blocked with 5% milk in tris-buffered saline with 0.1%
752
Tween (TBS-T, Boston Bioproduct) for 1 h. Membranes were incubated overnight with primary
753
antibodies at 4 °C (Supplemental Table 7). Blots were washed and incubated with secondary
754
antibodies for 2 h at room temperature. After washing, bands were visualized with ECL
755
chemiluminescence reagents (Genesee Scientific) using the iBright Imaging System (Thermo
756
Fisher Scientific). Band intensity was measured using the Image Studio Lite software (LI-COR
757
Biosciences). Protein expression level was normalized to β-actin or total Tau as appropriate.
758
WST-1 Assay and Neurite Length Measurement
759
Cell viability was measured by WST-1 reduction assay (Sigma-Aldrich). For the assay, all medium
760
was removed and replaced with 1X WST-1 reagent dissolved in complete neurobasal medium,
761
followed by 3 hours of incubation at 37°C. The absorbance of the culture medium was measured
762
with a microplate reader at test and reference wavelengths of 450 nm and 630 nm, respectively.
763
Live cell imaging was performed using the IncuCyteTM Live-Cell Imaging System (Essen
764
BioScience). Cell confluency, cell body number, neurite length, and branching points were
765
monitored and quantified using the IncuCyteTM software.
766
79
Human iPSC-Neurons from NPC lines
767
Approval for work with human subjects and derived iPSCs was obtained under the Massachusetts
768
General Hospital/MGB-approved IRB Protocol (#2010P001611/MGH). The NPC line MGH-
769
2046-RC1 (P301L) was derived from a female individual in her 50s with FTD carrying the
770
autosomal dominant mutation P301L (c.C1907T NCBI NM_001123066, rs63751273). The NPC
771
line MGH-2069-RC1 (WT) was derived from a related female individual in her 40s carrying the
772
unaffected WT Tau. Fibroblasts from the two individuals were reprogrammed into iPSCs,
773
converted into cortical-enriched neural progenitor cells (NPCs), and differentiated into neuronal
774
cells over 6-8 weeks by growth factor withdrawal, as previously described 60.
775
iPSC-neurons compound treatment for western blot analysis and semi-quantitative analysis
776
NPCs were plated at an average density of 90,000 cells/cm2 of six-well plates or 96-well plates
777
coated with poly-ornithine and laminin (POL) in DMEM/F12-B27 media and differentiated for 6
778
weeks. Compound treatment was performed by removing half-volume of neuronal-conditioned
779
media from each well and adding half-volume of new media pre-mixed with the compound at 2X
780
final concentration, followed by incubation at 37 °C. After 24h or 72h, neurons were washed in
781
PBS, collected, and lysed. Western blot and densitometry quantifications were performed as
782
previously described35.
783
Tau Protein Solubility Analysis
784
Neuronal cell lysates and fractionation were prepared based on protein differential solubility to
785
detergents Triton-X100 and SDS, as previously described 61. Briefly, cell pellets corresponding to
786
~800,000 cells were lysed in 1% (v/v) Triton-X100 buffer (Fisher Scientific) in DPBS
787
supplemented with 1% (v/v) Halt Protease/Phosphatase inhibitors (Thermo Fisher Scientific),
788
1:5000 Benzonase (Sigma) and 10 mM DTT (New England BioLabs). Lysates were centrifugated
789
80
at 14,000 g for 10 min at 4°C. The supernatants containing Trion-soluble proteins (S fractions)
790
were transferred to new tubes for western blot analysis. The pellets were resuspended in 5% (v/v)
791
SDS (Sigma) in RIPA buffer supplemented with 1% (v/v) Halt Protease/Phosphatase inhibitors
792
(Thermo Fisher Scientific), 1:5000 Benzonase (Sigma) and 10 mM DTT (New England BioLabs),
793
and centrifugated at 20,000 g for 2 min at room temperature. These supernatants contained proteins
794
of lower solubility/insoluble (P fractions). SDS-PAGE western blot was performed by loading 20
795
μg of each S-fraction and double the volume of the P-fraction onto pre-cast Tris-Acetate SDS-
796
PAGE (Novex, Invitrogen). Western blots were performed as before. Densitometry quantification
797
(pixel mean intensity in arbitrary units, a.u.) was done with the Histogram function of Adobe
798
Photoshop 2022, normalized to the respective GAPDH intensity in the S-fraction, followed by
799
normalization to Vehicle.
800
Neuronal viability assays
801
For cardiac glycoside’s dose-dependent effects on viability, NPCs were plated (~90,000 cells/cm2)
802
and differentiated in 96-well plates for 8 weeks. After treatment with cardiac glycosides, viability
803
was measured with the Alamar Blue HS Cell viability reagent (Life Technologies) at 1:10 dilution,
804
after 4h incubation at 37°C and according to the manufacturer’s instructions. Readings were done
805
in the EnVision Multilabel Plate Reader (Perkin Elmer).
806
For stress vulnerability assays, 1 µM or 5 µM of digoxin, oleandrin, or proscillaridin A was added
807
to the culture media and incubated for 6h at 37 °C. Then, either 30 μM Aβ(1-42), 5 μM rotenone,
808
400 μM NMDA, or vehicle (DMSO) alone, was added to each well for an additional 18h of
809
incubation. At 24h, viability was measured with the Alamar Blue HS Cell Viability reagent (Life
810
Technologies) and the EnVision Multilabel Plate Reader (Perkin Elmer).
811
81
Immunofluorescence of neuronal cells
812
NPCs were plated at a starting density of ~90,000 cells/cm2 in black, clear flat bottom, POL-coated
813
96-well plates (Corning) in DMEM/F12-B27 media and differentiated for six weeks, followed by
814
compound treatment. Neurons were fixed with 4% (v/v) formaldehyde-PBS (Tousimis) for 30 min,
815
washed in PBS (Corning), incubated in blocking/permeabilization buffer [10 mg/mL BSA
816
(Sigma), 0.05% (v/v) Tween-20 (Bio-Rad), 2% (v/v) goat serum (Life Technologies), 0.1% Triton
817
X-100 (Bio-Rad), in PBS] for 2h, and incubated with primary antibodies overnight (Tau K9JA at
818
1:1000, MAP2 at 1:1000, Hoechst-33342 at 1:2500). Cells were washed with PBS and incubated
819
with the corresponding AlexaFluor-conjugated secondary antibodies at 1:500 dilution (Life
820
Technologies). Image acquisition was done with a Zeiss AxioVert 200 inverted fluorescence
821
microscope.
822
Data Analysis
823
Data management and calculations were performed using Prism 9 (GraphPad). Comparisons
824
between two groups were done using the unpaired two-tailed student t-test. For the comparison of
825
more than two groups, a one-way analysis of variance (ANOVA), followed by post hoc test, was
826
performed. A P value < 0.05 was considered statistically significant, and the following notations
827
are used in all figures: *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001. All error bars
828
shown represent standard deviation (SD) unless otherwise stated.
829
82
References
830
1.
Cummings, J., Lee, G., Ritter, A., Sabbagh, M. & Zhong, K. Alzheimer's disease drug development
831
pipeline: 2020. Alzheimers Dement (N Y) 6, e12050 (2020).
832
2.
Wang, F., Zuroske, T. & Watts, J.K. RNA therapeutics on the rise. Nat Rev Drug Discov 19, 441-
833
442 (2020).
834
3.
Lekka, E. & Hall, J. Noncoding RNAs in disease. FEBS Lett 592, 2884-2900 (2018).
835
4.
Warner, K.D., Hajdin, C.E. & Weeks, K.M. Principles for targeting RNA with drug-like small
836
molecules. Nat Rev Drug Discov 17, 547-558 (2018).
837
5.
Gebert, L.F.R. & MacRae, I.J. Regulation of microRNA function in animals. Nat Rev Mol Cell Biol
838
20, 21-37 (2019).
839
6.
Ha, T.Y. MicroRNAs in Human Diseases: From Cancer to Cardiovascular Disease. Immune Netw
840
11, 135-154 (2011).
841
7.
Salta, E. & De Strooper, B. microRNA-132: a key noncoding RNA operating in the cellular phase
842
of Alzheimer's disease. FASEB J 31, 424-433 (2017).
843
8.
Pichler, S., et al. The miRNome of Alzheimer's disease: consistent downregulation of the miR-
844
132/212 cluster. Neurobiol Aging 50, 167 e161-167 e110 (2017).
845
9.
Cha, D.J., et al. miR-212 and miR-132 Are Downregulated in Neurally Derived Plasma Exosomes
846
of Alzheimer's Patients. Front Neurosci 13, 1208 (2019).
847
10.
Lau, P., et al. Alteration of the microRNA network during the progression of Alzheimer's disease.
848
EMBO Mol Med 5, 1613-1634 (2013).
849
11.
Wong, H.K., et al. De-repression of FOXO3a death axis by microRNA-132 and -212 causes
850
neuronal apoptosis in Alzheimer's disease. Hum Mol Genet 22, 3077-3092 (2013).
851
12.
Hernandez-Rapp, J., et al. microRNA-132/212 deficiency enhances Abeta production and senile
852
plaque deposition in Alzheimer's disease triple transgenic mice. Sci Rep 6, 30953 (2016).
853
13.
Salta, E., Sierksma, A., Vanden Eynden, E. & De Strooper, B. miR-132 loss de-represses ITPKB and
854
aggravates amyloid and TAU pathology in Alzheimer's brain. EMBO Mol Med 8, 1005-1018 (2016).
855
14.
Smith, P.Y., et al. miR-132/212 deficiency impairs tau metabolism and promotes pathological
856
aggregation in vivo. Hum Mol Genet 24, 6721-6735 (2015).
857
15.
Wang, Y., et al. Downregulation of miR-132/212 impairs S-nitrosylation balance and induces tau
858
phosphorylation in Alzheimer's disease. Neurobiol Aging 51, 156-166 (2017).
859
16.
El Fatimy, R., et al. MicroRNA-132 provides neuroprotection for tauopathies via multiple
860
signaling pathways. Acta Neuropathol 136, 537-555 (2018).
861
17.
Walgrave, H., et al. Restoring miR-132 expression rescues adult hippocampal neurogenesis and
862
memory deficits in Alzheimer's disease. Cell Stem Cell 28, 1805-1821 e1808 (2021).
863
18.
Chen, Y., Gao, D.Y. & Huang, L. In vivo delivery of miRNAs for cancer therapy: challenges and
864
strategies. Adv Drug Deliv Rev 81, 128-141 (2015).
865
19.
Garzon, R., Marcucci, G. & Croce, C.M. Targeting microRNAs in cancer: rationale, strategies and
866
challenges. Nat Rev Drug Discov 9, 775-789 (2010).
867
20.
Watts, J.K., Brown, R.H. & Khvorova, A. Nucleic Acid Therapeutics for Neurological Diseases.
868
Neurotherapeutics 16, 245-247 (2019).
869
21.
Nguyen, L.D., Chau, R.K. & Krichevsky, A.M. Small Molecule Drugs Targeting Non-Coding RNAs as
870
Treatments for Alzheimer's Disease and Related Dementias. Genes (Basel) 12 (2021).
871
22.
Van Meter, E.N., Onyango, J.A. & Teske, K.A. A review of currently identified small molecule
872
modulators of microRNA function. Eur J Med Chem 188, 112008 (2020).
873
83
23.
Lagomarsino, V.N., et al. Stem cell-derived neurons reflect features of protein networks,
874
neuropathology, and cognitive outcome of their aged human donors. Neuron 109, 3402-3420 e3409
875
(2021).
876
24.
Barberan-Soler, S., et al. Decreasing miRNA sequencing bias using a single adapter and
877
circularization approach. Genome Biol 19, 105 (2018).
878
25.
Magill, S.T., et al. microRNA-132 regulates dendritic growth and arborization of newborn
879
neurons in the adult hippocampus. Proc Natl Acad Sci U S A 107, 20382-20387 (2010).
880
26.
Vandesompele, J., et al. Accurate normalization of real-time quantitative RT-PCR data by
881
geometric averaging of multiple internal control genes. Genome Biol 3, RESEARCH0034 (2002).
882
27.
Vo, N., et al. A cAMP-response element binding protein-induced microRNA regulates neuronal
883
morphogenesis. Proc Natl Acad Sci U S A 102, 16426-16431 (2005).
884
28.
Van Kanegan, M.J., et al. BDNF mediates neuroprotection against oxygen-glucose deprivation by
885
the cardiac glycoside oleandrin. J Neurosci 34, 963-968 (2014).
886
29.
Garofalo, S., et al. The Glycoside Oleandrin Reduces Glioma Growth with Direct and Indirect
887
Effects on Tumor Cells. J Neurosci 37, 3926-3939 (2017).
888
30.
Yoshiyama, Y., et al. Synapse loss and microglial activation precede tangles in a P301S tauopathy
889
mouse model. Neuron 53, 337-351 (2007).
890
31.
Brines, M.L., Dare, A.O. & de Lanerolle, N.C. The cardiac glycoside ouabain potentiates
891
excitotoxic injury of adult neurons in rat hippocampus. Neurosci Lett 191, 145-148 (1995).
892
32.
Sun, Y., Dong, Z., Khodabakhsh, H., Chatterjee, S. & Guo, S. Zebrafish chemical screening reveals
893
the impairment of dopaminergic neuronal survival by cardiac glycosides. PLoS One 7, e35645 (2012).
894
33.
Silva, M.C., et al. Prolonged tau clearance and stress vulnerability rescue by pharmacological
895
activation of autophagy in tauopathy neurons. Nat Commun 11, 3258 (2020).
896
34.
Silva, M.C., et al. Discovery and Optimization of Tau Targeted Protein Degraders Enabled by
897
Patient Induced Pluripotent Stem Cells-Derived Neuronal Models of Tauopathy. Front Cell Neurosci 16,
898
801179 (2022).
899
35.
Silva, M.C., et al. Targeted degradation of aberrant tau in frontotemporal dementia patient-
900
derived neuronal cell models. Elife 8 (2019).
901
36.
Silva, M.C. & Haggarty, S.J. Tauopathies: Deciphering Disease Mechanisms to Develop Effective
902
Therapies. Int J Mol Sci 21 (2020).
903
37.
Pastuzyn, E.D., et al. The Neuronal Gene Arc Encodes a Repurposed Retrotransposon Gag
904
Protein that Mediates Intercellular RNA Transfer. Cell 172, 275-288 e218 (2018).
905
38.
Ashley, J., et al. Retrovirus-like Gag Protein Arc1 Binds RNA and Traffics across Synaptic Boutons.
906
Cell 172, 262-274 e211 (2018).
907
39.
Aruga, J. & Mikoshiba, K. Identification and characterization of Slitrk, a novel neuronal
908
transmembrane protein family controlling neurite outgrowth. Mol Cell Neurosci 24, 117-129 (2003).
909
40.
Gumireddy, K., et al. Small-molecule inhibitors of microrna miR-21 function. Angew Chem Int Ed
910
Engl 47, 7482-7484 (2008).
911
41.
Connelly, C.M., Boer, R.E., Moon, M.H., Gareiss, P. & Schneekloth, J.S., Jr. Discovery of Inhibitors
912
of MicroRNA-21 Processing Using Small Molecule Microarrays. ACS Chem Biol 12, 435-443 (2017).
913
42.
Young, D.D., Connelly, C.M., Grohmann, C. & Deiters, A. Small molecule modifiers of microRNA
914
miR-122 function for the treatment of hepatitis C virus infection and hepatocellular carcinoma. J Am
915
Chem Soc 132, 7976-7981 (2010).
916
43.
Velagapudi, S.P., et al. Design of a small molecule against an oncogenic noncoding RNA. Proc
917
Natl Acad Sci U S A 113, 5898-5903 (2016).
918
44.
Readhead, B., et al. Expression-based drug screening of neural progenitor cells from individuals
919
with schizophrenia. Nat Commun 9, 4412 (2018).
920
84
45.
Norkin, M., Ordonez-Moran, P. & Huelsken, J. High-content, targeted RNA-seq screening in
921
organoids for drug discovery in colorectal cancer. Cell Rep 35, 109026 (2021).
922
46.
Ye, C., et al. DRUG-seq for miniaturized high-throughput transcriptome profiling in drug
923
discovery. Nat Commun 9, 4307 (2018).
924
47.
Currie, G.M., Wheat, J.M. & Kiat, H. Pharmacokinetic considerations for digoxin in older people.
925
Open Cardiovasc Med J 5, 130-135 (2011).
926
48.
Wang, J.K.T., et al. Cardiac glycosides provide neuroprotection against ischemic stroke:
927
discovery by a brain slice-based compound screening platform. Proc Natl Acad Sci U S A 103, 10461-
928
10466 (2006).
929
49.
Dunn, D.E., et al. In vitro and in vivo neuroprotective activity of the cardiac glycoside oleandrin
930
from Nerium oleander in brain slice-based stroke models. J Neurochem 119, 805-814 (2011).
931
50.
Dvela-Levitt, M., Ami, H.C., Rosen, H., Shohami, E. & Lichtstein, D. Ouabain improves functional
932
recovery following traumatic brain injury. J Neurotrauma 31, 1942-1947 (2014).
933
51.
Kinoshita, P.F., et al. Signaling function of Na,K-ATPase induced by ouabain against LPS as an
934
inflammation model in hippocampus. J Neuroinflammation 11, 218 (2014).
935
52.
Song, H.L., Demirev, A.V., Kim, N.Y., Kim, D.H. & Yoon, S.Y. Ouabain activates transcription factor
936
EB and exerts neuroprotection in models of Alzheimer's disease. Mol Cell Neurosci 95, 13-24 (2019).
937
53.
Mann, C.N., et al. Astrocytic alpha2-Na(+)/K(+) ATPase inhibition suppresses astrocyte reactivity
938
and reduces neurodegeneration in a tauopathy mouse model. Sci Transl Med 14, eabm4107 (2022).
939
54.
Laudisio, A., et al. Digoxin and cognitive performance in patients with heart failure: a cohort,
940
pharmacoepidemiological survey. Drugs Aging 26, 103-112 (2009).
941
55.
Ni, D., et al. Murine pharmacokinetics and metabolism of oleandrin, a cytotoxic component of
942
Nerium oleander. J Exp Ther Oncol 2, 278-285 (2002).
943
56.
Benarroch, E.E. Na+, K+-ATPase: functions in the nervous system and involvement in neurologic
944
disease. Neurology 76, 287-293 (2011).
945
57.
Gupta, R.S., Chopra, A. & Stetsko, D.K. Cellular basis for the species differences in sensitivity to
946
cardiac glycosides (digitalis). J Cell Physiol 127, 197-206 (1986).
947
58.
Stoilov, P., Lin, C.H., Damoiseaux, R., Nikolic, J. & Black, D.L. A high-throughput screening
948
strategy identifies cardiotonic steroids as alternative splicing modulators. Proc Natl Acad Sci U S A 105,
949
11218-11223 (2008).
950
59.
Langmead, B., Trapnell, C., Pop, M. & Salzberg, S.L. Ultrafast and memory-efficient alignment of
951
short DNA sequences to the human genome. Genome Biol 10, R25 (2009).
952
60.
Silva, M.C., et al. Human iPSC-Derived Neuronal Model of Tau-A152T Frontotemporal Dementia
953
Reveals Tau-Mediated Mechanisms of Neuronal Vulnerability. Stem Cell Reports 7, 325-340 (2016).
954
61.
Guo, J.L. & Lee, V.M. Seeding of normal Tau by pathological Tau conformers drives pathogenesis
955
of Alzheimer-like tangles. J Biol Chem 286, 15317-15331 (2011).
956
957
| 2022 | Small Molecule Inducers of Neuroprotective miR-132 Identified by HTS-HTS in Human iPSC-derived Neurons | 10.1101/2022.11.01.514550 | [
"Nguyen Lien D.",
"Wei Zhiyun",
"Silva M. Catarina",
"Barberán-Soler Sergio",
"Rabinovsky Rosalia",
"Muratore Christina R.",
"Stricker Jonathan M. S.",
"Hortman Colin",
"Young-Pearse Tracy L.",
"Haggarty Stephen J.",
"Krichevsky Anna M."
] | null |
Title: Morphological Landscapes from High Content Imaging Identify Optimal Priming Strategies that Enhance
MSC Immunosuppression.
Authors: Seth H. Andrewsab, Matthew W. Klinkerc, Steven R. Bauerc1, Ross A. Markleinab1.
aSchool of Chemical, Materials and Biomedical Engineering, University of Georgia, Athens, GA, USA.
bRegenerative Bioscience Center, University of Georgia, Athens, GA, USA. cDivision of Cellular and Gene
Therapies, Center for Biologics Evaluation and Research, US Food and Drug Administration, Silver Spring,
MD, USA.
1To whom correspondence may be addressed. Email: steven.bauer@fda.hhs.gov and ross.marklein@uga.edu.
Abstract: Successful clinical translation of mesenchymal stromal cell (MSC) products has not been achieved in
the United States and may be in large part due to MSC functional heterogeneity. Efforts have been made to
identify ‘priming’ conditions that produce MSCs with consistent immunomodulatory function; however,
challenges remain with predicting and understanding how priming impacts MSC behavior. The purpose of this
study was to develop a high throughput, image-based approach to assess MSC morphology in response to
combinatorial priming treatments and establish morphological profiling as an effective approach to screen the
effect of manufacturing changes (i.e. priming) on MSC immunomodulation. We characterized the
morphological response of multiple MSC lines/passages to an array of Interferon-gamma (IFN-γ) and Tumor
Necrosis Factor alpha (TNF-⍺) priming conditions, as well as the effects of priming on MSC modulation of
activated T cells and MSC secretome. Although considerable functional heterogeneity, in terms of T cell
suppression, was observed between different MSC lines and at different passages, this heterogeneity was
significantly reduced with combined IFN-γ/TNF-⍺ priming. The magnitude of this change correlated strongly
with multiple morphological features and was also reflected by MSC secretion of immunomodulatory factors
e.g. PGE2, ICAM-1, and CXCL16. Overall, this study further demonstrates the ability of priming to enhance
MSC function, as well as the ability of morphology to better understand MSC heterogeneity and predict
changes in function due to manufacturing.
Keywords: Mesenchymal Stromal Cell, High Content Imaging, immunomodulation, single cell analysis, cell
manufacturing
1. INTRODUCTION
Mesenchymal stromal cells (MSCs) are widely studied as potential treatments for immune-mediated
diseases, such as osteoarthritis, diabetes, multiple sclerosis, and Alzheimer’s Disease1–3 via their modulation of
immune cells such as T cells, B cells, macrophages, microglia, and dendritic cells.1,4,5 This immunomodulation
has been shown to be mediated in vitro through the release of a multitude of immunomodulatory factors (termed
‘secretome’).6 MSCs can be used in an allogeneic, off-the-shelf manner, and have been studied in over 600
clinical trials with most studies reporting a good safety profile7; however, the manufacture of MSCs with
consistent and predictable quality has proven difficult.8 This is due to rare employment of robust quantitative
assays to assess their function, a lack of standardization in manufacturing processes, as well as MSC
heterogeneity.9,10 MSC functional heterogeneity can be attributed to a number of factors, including donor/tissue
sources, extended culture, and varying manufacturing processes.10
Although efforts have been made to develop improved assays for assessing and predicting MSC
function3,11,12, the majority of studies still predominantly characterize MSCs using the ISCT criteria established
in 2006.13 As the current prevailing hypothesis of MSC mechanism of action has centered on their
immunomodulatory capacity, new assays have been proposed that are more functionally relevant in the context
of immune diseases as compared to, for example, trilineage differentiation potential. For example, a number of
immunomodulatory factors (such as IDO and PD-L1)14,15 have been demonstrated to be predictive of MSC
suppression of activated T cells, which is one of the most common in vitro assays for MSC immunomodulatory
activity. However, these assays are typically performed at a population level and cannot determine differences
in MSC heterogeneity that may be present at the single cell level. Furthermore, single metrics may not fully
capture the complex, multifactorial mechanisms of MSC immunomodulation in terms of their secretome and the
diversity of target immune cells, and therefore multiple assays may be required to sufficiently capture MSC
immunomodulatory capacity.11,15 Finally, the mechanisms of action of MSCs are still being explored, adding
further challenge to the development of characterization assays.8
MSC heterogeneity also exists in terms of cell morphology, and MSC morphology has been shown to be
predictive
of
several
therapeutically
relevant
functions
(osteogenesis16,
chondrogenesis17,
and
immunosuppression18,19). As a characterization assay for MSCs, morphological profiling offers several
advantages. First, it can be performed in a rapid, low-cost, and high-throughput manner20,21. Additionally,
morphology is assessed at single-cell resolution, which is essential in recognizing and addressing MSC
heterogeneity. Finally, cell morphology can represent a summation of complex signaling pathways,22–24 possibly
serving as more effective critical quality attributes (CQAs i.e. predictors of quality) than expression of a single
protein or gene.
No standardized approaches exist for manufacturing MSCs, therefore differences in manufacturing
conditions (e.g. culture medium, vessels, cryopreservation) can significantly contribute to functional
heterogeneity.25 An increasingly used strategy to mitigate heterogeneity is to prime (i.e. precondition, pretreat)
MSCs in the presence of different microenvironmental signals to improve their immunomodulatory capacity.26
Examples of these priming signals include inflammatory cytokines such as IFN-γ and TNF-α, hypoxia, and certain
biomaterials and 3D cultures.27 The effects of priming MSCs with inflammatory cytokines has been extensively
studied, and can be assessed based on expression of factors such as IDO.28,29 However, to date studies
examining IFN-γ and TNF-α priming have often been done in a low throughput, binary manner26,30,31, limiting our
understanding of the interplay of different cytokines and dosing on MSC behavior.
The objective of this study was to develop and apply a high throughput morphological screening approach
to identify optimal priming conditions that enhance MSC immunomodulatory function in vitro. We accomplished
this by comprehensively profiling MSC morphological responses to a combinatorial array of cytokine priming
conditions (IFN-γ/TNF-⍺) using high content imaging and automated image analysis of single cell morphological
features (akin to image-based drug screens). IFN-γ/TNF-α priming ‘hits’ from morphological profiling were then
assayed for their T cell suppressive function to identify priming conditions with enhanced immunosuppression
as compared to unprimed MSCs. Finally, we examined MSC secretion of proteins relevant to immune activity
and correlated those with T cell suppression. We identified multiple MSC morphological features that predicted
the effects of IFN-γ and TNF-α priming in terms of enhanced T cell modulation. Additionally, these changes in
morphology and function due to priming were reflected by MSC secretion of cytokines and chemokines
associated with T cell activation. This further establishes morphology as a predictor of MSC function, as well as
presents a generalizable strategy for screening manufacturing conditions to improve the function of promising
cell therapies.
2. METHODS
2.1 MSC Manufacturing
Human bone marrow-derived MSCs were obtained from 4 different donors purchased from Lonza (Walkersville,
MD, USA), AllCells (Emeryville, CA, USA), and RoosterBio (Frederick, MD, USA) (see Table S1 for donor
information). MSC culture conditions for the Lonza and AllCells (8F3560, 110877, PCBM1662) were chosen
based on well-established protocols.32 Briefly, MSCs were continuously expanded in complete MSC growth
medium
(GM)
containing
10%
FBS
(JMBiosciences),
1%
L-glutamine
(Invitrogen),
and
1%
penicillin/streptomycin in alpha-MEM (both Invitrogen) at a seeding density of 60 cells/cm2 in T-175 flasks for a
total of 7 passages with MSCs cryopreserved at passages 3, 5, and 7. Passage 3 (P3) and passage 7 (P7)
MSCs from donors 8F3560, 110877, and PCBM1662 were used in this study. The RoosterBio cell-line RB9 was
expanded using RoosterBio’s recommended protocol, which consisted of seeding 10x106 MSCs in 12 T-225
flasks (3,704 cells/cm2) in RoosterBio growth medium and culturing until 80-90% confluency. RB9 MSCs were
continuously expanded with a portion of the harvested cells at each passage cryopreserved to create a cell bank
with passage 2 and passage 5 RB9 MSCs used in this study. All MSC lines used in this study have been
extensively characterized for their surface marker expression, genomic, epigenetic and proteomic profiles, as
well as performance in multiple functional bioassays.16,18,33–36 All cell-lines presented in this work possessed
viability >95% (based on Trypan Blue exclusion assay) prior to plating for morphological profiling,
immunosuppression assay, and secretomic profiling.
2.2 High Content Imaging, Morphological Profiling, and Morphological Landscapes
Morphological profiling was performed as described in 19 except modified to be performed in a high-throughput
96 well plate format. MSCs from each cell-line/passage experimental group were seeded at a density of 525
cells/cm2 in 96-well plates (Corning) and cultured for 24 hours in GM. GM was replaced with GM containing 64
different IFN-γ and TNF-⍺ (Life Technologies) priming conditions consisting of a full factorial design of 0, 0.5, 1,
2, 5, 10, 20, 50 ng/mL of each cytokine (n=4 replicate wells for each cell-line/passage/priming condition) and
cultured for an additional 24 hours. Following priming, MSCs were fixed with 4% paraformaldehyde for 15
minutes. Cell and nuclear morphology were assessed using 20 µM fluorescein-5-maleimide (Life Technologies)
and 10 µg/mL Hoechst (Sigma-Aldrich), respectively. Samples for morphological analysis were imaged at 10X
with a 6-by-6 stitched image captured for each well using an inverted Nikon Ti-S microscope with automated
stage (Prior) and filters (Chroma Technology). Automated quantification of cellular and nuclear shape features
was performed using CellProfiler v2.2.037 (pipeline available in File S1) to obtain high dimensional single cell
morphological data. An example of a segmented image output from the CellProfiler pipeline can be seen in
Figure S1.
Morphological landscapes (i.e. 3D surface plots) were created for both single morphological features and a
composite overall morphological score, which was created using principal component analysis (PCA). For each
landscape plot (for a given cell-line/passage), the median of each single cell morphological feature for each well
was averaged for quadruplicate wells and plotted for all 64 IFN-γ/TNF-⍺ priming combinations to create a 3D
surface plot using JMP Pro v14. An overall morphological landscape was plotted by first performing PCA on the
high dimensional morphological data on a per well basis with each well consisting of 21-dimensional
morphological data (definitions of each feature available in Table 2) selected based on features from our previous
work.19 Principal component 1 (PC1) was taken to be a composite morphological score for each well and
averaged across wells for all cell-lines/passages/priming conditions to create a 3D surface plot displaying
morphology (PC1) versus [IFN-γ] versus [TNF-⍺].
2.3 Assessment of MSC T cell Suppression
We quantitatively assessed MSC suppression of activated T cells as described in our previous work using a
MSC/PBMC (peripheral blood mononuclear cell) co-culture assay.18,19 For each experiment, MSCs from a given
cell-line/passage were first seeded at 10,000 cells/well in 96-well plates (n=5 wells per cell-line/passage/priming
condition) and cultured for 24 hours in GM. Then, GM was removed and the appropriate priming conditions were
added to replicate wells. Following 24 hours of culture (unprimed and primed), 105 PBMCs derived from a healthy
human donor were stimulated with T cell activating beads (anti-CD3/CD28 Dynabeads, ThermoFisher) at a 1:1
PBMC:bead ratio in each well containing MSCs. Following 3 days of co-culture, PBMCs were collected and their
activation assessed using flow cytometry (MACS-Quant, Miltenyi Biotec). Specifically, CD4+ and CD8+ T cells
were individually assessed by proliferation (CFSE dilution), CD25 expression, and production of IFN-γ and TNF-
α using FlowJo and compensation matrices generated using single-color control samples. All antibodies were
purchased from BioLegend (San Diego, CA) and their information listed in Table S3.
2.4 MSC Secretomic Profiling
Quantitative analysis of MSC secreted proteins was performed using antibody arrays and ELISAs. For all
secretion studies, MSCs from each cell-line at low passage were seeded in 12 well plates at a density of 104
cells/cm2. Following 24 hours of culture in GM, the medium was removed and replaced with different priming
conditions. After 24 hours of priming, the conditioned medium was collected, aliquoted into 1.5 mL
microcentrifuge tubes, and stored at -80 C. Initially, we comprehensively profiled the secretome of one MSC line
(PCBM1662 P3) using a 440-plex antibody array (Quantibody, Raybiotech, Norcross, GA). Frozen aliquots of
conditioned medium and control GM (triplicate samples for each group) were shipped frozen to Raybiotech and
analyzed using their array testing service, a Q440 Multiplex ELISA platform, which quantifies levels of 440
different human chemokines and cytokines. Secreted proteins found to significantly correlate with MSC T cell
suppression for PCBM1662 P3 were then assessed for all cell-lines using ELISAs (Raybiotech) for the following
target proteins: CXCL16, CXCL9, CXCL10, CXCL11, ICAM-1, CCL7, CCL8, CCL13, Legumain, Angiogenin
(ANG), PLGF, DKK1. Secretion of Prostaglandin E2 (PGE2), a well-established MSC immunomodulatory
factor38, was also quantified for each cell-line/priming condition using ELISA (Cayman Chemical, Ann Arbor, MI).
2.5 Statistical Analysis
All statistical tests were performed in GraphPad Prism v8 with the specific tests utilized for each experiment
described in the figure legends.
3. RESULTS
3.1 MSCs exhibit greater morphological response to IFN-γ priming vs TNF-α priming
First, we assessed the effect of priming with either IFN-γ or TNF-α alone on MSC morphology. CellProfiler was
used to measure 93 cell and nuclear morphological features of the MSCs. Selected morphological features for
low passage MSCs are displayed in Figure 1. As little as 10 ng/mL of IFN-γ priming resulted in significant
increases in cell major axis length, perimeter and cell aspect ratio after 24 hours compared to unprimed controls
for nearly all MSC lines at low passage (8F3560 the exception for cell perimeter), although increasing this as
high as 50 ng/mL did not have significant additional effects. On the other hand, cell solidity decreased with IFN-
γ priming (p<0.05) for three of the MSC lines (8F3560 again the exception). TNF-α had a much less pronounced
effect on morphological features with some distinct MSC line/passage responses e.g. decrease in cell aspect
ratio and increase in cell solidity for RB9 with 50 ng/mL TNF-⍺ priming. For high passage MSCs, IFN-γ again
had a more pronounced effect than TNF-α with increased cell major axis length and aspect ratio for all cell-lines
(Figure S2). Cell-line differences were observed in IFN-γ response for cell perimeter and solidity, and some
TNF-α dependent responses were observed e.g. decrease in cell perimeter for 110877 and increase in cell
solidity for RB9. Compared to day 0 (dotted lines, Figure 1), MSCs generally became larger (increased major
axis length, perimeter), more elongated (increased aspect ratio), and more complex (decreased solidity) upon
IFN-γ priming. For unstimulated and TNF-α only primed MSCs there was a notable decrease in major axis length
and perimeter with some cell-line dependent differences observed in terms of aspect ratio or solidity when
compared to day 0.
3.2 Synergistic effects of IFN-γ/TNF-α on MSC morphology
Next, we examined the combined effect of IFN-γ and TNF-α on MSC morphology. Early and late passage bone
marrow-derived MSCs from four donors were primed with 64 different combinations of IFN-γ (0-50 ng/mL) and
TNF-α (0-50 ng/mL) for 24 hours. Selected morphological features for low passage MSCs are displayed in
Figure 2. We observed the same single factor response as in Figure 1, but with noticeable morphological
Figure 1: MSCs possess greater morphological response to IFN-γ only vs TNF-α only. Cell major axis length, cell
perimeter, cell aspect ratio, and cell solidity of 4 MSC lines at low passage with different levels of IFN-γ and TNF-⍺
priming. Reference (dotted) lines show day 0 values for each cell-line. One-way ANOVA with Dunnett’s multiple
comparisons test vs unprimed control within the same cell-line. N = 4 wells per condition, *p < 0.05.
0-0
10-0
50-0
0-10
0-50
0
40
80
120
160
[IFNγ]-[TNF-α] (ng/mL)
Cell Major Axis Length (µm)
*
*
0-0
10-0
50-0
0-10
0-50
0
40
80
120
160
IFNγ-TNFα (ng/mL)
Cell Major Axis Length (µm)
*
*
0-0
10-0
50-0
0-10
0-50
0
40
80
120
160
[IFNγ]-[TNF-α] (ng/mL)
Cell Major Axis Length (µm)
*
*
0-0
10-0
50-0
0-10
0-50
0
40
80
120
160
[IFNγ]-[TNF-α] (ng/mL)
Cell Major Axis Length (µm)
*
*
0-0
10-0
50-0
0-10
0-50
0
100
200
300
400
500
[IFNγ]-[TNF-α] (ng/mL)
Cell Perimeter (µm)
*
*
0-0
10-0
50-0
0-10
0-50
0
100
200
300
400
500
[IFNγ]-[TNF-α] (ng/mL)
Cell Perimeter (µm)
0-0
10-0
50-0
0-10
0-50
0
100
200
300
400
500
[IFNγ]-[TNF-α] (ng/mL)
Cell Perimeter (µm)
*
*
0-0
10-0
50-0
0-10
0-50
0
100
200
300
400
500
[IFNγ]-[TNF-α] (ng/mL)
Cell Perimeter (µm)
*
*
0-0
10-0
50-0
0-10
0-50
1
2
3
4
5
6
[IFNγ]-[TNF-α] (ng/mL)
Cell Aspect Ratio
*
*
0-0
10-0
50-0
0-10
0-50
1
2
3
4
5
6
[IFNγ]-[TNF-α] (ng/mL)
Cell Aspect Ratio
*
*
0-0
10-0
50-0
0-10
0-50
1
2
3
4
5
6
[IFNγ]-[TNF-α] (ng/mL)
Cell Aspect Ratio
*
*
0-0
10-0
50-0
0-10
0-50
1
2
3
4
5
6
[IFNγ]-[TNF-α] (ng/mL)
Cell Aspect Ratio
*
*
*
0-0
10-0
50-0
0-10
0-50
0.0
0.4
0.4
0.6
0.8
1.0
[IFNγ]-[TNF-α] (ng/mL)
Cell Solidity
*
*
0-0
10-0
50-0
0-10
0-50
0.0
0.4
0.4
0.6
0.8
1.0
[IFNγ]-[TNF-α] (ng/mL)
Cell Solidity
0-0
10-0
50-0
0-10
0-50
0.0
0.4
0.4
0.6
0.8
1.0
[IFNγ]-[TNF-α] (ng/mL)
Cell Solidity
*
*
0-0
10-0
50-0
0-10
0-50
0.0
0.4
0.4
0.6
0.8
1.0
[IFNγ]-[TNF-α] (ng/mL)
Cell Solidity
*
*
*
110877
8F3560
PCBM1662
RB9
responses occurring when MSCs were
primed with both IFN-γ and TNF-⍺. For
example, the major axis length and
perimeter of the MSCs tended to
increase
with
increasing
IFN-γ
concentration, with the addition of TNF-
α contributing to a further increase. The
inverse was true for cell form factor and
solidity with both features decreasing
significantly (with a threshold response
observed at 5 ng/mL IFN-γ). While
trends were consistent across MSC
lines, the magnitude of the response
varied considerably. For example, the
perimeter of RB9 increased from ~300
µm to ~500 µm, compared to 110877,
which increased from ~250 µm to ~350
µm. When comparing low versus high
passage MSCs within a line, the
morphological response (in terms of perimeter, for example) followed the same overall trend; however, cell-line
dependent differences were observed due to different baseline/unstimulated morphologies for each cell-line
(Figure S3A). In all cases, the morphological response to priming was apparent at a threshold IFN-γ
concentration of approximately 5 ng/mL and TNF-⍺ concentration of 2 ng/mL. Increased priming past the
threshold with higher concentrations of IFN-γ or TNF-α had a diminishing effect on morphological features.
Overall, IFN-γ priming resulted in the most significant change in morphology, with TNF-α contributing in
a synergistic manner. Interestingly, in the case of cell solidity TNF-α appeared to ‘rescue’ MSCs from their IFN-
γ mediated decrease (Figure 2). We also assessed MSC proliferation during the 24 hours of priming relative to
their number prior to priming (Figure S3B,C). MSC proliferation followed a similar pattern to some of the
Figure 2: Synergistic effects of combined IFN-γ and TNF-α priming on
MSC morphology. Average values of cell major axis length, aspect ratio,
perimeter, solidity, and form factor of 4 MSC lines at low passage with
different levels of IFN-γ and TNF-⍺ priming. N = 4 wells for each IFN-γ/TNF-
⍺ priming condition with at least 200 cells analyzed per well.
morphological features in which it decreased with priming, the effect plateauing at higher concentrations. Like
cell solidity, a similar ‘rescue effect’ was observed for proliferation as IFN-γ priming alone decreased cell number,
but the addition of TNF-⍺ mitigated this loss in cell number (compared to unprimed controls). Most MSC lines
proliferated, although some high passage cell-lines had fewer cells after 24 hours.
3.3 Generation of a morphological landscape to visualize the overall morphological response
Given the high dimensionality of
the morphological data and the
fact that a single feature may not
fully capture the effect of priming,
we performed principal component
analysis
(PCA)
on
the
morphological
data
to
help
visualize
the
overall
MSC
morphological response to priming
(individual feature contributions to
PCA shown in Figure S4). PCA
was performed using the median
value of 21 morphological features
(Table
2)
for
each
MSC
line/passage/priming combination
(512 data points). As the first
principal component (PC1) generated from that analysis accounted for 62% of the variance in the data set, we
used it as a composite measure of the overall MSC morphological response to priming (Figure 3A, B). Similar
to the single morphological features, PC1 increased sharply with IFN-γ and TNF-α concentration before
plateauing at higher concentrations. These changes were primarily in response to IFN-γ treatment, which were
further augmented when combined with TNF-α concentrations at or above 2 ng/mL. Averaging PC1 across all
MSC lines and passages revealed a distinct morphological landscape that effectively summarizes 21-
dimensional morphological data from 4 cell-lines, two passages, and 64 different IFN-γ/TNF-⍺ priming conditions
Figure 3: Morphological landscapes enable visualization of overall
morphological response of MSCs to IFN-γ/TNF-⍺ priming. (A) PC1morphology vs
[TNF-α] and [IFN-γ] at low passage. N = 4 wells per priming condition. (B)
PC1morphology vs [TNF-α] and [IFN-γ] at high passage. N = 4 wells per IFN-γ/TNF-⍺
priming condition. (C) Average PC1morphology vs [TNF-α] and [IFN-γ]. (D) Selected
priming conditions (black boxes) for follow-up experiments summarized in tabular
form.
(Figure 3C). In order to investigate the effect of priming on MSC function, we then chose 10 priming conditions
based on key points within the morphological landscape to follow up on in future experiments (Figure 3D).
3.4 Effects of combinatorial IFN-γ/TNF-⍺ priming on T cell suppression
We then examined whether the different priming conditions identified from our screen had different/variable
effects on the ability of MSCs to suppress activated T cells. T cell activation was measured via proliferation
(%CFSE dilution), CD25 expression, TNF-α expression, and IFN-γ expression of CD4+ and CD8+ T cells that
had been stimulated with anti-CD3/CD28 Dynabeads. Generally, T cell suppression was lower for all MSC lines
at high passage (versus low passage) when MSCs were unprimed most notably for CD8+ T cell suppression
(Figure 4). Although all of these activation parameters were affected by priming, the effect on proliferation was
by far the most significant. While all MSCs suppressed CD4+ and CD8+ T cell proliferation, priming increased
this suppression compared to that of unprimed MSCs.
Figure 4: Effect of priming on MSC suppression of T cell proliferation. A(i) MSC suppression of CD4+ T cell
proliferation as measured by CFSE dilution with TNF-α and IFN-γ priming by cell line and passage. N = 5 wells per priming
condition. A(ii) MSC suppression of CD4+ and CD8+ T cell proliferation as measured by CFSE dilution vs [TNF-α] and [IFN-
γ] averaged across all cell lines and passages. B(i) MSC suppression of CD8+ T cell proliferation as measured by CFSE
dilution with TNF-α and IFN-γ priming by cell line and passage. N = 5 wells per priming condition. B(ii) MSC suppression
of CD4+ and CD8+ T cell proliferation as measured by CFSE dilution vs [TNF-α] and [IFN-γ] averaged across all cell lines
and passages. Reference dotted line represents activated PBMC-only control. Mean +/- SD. One-way ANOVA with Sidak’s
multiple comparisons test. * denotes p < 0.05 vs unprimed control. # denotes p < 0.05 vs 5-0.
Similar to the morphological effects, MSC T cell suppression was primarily affected by IFN-γ priming,
with effects becoming apparent at 5 ng/mL of IFN-γ only (Figure 4, *p<0.05). Conversely, TNF-α priming alone
did not significantly affect MSC suppression of T cell activation. However, with combined IFN-γ and TNF-α
priming, MSCs more effectively suppressed T cell proliferation than when primed by either cytokine alone
(#p<0.05). This effect was particularly pronounced in the case of CD8+ T cell proliferation, in which as little as 5
ng/mL IFN-γ and 2 ng/mL TNF-α priming was sufficient to decrease proliferation significantly more than unprimed
across all MSC lines and passages (p<0.05, Figure 4B(i)). The relationships between IFN-γ and TNF-α priming
and MSC suppression of CD4+ and CD8+ T cell proliferation across all cell-lines/passages is summarized by
Figure 4A(ii) and Figure 4B(ii), respectively. Priming had less pronounced effects on CD25, TNF-α, and IFN-γ
expression, although MSCs did generally suppress each of these measures (Figures S5-S7). Increased priming
of MSCs also decreased the standard deviation of their suppression of both CD4+ and CD8+ T cell proliferation
(i.e.
decreasing
functional
heterogeneity) across MSC lines and
passages (Figure 5) within a priming
condition. The homogeneity in MSC
function that resulted from maximal 50
ng/mL TNF-α + 50 ng/mL IFN-γ priming
was remarkable considering the MSCs
were from different donors/passages.
3.5 MSC morphological response to
priming is correlated with T cell
suppression
As CD4+ and CD8+ T cell proliferation
(as measured by CFSE dilution) had the
most variance across priming conditions
(Table S4), we used these functional
metrics of T cell suppression to correlate
with MSC morphological response to
Figure 5: Decrease in MSC functional heterogeneity with priming. (A)
Mean MSC suppression of CD4+ and CD8+ T cell proliferation as measured
by CFSE dilution with TNF-α and IFN-γ priming across all cell lines and
passages (N = 8). One-way ANOVA with Dunnett’s multiple comparisons
post-hoc test * p < 0.05 different from unprimed control. (B) Standard
deviations of mean MSC suppression of CD4+ and CD8+ T cell proliferation
as measured by CFSE dilution with TNF-α and IFN-γ priming across all cell
lines and passages.
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
CD4+ T Cells
[IFN-γ]-[TNF-α]
%CFSE Diluted
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
25
50
75
100
CD8+ T Cells
[IFN-γ]-[TNF-α]
%CFSE Diluted
* * * *
* * *
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
10
20
30
40
CD4+ CFSE Dilution SD
[IFN-γ]-[TNF-α]
STDEV (%CSE Diluted)
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
10
20
30
CD8+ CFSE Dilution SD
[IFN-γ]-[TNF-α]
STDEV (%CSE Diluted)
A
B
priming. Ten MSC morphological features were significantly correlated with function (Bonferonni-adjusted p-val
cutoff < 3.44x10-6, Table S5). The two morphological features (cell form factor and nuclear major axis length)
that exhibited the strongest correlations with T cell suppression are shown in Figure 6. As cell form factor
decreased (increased priming), CD4+ and CD8+ T cell proliferation decreased (enhanced suppression). On the
other hand, as the nuclear major axis length of MSCs increased (increased priming), proliferation of CD4+ and
CD8+ T cells decreased (enhanced suppression). These correlations effectively encompass our above reported
data for the effects of IFN-γ and TNF-α priming on MSC morphology and immunosuppressive function. The
effects of varied priming conditions on MSC immunosuppression are mirrored by changes in many of their
morphological features. However, there were some inconsistencies, notably PCBM1662 and RB9 at low
passages but this could partially be explained by the fact that these cell-lines had high functional capacity when
unstimulated and priming did not further enhance this function (Figure 4A, CD4+ T cells).
Figure 6: Morphological features correlate with MSC immunosuppression of activated T cells. (A) MSC suppression
of CD4+ and CD8+ T cell proliferation vs MSC cell form factor by MSC line for low (blue) and high (red) passages. Simple
linear regression, N = 10 priming conditions. (B) MSC suppression of CD4+ and CD8+ T cell proliferation vs MSC nuclear
major axis length by MSC line for low and high passages. Simple linear regression, N = 10 priming conditions. Correlation
coefficients for each graph are available in Table S5.
0.10
0.15
0.20
0.25
0
25
50
75
100
125
Cell Form Factor
% CFSE Diluted (CD4)
0.10
0.15
0.20
0.25
0
25
50
75
100
125
Cell Form Factor
% CFSE Diluted (CD4)
0.10
0.15
0.20
0.25
0
25
50
75
100
125
Cell Form Factor
% CFSE Diluted (CD4)
0.10
0.15
0.20
0.25
0
25
50
75
100
125
Cell Form Factor
% CFSE Diluted (CD4)
0.10
0.15
0.20
0.25
0
25
50
75
100
125
Cell Form Factor
%CFSE Diluted (CD8)
0.10
0.15
0.20
0.25
0
25
50
75
100
125
Cell Form Factor
%CFSE Diluted (CD8)
0.10
0.15
0.20
0.25
0
25
50
75
100
125
Cell Form Factor
%CFSE Diluted (CD8)
0.10
0.15
0.20
0.25
0
25
50
75
100
125
Cell Form Factor
%CFSE Diluted (CD8)
16
20
24
28
0
25
50
75
100
125
Nuclear Major Axis Length (µm)
% CFSE Diluted (CD4)
16
20
24
28
0
25
50
75
100
125
Nuclear Major Axis Length (µm)
% CFSE Diluted (CD4)
16
20
24
28
0
25
50
75
100
125
Nuclear Major Axis Length (µm)
% CFSE Diluted (CD4)
16
20
24
28
0
25
50
75
100
125
Nuclear Major Axis Length (µm)
% CFSE Diluted (CD4)
16
20
24
28
0
25
50
75
100
125
Nuclear Major Axis Length (µm)
%CFSE Diluted (CD8)
16
20
24
28
0
25
50
75
100
125
Nuclear Major Axis Length (µm)
%CFSE Diluted (CD8)
16
20
24
28
0
25
50
75
100
125
Nuclear Major Axis Length (µm)
%CFSE Diluted (CD8)
16
20
24
28
0
25
50
75
100
125
Nuclear Major Axis Length (µm)
%CFSE Diluted (CD8)
110877
8F3560
PCBM1662
RB9
A
B
Overall, MSCs responded similarly between cell-lines and passages. In terms of morphology, MSCs
became more spread and complex with priming, as shown by their increase in major axis length for both cell and
nucleus, increased perimeter, and decreased solidity/form factor. This is in line with our previously reported
results18 and has now been further demonstrated to be predictive of the effects of priming on MSC
immunomodulation.
3.6 Response of MSC secretome to priming is correlated with T cell suppression
To better understand possible mechanisms of action for MSC-mediated T cell suppression, we profiled the
secretome of MSCs primed with different combinations of IFN-γ/TNF-⍺. To identify target proteins, the
conditioned media from one MSC line at low passage (PCBM1662 P3) was screened for 440 proteins after being
primed using the 10 priming conditions identified from the morphological screen (Figure 3D). Proteins detected
at concentrations both above the limit of detection and the media only control are listed in Table S6.
Unsupervised hierarchical clustering performed on all priming conditions (Figure 7A) using a subset of proteins
secreted at levels significantly higher than control medium (132 total) resulted in unprimed MSCs clustered at
the top and maximally primed (50 ng/mL of both IFN-γ and TNF-⍺) clustered at the bottom of the heatmap.
Following this, we correlated each secreted factor with T cell suppression for a given priming condition to
determine whether any secreted factors could predict MSC function. 40 factors were significantly correlated with
suppression in terms of CD8+ T cell proliferation (Table S7), which was again selected due to the highest
observed CV (Table S4). From these, 12 of the factors with the highest correlation coefficients were selected as
targets for performing follow-up ELISAs on all 4 MSC lines at low passage.
Principal component analysis was performed on the secretory profiles generated from the ELISA follow-
up (Figure 7B). Viewing the first and second principal components shows that PC1secretion tracks similarly with
MSC function (colored by magnitude of CD8+ T cell proliferation). Plotting PC1secretion vs TNF-α and IFN-γ priming
conditions reveals a similar landscape pattern (Figure 7C) to the morphological and functional landscapes
shown in Figure 3C and Figure 4, respectively.
Furthermore, we correlated the secretion of these proteins with MSC suppression of CD8+ T cell
proliferation and found strong relationships within MSC lines (Figure 7D). Most secreted proteins (CXCL9,
CXCL10, CXCL11, CCL7, CCL8, CCL13, CXCL16, ICAM-1, Legumain, and ANG) increased with priming and
were positively correlated with T cell suppression (lower %CFSE dilution). Conversely, PLGF and DKK1 were
secreted at lower levels following priming and were thus negatively correlated with T cell suppression. The ELISA
follow-up served to validate the initial secretome screen with additional cell-lines, as well as revealing some
notable MSC line dependent correlations in the cases of CCL7, CCL8, CCL13, ANG, PLGF and DKK1. Secretion
of other proteins, such as CXCL16, ICAM-1, CXCL9, CXCL10, CXCL11, and Legumain were shown to be MSC
line independent in terms of their correlation with T cell suppression. Additionally, PGE2 (while not included in
the initial screening array) was included in the ELISA follow-up due to its previously described involvement with
Figure 7: MSC secretome screening and correlation with T cell suppression. A) heat map of protein screen of
PCBM1662 P3 conditioned media from 10 different priming conditions selected from morphological screen. N = 3 samples
per priming condition. B) Principal Component Analysis of protein levels in MSC conditioned media from ELISA follow-up
(12 proteins total) colored by CD8+ T cell proliferation. N = 40 (4 MSC lines x 10 priming conditions). C) PC1secretion vs [TNF-
α] and [IFN-γ]. N = 40. D) One phase decay, nonlinear fit of MSC suppression of CD8+ T cell proliferation vs levels of
secreted proteins. N = 10 priming conditions per MSC line. E) One phase decay, nonlinear fit of MSC suppression of CD8+
T cell proliferation vs secreted PGE2. N = 10 priming conditions per MSC line.
MSC T cell suppression.39 It was found to correlate well with function independent of MSC line and increased
with priming (Figure 7E).
4. DISCUSSION
This study explores in detail the response of MSCs to combinatorial TNF-α and IFN-γ priming and exemplifies
the use of MSC morphology to screen for manufacturing conditions that enhance function. While inflammatory
cytokine priming with TNF-α, IFN-γ (as well as other cytokines such as IL-1β) is a well-known method to improve
MSC immunomodulatory capabilities, many of these studies have investigated only binary priming conditions
(i.e. -/+ priming signals).18,28,31,40,41 Our work here encompasses four MSC lines (from three different commercial
sources) at multiple passages and 10 priming conditions, with their functional effects (T cell suppression) being
evaluated by eight different flow cytometry outcomes.
This comprehensive approach allowed us to identify a reduction of MSC heterogeneity across cell-lines
and passages with priming. MSC functional heterogeneity can be attributed to a number of factors: different
donor/tissue sources, extended culture, and manufacturing methods10. For example, a comparison of umbilical-
, bone marrow-, and adipose tissue-derived MSCs found that adipose MSCs were better able to suppress the
activation of PHA stimulated CD4+ and CD8+ T cells42. MSCs derived from the same tissue source but different
donors can exhibit markedly different responses to IFN-γ as determined by production of the immunomodulatory
enzyme IDO.43 Additionally, bone marrow-derived MSCs have been reported to have decreased secretion of
immunomodulatory cytokines IL-6, IL-8, and RANTES with increased passage44. Similarly, our group has shown
a decrease in the ability of MSCs to suppress the activation of CD4+ and CD8+ T cells with passaging.18,19
Functional heterogeneity of clonal MSC cultures has also been reported; however, much higher concentrations
of combined IFN-γ and TNF-α priming were used to enhance immunomodulatory function and mitigate the
observed heterogeneity31. The ability to reduce MSC heterogeneity – in terms of not only morphology, but also
immunomodulatory function and secreted factors - opens potential new manufacturing approaches in which
poorly performing MSC lines can be improved and their function effectively ‘rescued.’
This work provides a foundation for future studies to assess the effects of different priming conditions
and predict MSC immunomodulation based on single cell responses. Previous studies attempting to screen MSC
immunosuppressive function have utilized several approaches. Chinnadurai et al examined MSC secretome and
RNA content as indicators of MSC suppression of CD3+ T cell proliferation.15 They found strong correlations
between a number of cytokines in the media of PBMC-MSC cocultures and the observed MSC-mediated T cell
suppression. Specifically, CXCL9 and CXCL10 were upregulated (both in terms of secreted protein levels and
mRNA expression in cocultured and IFN-γ primed MSCs) and correlated with MSC suppression of T cells, which
also was the case in our study. In another study, small molecules were screened for their ability to prime MSCs
towards an immunosuppressive phenotype 45 using secretion of PGE2 as their target. Identified hits were
followed up by examining how primed MSCs attenuated TNF-α secretion, first by macrophages in vitro, and
finally in a mouse ear skin inflammation model. These approaches provide valuable information; however, they
assess MSCs on a population level, while morphological profiling allows for single-cell resolution and potential
identification of MSC functional subpopulations46. Additionally, relying on one functional measure (such as CD3+
T cell proliferation) or one analyte (such as PGE2) does not fully capture MSC multipotency i.e. their ability to
modulate different immune cells and exert functions through multiple mechanisms of action11.
Here we demonstrated that priming MSCs with IFN-γ alone versus TNF-⍺ alone induced a more
significant response in terms of MSC morphology, T cell suppression, and secretion. Additionally, while TNF-α
alone did not have a significant effect on MSC behavior, it did act synergistically with IFN-γ. Most studies
involving MSC priming have examined the effects of either one cytokine or a combination of two cytokines, rather
than a full factorial study examining different doses.28 The effects of binary priming MSCs with TNF-α and IFN-γ
on immunomodulation was also investigated using MSCs from both bone marrow and Wharton’s jelly tissue
sources.47 They found considerable heterogeneity between MSC donors and tissue sources in terms of the effect
of TNF-α and IFN-γ priming on MSC suppression of PBMC proliferation, but did note that MSCs tended to
become qualitatively larger and flatter with IFN-γ priming and more spindle-shaped upon exposure to TNF-α.
This is consistent with our finding that MSC cell perimeter and major axis length increased with exposure to IFN-
γ, but does not agree with our observations of TNF-α alone priming (Figure 1). In another study, Li et al reported
that TNF-α priming had much greater effects than IFN-γ priming on MSC ability to suppress T cell proliferation,
but also noted synergy when the two were used in combination 48. It is important to note that the priming in this
referenced study was done concurrently with T cell coculture while priming in our study was done prior to
coculture (i.e. pretreatment or preconditioning). Synergistic effects of 8 hour TNF-α/IFN-γ priming on MSC
suppression of T cell proliferation has also been reported 49. Timing and duration of priming varies considerably
between studies with demonstrated effects on MSC function, thus limiting the comparisons that can be made
between studies and further emphasizing the critical need for standardization of MSC characterization assays.29
The MSC secretome has been implicated as the primary mechanism by which MSCs exert their
immunomodulatory effects. Chinnadurai et al found significant correlations between MSC function as measured
by cocultured T cell proliferation and their secretion of various cytokines and morphogens.15 Interestingly, they
reported CXCL9 to have the same relationship with T cell proliferation i.e. reduced T cell proliferation/activation
with increased secretion. CXCL9, CXCL10, CXCL11, CXCL16, and CCL8 are known chemokines for T cells50–
54. Furthermore, CXCL9, CXCL10, and CXCL11 all bind CXCR3, which is primarily expressed on T cells and NK
cells50,52,55–58. These chemokines may recruit activated immune cells to be locally modulated by other secreted
or cell contact-mediated factors. CXCR3 is involved in regulatory T cell recruitment and migration as well, which
could be another possible avenue for MSC immunomodulation.59–61
Secretion of DKK1, which inversely correlated with T cell suppression, is an inhibitor of canonical Wnt
signalling62. PGE2, on the other hand, was positively correlated with MSC T cell suppression, and is a known
activator of canonical Wnt signaling63. The Wnt pathway is a potent regulator of cell differentiation, growth, and
migration, and its activation by primed MSCs may have far-ranging effects on the immune system. There is
considerable evidence supporting PGE2 as an effector of MSC immunomodulation64 as it can suppress T cell
activation and proliferation, and its secretion has been shown to be upregulated synergistically by IFN-γ/TNF-α
primed MSCs, which is further supported by our results.39,65 Additionally, levels of PGE2 secretion have been
shown to be predictive of MSC effects in a rat model of traumatic brain injury38. However, it is likely that MSCs
exert their effects through multiple pathways and target cells, and the assessment of a single factor may not fully
reflect MSC multipotency.11
The sensitivity, low cost and single-cell resolution of cell morphology could also be applied to many
aspects of MSC manufacturing in order to detect and predict functional changes. For example, MSCs can
respond to hypoxia or additional cytokines (e.g. IL-1β) through enhanced secretion of immunomodulatory factors
and altered migratory capacity and therefore may exhibit distinct morphological responses to microenvironment
signals besides IFN-γ and TNF-⍺.66–68 Additionally, our approach could be used to assess the impact of different
manufacturing methods on MSC immunomodulation. Cell culture substrates and biomaterials can be tuned to
direct MSC function and morphological profiling could be adapted to screen for biomaterial systems that further
enhance MSC function.69–72 It is well established that MSCs lose function and become senescent over the course
of ex vivo expansion and identification of soluble cues to include in defined growth medium could be another
application of this morphology-based approach.73,74 Additionally, cryopreserved MSCs undergo a recovery period
post-thaw, during which their function is impaired.40 Any observed differences in MSC immunosuppression
caused by changes in these manufacturing methods could be assessed by the techniques described here.
Beyond screening for optimal MSC manufacturing conditions, morphological profiling could be used to
assess manufacturing reagent batch-to-batch variability, which is an often overlooked challenge associated with
cell manufacturing.74–76 Given that the MSC response to IFN-γ is consistent and predictable in this study (as well
as in other studies), morphological profiling could be useful as a tool to assess the bioactivity of different batches
of recombinant IFN-γ, TNF-⍺, or other priming factors. Culture medium used for MSC expansion represents an
enormous source of variability when considering differences in supplement source (e.g. FBS vs. platelet lysate)
and defined medium components (growth factors and small molecules).74 The effects of manufacturing changes
on MSCs can be difficult to assess because there are no well-established CQAs associated with relevant MSC
functions. This issue becomes costlier and more difficult to address as manufacturers advance in clinical
development and often have to make significant changes (e.g. due to scaling or new vendor-sourced reagents)
prior to performing Phase 3 studies and submitting a Biologics License Application. As we have demonstrated
the ability of MSC morphology to predict reduction in functional heterogeneity with priming (Figure 5), we
anticipate this approach could be similarly applied to reduce (and predict) functional heterogeneity derived from
different media sources/compositions.
In summary, we have demonstrated that MSCs exhibit remarkably consistent morphological responses
following IFN-γ priming that can be further enhanced, in a synergistic manner, with the addition of TNF-α. These
morphological changes are strongly correlated with MSC immunosuppressive function, which in turn is reflected
by secretion of chemokines and other immunomodulatory factors associated with T cell activation and migration.
The morphological profiling approach presented in this work could be adapted and applied to improve
manufacturing of other cell-types and cell-derived products (e.g. extracellular vesicles), and further explored as
a means to better understand and control MSC functional heterogeneity.
5. Acknowledgments
The authors thank Drs. Zhaohui Ye, Saravanan Karumbayaram and Nirjal Bhattarai for their review of the
manuscript, and Drs. Jessica Lo Surdo, Johnny Lam, and Eva Rudikoff for technical assistance in manufacturing
the MSC lines used in this study. S.H.A. was supported by startup funds provided by the UGA College of
Engineering to R.A.M. administered by the UGA Office of the Vice Provost of Research (OVPR). M.W.K. was
supported in part by appointment to the Research Participation Program at the Center for Biologics Evaluation
and Research (CBER) administered by the Oak Ridge Institute for Science and Education through the US
Department of Energy and the US Food and Drug Administration (FDA). This work was also supported in part
by the FDA Modernizing Science grant program, a Biomedical Advanced Research and Development Authority
(BARDA) grant, and grant from the Medical Countermeasures Initiative and research funds from the Division of
Cell and Gene Therapies.
Disclosure of interests: S.H.A., M.W.K., S.R.B., and R.A.M have no commercial, proprietary or financial
interest in the products or companies described in this article.
REFERENCES
1.
Galieva LR, James V, Mukhamedshina YO, Rizvanov AA. Therapeutic Potential of Extracellular
Vesicles for the Treatment of Nerve Disorders. Front Neurosci. 2019;13(March):1–9.
2.
Galipeau J, Sensébé L. Mesenchymal Stromal Cells: Clinical Challenges and Therapeutic Opportunities.
Cell Stem Cell. 2018;22(6):824–33.
3.
Heathman TR, Nienow AW, McCall MJ, Coopman K, Kara B, Hewitt CJ. The translation of cell-based
therapies: Clinical landscape and manufacturing challenges. Vol. 10, Regenerative Medicine. 2015. p.
49–64.
4.
Kusuma GD, Carthew J, Lim R, Frith JE. Effect of the Microenvironment on Mesenchymal Stem Cell
Paracrine Signaling: Opportunities to Engineer the Therapeutic Effect. Stem Cells Dev. 2017;26(9):617–
31.
5.
Di Trapani M, Bassi G, Midolo M, Gatti A, Kamga PT, Cassaro A, et al. Differential and transferable
modulatory effects of mesenchymal stromal cell-derived extracellular vesicles on T, B and NK cell
functions. Sci Rep. 2016;6(December 2015):1–13.
6.
Prockop DJ. The exciting prospects of new therapies with mesenchymal stromal cells. Vol. 19,
Cytotherapy. Elsevier B.V.; 2017. p. 1–8.
7.
Olsen TR, Ng KS, Lock LT, Ahsan T, Rowley JA. Peak MSC-Are we there yet? Front Med.
2018;5(JUN).
8.
Gao F, Chiu SM, Motan DAL, Zhang Z, Chen L, Ji HL, et al. Mesenchymal stem cells and
immunomodulation: Current status and future prospects. Cell Death Dis. 2016;7(1).
9.
Phinney DG. Functional heterogeneity of mesenchymal stem cells: Implications for cell therapy. Vol.
113, Journal of Cellular Biochemistry. John Wiley & Sons, Ltd; 2012. p. 2806–12.
10.
Wilson A, Hodgson-Garms M, Frith JE, Genever P. Multiplicity of mesenchymal stromal cells: Finding
the right route to therapy. Vol. 10, Frontiers in Immunology. Frontiers Media S.A.; 2019. p. 1112.
11.
Galipeau J, Krampera M, Barrett J, Dazzi F, Deans RJ, DeBruijn J, et al. International Society for
Cellular Therapy perspective on immune functional assays for mesenchymal stromal cells as potency
release criterion for advanced phase clinical trials. Cytotherapy. 2016 Feb 17;18(2):151–9.
12.
de Wolf C, van de Bovenkamp M, Hoefnagel M. Regulatory perspective on in vitro potency assays for
human mesenchymal stromal cells used in immunotherapy. Cytotherapy. 2017;19(7):784–97.
13.
Dominici M, Le Blanc K, Mueller I, Slaper-Cortenbach I, Marini F, Krause D, et al. Minimal criteria for
defining multipotent mesenchymal stromal cells. The International Society for Cellular Therapy position
statement. Cytotherapy. 2006;8(4):315–7.
14.
Guan Q, Li Y, Shpiruk T, Bhagwat S, Wall DA. Inducible indoleamine 2,3-dioxygenase 1 and
programmed death ligand 1 expression as the potency marker for mesenchymal stromal cells.
Cytotherapy. 2018;
15.
Chinnadurai R, Rajan D, Qayed M, Arafat D, Garcia M, Liu Y, et al. Potency Analysis of Mesenchymal
Stromal Cells Using a Combinatorial Assay Matrix Approach. Cell Rep. 2018;22(9):2504–17.
16.
Marklein RA, Lo Surdo JL, Bellayr IH, Godil SA, Puri RK, Bauer SR. High Content Imaging of Early
Morphological Signatures Predicts Long Term Mineralization Capacity of Human Mesenchymal Stem
Cells upon Osteogenic Induction. Stem Cells. 2016;34(4):935–47.
17.
Lam J, Bellayr IH, Marklein RA, Bauer SR, Puri RK, Sung KE. Functional Profiling of Chondrogenically
Induced Multipotent Stromal Cell Aggregates Reveals Transcriptomic and Emergent Morphological
Phenotypes Predictive of Differentiation Capacity. Stem Cells Transl Med. 2018;7(9):664–75.
18.
Klinker MW, Marklein RA, Surdo JL Lo, Wei C-H, Bauer SR. Morphological features of IFN-γ–stimulated
mesenchymal stromal cells predict overall immunosuppressive capacity. Proc Natl Acad Sci.
2017;114(13):E2598–607.
19.
Marklein RA, Klinker MW, Drake KA, Polikowsky HG, Lessey-Morillon EC, Bauer SR. Morphological
profiling using machine learning reveals emergent subpopulations of interferon-γ–stimulated
mesenchymal stromal cells that predict immunosuppression. Cytotherapy. 2019;21(1):17–31.
20.
Carmen J, Burger SR, McCaman M, Rowley JA. Developing assays to address identity, potency, purity
and safety: Cell characterization in cell therapy process development. Regen Med. 2012;7(1):85–100.
21.
Singh S, Carpenter AE, Genovesio A. Increasing the content of high-content screening: An overview. J
Biomol Screen. 2014;19(5):640–50.
22.
Khetan S, Guvendiren M, Legant WR, Cohen DM, Chen CS, Burdick JA, et al. Degradation-mediated
cellular traction directs stem cell fate in covalently crosslinked three-dimensional hydrogels HHS Public
Access Author manuscript. Nat Mater. 2013;12(5):458–65.
23.
Huebsch N, Arany PR, Mao AS, Shvartsman D, Ali OA, Bencherif SA, et al. Harnessing traction-
mediated manipulation of the cell/matrix interface to control stem-cell fate. Nat Mater. 2010;9(6):518–26.
24.
Vega SL, Liu E, Patel PJ, Kulesa AB, Carlson AL, Ma Y, et al. High-content imaging-based screening of
microenvironment-induced changes to stem cells. J Biomol Screen. 2012;17(9):1151–62.
25.
Phinney DG, Galipeau J. Manufacturing mesenchymal stromal cells for clinical applications: A survey of
Good Manufacturing Practices at U.S. academic centers. Cytotherapy. 2019 Jul 1;21(7):782–92.
26.
Kouroupis D, Sanjurjo-Rodriguez C, Jones E, Correa D. Mesenchymal Stem Cell Functionalization for
Enhanced Therapeutic Applications. Tissue Eng - Part B Rev. 2019;25(1):55–77.
27.
Ferreira JR, Teixeira GQ, Santos SG, Barbosa MA, Almeida-Porada G, Gonçalves RM. Mesenchymal
Stromal Cell Secretome: Influencing Therapeutic Potential by Cellular Pre-conditioning. Front Immunol.
2018 Dec 4;9(December):2837.
28.
Noronha Nc NDC, Mizukami A, Caliári-Oliveira C, Cominal JG, Rocha JLM, Covas DT, et al. Priming
approaches to improve the efficacy of mesenchymal stromal cell-based therapies. Vol. 10, Stem Cell
Research and Therapy. BioMed Central Ltd.; 2019. p. 131.
29.
Boyt DT, Boland LK, Burand AJ, Brown AJ, Ankrum JA. Dose and duration of interferon γ pre-licensing
interact with donor characteristics to influence the expression and function of indoleamine-2,3-
dioxygenase in mesenchymal stromal cells. J R Soc Interface. 2020 Jun 17;17(167):20190815.
30.
Wobma HM, Kanai M, Ma SP, Shih Y, Li HW, Duran-Struuck R, et al. Dual IFN-γ/hypoxia priming
enhances immunosuppression of mesenchymal stromal cells through regulatory proteins and metabolic
mechanisms. J Immunol Regen Med. 2018;
31.
Szabó E, Fajka-Boja R, Kriston-Pál É, Hornung Á, Makra I, Kudlik G, et al. Licensing by Inflammatory
Cytokines Abolishes Heterogeneity of Immunosuppressive Function of Mesenchymal Stem Cell
Population. Stem Cells Dev. 2015;24(18):2171–80.
32.
Prockop DJ, Phinney DG, Bunnell BA. Mesenchymal Stem Cells: Methods and Protocols. Vol. 449,
Methods in Molecular Biology. 2008.
33.
Bellayr IH, Marklein RA, Lo Surdo JL, Bauer SR, Puri RK. Identification of Predictive Gene Markers for
Multipotent Stromal Cell Proliferation. Stem Cells Dev. 2016;25(11):861–73.
34.
Lo Surdo JL, Millis BA, Bauer SR. Automated microscopy as a quantitative method to measure
differences in adipogenic differentiation in preparations of human mesenchymal stromal cells.
Cytotherapy. 2013;15(12):1527–40.
35.
Lynch PJ, Thompson EE, McGinnis K, Rovira Gonzalez YI, Lo Surdo J, Bauer SR, et al. Chromatin
changes at the PPAR-γ2 promoter during bone marrow-derived multipotent stromal cell culture correlate
with loss of gene activation potential. Stem Cells. 2015;33(7):2169–81.
36.
Mindaye ST, Ra M, Lo Surdo JL, Bauer SR, Alterman MA. Global proteomic signature of
undifferentiated human bone marrow stromal cells: Evidence for donor-to-donor proteome
heterogeneity. Stem Cell Res. 2013;11(2):793–805.
37.
Carpenter AE, Jones TR, Lamprecht MR, Clarke C, Kang IH, Friman O, et al. CellProfiler: Image
analysis software for identifying and quantifying cell phenotypes. Genome Biol. 2006;7(10).
38.
Kota DJ, Prabhakara KS, Toledano-Furman N, Bhattarai D, Chen Q, DiCarlo B, et al. Prostaglandin E2
Indicates Therapeutic Efficacy of Mesenchymal Stem Cells in Experimental Traumatic Brain Injury. Stem
Cells. 2017;35(5):1416–30.
39.
Zafranskaya M, Nizheharodava D, Yurkevich M, Ivanchik G, Demidchik Y, Kozhukh H, et al. PGE 2
Contributes to In vitro MSC-Mediated Inhibition of Non-Specific and Antigen-Specific T Cell Proliferation
in MS Patients. Scand J Immunol. 2013 Nov;78(5):455–62.
40.
François M, Copland IB, Yuan S, Romieu-Mourez R, Waller EK, Galipeau J. Cryopreserved
mesenchymal stromal cells display impaired immunosuppressive properties as a result of heat-shock
response and impaired interferon-γ licensing. Cytotherapy. 2012;14(2):147–52.
41.
Zimmermann JA, Mcdevitt TC. Pre-conditioning mesenchymal stromal cell spheroids for
immunomodulatory paracrine factor secretion. Cytotherapy. 2014;16(3):331–45.
42.
Ribeiro A, Laranjeira P, Mendes S, Velada I, Leite C, Andrade P, et al. Mesenchymal stem cells from
umbilical cord matrix, adipose tissue and bone marrow exhibit different capability to suppress peripheral
blood B, natural killer and T cells. Stem Cell Res Ther. 2013 Oct 15;4(5):125.
43.
François M, Romieu-Mourez R, Li M, Galipeau J. Human MSC suppression correlates with cytokine
induction of indoleamine 2,3-dioxygenase and bystander M2 macrophage differentiation. Mol Ther.
2012;20(1):187–95.
44.
Russell AL, Lefavor R, Durand N, Glover L, Zubair AC. Modifiers of mesenchymal stem cell quantity and
quality. Transfusion. 2018 Jun 1;58(6):1434–40.
45.
Yang Z, Concannon J, Ng KS, Seyb K, Mortensen LJ, Ranganath S, et al. Tetrandrine identified in a
small molecule screen to activate mesenchymal stem cells for enhanced immunomodulation. Sci Rep.
2016;6(July):1–10.
46.
Marklein RA, Lam J, Guvendiren M, Sung KE, Bauer SR. Functionally-Relevant Morphological Profiling:
A Tool to Assess Cellular Heterogeneity. Trends Biotechnol. 2018;36(1):105–18.
47.
Prasanna SJ, Gopalakrishnan D, Shankar SR, Vasandan AB. Pro-inflammatory cytokines, IFNγ and
TNFα, influence immune properties of human bone marrow and Wharton jelly mesenchymal stem cells
differentially. PLoS One. 2010;5(2).
48.
Li W, Ren G, Huang Y, Su J, Han Y, Li J, et al. Mesenchymal stem cells: A double-edged sword in
regulating immune responses. Cell Death Differ. 2012;19(9):1505–13.
49.
Ren G, Zhang L, Zhao X, Xu G, Zhang Y, Roberts AI, et al. Mesenchymal Stem Cell-Mediated
Immunosuppression Occurs via Concerted Action of Chemokines and Nitric Oxide. Cell Stem Cell. 2008
Feb 7;2(2):141–50.
50.
Schoenborn JR, Wilson CB. Regulation of Interferon-γ During Innate and Adaptive Immune Responses.
Vol. 96, Advances in Immunology. Adv Immunol; 2007. p. 41–101.
51.
Dufour JH, Dziejman M, Liu MT, Leung JH, Lane TE, Luster AD. IFN-γ-Inducible Protein 10 (IP-10;
CXCL10)-Deficient Mice Reveal a Role for IP-10 in Effector T Cell Generation and Trafficking. J
Immunol. 2002 Apr 1;168(7):3195–204.
52.
Cole KE, Strick CA, Paradis TJ, Ogborne KT, Loetscher M, Gladue RP, et al. Interferon-inducible T cell
alpha chemoattractant (I-TAC): A novel non- ELR CXC chemokine with potent activity on activated T
cells through selective high affinity binding to CXCR3. J Exp Med. 1998 Jun 15;187(12):2009–21.
53.
Matloubian M, David A, Engel S, Ryan JE, Cyster JG. A transmembrane CXC chemokine is a ligand for
HIV-coreceptor Bonzo. Nat Immunol. 2000;1(4):298–304.
54.
Proost P, Wuyts A, Van Damme J. Human monocyte chemotactic proteins-2 and -3: Structural and
functional comparison with MCP-1. Vol. 59, Journal of Leukocyte Biology. Federation of American
Societies for Experimental Biology; 1996. p. 67–74.
55.
Booth V, Keizer DW, Kamphuis MB, Clark-Lewis I, Sykes BD. The CXCR3 binding chemokine IP-
10/CXCL10: Structure and receptor interactions. Biochemistry. 2002 Aug 20;41(33):10418–25.
56.
Blanpain C, Migeotte I, Lee B, Vakili J, Doranz BJ, Govaerts C, et al. CCR5 binds multiple CC-
chemokines: MCP-3 acts as a natural antagonist. Blood. 1999 Sep;94(6):1899–905.
57.
Ge B, Li J, Wei Z, Sun T, Song Y, Khan NU. Functional expression of CCL8 and its interaction with
chemokine receptor CCR3. BMC Immunol. 2017 Dec 28;18(1):54.
58.
Qin S, Rottman JB, Myers P, Kassam N, Weinblatt M, Loetscher M, et al. The chemokine receptors
CXCR3 and CCR5 mark subsets of T cells associated with certain inflammatory reactions. J Clin Invest.
1998 Feb 15;101(4):746–54.
59.
Paust HJ, Riedel JH, Krebs CF, Turner JE, Brix SR, Krohn S, et al. CXCR3+ regulatory T cells control
TH1 responses in crescentic GN. J Am Soc Nephrol. 2016 Jul 1;27(7):1933–42.
60.
Oo YH, Weston CJ, Lalor PF, Curbishley SM, Withers DR, Reynolds GM, et al. Distinct Roles for CCR4
and CXCR3 in the Recruitment and Positioning of Regulatory T Cells in the Inflamed Human Liver. J
Immunol. 2010 Mar 15;184(6):2886–98.
61.
Andrews SH, Maughon T, Marklein RA, Stice SL. Priming of MSCs with inflammation-relevant signals
affects extracellular vesicle biogenesis , surface markers , and modulation of T cell subsets. bioRxiv.
2020;
62.
Huang Y, Liu L, Liu A. Dickkopf-1: Current knowledge and related diseases. Vol. 209, Life Sciences.
Elsevier Inc.; 2018. p. 249–54.
63.
Goessling W, North TE, Loewer S, Lord AM, Lee S, Stoick-Cooper CL, et al. Genetic Interaction of
PGE2 and Wnt Signaling Regulates Developmental Specification of Stem Cells and Regeneration. Cell.
2009 Mar 20;136(6):1136–47.
64.
Najar M, Raicevic G, Boufker HI, Kazan HF, Bruyn C De, Meuleman N, et al. Mesenchymal stromal cells
use PGE2 to modulate activation and proliferation of lymphocyte subsets: Combined comparison of
adipose tissue, Wharton’s Jelly and bone marrow sources. Cell Immunol. 2010 Jan 1;264(2):171–9.
65.
Hegyi B, Kudlik G, Monostori É, Uher F. Activated T-cells and pro-inflammatory cytokines differentially
regulate prostaglandin E2 secretion by mesenchymal stem cells. Biochem Biophys Res Commun. 2012
Mar 9;419(2):215–20.
66.
Liu L, Gao J, Yuan Y, Chang Q, Liao Y, Lu F. Hypoxia preconditioned human adipose derived
mesenchymal stem cells enhance angiogenic potential via secretion of increased VEGF and bFGF. Cell
Biol Int. 2013;37(6):551–60.
67.
Huang C, Luo W-F, Ye Y-F, Lin L, Wang Z, Luo M-H, et al. Characterization of inflammatory factor-
induced changes in mesenchymal stem cell exosomes and sequencing analysis of exosomal
microRNAs. Vol. 11, World Journal of Stem Cells. 2019. 859–890 p.
68.
Fan H, Zhao G, Liu L, Liu F, Gong W, Liu X, et al. Pre-treatment with IL-1β enhances the efficacy of
MSC transplantation in DSS-induced colitis. Cell Mol Immunol. 2012 Nov;9(6):473–81.
69.
Ogle ME, Doron G, Levy MJ, Temenoff JS. Hydrogel Culture Surface Stiffness Modulates Mesenchymal
Stromal Cell Secretome and Alters Senescence. Tissue Eng Part A. 2020;1–13.
70.
Lei J, Trevino E, Temenoff J. Cell number and chondrogenesis in human mesenchymal stem cell
aggregates is affected by the sulfation level of heparin used as a cell coating. J Biomed Mater Res - Part
A. 2016 Jul 1;104(7):1817–29.
71.
Hansen TD, Koepsel JT, Le NN, Nguyen EH, Zorn S, Parlato M, et al. Biomaterial arrays with defined
adhesion ligand densities and matrix stiffness identify distinct phenotypes for tumorigenic and non-
tumorigenic human mesenchymal cell types. Biomater Sci. 2014;2(5):745–56.
72.
Solorio LD, Phillips LM, Mcmillan A, Cheng CW, Dang PN, Samorezov JE, et al. Spatially Organized
Differentiation of Mesenchymal Stem Cells within Biphasic Microparticle-Incorporated High Cell Density
Osteochondral Tissues. Adv Healthc Mater. 2015 Oct 1;4(15):2306–13.
73.
Wang J, Geesman GJ, Hostikka SL, Atallah M, Blackwell B, Lee E, et al. Inhibition of activated
pericentromeric SINE/Alu repeat transcription in senescent human adult stem cells reinstates self-
renewal. Cell Cycle. 2011 Sep 1;10(17):3016–30.
74.
Karnieli O, Friedner OM, Allickson JG, Zhang N, Jung S, Fiorentini D, et al. A consensus introduction to
serum replacements and serum-free media for cellular therapies. Cytotherapy. 2017;19(2):155–69.
75.
Lipsitz YY, Timmins NE, Zandstra PW. Quality cell therapy manufacturing by design. Nat Biotechnol.
2016;34(4):393–400.
76.
Martin I, De Boer J, Sensebe L. A relativity concept in mesenchymal stromal cell manufacturing.
Cytotherapy. 2016;18(5):613–20.
SUPPLEMENTAL FIGURES
Figure S1. Representative large-field (6x6) output segmented
image using custom CellProfiler pipeline (File S1). Scale bar =
1000 μm.
Figure S2: MSCs possess greater morphological response to IFN-γ only vs TNF-α only. Cell major axis length,
cell perimeter, cell aspect ratio, and cell solidity of 4 MSC lines at low passage with different levels of IFN-γ and TNF-⍺
priming. Reference (dotted) lines show day 0 values for each cell-line. One-way ANOVA with Dunnett’s multiple
comparisons test vs unprimed control within the same cell-line. N = 4, *p < 0.05.
0-0
10-0
50-0
0-10
0-50
0
40
80
120
160
200
[IFNγ]-[TNF-α] (ng/mL)
Cell Major Axis Length (µm)
*
0-0
10-0
50-0
0-10
0-50
0
200
400
600
800
[IFNγ]-[TNF-α] (ng/mL)
Cell Perimeter (µm)
*
*
0-0
10-0
50-0
0-10
0-50
1
2
3
4
5
6
7
[IFNγ]-[TNF-α] (ng/mL)
Cell Aspect Ratio
*
*
0-0
10-0
50-0
0-10
0-50
0.0
0.4
0.4
0.6
0.8
1.0
[IFNγ]-[TNF-α] (ng/mL)
Cell Solidity
*
0-0
10-0
50-0
0-10
0-50
0
40
80
120
160
200
IFNγ-TNFα (ng/mL)
Cell Major Axis Length (µm)
*
*
0-0
10-0
50-0
0-10
0-50
0
200
400
600
800
[IFNγ]-[TNF-α] (ng/mL)
Cell Perimeter (µm)
*
*
0-0
10-0
50-0
0-10
0-50
1
2
3
4
5
6
7
[IFNγ]-[TNF-α] (ng/mL)
Cell Aspect Ratio
*
*
0-0
10-0
50-0
0-10
0-50
0.0
0.4
0.4
0.6
0.8
1.0
[IFNγ]-[TNF-α] (ng/mL)
Cell Solidity
*
*
0-0
10-0
50-0
0-10
0-50
0
40
80
120
160
200
[IFNγ]-[TNF-α] (ng/mL)
Cell Major Axis Length (µm)
*
*
0-0
10-0
50-0
0-10
0-50
0
200
400
600
800
[IFNγ]-[TNF-α] (ng/mL)
Cell Perimeter (µm)
*
*
0-0
10-0
50-0
0-10
0-50
1
2
3
4
5
6
7
[IFNγ]-[TNF-α] (ng/mL)
Cell Aspect Ratio
*
*
0-0
10-0
50-0
0-10
0-50
0.0
0.4
0.4
0.6
0.8
1.0
[IFNγ]-[TNF-α] (ng/mL)
Cell Solidity
*
*
0-0
10-0
50-0
0-10
0-50
0
40
80
120
160
200
[IFNγ]-[TNF-α] (ng/mL)
Cell Major Axis Length (µm)
*
*
0-0
10-0
50-0
0-10
0-50
0
200
400
600
800
[IFNγ]-[TNF-α] (ng/mL)
Cell Perimeter (µm)
0-0
10-0
50-0
0-10
0-50
1
2
3
4
5
6
7
[IFNγ]-[TNF-α] (ng/mL)
Cell Aspect Ratio
*
*
0-0
10-0
50-0
0-10
0-50
0.0
0.4
0.4
0.6
0.8
1.0
[IFNγ]-[TNF-α] (ng/mL)
Cell Solidity
*
*
110877
8F3560
PCBM1662
RB9
Figure S3: Effects of passage and interplay of IFN-γ and TNF-α on MSC
growth and proliferation. (A) Average values of cell perimeter for low
(blue) and high (red) passage MSCs across 4 MSC lines. N = 4 wells for all
64 priming conditions. (B) Average values of cell count fold change vs day 0
control for 4 MSC lines at low passage vs [TNF-α] and [IFN-γ]. N = 4 wells
for all 64 priming conditions. (C) Average values of cell count fold change vs
day 0 control for 4 MSC lines at high passage vs [TNF-α] and [IFN-γ]. N = 4
wells for all 64 priming conditions.
Figure S4: PCA of selected MSC morphological features. (A) Plot of PC1 vs
PC2. 512 data points from 4 MSC lines, 2 passages, and 64 priming conditions.
(B) Loading plot of morphological features with respect to PC1 and PC2. (C)
Loading values of morphological features for PC1 and PC2.
Figure S5: Effect of priming on MSC suppression of T cell CD25 expression. (A) MSC suppression of CD4+ T cell
activation as measured by CD25 expression with TNF-α and IFN-γ priming by cell line and passage. (B) MSC
suppression of CD8+ T cell activation as measured by CD25 expression with TNF-α and IFN-γ priming by cell line and
passage. Reference line represents activated control PBMCs. Mean +/- SD. One-way ANOVA with Dunnett’s multiple
comparisons test vs unprimed control. * denotes p < 0.05. N = 5 wells per priming condition.
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD4+
*
* *
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD4+
*
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD4+
* * * * *
* * *
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD8+
*
*
*
*
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD8+
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD8+
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% CD25+ of CD8+
Low Passage
High Passage
Low Passage
High Passage
PCBM1662
110877
8F3560
RB9
A
B
Figure S6: Effect of priming on MSC suppression of T cell IFN-γ expression. (A) MSC suppression of CD4+ T cell
IFN-γ expression with TNF-α and IFN-γ priming by cell line and passage. (B) MSC suppression of CD8+ T cell IFN-γ
expression with TNF-α and IFN-γ priming by cell line and passage. Reference line represents activated control PBMCs.
Mean +/- SD. One-way ANOVA with Dunnett’s multiple comparisons test vs unprimed control. * denotes p < 0.05. N = 5
wells per priming condition.
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD4+
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD4+
* *
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD4+
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD4+
* * * * * * *
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD8+
*
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% IFN-g+ of CD8+
*
*
Low Passage
High Passage
Low Passage
High Passage
PCBM1662
110877
8F3560
RB9
A
B
Figure S7: Effect of priming on MSC suppression of T cell TNF-α expression. (A) MSC suppression of CD4+ T cell
TNF-α expression with TNF-α and IFN-γ priming by cell line and passage. (B) MSC suppression of CD8+ T cell TNF-α
expression with TNF-α and IFN-γ priming by cell line and passage. Reference line represents activated control PBMCs.
Mean +/- SD. One-way ANOVA with Dunnett’s multiple comparisons test vs unprimed control. * denotes p < 0.05. N = 5
wells per priming condition.
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD4+
*
*
* * * * *
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD4+
*
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD4+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD4+
*
* *
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD8+
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD8+
*
*
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD8+
*
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD8+
0-0
0-2
0-50
5-0
5-2
5-50
10-10
50-0
50-2
50-50
0
50
100
150
[IFNg]-[TNF-a] (ng/mL)
% TNF-a+ of CD8+
Low Passage
High Passage
Low Passage
High Passage
PCBM1662
110877
8F3560
RB9
A
B
SUPPLEMENTAL TABLES
Table S1. Donor information for all MSC lines used in this study.
Donor ID
Sex
Age
Passage
Vendor
8F3560
F
24
3, 7
Lonza
PCBM1662
F
31
3, 7
AllCells
110877
M
22
3, 7
Lonza
RB9
M
43
2, 5
RoosterBio
Table S2. Definitions for morphological features used in this study. All morphological features are
evaluated using CellProfiler’s MeasureObjectSizeShape module except for Aspect Ratio,
Perimeter/Area, and Nuclear/Cytoplasm Ratio (NC Ratio).
Table S3. Antibodies used for flow cytometric assessment of T cell activation.
Antigen
Clone
Conjugation
Manufacturer/catalog no.
Validation Profile
hCD4
RPA-T4
PerCP
BioLegend 300528
1DegreeBio
hCD8a
RPA-T8
APC
BioLegend 301049
1DegreeBio
hCD25
M-A251
Pacific Blue
BioLegend 356130
Antibodypedia
hIFN-γ
B27
PE/Cy7
BioLegend 506518
1DegreeBio
hTNF-α
Mab11
Pacific Blue
BioLegend 502920
1DegreeBio
Table S4: MSC suppression of T cell proliferation varies
with MSC priming. Coefficient of variance for measures of T
cell activation across all MSC lines/passages and priming
conditions (80 total).
Measure
CV
% CD25+ (CD4)
25.20363
% CFSE Diluted (CD4)
67.9451
% CD25+ (CD8)
3.295221
% CFSE Diluted (CD8)
95.40027
% IFNg+ (CD4))
28.32703
% TNFa+ (CD4)
30.00897
% IFNg+ (CD8)
39.65482
% TNFa+ (CD8)
12.55163
Table S5: MSC morphological features correlate with their suppression of T cell proliferation.
Significantly correlated MSC morphological/functional pairs. N = 10. Bonferonni-adjusted p value cutoff
calculated as p< 3.44x10-6 (0.05/14537 tests)
Functional measure
Morphological Feature
Correlation
Signif Prob
% CFSE Diluted (CD4)
24cell_FormFactor
0.6489
7.53E-11
% CFSE Diluted (CD8)
24cell_FormFactor
0.6419
1.40E-10
% CFSE Diluted (CD4)
24cell_MajorAxisLength
-0.6277
4.62E-10
% CFSE Diluted (CD8)
24cell_MajorAxisLength
-0.5513
1.16E-07
% CFSE Diluted (CD4)
24cell_Perimeter
-0.5666
4.29E-08
% CFSE Diluted (CD4)
24nuc_Area
-0.5863
1.10E-08
% CFSE Diluted (CD8)
24nuc_Area
-0.5368
2.85E-07
% CFSE Diluted (CD4)
24nuc_MajorAxisLength
-0.6096
1.96E-09
% CFSE Diluted (CD8)
24nuc_MajorAxisLength
-0.6494
7.22E-11
% CFSE Diluted (CD4)
24nuc_MeanRadius
-0.5667
4.25E-08
% CFSE Diluted (CD4)
24nuc_MinorAxisLength
-0.5344
3.29E-07
% CFSE Diluted (CD4)
24nuc_PerimAreaRatio
0.6023
3.41E-09
% CFSE Diluted (CD8)
24nuc_PerimAreaRatio
0.5361
2.98E-07
% CFSE Diluted (CD4)
24nuc_Perimeter
-0.6029
3.25E-09
% CFSE Diluted (CD8)
24nuc_Perimeter
-0.5893
8.91E-09
% CFSE Diluted (CD4)
24nuc_Solidity
-0.5047
1.81E-06
% CFSE Diluted (CD8)
24nuc_Solidity
-0.5625
5.61E-08
Table S6: List of proteins present in MSC cultures and control MSC growth medium. (A) proteins detected
in at least one conditioned medium sample from a single MSC line/priming condition that were also not detected in
medium-only controls. (B) proteins present in MSC growth medium only controls
A. Proteins present in MSC cultures not present in media only control in screen of MSC
secretome
6Ckine, Activin, A, ANG-1, Angiogenin, ANGPTL3, B2M, BMPR-II, BTC, CA19-9, CA9,
Cathepsin B, Cathepsin S, CCL28, CD58, CD99, CTLA4, CXCL16, Cystatin B, DKK-1, Dkk-
3, DNAM-1, ENA-78, Eotaxin-2, ESAM, Follistatin, Follistatin-like, 1, G-CSF R, Galectin-1,
Galectin-3, Galectin-9, GASP-2, GCP-2, GDF-15, GH, GM-CSF, GRO, HAI-2, hCGb, I-309,
ICAM-1, IGFBP-1, IGFBP-2, IGFBP-3, IGFBP-4, IGFBP-6, IL-1, F7, IL-1, F8, IL-1, F9, IL-10,
IL-11, IL-13, IL-13, R1, IL-15, IL-17B, R, IL-1a, IL-1b, IL-1ra, IL-2, IL-23, IL-31, IL-5, IL-6R, IL-
7, IL-8, IP-10, Kallikrein 5, LAP(TGFb1), Legumain, LIF, MCP-1, MCP-2, MCP-3, MCP-4,
MCSF, Mer, MIG, MIP-1a, MIP-1b, MIP-3a, MMP-13, MSP, NOV, OPG, PAI-1, PARC, PDGF-
AA, Pentraxin 3, PIGF, SDF-1a, Syndecan-1, TACI, TECK, TFPI, Thrombospondin-2, TIMP-1,
TIMP-2, TNF, RI, TNF, RII, TNFb, TRAIL, R2, TRANCE, ULBP-1, uPA, uPAR, VCAM-1, VE-
Cadherin, VEGF, VEGF-C.
B. Proteins present in medium only control in screen of MSC secretome
ACE-2, Aggrecan, Albumin, AMICA, ANG-2, B7-H1, B7-H3, BAFF, bIG-H3, Clusterin,
CRTAM, Cystatin A, Decorin, DLL1, Fetuin A, FGF-21, FLRG, Fractalkine, Furin, GASP-1,
Granulysin, I-TAC, IGF-2, IL-15, R, IL-17E, IL-27, IL-3, IL-32alpha, IL-6, L1CAM-2, LAG-3,
LRIG3, MIF, MMP-1, NCAM-1, Nidogen-1, Periostin, RANTES, ROBO3, S100A8, sFRP-3,
Syndecan-3, Tie-1, TIM-3, Troponin, I, TSP-1, WISP-1.
Table S7: Correlation of MSC secretome with immunosuppression.
Correlations of T cell proliferation with specific secreted factors identified from
the secretome profiling screen for low passage MSC line PCBM1662 ordered
by p-value (low to high) determined through linear regression. Bonferonni-
corrected cutoff p-value=0.05/132 = 3.8x10-4.
Variable
by Variable
Correlation Signif Prob
%CFSE Diluted (CD8) MCP-2
-0.9224
4.32E-13
%CFSE Diluted (CD8) IFNg
-0.9104
3.03E-12
%CFSE Diluted (CD8) IP-10
-0.9029
8.87E-12
%CFSE Diluted (CD8) DKK-1
0.8965
2.08E-11
%CFSE Diluted (CD8) PIGF
0.883
1.07E-10
%CFSE Diluted (CD8) CXCL16
-0.8323
1.19E-08
%CFSE Diluted (CD8) MCP-3
-0.819
3.16E-08
%CFSE Diluted (CD8) Legumain
-0.8087
6.42E-08
%CFSE Diluted (CD8) ANG-1
0.7922
1.83E-07
%CFSE Diluted (CD8) MCP-4
-0.7898
2.11E-07
%CFSE Diluted (CD8) I-TAC
-0.7878
2.38E-07
%CFSE Diluted (CD8) ICAM-1
-0.7874
2.44E-07
%CFSE Diluted (CD8) ANGPTL3
0.7711
6.13E-07
%CFSE Diluted (CD8) SDF-1a
0.7671
7.63E-07
%CFSE Diluted (CD8) MIG
-0.7668
7.73E-07
%CFSE Diluted (CD8) B2M
-0.7577
1.25E-06
%CFSE Diluted (CD8) MCSF
-0.7509
1.75E-06
%CFSE Diluted (CD8) TIMP-2
0.7298
4.73E-06
%CFSE Diluted (CD8) CTLA4
0.724
6.12E-06
%CFSE Diluted (CD8) CA9
0.724
6.12E-06
%CFSE Diluted (CD8) IL-1ra
-0.7222
6.62E-06
%CFSE Diluted (CD8) IL-11
-0.7218
6.74E-06
%CFSE Diluted (CD8) IL-2
-0.7035
1.45E-05
%CFSE Diluted (CD8) Cathepsin S
-0.7027
1.50E-05
%CFSE Diluted (CD8) IL-1 F9
0.697
1.87E-05
%CFSE Diluted (CD8) TNFb
-0.6903
2.43E-05
%CFSE Diluted (CD8) CD99
0.6888
2.57E-05
%CFSE Diluted (CD8) IL-13
-0.6847
3.00E-05
%CFSE Diluted (CD8) G-CSF R
0.6791
3.69E-05
%CFSE Diluted (CD8) IL-5
-0.6777
3.88E-05
%CFSE Diluted (CD8) VEGF
0.6655
5.99E-05
%CFSE Diluted (CD8) CA19-9
0.6622
6.72E-05
%CFSE Diluted (CD8) IL-7
-0.6615
6.89E-05
%CFSE Diluted (CD8) IL-6
-0.6593
7.42E-05
%CFSE Diluted (CD8) CCL28
-0.6543
8.78E-05
%CFSE Diluted (CD8) RANTES
-0.6533
9.07E-05
%CFSE Diluted (CD8) GCP-2
0.648
1.08E-04
%CFSE Diluted (CD8) IL-1a
-0.646
1.15E-04
%CFSE Diluted (CD8) ENA-78
0.6454
1.17E-04
%CFSE Diluted (CD8) TSP-1
0.6304
1.88E-04
| 2021 | Morphological Landscapes from High Content Imaging Identify Optimal Priming Strategies that Enhance MSC Immunosuppression | 10.1101/2021.02.23.432501 | [
"Andrews Seth H.",
"Klinker Matthew W.",
"Bauer Steven R.",
"Marklein Ross A."
] | creative-commons |
Crowdsourcing digital health measures to predict Parkinson’s disease severity: the
Parkinson’s Disease Digital Biomarker DREAM Challenge
Solveig K. Sieberts1, Jennifer Schaff2 , Marlena Duda3, Bálint Ármin Pataki4, Ming Sun5, Phil
Snyder1, Jean-Francois Daneault6,7, Federico Parisi6,8, Gianluca Costante6,8, Udi Rubin9, Peter
Banda10, Yooree Chae1, Elias Chaibub Neto1, Ray Dorsey11, Zafer Aydın12, Aipeng Chen13,
Laura L. Elo14, Carlos Espino9, Enrico Glaab10, Ethan Goan15, Fatemeh Noushin Golabchi6,
Yasin Görmez12, Maria K. Jaakkola14,16, Jitendra Jonnagaddala17,18, Riku Klén14, Dongmei Li19,
Christian McDaniel20,21, Dimitri Perrin15, Nastaran Mohammadian Rad22,23,24, Erin Rainaldi25,
Stefano Sapienza6, Patrick Schwab26, Nikolai Shokhirev9, Mikko S. Venäläinen14, Gloria
Vergara-Diaz6, Yuqian Zhang27, the Parkinson’s Disease Digital Biomarker Challenge
Consortium, Yuanjia Wang28, Yuanfang Guan3, Daniela Brunner9,29, Paolo Bonato6,8, Lara M.
Mangravite1, Larsson Omberg1
1 Sage Bionetworks, Seattle, WA 98121
2 Elder Research, Inc., Charlottesville, VA, 22903
3 Department of Computational Medicine and Bioinformatics, University of Michigan, Ann Arbor,
MI 48109
4 Department of Physics of Complex Systems, ELTE Eötvös Loránd University, Budapest,
Hungary
5 Google Inc, New York, NY, USA 10011
6 Dept of PM&R, Harvard Medical School, Spaulding Rehabilitation Hospital, Charlestown, MA,
02129
7 Dept of Kinesiology and Health, Rutgers University, New Brunswick, NJ 08901
8 Wyss Institute, Harvard University, Boston, MA, 02115
9 Early Signal, 311 W 43rd Street, New York, NY 10036
10 Luxembourg Centre for Systems Biomedicine, University of Luxembourg, Esch-sur-Alzette,
L-4362, Luxembourg
11 Center for Health + Technology, University of Rochester, Rochester, NY 14642
12 Department of Electrical and Computer Engineering, Abdullah Gul University, Kayseri, Turkey
13 Prince of Wales Clinical School, UNSW Sydney, Australia
14 Turku Bioscience Centre, University of Turku and Åbo Akademi University, Tykistökatu 6,
FI-20520 Turku, Finland
15 School of Electrical Engineering and Computer Science, Queensland University of
Technology, Brisbane, Queensland, Australia, 4000
16 Department of Mathematics and Statistics, University of Turku, FI-20014 Turku, Finland
17 School of Public Health and Community Medicine, UNSW Sydney, Australia
18 WHO Collaborating Centre for eHealth, UNSW Sydney, Australia
19 Clinical and Translational Science Institute, University of Rochester Medical Center,
Rochester, NY, USA, 14642
20 Artificial Intelligence, University of Georiga, Athens, GA, USA, 30602
21 Computer Science, University of Georiga, Athens, GA, USA, 30602
22 Institute for Computing and Information Sciences, Radboud University, Nijmegen, The
Netherlands, 6525EC
23 Fondazione Bruno Kessler (FBK), Via Sommarive 18, 38123, Povo, Trento, Italy
24 University of Trento, Italy, 38122 TN
25 Verily Life Sciences, 269 East Grand Ave, South San Francisco, CA 94080
26 Institute of Robotics and Intelligent Systems, ETH Zurich, Zurich, Switzerland, CH-8092
27 School of Biomedical Engineering, Shanghai Jiao Tong University, China
28 Department of Biostatistics, Mailman School of Public Health, Columbia University, 722
W168th Street, New York, NY 10032
29 Dept. of Psychiatry, Columbia University, New York, NY
Abstract
Mobile health, the collection of data using wearables and sensors, is a rapidly growing field in
health research with many applications. Deriving validated measures of disease and severity
that can be used clinically or as outcome measures in clinical trials, referred to as digital
biomarkers, has proven difficult. In part due to the complicated analytical approaches
necessary to develop these metrics. Here we describe the use of crowdsourcing to specifically
evaluate and benchmark features derived from accelerometer and gyroscope data in two
different datasets to predict the presence of Parkinson's Disease (PD) and severity of three PD
symptoms: tremor, dyskinesia and bradykinesia. 40 teams from around the world submitted
features, and achieved drastically improved predictive performance for PD (best AUROC=0.87),
as well as severity of tremor (best AUPR=0.75), dyskinesia (best AUPR=0.48) and bradykinesia
(best AUPR=0.95).
Mobile health and digital health, that is, the evaluation of health outside of the clinic using
wearables and smartphones, and, specifically, the collection of real world evidence using
sensors1 demonstrates great potential in understanding the lived experience of disease. These
efforts have been implemented using both research-grade wearable sensors and, increasingly,
through the use of smartphones, smartwatches, and consumer devices, which are readily
available to the general public. While most of this work has been in the context of exploratory
and feasibility studies, we are increasingly seeing evidence of their use as digital endpoints from
sensors in clinical trials.2 Digital measures provide the opportunity to more accurately monitor
the degree to which disease status and/or treatments affect an individual’s daily life, typically
through the capture of large amounts of longitudinal real world data. Development of sensitive
“digital biomarkers” extracted from these rich data offer the opportunity for better decision
making in both trials and health care.
One area of emerging digital biomarker development is Parkinson’s disease (PD), a
neurodegenerative disorder that conspicuously affects the motor coordination, along with other
domains such as cognitive function, mood, and sleep. Classic motor symptoms of the disease
include tremor, slowness of movement (bradykinesia), posture and walking perturbations, and
muscle rigidity. Additionally, motor symptoms can be common side effects of medical treatment,
chiefly involuntary movement, known as dyskinesia. Given the strong motor component of the
disease and treatment side-effects, multiple approaches have leveraged accelerometer and
gyroscope data from wearable devices for the development of digital biomarkers in PD (see for
example 3,4). However, they have yet to be translated into the clinic or as primary biomarkers.
The use of digital biomarkers as endpoints or measures of disease in the clinical or regulatory
setting requires robust evidence for their validity. Unfortunately, this work is both expensive and
difficult to perform, leading to often underpowered validation studies evaluated by a single
research group and, hence, subject to the self assessment trap.5 Pre-competitive efforts are
underway such as Critical Path’s Patient Reported Outcome (PRO) Consortium 6 and the Open
Wearables Initiative (OWI). Here we describe an open initiative to both competitively and
collaboratively evaluate analytical approaches for the assessment of disease severity in an
unbiased manner. The Parkinson’s Disease Digital Biomarker (PDDB) DREAM Challenge
(https://www.synapse.org/DigitalBiomarkerChallenge) benchmarked crowd-sourced methods of
processing sensor data (i.e. feature extraction), which can be used in the development of digital
biomarkers that are diagnostic of disease or can be used to assess symptom severity. In short,
the PDDB Challenge participants were provided with training data, which included sensor data
and disease status or symptom severity labels, as well as a test set, which contained sensor
data only. Given raw sensor data from two studies, participating teams engineered data features
that were evaluated on their ability to predict disease labels in models built using an
ensemble-based predictive modeling pipeline.
The challenge leveraged two different datasets--mPower7, a remote smartphone based study,
and the Levodopa (L-dopa) Response Study8,9, a multi-wearable clinical study --which were not
previously publicly available, so that evaluation could be performed in a blinded, unbiased
manner. For both studies, time-series data were recorded from sensors while participants
performed pre-specified motor tasks. In the mPower Study, accelerometer and gyroscope data
from a gait and balance test in 4,799 individuals were used to discriminate Parkinson’s patients
from controls using 76,039 total measures. In the L-dopa Response Study, accelerometer
recordings from GENEActiv and Pebble watches were captured on two separate days from 25
patients exhibiting motor-fluctuations10 (i.e. the side effects and return of symptoms after
administration of levodopa), as they were evaluated for symptom severity during the execution
of short (30 second) motor tasks designed to evaluate tremor, bradykinesia, and dyskinesia.
Data collection during the battery of tasks was repeated six to eight times over the course of
each day in 30 minutes blocks, resulting in 3-4 h-motor activity profiles reflecting changes in
symptom severity. In total 8,239 evaluations were collected across 3 different PD symptoms.
Results
We developed 4 sub-challenges using the two datasets; one using data from the mPower Study
and 3 using data from the L-dopa Response Study. Using the mPower data, we sought to
determine whether mobile sensor data from a walking/standing test could be used to predict PD
status (based on a professional diagnosis as self-reported by the study subjects) relative to age
matched controls from the mPower cohort (sub-challenge 1 (SC1)). The three remaining
sub-challenges used the L-dopa data to predict symptom severity as measured by: active limb
tremor severity (0-4 range) using mobile sensor data from 6 bilateral upper-limb activities
(sub-challenge 2.1 (SC2.1)); resting upper-limb dyskinesia (presence/absence) using bilateral
measurements of the resting limb while patients were performing tasks with the alternate arm
(sub-challenge 2.2 (SC2.2)); and presence/absence of active limb bradykinesia using data from
5 bilateral upper-limb activities (sub-challenge 2.3 (SC2.3)). Participants were asked to extract
features from the mobile sensor data, which were scored using a standard set of algorithms for
their ability to predict the disease or symptom severity outcome (Figure 1).
For SC1, we received 36 submissions from 20 unique teams, which were scored using
area under ROC curve (AUC) (see methods). For comparison, we also fit a ‘demographic’
model which included only age and gender. Of the 36 submissions, 2 scored significantly better
(unadjusted p-value ≤ 0.05) than the demographic and meta-data model (AUROC 0.627),
though this is likely due to the relatively small size of the test set used to evaluate the models.
The best model achieved an AUROC score of 0.868 (Figure 2A).
For SC2.1-SC2.3, we received 35 submissions from 21 unique teams, 37 submissions
from 22 unique teams, and 39 submissions from 23 unique teams, respectively (Figure 2B-D).
Due to the imbalance in severity classes, these sub-challenges were scored using the area
under precision-recall curve (AUPR). For prediction of tremor severity (SC2.1), 16 submissions
significantly outperformed baseline model developed using only meta-data at an unadjusted
p-value ≤ 0.05. The top performing submission achieved an AUPR of 0.750 (null expectation
0.432). For prediction of dyskinesia (SC2.2), 8 submissions significantly outperformed the
meta-data based baseline model. The top performing submission achieved an AUPR of 0.477
(null expectation 0.195). For prediction of bradykinesia (SC2.3), 22 submissions significantly
outperformed the baseline model. The top performing submission achieved an AUPR of 0.950
(null expectation 0.266). While this score is impressive, it is important to note that in this case
the meta-data based baseline model was also highly predictive (AUPR = 0.813).
The top performing team in SC1 used a deep learning model with data augmentation to
avoid overfitting (see Methods for details), and 4 of the top 5 models submitted to this
sub-challenge employed deep learning models. In contrast, each of the winning methods for
SC2.1-SC2.3 used signal processing approaches (see Methods). While there are differences in
the data sets used for the sub-challenges (e.g. size), which could contribute to differences in
which type of approach is ultimately most successful, we surveyed the landscape of approaches
taken to see if there was an overall trend relating approaches and better performance. Our
assessment, which included aspects of data used (e.g. outbound walk, inbound walk, and rest
for the mPower data), sensor data used (e.g. accelerometer, pedometer, or gyroscope), use of
pre- and post- data processing, as well as type of method used to generate features (e.g. neural
networks, statistical-, spectral-or energy- methods), found no methods or approaches which
were significantly associated with performance in any subchallenge. This lack of statistical
significance could be attributed to the large overlap in features, activities and sensors for
individual submissions in that, most teams used a combination of the different methods. We
also clustered submissions by similarity of their overall approaches based on the aspects
surveyed. While we found four distinct clusters for each sub-challenge no clusters associated
with better performance in either sub-challenge (Supplementary Figure 1).
We then turned our focus to the collection of features submitted by participants to
determine which individual features were best associated with disease status (SC1) or symptom
severity (SC2.1-2.3). For SC1, the 21 most associated individual features were from the two
submissions of the top performing team (which were ranked 1-2 among all submissions). These
21 features were also individually more informative (higher AUC) than any of the other teams
entire submission (Supplementary Figure 2B). Among the runner-up submissions,
approximately half of the top-performing features were derived using signal processing
techniques (36 out of 78 top features, see Supplementary Figure 2A) with a substantial
proportion specifically derived from the return phase of the walk. Interestingly, the performance
of individual features in the runner-up submissions did not always correspond to the final rank of
the team. For example the best individual feature of the second best performing team ranked
352 (out of 4546). Additionally, a well-performing individual feature did not guarantee good
performance of the submission (the best feature from runner-up submissions belongs to a team
with ranking 22 out of 36).
We then performed two-dimensional manifold projection and then clustered the
individual features to better understand the similarity of feature spaces across teams
(Supplementary Figure 3). One of the expected observations is that the relation between
features associated with the same team and the cluster membership is strongly significant
(p-value~0, mean Chi-Square=8461 for t-SNE and 5402 for MDS with k-means k > 2). This
suggests most of the teams had a tendency to design similar features such that within team
distances were smaller than across-team distances (on average 26% smaller for t-SNE and
16% smaller for MDS projections). We also found that cluster membership was significantly
associated with submission performance (mean p-value = 1.55E-11 for t-SNE and 1.11E-26 for
MDS with k-means k > 2). In other words, features from highly performing submissions tended
to cluster together. This enabled us to identify several high-performance hot-spots. For
example, in Supplementary Figure 3C a performance hot-spot is clearly identifiable and
contains 51% (respectively 39%) of the features from the two best teams in SC1: Yuanfang
Guan and Marlena Duda, and ethz-dreamers, which were the top performing teams, both of
which employed Convolutional Neural Net (CNN) modeling. Interactive visualizations of the
clusters are available online at https://ada.parkinson.lu/pdChallenge/clusters and
https://ada.parkinson.lu/pdChallenge/correlations.
For SC2.1-2.3, we found that the best performing individual feature was part of the
respective sub-challenge winning teams’ submission, and that these best performing individual
features were from submissions that have fewer features (Supplementary Figure 4B, 4D, 4F).
Similar to the observations in the SC1, the individual feature performance was typically not
correlated with overall performance (Pearson correlation = -0.05, 0.10 and 0.04 for SC2.1,
SC2.2 and SC2.3, respectively, p-values = 0.17, 0.0003, 0.44). Instead, individual features with
modest performance, when combined, achieved better performance than feature sets with
strong individual features. For SC2.1 and SC2.3 (tremor and bradykinesia), machine learning
approaches showed higher performing individual features than other methods, however, signal
processing based methods showed better performing individual features in SC2.2
(Supplementary Figure 4A, 4C, 4E). We also attempted to improve upon the best submissions
by searching among the space of submitted features for an optimal set. Attempts to optimally
select features using Random Forests or recursive feature elimination resulted in an AUPR of
0.38 and 0.35, respectively, in SC2.2, placing behind the top eight and twelve individual
submissions. An approach using the top principal components (PCs) of the feature space, fared
slightly better, outperforming the best model in SC2.2 (AUPR = 0.504 AUPR, above the top 5
feature submissions of 0.402-0.477), but failing to outperform the top models in SC2.1 and
SC2.3 (AUPR = 0.674, below the top five submission scores for SC2.1; and 0.907 AUPR,
within the range of the top 5 feature submissions of 0.903-0.950 for SC2.3).
Age, gender and medication effects in mPower
Because rich covariates were available in the mPower data set, we sought to explore the
prediction space created by the top submissions, in order to identify whether we could discern
any patterns with respect to available covariates or identify any indication that these models
could discern disease severity or medication effects (Supplementary Figure 5). To visualize this
complex space we employed topological data analysis (TDA)11 of the top SC1 submissions, to
explore grouping of subjects, firstly based on the fraction of cases with presence or absence of
PD. The algorithm outputs a topological representation of the data in network form (see
Methods) that maintains the local relationship represented within the data. Each node in the
network represents a closely related group of samples (individuals) where edges connect nodes
that share at least one sample. Next we used TDA clustering to explore whether the top models
showed any ability to discern symptom severity, as possibly captured by medication status
(Supplementary Figure 6). Specifically, we sought to identify whether PD patients "on-meds"
(right after taking medication) cases are more similar to controls as compared to patients who
were "off-meds" (right before taking medication or not taking at all). To this end, we created a
topological representation for both cases, treating on-med and off-med states separately for
each individual and comparing each case with the controls. Here we considered only subjects
with both on-med and off-med sessions, to ensure the comparison was between the same
population of subjects and using only 3 of the top six submissions (ethz-dreamers 1,
ethz-dreamers 2 and vmoroz), whose features values varied within individual. We observed no
separation between patients who were on-meds versus off-meds. This was consistent with the
statistical analysis which showed no significant difference in the predicted PD status for patients
who were “on-meds” versus “off-meds” at the time they performed their walking/balance test for
any of the top models, even among patients who have previously been shown to have motor
fluctuations 12,13.
We then explored whether the ability of the predictive models to correctly assess PD is
influenced by factors associated with the study participants’ demographics, such as their sex,
age, or the total number of walking activities they performed. We evaluated the relative
performance of the top features sets when applied to specific subsets of the test data. When
comparing the predictive models' performances in female subjects and male subjects aged 57
or older, we found that the predictive models' were on average more accurate in classifying
female subjects than male subjects with an average increase AUROC of 0.17 (paired t-test
p-value = 1.4e-4) across the top 14 models (i.e. those scoring better than the model using only
demographic data). We note that the magnitude of the relative change is likely driven by the
class balance differences between male and female subjects in the test set. In particular, a
larger fraction of the female subjects aged 57 or older had a prior professional PD diagnosis
than the male subjects. 80% of female subjects aged 57 or older (n=23) had PD, and 64% of
male subjects aged 57 or older (n=66). And indeed, when compared to the Demographic model,
several of the top submissions are actually performing worse than the Demographic model in
the female subjects, while almost all are outperforming the Demographic model in the male
subjects (Supplementary Figure 7). Generally, it appears that mobile sensor features are
contributing more to prediction accuracy in the male subjects than the female subjects.
We also compared the performance of the top 14 feature sets when applied to subjects
in various age groups, and found that the models performed similarly across age groups
(Supplementary Figure 7). However, in comparison to the Demographic model, the top
submissions perform relatively better in younger age groups (57 to 65) than in older age groups
(65 and up), and in particular, the Demographic model outperforms all of the top submissions in
the highest age bracket (75 and up). This implies that the mobile features do not contribute and
actually add noise in the older age brackets. Of note, the winning model by Yuanfang Guan and
Marlena Duda performs well in across most age and gender subgroups, but performs especially
poorly in oldest subgroup, which have the fewest samples.
To assess whether the total number of tasks performed by a subject had an impact on
predictive performance, we attempted to compare subjects that had performed more tasks with
those that had performed fewer. However, we found that in the mPower dataset the number of
walking activities performed was predictive in itself, i.e. PD cases on average performed more
tasks than the corresponding controls. We could therefore not conclusively determine whether
having more data from walking activities on a subject increased the performance of the
predictive models. Though, related work has shown that repeatedly performed smartphone
activities can capture symptom fluctuations in patients3.
Task performance across L-dopa sub-challenges
While the L-dopa data set had fewer patients, and thus was not powered to answer
questions about the models’ accuracy across demographic classes, the designed experiment
allowed us to examine the predictive accuracy of the different tasks performed in the L-dopa
data to understand which tasks showed the best accuracy with respect to predicting clinical
severity. We scored each submission separately by task applying the same model fitting and
scoring strategies used on the complete data set. For the prediction of tremor (SC2.1) and
bradykinesia (SC2.3), the different tasks showed markedly different accuracy as measured by
improvement in AUPR over null expectation (Supplementary Figure 8). We observe statistically
significant differences in improvement over expected value for tremor and bradykinesia
(Supplementary Table 1-2). For tremor, activities of everyday living such as folding laundry and
organizing paper perform better than UPDRS-based tasks such as finger-to-nose and
alternating hand movements (Supplementary Figure 8, Supplementary Table 1), and the
demographic model outperformed participant submissions in almost all cases. While the
assembling nuts and bolts task showed the highest improvement over the null expectation, the
demographic model also performed well, outperforming a substantial proportion of the
submissions. For bradykinesia, the expected AUPR varied widely (from 0.038 for pouring water
to 0.726 for alternating hand movements). For most tasks, the participant submissions
outperformed the demographic model, except in the case of the alternating hand movements
task. For dyskinesia, there was no statistical difference between finger-to-nose or alternating
hand movements, but since these were assessed on the resting limb, it is to be expected that
this is not affected by the task being performed on the active limb.
Discussion
Given the widespread availability of wearable sensors, there is significant interest in the
development of digital biomarkers and measures derived from these data with applications
ranging from their use as alternative outcomes of interest in clinical trials to basic disease
research1. Even given the interest and efforts toward this end, to-date, there are very few
examples where they have been deployed in practice beyond the exploratory endpoint or
feasibility study setting. This is partially due to a lack of proper validation and standard
benchmarks. Through a combination of competitive and collaborative effort we engaged
computational scientists around the globe to benchmark methods for extracting digital
biomarkers for the diagnostics and severity of PD. With this challenge we aimed to separate the
evaluation of methods from the data generation by creating two sets of challenges looking at
diagnostic and measures of severity in two separate datasets.
Participants in this challenge used an array of methods for feature extraction spanning
unsupervised machine learning to hand tuned signal processing. We did not, however, observe
associations between types of methods employed and performance with the notable exception
that the top two teams in the diagnostic biomarker challenge based on mPower data (SC1)
generated features using CNNs while top performing teams in SC2.1-2.3 that used the smaller
L-Dopa dataset derived features using signal processing (though a CNN-based feature set did
rank 2nd in SC2.3). The top performing team in SC1 significantly outperformed the submissions
of all remaining teams in the sub-challenge. This top performing team was unique in its use of
data augmentation, but otherwise used similar methods to the runner up team. And indeed deep
learning has previously been successfully applied in the context of detecting Parkinsonian gait14.
However, given it’s relatively poorer performance in SC2, which utilized a substantially smaller
dataset, we speculate that CNNs may be most effective in very large datasets. This was further
supported by the observation that the top SC1 model did not perform well in the oldest study
subjects which corresponds to the smallest age group. If sample size is indeed a driver of
success of CNNs, this suggests that applying these methods to most digital validation datasets
will not be possible as they currently tend to include dozens to hundreds of individuals in
contrast to the thousands available in the SC1 data and the typical deep learning dataset15.
Traditionally, biomarkers used clinically have a well-established biological or
physiological interpretation (e.g. temperature, blood pressure, serum LDL) allowing a clinician to
comprehend the relationship between the value of the marker and changes in phenotype or
disease state. Ideally, this would be the case for digital biomarkers as well, however, machine
learning models vary in their interpretability. In order to try to understand the features derived
from machine learning models, we computed correlations between the CNN derived features
submitted by teams with signal processing based features, which are often more physiologically
interpretable. We were unable to find any strong linearly related signal processing analogs.
Further work is necessary to try to interpret the effects being captured, though previous work
has identified several interpretable features including step length, walking speed, knee angle,
and vertical parameter of ground reaction force16, most of which are not directly measurable
given the available data available in mPower. Other work has suggested that Parkinsonian
freezing of gait is most pronounced at the start and during turns17–19.
Understanding the specific tasks and aspects of those activities which are most
informative helps researchers to optimize symptom assessments while reducing the burden on
study subjects and patients by focusing on shorter, more targeted tasks, ultimately aspiring to
models for tasks of daily living instead of prescribed tasks20. To this end, given the availability of
multiple tasks in SC2, we analyzed which tasks showed the best accuracy. For the tremor
severity for example, the most informative tasks were not designed to distinguish PD symptoms
specifically (pouring water, folding laundry and organizing sheets of paper) but mimic daily
activities. Whereas finger-to-nose and alternating hand movements, which are frequently used
in clinical assessments, showed the lowest predictive performance, and top models did not
outperform the demographic model for these tasks. For the assessment of bradykinesia, the
finger-to-nose, organizing paper and alternating hand movements tasks showed the best model
performance. However, in the case of alternating-hand-movements, the improved performance
could be fully explained by the demographic model.
We believe that there are opportunities to improve the submitted models further
specifically in the sub-populations where they performed worse. For example, given the
difference in performance between male and female in top submissions, as well the relatively
better performance in younger patients (57-65) it might be possible that different models and
features might be necessary to capture different aspects of the disease by age and gender. For
example, it stands to reason that the standard for normal gait differs in older people relative to
younger people. Given the heterogeneity of symptom manifestation in PD, there might be very
many sub-populations or even personalized differences in severity12. That is, the changes in
disease burden as explored in SC2 might best be learned by personalized models. To help
answer this question and to explore further the use of data collected in free living conditions, we
have recently launched a follow-up challenge looking at predicting personalized differences in
symptom burden from data collected passively during free living conditions.
Online Methods
The mPower Study
mPower7 is a longitudinal, observational iPhone-based study developed using Apple’s
ResearchKit library (http://researchkit.org/) and launched in March 2015 to evaluate the
feasibility of using mobile sensor-based phenotyping to track daily fluctuations in symptom
severity and response to medication in PD. The study was open to all US residents, above the
age of 18 who were able to download and access the study app from the Apple App Store, and
who demonstrated sufficient understanding of the study aims, participant rights, and data
sharing options to pass a 5-question quiz following the consent process. Study participants
participated from home, and completed study activities through their mobile device.
Once enrolled participants were posed with a one-time survey in which they were asked
to self report whether or not they had a professional diagnosis of PD, as well as demographic
(Table 1) and prior treatment information. On a monthly basis, they were asked to complete
standard PD surveys (Parkinson Disease Questionnaire 821 and a subset of questions from the
Movement Disorder Society Universal Parkinson Disease Rating Scale instrument22). They were
also presented daily with four separate activities: ‘memory’ (a memory-based matching game),
‘tapping’ (measuring the dexterity and speed of 2-finger tapping), ‘voice’ (measuring sustained
phonation by recording a 10-second sustained “Aaaahhh”), and ‘walking’ (measuring
participants’ gait and balance via the phone’s accelerometer and gyroscope). For the purposes
of this treatment, we focus on the ‘walking’ test, along with the initial demographic survey data.
The walking test instructed participants to walk 20 steps in a straight line, turn around,
and stand still for 30 seconds. In the first release of the app (version 1.0, build 7), they were
also instructed to walk 20 steps back, following the 30 second standing test, however
subsequent releases omitted this return walk. Participants could complete the four tasks,
including the walking test, up to three times a day. Participants who self-identified as having a
professional diagnosis of PD were asked to do the tasks (1) immediately before taking their
medication, (2) after taking their medication (when they are feeling at their best), and (3) at
some other time. Participants who self-identified as not having a professional diagnosis of PD
(the controls) could complete these tasks at any time during the day, with the app suggesting
that participants complete each activity three times per day.
The Levodopa Response Study
The L-dopa Response Study8,9 was an experiment with in-clinic and at-home
components, designed to assess whether mobile sensors could be used to track the unwanted
side-effects of prolonged treatment with L-dopa. Specifically, these side-effects, termed motor
fluctuations, include dyskinesia and waning effectiveness at controlling symptoms throughout
the day. In short, a total of 31 PD patients were recruited from 2 sites, Spaulding Rehabilitation
Hospital (Boston, MA) (n=19) and Mount Sinai Hospital (New York, NY) (n=12). Patients
recruited for the study came to the laboratory on Day 1 while on their usual medication schedule
where they donned multiple sensors: a GENEActiv sensor on the wrist of the most affected arm,
a Pebble smartwatch on the wrist of the least affected arm, and a Samsung Galaxy Mini
smartphone in a fanny pack worn in front at the waist. They then performed section III of the
MDS-UPDRS22. Thereafter, they performed a battery of motor tasks that included activities of
daily living and items of section III of the MDS-UPDRS. This battery of tasks lasted
approximately 20 minutes and was repeated 6-8 times at 30-minute intervals throughout the
duration of the first day. Study subjects returned 3 days later in a practically defined
off-medication state (medication withheld overnight for a minimum of 12 hours) and repeated
the same battery of tasks, taking their medication following the 1st round of activities. This study
also included data collection at home, between the two study visits, but these data were not
used for the purposes of this challenge.
During the completion of each motor task, clinical labels of symptom severity or
presence were assessed by a clinician with expertise in PD for each repetition. Limb-specific
(i.e. left arm, left leg, right arm, and right leg) tremor severity score (0-4), as well as upper-limb
and lower-limb presence of dyskinesia (yes or no) and bradykinesia (yes or no) were assessed.
For the purposes of this challenge, we used only the GENEActiv and Pebble sensor information
and upper limb clinical labels for a subset of the tasks: finger-to-nose for 15s (repeated twice
with each arm) (ftn), alternating hand movements for 15s (repeated twice with each arm) (ram),
opening a bottle and pouring water three times (drnkg), arranging sheets of paper in a folder
twice (orgpa), assembling nuts and bolts for 30s (ntblt), and folding a towel three times (fldng).
Accelerometer data for both devices were segmented by task repetition prior to use in this
challenge.
The Parkinson’s Disease Digital Biomarker Challenge
Using a collaborative modeling approach we ran a challenge to develop features that
can be used to predict PD status and disease severity using data from mPower and the L-dopa
Response Trial. The Challenge was divided up into 4 sub-challenges, based on different
phenotypes in the 2 different data sets. Sub-challenge 1 (SC1) focused on extraction of mobile
sensor features which distinguish between PD cases and controls using the mPower data.
Sub-challenges 2.1, 2.2, and 2.3 (SC2.1-SC2.3) focused on extraction of features which reflect
symptom severity for tremor, dyskinesia, and bradykinesia, respectively, using the L-dopa data.
In each case, participants were provided with a training set, containing mobile sensor data,
phenotypes for the individuals represented and all available demographics and metadata for the
data set in question. Using these data they were tasked with optimizing a set of features
extracted from the mobile sensor data, which best predicted the phenotype in question. They
were also provided a test set, containing only mobile sensor data, and upon challenge deadline
were required to return a feature matrix for both the training and test sets. Participants were
allowed a maximum of 2 submissions per sub-challenge, and could participate in any or all of
posed sub-challenges.
For extracting features which predict of PD status using the mPower data, participants
were provided with up to 30 seconds long recordings of approximately 100 Hz from an
accelerometer and gyroscope from 39,376 walking tasks as well as the associated 30 second
recordings of standing in place, representing 660 individuals with self-reported PD and 2,155
control subjects, as a training set. They were also provided with self reported covariates,
including PD diagnosis, year of diagnosis, smoking, surgical intervention, deep brain
stimulation, and medication usage, as well as demographic data, including age, gender, race,
education and marital status (Table 1)7. As a test data set, they were provided the same mobile
sensor data from 36,664 walking/standing tasks for 614 PD patients and 1,370 controls which
had not been publicly available previously, but were not provided any clinical or demographic
data for these individuals. Participants were asked to develop feature extraction algorithms for
the mobile sensor data which could be used to successfully distinguish PD patients from
controls, and were asked to submit features for all walking/standing activities in the training and
test sets.
For the prediction of symptom presence or severity (sub-challenges 2.1-2.3), participants
were provided with bilateral mobile sensor data for up to 14 repetitions of 12 separate tasks
(drining (drnkg), organizing papers(orgpa), nut ands bolts(ntblts), foolding laundry (fldng), and 2
bilateral repetitions of finger to nose(ftn) and rapid hand movements(ram)) from 27 subjects
from the L-dopa data. For 19 subjects, symptom severity (tremor) or presence (dyskinesia and
bradykinesia) were provided to participants as a training data set for a total of 3667
observations for tremor severity (2332, 878, 407, 38, and 12 for severity 0, 1, 2, 3, and 4,
respectively), 1556 observations for dyskinesia presence (1236 present), and 3016
observations for bradykinesia presence (2234 present). Participants were asked to provide
extracted features which are predictive of each symptom for these as well as the 1500, 660, and
1409 observations, for tremor, dyskinesia and bradykinesia, respectively, from the 8 test
individuals for which scores were not released.
It is important to note that for each data set, the training and test sets were split by
individual, that is that all tasks for a given individual fell exclusively into either the training or test
set to avoid inflation of prediction accuracy from the non-independence of repeated measures
on the same individual23.
The challenge website (https://www.synapse.org/DigitalBiomarkerChallenge) documents
the challenge results, including links to teams’ submission write-ups and code, and links to the
the public repositories for the mPower and L-dopa data.
Submission Scoring
For SC1, feature set submissions were evaluated by fitting an ensemble machine
learning algorithm to the training observations, and predicting on the test observations. To
minimize undue influence from subjects who completed large numbers of walking/standing
tests, features were first summarized using the median of each feature across all observations,
so that each subject occured once in the training or test set. Aggregation via maximum showed
similar results as median. For each submission, elastic net (glmnet), random forests, support
vector machines (SVM) with linear kernel, k-nearest neighbors, and neural nets models were
optimized using 50 bootstrap with AUROC as the optimization metric, and combined using a
greedy ensemble in caretEnsemble in R. Age and sex were added as potential predictors in
every submission. A subset of the provided data was used to minimize age differences between
cases and controls as well as to minimize biases in study enrollment date, resulting in a training
set of 48 cases and 64 controls and a testing set of 21 cases and 68 controls. Feature sets were
ranked by the area under the receiver operator characteristic curve (AUROC) of the test
predictions. Each team was allowed two submissions.
For SC2.1-2.3, the feature sets were evaluated using a soft-voting ensemble — which
averages the predicted class probabilities across models — of predictive models consisting of a
random forest, logistic regression with L2 regularization, and support vector machine (RBF
kernel) as implemented in the scikit-learn Python package (0.20.0) 24. The random forest
consisted of 500 trees each trained on a bootstrapped sample equal in size to the training set,
the logistic regression model used 3-fold cross-validation, and the support vector machine
trained directly on the training set with no cross-validation and outputted probability estimates,
rather than the default behavior of class scores. Other parameters were set to the default value
as specified in the scikit-learn v0.20 documentation. Due to imbalance of the class labels, we
adopted the area under the precision-recall curve (AUPR) as the performance metric for the
L-dopa sub-challenges. Non-linear interpolation was used to compute AUPR25. SC2.1 (active
limb tremor) represents a multiclass classification problem. In order to calculate a multiclass
AUPR we transformed the multiclass problem into multiple binary classification problems using
the “one-vs-rest” approach (where we trained a single classifier per class, with the samples of
that class as positive cases and remaining samples as negative cases). For each of these
binary classification problems, we computed AUPR values and combined them into a single
metric by taking their weighted mean, weighted by the class distribution of the test set. SC2.2
and SC2.3 are binary classification problems, and we employed the AUPR metric directly.
For all 4 subchallenges, 1000 bootstraps of the predicted labels were used to assess the
confidence of the score, and to compute the p-value relative to the demographic only model.
Description of winning methods
Along with their feature submissions, challenge participants provided methods
description and computational code to reproduce their features. Below we provide brief
descriptions of the winning models.
Subchallenge 1: Team Yuanfang Guan and Marlena Duda
The winning method by Team ‘Yuanfang Guan and Marlena Duda’ used an end-to-end
deep learning architecture to directly predict PD diagnosis utilizing the rotation rate records.
Separate models were nested-trained for balance and gait data, and the predictions were
pooled by average when both are available. RotationRate x, y and z were used as three
channels in the network. Each record was centered and scaled by standard deviation, then
standardized to contain 4000 time points by 0-padding. Data augmentation was key to prevent
overfitting to training data, and was the primary difference in performance to the next deep
learning model by ‘ethz-dreamers’. The following data augmentation techniques were included
to address the overfitting problem: a) simulating people holding phones at different directions by
3D random rotation of the signal in space based on the Euler rotation formula for standard rigid
body, vertex normalized to unit=1, b) time-wise noise-injection (0.8-1.2) to simulate a person
walks faster or slower and c) magnitude augmentation to account for tremors at higher
frequency and the sensor discrepancies when phones were outsourced to different
manufacturers.
The network architecture was structured as 8 successive pairs of convolution and max
pool layers. The last layer of prediction was provided as features for the Challenge. Parameters
were batch size = 4, learning rate = 5x10-4, epoch = 50*(~half of sample size). This CNN was
applied to OUTBOUND walk and REST. The networks were reseeded 10 times each. In each
reseeding, half of the examples were used as training, the other half were used as validation set
to call back the best mode by performance on the validation set. This resulted in multiple, highly
correlated features for each task.
Subchallenge 2.1 (Tremor): Balint Armin Pataki
The creation of the winning features by team ‘Balint Armin Pataki’ was based on signal
processing techniques. As the tremor of PD is a repetitive action added to the normal hand
movements of a person, it can be described well in the frequency space via Fourier
transformation. The main created features were the intensities of the Fourier spectrum at
frequencies between 4 and 20 Hz. Observing high intensities at a given frequency suggests that
there is a strong hand movement which repeats at that given frequency. Additionally, hundreds
of features were extracted from the accelerometer tracks via the tsfresh package26. Finally,
clinical feature descriptors were created by mean-encoding and feature-binarizing the
categorical clinical data provided via the scikit-learn package24. This resulted in 20
clinical-derived features, 99 Fourier spectrum-based features, and 2896 features derived from
tsfresh. In order to eliminate those which were irrelevant, a Random Forest classifier was
applied, which selected 81 features (3 clinical-derived, 6 Fourier-derived and 72 tsfresh-derived)
from the ~3000 generated.
Subchallenge 2.2 (Dyskinesia): Jennifer Schaff
Data was captured using GeneActive and Pebble watch devices along several axes of
motion, including the movement to the right (Y-axis). Because either of these devices could be
worn on the right or left wrist, an additional ‘axis’ of data was created to capture motion relative
to movement towards or away the center of the body. This Y-axis-alt data was calculated by
multiplying the Y-axis by -1 in patients that wore the device on the wrist for which the particular
device (GeneActive or Pebble) occurred less frequently. In other words, if the GeneActive was
most frequently worn on the right wrist, Y-axis measurements for left-worn measurements were
multiplied by -1.
To distinguish between choreic and purpose driven movements, summary statistics of
movement along each axis per approximate second were generated, and a selection process to
identify features that had predictive potential for dyskinesia was applied. For each separately
recorded task (set of patient, visit, session, and task), the absolute value of the lagged data
point for each axis was calculated, and the standard deviation, variance, minimum value,
maximum value, median, and sum were recorded for all variables over each approximate rolling
second (51 data points). Additional features were derived by log transformation of the previously
generated individual-second features. To summarize across the 51 individual-second values for
a given task, the individual-second features were aggregated using the mean, median, sum,
standard deviation, the median absolute deviation, the max, as well as each statistic taken over
the absolute value of each observation for each variable (both original and calculated), resulting
in approximately 1966 variables as potential features.
Random Forest model selection, as implemented Boruta package 27 in R, was used to
reduce the number of features while still retaining any variable the algorithm found to have
predictive value. Any feature that was chosen by Boruta in more than 10 of 25 Boruta iterations
was selected for submission, resulting in 389 variables. ‘Site’, ‘visit’, ‘session’, ‘device’, and
‘deviceSide’ as well as an indicator of medication usage were including bringing the number of
variables to 395. Features were calculated and selected for each device separately (to reduce
dependency on computational resources).
Subchallenge 2.3 (Bradykinesia): Team Vision
The method by team ‘Vision’ derived features using spectral decomposition for time series
and applied a hybrid logistic regression model to adjust for the imbalance in number of repetitions across
different tasks. Spectral analysis was chosen for its ability to decompose each time series into periodic
components and generate the spectral density of each frequency band, and determine those frequencies
that appear particularly strong or important. Intuitively, the composition of frequencies of periodic
components should shed light on the existence of Bradykinesia, if certain range of frequencies stand out
from the frequency of noise. Spectral decomposition was applied to the acceleration data on three axes X
(forward/backward), Y (side-to-side), Z (up/down). Each time series was first detrended using smoothing
spline with a fixed tuning parameter. The tuning parameter was set to be relatively large to ensure a
smooth fitted trend, so that the detrended data keep only important fluctuations. Specifically, the ‘spar’
parameter was set to 0.5 in smooth.spline function. It was selected by cross validation, and the error was
not sensitive with spar bigger than 0.5. The tuning parameter was set the same across the tasks and
selected by cross-validation. The detrended time series were verified to be consistent with an
autoregressive-moving-average (ARMA) model to ensure process stationarity. Following spectral
decomposition, the generated features were summarized as the maximum, mean and area of estimated
spectral density within five intervals of frequency bands [0, 0.05), [0.05, 0.1), [0.1, 0.2), [0.2, 0.3), [0.3, 0.4),
[0.4, 0.5]. These intervals cover the full range of the spectral density. Because the importance of each
feature is different for each task, features were normalized by the estimated coefficient derived by fitting
separate multivariate logistic regression models for each task. Class prediction was then made based on
the normalized features using logistic regression.
Analysis of methods used by participants
We surveyed participants regarding approaches used. Questions in the survey pertained
to the activities used (e.g. walking outbound, inbound or rest for the mPower data), the sensor
data used (e.g. device motion, user acceleration, gyroscope, pedometer, etc), and the methods
for extracting features from the selected data types, including pre-processing, feature
generation and post-processing steps. A one-way ANOVA was conducted to determine if any
the use of a particular sensor, activity or approach was associated with better performance in
the challenge. Significance thresholds were adjusted for multiple test correction using a
Bonferroni correction factor of 4, and no significant associations were found in any subchallenge
(p-value > 0.05 for all comparisons). We further clustered teams based on overall approach
incorporating all of the dimensions surveyed. Hierarchical clustering was performed in R using
the ward.d2 method and Manhattan distance. Four and three clusters were identified in SC1
and SC2, respectively. One-way ANOVA was then used to determine whether any cluster
groups showed significantly different performance. No significant difference in mean scores
across clusters was identified (p-value > 0.05 for all tests).
Saliency mapping of ‘Yuanfang Guan and Marlena Duda’ model
We applied saliency mapping28, a simple approach for characterization of patterns learned by
convolutional neural network (CNN) models which provides interpretability to these otherwise
“black box” models, to the winning CNN model for SC1 for all data samples in both the training
and testing sets of both the outbound and rest tasks in order to understand which aspects of the
walking and rest data were most informative in the prediction of PD status. The salience values
were computed as the gradient of the model output with respect to the model input, and “high
saliency” regions were identified by applying windowed maximum thresholding using a window
size of 30, a step size of 30 and a threshold of 0.1 to define highly salient regions. These
represent the time windows for each task for which a small change in the input value results in a
large change in the model output.
Univariate analysis of submitted features
A univariate analysis of all submitted features was performed by, on a feature-by-feature
basis, fitting a generalized linear model (GLM), either logistic for SC1, SC2.2 and SC2.3 or
multi-class logistic model for SC2.1, using the training samples, and predicting in the test
samples. AUROC was used to measure accuracy in SC1 whereas AUPR was used in
SC2.1-2.3. For SC2.1-2.3 only features from the top 10 teams were assessed. Features
occurring in multiple submissions (e.g. present in both submissions from the same team) were
evaluated only once to avoid double counting.
Identification of optimal feature sets
In total, thousands of features were submitted for each challenge. To determine if an
optimal subset of features (as defined by having a better AUPR than that achieved by individual
teams) could be derived from the set of all submitted features, two different feature selection
approaches were taken to identify whether choosing from all the submitted features could result
in better predictive performance. These feature selection approaches were applied using only
the training data to optimize the selection, and were evaluated in the test set according to the
Challenge methods.
First, the Boruta random forest algorithm 27 was tested on the entire set of submitted
features for SC2.2 (2,865), and 334 all-relevant features were selected in at least ten of 25
iterations. Recursive Feature Elimination (RFE) (i.e. simple backward selection) using accuracy
as the selection criteria as implemented in the caret package29 of R was then applied to the
downsized feature set and selected four of the 334 features as a minimal set of features. The
feature sets were then scored in the testing set per the Challenge scoring algorithms, achieving
AUPR of 0.38 and 0.35 for the larger and smaller sets, respectively, placing behind the top eight
and twelve individual submissions for SC2.2.
A second approach applied PCA (Principal Component Analysis) to the entire sets of
features submitted for sub-challenges 2.1, 2.2, and 2.3 separately. Non-varying features were
removed prior to application of PCA. Each PC imparted only an incremental value towards the
cumulative proportion of variance (CPV) explained ([maximum, 2nd, 3rd,..., median] value [14%,
7%, 4%,..., 0.0027%], [15%, 13%, 5%,..., 0.0014%] and [15%, 7%, 6%,..., 0.00039%] for SC2.1,
SC2.2 and SC2.3, respectively), suggesting wide variability in the feature space, and the top 20
PCs from each sub-challenge explained 49%, 66% and 61% of the cumulative variance for
SC2.1, SC2.2 and SC2.3, respectively. Then used the top number of PCs explaining
approximately ⅔ of the variation PCs as meta-features in each subchallenge (50, 20 and 30 for
SC2.1, SC2.2 and SC2.3, respectively), scoring against the Challenge test set. These achieved
an AUPR of 0.674 for SC2.1 (below the top five submission scores of 0.730-0.750), an AUPR of
0.504 AUPR for SC2.2 (above the top 5 feature submissions of 0.402-0.477) and an AUPR of
0.907 for SC2.3 (within the range of the top 5 feature submissions of 0.903-0.950).
Clustering of features
We performed a clustering analysis of all the features from SC1 using k-means and
bisecting k-means with random initialization to understand the landscape of features. To map
the input feature space to two dimensions for visualization while preserving the local distances,
we employed two manifold projection techniques: metric Multi-Dimensional Scaling (MDS) 30
and t-Distributed Stochastic Neighbor Embedding (t-SNE) 31 with various settings for perplexity,
PCA dimensions, and feature standardization. The outcomes of these projections were then
clustered with k-means and bisecting k-means with k = 2, 5, 10, and 20, using silhouette width 32
as a cluster validity index to select the optimal number of clusters. A Kruskal-Wallis rank sum
test was used to associate cluster membership with a feature’s submission score taken as the
performance of it’s associated feature set, however individual feature scores were also
examined. Hot-spots were identified by binning the projected plane and smoothing the
performance by a simple mean. The significance of association between the team associated
with a feature (as well as the predictive performance) with the cluster membership tends to
generally increases with the number of clusters used. Clustering without PCA gives more
compact and well separated clusters and the optimal k tested by the silhouette validity index is
estimated to be around 10. The clusters visualized as interactive charts are available online at
https://ada.parkinson.lu/pdChallenge/clusters and the correlation networks at
https://ada.parkinson.lu/pdChallenge/correlations.
Visualizations of feature clusters and aggregated correlations were carried out by Ada
Discovery Analytics (https://ada-discovery.github.io), a performant and highly customizable data
integration and analysis platform.
Topological Data Analysis of mPower features
To construct the topological representation, we leveraged the open source R
implementation of the mapper algorithm11 (https:// github.com/paultpearson/TDAmapper). As a
preprocessing step, we considered only the features (median value per subject) from the six top
performing submissions in SC1, and centered and scaled each feature to obtain a z-score. We
then reduced the space to two dimensions using multi-dimensional scaling (MDS) and binned
the space into 100 (10x10) equally sized two-dimensional regions. The size of the bins was
selected so that they have 15% overlap in each axis. A pairwise dissimilarity matrix based on
Pearson correlation was calculated as 1-r from the original multi-dimensional space, and used
to cluster the samples in each bin individually (using hierarchical single-linkage clustering). A
network was generated considering each cluster as a node while forming edges between nodes
that share at least one sample. Finally, we pruned the network by removing duplicate nodes and
terminal nodes which only contain samples that are already accounted for (not more than once)
in a paired node. We used the igraph R package (http://igraph.org/r/) to store the network data
structure and Plotly's R graphing library (https://plot.ly/r/) to render the network visualization.
Medication effects in mPower
For each submitted model to SC1, PD status was predicted for all individual walking
tests in the mPower Study, regardless of reported medication status. We tested whether
predicted PD status differed between Parkinson’s patients on medication (self reported status:
“Just after Parkinson medication (at your best)") or off medication (self reported status:
"Immediately before Parkinson medication" or "I don't take Parkinson medications") using a
linear mixed model with healthCode (individual) as a random effect to account for repeated
measures. We also obtained a list of individuals for whom medication status could reliably be
predicted (at 5% and 10% FDR)13, and repeated the analysis in this subset of individuals.
Results were not significant using the full set, as well as the two subsets, for any of the top 10
models, which implies that the models optimized to predict PD status could not be immediately
extrapolated to predict medication status.
Demographic subgroup analysis in mPower
For each feature set, the predicted class probabilities generated by the scoring algorithm
(see ‘Submission Scoring’) were used to compute AUROC within demographic subgroups by
subject age group (57-60, 60-65, 65-70, and 75+) and gender (Female and Male). The same
approach was used to assess the Demographic model against which the feature sets were
compared. For the purposes of this analysis, we only considered submissions which
outperformed the Demographic model.
Medication effects in L-dopa
Medication effect on prediction accuracy in L-dopa data (Supplementary Figure 8) was
evaluated by investigating how prediction accuracy changed as medication took effect or wore
off over sessions during the two visits. For each task repetition, average prediction accuracy
was defined as the average of absolute differences between known and predicted scores over
submissions that outperformed demographic baseline model. In SC2.2-2.3, the symptom
probabilities generated by the challenge scoring model (see ‘Scoring’ on the Online Methods)
were used as predicted scores, whereas in SC2.1, the predicted score was calculated as the
expectation.
Analysis of study tasks in L-dopa
For SC2.1-SC2.3, each feature set was re-fit and rescored within task. 1000 bootstrap
iterations were performed to assess the variability of each task score for each submission. On
each iteration, expected AUPR was computed based on the class distributions of the bootstrap
sample. For comparison of 2 tasks for a given submission, a bootstrap p-value was computed
as the proportion of bootstrap iterations in which AUPR(task1)-E[AUPR(task1)] >
AUPR(task2)-E[AUPR(task2)], and the overall significance of the comparison between task1
and task2 was assessed via one-sided Kolmogorov-Smirnov test of the distribution, across
submissions, of the p-values vs a U[0,1] distribution.
Acknowledgements
The Parkinson’s Disease Digital Biomarker challenge was funded by the Robert Wood Johnson
Foundation and the Michael J. Fox Foundation. Data were contributed by users of the
Parkinson mPower mobile application as part of the mPower study developed by Sage
Bionetworks and described in Synapse [doi:10.7303/syn4993293].
Author Contributions
JFD, FP, GC, FNG, SS, GVD, and PB designed the L-dopa Response Study and collected the
data.
SKS, PS, JFD, FP, GC, UR, PB, YC, ECN, RD, FNG, ER, GVD, DB, PB, LM, and LO designed
the challenge.
SKS, JS, MD, BAP, MS, PS, UR, PB, ZA, AC, LLE, CE, EG1, EG2, YG, MKJ, JJ, RK, DL, CMD,
DP, NMR, PS, NS, MSV, YZ, the Parkinson’s Disease Digital Biomarker Challenge Consortium,
YW, YG, DB, and LO analyzed the data.
The Parkinson’s Disease Digital Biomarker Challenge Consortium
Avner Abrami1, Aditya Adhikary2, Carla Agurto1, Sherry Bhalla2, Halil Bilgin3, Vittorio Caggiano1,
Jun Cheng4, Eden Deng5, Qiwei Gan6, Rajan Girsa2, Zhi Han7,8, Stephen Heisig1, Kun Huang7,
Samad Jahandideh9, Wolfgang Kopp10, Christoph F. Kurz11,12, Gregor Lichtner13, Raquel Norel1,
G.P.S Raghava2, Tavpritesh Sethi2, Nicholas Shawen14,15, Vaibhav Tripathi2, Matthew Tsai5,
Tongxin Wang16, Yi Wu7, Jie Zhang17, Xinyu Zhang18
1 IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA
2 Centre for Computational Biology, Indraprastha Institute of Information Technology Delhi, New
Delhi, Delhi, India, 110020
3 Department of Computer Engineering, Abdullah Gul University, Kayseri, Turkey, 38090
4 School of Biomedical Engineering, Shenzhen University, Shenzhen, Guangdong, China, 518055
5 Canyon Crest Academy, San Diego, CA 92130, USA
6 Department of Management Information Systems, Utah State University, Old Main Hill
Logan, Utah 84322, USA
7 Department of Medicine, Indiana University School of Medicine, Indianapolis, Indiana 46202, USA
8 Regenstrief Institute, Indianapolis, Indiana, 46202, USA
9 Predex Pharma LLC, Gaithersburg, MD, USA
10 BIMSB, Max Delbrueck Center for molecular medicine, Berlin, Germany, 10115
11 Institute of Health Economics and Health Care Management, Helmholtz Zentrum München, 85764
Neuherberg, Germany
12 Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and
Women’s Hospital and Harvard Medical School, Boston, MA, USA
13 Charité – Universitätsmedizin Berlin, Klinik für Anästhesiologie mit Schwerpunkt operative
Intensivmedizin (CCM, CVK), Berlin, Germany, 10117
14 Rehabilitation Technologies and Outcomes Lab, Shirley Ryan AbilityLab, Chicago, Illinois 60611, USA
15 Medical Scientist Training Program, Northwestern University Feinberg School of Medicine, Chicago,
Illinois 60611, USA
16 Department of Computer Science, Indiana University Bloomington, Bloomington, Indiana 47408, USA
17 Department of Medical and Molecular Genetics, Indiana University School of Medicine, Indianapolis,
Indiana 46202, USA
18 Department of Psychiatry, Yale School of Medicine, New Haven, CT 06511, USA
Competing Interests
References
1.
Sherman, R. E. et al. Real-World Evidence - What Is It and What Can It Tell Us? N. Engl. J.
Med. 375, 2293–2297 (2016).
2.
Goldsack, J. et al. Digital endpoints library can aid clinical trials for new medicines - STAT.
STAT
https://www.statnews.com/2019/11/06/digital-endpoints-library-clinical-trials-drug-developm
ent/ (2019).
3.
Zhan, A. et al. Using Smartphones and Machine Learning to Quantify Parkinson Disease
Severity: The Mobile Parkinson Disease Score. JAMA Neurol. 75, 876–880 (2018).
4.
Lipsmeier, F. et al. Evaluation of smartphone-based testing to generate exploratory
outcome measures in a phase 1 Parkinson’s disease clinical trial. Movement Disorders vol.
33 1287–1297 (2018).
5.
Norel, R., Rice, J. J. & Stolovitzky, G. The self-assessment trap: can we all be better than
average? Molecular Systems Biology vol. 7 537 (2011).
6.
PRO Consortium | Critical Path Institute. https://c-path.org/programs/proc/.
7.
Bot, B. M. et al. The mPower study, Parkinson disease mobile data collected using
ResearchKit. Sci Data 3, 160011 (2016).
8.
Daneault, J.-F. et al. The Levodopa Response Study: Part I - Data Collected with a
Minimum Set of Wearable Sensors. Nature Scientific Data.
9.
Vergara-Diaz, G. et al. The Levodopa Response Study: Part II - Data Collected with
Wearable Sensors on the Limbs and Trunk. Nature Scientific Data.
10. Ahlskog, J. E. & Muenter, M. D. Frequency of levodopa-related dyskinesias and motor
fluctuations as estimated from the cumulative literature. Mov. Disord. 16, 448–458 (2001).
11. Singh, G., Mémoli, F. & Carlsson, G. E. Topological Methods for the Analysis of High
Dimensional Data Sets and 3D Object Recognition. in SPBG (2007).
12. Chaibub Neto, E. et al. Personalized Hypothesis Tests For Detecting Medication Response
in Parkinsons Disease Patients using iPhone Sensor Data. Pac. Symp. Biocomput. 21,
273–284 (2016).
13. Omberg, L. et al. mPower: a smartphone approach to remotely monitor Parkinson Disease
and individual response to therapy. Nature Biotech.
14. Camps, J. et al. Deep learning for freezing of gait detection in Parkinson’s disease patients
in their homes using a waist-worn inertial measurement unit. KNOWLEDGE-BASED
SYSTEMS 139, 119–131 (2017).
15. Goodfellow, I., Bengio, Y. & Courville, A. Deep Learning. (MIT Press, 2016).
16. Manap, H. H., Tahir, N. M. & Yassin, A. I. M. Statistical analysis of parkinson disease gait
classification using Artificial Neural Network. in 2011 IEEE International Symposium on
Signal Processing and Information Technology (ISSPIT) 060–065 (2011).
17. Factor, S. A., Jennings, D. L., Molho, E. S. & Marek, K. L. The Natural History of the
Syndrome of Primary Progressive Freezing Gait. Arch. Neurol. 59, 1778–1783 (2002).
18. Bartels, A. L. et al. Relationship between freezing of gait (FOG) and other features of
Parkinson’s: FOG is not correlated with bradykinesia. J. Clin. Neurosci. 10, 584–588
(2003).
19. Schaafsma, J. D. et al. Characterization of freezing of gait subtypes and the response of
each to levodopa in Parkinson’s disease. Eur. J. Neurol. 10, 391–398 (2003).
20. Powers, I. I. I. et al. PASSIVE TRACKING OF DYSKINESIA/TREMOR SYMPTOMS. Patent
(2019).
21. Jenkinson, C., Fitzpatrick, R., Peto, V., Greenhall, R. & Hyman, N. The PDQ-8:
Development and validation of a short-form parkinson’s disease questionnaire. Psychol.
Health 12, 805–814 (1997).
22. Goetz, C. G. et al. Movement Disorder Society-sponsored revision of the Unified
Parkinson’s Disease Rating Scale (MDS-UPDRS): scale presentation and clinimetric testing
results. Mov. Disord. 23, 2129–2170 (2008).
23. Neto, E. C. et al. Learning Disease vs Participant Signatures: a permutation test approach
to detect identity confounding in machine learning diagnostic applications. arXiv [stat.AP]
(2017).
24. Pedregosa, F. et al. Scikit-learn: Machine Learning in Python. J. Mach. Learn. Res. 12,
2825–2830 (2011).
25. Davis, J. & Goadrich, M. The relationship between Precision-Recall and ROC curves. in
Proceedings of the 23rd international conference on Machine learning - ICML ’06 233–240
(ACM Press, 2006).
26. Christ, M., Braun, N., Neuffer, J. & Kempa-Liehr, A. W. Time Series FeatuRe Extraction on
basis of Scalable Hypothesis tests (tsfresh – A Python package). Neurocomputing 307,
72–77 (2018).
27. Kursa, M. B., Rudnicki, W. R. & Others. Feature selection with the Boruta package. J. Stat.
Softw. 36, 1–13 (2010).
28. Simonyan, K., Vedaldi, A. & Zisserman, A. Deep Inside Convolutional Networks: Visualising
Image Classification Models and Saliency Maps. arXiv [cs.CV] (2013).
29. Kuhn, M. Building Predictive Models in R Using the caret Package. Journal of Statistical
Software, Articles 28, 1–26 (2008).
30. Cox, T. F. & Cox, M. A. A. Multidimensional Scaling, Second Edition. (CRC Press, 2000).
31. van der Maaten, L. J. P. & Hinton, G. E. Visualizing High-Dimensional Data Using t-SNE. J.
Mach. Learn. Res. 9 (No), 2579–2605 (2008).
32. Rousseeuw, P. J. Silhouettes: A graphical aid to the interpretation and validation of cluster
analysis. J. Comput. Appl. Math. 20, 53–65 (1987).
(6 display items)
Figure Legends
Tables
Table 1: mPower data demographics
Training
Test
PD
Control
PD
Control
Age
60.6 +/- 10.7
34.7 +/- 14.2
60.4 +/- 11.9
34.9 +/- 14.4
Sex
Male
439 (66.5%)
1755 (81.4%)
377 (61.4%)
1071 (78.2%)
Female
219 (33.2%)
397 (18.4%)
226 (36.8%)
285 (20.8%)
Unspecified
2 (0.3%)
3 (0.1%)
11 (1.8%)
14 (1.0%)
Race
Caucasian
586 (88.8%)
1521 (70.6%)
533 (86.8%)
870 (63.5%)
Other or Mixed
74 (11.2%)
634 (29.4%)
81 (13.2%)
500 (36.5%)
Marital Status
Single
30 (4.5%)
993 (46.1%)
17 (2.8%)
628 (45.8%)
Married/Domestic Partnership
534 (80.9%)
1022 (47.4%)
271 (44.1%)
571 (41.7%)
Divorced/Separated/Widowed
87 (13.2%)
112 (5.2%)
41 (6.7%)
68 (5.0%)
Other/Unreported
9 (1.4%)
28 (1.3%)
285 (46.4%)
103 (7.5%)
Education
High School or less
45 (6.8%)
278 (12.9%)
44 (7.1%)
224 (16.4%)
College or college degree
281 (42.6%)
1227 (56.9%)
270 (44.0%)
727 (53.1%)
Graduate school or degree
334 (50.6%)
650 (30.1%)
300 (48.9%)
419 (30.6%)
Figures
Figure 1: For each subchallenge, data were split into training and test portions. Participants
were provided with the mobile sensor data for both the training and test portions, along with the
demographic and meta-data, and diagnosis or severity labels for the training portion of the data
only. Participants were asked to derive features from the mobile sensor data for both the
training and test portions of the data. These features were then used to train a classifier, using a
standard suite of algorithms, to predict disease status or symptom severity, and predict labels in
the testing portion of the data. Submissions were scored based on the accuracy of the resulting
predictions.
Figure 2: Bootstraps of the submissions for (A) SC1, (B) SC2.1, (C) SC2.2, and (D) SC2.3
ordered by submission rank. For each subchallenge, a model using only demographic and
meta-data is displayed in red as a benchmark.
Supplementary Material
Supplementary Table 1: Tremor subtask p-values (bonferroni corrected)
fldng
drnkg
ntblt
ram
ftn
orgpa
1.90E-09
1.30E-17
8.46E-26
9.17E-28
8.01E-28
fldng
5.34E-3
7.10E-12
2.39E-20
8.08E-24
drnkg
1.39E-09
7.38E-19
2.87E-21
ntblt
4.00E-3
5.30E-06
ram
1
Supplementary Table 2: Bradykinesia subtask p-values (bonferroni corrected)
Task1
ftn
ram
fldng
drnkg
orgpa
1.34E-3
8.69E-10
3.67E-10
1.07E-11
ftn
1.40E-10
1.89E-09
7.50E-11
ram
0.605
1.16E-4
fldng
0.152
Supplementary Figure 1: Clustering of methodological approach for (A) SC1 and (B) SC2.1-2.3
shows no association with submission performance.
Supplementary Figure 2: AUROC score of the top 100 single features in SC1 sorted by rank.
Dots are colored by method (A) and by team (B).
Supplementary Figure 3: Two-dimensional t-SNE projections of mPower features grouped to (A)
10 clusters produced by k-means clustering algorithm for the 35 top submissions. In (B) the
same projection is displayed with points colored by associated team, and in (C) a 20-by-20
mean-aggregated performance (AUROC) heatmap shows a visible hot-spot in the top-right
corner.
Supplementary Figure 4: AUPR score of the top 100 single features in SC2.1 (A-B), SC2.2
(C-D) and SC2.3 (E-F) sorted by rank. Dots are colored by method (A,C,E) and by team (B,D,F).
Supplementary Figure 5: Topological representation of the features space from the top six SC1
submissions labeled by professional diagnosis. Each node corresponds to a group of subjects
with similar feature space and edges connect nodes that share at least one subject. Nodes are
colored by the professional diagnosis ratio in each node, where blue represents controls and
red are PD subjects. Node size represents the number of samples within each node.
Supplementary Figure 6: Topological representation of the features space from the top six SC1
submissions labeled by professional diagnosis split into two sets: (a) the on-meds set which
includes sessions in which the subjects have just taken their medicine and (b) off-meds set as
defined by sessions in which the subjects were tested right before taking medication or not
taking medication at all. Given that three of the top six submissions (Yuanfang Guan and
Marlena Duda 1, Yuanfang Guan and Marlena Duda 2 and wangsijia1990) have the same
values for the features on both sets, and therefore are a confounding factor when looking for
differences between the two sets, we only considered the remaining 3 (ethz-dreamers 1,
ethz-dreamers 2 and vmorozov). Both sets included the same control population. Nodes are
colored by the professional diagnosis ratio in each node, where blue represents controls and
red are PD subjects. Node size represents the number of samples within each node. There are
no apparent medication effects.
Supplementary Figure 7: Performance of top models (those outperforming the
demographics-only model) in demographic subgroups by age and gender. The red circle
indicates the performance of the top-performing model by team Yuanfang Guan and Marlenda
Duda, and the red star indicates the score in the Demographic-only model. These top models
perform best, relative to the Demographic model, in younger age groups and in Male subjects.
The winning model performs well in well-represented subgroups, but performs especially poorly
in oldest subgroups, which have the fewest samples.
Supplementary Figure 8: Improvement over null expectation as a fraction of maximum possible
increase (i.e. (AUPR-E[AUPR])/(1-E[AUPR])) by subtask for all submissions for (A) SC2.1, (B)
SC2.2 and (C) SC2.3 for tasks: pouring water and drinking (drnkg), folding laundry (fldng),
finger-to-nose (ftn), assembling nuts and bolts (ntblt), organizing papers (orgpa), and alternating
hand movements (ram). The red star indicates the model containing only demographic and
meta-data. For prediction of tremor severity, practical tasks like assembling folding laundry and
pouring water were more predictive than contrived tasks like finger-to-nose and alternating hand
movements. For Bradykinesia, finger-to-nose and organizing paper showed the best
improvement over expectation as well as over the demographic model. For dyskinesia, in which
the resting hand was used to classify symptom presence, both tasks performed equally well.
| 2020 | Crowdsourcing digital health measures to predict Parkinson’s disease severity: the | 10.1101/2020.01.13.904722 | null | creative-commons |
Submitted Manuscript: Confidential
Template revised February 2021
1
Title: Three-dimensional genome re-wiring in loci with Human Accelerated
Regions
Authors: Kathleen C. Keough1,2,3, Sean Whalen1, Fumitaka Inoue2,3†, Pawel F. Przytycki1^,
Tyler Fair4,5, Chengyu Deng2,3, Marilyn Steyert4,5,6,7, Hane Ryu2,3, Kerstin Lindblad-Toh8,9,
Elinor Karlsson9,10,11, Zoonomia Consortium, Tomasz Nowakowski4,5,6,7, Nadav Ahituv2,3, Alex
Pollen7,12, Katherine S. Pollard1,3,13,14*
Affiliations:
1Gladstone Institute of Data Science and Biotechnology; San Francisco, California, USA.
2Department of Bioengineering and Therapeutic Sciences, University of California San
Francisco, San Francisco, California, USA
3Institute for Human Genetics, University of California San Francisco, San Francisco,
California, USA
4Department of Neurological Surgery, University of California, San Francisco, CA, USA
5Department of Anatomy, University of California, San Francisco, CA, USA
6Department of Psychiatry & Behavioral Sciences, University of California, San Francisco,
CA, USA
7Eli and Edythe Broad Center for Regeneration Medicine and Stem Cell Research, University
of California, San Francisco, CA, USA
8Science for Life Laboratory, Department of Medical Biochemistry and Microbiology,
Uppsala University; Uppsala, 751 32, Sweden
9Broad Institute of MIT and Harvard; Cambridge, MA 02139, USA
10Program in Bioinformatics and Integrative Biology, UMass Chan Medical School;
Worcester, MA 01605, USA
11Program in Molecular Medicine, UMass Chan Medical School; Worcester, MA 01605,
USA
12Department of Neurology, University of California, San Francisco, CA, USA
13Department of Epidemiology & Biostatistics and Bakar Institute for Computational Health
Sciences, University of California, San Francisco, CA, USA
14Chan Zuckerberg Biohub, San Francisco, CA, USA
†Present address: Institute for the Advanced Study of Human Biology (WPI-ASHBi), Kyoto
University, Kyoto, Japan
^Present address: Faculty of Computing & Data Sciences, Boston University, Boston, MA,
USA
*Corresponding author. Email: katherine.pollard@gladstone.ucsf.edu
Submitted Manuscript: Confidential
Template revised February 2021
2
Abstract: Human Accelerated Regions (HARs) are conserved genomic loci that evolved at an
accelerated rate in the human lineage and may underlie human-specific traits. We generated
HARs and chimpanzee accelerated regions with the largest alignment of mammalian genomes to
date. To facilitate exploration of accelerated evolution in other lineages, we implemented an
open-source Nextflow pipeline that runs on any computing platform. Combining deep-learning
with chromatin capture experiments in human and chimpanzee neural progenitor cells, we
discovered a significant enrichment of HARs in topologically associating domains (TADs)
containing human-specific genomic variants that change three-dimensional (3D) genome
organization. Differential gene expression between humans and chimpanzees at these loci in
multiple cell types suggests rewiring of regulatory interactions between HARs and
neurodevelopmental genes. Thus, comparative genomics together with models of 3D genome
folding revealed enhancer hijacking as an explanation for the rapid evolution of HARs.
One-Sentence Summary: Human-specific changes to 3D genome organization may have
contributed to rapid evolution of mammalian-conserved loci in the human genome.
Submitted Manuscript: Confidential
Template revised February 2021
3
Main Text: Human accelerated regions (HARs) are genomic loci that were conserved over
millions of years of vertebrate evolution but evolved quickly in the human lineage, and thus are
of great interest based on their potential to underlie human-specific traits (1–8). Many HARs are
predicted to function as gene enhancers, particularly for genes implicated in neural development
(9). Furthermore, most HARs appear to have evolved under positive selection due to having
more human substitutions than expected given the local neutral rate (10), an indication that the
sequence changes were beneficial to ancient humans. However, the mechanisms facilitating their
shift in selective pressure after millions of years of constraint remains to be determined.
Structural variation is a substantial driver of genome evolution. The majority of genomic
differences between humans and our closest extant relative, the chimpanzee, derive from
structural variation, largely in the noncoding genome (11). Changes to genome organization
mediated by structural variants can rewire gene regulatory networks through “enhancer
hijacking”, or “enhancer adoption”, through which genes gain or lose regulatory signals,
affecting spatiotemporal gene expression (12–14). Enhancer hijacking has been identified as a
contributing factor to cancer and other human diseases (12, 15–17), and previous work proposed
that it may be a driver of species evolution (7, 18, 19). For example, the locus containing the
cluster of Hox genes is encompassed in a single topologically associating domain (TAD) in the
bilaterian ancestor, but vertebrates have two separate TADs; this difference may have driven
evolutionary innovations in developmental body patterning specific to vertebrates (18, 20, 21).
Recent work comparing multiple great ape genomes identified a high quality set of 17,789
human-specific structural variants (hsSVs) (22). We hypothesized that some HARs were
hijacked due to hsSVs, changing their target gene repertoire and subjecting them to different
selective pressures in humans, thus driving their human-specific accelerated evolution.
To test this hypothesis, we leverage the largest alignment of mammalian genomes to date,
Zoonomia (23). We first identify an updated set of HARs (zooHARs) and chimpanzee
accelerated regions (zooCHARs), and develop an open-source Nextflow pipeline for
reproducible and streamlined identification of accelerated regions (ARs) in any lineage using
large multiple sequence alignments. We find that TADs containing hsSVs are enriched for
zooHARs. Using Akita, a deep learning model of three-dimensional (3D) genome folding, we
predict that multiple hsSVs change the chromatin interactions of zooHARs and zooCHARs. We
then validate these predictions by generating high-resolution chromatin capture (Hi-C) data from
human and chimpanzee induced pluripotent stem cell derived neural progenitor cells (NPCs) at
matched developmental time points and show that differentially expressed genes from NPCs (24)
and cerebral organoids (25) are enriched in TADs containing zooHARs and hsSVs (Chi-squared
p-value < 0.05). By integrating a machine learning model of enhancer activity, a network-based
cell type labeling method, and a massively parallel reporter assay (MPRA) performed on primary
cells from the human mid-gestation telencephalon, we characterize the regulatory activity of
zooHARs and zooCHARs in specific neuronal cell types. Taken together, these results implicate
enhancer hijacking as a genetic mechanism to explain the lineage-specific accelerated evolution
of many HARs, potentially underlying human-specific neurodevelopmental phenotypes.
Submitted Manuscript: Confidential
Template revised February 2021
4
Human accelerated regions are enriched in 3D topological associating domains with
human-specific structural variants
The identification of species-specific accelerated regions in alignments containing many species
with large genomes requires significant computational resources. Pipeline management software
enables analyses like these to be made portable to different parallel computing environments
(26). Therefore, we compiled previously developed methods for detecting accelerated regions (1,
27–29) into a new Nextflow pipeline and optimized modeling parameters in the Phylogenetic
Analysis with Space/Time models (PHAST) software package for large multiple sequence
alignments, creating a scalable software tool for identification of lineage-specific accelerated
elements in any species on any computing platform (Fig. 1A, Supplemental Text).
We then leveraged the Zoonomia alignment of 241 mammal genomes to identify 312 zooHARs
and 141 zooCHARs (Table S1, Table S2). These ARs demonstrate similar features to previous
sets of HARs, including being mainly noncoding, having signatures of positive selection (82% of
zooHARs and 86% of zooCHARs), and being located near genes involved in developmental and
neurological processes (Fig. S1-3) (6, 9, 10). Approximately one-third of zooHARs and
zooCHARs are transcribed in the developing human neocortex (Fig. S1E-F). The median
distance between zooHARs and zooCHARs is significantly less than expected (1.05Mb,
bootstrap p-value=0.02, both in hg38), as observed in previous sets of primate accelerated
regions (30). Genes near both zooHARs and zooCHARs are significantly enriched for roles in
transcriptional regulation (hypergeometric tests (31, 32); Fig. S2, 3). As human and chimpanzee
ARs demonstrate similar characteristics, the smaller number of zooCHARs is likely attributable
to the lower quality of the chimpanzee reference genome and the strict filtering we performed,
though the annotations of genes nearby zooHARs suggest connections to a broader diversity of
developmental processes compared to zooCHARs. Together these analyses demonstrate that
zooHARs identified from an alignment of 241 mammals demonstrate features consistent with
previous studies proposing gene regulatory functionality, particularly in neurodevelopment.
Genomic loci near duplicated genes have been shown to evolve rapidly, suggesting synergy
between structural variation and sequence-based genome evolution (33). To explore this, we
sought to determine whether zooHARs and hsSVs tended to co-locate in the context of the 3D
genome. Using a high-quality set of TADs from lymphoblastoid cells (34), we found that
zooHARs are strongly enriched in TADs with hsSVs relative to the set of conserved (phastCons)
elements from which zooHARs are identified (odds ratio = 3.0, bootstrap p-value < 0.001, Fig.
1B). This enrichment is robust to repeating the analysis with TADs from other cell types,
including primary mid-gestation telencephalon, and a different TAD-calling method, but it is not
observed with random genomic windows (Fig. S4). To determine whether the enrichment is
simply driven by localization of hsSVs near zooHARs in the 1D genome sequence, we replaced
the TADs with random size-matched windows and found that zooHARs were not significantly
enriched in this context relative to phastCons elements (fig. S4D-E). Thus, we conclude that
zooHARs are specifically enriched in TADs with hsSVs, suggesting a role for 3D genome
organization and structural variation in the accelerated evolution of HARs.
Submitted Manuscript: Confidential
Template revised February 2021
5
Human-specific structural variants are predicted to have changed the 3D chromatin
environment of zooHARs
Structural variants are the main contributor to genome-wide genetic divergence between the
human and chimpanzee genomes (11), and they have the potential to generate large changes in
3D genome organization through disruption of insulating boundaries or other structural motifs
(35). Based on our observation that zooHARs are enriched in TADs with hsSVs, we sought to
determine whether hsSVs may have generated changes in the 3D genome near zooHARs. Using
Akita, a neural network-based machine learning model trained on six cell types to predict 3D
genome contact matrices from DNA sequence (36), we assessed the impact of hsSVs (Table S3).
For each variant, we predicted the chromatin contact matrices for the DNA sequence with and
without the variant and computed the mean squared distance between the two matrices. Many
hsSVs are predicted to change 3D genome organization near zooHARs and zooCHARs; 30% of
zooHARs and 27% of zooCHARs occur within 500 kb of a hsSV with a disruption score in the
top decile of all disruption scores for hsSVs. These results suggest that human-specific 3D
genome structures are encoded in DNA sequence and modified through hsSVs.
High-resolution Hi-C data from human and chimpanzee validates 3D genome
reorganization near zooHARs and zooCHARs
In order to validate the predicted changes to 3D genome organization mediated by hsSVs near
zooHARs, we generated Hi-C data from NPCs differentiated from two human and two
chimpanzee induced pluripotent stem cell lines, together generating over 3.4 billion uniquely
mapped chromatin contacts (Table S4)(37). All lines were from male individuals, and two
replicates were generated per sample. Stratum-adjusted correlation coefficients (38)
demonstrated high concordance of data between replicates and individuals from the same species
(Fig. S5), so we merged data from replicates and samples from the same species for downstream
analyses. The cis/trans interaction ratio and distance-dependent interaction frequency decay
indicate that the data is high quality (Table S4, Fig. S6).
Conservation of 3D genome structures, such as A and B compartments and TAD boundaries, has
been demonstrated in various species, however our understanding of the extent of this
conservation is still developing (34, 39–44). We found 10% of TAD boundaries to be species-
specific (Table S5), slightly less than the 14% identified in a recent study comparing human and
macaque chromatin organization (42), likely due to chimpanzees being more closely related to
humans than are macaques. The majority of chromatin loops, also termed ‘dots’ or ‘peaks’ (45),
are conserved or partially conserved (Table S5, Fig. S7) (46, 47). These results support the idea
of conservation of large-scale chromatin structures between human and chimpanzee, though
differences are detectable in specific loci.
We next confirmed the enrichment of zooHARs in TADs containing hsSVs in our Hi-C data
from human NPCs (Fig. S4C, Table S5). This enrichment was also observed between
zooCHARs and chimpanzee-specific structural variants (22) in TADs from the chimpanzee data
(odds ratio=4.8, bootstrap p-value=0.04), indicating that co-location of lineage-specific structural
variants and ARs is not a human-specific phenomenon. As SVs and Hi-C data are generated for
more species, it will be possible to use the tools from this study to quantify this striking
Submitted Manuscript: Confidential
Template revised February 2021
6
association across Eukaryotes. Finally, we used our NPC Hi-C data to associate zooHARs and
zooCHARs with genes and found significant enrichment for transcriptional regulators of
developmental processes, confirming and extending our GO results based on nearby genes.
Hijacked zooHARs and zooCHARs are associated with differentially expressed genes
We next used gene expression data from NPCs (24) and cerebral organoids (25) derived from
human and chimpanzee induced pluripotent stem cells to test if zooHARs with altered chromatin
interactions are associated with altered gene regulation. We observed that differentially
expressed genes in both datasets are enriched in TADs containing zooHARs and hsSVs (chi-
squared p-values < 0.05). In contrast, genes differentially expressed between human and
chimpanzee adult brain tissue (48), induced pluripotent stem cells (iPSC), iPSC-derived
cardiomyocytes, and heart tissue (49) are not enriched in TADs containing zooHARs and hsSVs,
suggesting that the effects of enhancer hijacking may be developmental stage and cell type
specific.
The loci encompassing zooHAR.126 and zooHAR.15 are two clear examples of how hsSVs can
alter 3D regulatory interactions between HAR enhancers and neurodevelopmental genes. Each
locus has a strong Akita prediction of altered genome folding in the presence of an hsSV, which
is highly similar to the differences observed in NPC Hi-C data (Fig. 2A, B) (36). The average
disruption peaks at specific genomic elements within the 1Mb region (Fig. 2C, D), including at
species-specific loops and the promoters of genes differentially expressed between humans and
chimpanzees (Fig. 2E, F). For example, the Tourette’s syndrome gene NECTIN3 (50) is in the
same TAD with an hsSV and zooHAR.126, and it is downregulated in human versus chimpanzee
NPCs (24). Similarly, the developmental gene MAF, implicated in Ayme-Gripp syndrome, is
differentially expressed between human and chimpanzee in inhibitory neurons, NPCs, iPSCs,
iPSC-derived cardiomyocyte progenitors (24, 25, 49), and it is in a TAD encompassing a hsSV
and zooHAR.15, which overlaps previously identified 2xHAR.21 (51). In order to determine
with higher confidence that the observed changes in 3D structure at these loci were human-
derived, we assessed the orthologous loci in previously published rhesus macaque fetal brain
cortex plate (42). For both loci, the human-specific changes to 3D genome organization
described here were not observed in rhesus macaque data, suggesting that they are human-
derived as a result of the hsSVs, as predicted by Akita (Fig. S8) (36). Together, these results
establish that the 3D genome changes in these loci are human-specific, associated with gene
expression changes and likely caused by the hsSVs.
Many zooHARs are neurodevelopmental enhancers with cell type-specific activity
In order to define the cell types and tissues that may be impacted by hijacked HARs, we
expanded on previous work demonstrating enhancer-associated epigenomic signatures of HARs
in specific cell types and tissues and predicting enhancer activity (52) by including recently
generated data from 61 ATAC-seq, 40 DNase-seq and 204 ChIP-seq datasets in 44 cell types
including multiple brain regions from specific developmental timepoints (53–60). Even against a
stringent background set of phastCons elements, which themselves tend to be enriched for gene
regulatory marks related to development (9), zooHARs are enriched for markers indicative of
neurodevelopmental regulatory activity including ATAC-seq peaks and promoter capture Hi-C
Submitted Manuscript: Confidential
Template revised February 2021
7
interactions in multiple neuronal cell types (bootstrap p < 0.05; Fig. S9). For example,
zooHAR.126 overlaps numerous regulatory epigenomic marks and footprints for seven
transcription factors (Fig. 3A). Over all zooHAR footprints, enriched transcription factors
included inhibitory neuron specifier DLX1 (61), master brain regulator and telencephalon marker
FOXG1, and cortical and striatal projection neuron marker MEIS2 (62, 63) (Fig. 3B, Table S6).
Using these datasets as features, we trained a new machine learning model on in vivo validated
VISTA enhancers (64) and used it to predict that 197/312 zooHARs (63.1%) function as
neurodevelopmental enhancers based on their epigenetic profiles. This increases the proportion
of HARs with predicted regulatory activity in the brain relative to previous work (Table S1) (9,
54).
To further specify cell types in the human brain where zooHARs likely function as regulatory
elements, we applied the CellWalker method to map them to cell types using single-cell ATAC-
seq with RNA-seq from the developing human telencephalon surveyed at mid-gestation (59, 65–
67). We found the highest number of zooHARs assigned to newborn interneurons, radial glia,
excitatory neurons from the prefrontal cortex, and medial ganglionic eminence intermediate
progenitors (Fig. 3C, Table S7)(59). Repeating this analysis for zooCHARs, cell types were
largely similar to those assigned to zooHARs, but many fewer zooCHARs mapped to excitatory
neurons from the prefrontal cortex. This difference may provide clues towards the mechanisms
underlying species-specific neurodevelopmental traits, such as increased plasticity and protracted
maturation in the human brain. However, these results must be interpreted with the caveat that
cell-type assignments were made from human data as parallel chimpanzee data are not available
(Fig. S9, Table S7). Finally, we repeated the CellWalker analysis using single-cell ATAC-seq
and RNA-seq from the human adult brain (68, 69) and heart (70). Very few ARs mapped to adult
heart cell types. In the adult brain, fewer zooCHARs were assigned cell types compared to
zooHARs, with the largest species difference being in excitatory neurons, mirroring our finding
in mid-gestation brain (Fig. S10, Table S7).
Massively parallel validation of zooHARs in human primary cortical cells
To validate these predictions, we performed a massively parallel reporter assay (MPRA) to test
the enhancer activity of zooHARs in five replicates of human primary cells from mid-gestation
(gestation week 18) telencephalon. Of the 175 zooHARs predicted to function as
neurodevelopmental enhancers and passing MPRA quality control, 88 (50.3%) drove reporter
gene expression to a level indicative of enhancer activity (Methods; Table S6). This high-
confidence set of human accelerated enhancers active in human neurodevelopment includes
zooHAR.1, zooHAR.133, zooHAR.138, and zooHAR.156, all of which are in TADs with
developmental genes (GBX2, EFNA5, EN1, and PBX3, respectively) that have differential
contacts in our human versus chimpanzee NPC Hi-C data. Prior studies precisely reconstructing
human-specific mutations at the endogenous locus in mouse validated zooHAR.1 (also known as
HACNS1, HAR2, 2xHAR.3) as an enhancer of GBX2 and zooHAR.138 (2xHAR.20, HAR19,
HAR80) as an enhancer of EN1. Other zooHARs with enhancer-like epigenetic signatures but
lower MPRA activity may function in different developmental stages or in cell types poorly
represented in our telencephalon samples, or their activity may be underestimated by MPRA due
to using 270-bp sequences and random integration sites. Despite these limitations, our MPRA
data strongly support the conclusion that many zooHARs function as enhancers in cell types of
Submitted Manuscript: Confidential
Template revised February 2021
8
the developing brain. Altogether, this work demonstrates that hsSVs cluster in TADs with HARs
that likely function as regulatory elements in neurodevelopment, and these hsSVs can change 3D
regulatory interactions of HARs.
Discussion
Lineage-specific ARs represent sequence-based evolutionary innovations in the genome that may
underlie traits that define each species. The Nextflow pipeline introduced in this work enables
reproducible identification of ARs in any species in very large alignments, as demonstrated with
the Zoonomia dataset of 241 mammals (23). Integration of dozens of public and novel datasets
refined our understanding of which HARs may function as regulatory elements, at which
developmental stages, and in what cell types. Viewing ARs through the lens of 3D genome
organization revealed an enrichment of HARs and CHARs in TADs containing species-specific
SVs. Generation of the highest resolution cross-species Hi-C dataset to date in matched NPCs
from human and chimpanzee enabled further discovery that hsSVs predicted by a deep-learning
model to change 3D genome organization nearby HARs and CHARs correspond to true
differences between human and chimpanzee NPCs. HARs are active enhancers in diverse cell
types and the majority contact putative target genes in a cell type-specific manner (71), so future
investigation of more cell types may uncover further perturbations.
It is interesting to ask about the sequence of genomic events in loci with hsSVs and HARs. One
intriguing possibility is that in some cases the hsSV altered the 3D chromatin contacts of a
conserved regulatory element that then underwent rapid adaptation through point mutations in
the same species to adjust to its new target genes. With available data, however, we cannot rule
out the possibility that the accelerated region changed prior to the structural variant. Nor can we
confidently infer that the structural variant and 3D genome changes caused accelerated sequence
evolution of the regulatory element. It is also important to note that the vast majority of TADs
containing hsSVs with high disruption scores do not contain zooHARs, and about a third contain
phastCons elements that are not human-accelerated. Nonetheless, our integrative data analysis
points to enhancer hijacking as a potential genetic mechanism to explain HARs and other
lineage-accelerated conserved non-coding regions. Further experimentation will be needed to
ascertain the validity of this hypothesis. However, it is clear that the evolution of genome
sequence and 3D organization do not occur in isolation.
Submitted Manuscript: Confidential
Template revised February 2021
9
Supplementary Materials for
Three-dimensional genome re-wiring in loci with Human Accelerated Regions
Kathleen C. Keough1,2,3, Sean Whalen1, Fumitaka Inoue2,3†, Pawel F. Przytycki1^, Tyler Fair4,5,
Chengyu Deng2,3, Marilyn Steyert4,5,6,7, Hane Ryu2,3, Kerstin Lindblad-Toh8,9, Elinor
Karlsson9,10,11, Zoonomia Consortium, Tomasz Nowakowski4,5,6,7, Nadav Ahituv2,3, Alex
Pollen7,12, Katherine S. Pollard1,3,13,14*
Correspondence to: katherine.pollard@gladstone.ucsf.edu
This PDF file includes:
Materials and Methods
Supplementary Text
Figs. S1 to S12
Captions for Tables S1 to S7
References (31 to 74)
Other Supplementary Materials for this manuscript include the following:
Tables S1 to S7
Submitted Manuscript: Confidential
Template revised February 2021
10
Materials and Methods
Automated identification of human and chimpanzee accelerated regions
To facilitate detection of accelerated regions in any lineage on any computing
infrastructure, we developed a pipeline implemented in Nextflow (26) (Fig. 1A). To date,
identification of lineage-specific accelerated regions has used custom scripts that call the
PHAST/RPHAST packages (27, 29, 72) or similar software to identify conserved elements with
increased rates of nucleotide substitutions in a given part of a phylogeny using a multiple
sequence alignment of the species in the tree. Highly conserved elements are likely to be
functional, and they have higher power for detecting accelerated substitution rates on short (e.g.,
human, chimpanzee) branches as compared to less conserved elements. Our pipeline
AcceleratedRegionsNF, available at github.com/keoughkath/AcceleratedRegionsNF, automates
these analyses, including tuning run time parameters for large alignments and parallelizing
compute over genomic regions (see Supplementary Text). Users provide a multiple sequence
alignment in MAF format, a Newick-formatted, bifurcating species tree, and a neutral model.
Users may analyze a subset of species in the multiple alignment by also submitting a species list.
They simply change the configuration file to describe their computing environment, and the
analysis pipeline will run beginning to end. The pipeline generates a BED-formatted file of
accelerated regions at a user-defined false discovery rate (FDR) and a table of phastCons
elements with phyloP scores and p-values (raw and Benjamini-Hochberg adjusted), enabling the
user to adjust the acceleration FDR if desired after running the pipeline. Run time of the pipeline
changes based on the size of the computing environment and the size of the input MAF files.
Splitting the MAF files into smaller segments (e.g., 10 megabases each) speeds up the runtime
significantly.
The human (zooHAR) and chimpanzee (zooCHAR) accelerated regions described in this
work were identified using the Zoonomia 241-mammal human-referenced MAF-formatted
multiple alignments, a neutral model based on ancestral repeats and the Zoonomia chromosome
X species tree (23). Because the multiple alignment was human-referenced, zooHARs and
zooCHARs were both initially identified in the human reference genome (hg38). Using the
Nextflow pipeline described above, conserved elements in all species in the multiple alignment
were identified using phastCons (72) with the human or chimpanzee sequence masked, these
elements were filtered for level one or two synteny with rhesus macaque, dog, and mouse (73).
Duplications, pseudogenes from Gencode v29, self-chain and repetitive regions were filtered out
(73). Elements with a phastCons log odds score in the bottom three deciles were removed, as
well as any elements less than 50 base pairs (bp) long. We note that multiple ~100-bp phastCons
elements often occur near each other, because a functionally constrained element (e.g., exon,
enhancer) may be composed of highly conserved regions broken up by several less conserved
alignment columns that cause phastCons to annotate separate conserved regions.
Accelerated elements in human or chimpanzee were identified using phyloP (28). Elements
with a Benjamini-Hochberg false discovery rate less than 0.05 were retained as accelerated
regions. When several phastCons elements are adjacent pieces of a larger enhancer-like element,
they were separately tested using phyloP and hence may not all be accelerated.
Characterization of zooHARs and zooCHARs
zooHAR distribution relative to gene annotations was performed using GENCODE v37
annotation in reference human genome assembly hg38 (74). Selection and clustering analyses
were conducted as previously described (10, 30). Enriched ontology terms for genes proximal to
zooHARs were identified using GREAT (31). Functional modules associated with zooHAR-
linked genes were detected using HumanBase tissue-specific networks (32). Further gene
Submitted Manuscript: Confidential
Template revised February 2021
11
ontology analysis of genes that co-occur in chromatin loops (10kb resolution) with zooHARs
was conducted using DAVID (75). Epigenomic annotations were performed on the midpoint of
each zooHAR extended upstream and downstream by 750bp, and the decision threshold for
enhancer predictions adjusted to 0.3, in order to more closely match the properties of validated
VISTA enhancers (64). zooHAR brain cell types were identified by CellWalker (65) as
implemented in the CellWalkR package (version 0.99, default parameters, with Jaccard
similarity used for cell edges, gene accessibility used for label edges, and the label edge weight
parameter set to one)(66) applied to data from the developing human telencephalon (59, 65).
zooHAR expression was assessed by overlap with transcripts from (76) lifted over to hg38 (77).
Enrichment of zooHARs in chromatin contact domains with human-specific structural variants
(hsSVs) was performed by calculating the odds ratio of a chromatin contact domain containing a
zooHAR and an hsSV. A p-value was generated by comparing that odds ratio to a null
distribution of 1000 odds ratios calculated the same way, except with a random draw of N
phastCons elements, where N is the number of zooHARs. Various computational analyses
utilized GNU parallel (78). To characterize chimpanzee accelerated regions, the above analyses
were repeated with zooCHARs in place of zooHARs.
Prediction in silico of human-specific structural variant impacts
Prediction of hsSV effects was performed using Akita, a deep learning model that predicts
chromatin contact matrices from DNA sequence (36). To predict the impact of hsSVs on the 3D
genome, we submitted two 1Mb sequences to Akita, one with and one without the hsSV. We
used the human (hg38) sequence if the hsSV was an insertion and chimpanzee (pantro6)
sequence if the hsSV was a deletion or inversion. We then calculated the mean squared error
(“disruption score”) between these two contact matrices.
NPC generation, differentiation, validation
Two human (WTC11 and HS1) and two chimpanzee (C3649 and Pt2a) induced pluripotent
cell lines (iPSCs) were cultured in Matrigel-coated plates with mTeSR media (WTC11 and
C3649 were cultured in StemFlex) in an undifferentiated state. Cells were propagated at a 1:3
ratio by treatment with 200 U/mL collagenase IV (or PBS-EDTA) and mechanical dissection.
WTC11 and C3649 iPSCs were differentiated to neural progenitor cells (NPCs) and
validated as previously described (37). Briefly, 2-2.5×10⁵ cells per cm² were seeded on Matrigel-
coated wells in StemFlex containing 2 μM Thiazovivin. The following day (Day 0), medium was
replaced with E6 containing 500 nM LDN193189 (Selleckchem), 10 μM SB431542
(Selleckchem), and 5 μM XAV-939 (Selleckchem). Starting on Day 3, medium was replaced
with E6 containing 500 nM LDN193189 and 10 μM SB431542 every 48 hrs. Starting on Day 12,
medium was replaced with Neurobasal containing 2 mM GlutaMAX, 60 μg per ml L-Ascorbic
acid 2-phosphate, N2, and B27 without Vitamin A every 48 hours. Around Day 16, cells were
washed with PBS, dissociated with Accutase, pelleted and resuspended in Neurobasal containing
2 mM GlutaMAX, 60 μg per ml L-Ascorbic acid 2-phosphate, N2, and B27 without Vitamin A,
10 ng per ml fibroblast growth factor 2, and 10 ng per ml epidermal growth factor, and seeded on
poly-L-ornithine-, fibronectin-, and laminin-coated wells. Cells were collected for HiC at
passage 5-7.
To differentiate HS1 and Pt2a iPSCs into NPCs, cells were split with EDTA at 1:5 ratios in
culture dishes coated with matrigel and culture in N2B27 medium (comprised of DMEM/F12
medium (Invitrogen) supplemented with 1% MEM-nonessential amino acids (Invitrogen), 1 mM
L-glutamine, 1% penicillin-streptomycin, 50 ng/mL bFGF (FGF-2) (Millipore), 1x N2
supplement, and 1 x B27 supplement without Vitamin A (Invitrogen)) supplemented with 100
ng/ml mouse recombinant Noggin (R&D systems). Cells at passages 1-3 were split by
Submitted Manuscript: Confidential
Template revised February 2021
12
collagenase into small clumps, and continuously cultured in N2B27 medium with Noggin. After
passage 3, cells were plated at the density of 5×10⁵ cells/cm² after disassociation by TrypLE
express (Invitrogen) into single-cell suspension, and cultured in N2B27 medium supplemented
with 20 ng/mL bFGF and EGF. Cells were maintained and collected at passage 18-20. Our use of
two differentiation protocols reflects rapid progress in stem cell research during the course of this
study. Cells from the same populations were validated and used in a previous study (37). We
verified that the chromatin interactions in the resulting Hi-C data did not show a batch effect
across protocols.
Hi-C data generation
Hi-C was performed using the Arima Hi-C kit (Arima Genomics) according to the
manufacturer’s instructions. 10 million cells were used. The sequencing library was prepared
using Accel-NGS 2S Plus DNA Library Kit (Swift Biosciences) according to the manufacturer's
protocol. Two independent biological replicates were prepared for each cell line. In total eight
libraries were pooled and sequenced with paired-end 150-bp reads using two lanes of a
NovaSeq6000 S2 (Illumina) at the Chan Zuckerberg Biohub.
Hi-C data processing
Adapters were trimmed from raw FASTQ files using TrimGalore [v0.6.5] with options --
illumina --paired. The data were then processed from adapter-trimmed FASTQ files to Hi-C
contacts as cooler files using Distiller [v0.3.3] (79). This processing includes read mapping with
BWA-MEM (80), filtering (MAPQ >= 30), contact pair processing with pairtools (81) and
normalization via matrix balancing (82). Samples were processed both per replicate, per
individual and per species. For easier comparison of samples in some analyses, we mapped the
data from each species to the reference genome of the other species (human to pantro6 and
chimp to hg38). Cis/trans ratio was calculated as the ratio of cis to trans contacts for each
replicate (83). Distance-dependent interaction frequency decay was computed using cooltools
with 100-kilobase (kb) bins (83, 84).
A and B compartments were identified by eigenvector decomposition of the contact
matrices, phased by GC content with A compartment having higher GC content than B
compartment using cooltools (79). We assessed conservation between TAD boundaries based on
the method from (42). We identified boundaries by calculating the insulation score at a
resolution of 50kb and using a 800-kb sliding window, considering bins with boundary strength
greater than 0.1 and insulation score less than zero as boundaries. Boundaries were considered
conserved if they were within two bins (100 kb) of a boundary in the other species, and species-
specific if they were more than five bins (500 kb) from the nearest boundary in the other species
after liftOver (42). TADs for the zooHAR enrichment analyses were identified using a 400-kb
window and 10-kb bin size, with boundary strength greater than 0.1 and insulation score less
than zero as boundaries. Loops were identified using Mustache at 5-kb resolution (46).
Conservation of loop anchors was conducted using mapLoopLoci (47).
Massively parallel reporter assay
We designed 270-bp oligos centered on zooHARs and positive control enhancer
sequences. For zooHARs longer than 270 bp, we tiled oligos across the element. A 31-bp
minimal promoter (minP) and 15-bp random barcodes were placed downstream of the
synthesized oligos via PCR and cloned into an MPRA vector as previously described (85). The
library was packaged into lentivirus and used to infect human primary cortical cells dissociated
from two fresh tissue samples (gestational week 18). Cells were cultured for two days prior to
infection and 3 days following infection in a DMEM-based media containing B27, N2, and Pen-
Strep. Cells were harvested, then DNA and RNA were obtained for sequencing. For each oligo,
Submitted Manuscript: Confidential
Template revised February 2021
13
we quantified enhancer activity using the ratio of barcode abundance in RNA versus DNA
normalized and batch corrected across replicates. A zooHAR was determined to be active if its
maximally active tile had an RNA/DNA value exceeding the median of a set of positive control
enhancer sequences that we included in the MPRA library.
Supplementary Text
Impact of number and choice of species in the alignment
Previous analyses to identify human accelerated regions (HARs) have generally used
alignments of fewer than 30 species (1, 2, 51). The Zoonomia multiple alignment analyzed in
this work, as well as other commonly used multiple alignments of vertebrates, such as the 100-
way UCSC alignment, are much larger. Additionally, genome quality and completeness for
many species have improved greatly since early HAR analyses. Therefore, we systematically
assessed each step of the HAR identification analysis laid out in earlier work to determine
whether changes needed to be made.
In order to assess the variability of HARs and phastCons elements per species number and
set, we identified HARs and phastCons elements from the 100-way hg38 UCSC multiple
alignment of vertebrates using sets of ten to ninety randomly selected species with three
replicates of random species selection per species number. Each species set included human and
chimpanzee, but was otherwise randomly selected from the full set of species in the UCSC 100-
way alignment. These HARs were identified using a neutral model based on 4-fold degenerate
sites, phastCons parameters rho=0.3, omega=45, gamma=0.3 with a Benjamini-Hochberg FDR <
0.01 for phyloP acceleration. We found that with increasing numbers of species, the number of
elements identified, genome coverage, and size of elements all decreased (Fig. S11). These
trends were consistent across other FDR thresholds. We next compared the HARs analyzed in
the main text using the Zoonomia 241-mammal alignment (zooHARs) to HARs identified from
the subset of all mammals (UCSC mammal) and from the full set of species (UCSC vertebrate)
in the hg38 100-way MULTIZ alignment from UCSC, in each case based on a neutral model
derived from ancestral repeats. Most UCSC vertebrate HARs were a subset of the UCSC
mammal HARs or zooHARs (Fig. S12), while UCSC mammal HARs and zooHARs shared
about half of their elements and base pairs. These results indicate that the alignments used to
identify phastCons elements have a big impact on the resulting set of ARs, and including non-
mammal vertebrates decreases the number of ARs discovered.
Using a subset of high-quality species for HAR identification
We explored the strategy of using a subset of “high-quality” species genomes for HAR
identification, with the rationale that this may help avoid false positives caused by spurious
alignments or miscalled regions in genomes. A barrier to this approach was that genomes from
different species were assembled using different sequencing technologies and methodologies,
making it difficult to establish a set of objective standards for inclusion. Additionally, many of
the “higher quality” genomes are in the primate clade, thus skewing the phylogenetic
representation of the species set. Due to these constraints, we were not able to curate an optimal
species set based on maximizing stability of the HARs identified. Therefore, we decided to
proceed with the full set of species to identify zooHARs. However, these results emphasize the
importance of careful species set selection in AR analyses depending on the research goals. To
this end, in the AR-identification pipeline described in this paper, we enable the researcher to
submit a list of species in order to analyze a subset of the species present in the multiple
sequence alignment.
Tuning phastCons parameters in assemblies with hundreds of species
Submitted Manuscript: Confidential
Template revised February 2021
14
The methods to identify HARs were developed using alignments with less than 30 species
and older versions of genome assemblies. As multiple species alignments have grown and
assemblies have improved and become more complete, we systematically assessed the parameter
choice for identifying the set of conserved elements from which HARs would be drawn. The
tuning parameters in phastCons that we assessed included ⍴, a scaling factor describing the
extent to which a neutral tree should be shrunk to approximate the conserved state, ⍵, the
estimated length of conserved elements, and ɣ, the estimated genome coverage by conserved
elements. Of these parameters, the most obvious candidate to be adjusted was ɣ, as this
parameter is inversely proportional to the proportion of the reference genome in the multiple
alignment blocks. In previous alignments, only 16.5% of the human genome was represented in
alignment blocks, whereas in the Zoonomia 241-mammal alignment that has increased to 97.7%.
Therefore, based on (72) and an expected genome coverage of 5% by conserved elements, ɣ is
approximately 0.05. As another method of checking these parameters, we estimated ⍴, ⍵ and ɣ
by maximum likelihood using the phastCons program. The parameters were estimated based on
100 1-Mb windows of the UCSC 100-way alignment, using a neutral model estimated from
ancestral repeats. The median values identified were ɣ=0.06, ⍴=0.27 and ⍵=4.05. Thus, we
decided to proceed with parameters ɣ=0.05 and ⍴=0.3 based on these estimates, but we used
⍵=45 as done in previous work with the goal of increasing the size of the conserved elements
identified, which increases power in downstream phyloP tests for acceleration (27) and
eliminates the need to develop ad hoc methods to merge adjacent phastCons elements.
Additionally, we implemented a threshold for the phastCons log odds score, requiring that
phastCons elements considered for acceleration were above the third decile of length-normalized
log odds scores, thus removing elements with the weakest signatures of conservation from
consideration.
Automated identification of human- and chimpanzee-specific accelerated regions
Genome-wide analyses of large multiple-species alignments typically require cluster
computing, which hinders reproducibility and accessibility. To enable automated detection of
accelerated regions in any lineage on any computing infrastructure, we implemented our analysis
pipeline in Nextflow (26) Given a species tree, neutral model, and multiple sequence alignment,
this open-source software uses PHAST to identify lineage-specific accelerated regions for any
species of interest (Fig. S1A). This pipeline enables simplified, portable and reproducible
identification of lineage-specific accelerated regions.
zooHAR and zooCHAR characterization
Accelerated regions cluster and are mostly noncoding
Using the Zoonomia 241-mammal alignments, we identified 312 zooHARs and 141
chimpanzee accelerated regions (zooCHARs) (Benjamini-Hochberg FDR < 0.05, Tables S1, 2).
Median length was 117.5 base pairs (bp) for zooHARs (IQR: 110.5 bp) and 108.0 bp for
zooCHARs (IQR: 90 bp), similar to prior studies. 32.4% of zooHARs overlap previous lists of
HARs identified by similar methods (1, 6, 51), and 5.5% of a merged group of previous sets of
HARs identified by varying methods (1, 3–5, 9, 51) overlap zooHARs, agreeing with prior
analyses which found that differing methodologies and underlying datasets render most HAR
sets unique from one another, and thus we do not claim this set to be superior to others (9).
zooHARs and zooCHARs were identified on all autosomes and chromosome X. Each set is
clustered along the linear genome so that specific loci harbor more zooHARs (p = 0.01) or more
zooCHARs (p = 0.01) than expected given the density of conserved (phastCons) regions
(225,317 phastCons elements from which zooHARs and 225,287 from which zooCHARs were
identified). zooHARs and zooCHARs show a similar genomic distribution to previous HAR sets
Submitted Manuscript: Confidential
Template revised February 2021
15
with respect to genomic features. The majority are intergenic, although some overlap protein-
coding features and noncoding RNAs (Fig. S1B, C).
Genes near zooHARs are involved in transcriptional regulation, forebrain development and
morphogenesis, and multiple other developmental terms based on GREAT analysis (Fig. S2A)
(31). GREAT analysis also revealed enrichment of zooHARs nearby genes involved in mouse
neonatal lethality with complete penetrance, and multiple abnormal developmental events in
mouse (Fig. S2B). GREAT analysis of zooCHARs revealed an enrichment of nearby genes for
transcriptional regulation and sequence specific DNA binding (Fig. S3) and neonatal lethality
(31). However, GREAT analyses are based on genes nearest HARs and CHARs, which may not
be the target genes of these elements. Therefore we also identified ontology terms enriched in
genes that are associated with HARs and CHARs via 3D chromatin loops from the Hi-C data in
NPCs generated in this study. Enriched gene ontology (GO) terms included multiple
developmental terms, including “heart development”, “positive regulation of developmental
process”. Thus, regardless of the method for associating zooHARs and zooCHARs with target
genes, we see a clear enrichment for loci with transcription factors in both species.
Developmental processes are also enriched, particularly for zooHARs. The stronger signal for
diverse developmental processes in zooHAR loci as compared to zooCHAR loci may be due to
higher power with the larger set of zooHARs, but it could also reflect biological differences in
the function of these elements in the two species, consistent with adaptation of each species to
their distinct environmental niches.
Most zooHARs and zooCHARs are under positive selection
Accelerated evolution is not synonymous with positive selection. Positive selection
indicates a rate of nucleotide substitutions that is faster than the (local or genome-wide) neutral
rate, indicating that the sequence changes are beneficial. Acceleration means a rate of nucleotide
substitutions that is faster than expected given the rate in the rest of the tree, which could be
faster, slower or equal to the neutral rate. The rest of the tree is evolving slower than the neutral
rate for the accelerated regions in this study, so the lineage of interest (human or chimpanzee)
could be less slow but still below the neutral rate, equal to the neutral rate or faster than the
neutral rate. GC-biased gene conversion (GBGC) can mimic positive selection (86), but the
substitutions are biased towards A/T to G/C changes. To infer the evolutionary forces that
shaped the accelerated regions in this study, we applied a method that uses likelihood ratio tests
to assess loci for evidence of positive selection, GBGC, or both (10). This method controls for
local variation in the neutral rate of evolution by comparing each element to the surrounding 1
Mb of genome rather than the rate of evolution in the other species without the element itself
(based on rescaling a phylogeny built using the genome-wide neutral rate). This analysis
estimated that 82% of zooHARs and 86% of zooCHARs are under positive selection, though 7%
of zooHARs and zooCHARs show strong evidence for GBGC, and 5% of zooHARs may have
been shaped by a combination of selection and GBGC (Fig. 1D, E; Tables S1, S2).
zooHARs and zooCHARs are transcribed in the developing human brain
Some HARs have been shown to function as noncoding RNAs, including the original
HAR1 (2), therefore we investigated the noncoding RNA potential of zooHARs. Additionally,
many active enhancers are transcribed (eRNAs). We assessed the expression of zooHARs and
zooCHARs in RNAseq data from the developing human neocortex (76), including both poly-A
and total RNA, enabling the study of non-protein-coding RNA transcripts (76) and eRNAs. We
found that 100 of 312 zooHARs (32%) and 41 of 141 zooCHARs (29%) were expressed in the
total RNA dataset (TPM>5, Fig. S1E, F). Twenty of the expressed zooHARs overlapped gene
exons, including ERC2, involved in neurotransmitter release (87), and TNIK, implicated in
Submitted Manuscript: Confidential
Template revised February 2021
16
neurological disorders, neurogenesis and cell proliferation (88). Of the expressed zooHARs 88
overlapped gene introns, 12 overlapped annotated noncoding RNAs, and 13 do not overlap any
currently annotated elements, and therefore could represent uncharacterized noncoding RNAs or
eRNAs.
Submitted Manuscript: Confidential
Template revised February 2021
17
Fig. S1. zooHARs demonstrate similar characteristics to prior HAR sets.
(A, B) Genic distribution of zooHARs (A) and zooCHARs (B) (both in reference hg38), based
on Gencode V37 annotations. (C, D) Selective forces acting on zooHARs (C) and zooCHARs.
(D) from pipeline described in (10). Positive=positive selection, GBGC=GC-biased gene
conversion, hc=high-confidence. (E, F) Transcription of zooHARs from the positive (E) and
negative (F) strand in the developing human neocortex at five mid-gestation time points (76).
Whiskers extend to 1.5 times the inter-quartile range (IQR). TPM=transcripts per million.
Submitted Manuscript: Confidential
Template revised February 2021
18
Fig. S2. GREAT analysis of genes near zooHARs.
(A) GREAT (31) gene ontology enrichment analysis of zooHARs. (B) GREAT mouse
phenotype (single knockout) enrichment analysis of zooHARs.
Submitted Manuscript: Confidential
Template revised February 2021
19
Fig. S3. GREAT analysis of genes near zooCHARs.
GREAT (31) gene ontology analysis of zooCHARs.
Submitted Manuscript: Confidential
Template revised February 2021
20
Fig. S4. Enrichment of zooHARs in TADs with hsSVs compared to random windows.
(A) Odds ratio of TADs from human iPSCs called with the Arrowhead algorithm (41) containing
one of the 17,789 human-specific structural variants (hsSVs) and one of the 312 zooHARs
(green line) compared to a null distribution based on 1000 random draws of 312 phastCons
elements (blue shaded area). (B-F) Same analysis as in A, but with (B) TADs from mid-gestation
developing human cerebral cortex (cortical plate and germinal zone) based on insulation scores
(89); (C) TADs from human NPCs (25); (D) random 185-kb windows, the median size of contact
domains from (A); (E) random 185-kb windows, the median size of contact domains from (B);
(F) 280-kb random windows, the median size of the human NPC TADs from (C).
Submitted Manuscript: Confidential
Template revised February 2021
21
Fig. S5. Hi-C correlation values per Hi-C sample.
Stratum adjusted correlation coefficients (SCC) between all samples mapped to hg38. The SCC
statistic is calculated by stratifying the data by genomic distance, then computing a Pearson
correlation coefficient for each stratum and then aggregating the stratum-specific correlation
coefficients using a weighted average, with the weights derived from the generalized Cochran–
Mantel–Haenszel statistic (38).
Submitted Manuscript: Confidential
Template revised February 2021
22
Fig. S6. Distance-dependent contact decay per Hi-C sample.
Corrected (IC) contact frequency as a function of distance between all pairs of 100-kb bins for
each replicate.
Submitted Manuscript: Confidential
Template revised February 2021
23
Fig. S7. Loop conservation human to chimpanzee.
Loop conservation assessed with mapLoopLoci from (47) for human compared to chimpanzee
Hi-C data with (A, B) both mapped to hg38 and (C, D) each mapped to their respective species’
reference genome.
Submitted Manuscript: Confidential
Template revised February 2021
24
Fig. S8. Loci of interest in rhesus macaque.
The loci surrounding zooHAR.126 (A) and zooHAR.15 (B) in human and chimpanzee NPC Hi-
C generated in this work, compared to Hi-C from rhesus macaque cortex plate (42).
Log(observed/expected) values are shown in the heatmaps.
Submitted Manuscript: Confidential
Template revised February 2021
25
Fig. S9. Overlap of zooHARs with epigenomic marks from brain.
A majority of zooHARs overlap robust peaks (Irreproducible Discovery Rate 10%) from open
chromatin (ATAC-seq or DNase-seq) and activating histone modifications (ChIP-seq) from
neural cell lines or primary brain tissue (53–60). Bar plot on the y-axis (left) indicates the
number of zooHARs overlapping each epigenomic feature, bar plot on the x-axis (top) indicates
the number of zooHARs overlapping multiple epigenomic features, indicated by the shaded dots
in the center. The highest proportion of zooHARs overlap both activating ChIP-seq and ATAC-
seq peaks, followed by those that overlap only activating ChIP-seq peaks. Accounting for the
smaller number of DNase-seq datasets, many zooHARs that overlap activating ChIP-seq and
ATAC-seq also overlap DNase-seq peaks.
Submitted Manuscript: Confidential
Template revised February 2021
26
Fig. S10. CellWalker analysis of zooHARs and zooCHARs mapped to adult brain and
heart cell types.
As controls to compare to our analysis of mid-gestation telencephalon cell types, we ran
CellWalker using matched single-cell ATAC-seq and RNA-seq from adult brain (68, 69) and
adult heart (70) to associate each zooHAR and each zooCHAR with cell types in which they
appear to be active. The only heart cell type with any ARs is ventricular cardiomyocytes, which
was predicted as an active cell type for only a few zooHARs and zooCHARs. In both adult
tissues, cell types tend to have similar numbers of zooHARs and zooCHAR associations, with
the exception of excitatory neurons, which have many more zooHAR associations. This
enrichment mirrors what we observed in mid-gestation brain excitatory neurons (Fig. 3).
Submitted Manuscript: Confidential
Template revised February 2021
27
Fig. S11. Impact of species number on UCSC HARs and phastCons elements identified.
Number of elements, genome coverage, and element length as a function of the number of
species included in the analysis for UCSC HARs (A, B, C) and phastCons (D, E, F) based on the
UCSC 100-way alignment of vertebrates. Error bars indicate standard deviation based on three
sets of random draws of species (see Supplemental Text).
Submitted Manuscript: Confidential
Template revised February 2021
28
Fig. S12. Comparison of Zoonomia with UCSC HARs.
Overlap in base pairs (A) and elements (B) of HARs identified from the entire 100-way UCSC
alignment (UCSC vertebrate HARs), the 61-mammal subset of the 100-way UCSC alignment
(UCSC mammals HARs), and zooHARs.
Submitted Manuscript: Confidential
Template revised February 2021
29
Table S1. (separate file)
zooHAR coordinates (hg38), selection annotations and enhancer prediction scores.
Table S2. (separate file)
zooCHAR coordinates (hg38) and selection annotations.
Table S3. (separate file)
Predicted disruption scores for human-specific structural variants.
Table S4. (separate file)
Quality control information for the Hi-C data for human and chimpanzee NPCs generated in this
work, including sequenced read depths, uniquely mapped pairs and cis/trans ratios for each
sample.
Table S5. (separate file)
Loops and TADs for human and chimpanzee NPC Hi-C generated in this study.
Table S6. (separate file)
zooHAR overlaps with epigenomic annotations and enrichments compared to phastCons
elements, and zooHAR activity in an MPRA in primary human mid-gestation telencephalon
cells.
Table S7. (separate file)
zooHAR and zooCHAR CellWalker assignments based on data from the developing human
telencephalon.
References and Notes
1. K. S. Pollard, S. R. Salama, B. King, A. D. Kern, T. Dreszer, S. Katzman, A. Siepel, J. S.
Pedersen, G. Bejerano, R. Baertsch, K. R. Rosenbloom, J. Kent, D. Haussler, Forces
shaping the fastest evolving regions in the human genome. PLoS Genet. 2, e168 (2006).
2. K. S. Pollard, S. R. Salama, N. Lambert, M.-A. Lambot, S. Coppens, J. S. Pedersen, S.
Katzman, B. King, C. Onodera, A. Siepel, A. D. Kern, C. Dehay, H. Igel, M. Ares Jr, P.
Vanderhaeghen, D. Haussler, An RNA gene expressed during cortical development evolved
rapidly in humans. Nature. 443, 167–172 (2006).
3. S. Prabhakar, J. P. Noonan, S. Pääbo, E. M. Rubin, Accelerated evolution of conserved
noncoding sequences in humans. Science. 314, 786 (2006).
4. C. P. Bird, B. E. Stranger, M. Liu, D. J. Thomas, C. E. Ingle, C. Beazley, W. Miller, M. E.
Hurles, E. T. Dermitzakis, Fast-evolving noncoding sequences in the human genome.
Submitted Manuscript: Confidential
Template revised February 2021
30
Genome Biol. 8, R118 (2007).
5. E. C. Bush, B. T. Lahn, A genome-wide screen for noncoding elements important in
primate evolution. BMC Evol. Biol. 8, 17 (2008).
6. M. J. Hubisz, K. S. Pollard, Exploring the genesis and functions of Human Accelerated
Regions sheds light on their role in human evolution. Curr. Opin. Genet. Dev. 29, 15–21
(2014).
7. L. F. Franchini, K. S. Pollard, Human evolution: the non-coding revolution. BMC Biol. 15,
89 (2017).
8. S. K. Reilly, J. P. Noonan, Evolution of Gene Regulation in Humans. Annu. Rev. Genomics
Hum. Genet. 17, 45–67 (2016).
9. J. A. Capra, G. D. Erwin, G. McKinsey, J. L. R. Rubenstein, K. S. Pollard, Many human
accelerated regions are developmental enhancers. Philos. Trans. R. Soc. Lond. B Biol. Sci.
368, 20130025 (2013).
10. D. Kostka, M. J. Hubisz, A. Siepel, K. S. Pollard, The role of GC-biased gene conversion in
shaping the fastest evolving regions of the human genome. Mol. Biol. Evol. 29, 1047–1057
(2012).
11. Chimpanzee Sequencing and Analysis Consortium, Initial sequence of the chimpanzee
genome and comparison with the human genome. Nature. 437, 69–87 (2005).
12. D. Hnisz, A. S. Weintraub, D. S. Day, A.-L. Valton, R. O. Bak, C. H. Li, J. Goldmann, B.
R. Lajoie, Z. P. Fan, A. A. Sigova, J. Reddy, D. Borges-Rivera, T. I. Lee, R. Jaenisch, M.
H. Porteus, J. Dekker, R. A. Young, Activation of proto-oncogenes by disruption of
chromosome neighborhoods. Science. 351, 1454–1458 (2016).
13. M. Affer, M. Chesi, W.-D. G. Chen, J. J. Keats, Y. N. Demchenko, A. V. Roschke, S. Van
Wier, R. Fonseca, P. L. Bergsagel, W. M. Kuehl, Promiscuous MYC locus rearrangements
hijack enhancers but mostly super-enhancers to dysregulate MYC expression in multiple
myeloma. Leukemia. 28, 1725–1735 (2014).
14. M. W. Zimmerman, Y. Liu, S. He, A. D. Durbin, B. J. Abraham, J. Easton, Y. Shao, B. Xu,
S. Zhu, X. Zhang, Z. Li, N. Weichert-Leahey, R. A. Young, J. Zhang, A. Thomas Look,
MYC Drives a Subset of High-Risk Pediatric Neuroblastomas and Is Activated through
Mechanisms Including Enhancer Hijacking and Focal Enhancer Amplification. Cancer
Discovery. 8 (2018), pp. 320–335.
15. D. G. Lupiáñez, M. Spielmann, S. Mundlos, Breaking TADs: How Alterations of
Chromatin Domains Result in Disease. Trends Genet. 32, 225–237 (2016).
16. J. Ibn-Salem, S. Köhler, M. I. Love, H.-R. Chung, N. Huang, M. E. Hurles, M. Haendel, N.
L. Washington, D. Smedley, C. J. Mungall, S. E. Lewis, C.-E. Ott, S. Bauer, P. N.
Schofield, S. Mundlos, M. Spielmann, P. N. Robinson, Deletions of chromosomal
regulatory boundaries are associated with congenital disease. Genome Biol. 15, 423 (2014).
Submitted Manuscript: Confidential
Template revised February 2021
31
17. M. Franke, D. M. Ibrahim, G. Andrey, W. Schwarzer, V. Heinrich, R. Schöpflin, K. Kraft,
R. Kempfer, I. Jerković, W.-L. Chan, M. Spielmann, B. Timmermann, L. Wittler, I. Kurth,
P. Cambiaso, O. Zuffardi, G. Houge, L. Lambie, F. Brancati, A. Pombo, M. Vingron, F.
Spitz, S. Mundlos, Formation of new chromatin domains determines pathogenicity of
genomic duplications. Nature. 538, 265–269 (2016).
18. R. D. Acemel, I. Maeso, J. L. Gómez-Skarmeta, Topologically associated domains: a
successful scaffold for the evolution of gene regulation in animals. Wiley Interdisciplinary
Reviews: Developmental Biology. 6 (2017), p. e265.
19. I. Maeso, R. D. Acemel, J. L. Gómez-Skarmeta, Cis-regulatory landscapes in development
and evolution. Curr. Opin. Genet. Dev. 43, 17–22 (2017).
20. N. Lonfat, D. Duboule, Structure, function and evolution of topologically associating
domains (TADs) at HOX loci. FEBS Lett. 589, 2869–2876 (2015).
21. R. D. Acemel, J. J. Tena, I. Irastorza-Azcarate, F. Marlétaz, C. Gómez-Marín, E. de la
Calle-Mustienes, S. Bertrand, S. G. Diaz, D. Aldea, J.-M. Aury, S. Mangenot, P. W. H.
Holland, D. P. Devos, I. Maeso, H. Escrivá, J. L. Gómez-Skarmeta, A single three-
dimensional chromatin compartment in amphioxus indicates a stepwise evolution of
vertebrate Hox bimodal regulation. Nat. Genet. 48, 336–341 (2016).
22. Z. N. Kronenberg, I. T. Fiddes, D. Gordon, S. Murali, S. Cantsilieris, O. S. Meyerson, J. G.
Underwood, B. J. Nelson, M. J. P. Chaisson, M. L. Dougherty, K. M. Munson, A. R. Hastie,
M. Diekhans, F. Hormozdiari, N. Lorusso, K. Hoekzema, R. Qiu, K. Clark, A. Raja, A. E.
Welch, M. Sorensen, C. Baker, R. S. Fulton, J. Armstrong, T. A. Graves-Lindsay, A. M.
Denli, E. R. Hoppe, P. Hsieh, C. M. Hill, A. W. C. Pang, J. Lee, E. T. Lam, S. K. Dutcher,
F. H. Gage, W. C. Warren, J. Shendure, D. Haussler, V. A. Schneider, H. Cao, M. Ventura,
R. K. Wilson, B. Paten, A. Pollen, E. E. Eichler, High-resolution comparative analysis of
great ape genomes. Science. 360 (2018), doi:10.1126/science.aar6343.
23. Genereux, Zoonomia consortium., Evolutionary innovation of eutherian mammals. In
preparation/submission (Flagship 1).
24. M. C. Marchetto, B. Hrvoj-Mihic, B. E. Kerman, D. X. Yu, K. C. Vadodaria, S. B. Linker,
I. Narvaiza, R. Santos, A. M. Denli, A. P. Mendes, R. Oefner, J. Cook, L. McHenry, J. M.
Grasmick, K. Heard, C. Fredlender, L. Randolph-Moore, R. Kshirsagar, R. Xenitopoulos,
G. Chou, N. Hah, A. R. Muotri, K. Padmanabhan, K. Semendeferi, F. H. Gage, Species-
specific maturation profiles of human, chimpanzee and bonobo neural cells. Elife. 8 (2019),
doi:10.7554/eLife.37527.
25. A. A. Pollen, A. Bhaduri, M. G. Andrews, T. J. Nowakowski, O. S. Meyerson, M. A.
Mostajo-Radji, E. Di Lullo, B. Alvarado, M. Bedolli, M. L. Dougherty, I. T. Fiddes, Z. N.
Kronenberg, J. Shuga, A. A. Leyrat, J. A. West, M. Bershteyn, C. B. Lowe, B. J. Pavlovic,
S. R. Salama, D. Haussler, E. E. Eichler, A. R. Kriegstein, Establishing Cerebral Organoids
as Models of Human-Specific Brain Evolution. Cell. 176, 743–756.e17 (2019).
26. P. Di Tommaso, M. Chatzou, E. W. Floden, P. P. Barja, E. Palumbo, C. Notredame,
Submitted Manuscript: Confidential
Template revised February 2021
32
Nextflow enables reproducible computational workflows. Nat. Biotechnol. 35, 316–319
(2017).
27. M. J. Hubisz, K. S. Pollard, A. Siepel, PHAST and RPHAST: phylogenetic analysis with
space/time models. Brief. Bioinform. 12, 41–51 (2011).
28. A. Siepel, K. S. Pollard, D. Haussler, New Methods for Detecting Lineage-Specific
Selection. Lecture Notes in Computer Science (2006), pp. 190–205.
29. K. S. Pollard, M. J. Hubisz, K. R. Rosenbloom, A. Siepel, Detection of nonneutral
substitution rates on mammalian phylogenies. Genome Res. 20, 110–121 (2010).
30. D. Kostka, A. K. Holloway, K. S. Pollard, Developmental Loci Harbor Clusters of
Accelerated Regions That Evolved Independently in Ape Lineages. Mol. Biol. Evol. 35,
2034–2045 (2018).
31. C. Y. McLean, D. Bristor, M. Hiller, S. L. Clarke, B. T. Schaar, C. B. Lowe, A. M. Wenger,
G. Bejerano, GREAT improves functional interpretation of cis-regulatory regions. Nat.
Biotechnol. 28, 495–501 (2010).
32. C. S. Greene, A. Krishnan, A. K. Wong, E. Ricciotti, R. A. Zelaya, D. S. Himmelstein, R.
Zhang, B. M. Hartmann, E. Zaslavsky, S. C. Sealfon, D. I. Chasman, G. A. FitzGerald, K.
Dolinski, T. Grosser, O. G. Troyanskaya, Understanding multicellular function and disease
with human tissue-specific networks. Nat. Genet. 47, 569–576 (2015).
33. D. Kostka, M. W. Hahn, K. S. Pollard, Noncoding sequences near duplicated genes evolve
rapidly. Genome Biol. Evol. 2, 518–533 (2010).
34. S. S. P. Rao, M. H. Huntley, N. C. Durand, E. K. Stamenova, I. D. Bochkov, J. T.
Robinson, A. L. Sanborn, I. Machol, A. D. Omer, E. S. Lander, E. L. Aiden, A 3D map of
the human genome at kilobase resolution reveals principles of chromatin looping. Cell. 159,
1665–1680 (2014).
35. M. Spielmann, D. G. Lupiáñez, S. Mundlos, Structural variation in the 3D genome. Nat.
Rev. Genet. 19, 453–467 (2018).
36. G. Fudenberg, D. R. Kelley, K. S. Pollard, Predicting 3D genome folding from DNA
sequence with Akita. Nat. Methods. 17, 1111–1117 (2020).
37. S. Whalen, F. Inoue, H. Ryu, T. Fair, E. Markenscoff-Papadimitriou, K. Keough, M.
Kircher, B. Martin, B. Alvarado, O. Elor, D. L. Cintron, A. Williams, M. A. H. Samee, S.
Thomas, R. Krencik, E. M. Ullian, A. R. Kriegstein, J. Shendure, A. Pollen, N. Ahituv, K.
S. Pollard, Machine-Learning Dissection of Human Accelerated Regions in Primate
Neurodevelopment (2022), (available at https://papers.ssrn.com/abstract=4149954).
38. T. Yang, F. Zhang, G. G. Yardımcı, F. Song, R. C. Hardison, W. S. Noble, F. Yue, Q. Li,
HiCRep: assessing the reproducibility of Hi-C data using a stratum-adjusted correlation
coefficient. Genome Res. 27, 1939–1949 (2017).
Submitted Manuscript: Confidential
Template revised February 2021
33
39. J. R. Dixon, S. Selvaraj, F. Yue, A. Kim, Y. Li, Y. Shen, M. Hu, J. S. Liu, B. Ren,
Topological domains in mammalian genomes identified by analysis of chromatin
interactions. Nature. 485, 376–380 (2012).
40. M. Vietri Rudan, M. V. Rudan, C. Barrington, S. Henderson, C. Ernst, D. T. Odom, A.
Tanay, S. Hadjur, Comparative Hi-C Reveals that CTCF Underlies Evolution of
Chromosomal Domain Architecture. Cell Reports. 10 (2015), pp. 1297–1309.
41. I. E. Eres, K. Luo, C. J. Hsiao, L. E. Blake, Y. Gilad, Reorganization of 3D genome
structure may contribute to gene regulatory evolution in primates. PLoS Genet. 15,
e1008278 (2019).
42. X. Luo, Y. Liu, D. Dang, T. Hu, Y. Hou, X. Meng, F. Zhang, T. Li, C. Wang, M. Li, H.
Wu, Q. Shen, Y. Hu, X. Zeng, X. He, L. Yan, S. Zhang, C. Li, B. Su, 3D Genome of
macaque fetal brain reveals evolutionary innovations during primate corticogenesis. Cell.
184, 723–740.e21 (2021).
43. I. E. Eres, Y. Gilad, A TAD Skeptic: Is 3D Genome Topology Conserved? Trends Genet.
37, 216–223 (2021).
44. C. Hoencamp, O. Dudchenko, A. M. O. Elbatsh, S. Brahmachari, J. A. Raaijmakers, T. van
Schaik, Á. Sedeño Cacciatore, V. G. Contessoto, R. G. H. P. van Heesbeen, B. van den
Broek, A. N. Mhaskar, H. Teunissen, B. G. St Hilaire, D. Weisz, A. D. Omer, M. Pham, Z.
Colaric, Z. Yang, S. S. P. Rao, N. Mitra, C. Lui, W. Yao, R. Khan, L. L. Moroz, A. Kohn, J.
St Leger, A. Mena, K. Holcroft, M. C. Gambetta, F. Lim, E. Farley, N. Stein, A. Haddad, D.
Chauss, A. S. Mutlu, M. C. Wang, N. D. Young, E. Hildebrandt, H. H. Cheng, C. J. Knight,
T. L. U. Burnham, K. A. Hovel, A. J. Beel, P.-J. Mattei, R. D. Kornberg, W. C. Warren, G.
Cary, J. L. Gómez-Skarmeta, V. Hinman, K. Lindblad-Toh, F. Di Palma, K. Maeshima, A.
S. Multani, S. Pathak, L. Nel-Themaat, R. R. Behringer, P. Kaur, R. H. Medema, B. van
Steensel, E. de Wit, J. N. Onuchic, M. Di Pierro, E. Lieberman Aiden, B. D. Rowland, 3D
genomics across the tree of life reveals condensin II as a determinant of architecture type.
Science. 372, 984–989 (2021).
45. L. A. Mirny, M. Imakaev, N. Abdennur, Two major mechanisms of chromosome
organization. Current Opinion in Cell Biology. 58 (2019), pp. 142–152.
46. A. Roayaei Ardakany, H. T. Gezer, S. Lonardi, F. Ay, Mustache: multi-scale detection of
chromatin loops from Hi-C and Micro-C maps using scale-space representation. Genome
Biol. 21, 256 (2020).
47. A. G. Diehl, N. Ouyang, A. P. Boyle, Transposable elements contribute to cell and species-
specific chromatin looping and gene regulation in mammalian genomes. Nat. Commun. 11,
1796 (2020).
48. S. Kanton, M. J. Boyle, Z. He, M. Santel, A. Weigert, F. Sanchís-Calleja, P. Guijarro, L.
Sidow, J. S. Fleck, D. Han, Z. Qian, M. Heide, W. B. Huttner, P. Khaitovich, S. Pääbo, B.
Treutlein, J. G. Camp, Organoid single-cell genomic atlas uncovers human-specific features
of brain development. Nature. 574, 418–422 (2019).
Submitted Manuscript: Confidential
Template revised February 2021
34
49. B. J. Pavlovic, L. E. Blake, J. Roux, C. Chavarria, Y. Gilad, A Comparative Assessment of
Human and Chimpanzee iPSC-derived Cardiomyocytes with Primary Heart Tissues. Sci.
Rep. 8, 15312 (2018).
50. S. K. Sundaram, A. M. Huq, Z. Sun, W. Yu, L. Bennett, B. J. Wilson, M. E. Behen, H. T.
Chugani, Exome sequencing of a pedigree with tourette syndrome or chronic tic disorder.
Annals of Neurology. 69 (2011), pp. 901–904.
51. K. Lindblad-Toh, M. Garber, O. Zuk, M. F. Lin, B. J. Parker, S. Washietl, P. Kheradpour, J.
Ernst, G. Jordan, E. Mauceli, L. D. Ward, C. B. Lowe, A. K. Holloway, M. Clamp, S.
Gnerre, J. Alföldi, K. Beal, J. Chang, H. Clawson, J. Cuff, F. Di Palma, S. Fitzgerald, P.
Flicek, M. Guttman, M. J. Hubisz, D. B. Jaffe, I. Jungreis, W. J. Kent, D. Kostka, M. Lara,
A. L. Martins, T. Massingham, I. Moltke, B. J. Raney, M. D. Rasmussen, J. Robinson, A.
Stark, A. J. Vilella, J. Wen, X. Xie, M. C. Zody, Broad Institute Sequencing Platform and
Whole Genome Assembly Team, J. Baldwin, T. Bloom, C. W. Chin, D. Heiman, R. Nicol,
C. Nusbaum, S. Young, J. Wilkinson, K. C. Worley, C. L. Kovar, D. M. Muzny, R. A.
Gibbs, Baylor College of Medicine Human Genome Sequencing Center Sequencing Team,
A. Cree, H. H. Dihn, G. Fowler, S. Jhangiani, V. Joshi, S. Lee, L. R. Lewis, L. V. Nazareth,
G. Okwuonu, J. Santibanez, W. C. Warren, E. R. Mardis, G. M. Weinstock, R. K. Wilson,
Genome Institute at Washington University, K. Delehaunty, D. Dooling, C. Fronik, L.
Fulton, B. Fulton, T. Graves, P. Minx, E. Sodergren, E. Birney, E. H. Margulies, J. Herrero,
E. D. Green, D. Haussler, A. Siepel, N. Goldman, K. S. Pollard, J. S. Pedersen, E. S.
Lander, M. Kellis, A high-resolution map of human evolutionary constraint using 29
mammals. Nature. 478, 476–482 (2011).
52. R. N. Doan, B.-I. Bae, B. Cubelos, C. Chang, A. A. Hossain, S. Al-Saad, N. M. Mukaddes,
O. Oner, M. Al-Saffar, S. Balkhy, G. G. Gascon, Homozygosity Mapping Consortium for
Autism, M. Nieto, C. A. Walsh, Mutations in Human Accelerated Regions Disrupt
Cognition and Social Behavior. Cell. 167, 341–354.e12 (2016).
53. B. Castelijns, M. L. Baak, I. S. Timpanaro, C. R. M. Wiggers, M. W. Vermunt, P. Shang, I.
Kondova, G. Geeven, V. Bianchi, W. de Laat, N. Geijsen, M. P. Creyghton, Hominin-
specific regulatory elements selectively emerged in oligodendrocytes and are disrupted in
autism patients. Nat. Commun. 11, 301 (2020).
54. E. Markenscoff-Papadimitriou, S. Whalen, P. Przytycki, R. Thomas, F. Binyameen, T. J.
Nowakowski, A. R. Kriegstein, S. J. Sanders, M. W. State, K. S. Pollard, J. L. Rubenstein,
A Chromatin Accessibility Atlas of the Developing Human Telencephalon. Cell. 182, 754–
769.e18 (2020).
55. M. P. Forrest, H. Zhang, W. Moy, H. McGowan, C. Leites, L. E. Dionisio, Z. Xu, J. Shi, A.
R. Sanders, W. J. Greenleaf, C. A. Cowan, Z. P. Pang, P. V. Gejman, P. Penzes, J. Duan,
Open Chromatin Profiling in hiPSC-Derived Neurons Prioritizes Functional Noncoding
Psychiatric Risk Variants and Highlights Neurodevelopmental Loci. Cell Stem Cell. 21,
305–318.e8 (2017).
56. C. C. Funk, A. M. Casella, S. Jung, M. A. Richards, A. Rodriguez, P. Shannon, R.
Donovan-Maiye, B. Heavner, K. Chard, Y. Xiao, G. Glusman, N. Ertekin-Taner, T. E.
Submitted Manuscript: Confidential
Template revised February 2021
35
Golde, A. Toga, L. Hood, J. D. Van Horn, C. Kesselman, I. Foster, R. Madduri, N. D. Price,
S. A. Ament, Atlas of Transcription Factor Binding Sites from ENCODE DNase
Hypersensitivity Data across 27 Tissue Types. Cell Rep. 32, 108029 (2020).
57. M. Song, X. Yang, X. Ren, L. Maliskova, B. Li, I. R. Jones, C. Wang, F. Jacob, K. Wu, M.
Traglia, T. W. Tam, K. Jamieson, S.-Y. Lu, G.-L. Ming, Y. Li, J. Yao, L. A. Weiss, J. R.
Dixon, L. M. Judge, B. R. Conklin, H. Song, L. Gan, Y. Shen, Mapping cis-regulatory
chromatin contacts in neural cells links neuropsychiatric disorder risk variants to target
genes. Nat. Genet. 51, 1252–1262 (2019).
58. M. Song, M.-P. Pebworth, X. Yang, A. Abnousi, C. Fan, J. Wen, J. D. Rosen, M. N. K.
Choudhary, X. Cui, I. R. Jones, S. Bergenholtz, U. C. Eze, I. Juric, B. Li, L. Maliskova, J.
Lee, W. Liu, A. A. Pollen, Y. Li, T. Wang, M. Hu, A. R. Kriegstein, Y. Shen, Cell-type-
specific 3D epigenomes in the developing human cortex. Nature. 587, 644–649 (2020).
59. R. S. Ziffra, C. N. Kim, J. M. Ross, A. Wilfert, T. N. Turner, M. Haeussler, A. M. Casella,
P. F. Przytycki, K. C. Keough, D. Shin, D. Bogdanoff, A. Kreimer, K. S. Pollard, S. A.
Ament, E. E. Eichler, N. Ahituv, T. J. Nowakowski, Single-cell epigenomics reveals
mechanisms of human cortical development. Nature. 598, 205–213 (2021).
60. Roadmap Epigenomics Consortium, A. Kundaje, W. Meuleman, J. Ernst, M. Bilenky, A.
Yen, A. Heravi-Moussavi, P. Kheradpour, Z. Zhang, J. Wang, M. J. Ziller, V. Amin, J. W.
Whitaker, M. D. Schultz, L. D. Ward, A. Sarkar, G. Quon, R. S. Sandstrom, M. L. Eaton,
Y.-C. Wu, A. R. Pfenning, X. Wang, M. Claussnitzer, Y. Liu, C. Coarfa, R. A. Harris, N.
Shoresh, C. B. Epstein, E. Gjoneska, D. Leung, W. Xie, R. D. Hawkins, R. Lister, C. Hong,
P. Gascard, A. J. Mungall, R. Moore, E. Chuah, A. Tam, T. K. Canfield, R. S. Hansen, R.
Kaul, P. J. Sabo, M. S. Bansal, A. Carles, J. R. Dixon, K.-H. Farh, S. Feizi, R. Karlic, A.-R.
Kim, A. Kulkarni, D. Li, R. Lowdon, G. Elliott, T. R. Mercer, S. J. Neph, V. Onuchic, P.
Polak, N. Rajagopal, P. Ray, R. C. Sallari, K. T. Siebenthall, N. A. Sinnott-Armstrong, M.
Stevens, R. E. Thurman, J. Wu, B. Zhang, X. Zhou, A. E. Beaudet, L. A. Boyer, P. L. De
Jager, P. J. Farnham, S. J. Fisher, D. Haussler, S. J. M. Jones, W. Li, M. A. Marra, M. T.
McManus, S. Sunyaev, J. A. Thomson, T. D. Tlsty, L.-H. Tsai, W. Wang, R. A. Waterland,
M. Q. Zhang, L. H. Chadwick, B. E. Bernstein, J. F. Costello, J. R. Ecker, M. Hirst, A.
Meissner, A. Milosavljevic, B. Ren, J. A. Stamatoyannopoulos, T. Wang, M. Kellis,
Integrative analysis of 111 reference human epigenomes. Nature. 518, 317–330 (2015).
61. M. A. Petryniak, G. B. Potter, D. H. Rowitch, J. L. R. Rubenstein, Dlx1 and Dlx2 control
neuronal versus oligodendroglial cell fate acquisition in the developing forebrain. Neuron.
55, 417–433 (2007).
62. M. T. Schmitz, K. Sandoval, C. P. Chen, M. A. Mostajo-Radji, W. W. Seeley, T. J.
Nowakowski, C. J. Ye, M. F. Paredes, A. A. Pollen, The development and evolution of
inhibitory neurons in primate cerebrum. Nature. 603, 871–877 (2022).
63. M. Shibata, K. Pattabiraman, B. Lorente-Galdos, D. Andrijevic, X. Xing, A. M. M. Sousa,
G. Santpere, N. Sestan, Regulation of Prefrontal Patterning, Connectivity and
Synaptogenesis by Retinoic Acid. bioRxiv (2019), p. 2019.12.31.891036.
Submitted Manuscript: Confidential
Template revised February 2021
36
64. A. Visel, S. Minovitsky, I. Dubchak, L. A. Pennacchio, VISTA Enhancer Browser—a
database of tissue-specific human enhancers. Nucleic Acids Res. 35, D88–D92 (2006).
65. P. F. Przytycki, K. S. Pollard, CellWalker integrates single-cell and bulk data to resolve
regulatory elements across cell types in complex tissues. Genome Biol. 22, 61 (2021).
66. P. F. Przytycki, K. S. Pollard, CellWalkR: An R Package for integrating and visualizing
single-cell and bulk data to resolve regulatory elements. Bioinformatics (2022),
doi:10.1093/bioinformatics/btac150.
67. T. J. Nowakowski, A. Bhaduri, A. A. Pollen, B. Alvarado, M. A. Mostajo-Radji, E. Di
Lullo, M. Haeussler, C. Sandoval-Espinosa, S. J. Liu, D. Velmeshev, J. R. Ounadjela, J.
Shuga, X. Wang, D. A. Lim, J. A. West, A. A. Leyrat, W. J. Kent, A. R. Kriegstein,
Spatiotemporal gene expression trajectories reveal developmental hierarchies of the human
cortex. Science. 358, 1318–1323 (2017).
68. R. D. Hodge, T. E. Bakken, J. A. Miller, K. A. Smith, E. R. Barkan, L. T. Graybuck, J. L.
Close, B. Long, N. Johansen, O. Penn, Z. Yao, J. Eggermont, T. Höllt, B. P. Levi, S. I.
Shehata, B. Aevermann, A. Beller, D. Bertagnolli, K. Brouner, T. Casper, C. Cobbs, R.
Dalley, N. Dee, S.-L. Ding, R. G. Ellenbogen, O. Fong, E. Garren, J. Goldy, R. P. Gwinn,
D. Hirschstein, C. D. Keene, M. Keshk, A. L. Ko, K. Lathia, A. Mahfouz, Z. Maltzer, M.
McGraw, T. N. Nguyen, J. Nyhus, J. G. Ojemann, A. Oldre, S. Parry, S. Reynolds, C.
Rimorin, N. V. Shapovalova, S. Somasundaram, A. Szafer, E. R. Thomsen, M. Tieu, G.
Quon, R. H. Scheuermann, R. Yuste, S. M. Sunkin, B. Lelieveldt, D. Feng, L. Ng, A.
Bernard, M. Hawrylycz, J. W. Phillips, B. Tasic, H. Zeng, A. R. Jones, C. Koch, E. S. Lein,
Conserved cell types with divergent features in human versus mouse cortex. Nature. 573,
61–68 (2019).
69. B. Tasic, Z. Yao, L. T. Graybuck, K. A. Smith, T. N. Nguyen, D. Bertagnolli, J. Goldy, E.
Garren, M. N. Economo, S. Viswanathan, O. Penn, T. Bakken, V. Menon, J. Miller, O.
Fong, K. E. Hirokawa, K. Lathia, C. Rimorin, M. Tieu, R. Larsen, T. Casper, E. Barkan, M.
Kroll, S. Parry, N. V. Shapovalova, D. Hirschstein, J. Pendergraft, H. A. Sullivan, T. K.
Kim, A. Szafer, N. Dee, P. Groblewski, I. Wickersham, A. Cetin, J. A. Harris, B. P. Levi, S.
M. Sunkin, L. Madisen, T. L. Daigle, L. Looger, A. Bernard, J. Phillips, E. Lein, M.
Hawrylycz, K. Svoboda, A. R. Jones, C. Koch, H. Zeng, Shared and distinct transcriptomic
cell types across neocortical areas. Nature. 563, 72–78 (2018).
70. J. D. Hocker, O. B. Poirion, F. Zhu, J. Buchanan, K. Zhang, J. Chiou, T.-M. Wang, Q.
Zhang, X. Hou, Y. E. Li, Y. Zhang, E. N. Farah, A. Wang, A. D. McCulloch, K. J. Gaulton,
B. Ren, N. C. Chi, S. Preissl, Cardiac cell type–specific gene regulatory programs and
disease risk association. Science Advances. 7, eabf1444 (2021).
71. H. Won, J. Huang, C. K. Opland, C. L. Hartl, D. H. Geschwind, Human evolved regulatory
elements modulate genes involved in cortical expansion and neurodevelopmental disease
susceptibility. Nat. Commun. 10, 2396 (2019).
72. A. Siepel, G. Bejerano, J. S. Pedersen, A. S. Hinrichs, M. Hou, K. Rosenbloom, H.
Clawson, J. Spieth, L. W. Hillier, S. Richards, G. M. Weinstock, R. K. Wilson, R. A. Gibbs,
Submitted Manuscript: Confidential
Template revised February 2021
37
W. J. Kent, W. Miller, D. Haussler, Evolutionarily conserved elements in vertebrate, insect,
worm, and yeast genomes. Genome Res. 15, 1034–1050 (2005).
73. D. Karolchik, A. S. Hinrichs, T. S. Furey, K. M. Roskin, C. W. Sugnet, D. Haussler, W. J.
Kent, The UCSC Table Browser data retrieval tool. Nucleic Acids Res. 32, D493–6 (2004).
74. A. Frankish, M. Diekhans, A.-M. Ferreira, R. Johnson, I. Jungreis, J. Loveland, J. M.
Mudge, C. Sisu, J. Wright, J. Armstrong, I. Barnes, A. Berry, A. Bignell, S. C. Sala, J.
Chrast, F. Cunningham, T. Di Domenico, S. Donaldson, I. T. Fiddes, C. G. Girón, J. M.
Gonzalez, T. Grego, M. Hardy, T. Hourlier, T. Hunt, O. G. Izuogu, J. Lagarde, F. J. Martin,
L. Martínez, S. Mohanan, P. Muir, F. C. P. Navarro, A. Parker, B. Pei, F. Pozo, M. Ruffier,
B. M. Schmitt, E. Stapleton, M.-M. Suner, I. Sycheva, B. Uszczynska-Ratajczak, J. Xu, A.
Yates, D. Zerbino, Y. Zhang, B. Aken, J. S. Choudhary, M. Gerstein, R. Guigó, T. J. P.
Hubbard, M. Kellis, B. Paten, A. Reymond, M. L. Tress, P. Flicek, GENCODE reference
annotation for the human and mouse genomes. Nucleic Acids Research. 47 (2019), pp.
D766–D773.
75. X. Jiao, B. T. Sherman, D. W. Huang, R. Stephens, M. W. Baseler, H. C. Lane, R. A.
Lempicki, DAVID-WS: a stateful web service to facilitate gene/protein list analysis.
Bioinformatics. 28, 1805–1806 (2012).
76. S. J. Liu, T. J. Nowakowski, A. A. Pollen, J. H. Lui, M. A. Horlbeck, F. J. Attenello, D. He,
J. S. Weissman, A. R. Kriegstein, A. A. Diaz, D. A. Lim, Single-cell analysis of long non-
coding RNAs in the developing human neocortex. Genome Biol. 17, 67 (2016).
77. A. S. Hinrichs, D. Karolchik, R. Baertsch, G. P. Barber, G. Bejerano, H. Clawson, M.
Diekhans, T. S. Furey, R. A. Harte, F. Hsu, J. Hillman-Jackson, R. M. Kuhn, J. S. Pedersen,
A. Pohl, B. J. Raney, K. R. Rosenbloom, A. Siepel, K. E. Smith, C. W. Sugnet, A. Sultan-
Qurraie, D. J. Thomas, H. Trumbower, R. J. Weber, M. Weirauch, A. S. Zweig, D.
Haussler, W. J. Kent, The UCSC Genome Browser Database: update 2006. Nucleic Acids
Res. 34, D590–8 (2006).
78. O. Tange, GNU Parallel 20200922 (’Ginsburg') (2020), doi:10.5281/zenodo.4045386.
79. S. Venev, N. Abdennur, A. Goloborodko, I. Flyamer, G. Fudenberg, J. Nuebler, A.
Galitsyna, B. Akgol, S. Abraham, P. Kerpedjiev, M. Imakaev, mirnylab/cooltools: v0.3.2
(Zenodo, 2020; https://zenodo.org/record/3787004).
80. H. Li, Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM
(2013), (available at http://arxiv.org/abs/1303.3997).
81. open2c, open2c/pairtools, (available at https://github.com/open2c/pairtools).
82. M. Imakaev, G. Fudenberg, R. P. McCord, N. Naumova, A. Goloborodko, B. R. Lajoie, J.
Dekker, L. A. Mirny, Iterative correction of Hi-C data reveals hallmarks of chromosome
organization. Nat. Methods. 9, 999–1003 (2012).
83. B. R. Lajoie, J. Dekker, N. Kaplan, The Hitchhiker’s guide to Hi-C analysis: Practical
guidelines. Methods. 72 (2015), pp. 65–75.
Submitted Manuscript: Confidential
Template revised February 2021
38
84. open2c, open2c/open2c_examples, (available at
https://github.com/open2c/open2c_examples).
85. M. G. Gordon, F. Inoue, B. Martin, M. Schubach, V. Agarwal, S. Whalen, S. Feng, J. Zhao,
T. Ashuach, R. Ziffra, A. Kreimer, I. Georgakopoulos-Soares, N. Yosef, C. J. Ye, K. S.
Pollard, J. Shendure, M. Kircher, N. Ahituv, lentiMPRA and MPRAflow for high-
throughput functional characterization of gene regulatory elements. Nat. Protoc. 15, 2387–
2412 (2020).
86. L. Duret, N. Galtier, Biased gene conversion and the evolution of mammalian genomic
landscapes. Annu. Rev. Genomics Hum. Genet. 10, 285–311 (2009).
87. E. Takao-Rikitsu, S. Mochida, E. Inoue, M. Deguchi-Tawarada, M. Inoue, T. Ohtsuka, Y.
Takai, Physical and functional interaction of the active zone proteins, CAST, RIM1, and
Bassoon, in neurotransmitter release. J. Cell Biol. 164, 301–311 (2004).
88. M. P. Coba, N. H. Komiyama, J. Nithianantharajah, M. V. Kopanitsa, T. Indersmitten, N.
G. Skene, E. J. Tuck, D. G. Fricker, K. A. Elsegood, L. E. Stanford, N. O. Afinowi, L. M.
Saksida, T. J. Bussey, T. J. O’Dell, S. G. N. Grant, TNiK is required for postsynaptic and
nuclear signaling pathways and cognitive function. J. Neurosci. 32, 13987–13999 (2012).
89. H. Won, L. de la Torre-Ubieta, J. L. Stein, N. N. Parikshak, J. Huang, C. K. Opland, M. J.
Gandal, G. J. Sutton, F. Hormozdiari, D. Lu, C. Lee, E. Eskin, I. Voineagu, J. Ernst, D. H.
Geschwind, Chromosome conformation elucidates regulatory relationships in developing
human brain. Nature. 538, 523–527 (2016).
Submitted Manuscript: Confidential
Template revised February 2021
39
Acknowledgments: We thank Maureen Pittman, Geoffrey Fudenberg, Abigail Lind, Evonne
McArthur, Ryan Ziffra, Tony Capra and Svetlana Lyalina for helpful discussions, sharing code
and suggestions towards the results shown in this work. We thank Giovanni Maki for assistance
with figures and visualization. This project was funded by NIH NHGRI R01HG008742 and the
Swedish Research Council Distinguished Professor Award.
Funding:
Discovery Fellowship (KCK)
National Institutes of Health GR-01125 (KCK, KSP)
National Institute of Mental Health R01MH109907, U01MH116438 (NA, KSP)
Gladstone Institutes (KSP)
NIH DP2MH122400-01 (AAP, TF)
Schmidt Futures Foundation (AAP, TF)
Shurl and Kay Curci Foundation (AAP, TF)
NIH NHGRI R01HG008742 (EK)
Swedish Research Council Distinguished Professor Award (KLT)
Author contributions:
Conceptualization: KCK, KSP
Methodology: KCK, SW, PFP, TF, FI, HR, NA, AP, ZC, KSP
Investigation: KCK, SW, PFP, TF, FI, HR
Visualization: KCK, PFP
Funding acquisition: NA, KSP
Supervision: NA, AP, KSP
Writing - original draft: KCK, KSP
Writing - review & editing: All authors
Competing interests: Authors declare they have no competing interests.
Data and materials availability: The Zoonomia data are available at
https://zoonomiaproject.org/the-project/. The Nextflow pipeline to identify lineage-specific
accelerated regions is available at https://github.com/keoughkath/AcceleratedRegionsNF.
The Hi-C data are available at GSE183137. All other data are available in the main text or
the supplementary materials.
Supplementary Materials
Materials and Methods
Supplementary Text
Figs. S1 to S11
Tables S1 to S7
Submitted Manuscript: Confidential
Template revised February 2021
40
References (31–74)
Submitted Manuscript: Confidential
Template revised February 2021
41
Submitted Manuscript: Confidential
Template revised February 2021
42
Fig. 1. Human-specific structural variants are enriched in zooHAR chromatin domains and
predicted to change the 3D genome. (A) Pipeline to identify lineage-specific accelerated
regions. Blue circles indicate initial input data, purple hexagons are intermediate results, and the
green square is the final output. (B) Odds ratio of chromatin contact domains in GM12878 cells
(34) containing hsSVs and zooHARs (green line) compared to a null distribution (shaded blue
region) of odds ratios for chromatin contact domains containing conserved (phastCons) elements
and hsSVs from 1000 random draws of phastCons equaling the number of zooHARs. (C) Akita
prediction of a locus (hg38.chr4:26614489-27531993, hsSV: human-specific insertion
O_000012F_1_28503465_quiver_pilon_11099913_11099913 from (22)) with a human-specific
insertion (“Original”), with the human-specific insertion deleted in silico (“hsSV deleted”) and a
subtraction matrix (“Original - hsSV deleted”) comparing the chromatin contact matrices with
and without the human-specific insertion. White boxes indicate regions that change in the
“Original” compared to the “hsSV deleted” sequences. Log(observed/expected) contact values
are shown in the heatmaps.
Submitted Manuscript: Confidential
Template revised February 2021
43
Fig. 2. Human-specific structural variants change the 3D genome around zooHARs and
zooCHARs. White boxes highlight differences between the species. Log(observed/expected)
values are shown in the heatmaps. (A, B): Subtraction matrices for the in silico predicted change
due to the human-specific insertion (left) and observed chromatin contact maps in human
compared to chimpanzee NPC Hi-C (right) for the loci containing zooHAR.126
(hg38.chr4:26614489-27531993; hsSV:
chr4_27070203_DEL_chimpanzee_000012F_1_28503465_quiver_pilon_11099913_11099913
from (22))) and zooHAR.15 (hg38.chr16:79237694-80155198; hsSV:
chr16_79695894_DEL_chimpanzee_000093F_1_10181781_quiver_pilon_1690619_1690619
from (22)), respectively. (C, D): Human (top) and chimpanzee (bottom) log(observed/expected)
Hi-C contact frequencies in each locus, with the disruption score (10 kilobase resolution) in
between. (E, F): zooHAR locations denoted by vertical lines adjacent to their names. Conserved
Submitted Manuscript: Confidential
Template revised February 2021
44
(blue), chimpanzee-specific (green), and human-specific (orange) loops (5 kilobase resolution,
loops called with Mustache (46))
Submitted Manuscript: Confidential
Template revised February 2021
45
Submitted Manuscript: Confidential
Template revised February 2021
46
Fig. 3. zooHARs in human brain development. (A) Transcription factor footprints (56) and
epigenomic marks (60) overlapping zooHAR.126. NSC: neural stem cell. (B) Subset of enriched
transcription factor footprints in zooHARs relative to phastCons elements (Fisher’s exact p-value
≤ 0.05). Full set available in Table S6. (C) Cell types in which zooHARs are predicted to
regulate gene expression based on CellWalker analysis of data from the developing human
telencephalon. (D) Cell type assignments for zooCHARs based on CellWalker analysis of data
from the developing human telencephalon. Unlike with HARs, no CHARs map to late stage
excitatory neurons. Abbreviations of cell types for (C, D); excitatory neurons (EN) derived from
primary visual cortex (V1) or prefrontal cortex (PFC), newborn excitatory neurons (nEN),
inhibitory cortical interneurons (IN-CTX) originating in the medial/caudal ganglionic eminence
(MGE/CGE), newborn interneurons (nIN), intermediate progenitor cells (IPC), and
truncated/ventral/outer radial glia (tRG/vRG/oRG). More cell type information is available at
https://cells.ucsc.edu/?ds=cortex-dev (59, 65).
Submitted Manuscript: Confidential
Template revised February 2021
47
Zoonomia Consortium Authors - Collaborators:
Gregory Andrews1, Joel C. Armstrong2, Matteo Bianchi3, Bruce W. Birren4, Kevin Bredemeyer5,
Ana M Breit6, Matthew J Christmas3, Joana Damas7, Mark Diekhans2, Michael X. Dong3,
Eduardo Eizirik8, Kaili Fan1, Cornelia Fanter9, Nicole M. Foley5, Karin Forsberg-Nilsson10,
Carlos J. Garcia11, John Gatesy12, Steven Gazal13, Diane P. Genereux4, Daniel Goodman14, Linda
Goodman15, Jenna Grimshaw11, Michaela K. Halsey11, Andrew Harris5, Glenn Hickey16, Michael
Hiller17, Allyson Hindle9, Robert M. Hubley18, Graham Hughes19, Jeremy Johnson4, David
Juan20, Irene M. Kaplow21,22, Elinor K. Karlsson1,4, Kathleen C. Keough23,24, Bogdan
Kirilenko17, Jennifer M. Korstian11, Sergey V. Kozyrev3, Alyssa J. Lawler25, Colleen Lawless19,
Danielle L. Levesque6, Harris A. Lewin 7,26,27, Xue Li1,4 , Abigail Lind23,24, Kerstin Lindblad-
Toh3,4, Voichita D. Marinescu3, Tomas Marques-Bonet20, Victor Mason28, Jennifer R. S.
Meadows3, Jill Moore1, Diana D. Moreno-Santillan11, Kathleen M. Morrill1,4, Gerhard
Muntané20, William Murphy5, Arcadi Navarro20, Martin Nweeia29,30,31,32, Austin Osmanski11,
Benedict Paten2, Nicole S. Paulat11, Eric Pederson3, Andreas R. Pfenning21,22, BaDoi N. Phan21,
Katherine S. Pollard23,24,33, Kavya Prasad21, Henry Pratt1, David A. Ray11, Jeb Rosen18, Irina Ruf
34, Louise Ryan19, Oliver Ryder35,36, Daniel Schäffer21, Aitor Serres20, Beth Shapiro37,38, Arian F.
A. Smit18, Mark Springer39, Chaitanya Srinivasan21, Cynthia Steiner40, Jessica M. Storer18,
Patrick F. Sullivan41, Kevin A. M. Sullivan10, Elisabeth Sundström3, Megan A Supple38, Ross
Swofford4, Joy-El Talbot42, Emma Teeling19, Jason Turner-Maier4, Alejandro Valenzuela20,
Franziska Wagner34, Ola Wallerman3, Chao Wang3, Juehan Wang13, Zhiping Weng1, Aryn P.
Wilder35, Morgan E. Wirthlin21,22, Shuyang Yao43, Xiaomeng Zhang2
Submitted Manuscript: Confidential
Template revised February 2021
48
Zoonomia Author Affiliations:
1 Program in Bioinformatics and Integrative Biology, University of Massachusetts Medical
School, Worcester, MA 01605, USA
2 Genomics Institute, UC Santa Cruz, 1156 High Street, Santa Cruz, CA 95064, USA
3 Science for Life Laboratory, Department of Medical Biochemistry and Microbiology, Uppsala
University, Uppsala, 751 32, Sweden
4 Broad Institute of MIT and Harvard, Cambridge MA 02139, USA
5 Veterinary Integrative Biosciences, Texas A&M University, College Station, TX 77843, USA
6 School of Biology and Ecology, University of Maine, Orono, Maine 04469, USA
7 The Genome Center, University of California Davis, Davis, CA 95616, USA
8 School of Health and Life Sciences, Pontifical Catholic University of Rio Grande do Sul, Porto
Alegre, 90619-900, Brazil
9 School of Life Sciences, University of Nevada Las Vegas, Las Vegas, NV 89154, USA
10 Department of Immunology, Genetics and Pathology, Science for Life Laboratory, Uppsala
University, Uppsala, 751 85, Sweden
11 Department of Biological Sciences, Texas Tech University, Lubbock, TX 79409, USA
12 Division of Vertebrate Zoology, American Museum of Natural History, New York, NY
10024, USA
13 Keck School of Medicine, University of Southern California, Los Angeles, CA 90033, USA
14 University of California San Francisco, San Francisco, CA 94143 USA
15 Fauna Bio Inc., Emeryville, CA 94608, USA
16 Baskin School of Engineering, University of California Santa Cruz, Santa Cruz, CA 95064,
USA
17 Max Planck Institute of Molecular Cell Biology and Genetics, 01307, Dresden, Germany
18 Institute for Systems Biology, Seattle, WA 98109, USA
19 School of Biology and Environmental Science, University College Dublin, Belfield, Dublin 4,
Ireland
20 Institute of Evolutionary Biology (UPF-CSIC), Department of Experimental and Health
Sciences, Universitat Pompeu Fabra, Barcelona, 08003, Spain
Submitted Manuscript: Confidential
Template revised February 2021
49
21 Department of Computational Biology, School of Computer Science, Carnegie Mellon
University, Pittsburgh, PA 15213, USA
22 Neuroscience Institute, Carnegie Mellon University, Pittsburgh, PA 15213, USA
23 Gladstone Institutes, San Francisco, CA 94158, USA
24 Department of Epidemiology & Biostatistics, University of California, San Francisco, CA
94158, USA
25 Department of Biology, Carnegie Mellon University, Pittsburgh, PA 15213, USA
26 Department of Evolution and Ecology, University of California, Davis, CA 95616, USA
27 John Muir Institute for the Environment, University of California, Davis, CA 95616, USA
28 Institute of Cell Biology, University of Bern, 3012 Bern, Switzerland
29 Narwhal Genome Initiative, Department of Restorative Dentistry and Biomaterials Sciences,
Harvard School of Dental Medicine, Boston, MA 02115, USA
30 Department of Comprehensive Care, School of Dental Medicine, Case Western Reserve
University, Cleveland, OH 44106, USA
31 Department of Vertebrate Zoology, Smithsonian Institution, Washington, DC 20002, USA
32 Department of Vertebrate Zoology, Canadian Museum of Nature, Ottawa, Ontario K2P 2R1,
Canada
33 Chan Zuckerberg Biohub, San Francisco, CA 94158, USA
34 Division of Messel Research and Mammalogy, Senckenberg Research Institute and Natural
History Museum Frankfurt, 60325 Frankfurt am Main, Germany
35 Conservation Genetics, San Diego Zoo Wildlife Alliance, Escondido, CA 92027, USA
36 Department of Evolution, Behavior and Ecology, Division of Biology, University of
California, San Diego, La Jolla, CA 92039 USA
37 Howard Hughes Medical Institute, University of California Santa Cruz, Santa Cruz, CA
95064, USA
38 Department of Ecology and Evolutionary Biology, University of California Santa Cruz, Santa
Cruz, CA 95064, USA
39 Department of Evolution, Ecology and Organismal Biology, University of California,
Riverside, CA 92521, USA
40 Conservation Science Wildlife Health, San Diego Zoo Wildlife Alliance, Escondido CA
92027, USA
Submitted Manuscript: Confidential
Template revised February 2021
50
41 Department of Genetics, University of North Carolina Medical School, Chapel Hill, NC
27599, USA
42 Iris Data Solutions, LLC, Orono, ME 04473, USA
43 Department of medical Epidemiology and Biostatistics, Karolinska Institute, Stockholm, 171
77, Sweden
| 2022 | Three-dimensional genome re-wiring in loci with Human Accelerated Regions | 10.1101/2022.10.04.510859 | null | creative-commons |
1
Chromatin sensing by the auxiliary domains of KDM5C
regulates its demethylase activity and is disrupted by X-
linked intellectual disability mutations
Fatima S. Ugur1,3, Mark J. S. Kelly2, and Danica Galonić Fujimori2,3,4*
1 Chemistry and Chemical Biology Graduate Program, 2 Department of Pharmaceutical
Chemistry, 3 Department of Cellular and Molecular Pharmacology, 4 Quantitative Biosciences
Institute, University of California, San Francisco
600 16th St., San Francisco, CA 94158, USA
*To whom correspondence should be addressed: danica.fujimori@ucsf.edu
2
ABSTRACT
The H3K4me3 chromatin modification, a hallmark of promoters of actively transcribed genes, is
dynamically removed by the KDM5 family of histone demethylases. The KDM5 demethylases
have a number of accessory domains, two of which, ARID and PHD1, lie between the segments
of the catalytic domain. KDM5C, which has a unique role in neural development, harbors a
number of mutations adjacent to its accessory domains that cause X-linked intellectual disability
(XLID). The roles of these accessory domains remain unknown, limiting an understanding of
how XLID mutations affect KDM5C activity. Through in vitro binding and kinetic studies using
nucleosomes, we find that while the ARID domain is required for efficient nucleosome
demethylation, the PHD1 domain alone has an inhibitory role in KDM5C catalysis. In addition,
the unstructured linker region between the ARID and PHD1 domains interacts with PHD1 and is
necessary for nucleosome binding. Our data suggests a model in which the PHD1 domain
inhibits DNA recognition by KDM5C. This inhibitory effect is relieved by the H3 tail, enabling
recognition of flanking DNA on the nucleosome. Importantly, we find that XLID mutations
adjacent to the ARID and PHD1 domains break this regulation by enhancing DNA binding,
resulting in the loss of specificity of substrate chromatin recognition and rendering demethylase
activity lower in the presence of flanking DNA. Our findings suggest a model by which specific
XLID mutations could alter chromatin recognition and enable euchromatin-specific dysregulation
of demethylation by KDM5C.
Keywords: histone demethylase, nucleosome, NMR, reader domain, intrinsically disordered
region
3
INTRODUCTION
The methylation of lysine 4 on histone H3 is a chromatin modification found on
euchromatin, where H3K4 trimethylation (H3K4me3) is present at gene promoter regions
associated with active transcription, and where H3K4 monomethylation (H3K4me1) is found at
active enhancer regions [1]. While H3K4me1/2 is demethylated by the KDM1/LSD family,
H3K4me1/2/3 is dynamically regulated by the KDM5/JARID1 subfamily of Jumonji histone
demethylases [2–7]. This demethylase family harbors unique auxiliary domains in addition to its
catalytic domain comprised of the JmjN and JmjC segments that form a composite active site
for demethylation [8,9]. KDM5A (RBP2, JARID1A), KDM5B (PLU-1, JARID1B), KDM5C (SMCX,
JARID1C), and KDM5D (SMCY, JARID1D) all contain an AT-rich interaction domain (ARID),
C5HC2 zinc finger domain (ZnF), and 2-3 plant homeodomains (PHD1-3). Unique to the KDM5
family is the insertion of the ARID and PHD1 domains between the JmjN and JmjC segments of
the catalytic domain, and ARID and PHD1 are required for demethylase activity in vivo [4,10–
13]. ARID domains are DNA binding domains, and ARID of KDM5A/B has been shown to bind
to GC-rich DNA [13–15]. PHD domains are H3K4 methylation reader domains with varying
specificity towards unmethylated and methylated H3K4 states [16–22]. PHD1 of KDM5A/B
preferentially binds the unmethylated H3 tail, and this recognition of the demethylation product
allosterically stimulates demethylase activity of KDM5A in vitro [23–28]. In contrast, the ARID
and PHD1 domains have not been extensively studied in KDM5C, which possesses a unique
function in neural development and has nonredundant demethylase activity [2,29].
KDM5C is ubiquitously expressed but has highest expression levels in the brain [30,31].
This demethylase is important for neural development and dendrite morphogenesis, and
KDM5C knockout mice have abnormal dendritic branching and display memory defects,
impaired social behavior, and aggression [2,29]. KDM5C fine-tunes the expression of
4
neurodevelopmental genes, as gene expression levels only change less than 2 fold upon
knockout of KDM5C in mice [29,32]. KDM5C localizes to enhancers in addition to promoter
regions and has been shown to also demethylate spurious H3K4me3 at enhancers during
neuronal maturation [29,32–34]. In line with its neurodevelopmental function, a number of
missense and nonsense mutations that cause X-linked intellectual disability (XLID) are found
throughout KDM5C [31,35–39]. As KDM5C is located on the X-chromosome and the Y paralog
KDM5D cannot compensate for its function, males with KDM5C XLID mutations are primarily
affected with a range of mild to severe symptoms of limitations in cognition, memory, and
adaptive behavior [30,31,37,38,40]. Some functionally characterized mutations have been
shown to reduce demethylase activity despite not occurring in the catalytic domains, and a
select few mutations have been shown to not affect demethylase activity, disrupting
nonenzymatic functions instead [2,11,39,41,42]. The consequences of these XLID mutations on
KDM5C at its target regions within chromatin and their impact on gene expression during neural
development is not fully understood. Interestingly, a number of XLID mutations are present
throughout and in between the accessory domains of KDM5C, suggesting potential disruption of
their regulatory functions. The impact of these mutations on demethylase regulation is hindered
by the limited understanding of the accessory domain roles in KDM5C.
Here, we sought to determine the functions of the ARID and PHD1 auxiliary domains in
KDM5C and evaluate whether these functions might be disrupted by XLID mutations. We
approached these questions by interrogating the recognition and demethylation of nucleosomes
by KDM5C, as nucleosome substrates enable extended interactions by multiple domains of the
demethylase. Our findings reveal that the ARID and PHD1 domains, as well as the linker
between them, regulate nucleosome demethylation and chromatin recognition by KDM5C. We
find that DNA recognition by ARID contributes to nucleosome demethylation but not
nucleosome binding, which is instead driven by the unstructured linker between ARID and
5
PHD1. In contrast, we find that PHD1 inhibits demethylation. Furthermore, we find that XLID
mutations near these regulatory domains disrupt interdomain interactions and enhance affinity
towards nucleosomes, resulting in nonproductive chromatin binding and inhibition of
demethylation in the presence of flanking DNA. Our findings define functional roles of the ARID
and PHD1 domains in the regulation of KDM5C and provide rationale for disruption of this
regulation by mutations in X-linked intellectual disability.
RESULTS
ARID & PHD1 region contributes to productive nucleosome demethylation
Previous work has demonstrated that KDM5C is capable of demethylating H3K4me3
peptides and that the catalytic JmjN-JmjC domain and zinc finger domain are necessary for
demethylase activity [2,8,42]. To evaluate the contributions of the ARID and PHD1 domains, we
sought to interrogate the recognition and demethylation of nucleosomes, given the expected
interactions of these domains with DNA and histone tails, respectively. We utilized an N-terminal
fragment of KDM5C containing the residues 1 to 839 necessary to monitor demethylation in
vitro (KDM5C1-839), as well as an analogous construct where the ARID and PHD1 region
(residues 83 to 378) is replaced by a short linker (KDM5C1-839 ∆AP) (Figure 1A) [8]. We
measured binding affinities of these constructs to both unmodified and substrate H3K4me3 core
nucleosomes containing 147 bp DNA by electrophoretic mobility shift assay. KDM5C binds
nucleosomes with weak affinity and with approximately a two fold affinity gain towards substrate
nucleosomes, with Kdapp of ~7 µM for the H3K4me3 nucleosome and ~13 µM for the unmodified
nucleosome (Figure 1B). Surprisingly, the ARID and PHD1 domains have a modest contribution
to nucleosome binding, as KDM5C1-839 ∆AP displays only a ~3 fold reduction in nucleosome
affinity and retains the two fold preference towards the substrate nucleosome (Figure 1B). The
6
absence of a significant enhancement of nucleosome binding through ARID and PHD1 domain-
mediated interactions suggests a more complex role of these domains rather than simply
facilitating chromatin recruitment.
Figure 1. The ARID & PHD1 region of KDM5C contributes to efficient nucleosome demethylation and has a
modest contribution to nucleosome binding.
(A) Domain architecture of KDM5C and KDM5C constructs used in this study. (B) Unmodified and substrate
nucleosome binding by KDM5C constructs with apparent dissociation constants (Kdapp) measured by EMSA (binding
curves in Figure S1B). Due to unattainable saturation of binding for the unmodified nucleosome, a lower limit for the
dissociation constant is presented. (C) Demethylation kinetics of the H3K4me3 substrate nucleosome by KDM5C
constructs under single turnover conditions (enzyme in excess of substrate). Observed rates are fit to a cooperative
kinetic model, with n denoting the Hill coefficient. Representative kinetic traces used to determine observed
demethylation rates are in Figure S1C. All error bars represent SEM of at least three independent experiments
(n ≥ 3).
We next interrogated the demethylase activity of KDM5C towards the H3K4me3
substrate nucleosome in vitro by utilizing a TR-FRET based kinetic assay that detects formation
of the H3K4me1/2 product nucleosome. In order to measure true catalytic rates (kmax),
demethylation was performed under single turnover conditions with enzyme in excess [43]. We
find that KDM5C1-839 demethylates the substrate nucleosome with an observed catalytic rate of
~0.09 min-1 and KDM5C1-839 ∆AP with a 3-fold lower catalytic rate of ~0.03 min-1 (Figure 1C),
A
B
C
1
1560
JmjN ARID
PHD1 JmjC Zf
PHD2
KDM5C
1
839
JmjN ARID
PHD1 JmjC Zf
KDM5C1-839
1
839
JmjN JmjC Zf
KDM5C1-839 ���
���������
KDM5C1-839
KDM5C1-839 AP
0
10
20
30
40
Kd
apparent ( M)
6.9
35.3
12.5
19.6
unmod nuc
H3K4me3 nuc
WT
���
0
5
10
15
20
0.00
0.02
0.04
0.06
[KDM5C construct] ( M)
kobs (min-1)
H3K4me3 nucleosome
kmax (min-1)
Km
app ( M)
KDM5C1-839
KDM5C1-839 ��
5.7 � 1.7
0.087 ��0.018
18.1 � 2.8
0.032 ��0.007
n
1.5
3.1
7
indicating that the ARID and PHD1 region contributes to productive catalysis on nucleosomes.
The contribution of the ARID and PHD1 domain region towards efficient demethylation appears
to be through interactions of these domains with the nucleosome, as the catalytic efficiency
(kmax/Kmapp) of KDM5C1-839 ∆AP relative to wild type is only 3-fold lower on the substrate
H3K4me3 peptide (Figure S1A), as opposed to the 9-fold reduction in catalytic efficiency on the
substrate nucleosome. As the ARID and PHD1 domains are poorly functionally characterized in
KDM5C, we sought to next investigate the features of the nucleosome that they recognize.
PHD1 domain inhibits KDM5C catalysis
The PHD1 domain of KDM5C has been previously shown to bind to H3K9me3 through
peptide pull down [2]. To interrogate the histone binding and specificity of PHD1, we purified the
PHD1 domain and quantified binding to histone peptides by nuclear magnetic resonance (NMR)
spectroscopy and bio-layer interferometry (BLI). We observe near identical binding between H3
and H3K9me3 tail peptides, indicating no specific binding of PHD1 towards the H3K9me3
modification (Figure S2A). Furthermore, we observe biphasic binding kinetics of PHD1 binding
the H3 tail peptide, indicative of a two step binding mechanism (Figure S2B). Upon titration of
the H3 tail, large chemical shift changes occur in the two-dimensional heteronuclear single
quantum coherence (HSQC) NMR spectrum of a majority of assigned residues in PHD1 (Figure
2A, Figure S2C). The observed affinity of PHD1 towards the H3 tail is surprisingly weak with a
dissociation constant of 130 µM, about 100 fold weaker than the affinity of the homologous
PHD1 of KDM5A towards the H3 tail (Figure 2B) [25,28]. Despite this difference in affinity,
PHD1 of KDM5C retains similar specificity towards the unmodified H3 tail over H3K4 methylated
tail peptides as observed in the PHD1 domains of KDM5A/B (Figure 2B).
8
Figure 2. The PHD1 domain of KDM5C preferentially binds the unmodified H3 tail and has an inhibitory role
towards nucleosome demethylation.
(A) 2D 1H-15N HSQC spectra of PHD1 titrated with increasing amounts of H3 (1-18) peptide with indicated molar
ratios (top). Backbone assignments of residues in PHD1 are labeled. Corresponding chemical shift change (Δδ) of
PHD1 residues upon binding of the H3 (1-18) tail peptide at 1:5 molar ratio (PHD:peptide) (bottom). The chemical
shift change of G364 (* denoted by asterisk) could not be determined due to broadened chemical shift when bound.
Dashed lines indicate 25th, 50th, and 75th percentile rankings, and residues are colored by a gradient from
unperturbed (yellow) to significantly perturbed (maroon). Perturbations colored by the gradient and mapped to
homologous residues in the structure of KDM5D PHD1 are in Figure S2C. (B) 2D 1H-15N HSQC of I361 in PHD1
upon titration of H3K4me0/1/2/3 (1-18) peptides (bottom) with dissociation constants determined from the chemical
shift change (Δδ) of I361 with standard error (top). Due to incomplete saturation of binding, a lower limit for the
dissociation constant is presented for the H3K4me2/3 peptides. Dissociation constants determined from chemical
shift changes of several PHD1 residues are in Figure S2H. (C) Binding of the H3 (1-18) tail peptide by PHD1 and
PHD1 D343A mutant measured by NMR titration HSQC experiments. The chemical shift change (Δδ) of I361 in
PHD1 was fit to obtain dissociation constants with standard error. Due to incomplete saturation of binding by the
D343A mutant, a lower limit for the dissociation constant is presented. (D) Demethylation kinetics of the H3K4me3
substrate nucleosome by wild type and PHD1 mutant KDM5C1-839 under single turnover conditions. Observed rates
are fit to a cooperative kinetic model, with n denoting the Hill coefficient. Wild type kinetic curve replotted from Figure
1C for comparison. Error bars represent SEM of at least three independent experiments (n ≥ 3).
1H-15N HSQC - PHD1 & H3 (1-18)
1H (ppm)
15N (ppm)
6
7
8
9
10
110
115
120
125
130
PHD: H3
1:0
1:0.25
1:0.5
1:1.25
1:1.85
1:3
1:4
1:5
D337
H350
D334
L339
G344
G333
G364
F352
E360
L354
R367
E335
M329
I351
V326
S331
S324
D336
Y349
N348
C330
D346
K370
E375
M373
C376
C345
L355
V365
K338
Q320
R332
V372
F321
K363
D347
L358A374
C353
C327
I322
Y325
K377
L341
E323
C371
I361
R378
R328
L340
C342
C368
W366
D343
A
1H (ppm)
15N (ppm)
8.7
122.5
123.0
123.5
124.0
8.7
8.7
8.7
H3K4
1H-15N HSQC - PHD1 & H3K4me (1-18)
Ile361
(1:0 to 1:5)
H3K4me1
H3K4me2
H3K4me3
B
N A Q F I E S Y V C R M C S R G D E D D K L L L C D G C D D N Y H I F C L L P P L P E I P K G V W R C P K C V M A E C K R
0.0
0.5
1.0
50%
75%
25%
(ppm)
PHD1 & H3 (1-18)
(bound at 1:5 ratio)
*
318
378
100
1000
0.0
0.1
0.2
0.3
[H3K4me (1-18) peptide] ( M)
(ppm)
PHD1 Ile361
H3
H3K4me1
Kd
130 � 7 M
H3K4me2
H3K4me3
313 � 10 M
900 � 80 M
1500 � 100 M
C
100
1000
0.0
0.1
0.2
0.3
[H3 (1-18) peptide] ( M)
(ppm)
Ile361
PHD1
PHD1 D343A
Kd
130 � 7 M
1450 � 40 M
D
1
839
JmjN ARID
PHD1 JmjC Zf
KDM5C1-839
D343A
WT
D343A
0
2
4
6
8
0.0
0.1
0.2
0.3
[KDM5C construct] ( M)
kobs (min-1)
H3K4me3 nucleosome
kmax (min-1)
Km
app ( M)
KDM5C1-839
KDM5C1-839 D343A
5.7 � 1.7
0.087 ��0.018
4.1 � 1.4
0.44 ��0.10
n
1.5
1.3
9
In order to investigate the function of PHD1 binding to the H3 tail in KDM5C catalysis,
we sought to disrupt the PHD1-H3 interaction through mutagenesis. One of the largest chemical
shift perturbations that occurs in PHD1 upon H3 tail binding is at the D343 residue, a residue
homologous to D312 in PHD1 of KDM5A where this residue is involved in H3R2 recognition
(Figure S2D) [44]. Similarly to PHD1 of KDM5A, we observe a dependence of histone tail
binding on recognition of the H3R2 residue by PHD1 of KDM5C (Figure S2E). Like the mutation
of D312 in KDM5A, the D343A mutation decreases the affinity of KDM5C PHD1 to the H3 tail at
least 10 fold (Figure 2C) [25]. When introduced into the KDM5C1-839 enzyme, the D343A
mutation does not affect the catalytic rate of H3K4me3 peptide demethylation (Figure S2F).
Surprisingly, the D343A PHD1 mutant enzyme demethylates the H3K4me3 nucleosome more
rapidly than wild type KDM5C1-839, with a ~5 fold increase of kmax (Figure 2D). No significant
change in nucleosome binding due to the D343A mutation in KDM5C1-839 was observed (Figure
S2G). This data supports an inhibitory role of the PHD1 domain in nucleosome demethylation
by KDM5C. This inhibitory role is in stark contrast to that observed for the PHD1 domain in
KDM5A, where the PHD1 domain has a stimulatory role in catalysis [25,28].
ARID domain contributes to nucleosome demethylation by KDM5C
In contrast to the inhibition of KDM5C demethylation by the PHD1 domain alone,
together the ARID and PHD1 domains provide catalytic enhancement on nucleosomes (Figure
1C). We hypothesize that this effect may be due to the ability of the ARID domain to interact
with DNA, similarly to the previously demonstrated DNA recognition by the ARID domains of
KDM5A/B [13–15]. To test this hypothesis, we interrogated binding of KDM5C1-839 towards
nucleosomes containing 20 bp flanking DNA on both ends (187 bp nucleosome). Strikingly, we
observe a 3-fold gain in affinity towards the 187 bp nucleosome compared to the core (147 bp)
nucleosome (Figure 3A), demonstrating that KDM5C is capable of recognizing flanking DNA.
10
KDM5C1-839 ∆AP has similar affinity towards both the flanking DNA-containing and core
nucleosome (Figure 3B), indicating that the ARID and PHD1 region is responsible for the
recognition of flanking DNA.
To further analyze DNA recognition, we purified the KDM5C ARID domain and
interrogated its ability to bind the flanking DNA present in the 187 bp nucleosome used in this
study. We find that the ARID domain binds the 5’ flanking DNA fragment, with a dissociation
constant of 10 µM (Figure S3A). Minimal binding was observed for the 3’ flanking DNA fragment
(Figure S3A), suggesting sequence specificity in DNA binding by ARID. We utilized NMR
spectroscopy to identify the residues of the ARID domain involved in DNA binding. Previously
determined assignments for the ARID domain were reliably transferred to a majority of
resonances observed in the 1H-15N HSQC of ARID, and modest chemical shift changes of select
ARID residues were observed upon titration of the 5’ flanking DNA fragment (Figure S3B, Figure
S3C) [45]. The perturbed residues localize to a surface on the structure of KDM5C ARID (Figure
3C), with the most notable chemical shift changes at the K101, V105, E106, R107, and N148
residues [45].
We interrogated the contributions of several identified residues, K101, R107, and N148,
towards DNA binding through mutagenesis, where we tested binding to the 147 bp 601 core
nucleosome positioning sequence (Figure 3D). We find the N148A mutation does not
significantly affect DNA binding by ARID, while the K101A and R107A mutations reduce DNA
binding by 4-6 fold (Figure 3D). A further 22-fold reduction in DNA binding was observed upon
the K101A/R107A double mutation in ARID (Figure 3D), indicating that the K101 and R107
residues are significantly involved in DNA recognition. These residues parallel those identified in
the ARID domains of KDM5A/B where the homologous residues, R112 of KDM5A and K119 &
11
R125 of KDM5B, contribute to DNA binding, suggesting conservation of DNA binding residues
in the KDM5 family [13,15].
KDM5C ARID
PDB: 2JRZ
���(ppm)
1.0
N/A
0.05
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
���(ppm)
0.1
N/A
0.01
0.015
0.03
1
839
JmjN ARID
PHD1 JmjC Zf
KDM5C1-839
K101A/R107A
E
B
A
0
1.4
2.7
5.4
10.9 µM
KDM5C1-839
187 bp unmodified
nucleosome
nuc
KDM5C
-nuc
complex
C
D
WT
K101A/R107A
F
G
0
5
0.00
0.02
0.04
0.06
[KDM5C construct] ( M)
kobs (min-1)
H3K4me3 nucleosome
kmax (min-1)
Km
app ( M)
KDM5C1-839
KDM5C1-839 K101A/R107A
5.7 � 1.7
0.087 ��0.018
4.2 � 0.5
0.029 ��0.003
n
1.5
2.7
0
1
10
0.0
0.5
1.0
[KDM5C] ( M)
Fraction unbound nuc
KDM5C1-839
Kd
app
n
�12.5 � 1.2 M
4.3 � 0.2 M
2.1
2.6
147 bp
nuc
187 bp
nuc
0
1
10
100
0.0
0.5
1.0
[KDM5C AP] ( M)
Fraction unbound nuc
KDM5C1-839 AP
Kd
app
n
�35 � 3 M
�42 � 10 M
1.4
1.2
147 bp
nuc
187 bp
nuc
0
1
10
100
0.0
0.5
1.0
[ARID construct] ( M)
Fraction unbound DNA
Kd
app
Kd relative
to WT
7.8 � 0.4 M
11.2 � 0.4 M
1.0
1.4
ARID
ARID N148A
ARID R107A
33.2 � 2.0 M
4.3
44.1 � 2.5 M
5.7
ARID K101A
ARID K101A/R107A
169 � 8 M
22
0
1
10
100
0.0
0.5
1.0
[ARID] ( M)
Fraction unbound nuc
ARID domain
Kd
app
n
31 � 3 M
22 � 2 M
1.6
1.4
147 bp
nuc
187 bp
nuc
0
1
10
0.0
0.5
1.0
[KDM5C K101A/R107A] ( M)
Fraction unbound nuc
KDM5C1-839 K101A/R107A
Kd
app
n
�15.1 � 0.9 M
6.3 � 0.3 M
3.8
2.8
147 bp
nuc
187 bp
nuc
0
2.5
5
9.9
19.8
µM
KDM5C1-839 ���
187 bp unmodified
nucleosome
nuc
12
Figure 3. DNA recognition by the ARID domain is needed for nucleosome demethylation but not nucleosome
binding by KDM5C.
(A) Binding of KDM5C1-839 to unmodified nucleosomes with and without 20 bp flanking DNA. Representative gel shift
of KDM5C binding to the 187 bp nucleosome (left). Nucleosome binding curves measured by EMSA and fit to a
cooperative binding model to determine apparent dissociation constants (Kdapp), with n denoting the Hill coefficient
(right). Due to unattainable saturation of binding, a lower limit for the dissociation constant is presented for the
unmodified core nucleosome. (B) Binding of KDM5C1-839 ∆AP to unmodified nucleosomes with and without 20 bp
flanking DNA. Representative gel shift of KDM5C ∆AP binding to the 187 bp nucleosome (left) and nucleosome
binding curves of KDM5C1-839 ∆AP (right). Due to unattainable saturation of binding, a lower limit for the dissociation
constant is presented. (C) Chemical shift changes of ARID binding to 20 bp 5’ flanking DNA colored by the gradient
and mapped to the KDM5C ARID structure (PDB: 2JRZ) of residues with backbone assignments in the 1H-15N HSQC
spectrum. Significantly perturbed residues are labeled. (D) DNA (147 bp 601 core nucleosome positioning sequence)
binding by ARID and ARID mutants. Binding curves were measured by EMSA and fit to a cooperative binding model
to determine apparent dissociation constants (Kdapp). (E) Nucleosome binding curves of the ARID domain binding to
unmodified nucleosomes with and without 20 bp flanking DNA. (F) Nucleosome binding curves of ARID mutant
KDM5C1-839 K101A/R107A binding to unmodified nucleosomes with and without 20 bp flanking DNA. (G)
Demethylation kinetics of the H3K4me3 core substrate nucleosome by wild type and ARID mutant KDM5C1-839 under
single turnover conditions. Observed rates are fit to a cooperative kinetic model, with n denoting the Hill coefficient.
Wild type kinetic curve replotted from Figure 1C for comparison. All error bars represent SEM of at least three
independent experiments (n ≥ 3).
We next interrogated DNA binding by ARID in the context of the 147 bp and 187 bp
nucleosomes. We find the ARID domain does not display a strong binding preference for the
flanking DNA-containing nucleosome and instead binds both nucleosomes with a similar weak
affinity (Figure 3E). The observed binding corresponds to a 3-4 fold reduction in affinity relative
to 147 bp non-nucleosomal DNA (Figure 3D, Figure 3E).
We then investigated the function of ARID in the context of the KDM5C enzyme towards
nucleosome binding and demethylation by introducing the K101A/R107A double mutation into
KDM5C1-839. We find that ARID mutant KDM5C1-839 retains a similar binding affinity as wild type
KDM5C1-839 towards both the flanking DNA-containing and core nucleosome (Figure 3F, Figure
3A). This indicates that the ARID domain does not contribute to nucleosome binding or to
recognition of flanking DNA by KDM5C, in contrast to our original hypothesis. However, ARID
mutant KDM5C1-839 has a reduced ability to demethylate the H3K4me3 nucleosome, with a 3-
fold reduction in kmax relative to wild type KDM5C1-839 (Figure 3G). Reduced catalysis by the
ARID mutant enzyme is only observed on the nucleosome, as the K101A/R107A double
mutation does not reduce the catalytic rate of H3K4me3 peptide demethylation (Figure S3D).
13
The similarity of catalytic rates of nucleosome demethylation between ARID mutant KDM5C1-839
and KDM5C1-839 ∆AP (0.029 min-1 and 0.032 min-1, respectively) implicates the ARID-DNA
interaction as the significant contributor in the ARID and PHD1 region towards catalysis rather
than nucleosome recognition (Figure 3G, Figure 1C).
PHD1 regulates the ability of KDM5C to recognize flanking DNA on the nucleosome
Unlike wild type (Figure 3A) and ARID mutant KDM5C (Figure 3F), KDM5C1-839 ∆AP has
reduced nucleosome binding and a loss in the ability to discriminate between the 147 bp and
187 bp nucleosome (Figure 3B). To better understand elements of KDM5C that contribute to its
ability to bind DNA in the context of the nucleosome, we focused on the linker region between
ARID and PHD1. The ARID-PHD1 linker region of KDM5C is the longest among KDM5 family
members and contains many basic residues (Figure S4A). This linker region also has low
conservation in the KDM5 family and is predicted to be disordered in KDM5C (Figure S4A,
Figure S4B). We generated a construct where the linker region (residues 176 to 317) is
replaced by a short (GGS)5 linker sequence (KDM5C1-839 ∆linker) (Figure 4A). KDM5C1-839
∆linker possesses similar catalytic efficiencies as wild type KDM5C1-839 on both the H3K4me3
nucleosome and H3K4me3 peptide substrate (Figure S4C, Figure S4D), indicating that the
enzyme without the ARID-PHD1 linker is functionally active. We then assessed binding of
KDM5C1-839 ∆linker to the 147 bp and 187 bp nucleosome and surprisingly did not detect any
nucleosome binding (Figure 4A), suggesting that the linker region affects nucleosome and
flanking DNA recognition by KDM5C.
We next interrogated recognition of flanking DNA on the nucleosome in the presence of
the H3K4me3 substrate, as recognition of both could facilitate recruitment of KDM5C to its
14
target sites in euchromatin [29]. Intriguingly, KDM5C1-839 has similar binding affinity for both the
core and flanking DNA-containing H3K4me3 nucleosome, with Kdapp of ~7 µM, indicating no
engagement of flanking DNA in the presence of the H3K4me3 substrate (Figure 4B). This is in
contrast to unmodified nucleosome binding, where KDM5C has a clear preference for
nucleosomes with flanking DNA (Figure 4B).
A
E
C
ZnF
JmjC
JmjN
PHD1
ARID
X
PHD1
JmjC
ARID
JmjN
ZnF
PHD1 mutation
WT
restriction of
DNA binding
by PHD1
D343A
abolished
inhibition
H3K4me3
D
KDM5C1-839 �������
����������
1
839
JmjN ARID PHD1 JmjC Zf
147 bp
nuc
187 bp
nuc
0
1
10
100
0.0
0.5
1.0
[KDM5C linker] ( M)
Fraction unbound nuc
KDM5C1-839 linker
unmod nuc H3K4me3 nuc
187 bp
147 bp
187 bp
147 bp
5
10
15
Kd
apparent ( M)
KDM5C1-839
12.5
4.3
7.3
6.9
unmod nuc H3K4me3 nuc
187 bp
147 bp
187 bp
147 bp
0
5
10
Kd
apparent ( M)
KDM5C1-839 D343A
8.4
3.2
4.7
8.9
������a����
�������a����
pep����
kobs1 (s
-1�
kobs2 (s
-1�
kobs1 (s
-1�
kobs2 (s
-1�
H3 (1-���
3.16 ± 0.09
0.022 ± 0.001
0.69 ± 0.01
0.018 ± 0.001
KDM5C (199-����
2.77 ± 0.10
0.015 ± 0.002
1.45 ± 0.04
0.020 ± 0.003
KDM5C (295-����
1.88 ± 0.13
0.013 ± 0.001
2.20 ± 0.15
0.040 ± 0.004
KDM5C (239-����
2.68 ± 0.45
0.012 ± 0.001
3.23 ± 0.59
0.015 ± 0.003
100
1000
0.0
0.2
0.4
0.6
0.8
[peptide] ( M)
(ppm)
PHD1 Asp343
H3 (1-18)
KDM5C (199-218)
Kd
140 � 10 M
440 � 20 M
B
F
0
60
120
180
240
300
360
0.0
0.2
0.4
0.6
Time (sec)
Response (nm)
PHD1 & KDM5C linker peptides
H3 (1-20)
KDM5C (199-218)
KDM5C (295-314)
H3 (1-20)
KDM5C (175-194)
KDM5C (183-202)
KDM5C (191-210)
KDM5C (199-218)
KDM5C (207-226)
KDM5C (215-234)
KDM5C (223-242)
KDM5C (231-250)
KDM5C (239-258)
KDM5C (247-266)
KDM5C (255-274)
KDM5C (263-282)
KDM5C (271-290)
KDM5C (279-298)
KDM5C (287-306)
KDM5C (295-314)
15
Figure 4. KDM5C recognizes flanking DNA in the absence of H3K4me3 due to regulation by PHD1.
(A) Binding of KDM5C1-839 ∆linker to unmodified nucleosomes with and without 20 bp flanking DNA. Nucleosome
binding curves were measured by EMSA. (B) Nucleosome binding by KDM5C1-839 with apparent dissociation
constants (Kdapp) measured by EMSA and fit to a cooperative binding model (substrate nucleosome binding curves in
Figure S4H). Select dissociation constants replotted from Figure 1B and Figure 3A for comparison. Due to
unattainable saturation of binding, a lower limit for the dissociation constant is presented for the unmodified core
nucleosome. (C) Nucleosome binding by PHD1 mutant KDM5C1-839 D343A with apparent dissociation constants
(Kdapp) measured by EMSA (binding curves in Figure S4I). (D) Model for KDM5C inhibition, where PHD1 prevents
flanking DNA recognition in the presence of H3K4me3, and its relief by the PHD1 mutation that disrupts the inhibition.
(E) Binding kinetic trace of immobilized Avitag-PHD1 binding to H3 (1-20) tail peptide and KDM5C ARID-PHD1 linker
fragment 20-mer peptides measured by bio-layer interferometry. Observed rates (kobs) of association and dissociation
by peptides with detectable binding were obtained from fitting kinetic traces to a two phase exponential function.
KDM5C linker fragment peptides have acetylated N-termini and amidated C-termini. Identified PHD1-binding KDM5C
peptide
sequences
are
KDM5C
(199-218):
QSVQPSKFNSYGRRAKRLQP
and
KDM5C
(295-314):
KEELSHSPEPCTKMTMRLRR. (F) Binding of the H3 (1-18) tail peptide and KDM5C (199-218) peptide by PHD1
measured by NMR titration HSQC experiments. The chemical shift change (Δδ) of D343 in PHD1 was fit to obtain
dissociation constants with standard error. All error bars represent SEM of at least three independent experiments
(n ≥ 3).
Since KDM5C recognizes flanking DNA only in the context of the unmodified
nucleosome, we considered the possibility that the ability to engage flanking DNA is coupled to
binding of the H3 tail product to the PHD1 domain. To test this model, we interrogated the effect
of the PHD1 D343A mutation, which abrogates H3 binding, on the recognition of flanking DNA
by KDM5C. We find that PHD1 mutant KDM5C1-839 D343A still retains the ~3-fold affinity gain
towards the unmodified 187 bp nucleosome (Kdapp = 3.2 µM) compared to the unmodified core
nucleosome (Kdapp = 8.4 µM) (Figure 4C). In addition, PHD1 mutant KDM5C displays a ~2 fold
affinity gain towards the 187 bp H3K4me3 nucleosome (Kdapp = 4.7 µM), relative to the
H3K4me3 core nucleosome (Kdapp = 8.9 µM) (Figure 4C). Although modest, this improved
binding demonstrates that, unlike wild type KDM5C, PHD1 mutant KDM5C can recognize
flanking DNA in the presence of the H3K4me3 substrate. This observation lead us to
hypothesize that, beyond disruption of H3 tail binding, the D343A mutation may also disrupt
intramolecular interactions within the demethylase which restrict the ability of ARID and the
ARID-PHD1 linker to interact with DNA (Figure 4D). This PHD1-imposed inhibition model is
consistent with the strong catalytic enhancement observed with the PHD1 mutant demethylase
under single turnover conditions (Figure 2D).
16
To further test this model, we examined whether PHD1 is capable of engaging in
intramolecular interactions within KDM5C, which could impede its ability to interact with the H3
tail. An intramolecular interaction would necessitate that PHD1 is able to interact with ligands
that do not have a free N-terminus, in contrast to typical PHD-H3 interactions [22]. We first
tested whether a free N-terminus is required for H3 tail recognition by PHD1 and find that N-
terminal acetylation of the H3 tail peptide slightly reduces but does not abrogate binding by
PHD1 (Figure S4E), indicating permissibility for recognition of an internal protein sequence. No
interaction was detected between PHD1 and the ARID domain (Figure S4F). Using tiled
peptides, we then tested binding to peptide fragments of the ARID-PHD1 linker region, each
consisting of 20 amino acids with an acetylated N-terminus and amidated C-terminus. PHD1
exhibits most notable binding to the fragment of the ARID-PHD1 linker spanning residues 199-
218, which contains a polybasic segment reminiscent of H3 (Figure 4E). Using NMR, titration of
PHD1 with the KDM5C (199-218) peptide engages a subset of residues that participate in H3
tail binding, including D343 (Figure S4G, Figure 2A). The affinity of PHD1 towards KDM5C
(199-218) is ~3-fold lower than that of the H3 tail (Figure 4F). These findings indicate PHD1
could interact with the ARID-PHD1 linker within KDM5C, an interaction that can be outcompeted
by its H3 tail ligand.
X-linked intellectual disability mutations alter nucleosome recognition and demethylation
by KDM5C
Our proposed regulatory model provides a mechanistic framework for querying the
effects of mutations in KDM5C that cause XLID (Figure 5A). Specifically, we sought to
investigate the D87G and A388P mutations found at the beginning of ARID and immediately
downstream of PHD1, respectively. The D87G mutation, associated with mild intellectual
disability, has been demonstrated to have no effect on global H3K4me3 levels in vivo [42]. The
17
A388P mutation, associated with moderate intellectual disability, has also been shown to have
no effect on global H3K4me3 levels in vivo but has been reported to reduce demethylase
activity in vitro [2,46]. We initially interrogated nucleosome binding by KDM5C1-839 D87G and
A388P. Strikingly, relative to wild type KDM5C1-839, we observe 4-7 fold enhanced binding of the
XLID mutants to the unmodified core nucleosome (Figure 5B), suggesting that these mutations
enable enhanced nucleosome engagement. The ARID and PHD1 region is required for this
enhanced nucleosome binding, as there is no gain in nucleosome affinity due to the A388P
mutation when the ARID and PHD1 region is removed (Figure S5A). Importantly, the gain in
nucleosome affinity of the XLID mutants is more prominent on the unmodified core nucleosome
than the substrate H3K4me3 core nucleosome, resulting in loss of binding specificity towards
H3K4me3 by KDM5C due to the D87G and A388P mutations (Figure 5C).
As the XLID mutations cause an overall affinity gain towards both unmodified and
substrate nucleosomes, we reasoned that the recognition of the shared common epitope of
DNA, rather than the H3 tail, is altered in the mutants. Indeed, relative to wild type KDM5C1-839,
we observe a similar 3-5 fold gain in affinity by the XLID mutants towards the 187 bp unmodified
nucleosome with flanking DNA, with both D87G and A388P mutants converging to a high
nucleosome affinity of Kdapp ~1 µM (Figure 5D). As flanking DNA recognition by KDM5C appears
to be regulated by PHD1 (Figure 4C), we further interrogated recognition of the 187 bp
substrate nucleosome by the D87G and A388P mutants. Both KDM5C1-839 D87G and A388P
are capable of recognizing flanking DNA in the presence of H3K4me3, with a ~2 fold gain in
affinity towards the 187 bp H3K4me3 nucleosome over the H3K4me3 core nucleosome (Figure
5E). These findings suggest that, similarly to the D343A PHD1 mutation (Figure 4C), the XLID
mutations may disrupt the PHD1-mediated inhibition of DNA binding. Our findings are consistent
with the model that these XLID mutations are altering the ARID and PHD1 region to relieve the
inhibition of DNA binding, enabling unregulated binding to the nucleosome.
18
Figure 5. X-linked intellectual disability mutations enhance nucleosome binding by KDM5C and reduce
demethylase activity in the presence of flanking DNA.
(A) XLID mutations found in KDM5C (top) and the XLID mutations investigated in this study (bottom). (B) Unmodified
core nucleosome binding by KDM5C1-839 wild type (WT), D87G, and A388P. Nucleosome binding was measured by
EMSA and curves fit to a cooperative binding model to determine apparent dissociation constants (Kdapp), with n
denoting the Hill coefficient. WT binding curve replotted from Figure 3A for comparison. Due to unattainable
saturation of binding, a lower limit for the dissociation constant is presented for WT KDM5C binding the unmodified
nucleosome. (C) Apparent dissociation constants (Kdapp) of binding by KDM5C1-839 WT, D87G, and A388P to
unmodified and substrate core nucleosomes and resulting H3K4me3 fold binding specificity. Select dissociation
A
B
C
D
F
E
G
1
839
JmjN ARID
PHD1 JmjC Zf
KDM5C1-839
D87G
A388P
1
1560
JmjN ARID
PHD1 JmjC Zf
PHD2
KDM5C
M1T
R68fsA77T
D87G
R332*
E468fs
R694*
C724*
V1075Yfs
K1087fs
W1288*
R1481fs
A388P
D402Y
S451R
P480L
V504M
P556T
F642L
E698K
L731F
R750W
Y751C
R766W
R1115H
WT
D87G
A388P
D87G
WT
A388P
0
0.1
1
10
0.0
0.5
1.0
[KDM5C construct] ( M)
Fraction unbound nuc
147 bp unmodified nucleosome
WT
D87G
Kd
app
n
�12.5 � 1.2 M
1.8 � 0.03 M
3.1 � 0.2 M
2.1
2.0
2.5
A388P
0
0.1
1
10
0.0
0.5
1.0
[KDM5C construct] ( M)
Fraction unbound nuc
187 bp unmodified nucleosome
WT
D87G
Kd
app
n
4.3 � 0.2 M
0.9 � 0.1 M
1.3 � 0.04 M
2.6
2.0
2.5
A388P
WT
A388P
D87G
0
5
10
Kd
apparent ( M)
6.9
7.3
1.8
3.7
1.3
2.3
187 bp H3K4me3 nuc
147 bp H3K4me3 nuc
Kd
app ( M)
12.5 � 1.2
6.9 ��0.8
3.1 � 0.2
3.7 ��0.6
H3K4me3
specificity
unmod nuc
H3K4me3 nuc
WT
D87G
1.8 � 0.03
2.3 � 0.2
1.8
0.8
0.8
A388P
0
2
4
6
8
0.00
0.02
0.04
0.06
[KDM5C construct] ( M)
kobs (min-1)
147 bp H3K4me3 nucleosome
kmax (min-1)
Km
app ( M)
WT
D87G
5.7 � 1.7
0.087 ��0.018
1.8 � 0.3
0.071 ��0.007
n
1.5
1.5
4.3 � 0.7
0.013 ��0.002
1.7
A388P
0
2
4
6
8
0.00
0.02
0.04
0.06
[KDM5C construct] ( M)
kobs (min-1)
187 bp H3K4me3 nucleosome
kmax (min-1)
Km
app ( M)
WT
D87G
9.6 � 2.3
0.092 ��0.015
1.6 � 0.3
0.033 ��0.003
n
1.3
1.9
3.6 � 3.6
0.003 ��0.001
1.0
A388P
19
constants are from Figure 1B and Figure 5B for comparison. (D) Binding curves of KDM5C1-839 WT, D87G, and
A388P binding to the unmodified 187 bp nucleosome with 20 bp flanking DNA. WT binding curve replotted from
Figure 3A for comparison. (E) Binding of KDM5C1-839 WT, D87G, and A388P to substrate nucleosomes with and
without 20 bp flanking DNA with apparent dissociation constants (Kdapp) measured by EMSA (binding curves in
Figure S5C). Select dissociation constants are replotted from Figure 4B and Figure 5C for comparison. (F)
Demethylation kinetics of the core substrate nucleosome by KDM5C1-839 WT, D87G, and A388P under single turnover
conditions. Observed rates are fit to a cooperative kinetic model, with n denoting the Hill coefficient. Wild type kinetic
curve replotted from Figure 1C for comparison. (G) Demethylation kinetics of the 187 bp substrate nucleosome by
KDM5C1-839 WT, D87G, and A388P under single turnover conditions. All error bars represent SEM of at least three
independent experiments (n ≥ 3).
We next measured the demethylase activity of KDM5C1-839 D87G and A388P towards
the H3K4me3 core nucleosome substrate. Despite these XLID mutants sharing similar
enhanced nucleosome binding, their effects on nucleosome demethylation differ. The A388P
mutation impairs KDM5C catalysis (kmax) by ~7 fold, while the D87G mutation increases catalytic
efficiency (kmax/Kmapp) ~3 fold through an enhanced Kmapp, indicating both nonproductive and
productive KDM5C states caused by these mutations (Figure 5F). The reduced demethylase
activity caused by the A388P mutation is consistent with previous findings of reduced in vitro
demethylation, with the 7-fold reduction we observe on nucleosomes exceeding the previously
reported 2-fold reduction on substrate peptide [2]. The reduced demethylase activity due to the
A388P mutation might be caused by impairment of the composite catalytic domain, as we
observe reduced demethylase activity in A388P mutant KDM5C1-839 ∆AP (Figure S5B). In
contrast, the D87G mutation does not appear to affect the catalytic domain, and instead the
improved catalytic efficiency reflects the enhancement in nucleosome binding.
Unlike wild type KDM5C, these XLID mutants recognize flanking DNA in the presence of
H3K4me3, prompting us to measure demethylase activity on the 187 bp H3K4me3 nucleosome.
Interestingly, while catalysis by the wild type enzyme is only slightly reduced, we find that
addition of flanking DNA to the substrate nucleosome results in strong inhibition of catalysis by
KDM5C1-839 A388P, with a 5-fold reduction in kmax relative to the core substrate nucleosome
(Figure 5G). Addition of flanking DNA also reduces catalysis by KDM5C1-839 D87G, although to a
20
lesser degree of ~2 fold (Figure 5G). Despite lower maximal catalysis (kmax) of the D87G mutant
relative to wild type KDM5C1-839 in the presence of flanking DNA, the D87G mutant is still ~2 fold
more efficient (kmax/Kmapp) due to its enhanced nucleosome binding. Regardless, enhanced DNA
recognition caused by the XLID mutations results in a reduction in the catalytic rate of H3K4me3
demethylation of nucleosomes with flanking DNA compared to core nucleosomes.
DISCUSSION
Different reader and regulatory domains within chromatin binding proteins and modifying
enzymes influence their activity and substrate specificity by recognizing distinct chromatin states
through distinguishing features on the nucleosome and surrounding DNA. Emerging structural
studies of chromatin modifying enzymes in complex with nucleosomes have highlighted these
multivalent interactions, with increasing observations of interactions with DNA contributing to
nucleosome engagement by histone modifying enzymes [47–59]. Despite the unique insertion
of the ARID and PHD1 reader domains in the composite catalytic domain, the function of
accessory domains within the KDM5 demethylase family has not been explored on
nucleosomes. Here, we describe a hierarchy of regulation by these domains by investigating
nucleosome recognition and demethylation in KDM5C, a unique member of the KDM5 family
involved in regulation of neuronal gene transcription. We find that there are opposing roles of
the ARID and PHD1 domains, with DNA recognition by ARID providing a beneficial interaction
for nucleosome demethylation and regulation by PHD1 inhibiting nucleosome recognition and
demethylation. We further demonstrate that DNA recognition is regulated by the PHD1 domain
through its interaction with the ARID-PHD1 linker, allowing for sensing of the H3K4me3
substrate. These regulatory interactions are disrupted by the D87G and A388P XLID mutations
adjacent to the ARID and PHD1 domains, resulting in enhanced DNA binding and loss of
H3K4me3 specificity. As enhanced flanking DNA recognition by XLID mutants is detrimental to
21
demethylase activity, our findings suggest dysregulation of KDM5C demethylation at
euchromatic loci, where this enzyme predominantly functions [29,32].
Our findings of KDM5C nucleosome recognition and demethylation can be best
explained by a regulatory model where PHD1 controls DNA recognition (Figure 6A). In the
ground state, binding of the ARID-PHD1 linker to PHD1 restricts the ability of the enzyme to
interact with DNA, attenuating catalysis (state I). Release of the PHD1-imposed constraint on
the ARID-PHD1 linker and ARID domain enables improved interaction with DNA, leading to
faster catalysis (state II). In our experiments, the D343A PHD1 mutation was used as a
mechanistic probe to release the PHD1-imposed restriction on DNA binding. In the context of
chromatin, this release of inhibition could be achieved through binding of the H3 tail to PHD1,
allowing for the regulation of demethylation by the surrounding chromatin environment.
Formation of the demethylated H3 product, and its binding to PHD1, further reinforces an
interaction of KDM5C with chromatin by enabling flanking DNA recognition, possibly through the
ARID-PHD1 linker region (state III). Alternatively, the PHD1 domain could act directly on the
catalytic domains to impair productive substrate nucleosome engagement.
ZnF
JmjC
JmjN
PHD1
ARID
H3K4me3 substrate
recognition
inhibited ground state
H3 product & flanking
DNA recognition
PHD1
JmjC
ARID
JmjN
ZnF
ZnF
JmjC
JmjN
PHD1
ARID
enhanced demethylation
I
II
III
catalytically active state
product bound state
ZnF
JmjC
JmjN
PHD1
ARID
XLID
altered
conformational state
H3 tail
A
B
22
Figure 6. Model of KDM5C regulation by the ARID-linker-PHD1 region and KDM5C dysregulation by XLID
mutations.
(A) KDM5C recognizes H3K4me3 and binds to substrate nucleosomes through the catalytic domain (pre-catalytic
and inhibited ground state I). DNA binding in the presence of H3K4me3 is attenuated due to an inhibitory role of
PHD1 on DNA recognition. During demethylation, ARID makes transient interactions with nucleosomal DNA to orient
the catalytic domain towards the H3K4me3 tail for efficient demethylation. H3 tail binding to PHD1 releases the PHD1
interaction constraining the ARID-PHD1 linker and ARID domain, enabling ARID interactions with DNA to further
enhance demethylation (catalytically active state II). After demethylation, binding of the product H3 tail to PHD1
enables flanking DNA binding by the ARID-PHD1 linker region (post-catalytic and product bound state III). (B)
Proposed altered conformational state of the ARID and PHD1 region in KDM5C due to XLID mutations in this region
disrupting hypothesized intramolecular interactions.
Intriguingly, we observe cooperativity (Hill coefficients > 1) in nucleosome binding and
demethylation (Figure 1C, Figure S1B). In addition, cooperativity occurs in peptide
demethylation by wild type KDM5C1-839 but not by KDM5C1-839 ∆AP under single turnover
conditions (Figure S1A), suggesting that cooperativity might arise both from the state of KDM5C
and from the nucleosome possessing two H3 tails.
Our finding of the beneficial role of the ARID domain towards KDM5C catalysis on
nucleosomes can be rationalized by favorable transient interactions of the ARID domain with
nucleosomal DNA to better orient the catalytic domains for demethylation and could make the
substrate H3K4me3 more accessible through disrupting histone tail-DNA interactions [60–63].
This is supported by the previous observation that the ARID domain of KDM5C is required for its
demethylase activity in vivo but not for its chromatin association [11]. This role of the ARID
domain in productive nucleosome demethylation may be conserved within the KDM5 family, as
the ARID domain has also been found to be required for in vivo demethylation by KDM5A/B and
the Drosophila KDM5 homolog Lid [4,10,12,13]. The ARID domain may be required for
nucleosome demethylation in order to displace the H3K4me3 tail from interacting with DNA,
making it accessible for engagement by the catalytic domain. This histone tail displacement
function has been proposed for DNA binding reader domain modules and for the LSD1/CoREST
complex, where the SANT2 domain interacts with nucleosomal DNA to displace the H3 tail for
engagement by the LSD1 active site [55,63–65].
23
In contrast to the beneficial role of the ARID domain, we observe an unexpected
inhibitory role of PHD1 towards KDM5C demethylation on nucleosomes. This finding suggests
differential regulation by PHD1 in the KDM5 family, as PHD1 binding has a stimulatory role
towards in vitro demethylation in KDM5A/B, and PHD1 has been previously shown to be
required for demethylase activity in vivo for KDM5B and Lid [4,10,25,26,28]. Our data suggests
this inhibitory role is mediated by the ability of PHD1 to inhibit KDM5C’s engagement of DNA on
the nucleosome (Figure 6A). With weak affinity and indifference for a free N-terminus (Figure
S4E), ligand recognition by PHD1 in KDM5C is strikingly different from that observed for the
PHD1 domains in KDM5A and KDM5B. While further work is needed to identify how PHD1
restricts DNA binding, our findings indicate that this could be achieved through an interaction
between PHD1 and the unstructured ARID-PHD1 linker region, possibly mediated by the basic
residues within the linker. This unique ARID-PHD1 linker (Figure S4A) may contribute to distinct
regulation by PHD1 in KDM5C. Although we are unable to directly test the effect of H3 tail
binding to PHD1 on DNA recognition due to the low affinity regime, we hypothesize that the
resulting binding could release inhibition, allowing for the regulation of KDM5C activity by
different chromatin environments. As a consequence, H3 tail binding by PHD1 might stimulate
demethylation, as observed upon PHD1 binding in KDM5A/B, through a mechanistically distinct
relief of negative regulation in KDM5C (Figure 6A).
Unlike the ARID domain, whose DNA recognition is needed for nucleosome
demethylation but not nucleosome binding, the ARID-PHD1 linker region contributes towards
nucleosome binding but does not appear to contribute to demethylation by KDM5C.
Surprisingly, we observe diminished nucleosome binding upon deletion of the ARID-PHD1 linker
as opposed to a ~3-fold decrease in nucleosome binding upon deletion of the entire ARID and
PHD1 region. While the molecular basis for these effects requires further studies, this observed
discrepancy could result from the ARID and PHD1 domains affecting nucleosome binding by the
24
catalytic and zinc finger domains of KDM5C. Our findings add to the reports of intrinsically
disordered regions as functional elements within chromatin binding proteins [66–69].
Unexpectedly, KDM5C recognizes flanking DNA around the nucleosome in the presence
of the unmodified H3 tail but not in the presence of the H3K4me3 substrate. Linker DNA
recognition may serve to retain KDM5C at its target promoter and enhancer sites within open
chromatin after demethylation. It may also enable processive demethylation of adjacent
nucleosomes in euchromatin by KDM5C. Interestingly, the recognition of linker DNA has been
observed in the mechanistically unrelated H3K4me1/2 histone demethylase LSD1/KDM1A,
where demethylase activity is in contrast stimulated by linker DNA [47,70]. The H3K36me1/2
demethylase KDM2A is also capable of recognizing linker DNA, where it is specifically recruited
to unmethylated CpG islands at gene promoters through its ZF-CxxC domain [71,72]. These
findings suggest that recognition of the chromatin state with accessible linker DNA may be
utilized by histone modifying enzymes that function on euchromatin. While the sequence
specificity of linker DNA recognition requires further investigation, it is evident that the sensing
of the H3K4me3 substrate tail by KDM5C is preferred over recognition of linker DNA, a feature
accessible in open chromatin. This observed hierarchy, coupled with KDM5C’s overall weak
affinity towards nucleosomes and dampened demethylase activity due to regulation by PHD1,
suggests tunable demethylation by KDM5C. Thus, this multi-domain regulation might serve to
establish H3K4me3 surveillance through KDM5C-catalyzed demethylation, which is well suited
for the physiological role of this enzyme in fine tuning gene expression through H3K4me3
demethylation at enhancers and promoters of genes, as well as its role in genome surveillance
by preventing activation of non-neuronal genes in adult neurons [29,32].
Our findings show that the regulation of DNA recognition by KDM5C is disrupted by the
D87G and A388P XLID mutations adjacent to the ARID and PHD1 domains, such that
25
nucleosome binding is significantly enhanced, H3K4me3 specificity is lost, and demethylase
activity is sensitized to inhibition by linker DNA. The location of these mutations lends support to
our model, where the XLID mutations relieve the inhibition of DNA recognition, enabling
enhanced nucleosome binding irrespective of the methylation status of the nucleosome, by
altering the conformational state of the ARID and PHD1 region (Figure 6B). Beyond disruption
of histone demethylase activity, our findings suggest an additional mechanism of dysregulation
of KDM5C in XLID, that of enhanced nonproductive chromatin engagement and differential
dysregulation of demethylation at different loci depending on the accessibility of linker DNA.
Despite the reduced in vitro activity of KDM5C due to the A388P mutation, global H3K4me3
levels are unaffected with human KDM5C A388P in vivo [2,46]. In contrast, increased global
H3K4me3 levels have been observed in a Drosophila intellectual disability model with A512P
mutant Lid, signifying that further work is needed to profile H3K4me3 levels at genomic target
regions affected by XLID mutations in human KDM5C [73]. Furthermore, we observe that the
demethylase activity of KDM5C D87G varies relative to wild type depending on the presence of
flanking DNA, which might account for the unaffected global H3K4me3 levels previously
observed with this D87G mutation [42]. Our findings suggest that the chromatin environment, in
particular the presence of accessible linker DNA, could govern altered demethylation and
nonproductive chromatin recognition by KDM5C in XLID. Euchromatin-specific dysregulation of
KDM5C demethylation might account for the hard-to-reconcile discrepancies between reported
in vitro demethylase activities of KDM5C XLID mutants and their effect on global H3K4me3
levels.
26
MATERIALS AND METHODS
Generation of KDM5C constructs
Human KDM5C gene was obtained from Harvard PlasmID (HsCD00337804) and Q175 was
inserted to obtain the canonical isoform (NP_004178.2). KDM5C residues 1 to 839 were cloned
into a pET28b His-Smt3 vector to produce 6xHis-SUMO-KDM5C and was mutated by site-
directed mutagenesis for point mutants. The KDM5C1-839 ∆AP construct was cloned by replacing
residues 83-378 with a 4xGly linker. The KDM5C1-839 ∆linker construct was cloned by replacing
residues 176-317 with a (GGS)5 linker.
Purification of KDM5C constructs
Recombinant His-tagged SUMO-KDM5C constructs were expressed in BL21(DE3) E. coli in LB
media containing 50 µM ZnCl2 and 100 µM FeCl3 through induction at OD600 ~0.6 using 100 µM
IPTG followed by expression at 18 ºC overnight. Collected cells were resuspended in 50 mM
HEPES pH 8, 500 mM KCl, 1 mM BME, 5 mM imidazole, and 1 mM PMSF, supplemented with
EDTA-free Pierce protease inhibitor tablets (Thermo Fisher Scientific) and benzonase, and
lysed by microfluidizer. Lysate was clarified with ultracentrifugation and the recovered
supernatant was then purified by TALON metal affinity resin (total contact time under 2 hrs) at 4
ºC. The His-SUMO tag was then cleaved by SenP1 during overnight dialysis at 4 ºC in 50 mM
HEPES pH 7.5, 150 mM KCl, and 5 mM BME. KDM5C constructs were then purified by anion
exchange (MonoQ, GE Healthcare) and subsequent size exclusion (Superdex 200, GE
Healthcare) chromatography in 50 mM HEPES pH 7.5 and 150 mM KCl. Fractions were
concentrated and aliquots snap frozen in liquid nitrogen for storage at -80 ºC.
27
Nucleosomes and DNA
Recombinant human 5’ biotinylated unmodified 147 bp mononucleosomes (16-0006),
unmodified 187 bp mononucleosomes (16-2104), 5’ biotinylated H3K4me3 147 bp
mononucleosomes (16-0316), and 5’ biotinylated H3K4me3 187 bp mononucleosomes (16-
2316) were purchased from Epicypher, Inc., in addition to biotinylated 147 bp 601 sequence
DNA
(18-005).
187
bp
nucleosomes
contain
the
20
bp
sequences
5’
GGACCCTATACGCGGCCGCC and GCCGGTCGCGAACAGCGACC 3’ flanking the core 601
positioning sequence. 20 bp flanking DNA duplex fragments were synthesized by Integrated
DNA Technologies, Inc. For use in binding and kinetic assays, stock nucleosomes were buffer
exchanged into corresponding assay buffer using a Zeba micro spin desalting column (Thermo
Scientific).
Nucleosome and DNA binding assays
Nucleosome and DNA binding was assessed by EMSA. 100 nM nucleosomes (0.5 pmol) and
various concentrations of KDM5C were incubated in binding buffer (50 mM HEPES pH 7.5, 50
mM KCl, 1mM BME, 0.01% Tween-20, 0.01% BSA, 5% sucrose) for 1 hr on ice prior to analysis
by native 7.5% PAGE. For DNA binding, 100 nM 147 bp 601 sequence DNA or 500 nM 20 bp
linker DNA fragments were incubated with various concentrations of ARID. Samples were
separated using pre-run gels by electrophoresis in 1xTris-Glycine buffer at 100V for 2 hrs at 4
ºC, stained using SYBR Gold for DNA visualization, and imaged using the ChemiDoc imaging
system (Bio-Rad Laboratories). Bands were quantified using Bio-Rad Image Lab software to
determine the fraction of unbound nucleosome to calculate apparent dissociation constants by
fitting to the cooperative binding equation Y=(X^n)/(Kd^n + X^n), where X is the concentration of
KDM5C, n is the Hill coefficient, and Kd is the concentration of KDM5C at which nucleosomes
are half bound.
28
Single turnover nucleosome demethylation kinetics
The demethylation of biotinylated H3K4me3 nucleosome was monitored under single turnover
conditions (>10 fold excess of KDM5C over substrate) through the detection of H3K4me1/2
product nucleosome formation over time by TR-FRET of an anti-H3K4me1/2 donor with an anti-
biotin acceptor reagent. Various concentrations of KDM5C were reacted with 25 nM 5’
biotinylated H3K4me3 nucleosome in 50 mM HEPES pH 7.5, 50 mM KCl, 0.01% Tween-20,
0.01% BSA, 50 µM alpha-ketoglutarate, 50 µM ammonium iron(II) sulfate, and 500 µM ascorbic
acid at room temperature. 5 µL time points were taken and quenched with 1.33 mM EDTA then
brought to 20 µL final volume for detection using 1 nM LANCE Ultra Europium anti-H3K4me1/2
antibody (TRF0402, PerkinElmer) and 50 nM LANCE Ultra Ulight-Streptavidin (TRF0102,
PerkinElmer) in 0.5X LANCE detection buffer. Detection reagents were incubated with reaction
time points for 2 hours at room temperature in 384 well white microplates (PerkinElmer
OptiPlate-384) then TR-FRET emission at 665 nm and 615 nm by 320 nm excitation with 50 µs
delay and 100 µs integration time was measured using a Molecular Devices SpectraMax M5e
plate reader. TR-FRET was calculated as the 665/615 nm emission ratio and kinetic curves
were fit to a single exponential function to determine kobs of demethylation. kobs parameters
were then plotted as a function of KDM5C concentration and fit to the sigmoidal kinetic equation
Y=kmax*X^n/(Khalf^n + X^n) using GraphPad Prism to determine kmax and Kmapp parameters of
demethylation.
Purification of PHD1 for NMR
PHD1 (KDM5C residues 318-378) was cloned into a pET28b His-Smt3 vector to express
recombinant 6xHis-SUMO-PHD1 in BL21(DE3) E. coli in metal supplemented M9 minimal
medium containing 15NH4Cl (Cambridge Isotope Laboratories). 13C-glucose (Cambridge Isotope
Laboratories) was used in medium for expression of 15N, 13C-labeled PHD1. Expression was
induced at OD600 ~0.6 using 1 mM IPTG for expression at 18 ºC overnight. Collected cells were
29
resuspended in 50 mM HEPES pH 8, 500 mM KCl, 5 mM BME, 10 mM imidazole, and 1 mM
PMSF, supplemented with benzonase, and lysed by sonication. Lysate was clarified with
ultracentrifugation and the recovered supernatant was then purified by Ni-NTA affinity resin. The
His-SUMO tag was then cleaved by SenP1 during overnight dialysis at 4 ºC in 50 mM HEPES
pH 7.5, 150 mM KCl, 50 µM ZnCl2 and 10 mM BME. Cleaved His-SUMO tag and SenP1 was
captured by passing through Ni-NTA affinity resin and flow-through was then purified by anion
exchange (MonoQ) chromatography in starting buffer of 50 mM HEPES pH 7.5, 150 mM KCl,
50 µM ZnCl2 and 10 mM BME. Flow-through MonoQ fractions containing PHD1 were
concentrated and aliquots snap frozen in liquid nitrogen for storage at -80 ºC.
PHD1 NMR and histone peptide NMR titrations
For backbone assignment of KDM5C PHD1, 400 µM 15N, 13C-labeled PHD1 in 50 mM HEPES
pH 7.5, 50 mM KCl, 5 mM BME, 50 µM ZnCl2, and 5% D2O was used to perform 3D triple-
resonance CBCA(CO)NH and CBCANH experiments at 298K using a 500 MHz Bruker
spectrometer equipped with a cryoprobe. Triple-resonance experiments were also performed
using 400 µM 15N, 13C-labeled PHD1 bound to 2 mM H3 (1-18) peptide (1:5 ratio) to assign
broadened backbone residues in apo spectra. 3D spectra were processed using NMRPipe then
analyzed and assigned using CcpNMR Analysis. Out of 56 assignable residues, 54 in apo
PHD1 and 53 residues in H3 bound PHD1 were assigned.
For 2D 1H-15N HSQC spectra of KDM5C PHD1, 200 µM 15N-labeled PHD1 in 50 mM HEPES pH
7.5, 50 mM KCl, 5 mM BME, 50 µM ZnCl2, and 5% D2O was used to obtain 2D spectra at 298K
using a 800 MHz Bruker spectrometer equipped with a cryoprobe. Chemical shift perturbation
experiments were performed by obtaining HSQC spectra with increasing concentrations of
histone tail peptides (GenScript) up to 1:5 molar ratio of PHD1:peptide. Data were processed
using Bruker TopSpin and analyzed using CcpNMR Analysis. Chemical shifts were scaled and
30
calculated as Δδ = sqrt(((ΔδH)^2+(ΔδN/5)^2) / 2). Chemical shift values were then plotted as a
function of histone peptide concentration and fit to the quadratic binding equation Y=((X+PT+Kd)-
sqrt((X+PT+Kd)^2-4*PT*X))*(Ymax-Ymin)/(2*PT), where X is the concentration of peptide and PT is
the concentration of PHD1, using GraphPad Prism to determine Kd values.
Purification of ARID for NMR
ARID (KDM5C residues 73-188) was cloned into a pET28b His-Smt3 vector to express
recombinant 6xHis-SUMO-ARID in BL21(DE3) E. coli in metal supplemented M9 minimal
medium containing 15NH4Cl. Expression was induced at OD600 ~0.6 using 1 mM IPTG for
expression at 18 ºC overnight. Collected cells were resuspended in 50 mM HEPES pH 8, 500
mM KCl, 1 mM BME, 10 mM imidazole, and 1 mM PMSF, supplemented with EDTA-free Pierce
protease inhibitor tablets and benzonase, and lysed by microfluidizer. Lysate was clarified with
ultracentrifugation and the recovered supernatant was then purified by Ni-NTA affinity resin. The
His-SUMO tag was then cleaved by SenP1 during overnight dialysis at 4 ºC in 50 mM HEPES
pH 7.5, 500 mM KCl, and 5 mM BME. Cleaved His-SUMO tag and SenP1 was captured by
passing through Ni-NTA affinity resin and flow-through was then purified by size exclusion
(Superdex 75, GE Healthcare) chromatography in 50 mM HEPES pH 7, 150 mM KCl, and 5 mM
BME. Fractions were buffer exchanged into 50 mM HEPES pH 7, 50 mM KCl, and 5 mM BME
then concentrated and aliquots snap frozen in liquid nitrogen for storage at -80 ºC.
ARID and DNA NMR titration
For 2D 1H-15N HSQC spectra of KDM5C ARID, 100 µM 15N-labeled ARID in 50 mM HEPES pH
7, 50 mM KCl, 5 mM BME, and 5% D2O was used to obtain 2D spectra at 298K using a 800
MHz Bruker spectrometer equipped with a cryoprobe. Chemical-shift perturbation experiments
were performed by obtaining HSQC spectra with increasing concentrations of the 5’ linker DNA
20 bp fragment up to 1:1 molar ratio of ARID:DNA. For the PHD1 titration experiment, 50 µM
31
15N-labeled ARID in 50 mM HEPES pH 7, 50 mM KCl, 5 mM BME, 50 µM ZnCl2, and 5% D2O
was used with increasing concentrations of PHD1 up to 1:3 molar ratio of ARID:PHD1. Data
were processed using Bruker TopSpin and analyzed using CcpNMR Analysis. Chemical shifts
were scaled and calculated as Δδ = sqrt(((ΔδH)^2+(ΔδN/5)^2) / 2). Previously determined
assignments (BMRB: 15348) were transferred to a majority of resonances observed in the
HSQC spectra of ARID [45].
Purification of ARID mutants
Recombinant His-tagged SUMO-ARID mutants were expressed in BL21(DE3) E. coli in 2xTY
media through induction at OD600 ~0.6 using 1 mM IPTG followed by expression at 18 ºC
overnight. Collected cells were resuspended in 50 mM HEPES pH 8, 500 mM KCl, 1 mM BME,
10 mM imidazole, and 1 mM PMSF, supplemented with benzonase, and lysed by sonication.
Lysate was clarified with centrifugation and the recovered supernatant was then purified by Ni-
NTA affinity resin. The His-SUMO tag was then cleaved by SenP1 for 2 hours at 4 ºC in 50 mM
HEPES pH 7, 500 mM KCl, and 5 mM BME. Cleaved His-SUMO tag and SenP1 was captured
by passing through Ni-NTA affinity resin. The flow-through was buffer exchanged into 50 mM
HEPES pH 7, 50 mM KCl, and 5 mM BME then concentrated and aliquots snap frozen in liquid
nitrogen for storage at -80 ºC.
Single turnover peptide demethylation kinetics
The demethylation of biotinylated H3K4me3 peptide was monitored under single turnover
conditions (>10 fold excess of KDM5C over substrate) through the detection of H3K4me3
substrate loss over time by TR-FRET of an anti-rabbit IgG donor, recognizing an anti-H3K4me3
rabbit antibody, with an anti-biotin acceptor reagent. Various concentrations of KDM5C were
reacted with 25 nM H3K4me3 (1-21)-biotin peptide (AS-64357, AnaSpec) in 50 mM HEPES pH
7.5, 50 mM KCl, 0.01% Tween-20, 0.01% BSA, 50 µM alpha-ketoglutarate, 50 µM ammonium
32
iron(II) sulfate, and 500 µM ascorbic acid at room temperature. 2.5 µL time points were taken
and quenched with 2 mM EDTA then brought to 20 µL final volume for detection using 1:500
dilution anti-H3K4me3 antibody (05-745R, EMD Millipore), 1 nM LANCE Ultra Europium anti-
rabbit IgG antibody (PerkinElmer AD0082), and 50 nM LANCE Ultra Ulight-Streptavidin
(PerkinElmer TRF0102) in 0.5X LANCE detection buffer. Detection reagents were added
stepwise with 30 min incubation of anti-H3K4me3 antibody and Ulight-Streptavidin with reaction
time points followed by 1 hr incubation with Europium anti-rabbit antibody in 384 well white
microplates (PerkinElmer OptiPlate-384). TR-FRET emission at 665 nm and at 615 nm by 320
nm excitation with 50 µs delay and 100 µs integration time was measured using a Molecular
Devices SpectraMax M5e plate reader. TR-FRET was calculated as the 665/615 nm emission
ratio then subject to normalization to H3K4me3 substrate signal before demethylation. Kinetic
curves were fit to a single exponential function, with the plateau set to nonspecific background
of H3K4me2 product detection, to determine kobs of the H3K4me3 demethylation step. kobs
parameters were then plotted as a function of KDM5C concentration and fit to the sigmoidal
kinetic equation Y=kmax*X^n/(Khalf^n + X^n) using GraphPad Prism to determine kmax and Km’
parameters of demethylation.
Multiple turnover peptide demethylation kinetics
A fluorescence-based enzyme coupled assay was used to detect the formaldehyde product of
demethylation of H3K4me3 peptide under multiple turnover conditions (excess of substrate
peptide over KDM5C). Various concentrations of H3K4me3 (1-21) substrate peptide (GenScript)
were added with 1mM alpha-ketoglutarate to initiate demethylation by ~1 µM KDM5C in 50 mM
HEPES pH 7.5, 50 mM KCl, 50 µM ammonium iron(II) sulfate, 2 mM ascorbic acid, 2 mM
NAD+, and 0.05 U formaldehyde dehydrogenase (Sigma-Aldrich) at room temperature. Upon
initiation, fluorescence (350 nm excitation, 460 nm emission) was measured in 20 sec intervals
over 30 min using a Molecular Devices SpectraMax M5e plate reader. NADH standards were
33
used to convert fluorescence to the rate of product concentration formed. Initial rates of the first
3 min of demethylation were plotted as a function of substrate concentration and fit to the tight-
binding quadratic velocity equation Y=Vmax*((X+ET+Km)-sqrt((X+ET+Km)^2-4*ET*X))/(2*ET) using
GraphPad Prism to determine Michaelis-Menten kinetic parameters of demethylation.
Histone peptide binding kinetics
Bio-layer interferometry was used to measure binding kinetics of histone peptides to biotinylated
Avitag-PHD1. Avitag followed by a linker was inserted into pET28b His-Smt3-PHD1318-378 to
generate recombinant endogenously biotinylated 6xHis-SUMO-Avitag-(GS)2-PHD1 through
coexpression with BirA in BL21(DE3) E. coli in 2xTY media containing 50 µM ZnCl2 and 50 µM
biotin. Expression was induced at OD600 ~0.7 using 0.4 mM IPTG for expression at 18 ºC
overnight. Collected cells were resuspended in 50 mM HEPES pH 8, 500 mM KCl, 5 mM BME,
10 mM imidazole, 50 µM biotin, and 1 mM PMSF, supplemented with benzonase, and lysed by
sonication. Lysate was clarified with ultracentrifugation and the recovered supernatant was then
purified by Ni-NTA affinity resin. The His-SUMO tag was then cleaved by SenP1 during
overnight dialysis at 4 ºC in 50 mM HEPES pH 8, 150 mM KCl, 50 µM ZnCl2 and 10 mM BME.
Cleaved His-SUMO tag and SenP1 was captured by passing through Ni-NTA affinity resin and
flow-through was then purified by anion exchange (MonoQ) chromatography in starting buffer of
50 mM HEPES pH 8, 150 mM KCl, 50 µM ZnCl2 and 10 mM BME. Flow-through MonoQ
fractions containing Avitag-PHD1 were analyzed by western blotting to identify biotinylated
fractions, which were then concentrated and aliquots snap frozen in liquid nitrogen for storage at
-80 ºC. Using the Octet Red384 system (ForteBio) at 1000 rpm and 25 ºC, 100 nM Avitag-PHD1
was loaded onto streptavidin biosensors (ForteBio) for 10 min in assay buffer (50 mM HEPES
pH 8, 50 mM KCl, 50 µM ZnCl2, 5 mM BME, and 0.05% Tween-20) followed by 120 sec
baseline then association and dissociation of 100 µM peptide (GenScript) in assay buffer. Data
were processed by subtracting a single reference experiment of loaded Avitag-PHD1 without
34
peptide. A two phase exponential function was used to fit the biphasic kinetic data using Origin
software. For the ARID-PHD1 linker peptides tested to bind PHD1, the KDM5C (263-282) and
KDM5C (271-290) peptides were found to nonspecifically associate with biosensors, and loaded
biosensors were associated with these peptides prior to obtaining their binding traces.
35
ACKNOWLEDGMENTS
We thank Barbara Panning, Geeta Narlikar, John Gross, Daniele Canzio, Ryan Tibble, Cynthia
Chio, and members of the Fujimori laboratory for helpful discussions and guidance.
FUNDING AND ADDITIONAL INFORMATION
This work was supported by the UCSF Discovery Fellows program and National Science
Foundation Graduate Research Fellowship to F. S. U. and by the National Institutes of Health
(R01 GM114044, R01 GM114044-03S1, and R01 CA250459) to D. G. F. The content is solely
the responsibility of the authors and does not necessarily represent the official views of the
National Institutes of Health.
CONFLICT OF INTEREST
The authors declare that they have no conflicts of interest.
SUPPLEMENTAL INFORMATION
This article contains supporting information.
36
REFERENCES
[1] H. Kimura, Histone modifications for human epigenome analysis, Journal of Human
Genetics. 58 (2013) 439–445.
[2] S. Iwase, F. Lan, P. Bayliss, L. de la Torre-Ubieta, M. Huarte, H.H. Qi, J.R. Whetstine, A.
Bonni, T.M. Roberts, Y. Shi, The X-linked mental retardation gene SMCX/JARID1C defines
a family of histone H3 lysine 4 demethylases, Cell. 128 (2007) 1077–1088.
[3] R.J. Klose, Q. Yan, Z. Tothova, K. Yamane, H. Erdjument-Bromage, P. Tempst, D.G.
Gilliland, Y. Zhang, W.G. Kaelin Jr, The retinoblastoma binding protein RBP2 is an H3K4
demethylase, Cell. 128 (2007) 889–900.
[4] K. Yamane, K. Tateishi, R.J. Klose, J. Fang, L.A. Fabrizio, H. Erdjument-Bromage, J.
Taylor-Papadimitriou, P. Tempst, Y. Zhang, PLU-1 is an H3K4 demethylase involved in
transcriptional repression and breast cancer cell proliferation, Mol. Cell. 25 (2007) 801–812.
[5] M.G. Lee, J. Norman, A. Shilatifard, R. Shiekhattar, Physical and functional association of a
trimethyl H3K4 demethylase and Ring6a/MBLR, a polycomb-like protein, Cell. 128 (2007)
877–887.
[6] Y. Shi, F. Lan, C. Matson, P. Mulligan, J.R. Whetstine, P.A. Cole, R.A. Casero, Y. Shi,
Histone demethylation mediated by the nuclear amine oxidase homolog LSD1, Cell. 119
(2004) 941–953.
[7] J. Christensen, K. Agger, P.A.C. Cloos, D. Pasini, S. Rose, L. Sennels, J. Rappsilber, K.H.
Hansen, A.E. Salcini, K. Helin, RBP2 belongs to a family of demethylases, specific for tri-
and dimethylated lysine 4 on histone 3, Cell. 128 (2007) 1063–1076.
[8] J.R. Horton, A. Engstrom, E.L. Zoeller, X. Liu, J.R. Shanks, X. Zhang, M.A. Johns, P.M.
Vertino, H. Fu, X. Cheng, Characterization of a Linked Jumonji Domain of the
KDM5/JARID1 Family of Histone H3 Lysine 4 Demethylases, J. Biol. Chem. 291 (2016)
2631–2646.
[9] C. Johansson, S. Velupillai, A. Tumber, A. Szykowska, E.S. Hookway, R.P. Nowak, C.
Strain-Damerell, C. Gileadi, M. Philpott, N. Burgess-Brown, N. Wu, J. Kopec, A. Nuzzi, H.
Steuber, U. Egner, V. Badock, S. Munro, N.B. LaThangue, S. Westaway, J. Brown, N.
Athanasou, R. Prinjha, P.E. Brennan, U. Oppermann, Structural analysis of human KDM5B
guides histone demethylase inhibitor development, Nat. Chem. Biol. 12 (2016) 539–545.
[10] L. Li, C. Greer, R.N. Eisenman, J. Secombe, Essential functions of the histone demethylase
lid, PLoS Genet. 6 (2010) e1001221.
[11] E. Brookes, B. Laurent, K. Õunap, R. Carroll, J.B. Moeschler, M. Field, C.E. Schwartz, J.
Gecz, Y. Shi, Mutations in the intellectual disability gene KDM5C reduce protein stability
and demethylase activity, Hum. Mol. Genet. 24 (2015) 2861–2872.
[12] Y. Xiang, Z. Zhu, G. Han, X. Ye, B. Xu, Z. Peng, Y. Ma, Y. Yu, H. Lin, A.P. Chen, C.D.
Chen, JARID1B is a histone H3 lysine 4 demethylase up-regulated in prostate cancer,
Proc. Natl. Acad. Sci. U. S. A. 104 (2007) 19226–19231.
[13] S. Tu, Y.-C. Teng, C. Yuan, Y.-T. Wu, M.-Y. Chan, A.-N. Cheng, P.-H. Lin, L.-J. Juan, M.-D.
Tsai, The ARID domain of the H3K4 demethylase RBP2 binds to a DNA CCGCCC motif,
Nat. Struct. Mol. Biol. 15 (2008) 419–421.
[14] A.G. Scibetta, S. Santangelo, J. Coleman, D. Hall, T. Chaplin, J. Copier, S. Catchpole, J.
Burchell, J. Taylor-Papadimitriou, Functional analysis of the transcription repressor PLU-
1/JARID1B, Mol. Cell. Biol. 27 (2007) 7220–7235.
[15] W. Yao, Y. Peng, D. Lin, The flexible loop L1 of the H3K4 demethylase JARID1B ARID
domain has a crucial role in DNA-binding activity, Biochem. Biophys. Res. Commun. 396
(2010) 323–328.
[16] F. Lan, R.E. Collins, R. De Cegli, R. Alpatov, J.R. Horton, X. Shi, O. Gozani, X. Cheng, Y.
Shi, Recognition of unmethylated histone H3 lysine 4 links BHC80 to LSD1-mediated gene
37
repression, Nature. 448 (2007) 718–722.
[17] B.J. Klein, X. Wang, G. Cui, C. Yuan, M.V. Botuyan, K. Lin, Y. Lu, X. Wang, Y. Zhao, C.J.
Bruns, G. Mer, X. Shi, T.G. Kutateladze, PHF20 Readers Link Methylation of Histone H3K4
and p53 with H4K16 Acetylation, Cell Rep. 17 (2016) 1158–1170.
[18] X. Shi, T. Hong, K.L. Walter, M. Ewalt, E. Michishita, T. Hung, D. Carney, P. Peña, F. Lan,
M.R. Kaadige, N. Lacoste, C. Cayrou, F. Davrazou, A. Saha, B.R. Cairns, D.E. Ayer, T.G.
Kutateladze, Y. Shi, J. Côté, K.F. Chua, O. Gozani, ING2 PHD domain links histone H3
lysine 4 methylation to active gene repression, Nature. 442 (2006) 96–99.
[19] H. Li, S. Ilin, W. Wang, E.M. Duncan, J. Wysocka, C.D. Allis, D.J. Patel, Molecular basis for
site-specific read-out of histone H3K4me3 by the BPTF PHD finger of NURF, Nature. 442
(2006) 91–95.
[20] P.V. Peña, F. Davrazou, X. Shi, K.L. Walter, V.V. Verkhusha, O. Gozani, R. Zhao, T.G.
Kutateladze, Molecular mechanism of histone H3K4me3 recognition by plant homeodomain
of ING2, Nature. 442 (2006) 100–103.
[21] J. Wysocka, T. Swigut, H. Xiao, T.A. Milne, S.Y. Kwon, J. Landry, M. Kauer, A.J. Tackett,
B.T. Chait, P. Badenhorst, C. Wu, C.D. Allis, A PHD finger of NURF couples histone H3
lysine 4 trimethylation with chromatin remodelling, Nature. 442 (2006) 86–90.
[22] R. Sanchez, M.-M. Zhou, The PHD finger: a versatile epigenome reader, Trends Biochem.
Sci. 36 (2011) 364–372.
[23] B.J. Klein, L. Piao, Y. Xi, H. Rincon-Arano, S.B. Rothbart, D. Peng, H. Wen, C. Larson, X.
Zhang, X. Zheng, M.A. Cortazar, P.V. Peña, A. Mangan, D.L. Bentley, B.D. Strahl, M.
Groudine, W. Li, X. Shi, T.G. Kutateladze, The histone-H3K4-specific demethylase KDM5B
binds to its substrate and product through distinct PHD fingers, Cell Rep. 6 (2014) 325–
335.
[24] Y. Zhang, H. Yang, X. Guo, N. Rong, Y. Song, Y. Xu, W. Lan, X. Zhang, M. Liu, Y. Xu, C.
Cao, The PHD1 finger of KDM5B recognizes unmodified H3K4 during the demethylation of
histone H3K4me2/3 by KDM5B, Protein Cell. 5 (2014) 837–850.
[25] I.O. Torres, K.M. Kuchenbecker, C.I. Nnadi, R.J. Fletterick, M.J.S. Kelly, D.G. Fujimori,
Histone demethylase KDM5A is regulated by its reader domain through a positive-feedback
mechanism, Nat. Commun. 6 (2015) 6204.
[26] S. Zhao, K.N. Chuh, B. Zhang, B.E. Dul, R.E. Thompson, L.A. Farrelly, X. Liu, N. Xu, Y.
Xue, R.G. Roeder, I. Maze, T.W. Muir, H. Li, Histone H3Q5 serotonylation stabilizes H3K4
methylation and potentiates its readout, Proc. Natl. Acad. Sci. U. S. A. 118 (2021).
https://doi.org/10.1073/pnas.2016742118.
[27] G.G. Wang, J. Song, Z. Wang, H.L. Dormann, F. Casadio, H. Li, J.-L. Luo, D.J. Patel, C.D.
Allis, Haematopoietic malignancies caused by dysregulation of a chromatin-binding PHD
finger, Nature. 459 (2009) 847–851.
[28] J.E. Longbotham, C.M. Chio, V. Dharmarajan, M.J. Trnka, I.O. Torres, D. Goswami, K.
Ruiz, A.L. Burlingame, P.R. Griffin, D.G. Fujimori, Histone H3 binding to the PHD1 domain
of histone demethylase KDM5A enables active site remodeling, Nat. Commun. 10 (2019)
94.
[29] S. Iwase, E. Brookes, S. Agarwal, A.I. Badeaux, H. Ito, C.N. Vallianatos, G.S. Tomassy, T.
Kasza, G. Lin, A. Thompson, L. Gu, K.Y. Kwan, C. Chen, M.A. Sartor, B. Egan, J. Xu, Y.
Shi, A Mouse Model of X-linked Intellectual Disability Associated with Impaired Removal of
Histone Methylation, Cell Rep. 14 (2016) 1000–1009.
[30] L.R. Jensen, H. Bartenschlager, S. Rujirabanjerd, A. Tzschach, A. Nümann, A.R. Janecke,
R. Spörle, S. Stricker, M. Raynaud, J. Nelson, A. Hackett, J.-P. Fryns, J. Chelly, A.P. de
Brouwer, B. Hamel, J. Gecz, H.-H. Ropers, A.W. Kuss, A distinctive gene expression
fingerprint in mentally retarded male patients reflects disease-causing defects in the histone
demethylase KDM5C, Pathogenetics. 3 (2010) 2.
[31] L.R. Jensen, M. Amende, U. Gurok, B. Moser, V. Gimmel, A. Tzschach, A.R. Janecke, G.
38
Tariverdian, J. Chelly, J.-P. Fryns, H. Van Esch, T. Kleefstra, B. Hamel, C. Moraine, J.
Gecz, G. Turner, R. Reinhardt, V.M. Kalscheuer, H.-H. Ropers, S. Lenzner, Mutations in
the JARID1C gene, which is involved in transcriptional regulation and chromatin
remodeling, cause X-linked mental retardation, Am. J. Hum. Genet. 76 (2005) 227–236.
[32] M. Scandaglia, J.P. Lopez-Atalaya, A. Medrano-Fernandez, M.T. Lopez-Cascales, B. Del
Blanco, M. Lipinski, E. Benito, R. Olivares, S. Iwase, Y. Shi, A. Barco, Loss of Kdm5c
Causes Spurious Transcription and Prevents the Fine-Tuning of Activity-Regulated
Enhancers in Neurons, Cell Rep. 21 (2017) 47–59.
[33] N.S. Outchkourov, J.M. Muiño, K. Kaufmann, W.F.J. van Ijcken, M.J. Groot Koerkamp, D.
van Leenen, P. de Graaf, F.C.P. Holstege, F.G. Grosveld, H.T.M. Timmers, Balancing of
histone H3K4 methylation states by the Kdm5c/SMCX histone demethylase modulates
promoter and enhancer function, Cell Rep. 3 (2013) 1071–1079.
[34] H. Shen, W. Xu, R. Guo, B. Rong, L. Gu, Z. Wang, C. He, L. Zheng, X. Hu, Z. Hu, Z.-M.
Shao, P. Yang, F. Wu, Y.G. Shi, Y. Shi, F. Lan, Suppression of Enhancer Overactivation by
a RACK7-Histone Demethylase Complex, Cell. 165 (2016) 331–342.
[35] T.F. Gonçalves, A.P. Gonçalves, N. Fintelman Rodrigues, J.M. dos Santos, M.M.G.
Pimentel, C.B. Santos-Rebouças, KDM5C mutational screening among males with
intellectual disability suggestive of X-Linked inheritance and review of the literature, Eur. J.
Med. Genet. 57 (2014) 138–144.
[36] A. Tzschach, S. Lenzner, B. Moser, R. Reinhardt, J. Chelly, J.-P. Fryns, T. Kleefstra, M.
Raynaud, G. Turner, H.-H. Ropers, A. Kuss, L.R. Jensen, Novel JARID1C/SMCX mutations
in patients with X-linked mental retardation, Hum. Mutat. 27 (2006) 389.
[37] C. Santos, L. Rodriguez-Revenga, I. Madrigal, C. Badenas, M. Pineda, M. Milà, A novel
mutation in JARID1C gene associated with mental retardation, Eur. J. Hum. Genet. 14
(2006) 583–586.
[38] F.E. Abidi, L. Holloway, C.A. Moore, D.D. Weaver, R.J. Simensen, R.E. Stevenson, R.C.
Rogers, C.E. Schwartz, Mutations in JARID1C are associated with X-linked mental
retardation, short stature and hyperreflexia, J. Med. Genet. 45 (2008) 787–793.
[39] S. Rujirabanjerd, J. Nelson, P.S. Tarpey, A. Hackett, S. Edkins, F. Lucy Raymond, C.E.
Schwartz, G. Turner, S. Iwase, Y. Shi, P. Andrew Futreal, M.R. Stratton, J. Gecz,
Identification and characterization of two novel JARID1C mutations: suggestion of an
emerging genotype–phenotype correlation, European Journal of Human Genetics. 18
(2010) 330–335. https://doi.org/10.1038/ejhg.2009.175.
[40] J. Xu, P.S. Burgoyne, A.P. Arnold, Sex differences in sex chromosome gene expression in
mouse brain, Human Molecular Genetics. 11 (2002) 1409–1419.
https://doi.org/10.1093/hmg/11.12.1409.
[41] C.N. Vallianatos, C. Farrehi, M.J. Friez, M. Burmeister, C.E. Keegan, S. Iwase, Altered
Gene-Regulatory Function of KDM5C by a Novel Mutation Associated With Autism and
Intellectual Disability, Front. Mol. Neurosci. 11 (2018) 104.
[42] M. Tahiliani, P. Mei, R. Fang, T. Leonor, M. Rutenberg, F. Shimizu, J. Li, A. Rao, Y. Shi,
The histone H3K4 demethylase SMCX links REST target genes to X-linked mental
retardation, Nature. 447 (2007) 601–605.
[43] B.N. Jones, D.-U. Quang-Dang, Y. Oku, J.D. Gross, A kinetic assay to monitor RNA
decapping under single- turnover conditions, Methods Enzymol. 448 (2008) 23–40.
[44] J.E. Longbotham, M.J.S. Kelly, D.G. Fujimori, Recognition of Histone H3 Methylation States
by the PHD1 Domain of Histone Demethylase KDM5A, ACS Chem. Biol. (2021).
https://doi.org/10.1021/acschembio.0c00976.
[45] C. Koehler, S. Bishop, E.F. Dowler, P. Schmieder, A. Diehl, H. Oschkinat, L.J. Ball,
Backbone and sidechain 1H, 13C and 15N resonance assignments of the Bright/ARID
domain from the human JARID1C (SMCX) protein, Biomol. NMR Assign. 2 (2008) 9–11.
[46] L. Poeta, A. Padula, M.B. Lioi, H. van Bokhoven, M.G. Miano, Analysis of a Set of KDM5C
39
Regulatory Genes Mutated in Neurodevelopmental Disorders Identifies Temporal
Coexpression Brain Signatures, Genes. 12 (2021). https://doi.org/10.3390/genes12071088.
[47] S.-A. Kim, J. Zhu, N. Yennawar, P. Eek, S. Tan, Crystal Structure of the LSD1/CoREST
Histone Demethylase Bound to Its Nucleosome Substrate, Mol. Cell. 78 (2020) 903–
914.e4.
[48] V. Kasinath, C. Beck, P. Sauer, S. Poepsel, J. Kosmatka, M. Faini, D. Toso, R. Aebersold,
E. Nogales, JARID2 and AEBP2 regulate PRC2 in the presence of H2AK119ub1 and other
histone modifications, Science. 371 (2021). https://doi.org/10.1126/science.abc3393.
[49] K. Finogenova, J. Bonnet, S. Poepsel, I.B. Schäfer, K. Finkl, K. Schmid, C. Litz, M. Strauss,
C. Benda, J. Müller, Structural basis for PRC2 decoding of active histone methylation
marks H3K36me2/3, Elife. 9 (2020). https://doi.org/10.7554/eLife.61964.
[50] E.J. Worden, X. Zhang, C. Wolberger, Structural basis for COMPASS recognition of an
H2B-ubiquitinated nucleosome, Elife. 9 (2020). https://doi.org/10.7554/eLife.53199.
[51] W. Li, W. Tian, G. Yuan, P. Deng, D. Sengupta, Z. Cheng, Y. Cao, J. Ren, Y. Qin, Y. Zhou,
Y. Jia, O. Gozani, D.J. Patel, Z. Wang, Molecular basis of nucleosomal H3K36 methylation
by NSD methyltransferases, Nature. 590 (2021) 498–503.
[52] S. Bilokapic, M.J. Suskiewicz, I. Ahel, M. Halic, Bridging of DNA breaks activates PARP2–
HPF1 to modify chromatin, Nature. 585 (2020) 609–613. https://doi.org/10.1038/s41586-
020-2725-7.
[53] S. Bilokapic, M. Halic, Nucleosome and ubiquitin position Set2 to methylate H3K36, Nat.
Commun. 10 (2019) 3795.
[54] P.L. Hsu, H. Shi, C. Leonen, J. Kang, C. Chatterjee, N. Zheng, Structural Basis of H2B
Ubiquitination-Dependent H3K4 Methylation by COMPASS, Mol. Cell. 76 (2019) 712–
723.e4.
[55] C. Marabelli, B. Marrocco, S. Pilotto, S. Chittori, S. Picaud, S. Marchese, G. Ciossani, F.
Forneris, P. Filippakopoulos, G. Schoehn, D. Rhodes, S. Subramaniam, A. Mattevi, A Tail-
Based Mechanism Drives Nucleosome Demethylation by the LSD2/NPAC Multimeric
Complex, Cell Rep. 27 (2019) 387–399.e7.
[56] S. Poepsel, V. Kasinath, E. Nogales, Cryo-EM structures of PRC2 simultaneously engaged
with two functionally distinct nucleosomes, Nat. Struct. Mol. Biol. 25 (2018) 154–162.
[57] S.H. Park, A. Ayoub, Y.-T. Lee, J. Xu, H. Kim, W. Zheng, B. Zhang, L. Sha, S. An, Y.
Zhang, M.A. Cianfrocco, M. Su, Y. Dou, U.-S. Cho, Cryo-EM structure of the human MLL1
core complex bound to the nucleosome, Nat. Commun. 10 (2019) 1–13.
[58] H. Xue, T. Yao, M. Cao, G. Zhu, Y. Li, G. Yuan, Y. Chen, M. Lei, J. Huang, Structural basis
of nucleosome recognition and modification by MLL methyltransferases, Nature. 573 (2019)
445–449.
[59] S. Hatazawa, J. Liu, Y. Takizawa, M. Zandian, L. Negishi, T.G. Kutateladze, H.
Kurumizaka, Structural basis for binding diversity of acetyltransferase p300 to the
nucleosome, iScience. 25 (2022) 104563. https://doi.org/10.1016/j.isci.2022.104563.
[60] J. Gatchalian, X. Wang, J. Ikebe, K.L. Cox, A.H. Tencer, Y. Zhang, N.L. Burge, L. Di, M.D.
Gibson, C.A. Musselman, M.G. Poirier, H. Kono, J.J. Hayes, T.G. Kutateladze, Accessibility
of the histone H3 tail in the nucleosome for binding of paired readers, Nat. Commun. 8
(2017) 1489.
[61] E.A. Morrison, S. Bowerman, K.L. Sylvers, J. Wereszczynski, C.A. Musselman, The
conformation of the histone H3 tail inhibits association of the BPTF PHD finger with the
nucleosome, Elife. 7 (2018). https://doi.org/10.7554/eLife.31481.
[62] A. Stützer, S. Liokatis, A. Kiesel, D. Schwarzer, R. Sprangers, J. Söding, P. Selenko, W.
Fischle, Modulations of DNA Contacts by Linker Histones and Post-translational
Modifications Determine the Mobility and Modifiability of Nucleosomal H3 Tails, Mol. Cell.
61 (2016) 247–259.
[63] Y. Peng, S. Li, A. Onufriev, D. Landsman, A.R. Panchenko, Binding of regulatory proteins
40
to nucleosomes is modulated by dynamic histone tails, Nat. Commun. 12 (2021) 5280.
[64] T.M. Weaver, E.A. Morrison, C.A. Musselman, Reading More than Histones: The
Prevalence of Nucleic Acid Binding among Reader Domains, Molecules. 23 (2018).
https://doi.org/10.3390/molecules23102614.
[65] S. Pilotto, V. Speranzini, M. Tortorici, D. Durand, A. Fish, S. Valente, F. Forneris, A. Mai,
T.K. Sixma, P. Vachette, A. Mattevi, Interplay among nucleosomal DNA, histone tails, and
corepressor CoREST underlies LSD1-mediated H3 demethylation, Proc. Natl. Acad. Sci. U.
S. A. 112 (2015) 2752–2757.
[66] C.A. Musselman, T.G. Kutateladze, Characterization of functional disordered regions within
chromatin-associated proteins, iScience. 24 (2021) 102070.
[67] S. Contreras-Martos, A. Piai, S. Kosol, M. Varadi, A. Bekesi, P. Lebrun, A.N. Volkov, K.
Gevaert, R. Pierattelli, I.C. Felli, P. Tompa, Linking functions: an additional role for an
intrinsically disordered linker domain in the transcriptional coactivator CBP, Sci. Rep. 7
(2017) 4676.
[68] L. Jiao, M. Shubbar, X. Yang, Q. Zhang, S. Chen, Q. Wu, Z. Chen, J. Rizo, X. Liu, A
partially disordered region connects gene repression and activation functions of EZH2,
Proc. Natl. Acad. Sci. U. S. A. 117 (2020) 16992–17002.
[69] M.M. Keenen, D. Brown, L.D. Brennan, R. Renger, H. Khoo, C.R. Carlson, B. Huang, S.W.
Grill, G.J. Narlikar, S. Redding, HP1 proteins compact DNA into mechanically and
positionally stable phase separated domains, Elife. 10 (2021).
https://doi.org/10.7554/eLife.64563.
[70] S.-A. Kim, N. Chatterjee, M.J. Jennings, B. Bartholomew, S. Tan, Extranucleosomal DNA
enhances the activity of the LSD1/CoREST histone demethylase complex, Nucleic Acids
Res. 43 (2015) 4868–4880.
[71] J.C. Zhou, N.P. Blackledge, A.M. Farcas, R.J. Klose, Recognition of CpG island chromatin
by KDM2A requires direct and specific interaction with linker DNA, Mol. Cell. Biol. 32 (2012)
479–489.
[72] N.P. Blackledge, J.C. Zhou, M.Y. Tolstorukov, A.M. Farcas, P.J. Park, R.J. Klose, CpG
islands recruit a histone H3 lysine 36 demethylase, Mol. Cell. 38 (2010) 179–190.
[73] S. Zamurrad, H.A.M. Hatch, C. Drelon, H.M. Belalcazar, J. Secombe, A Drosophila Model
of Intellectual Disability Caused by Mutations in the Histone Demethylase KDM5, Cell Rep.
22 (2018) 2359–2369.
[74] G. Erdős, M. Pajkos, Z. Dosztányi, IUPred3: prediction of protein disorder enhanced with
unambiguous experimental annotation and visualization of evolutionary conservation,
Nucleic Acids Res. 49 (2021) W297–W303.
41
ABBREVIATIONS
XLID (X-linked intellectual disability), KDM (lysine demethylase), H3 (histone 3), H3K4 (histone
3 lysine 4), H3K4me1 (monomethylated lysine 4 of histone 3), H3K4me2 (dimethylated lysine 4
of histone 3), H3K4me3 (trimethylated lysine 4 of histone 3), PHD (plant homeodomain), ARID
(AT-rich interaction domain), JmjN (Jumonji N domain), JmjC (Jumonji C domain), ZnF (zinc
finger domain), KDM5C1-839 ∆AP (KDM5C protein containing residues 1 to 839 with truncation of
the ARID and PHD1 region), KDM5C1-839 ∆linker (KDM5C protein containing residues 1 to 839
with truncation of the linker region between ARID and PHD1), HSQC (heteronuclear single
quantum coherence), EMSA (electrophoretic mobility shift assay), TR-FRET (time-resolved
fluorescence resonance energy transfer), nuc (nucleosome)
42
SUPPLEMENTAL INFORMATION
0
2
4
6
8
0
5
10
[KDM5C construct] (µM)
kobs (min-1)
H3K4me3 (1-21) demethylation
single turnover
kmax (min-1)
Km' ( M)
KDM5C1-839
KDM5C1-839 AP
4.2 � 0.1
10.6 ��0.3
3.7 � 0.6
4.2 ��0.3
n
2.5
1.0
Figure 1. supplemental
A
B
0
200
400
0
1
2
[H3K4me3 (1-21)] ( M)
Rate (min-1)
H3K4me3 (1-21) demethylation
multiple turnover
kcat (min-1)
Km ( M)
KDM5C1-839
KDM5C1-839 AP
2.6 � 0.7
1.34 ��0.08
3.6 � 0.7
2.22 ��0.07
0
1.4
2.7
5.4
10.9
µM
KDM5C1-839
H3K4me3 nucleosome
nuc
KDM5C
-nuc
complex
0
1.4
2.7
5.4
10.9
µM
KDM5C1-839
unmodified nucleosome
nuc
KDM5C
-nuc
complex
0
2.5
5
9.9
19.8
µM
KDM5C1-839 ���
H3K4me3 nucleosome
nuc
0
2.5
5
9.9
19.8
µM
KDM5C1-839 ���
unmodified nucleosome
nuc
0
1
10
0.0
0.5
1.0
[KDM5C] ( M)
Fraction unbound nuc
KDM5C1-839
Kd
app
n
6.9 � 0.8 M
�12.5 � 1.2 M
2.3
2.1
H3K4me3 nuc
unmod nuc
0
1
10
100
0.0
0.5
1.0
[KDM5C ��] ( M)
Fraction unbound nuc
KDM5C1-839 AP
Kd
app
n
20 � 1 M
�35 � 3 M
2.4
1.4
H3K4me3 nuc
unmod nuc
43
Figure S1. Related to Figure 1.
(A) H3K4me3 substrate peptide demethylation by KDM5C constructs. Left: Demethylation kinetics of the H3K4me3
(1-21) substrate peptide by KDM5C constructs under single turnover conditions measured by a TR-FRET based
kinetic assay. Observed rates are fit to a cooperative kinetic model, with n denoting the Hill coefficient.
Representative kinetic traces used to determine observed demethylation rates are in Figure S1D. Right:
Demethylation kinetics of the H3K4me3 (1-21) substrate peptide by KDM5C constructs under multiple turnover
conditions measured by a formaldehyde release based kinetic assay. Deletion of the ARID and PHD1 region results
in higher demethylase activity on the substrate peptide under multiple turnover conditions due to loss of substrate
inhibition caused by this region. (B) Unmodified and substrate core nucleosome binding by KDM5C1-839 and KDM5C1-
839 ∆AP. Nucleosome binding curves were measured by EMSA and fit to a cooperative binding model to determine
apparent dissociation constants (Kdapp), with n denoting the Hill coefficient (top). Representative gel shifts of KDM5C
binding to nucleosomes (bottom). Due to unattainable saturation of binding, a lower limit for the dissociation constant
is presented for the unmodified nucleosome. (C) Representative demethylation kinetic traces of substrate
nucleosome demethylation by KDM5C constructs (left: KDM5C1-839, right: KDM5C1-839 ∆AP) under single turnover
conditions using TR-FRET based kinetic assay detecting formation of the H3K4me1/2 product nucleosome over time.
Observed rates (kobs) are obtained by fitting kinetic traces to an exponential function. (D) Representative
demethylation kinetic traces of substrate peptide demethylation by KDM5C constructs (left: KDM5C1-839, right:
KDM5C1-839 ∆AP) under single turnover conditions using TR-FRET based kinetic assay detecting loss of the
H3K4me3 substrate peptide over time. Observed rates (kobs) are obtained by fitting kinetic traces to an exponential
function. All error bars represent SEM of at least two independent experiments (n ≥ 2).
C
0
30
60
90
120
0.00
0.05
0.10
0.15
0.20
Time (min)
H3K4me1/2 TR-FRET
(665/615 nm ratio)
KDM5C1-839
H3K4me3 nucleosome demethylation
0.5 M
1 M
[KDM5C1-839]
4.1 M
6.1 M
8.1 M
2 M
0
1
2
3
4
5
0.0
0.5
1.0
Time (min)
H3K4me3 TR-FRET
(relative to t=0)
KDM5C1-839
H3K4me3 (1-21) demethylation
0.25 M
0.5 M
1 M
2 M
4.1 M
8.1 M
[KDM5C1-839]
D
0
60
120
180
240
0.00
0.05
0.10
0.15
0.20
Time (min)
H3K4me1/2 TR-FRET
(665/615 nm ratio)
KDM5C1-839����
H3K4me3 nucleosome demethylation
5.6 M
11.2 M
16.7 M
19.5 M
22.3 M
13.9 M
[KDM5C1-839 AP]
0
1
2
3
4
5
0.0
0.5
1.0
Time (min)
H3K4me3 TR-FRET
(relative to t=0)
KDM5C1-839����
H3K4me3 (1-21) demethylation
0.25 M
0.5 M
1 M
2 M
3.5 M
7 M
[KDM5C1-839 AP]
�
�
44
Figure 2. supplemental
���(ppm)
1.0
N/A
0.05
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
���(ppm)
1.0
N/A
0.05
0.2
0.4
0.6
0.8
A
B
C
E
D
KDM5C
KDM5D
KDM5A
KDM5B
307
297
276
295
388
378
357
373
KM T M R L R R N H SN AQ F I E SY V C RM C SR GD ED D K L L L C D GC D D N Y H I F C L L P P L P E I P K GVWR C P K C VM A EC K R P P EA F G F EQ A
C
C
C
C
H
C
C
C
K T T MQ L R K N H S SAQ F I D SY I CQ V C SR GD ED D K L L F C D GC D D N Y H I F C L L P P L P E I P R G I WR C P K C I L A EC KQ P P EA F G F EQ A
C
C
C
C
H
C
C
C
NMQM RQ R K GT L SV N F V D L Y V CM F C GR GN N ED K L L L C D GC D D SY H T F C L I P P L P D V P K GDWR C P K C V A E EC SK P R EA F G F EQ A
C
C
C
C
H
C
C
C
K P K SR SK K A T - - - N A V D L Y V C L L C G SGN D ED R L L L C D GC D D SY H T F C L I P P L H D V P K GDWR C P K C L AQ EC SK PQ EA F G F EQ A
C
C
C
C
H
C
C
C
F
0
200
400
0.0
0.5
1.0
1.5
[H3K4me3 (1-21)] ( M)
Rate (min-1)
H3K4me3 (1-21) demethylation
multiple turnover
kcat (min-1)
Km ( M)
KDM5C1-839
KDM5C1-839 D343A
2.6 � 0.7
1.34 ��0.08
4.0 � 0.6
1.52 ��0.07
0
60
120
180
240
300
360
0.0
0.2
0.4
Time (sec)
Response (nm)
H3K4me0 (1-18)
H3K4me1 (1-18)
H3K4me3 (1-18)
H3K4me2 (1-18)
!
!
!
!
!
!
Association
Dissociation
H3 (1-18) peptide
kobs1 (s
-1)
kobs2 (s
-1)
kobs1 (s
-1)
kobs2 (s
-1)
H3K4me0
3.07 ± 0.01
0.014 ± 0.0005
0.87 ± 0.01
0.017 ± 0.0007
H3K4me1
2.42 ± 0.02
0.018 ± 0.0003
1.34 ± 0.03
0.010 ± 0.0002
H3K4me2
2.02 ± 0.09
0.026 ± 0.0005
1.37 ± 0.04
0.014 ± 0.0003
H3K4me3
2.73 ± 0.11
0.019 ± 0.0008
1.55 ± 0.03
0.019 ± 0.0007
!
!
!
!
!
!
!
!
!
!
0
60
120
180
240
300
360
0.0
0.2
0.4
Time (sec)
Response (nm)
H3 (1-18)
H3K9me3 (1-18)
!
!
!
!
!
Association
Dissociation
H3 (1-18) peptide
kobs1 (s
-1)
kobs2 (s
-1)
kobs1 (s
-1)
kobs2 (s
-1)
H3
3.23 ± 0.11
0.018 ± 0.001
0.87 ± 0.01
0.015 ± 0.001
H3K9me3
3.32 ± 0.18
0.015 ± 0.001
0.95 ± 0.01
0.016 ± 0.001
!
!
0
60
120
180
240
300
360
0.0
0.2
0.4
Time (sec)
Response (nm)
H3 (1-18)
H3R2A (1-18)
H3K4A (1-18)
!
!
!
Association
Dissociation
H3 (1-18) peptide
kobs1 (s
-1)
kobs2 (s
-1)
kobs1 (s
-1)
kobs2 (s
-1)
H3
3.23 ± 0.11
0.018 ± 0.0007
0.87 ± 0.01
0.015 ± 0.001
H3R2A
3.32 ± 0.36
0.025 ± 0.0026
2.09 ± 0.13
0.021 ± 0.002
H3K4A
3.61 ± 0.17
0.017 ± 0.0007
0.76 ± 0.01
0.016 ± 0.001
!
!
0
2
4
6
8
0
5
10
[KDM5C construct] (µM)
kobs (min-1)
H3K4me3 (1-21) demethylation
single turnover
kmax (min-1)
Km' ( M)
KDM5C1-839
KDM5C1-839 D343A
4.2 � 0.1
10.6 ��0.3
1.3 � 0.1
8.6 ��0.4
n
2.5
5.8
KDM5D PHD1
PDB: 2E6R
D343A
45
Figure S2. Related to Figure 2.
(A) Binding kinetic trace of immobilized Avitag-PHD1 binding to H3 (1-18) and H3K9me3 (1-18) tail peptides
measured by bio-layer interferometry (BLI). Observed rates (kobs) of association and dissociation are obtained by
fitting kinetic traces to a two phase exponential function. (B) Binding kinetic trace of immobilized Avitag-PHD1 binding
to H3K4me0/1/2/3 (1-18) tail peptides measured by BLI. Biphasic kinetic binding by PHD1 is modulated by the
H3K4me state. (C) Chemical shift perturbations of PHD1 residues upon binding of the H3 (1-18) tail peptide (Figure
2A) colored by the gradient, unperturbed (yellow) to significantly perturbed (maroon), mapped to homologous
residues in KDM5D PHD1 structure (PDB: 2E6R). Significantly perturbed residues are labeled. (D) Binding kinetic
trace of immobilized Avitag-PHD1 binding to H3 (1-18) and H3 mutant (1-18) tail peptides (H3R2A and H3K4A)
measured by BLI. Recognition of the H3 tail by PHD1 depends on the R2 residue but not K4 residue in H3. (E)
Sequence alignment of PHD1 domains in KDM5A-D. The H3R2 recognizing residues D312 and D315 of KDM5A are
indicated in red, and the PHD1 mutation D343A from this study is denoted above KDM5C. Zinc coordinating residues
G
0
1
10
0.0
0.5
1.0
[KDM5C construct] ( M)
Fraction unbound nuc
H3K4me3 nucleosome
Kd
app
n
6.9 � 0.8 M
8.9 � 0.7 M
2.3
2.2
KDM5C1-839
KDM5C1-839 D343A
0
1
10
0.0
0.5
1.0
[KDM5C construct] ( M)
Fraction unbound nuc
Unmodified nucleosome
Kd
app
n
�12.5 � 1.2 M
8.4 � 0.5 M
2.1
2.5
KDM5C1-839
KDM5C1-839 D343A
H
100
1000
0.0
0.2
0.4
0.6
0.8
[H3 (1-18) peptide] ( M)
(ppm)
PHD1 & H3 (1-18)
D346
D343
I361
L341
D336
C345
Kd = 127 � 5 M
100
1000
0.0
0.2
0.4
0.6
[H3K4me1 (1-18) peptide] ( M)
(ppm)
PHD1 & H3K4me1 (1-18)
D346
D343
I361
L341
C342
D336
C345
Kd = 310 � 4 M
100
1000
0.0
0.1
0.2
0.3
[H3K4me2 (1-18) peptide] ( M)
(ppm)
PHD1 & H3K4me2 (1-18)
D346
D343
I361
L341
C342
D336
C345
Kd 882 � 22 M
100
1000
0.0
0.1
0.2
[H3K4me3 (1-18) peptide] ( M)
(ppm)
PHD1 & H3K4me3 (1-18)
D346
D343
I361
L341
C342
D336
C345
Kd 1325 � 51 M
46
are highlighted in gray. (F) H3K4me3 substrate peptide demethylation by PHD1 mutant KDM5C1-839 relative to wild
type. Left: Demethylation kinetics of the H3K4me3 (1-21) substrate peptide under single turnover conditions
measured by a TR-FRET based kinetic assay. Observed rates are fit to a cooperative kinetic model, with n denoting
the Hill coefficient. Unlike on the substrate nucleosome, the D343A PHD1 mutation does not increase catalytic rate
on the substrate peptide, but does increase overall catalytic efficiency. Right: Demethylation kinetics of the H3K4me3
(1-21) substrate peptide under multiple turnover conditions measured by a formaldehyde release based kinetic assay.
The D343A PHD1 mutation does not affect catalysis on the substrate peptide under these conditions, nor does it
significantly affect substrate inhibition. (G) Unmodified and substrate core nucleosome binding by PHD1 mutant
KDM5C1-839 relative to wild type. Nucleosome binding curves were measured by EMSA and fit to a cooperative
binding model to determine apparent dissociation constants (Kdapp), with n denoting the Hill coefficient. Due to
unattainable saturation of binding, a lower limit for the dissociation constant is presented for WT KDM5C binding the
unmodified nucleosome. (H) Binding of the H3K4me0/1/2/3 (1-18) tail peptides by PHD1 by NMR titration HSQC
experiments of indicated PHD1 residues that localize to the H3 binding surface (Figure S2C). Average dissociation
constants with standard error for each ligand were determined from dissociation constants obtained from chemical
shift changes (Δδ) of individual PHD1 residues. Due to incomplete saturation of binding, a lower limit for the
dissociation constant is presented for the H3K4me2/3 peptides. All error bars represent SEM of at least two
independent experiments (n ≥ 2).
47
Figure S3. Related to Figure 3.
(A) 20 bp linker DNA fragment binding by the ARID domain. Fragments contain 5’ and 3’ flanking DNA sequences
used in the 187 bp nucleosome. Binding curves were measured by EMSA and fit to a binding model to determine
apparent dissociation constants (Kdapp) (left). Representative gel shifts of ARID binding to 20 bp flanking linker DNA
fragments (right). (B) 2D 1H-15N HSQC spectra of ARID titrated with increasing amounts of the 5’ linker DNA 20 bp
fragment with indicated molar ratios. Assignments of most perturbed residues in ARID are labeled. (C) Chemical shift
change (Δδ) of ARID residue backbone assignments upon binding of the 5’ linker DNA 20 bp fragment at 1:1 molar
ratio measured by NMR. ARID backbone assignments could not be reliably transferred to a subset of residues and
thus chemical shifts could not be determined (indicated by no values). Dashed lines indicate 25th, 50th, and 75th
percentile rankings, and residues are colored by a gradient from unperturbed (light blue) to significantly perturbed
(navy). (D) Demethylation kinetics of the H3K4me3 (1-21) substrate peptide by wild type and ARID mutant KDM5C1-
839 under single turnover conditions. Observed rates are fit to a cooperative kinetic model, with n denoting the Hill
coefficient. Unlike on the substrate nucleosome, the K101A/R107A ARID double mutation does not decrease catalytic
rate on the substrate peptide, but does increase overall catalytic efficiency. All error bars represent SEM of at least
two independent experiments (n ≥ 2).
!
0
1
10
0.0
0.5
1.0
1.5
[ARID] ( M)
Fraction unbound DNA
20 bp linker DNA
Kd
app
9.9 � 0.4 M
-
5' linker
3' linker
0
1
2
4
16
ARID
20 bp 5' linker DNA
8
25
15
35
50
200
bp
0
1
2
4
16
µM
ARID
20 bp 3' linker DNA
DNA
8
0
2
4
6
8
0
5
10
15
20
[KDM5C construct] (µM)
kobs (min-1)
H3K4me3 (1-21) demethylation
single turnover
kmax (min-1)
Km' ( M)
KDM5C1-839
KDM5C1-839 K101A/R107A
4.2 � 0.1
10.6 ��0.3
0.8 � 0.2
16.3 ��1.5
n
2.5
2.1
A
B
NE L EAQTRVK L NY L DQ I AK FWE I QGSS L K I PNVERR I L D L YS L SK I VVEEGGYEA I CKDRRWARVAQR L NYPPGKN I GS L L RSHYER I VYPYEMYQSGANL VQCN TRP FDNEEKDK
0.00
0.05
0.10
25%
50%
75%
(ppm)
ARID & 20 bp 5' linker DNA
(bound at 1:1 ratio)
73
188
D
C
1H-15N HSQC - ARID & 20 bp 5' linker DNA
1H (ppm)
15N (ppm)
6
7
8
9
10
105
110
115
120
125
130
ARID: DNA
1:0
1:0.1
1:0.25
1:0.4
1:0.6
1:1
R107
V105
K101
L152
V137
V174
I149
A138
R108
S151
S169
S155
N109
N148
R133
N104
D131
48
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
KDM5C
KDM5D
KDM5A
KDM5B
1
1
1
1
76
76
81
94
M E- - - - - - - - - - - - - - - - - P G S- - D D F L P P P EC P V F EP SWA E F R D P L GY I A K I R P I A EK SG I C K I R P P A DWQ P P F A V EV D N F R F T P R I Q R L N E L E
M E- - - - - - - - - - - - - - - - - P GC - - D E F L P P P EC P V F EP SWA E FQ D P L GY I A K I R P I A EK SG I C K I R P P A DWQ P P F A V EV D N F R F T P R VQ R L N E L E
M A GV - - - - - - - - - - - - - - GP GGY A A E F V P P P EC P V F EP SW E E F T D P L S F I GR I R P L A EK T G I C K I R P P K DWQ P P F A C EV K S F R F T P R VQ R L N E L E
M EA A T T L H P GP R P A L P L GGP GP L - G E F L P P P EC P V F EP SW E E F A D P F A F I H K I R P I A EQ T G I C K V R P P P DWQ P P F A C D V D K L H F T P R I Q R L N E L E
77
77
82
95
171
171
176
189
AQ T R V K L N Y L DQ I A K FW E I Q G S S L K I P N V ER R I L D L Y S L SK I V V E EGGY EA I C K D R RWA R V AQ R L N Y P P GK N I G S L L R SH Y ER I V Y P Y EM YQ SGA
AQ T R V K L N Y L DQ I A K FW E I Q G S S L K I P N V ER K I L D L Y S L SK I V I E EGGY EA I C K D R RWA R V AQ R L H Y P P GK N I G S L L R SH Y ER I I Y P Y EM FQ SGA
AM T R V R L D F L DQ L A K FW E LQ G ST L K I P V V ER K I L D L Y A L SK I V A SK GG F EM V T K EK KW SK V G SR L GY L P GK GT G S L L K SH Y ER I L Y P Y E L FQ SGV
AQ T R V K L N F L DQ I A K YW E LQ G ST L K I P H V ER K I L D L FQ L N K L V A E EGG F A V V C K D R KWT K I A T KM G F A P GK A V G SH I R GH Y ER I L N P Y N L F L SGD
172
172
177
190
258
260
263
283
N L VQ C N T R P F D N E EK D K EY K P H S I P L RQ SVQ P SK F N SY GR R A K R LQ P D P - - - - - EP T E ED I EK N P E L K K LQ I Y GA GP KMM G- L G LM A K D K - - T L R
N H VQ C N T H P F D N EV K D K EY K P H S I P L RQ SVQ P SK F S SY SR R A K R LQ P D P - - - - - EP T E ED I EK H P E L K K LQ I Y GP GP KMM G- L G LM A K D K D K T V H
S LM GVQM P N L D L K EK V E- - - - - - - P EV L ST D T Q T SP EP GT RM N I L P K R T R R V K T Q S E SGD V SR N T E L K K LQ I F GA GP K V V G- L AM GT K D K ED EV T
S L R C LQ K P N L T T D T K D K EY K P H D I PQ RQ SVQ P S ET C P P A R R A K RM R A EAM N I K I EP E ET T EA R T H N L R R - RM GC P T P K C EN EK EM K S S I KQ EP I E
259
261
264
284
353
343
322
338
K K D K EGP EC P P T V V V K E E L GGD V K V E ST SP K T F L E SK E E L SH SP EP C T KM T M R L R R N H SN AQ F I E SY V C RM C SR GD ED D K L L L C D GC D D N Y H I F C
K K V T - - - - C P P T V T V K D EQ SGGGN V S ST L L KQ H L - - - - - - - - S L EP C T K T T MQ L R K N H S SAQ F I D SY I CQ V C SR GD ED D K L L F C D GC D D N Y H I F C
R R R K - - - - - - - - V T N R SD A - - - - - - - - - - - - - - - - - - - - - - - - - - - - F NMQM RQ R K GT L SV N F V D L Y V CM F C GR GN N ED K L L L C D GC D D SY H T F C
R K D Y - - - - - - - - I V EN EK E- - - - - - - - - - - - - - - - - - - - - - - - - - - - - K P K SR SK K A T - - - N A V D L Y V C L L C G SGN D ED R L L L C D GC D D SY H T F C
354
344
323
339
448
438
417
433
L L P P L P E I P K GVWR C P K C VM A EC K R P P EA F G F EQ A T R EY T LQ S F G EM A D S F K A D Y F NM P V HM V P T E L V EK E FWR L V N S I E ED V T V EY GA D I H SK E
L L P P L P E I P R G I WR C P K C I L A EC KQ P P EA F G F EQ A T Q EY S LQ S F G EM A D S F K SD Y F NM P V HM V P T E L V EK E FWR L V S S I E ED V T V EY GA D I H SK E
L I P P L P D V P K GDWR C P K C V A E EC SK P R EA F G F EQ A V R EY T LQ S F G EM A D N F K SD Y F NM P V HM V P T E L V EK E FWR L V S S I E ED V I V EY GA D I S SK D
L I P P L H D V P K GDWR C P K C L AQ EC SK PQ EA F G F EQ A A R D Y T L R T F G EM A D A F K SD Y F NM P V HM V P T E L V EK E FWR L V ST I E ED V T V EY GA D I A SK E
449
439
418
434
543
533
512
528
F G SG F P V SD SK R H L T P E E E EY A T SGWN L N VM P V L EQ SV L C H I N A D I SGM K V PW L Y V GM V F SA F CWH I ED HW SY S I N Y L HWG EP K T WY GV P S L A A E
F G SG F P V SN SKQ N L SP E EK EY A T SGWN L N VM P V L DQ SV L C H I N A D I SGM K V PW L Y V GM V F SA F CWH I ED HW SY S I N Y L HWG EP K T WY GV P S L A A E
F G SG F P V K D GR R K I L P E E E EY A L SGWN L N NM P V L EQ SV L A H I N V D I SGM K V PW L Y V GM C F S S F CWH I ED HW SY S I N Y L HWG EP K T WY GV P SH A A E
F G SG F P V R D GK I K L SP E E E EY L D SGWN L N NM P VM EQ SV L A H I T A D I C GM K L PW L Y V GM C F S S F CWH I ED HW SY S I N Y L HWG EP K T WY GV P GY A A E
544
534
513
529
638
628
607
623
H L E EVM K K L T P E L F D SQ P D L L HQ L V T LM N P N T LM SH GV P V V R T NQ C A G E F V I T F P R A Y H SG F NQ GY N F A EA V N F C T A DW L P A GRQ C I EH Y R R L R R
H L E EVM KM L T P E L F D SQ P D L L HQ L V T LM N P N T LM SH GV P V V R T NQ C A G E F V I T F P R A Y H SG F NQ GY N F A EA V N F C T A DW L P A GRQ C I EH Y R R L R R
Q L E EVM R E L A P E L F E SQ P D L L HQ L V T I M N P N V LM EH GV P V Y R T NQ C A G E F V V T F P R A Y H SG F NQ GY N F A EA V N F C T A DW L P I GRQ C V N H Y R R L R R
Q L EN VM K K L A P E L F V SQ P D L L HQ L V T I M N P N T LM T H EV P V Y R T NQ C A G E F V I T F P R A Y H SG F NQ G F N F A EA V N F C T V DW L P L GRQ C V EH Y R L L H R
639
629
608
624
733
723
702
718
Y C V F SH E E L I C KM A A C P EK L D L N L A A A V H K EM F I M VQ E ER R L R K A L L EK G I T EA ER EA F E L L P D D ERQ C I K C K T T C F L SA L A C Y D C P D G L V C L SH
Y C V F SH E E L I C KM A A F P ET L D L N L A V A V H K EM F I M VQ E ER R L R K A L L EK GV T EA ER EA F E L L P D D ERQ C I K C K T T C F L SA L A C Y D C P D G L V C L SH
H C V F SH E E L I F KM A A D P EC L D V G L A AM V C K E L T LM T E E ET R L R E SV VQM GV LM S E E EV F E L V P D D ERQ C SA C R T T C F L SA L T C SC N P ER L V C L Y H
Y C V F SH D EM I C KM A SK A D V L D V V V A ST VQ K DM A I M I ED EK A L R ET V R K L GV I D S ERM D F E L L P D D ERQ C V K C K T T C FM SA I SC SC K P G L L V C L H H
734
724
703
719
828
818
797
813
I N D L C K C S S SRQ Y L R Y R Y T L D E L P AM L H K L K V R A E S F D T WA N K V R V A L EV ED GR K R S L E E L R A L E S EA R ER R F P N S E L LQQ L K N C L S EA EA C V SR
I N D L C K C S S SRQ Y L R Y R Y T L D E L P T M L H K L K I R A E S F D T WA N K V R V A L EV ED GR K R S F E E L R A L E S EA R ER R F P N S E L LQ R L K N C L S EV EA C I AQ
P T D L C P C PMQ K K C L R Y R Y P L ED L P S L L Y GV K V R AQ SY D T WV SR V T EA L SA N F N H K K D L I E L R VM L ED A ED R K Y P EN D L F R K L R D A V K EA ET C A SV
V K E L C SC P P Y K Y K L R Y R Y T L D D L Y PMM N A L K L R A E SY N EWA L N V N EA L EA K I N K K K S L V S F K A L I E E S EM K K F P D N D L L R H L R L V T Q D A EK C A SV
829
819
798
814
919
906
892
908
A L G L V SGQ EA GP H R V A G- - - - LQM T L T E L R A F L DQM N N L P C AM HQ I GD V K GV L EQ V EA YQ A EA R EA L A S L P S SP G L LQ S L L ER GRQ L GV EV P EAQ
V L G L V SGQ V A - - - RM D T - - - - PQ L T L T E L R V L L EQM G S L P C AM HQ I GD V K D V L EQ V EA YQ A EA R EA L A T L P S SP G L L R S L L ER GQQ L GV EV P EA H
AQ L L L SK KQ K H RQ SP D SGR T R T K L T V E E L K A F VQQ L F S L P C V I SQ A RQ V K N L L D D V E E F H ER AQ EAMM D ET P D S SK LQM L I DM G S S L Y V E L P E L P
AQQ L L N GK RQ T R Y R SGGGK SQ NQ L T V N E L RQ F V T Q L Y A L P C V L SQ T P L L K D L L N R V ED FQQ H SQ K L L S E ET P SA A E LQ D L L D V S F E F D V E L PQ L A
920
907
893
909
1014
1001
986
1002
Q LQ RQ V EQ A RW L D EV K R T L A P SA R R GT L A VM R G L L V A GA SV A P SP A V D K AQ A E LQ E L L T I A ERW E EK A H L C L EA RQ K H P P A T L EA I I R EA EN I P V
Q LQQQ V EQ AQW L D EV KQ A L A P SA H R G S L V I MQ G L L VM GA K I A S SP SV D K A R A E LQ E L L T I A ERW E EK A H F C L EA RQ K H P P A T L EA I I R ET EN I P V
R L KQ E LQQ A RW L D EV R L T L S- D PQQ V T L D VM K K L I D SGV G L A P H H A V EK AM A E LQ E L L T V S ERW E EK A K V C LQ A R P R H SV A S L E S I V N EA K N I P A
EM R I R L EQ A RW L E EVQQ A C L - D P S S L T L D DM R R L I D L GV G L A P Y SA V EK AM A R LQ E L L T V S EHWD D K A K S L L K A R P R H S L N S L A T A V K E I E E I P A
1015
1002
987
1003
1109
1096
1081
1097
H L P N I Q A L K EA L A K A R AW I A D V D E I Q N GD H Y P C L D D L EG L V A V GR D L P V G L E E L RQ L E LQ V L T A H SWR EK A SK T F L K K N SC Y T L L EV L C P C A D A G
H L P N I Q A L K EA L T K AQ AW I A D V D E I Q N GD H Y P C L D D L EG L V A V GR D L P V G L E E L RQ L E LQ V L T A H SWR EK A SK T F L K K N SC Y T L L EV L C P C A D A G
F L P N V L S L K EA LQ K A R EWT A K V EA I Q SG SN Y A Y L EQ L E S L SA K GR P I P V R L EA L PQ V E SQ V A A A R AWR ER T GR T F L K K N S SH T L LQ V L SP R T D I G
Y L P N GA A L K D SVQ R A R DW LQ D V EG LQ A GGR V P V L D T L I E L V T R GR S I P V H L N S L P R L ET L V A EVQ AWK EC A V N T F L T EN SP Y S L L EV L C P R C D I G
1110
1097
1082
1098
1196
1183
1172
1187
SD ST - K R SRWM EK E L - - G L Y K SD T E L L G L S- - - - - AQ D L R D P G SV I V A F K EG EQ K EK EG I LQ L R R T N SA K P SP L A S S ST A S ST T S I C V C GQ V L A G
SD ST - K R SRWM EK A L - - G L YQ C D T E L L G L S- - - - - AQ D L R D P G SV I V A F K EG EQ K EK EG I LQ L R R T N SA K P SP L A P S LM A S SP T S I C V C GQ V P A G
V Y G SGK N R R K K V K E L I EK EK EK D L D L EP L SD L E EG L E ET R D T AM V V A V F K ER EQ K E I EAM H S L R A A N L A K - - - - M T M V D R I E EV K F C I C R K T A SG
L L G L - K R KQ R K L K EP L P N GK K K ST K L E S L SD L ER A L T E SK ET A SAM A T L G EA R L R EM EA LQ S L R L A N EGK - - - - L L SP LQ D V D I K I C L CQ K A P A A
1197
1184
1173
1188
1291
1278
1259
1265
A GA LQ C D L CQ DW F H GR C V SV P R L L S SP R P N P T S SP L L AWW EWD T K F L C P L CM R SR R P R L ET I L A L L V A LQ R L P V R L P EG EA LQ C L T ER A I SWQ GR
V GV LQ C D L CQ DW F H GQ C V SV P H L L T SP K P S L T S SP L L AWW EWD T K F L C P L CM R SR R P R L ET I L A L L V A LQ R L P V R L P EG EA LQ C L T ER A I GWQ D R
F - M LQ C E L C K DW F H N SC V P L P K S S SQ K K G S- - - - - - - SWQ A K EV K F L C P L CM R SR R P R L ET I L S L L V S LQ K L P V R L P EG EA LQ C L T ER AM SWQ D R
P - M I Q C E L C R D A F H T SC V A V P S I SQ G L R - - - - - - - - - I W- - - - - - - L C P H C R R S EK P P L EK I L P L L A S LQ R I R V R L P EGD A L R YM I ER T V NWQ H R
1292
1279
1260
1266
1339
1326
1352
1324
A RQ A L A S ED V T A L L GR L A E- - L RQ R LQ A E- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - P R P E EP P N Y - - P A A P A SD P L R EG
A R K A L A S ED V T A L L RQ L A E- - L RQQ LQ A K - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - P R P E EA SV Y - - T SA T A C D P I R EG
A RQ A L A T D E L S SA L A K L SV - - L SQ RM V EQ A A R EK T EK I I SA E LQ K A A A N P D LQ GH L P S FQQ SA F N R V V S SV S S SP RQ T M D Y D D E ET D SD ED I R ET
AQQ L L S SGN L K F VQ D R V G SG L L Y SRWQ A SA G- - - - - - - - - - - - - - - - - - - - - - - - - - - - Q V SD T N K V - - - - - SQ P P GT T S F - - - S L P D DWD N R T S
1340
1327
1353
1325
1401
1385
1447
1348
SGK DM P - - - - - K VQ G L L E- - - - - - - - - - - - - - - - - - - - - - - - - - - N GD SV T SP EK V A P E EG SGK R D L E L L S S L L PQ - L T GP V L E L P EA T R A P L E E
SGN N I S- - - - - K VQ G L L E- - - - - - - - - - - - - - - - - - - - - - - - - - - N GD SV T SP ENM A P GK G S- - - D L E L L S S L L PQ - L T GP V L E L P EA I R A P L E E
Y GY DM K D T A SV K S S S S L EP N L F C D E E I P I K S E EV V T HMWT A P S F C A EH A Y S SA SK SC SQ G S ST P R KQ P R K SP L V P R S L EP P V L E L SP GA K AQ L E E
Y - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - L H SP F ST GR SC - - - - - - - - - - - - - - - - I P L - - - - - - - - - - H GV SP EV N E
1402
1386
1448
1349
1476
1461
1537
1437
LMM EGD L L EV T L D EN H S I WQ L LQ A GQ P P D L ER I R T L L E L EK A ER H G SR A R GR - - - - - A L ER R R R R - K V D R - - - - - - - - - - - - - - GG EGD D P A R E E
LMM EGD L L EV T L D EN H S I WQ L LQ A GQ P P D L D R I R T L L E L EK F EHQ G SR T R SR - - - - - A L ER R R R RQ K V DQ - - - - - - - - - - - - - - GR N V EN L VQQ E
LMM V GD L L EV S L D ET Q H I WR I LQ A T H P P S ED R F L H I M ED D SM E EK P L K V K GK - - - - D S S EK K R K R - K L EK V EQ L F G EGKQ K SK E L K KM D K P R K K K
L LM EAQ L LQ V S L P E I Q E L YQ T L L A K P SP AQQ T D R S SP V R P S S EK N D C - C R GK R D G I N S L ER K L K R - R L ER - - - - EG L S S ERW ER V K KM R T P K K K K
1477
1462
1538
1438
1515
1500
1632
1509
L EP - - - - - - - - - - - - - K R V R S S- - - - - - - - - - - - - - - - - GP EA E EVQ E E E E L - - - - - - - - - - E E ET GG EGP P A P I P T T G- - - - - - - - - - - - - - - -
LQ S- - - - - - - - - - - - - K R A R S S- - - - - - - - - - - - - - - - - G I M SQ V GR E E EH Y - - - - - - - - - - Q EK A D R ENM F L T P ST D H - - - - - - - - - - - - - - - -
L K L GA D K SK E L N K L A K K L A K E E ER K K K K EK A A A A K V E L V K E ST EK K R EK K V L D I P SK Y DW SGA E E SD D EN A V C A AQ N CQ R P C K D K V DWVQ C D GGC
I K L - - SH P K DM N N F - - K L ER ER - - - - - - - - - - - - - - - - - - - SY E L V R SA ET H S L P SD T SY S EQ ED S ED ED A I C P A V SC LQ P EGD EV DWVQ C D G SC
1516
1501
1633
1510
1560
1539
1690
1544
- - - - - - - - - - - SP ST Q ENQ N G L EP A EGT T SGP SA P F ST L T P R L H L P C PQQ P - PQQQ L - - - -
- - - - - - - - - - - SP F L K GNQ N S L - - - Q H K D SG S SA A C P S LM P L LQ L - - - SY S- D EQQ L - - - -
D EW F HQ V C V GV SP EM A EN ED Y I C I N C A K KQ GP V SP GP A P P P S F I M - - - SY K L PM ED L K ET S
NQW F HQ V C V GV SP EM A EK ED Y I C V R C T V K D A P - - - - - - - - - - - - - - - - SR K - - - - - - - - - -
JmjN
ARID
PHD1
PHD1
JmjC
JmjC
ZnF
ZnF
PHD2
PHD2
PHD3 (KDM5A/B)
PHD3 (KDM5A/B)
A
49
B
D
0
60
120
180
240
300
360
0.0
0.2
0.4
Time (sec)
Response (nm)
PHD1 & H3 peptides
H3 (1-18)
Ac-H3 (1-18)
H3 (5-18)
!
!
!
Association
Dissociation
H3 peptide
kobs1 (s
-1)
kobs2 (s
-1)
kobs1 (s
-1)
kobs2 (s
-1)
H3 (1-18)
2.79 ± 0.08
0.015 ± 0.0006
0.84 ± 0.01
0.015 ± 0.0009
Ac-H3 (1-18)
2.51 ± 0.08
0.016 ± 0.0011
1.63 ± 0.03
0.022 ± 0.0016
H3 (5-18)
2.14 ± 0.26
0.063 ± 0.0217
2.34 ± 0.48
0.015 ± 0.0040
!
0
100
200
300
400
500
600
700
800
900
1000
1100
1200
1300
1400
1500
0.0
0.5
1.0
Position
IUPRED Score
KDM5C
1
1560
JmjN ARID
PHD1 JmjC Zf
PHD2
KDM5C
C
0
5
10
0.00
0.02
0.04
0.06
0.08
[KDM5C construct] (µM)
kobs (min-1)
H3K4me3 nucleosome
kmax (min-1)
Km
app ( M)
KDM5C1-839
KDM5C1-839 linker
5.7 � 1.7
0.087 ��0.018
14.2 � 5.3
0.168 ��0.061
n
1.5
1.6
WT
�������
H
E
0
1
10
0.0
0.5
1.0
[KDM5C] ( M)
Fraction unbound nuc
KDM5C1-839
Kd
app
n
6.9 � 0.8 M
7.3 � 0.9 M
2.3
1.7
147 bp
H3K4me3 nuc
187 bp
H3K4me3 nuc
0
2
4
6
8
0
5
10
[KDM5C construct] ( M)
kobs (min-1)
H3K4me3 (1-21) demethylation
single turnover
kmax (min-1)
Km' ( M)
KDM5C1-839
KDM5C1-839 linker
4.2 � 0.1
10.6 ��0.3
4.9 � 1.3
9.7 ��1.3
n
2.5
1.0
50
1H-15N HSQC - PHD1 & KDM5C (199-218)
1H (ppm)
15N (ppm)
6
7
8
9
10
110
115
120
125
130
PHD: peptide
1:0
1:0.25
1:0.5
1:1
1:1.5
1:2.5
1:3.5
1:5
G344
V326
Y349
D346
C345
D347
L340
C342
R328
D343
F
N A Q F I E S Y V C R M C S R G D E D D K L L L C D G C D D N Y H I F C L L P P L P E I P K G V W R C P K C V M A E C K R
0.0
0.1
0.2
50%
75%
25%
(ppm)
PHD1 & KDM5C (199-218)
(bound at 1:5 ratio)
318
378
G
6
7
8
9
10
105
110
115
120
125
130
1H-15N HSQC - ARID & PHD1
1H (ppm)
15N (ppm)
ARID:PHD1
1:0
1:0.5
1:1
1:2
1:3
51
Figure S4. Related to Figure 4.
(A) Sequence alignment of human KDM5A-D with annotated domains. KDM5C has a different and extended linker
region between ARID and PHD1 (boxed in red). (B) IUPred profile [74] of predicted disorder in KDM5C (top) and
annotated domain architecture of KDM5C (bottom). The linker between ARID and PHD1 is predicted to be
disordered. (C) Demethylation kinetics of the H3K4me3 substrate nucleosome by wild type and KDM5C1-839 ∆linker
under single turnover conditions. Observed rates are fit to a cooperative kinetic model, with n denoting the Hill
coefficient. Deletion of the ARID-PHD1 linker does not significantly affect the catalytic efficiency of substrate
nucleosome demethylation. (D) Demethylation kinetics of the H3K4me3 (1-21) substrate peptide by wild type and
KDM5C1-839 ∆linker under single turnover conditions. Observed rates are fit to a cooperative kinetic model, with n
denoting the Hill coefficient. Similarly to nucleosomes, deletion of the ARID-PHD1 linker does not significantly affect
the catalytic efficiency of substrate peptide demethylation. (E) Binding kinetic trace of immobilized Avitag-PHD1
binding to H3 (1-18), N-terminally acetylated H3 (1-18), and H3 (5-18) tail peptides measured by bio-layer
interferometry. Observed rates (kobs) of association and dissociation are obtained by fitting kinetic traces to a two
phase exponential function. Recognition of the H3 tail by PHD1 does not strongly depend on the H3 N-terminus but
does depend on the first 4 residues of H3 (ARTK). (F) 2D 1H-15N HSQC spectra of ARID titrated with increasing
amounts of PHD1 with indicated molar ratios. PHD1 does not appear to bind the ARID domain. (G) 2D 1H-15N HSQC
spectra of PHD1 titrated with increasing amounts of KDM5C (199-218) peptide with indicated molar ratios (top).
Assignments of most perturbed residues in PHD1 are labeled. Corresponding chemical shift change (Δδ) of PHD1
residues upon binding of the KDM5C (199-218) peptide at 1:5 molar ratio (PHD:peptide) (bottom). Dashed lines
indicate 25th, 50th, and 75th percentile rankings. (H) Nucleosome binding curves of KDM5C1-839 binding to substrate
nucleosomes with and without 20 bp flanking DNA. Nucleosome binding curves were measured by EMSA and fit to a
cooperative binding model to determine apparent dissociation constants (Kdapp), with n denoting the Hill coefficient. (I)
Binding curves of PHD1 mutant KDM5C1-839 binding to unmodified and substrate nucleosomes with and without 20 bp
flanking DNA. All error bars represent SEM of at least two independent experiments (n ≥ 2).
I
0
1
10
0.0
0.5
1.0
[KDM5C D343A] ( M)
Fraction unbound nuc
KDM5C1-839 D343A
Kd
app
n
8.4 � 0.5 M
3.2 � 0.4 M
2.5
2.3
147 bp
nuc
187 bp
nuc
0
1
10
0.0
0.5
1.0
[KDM5C D343A] ( M)
Fraction unbound nuc
KDM5C1-839 D343A
Kd
app
n
8.9 � 0.7 M
4.7 � 0.3 M
2.2
2.4
147 bp
H3K4me3 nuc
187 bp
H3K4me3 nuc
52
Figure S5. Related to Figure 5.
(A) Unmodified core nucleosome binding by KDM5C1-839 ∆AP wild type and A388P. Nucleosome binding curves were
measured by EMSA and fit to a cooperative binding model to determine apparent dissociation constants (Kdapp), with
n denoting the Hill coefficient. Due to unattainable saturation of binding, a lower limit for the dissociation constant is
presented. The A388P mutation does not enhance nucleosome binding in the absence of the ARID and PHD1 region,
indicating this region in KDM5C is altered by the A388P mutation to enable enhanced binding. (B) Demethylation
kinetics of the H3K4me3 (1-21) substrate peptide by KDM5C1-839 ∆AP wild type and A388P under multiple turnover
conditions measured by a formaldehyde release based kinetic assay. The A388P mutation reduces demethylase
activity of the catalytic domain alone, indicating distal structural disruption of the catalytic domain by this mutation. (C)
Binding curves of KDM5C1-839 D87G and A388P binding to substrate nucleosomes with and without 20 bp flanking
DNA. All error bars represent SEM of at least two independent experiments (n ≥ 2).
0
200
400
0
1
2
[H3K4me3 (1-21)] ( M)
Rate (min-1)
H3K4me3 (1-21) demethylation
kcat (min-1)
Km ( M)
KDM5C1-839 AP
KDM5C1-839 AP A388P
3.6 � 0.7
2.22 ��0.07
26.6 � 2.0
0.95 ��0.02
A
B
1
839
JmjN JmjC Zf
KDM5C1-839 ���
���������
�����
C
Figure 5. supplemental
0
1
10
100
0.0
0.5
1.0
[KDM5C construct] ( M)
Fraction unbound nuc
Unmodified nucleosome
Kd
app
n
�35 � 3 M
�45 � 3 M
1.4
1.7
KDM5C1-839 AP
KDM5C1-839 AP A388P
0
0.1
1
10
0.0
0.5
1.0
[KDM5C D87G] ( M)
Fraction unbound nuc
KDM5C1-839 D87G
Kd
app
n
2.3 � 0.2 M
1.3 � 0.1 M
1.8
2.1
147 bp
H3K4me3 nuc
187 bp
H3K4me3 nuc
0
1
10
0.0
0.5
1.0
[KDM5C A388P] ( M)
Fraction unbound nuc
KDM5C1-839 A388P
Kd
app
n
3.7 � 0.6 M
1.8 � 0.2 M
2.2
2.2
147 bp
H3K4me3 nuc
187 bp
H3K4me3 nuc
| 2022 | Chromatin sensing by the auxiliary domains of KDM5C regulates its demethylase activity and is disrupted by X-linked intellectual disability mutations | 10.1101/2022.01.13.476263 | [
"Ugur Fatima S.",
"Kelly Mark J. S.",
"Fujimori Danica Galonić"
] | null |
Removal of rare amplicon sequence variants from 16S rRNA gene
sequence surveys biases the interpretation of community structure
data
Patrick D. Schloss†
† To whom correspondence should be addressed:
5
pschloss@umich.edu
Department of Microbiology and Immunology University of Michigan Ann Arbor, MI 48109
Research article format
1
Abstract
Methods for remediating PCR and sequencing artifacts in 16S rRNA gene sequence collections are in con-
10
tinuous development and have significant ramifications on the inferences that can be drawn. A common
approach is to remove rare amplcon sequence variants (ASVs) from datasets. But, the definition of rarity is
generally selected without regard for the number of sequences in the samples or the variation in sequencing
depth across samples within a study. I analyzed the impact of removing rare ASVs on metrics of alpha and
beta diversity using samples collected across 12 published datasets. Removal of rare ASVs significantly
15
decreased the number of ASVs and operational taxonomic units as well as their diversity. Furthermore,
their removal increased the variation in community structure between samples. When simulating a known
effect size, removal of rare ASVs reduced the power to detect the effect relative to not removing rare ASVs.
Removal of rare ASVs did not affect the false detection rate when samples were randomized to simulate a
null model. However, the false detection rate increased when rare ASVs were removed using a null distri-
20
bution and assignment of samples to simulated treatment groups according to their sequencing depth. The
false detection rate did not vary when rare ASVs were retained. This analysis demonstrates the problems
inherent in removing rare ASVs. Researchers are encouraged to retain rare ASVs, to select approaches
that minimize PCR and sequencing artifacts, and to use rarefaction to control for uneven sequencing effort.
2
Importance
25
Removing rare amplicon sequence variants (ASVs) from 16S rRNA gene sequence collections is an ap-
proach that has grown in popularity for limiting PCR and sequencing artifacts. Yet, it is unclear what impact
an abundance-based filter has on downstream analyses. To investigate the effects of removing rare ASVs,
I analyzed the community distributions found in the samples of 12 published datasets. Analysis of these
data and simulations based on them showed that removal of rare ASVs distorts the representation of mi-
30
crobial communities. This has the effect of artificially making it more difficult to detect differences between
treatment groups. Also of concern was the observation that if sequencing depth is confounded with the
treatment, then the probability of falsely detecting a difference between the treatment groups increased
with the removal of rare ASVs. The practice of removing rare ASVs should stop, lest researcher adversely
affect the interpretation of their data.
35
3
Introduction
16S rRNA gene sequencing is a mainstay of microbial community analysis (1). Two elements that are
held in tension in the analysis of 16S rRNA gene sequence data are how to adequately remove PCR and
sequencing artifacts while decreasing the granularity of the taxonomic level that is used in the analysis.
When coarse taxonomic levels (e.g. genus level) are used, the effects of artifacts are minimized since the
40
genetic breadth of the level is wider than the diversity of artifacts. Conversely, with fine taxonomic levels
(e.g. amplicon sequence variants; ASVs) the effects of artifacts are significant since each artifact may
represent a new ASV.
Numerous studies have attempted to address the problem of removing or “denoising” artifacts from data
generated using Illumina’s MiSeq platform. In one approach, paired sequence reads are aligned and any
45
discrepencies between the reads are resolved based on the difference in quality score for the position in
question (2–4); quality scores are also used to curate single reads (5). In addition, a polishing step is
often used to identify ASVs based on the frequency and similarity of sequences (2, 3, 6). In a second
approach, the quality scores and types of errors are modelled to cluster sequence reads directly into ASVs
(7). Regardless of the approach, many pipelines advocate for abundance-based screening where rare
50
sequences are removed from each dataset prior to outputting the sequence data as ASVs (3, 6, 7). Some
algorithms recommend removing all ASVs that appear one (i.e. singletons) (7), eight (3), or ten (6) or
fewer times; these pipelines also vary in whether the minimum abundance threshold should be applied to
individual samples (6, 7) or the pool of samples in a study (3). A notable exception, the mothur-based
pipeline discourages the practice of removing rare sequences (2). After assigning sequences to ASVs,
55
ASVs are often analyzed as a taxonomic unit or clustered to generate operational taxonomic units (OTUs)
or phylotypes.
The abundance-based screening approach assumes that rare ASVs are more likely to be artifacts than
more abundant ASVs. Sequencing of mock communities confirms that artifacts tend to be rare (2, 5).
Proponents of abundance-based screening point to their ability to obtain the correct number of ASVs, OTUs,
60
or phylotypes with data generated from sequencing mock communities when rare ASVs are removed (5).
However, this approach effectively overfits the curation pipeline to data generated from a phylogenetically
simple community with an atypical community distribution that is often sequenced to a coverage that is not
achieved with biological samples. It is necessary to think more deeply about the practice of abundance-
based screening.
65
The minimum abundance thresholds that have been proscribed were developed and applied without regard
4
for the number of sequences generated from each sample. These recommendations appear to assume
that the number of reads per sample is consistent within and between studies. However, it is common that
the number of sequences generated from each sample may vary by two or three orders of magnitude
(e.g. Table 1 and Figure S1).
An ASV that appears once in a sample with 2,000 sequences is more
70
trustworthy than an ASV that appears once in a sample with 100,000 sequences since it has a 50-fold
higher relative abundance. But, according to the pipeline recommendations, they are treated as being
equally trustworthy. Rather than removing rare ASVs, the approach taken by the mothur pipeline applies
the classical ecological approach of rarefaction. Each sample is rarefied to the same sequencing depth so
that the number of artifacts that appears in each sample is controlled.
75
Experience sequencing biological samples also demonstrates that there are bona fide ASVs that may have
an abundance below the proscribed threshold. For example, the abundance of an ASV may be below the
threshold in some samples or time points and above the threshold in others. However, rarity, both in terms of
prevalence and incidence, is an important ecological concept (8). Removing rare ASVs likely hinders one’s
to ability to make inferences about the dynamics and nature of the populations that rare ASVs represent.
80
Furthermore, removing ASVs whose abundances are below the proscribed threshold also potentially biases
the community structure of the samples.
In the current study, I used published sequence data from 12 studies to investigate the nature of rare ASVs
(i.e. those that appear 10 or fewer times) and the effect that removing them has on downstream analysis
of microbial communities. The analysis was also performed using traditional OTUs, where ASVs subjected
85
to abundance-based screening were clustered such that the ASVs within an OTU were no more than 3%
different from each other. The results reject the assumptions built into abundance-based screening and
highlight the problems inherent in removing rare ASVs.
Results
Datasets. I collected 12 publicly available datasets that used the Illumina MiSeq platform to sequence
90
the V4 region of the 16S rRNA gene from a variety of environments (Table 1; Figure S1). After removing
poor quality and chimeric ASVs and samples that had uncharacteristically low number of sequences for
the dataset, these datasets included between 7 and 490 samples. The median number of sequences for
each dataset ranged between 6,477 and 193,464. Strikingly, aside from the relatively small marine and
soil datasets, the difference between the sample with the fewest sequences and the sample with the most
95
sequences for each dataset varied by between 7.4 and 96.6-fold.
5
The nature of singletons. Removal of rare ASVs is commonly justified as a method of removing ASVs
that are artifacts. If such ASVs are artifacts, then one would expect the number of singleton ASVs to ac-
cumulate with sequencing depth. Contrary to this expectation, the median percentage of sequences that
were discarded when singleton ASVs were removed from each dataset varied between 0.42 and 22.23%
100
(bioethanol and seagrass). In addition, with the exception of the samples from the marine and sediment
datasets (Spearman correlation, P>0.05), the fraction of singleton ASVs in samples was negatively cor-
related with the number of sequences in each sample with a range between -0.27 and -0.87 (rice and
bioethanol) (Figure 1A). This showed that with additional sequencing, the probability of seeing singleton
ASVs in multiple samples was greater than the probability of generating an artifact. This suggests that the
105
singleton ASVs are not as likely to be artifacts as previously thought. Furthermore, if singleton ASVs were
artifacts, then one would not expect to find them in other samples from the same dataset. In fact, singleton
ASVs from samples with fewer sequences were often found in samples with more sequences. At least
50% of the singleton ASVs found in the samples from the mice, rice, seagrass, and stream datasets were
found in another sample from the same dataset (Figure 1B). Considering the likelihood of finding an ASV
110
duplicated in another sample is confounded by the number of samples and inter-sample diversity, the high
coverage of singleton ASVs in these datasets was remarkable. The correlation between the number of se-
quences in a sample and the fraction of that sample’s singleton ASVs that were covered by another sample
in the dataset was significant and negative for 9 of the datasets ranging between -0.31 and -0.84 for the rice
and seagrass datasets, respectively (Figure 1C). The negative correlation indicated that the singleton ASVs
115
in the smaller samples were more likely to be covered by ASVs in the larger samples. Among the three
datasets without a significant correlation (Spearman correlation, P>0.05), the marine and soil datasets had
the fewest samples in our collection and the stream dataset already had a high level of coverage regardless
of the number of sequences. Contrary to the common motivation for removing rare ASVs, these results in-
dicate that this practice disproportionately impacts samples with fewer sequences and likely removes more
120
non-artifact ASVs than those that are artifacts.
The impact of removing rare ASVs on the information represented in each sample. Removing rare
ASVs will reduce the richness of ASVs (i.e. the number of ASVs per sample) and increase the relative abun-
dance of the remaining ASVs. To quantify the effect of removing rare ASVs on the information contained
within each sample, I varied the minimum abundance threshold to simulate removing ASVs of varying rarity
125
from each sample. The richness of ASVs in each sample decreased by between 34.4 and 86.2% (per-
omyscus and soil) when removing those ASVs that only appeared once and by between 76.0 and 95.6%
(sediment and soil) when removing those that appeared ten or fewer times from each sample (Figure 2A).
6
Similarly, the Shannon diversity decreased by between 1.8 and 15.9% (human and soil) when removing
ASVs that only appeared once and by between 5.4 and 35.4% (human and seagrass) when removing
130
ASVs that appeared ten or fewer times from each sample (Figure 2B). Next, I assigned the ASVs to OTUs,
which were defined as a group of ASVs that were more than 97% similar to each other to assess the impact
of removing rare ASVs on higher level taxonomic groupings that are commonly used in microbial ecology
studies. Although pooling similar ASVs into OTUs reduced the impact of removing the rare ASVs relative
to the ASV-based analysis, the minimum abundance threshold still decreased the richness of OTUs and
135
the diversity decreased relative to the full community (Figure S2AB). In contrast to the richness and diver-
sity measurements, the Kullback–Leibler divergence compares the relative abundance of specific ASVs or
OTUs between representations of the community. I calculated the Kullback–Leibler divergence between the
full communities and those where rare ASVs were removed. As the threshold for removing ASVs increased,
the amount of information lost also increased for both ASVs and OTUs (Figure 2C and Figure S2C). The
140
relative loss of information was generally smaller for OTUs than than it was for ASVs. Removing rare ASVs,
regardless of abundance threshold, had profound impacts on the representation of the communities.
Removing treatment group effects from community data. Because treatment effects often affect a
sample’s diversity and inter-sample variation, I generated null distributions for each study by randomizing,
without replacement, the number of times each ASV was observed in each sample such that the total
145
number of sequences in each sample and the total number of times each ASV was observed across all
samples in the study was the same as was originally observed. This effectively made every community
in a study a statistical sample of the study-wide composite community distribution. For example, after this
procedure, each of the 490 samples from the human dataset would be expected to have the same richness
and diversity of ASVs and one would not expect to find treatment-based effects between the samples.
150
Because of the risk of bias if only one representation of the null distribution was generated, I generated 100
randomized datasets for each study. The trends between removing rare ASVs and the richness, diversity,
and information loss that were identified using the observed community distribution data were also identified
with the data from the null distribution; however, the losses were larger when using the null distribution data
(Figure S3). The null distribution data were used in the remainder of the study to minimize the risk of bias.
155
The impact of removing rare ASVs on the information represented between samples. Considering
the loss of richness, diversity, and information when a community has its rarest ASVs removed, it seemed
likely that the relationship between communities would also be altered. To assess the impact of remov-
ing rare ASVs on measures of alpha diversity between samples I calculated the coefficients of variation
(COVs, i.e. the standard deviation divided by the mean) for richness and diversity for each study at multiple
160
7
abundance thresholds. The COVs for the richness of ASVs across the studies after removing singletons
were between 3.6 and 32.7-times larger than they were without removing singleton ASVs (mice and stream;
Figure 3A). Similarly, the COVs for the diversity of ASVs were between 1.8 and 20.4-times larger when sin-
gletons were removed than when they were not removed (mice and rice; Figure 3B). To assess the impact
of removing rare ASVs on measures of beta diversity between samples, I calculated the COVs of the Bray-
165
Curtis distances between samples within the same study at multiple abundance thresholds. The COVs
between Bray-Curtis distances within a study when singletons were removed was between 1.3 and 18.6-
times larger than when they were not removed (mice and stream; Figure 3C). When ASVs were clustered
into OTUs the difference in COVs was less than it was for the ASVs (Figure S4). These results indicate that
removing rare ASVs increases the dissimilarity between samples, which could have a significant impact on
170
the statistical power to detect differences between treatment groups.
The impact of removing rare ASVs on the ability to detect statistically significant differences be-
tween treatment groups. To test the effect of increased inter-sample variation, I randomly assigned sam-
ples to one of two treatment groups. In the first treatment group, communities were randomly sampled
from the null distribution as described above. For the second treatment group, I randomly selected 10% of
175
the ASVs in the pooled study distribution to increase their abundance by 5%. I randomly generated 100
simulated sets of treatment groups and samples. I then tested the ability to detect a difference between
the two treatment groups using alpha and beta diversity metrics. The fraction of significant tests was a
measurement of the statistical power to detect the difference between the treatment groups. When con-
sidering the differences in richness and diversity, the marine dataset yielded no simulated sets that were
180
statistically significant, which was likely due to the small number of samples in the study (N=7). Among the
remaining datasets, the power to detect a difference in the richness of ASVs ranged between 0.10 and 0.49
(sediment and stream) and between 0.10 and 0.53 (rainforest and stream) to detect a difference in diversity
when using a Wilcox test (Figure 4A). When singleton ASVs were removed, the power to detect a differ-
ence in the richness of ASVs dropped by between 27.3 and 92.9% (bioethanol and soil) and by between
185
40.0 and 93.3% for their diversity(rainforest and soil; Figure 4B). The effect of removing rare ASVs on the
richness of OTUs and their diversity was similar (Figure S5AB). I used the Bray-Curtis dissimilarity index
to compare the simulated communities within each dataset and calculated the power to detect differences
between the two simulated treatment groups using the analysis of molecular variance (Figure 4C and S5C).
Without removing rare sequences, the power to detect a difference between the two simulated treatment
190
groups varied between 0.41 and 1.00 (rainforest and rainforest). Aside from the bioethanol, human, and
mice datasets, the power to detect differences dropped by between 6.5 and 64.0% (soil and rice) when
8
singletons were removed. However, when ASVs that occurred 10 or fewer times were removed from each
sample, the power to detect differences dropped by 12.0 and 97.2% (human and peromyscus); similar re-
sults were observed when ASVs were clustered into OTUs. Removing rare ASVs reduced the ability to
195
detect simulated treatment effects using metrics commonly used to compare microbial communities.
The impact of removing rare ASVs on the probability of falsely detecting a difference between treat-
ment groups. I next asked whether removing rare ASVs could lead to falsely claiming that a treatment
effect had a significant effect on community diversity and structure. First, I sampled sequences from the
null distribution for each dataset and randomly assigned each sample to one of two treatment groups and
200
determined the richness and diversity of ASVs and OTUs. Testing at an experiment-wise error rate of 0.05,
I expected 5% of the iterations for each dataset to yield a significant test result. Indeed, there was no
evidence that removing rare ASVs resulted in an inflated experiment-wise error rate. The average fraction
of significant tests did not meaningfully vary from 0.05 across the minimum abundance threshold, dataset,
metric of describing sample alpha-diversity, or whether the abundance of ASVs or OTUs were used (Figure
205
5A and S6A). Similarly, the average fraction of significant tests did not meaningfully vary from 0.05 when
using analysis of molecular variance to compare communities using Bray-Curtis distances (Figure 5A and
S6A). Second, I again sampled sequences from the null distribution, but assigned samples to one of two
treatment groups based on the number of sequences in each sample. The samples with fewer than the
median number of sequences for the dataset were assigned to one group and those with more than the
210
median were assigned to the other. This exaggerated bias has been observed in comparisons of the lung
and oral microbiota because of the larger number of non-specific amplicons that can be sequenced from
lung samples relative to those in the oral cavity leading to a significant difference in sequencing depth be-
tween treatment groups (9). When rare sequences were not removed, the fraction of significant tests did
not differ from 5% for comparing the richness, their diversity, or Bray-Curtis distances (Figure 5B and S6B).
215
However, when rare taxa of any frequency were removed, the probability of falsely detecing a difference as
signifiant increased with the definition of rarity (Figure 5B and S6B). Not including the small marine dataset,
the average fraction the average fraction of falsely detecting a difference across datasets when only sin-
gletons were removed was 92.45%. If there is any relationship between the number of sequences and the
treatment group, the risk of falsely rejecting the null hypothesis is inflated when researchers use the strategy
220
of removing rare sequences. The most conservative approach is to not remove low abundance sequences.
9
Discussion
Removing rare sequences from 16S rRNA gene sequence data is a common practice that is used as a
heuristic to help remove residual PCR artifacts and low quality sequences. In this analysis, I have shown
that rare sequences are more common in samples with shallow sequencing than in those with deep se-
225
quencing and that rare sequences are frequently observed in multiple samples. These observations sug-
gest that many of the sequences being removed are actually good sequences. Becuase rarity is often
defined by a fixed number of observations per sample (e.g. sequences that only appear once in a sample,
regardless of the size of the sample), removing rare sequences has a disproportionate impact on sam-
ples with fewer sequences. Removing rare sequences resulted in a reduction in the alpha diversity and a
230
pronounced change in the structure of individual samples. The effect was an increase in the differences
observed between samples, which made it more difficult to detect differences between treatment groups
when differences actually existed. Furthermore, if the number of reads per sample was confounded with the
treatment groups, then removing rare sequences increased the probability of falsely detecting a difference
between the treatment groups. The practice of removing rare sequences from samples should be stopped.
235
The practice of removing rare sequences from samples seems to be a response to researchers prioritizing
the number of reads and length of sequences over their quality (5). Previous work has shown that assembly
of fully overlapping sequence reads results in the lowest sequencing error rates (2). The studies highlighted
in this analysis sequenced the V4 region of the 16S rRNA gene using the Illumina MiSeq sequencing plat-
form with their version 2 chemistry. The resulting data consists of two 250 nt reads that span a region that is
240
about 253 nt long. In contrast, there has been a movement to sequence longer regions with similar chem-
istry resulting in less overlap between the sequencing reads (10, 11). Alternatively, others have prioritized
increasing the number of sequences per sample by sequencing the V4 region, but with paired or single
reads that are 150 nt long (12, 13). Both practices result in a significantly higher error rate for the resulting
assmebly. Instead, researchers should prioritize the quality over the quantity and length of their data. For
245
these reasons, this analysis did not use lower quality data generated by alternative methods.
The impacts of removing rare sequences on the representation of communities are caused by the uneven
impact of applying a single abundance threshold across all samples. The number of sequences per sample
in the datasets highlighted in this analysis varied by as much as 100-fold (Table 1; Figure S1). In practice,
I suspect this range is actually larger since researchers may have opted against depositing samples with
250
fewer reads into the SRA. If the range in read coverage for a study is 100-fold, a singleton in a small
sample would have had a comparable relative abundance as a sequence that appeared 100 times in a
10
more densely sequenced sample. However, it would have been removed from the smaller sample and not
the larger sample. Thus, applying a single abundance threshold disproportionately impacted the samples
with fewer sequences. This seems to contradict the purpose of removing rare sequences since a singleton
255
in the smaller sample is more reliable than the singleton in the larger sample. A superior approach to
removing rare sequences is to use rarefaction to conrol for uneven sampling.
Even with the best sequencing approaches, PCR artifacts and sequencing errors persist. This may account
for some sequences not being observed other samples. Although this lack of inter-sample coverage could
have been due to treatment effects and natural variation, it is likely that some portion of the sequences that
260
were unique to samples were artifacts or contained errors. Researchers are encouraged to use rarefaction
to control for uneven sampling and the presence of spurious sequences. Previous work sequencing mock
communities has shown that the number of spurious sequences increases with sequencing depth (2, 14).
By rarefying data to a common number of sequences per sample, the number of spurious sequences can
be controlled. As shown in the data I presented, which was rarefied to a common number of reads per
265
sample within a dataset, when rare sequences were not removed the power to detect differences was the
highest and the false discovery rate was the expected 5% (Figures 5 and S6).
In addition to considerations of how to control for the presence of spurious sequences, researchers also
need to be mindful of how to interpret the results of their work. Because every dataset will contain residual
sequences that are spurious, measures of richness and diversity should be made on a relative basis. For
270
example, pronouncements that communities from an environment or treatment contain a specific number of
taxa are problematic. Instead, we should limit ourselves to indicating that samples from one treatment group
has more taxa than another without using absolute values of richness. Furthermore, we must take caution in
interpreting rare taxa. Although the data from the studies highlighted here suggest that most rare sequences
are not spurious, it is likely that some are. Therefore, researchers must approach rare sequences with
275
more skepticism than more abundant sequences. Reseachers should seek out other methods to confirm
inferences that they make about rare sequences. This is a standard that should be applied regardless of
their abundance (15).
How to curate and interpret rare sequences has been a significant challenge since microbial ecologists
transitioned away from Sanger sequencing of samples (16–18). Although the extent of the “rare biosphere”
280
is still an open question, it is important to appreciate the importance of rare populations in all communities.
Populations can be numerically rare but ubuiquitous or abundant and limited in their geographic range.
Alternatively, they can be numerically rare but temporally common or abundant but present infrequently.
Removing sequences from any of these settings will limit our ability to study the role of such populations or
11
the processes that drive their patchy distributions that are so common with microbial communities (8).
285
Materials and Methods
Data curation and analysis. To insure the highest possible data quality, datasets were limited to those
where the 500 cycle version 2 MiSeq chemistry was used to sequence the amplicons. The paired 250
nt reads resulted in near complete 2-fold sequencing coverage of every nucleotide in the ca.
253 nt-
long region. This region and sequencing platform were selected because previous work has shown that a
290
standard data analysis pipeline in mothur results in a sequencing error rate below 0.02% (2). All sequence
data were obtained from the Sequence Read Archive (SRA) and processed using a standard mothur-based
sequencing pipeline that resulted in ASVs as generated by the pre.cluster algorithm using a threshold of 2
nt (2, 19). ASVs were assigned to OTUs using a 3% distance threshold using mothur’s cluster function with
the OptiClust algorithm (20). To minimize the effects of uneven sampling effort, samples were rarefied to the
295
number of sequences in the smallest sample for each dataset. Because metrics of alpha diversity did not
consistently follow a normal distribution, I used the non-parametric Wilcoxon rank test as implemented in R.
For comparisons of Bray-Curtis distances, the amova function within mothur was used, which implements
the analysis of molecular variance algorithm (21).
Reproducibility. All analyses were performed using mothur (version 1.44.1) and R (version 4.0.2) with
300
the tidyverse (version 1.3.0), broom (version 0.7.0), data.table (version 1.13.4), and cowplot (version 1.1.0)
packages. All of the code that was used in this analysis as well as the Makefile for running the analysis
are available at https://github.com/SchlossLab/Schloss_Singletons_XXXXX_2019 along with the complete
version history of the project.
Acknowledgements
305
I endebted to the researchers who developed the 12 datasets used in this study for depositing their se-
quence data into the Sequence Read Archive. This work was supported in part by funding from the National
Institutes of Health (U01AI124255, P30DK034933, R01CA215574).
12
References
1.
Knight R, Vrbanac A, Taylor BC, Aksenov A, Callewaert C, Debelius J, Gonzalez A, Kosciolek
T, McCall L-I, McDonald D, Melnik AV, Morton JT, Navas J, Quinn RA, Sanders JG, Swaf-
ford AD, Thompson LR, Tripathi A, Xu ZZ, Zaneveld JR, Zhu Q, Caporaso JG, Dorrestein
PC. 2018. Best practices for analysing microbiomes. Nature Reviews Microbiology 16:410–422.
doi:10.1038/s41579-018-0029-9.
310
2.
Kozich JJ, Westcott SL, Baxter NT, Highlander SK, Schloss PD. 2013. Development of a dual-
index sequencing strategy and curation pipeline for analyzing amplicon sequence data on the MiSeq
Illumina sequencing platform. Applied and environmental microbiology 79:5112–5120.
3.
Edgar RC. 2016. UNOISE2: Improved error-correction for illumina 16S and ITS amplicon sequenc-
ing. bioRxiv. doi:10.1101/081257.
4.
Masella AP, Bartram AK, Truszkowski JM, Brown DG, Neufeld JD. 2012. PANDAseq: Paired-end
assembler for illumina sequences. BMC Bioinformatics 13:31. doi:10.1186/1471-2105-13-31.
5.
Bokulich NA, Subramanian S, Faith JJ, Gevers D, Gordon JI, Knight R, Mills DA, Caporaso
JG. 2013. Quality-filtering vastly improves diversity estimates from illumina amplicon sequencing.
Nature Methods 10:57–59. doi:10.1038/nmeth.2276.
6.
Amir A, McDonald D, Navas-Molina JA, Kopylova E, Morton JT, Xu ZZ, Kightley EP, Thompson
LR, Hyde ER, Gonzalez A, Knight R. 2017. Deblur rapidly resolves single-nucleotide community
sequence patterns. mSystems 2. doi:10.1128/msystems.00191-16.
315
7.
Callahan BJ, McMurdie PJ, Rosen MJ, Han AW, Johnson AJA, Holmes SP. 2016. DADA2:
High-resolution sample inference from illumina amplicon data.
Nature Methods 13:581–583.
doi:10.1038/nmeth.3869.
8.
Reid A, Buckley M. 2011. The rare biosphere.
9.
Morris A, Beck JM, Schloss PD, Campbell TB, Crothers K, Curtis JL, Flores SC, Fontenot
AP, Ghedin E, Huang L, Jablonski K, Kleerup E, Lynch SV, Sodergren E, Twigg H, Young VB,
Bassis CM, Venkataraman A, Schmidt TM, Weinstock GM. 2013. Comparison of the respiratory
microbiome in healthy nonsmokers and smokers. American Journal of Respiratory and Critical Care
Medicine 187:1067–1075. doi:10.1164/rccm.201210-1913oc.
10.
Fadrosh DW, Ma B, Gajer P, Sengamalay N, Ott S, Brotman RM, Ravel J. 2014. An improved
dual-indexing approach for multiplexed 16S rRNA gene sequencing on the illumina MiSeq platform.
Microbiome 2:6. doi:10.1186/2049-2618-2-6.
13
11.
Holm JB, Humphrys MS, Robinson CK, Settles ML, Ott S, Fu L, Yang H, Gajer P, He X, McComb
E, Gravitt PE, Ghanem KG, Brotman RM, Ravel J. 2019. Ultrahigh-throughput multiplexing and
sequencing of >500-base-pair amplicon regions on the Illumina HiSeq 2500 platform. mSystems 4.
doi:10.1128/msystems.00029-19.
320
12.
Thompson LR, Jon G. Sanders and, McDonald D, Amir A, Ladau J, Locey KJ, Prill RJ, Tripathi
A, Gibbons SM, Ackermann G, Navas-Molina JA, Janssen S, Kopylova E, Vázquez-Baeza Y,
González A, Morton JT, Mirarab S, Xu ZZ, Jiang L, Haroon MF, Kanbar J, Zhu Q, Song SJ,
Kosciolek T, Bokulich NA, Lefler J, Brislawn CJ, Humphrey G, Owens SM, Hampton-Marcell
J, Berg-Lyons D, McKenzie V, Fierer N, Fuhrman JA, Clauset A, Stevens RL, Shade A, Pollard
KS, Goodwin KD, Jansson JK, Gilbert JA, Knight R. 2017. A communal catalogue reveals earth’s
multiscale microbial diversity. Nature 551:457–463. doi:10.1038/nature24621.
14
13.
McDonald D, Hyde E, Debelius JW, Morton JT, Gonzalez A, Ackermann G, Aksenov AA,
Behsaz B, Brennan C, Chen Y, Goldasich LD, Dorrestein PC, Dunn RR, Fahimipour AK,
Gaffney J, Gilbert JA, Gogul G, Green JL, Hugenholtz P, Humphrey G, Huttenhower C, Jack-
son MA, Janssen S, Jeste DV, Jiang L, Kelley ST, Knights D, Kosciolek T, Ladau J, Leach
J, Marotz C, Meleshko D, Melnik AV, Metcalf JL, Mohimani H, Montassier E, Navas-Molina J,
Nguyen TT, Peddada S, Pevzner P, Pollard KS, Rahnavard G, Robbins-Pianka A, Sangwan N,
Shorenstein J, Smarr L, Song SJ, Spector T, Swafford AD, Thackray VG, Thompson LR, Tri-
pathi A, Vázquez-Baeza Y, Vrbanac A, Wischmeyer P, Wolfe E, Zhu Q, Knight R, Mann AE,
Amir A, Frazier A, Martino C, Lebrilla C, Lozupone C, Lewis CM, Raison C, Zhang C, Lauber
CL, Warinner C, Lowry CA, Callewaert C, Bloss C, Willner D, Galzerani DD, Gonzalez DJ, Mills
DA, Chopra D, Gevers D, Berg-Lyons D, Sears DD, Wendel D, Lovelace E, Pierce E, TerAvest
E, Bolyen E, Bushman FD, Wu GD, Church GM, Saxe G, Holscher HD, Ugrina I, German JB,
Caporaso JG, Wozniak JM, Kerr J, Ravel J, Lewis JD, Suchodolski JS, Jansson JK, Hampton-
Marcell JT, Bobe J, Raes J, Chase JH, Eisen JA, Monk J, Clemente JC, Petrosino J, Goodrich
J, Gauglitz J, Jacobs J, Zengler K, Swanson KS, Lewis K, Mayer K, Bittinger K, Dillon L,
Zaramela LS, Schriml LM, Dominguez-Bello MG, Jankowska MM, Blaser M, Pirrung M, Minson
M, Kurisu M, Ajami N, Gottel NR, Chia N, Fierer N, White O, Cani PD, Gajer P, Strandwitz P,
Kashyap P, Dutton R, Park RS, Xavier RJ, Mills RH, Krajmalnik-Brown R, Ley R, Owens SM,
Klemmer S, Matamoros S, Mirarab S, Moorman S, Holmes S, Schwartz T, Eshoo-Anton TW,
Vigers T, Pandey V, Treuren WV, Fang X, Xu ZZ, Jarmusch A, Geier J, Reeve N, Silva R, Kopy-
lova E, Nguyen D, Sanders K, Benitez RAS, Heale AC, Abramson M, Waldispühl J, Butyaev A,
Drogaris C, Nazarova E, Ball M, Gunderson B. 2018. American gut: An open platform for citizen
science microbiome research. mSystems 3. doi:10.1128/msystems.00031-18.
14.
Schloss
PD,
Gevers
D,
Westcott
SL.
2011.
Reducing
the
effects
of
PCR
ampli-
fication
and
sequencing
artifacts
on
16S
rRNA-based
studies.
PLoS
ONE
6:e27310.
doi:10.1371/journal.pone.0027310.
15.
Schloss PD. 2018. Identifying and overcoming threats to reproducibility, replicability, robustness,
and generalizability in microbiome research. mBio 9. doi:10.1128/mbio.00525-18.
16.
Sogin ML, Morrison HG, Huber JA, Welch DM, Huse SM, Neal PR, Arrieta JM, Herndl GJ. 2006.
Microbial diversity in the deep sea and the underexplored "rare biosphere".
Proceedings of the
National Academy of Sciences 103:12115–12120. doi:10.1073/pnas.0605127103.
325
17.
Huse SM, Welch DM, Morrison HG, Sogin ML. 2010. Ironing out the wrinkles in the rare biosphere
through improved OTU clustering. Environmental Microbiology 12:1889–1898. doi:10.1111/j.1462-
2920.2010.02193.x.
15
18.
Reeder J, Knight R. 2009.
The rare biosphere: A reality check.
Nature Methods 6:636–637.
doi:10.1038/nmeth0909-636.
19.
Schloss PD, Westcott SL, Ryabin T, Hall JR, Hartmann M, Hollister EB, Lesniewski RA, Oak-
ley BB, Parks DH, Robinson CJ, Sahl JW, Stres B, Thallinger GG, Horn DJV, Weber CF. 2009.
Introducing mothur: Open-source, platform-independent, community-supported software for describ-
ing and comparing microbial communities. Applied and Environmental Microbiology 75:7537–7541.
doi:10.1128/aem.01541-09.
20.
Westcott SL, Schloss PD. 2017. OptiClust, an improved method for assigning amplicon-based
sequence data to operational taxonomic units. mSphere 2. doi:10.1128/mspheredirect.00073-17.
21.
Anderson MJ. 2001. A new method for non-parametric multivariate analysis of variance. Austral
Ecology 26:32–46. doi:10.1111/j.1442-9993.2001.01070.pp.x.
330
22.
Li Q, Heist EP, Moe LA. 2015. Bacterial community structure and dynamics during corn-based
bioethanol fermentation. Microbial Ecology 71:409–421. doi:10.1007/s00248-015-0673-9.
23.
Baxter NT, Ruffin MT, Rogers MAM, Schloss PD. 2016.
Microbiota-based model improves
the sensitivity of fecal immunochemical test for detecting colonic lesions.
Genome Medicine 8.
doi:10.1186/s13073-016-0290-3.
24.
Beall BFN, Twiss MR, Smith DE, Oyserman BO, Rozmarynowycz MJ, Binding CE, Bourbon-
niere RA, Bullerjahn GS, Palmer ME, Reavie ED, Waters LMK, Woityra LWC, McKay RML.
2015. Ice cover extent drives phytoplankton and bacterial community structure in a large north-
temperate lake: Implications for a warming climate.
Environmental Microbiology 18:1704–1719.
doi:10.1111/1462-2920.12819.
25.
Henson MW, Pitre DM, Weckhorst JL, Lanclos VC, Webber AT, Thrash JC. 2016. Artificial sea-
water media facilitate cultivating members of the microbial majority from the gulf of mexico. mSphere
1. doi:10.1128/msphere.00028-16.
26.
Baxter NT, Wan JJ, Schubert AM, Jenior ML, Myers P, Schloss PD. 2014. Intra- and interindi-
vidual variations mask interspecies variation in the microbiota of sympatric peromyscus populations.
Applied and Environmental Microbiology 81:396–404. doi:10.1128/aem.02303-14.
335
27.
Levy-Booth DJ, Giesbrecht IJW, Kellogg CTE, Heger TJ, D’Amore DV, Keeling PJ, Hallam
SJ, Mohn WW. 2018. Seasonal and ecohydrological regulation of active microbial populations in-
volved in DOC, CO2, and CH4 fluxes in temperate rainforest soil. The ISME Journal 13:950–963.
doi:10.1038/s41396-018-0334-3.
28.
Edwards J, Johnson C, Santos-Medellín C, Lurie E, Podishetty NK, Bhatnagar S, Eisen JA,
Sundaresan V. 2015. Structure, variation, and assembly of the root-associated microbiomes of rice.
Proceedings of the National Academy of Sciences 112:E911–E920. doi:10.1073/pnas.1414592112.
29.
Ettinger CL, Williams SL, Abbott JM, Stachowicz JJ, Eisen JA. 2017. Microbiome succession
during ammonification in eelgrass bed sediments. PeerJ 5:e3674. doi:10.7717/peerj.3674.
16
30.
Graw MF, DAngelo G, Borchers M, Thurber AR, Johnson JE, Zhang C, Liu H, Colwell FS.
2018. Energy gradients structure microbial communities across sediment horizons in deep marine
sediments of the south china sea. Frontiers in Microbiology 9. doi:10.3389/fmicb.2018.00729.
31.
Johnston ER, Rodriguez-R LM, Luo C, Yuan MM, Wu L, He Z, Schuur EAG, Luo Y, Tiedje
JM, Zhou J, Konstantinidis KT. 2016.
Metagenomics reveals pervasive bacterial populations
and reduced community diversity across the alaska tundra ecosystem. Frontiers in Microbiology
7. doi:10.3389/fmicb.2016.00579.
340
32.
Hassell N, Tinker KA, Moore T, Ottesen EA. 2018. Temporal and spatial dynamics in microbial
community composition within a temperate stream network. Environmental Microbiology 20:3560–
3572. doi:10.1111/1462-2920.14311.
17
Table 1. Summary of studies used in the analysis. For all studies, the number of sequences used from
each study was rarefied to the smallest sample size. A graphical represenation of the distribution of sample
sizes for each study and the samples that were removed from each study are provided in Figure S1.
Study (Ref)
Samples
Total
sequences
Median
sequences
Range of
sequences
Fold-difference
between largest
and smallest sample
SRA study
accession
Bioethanol (22)
95
3,972,943
16,015
3,688-356,136
96.6
SRP055545
Human (23)
490
20,909,768
32,505
10,523-430,415
40.9
SRP062005
Lake (24)
52
3,169,868
69,041
15,347-112,871
7.4
SRP050963
Marine (25)
7
1,391,396
193,464
133,516-254,060
1.9
SRP068101
Mice (2)
348
2,813,747
6,477
1,804-30,565
16.9
SRP192323
Peromyscus (26)
111
1,555,545
12,446
4,464-33,644
7.5
SRP044050
Rainforest (27)
69
946,295
11,561
4,932-37,767
7.7
ERP023747
Rice (28)
490
22,591,168
43,216
2,776-193,464
69.7
SRP044745
Seagrass (29)
286
4,130,454
13,567
1,803-45,191
25.1
SRP092441
Sediment (30)
58
1,154,174
17,584
7,685-68,321
8.9
SRP097192
Soil (31)
18
956,656
51,844
47,806-59,956
1.3
ERP012016
Stream (32)
201
21,162,574
90,159
9,175-390,964
42.6
SRP075852
18
345
Figure 1. Singletons are more common in samples with fewer seqeunces and tend to be shared with
samples having more sequences. For each of the 12 datasets, Spearman correlation coefficients were
calcualted between the number of sequences in each sample and the number of singletons in the sample
(A) and the fraction of its singletons that were shared with another sample (C). Those correlations that were
not statisically significant had a P-value greater than 0.05. The faction of singletons shared across samples
350
(B) were calculated for each dataset. The median value is shown with a solid circle and the 95% confidence
interval is indicated by the solid line.
19
Figure 2. Removing rare sequences from samples alters their representation of alpha-diversity us-
ing amplicon sequence variants (ASVs). The average difference in the richness (A), Shannon diversity
355
(B), and Kullback-Leiber divergence (C) for each sample within a dataset was calculated between the origi-
nal community structures relative to applying different minimum abundance thresholds.
20
Figure 3. Removing rare sequences from samples increases the inter-sample variation for amplicon
sequence variants (ASVs). The coefficient of variation in richness (A), Shannon diversity (B), and Bray-
360
Curtis distances (C) for each dataset was calculated using the null distributed samples for each dataset with
varying minimum abundance thresholds.
21
Figure 4. Removing rare sequences from samples reduces the statistical power to detect differences
between empirically generated treatment groups when using amplicon sequence variants (ASVs).
365
The fraction of significant tests comparing the richness (A) and Shannon diversity (B) using a Wilcox test
and Bray-Curtis distances (C) using analysis of molecular variance for each dataset was calculated using
empirically generated treatment groups containing equal numbers of samples for each dataset with varying
minimum abundance thresholds. For each dataset and minimum abundance threshold, 100 randomizations
were peformed.
370
22
Figure 5. Removing rare sequences does not impact the false detection rate unless the number of
sequences per sample is confounded with the treatment groups when using amplicon sequence
variants (ASVs). The fraction of significant tests comparing the richness and Shannon diversity using a
Wilcox test and Bray-Curtis distances using analysis of molecular variance for each dataset was calculated.
375
Empirically generated treatment groups were generated containing equal numbers of samples where the
samples represented a null distribution. In one simulation the samples were randomly assigned to a treat-
ment group (A) and in the other the samples were assigned based on the number of sequences in each
sample (B). For each dataset and minimum abundance threshold, 100 randomizations were peformed.
23
380
Figure S1. Distribution of the number of sequences per sample in the 12 datasets included in this
study. A different minimum number of sequences per sample threshold was applied to each dataset based
on identifying natural breaks in the distribution of the number of sequences per sample.
24
Figure S2. Removing rare sequences from samples alters their representation of alpha-diversity
385
using operational taxonomic units (OTUs). The average difference in the richness (A), Shannon diver-
sity (B), and Kullback-Leiber divergence (C) for each sample within a dataset was calculated between the
original community structures relative to applying different minimum abundance thresholds.
25
Figure S3. Removing rare sequences from samples alters their representation of alpha-diversity
390
when regenerating samples using a null distribution for each dataset. The average difference in the
richness, Shannon diversity, and Kullback-Leiber divergence for each sample within a dataset was calcu-
lated between the original community structures relative to applying different minimum abundance thresh-
olds. Some values of Kullback-Leiber divergence are missing because undefined values were calculated
due to the removal of rare sequences. Data are shown for amplicon sequence variants (ASVs) and opera-
395
tional taxonomic units (OTUs)
26
Figure S4. Removing rare sequences from samples increases the inter-sample variation for op-
erational taxonomic units (OTUs). The coefficient of variation in richness (A), Shannon diversity (B),
and Bray-Curtis distances (C) for each dataset was calculated using the null distributed samples for each
400
dataset with varying minimum abundance thresholds.
27
Figure S5. Removing rare sequences from samples reduces the statistical power to detect differ-
ences between empirically generated treatment groups when using operational taxonomic units
(OTUs). The fraction of significant tests comparing the richness (A) and Shannon diversity (B) using a
405
Wilcox test and Bray-Curtis distances (C) using analysis of molecular variance for each dataset was calcu-
lated using empirically generated treatment groups containing equal numbers of samples for each dataset
with varying minimum abundance thresholds. For each dataset and minimum abundance threshold, 100
randomizations were peformed.
28
410
Figure S6. Removing rare sequences does not impact the false detection rate unless the number of
sequences per sample is confounded with the treatment groups when using operational taxonomic
units (OTUs). The fraction of significant tests comparing the richness and Shannon diversity using a Wilcox
test and Bray-Curtis distances using analysis of molecular variance for each dataset was calculated. Empir-
ically generated treatment groups were generated containing equal numbers of samples where the samples
415
represented a null distribution. In one simulation the samples were randomly assigned to a treatment group
(A) and in the other the samples were assigned based on the number of sequences in each sample (B).
For each dataset and minimum abundance threshold, 100 randomizations were peformed.
29
| 2020 | Removal of rare amplicon sequence variants from 16S rRNA gene sequence surveys biases the interpretation of community structure data | 10.1101/2020.12.11.422279 | null | creative-commons |
Why does a flying fish taxi on sea surface before take-off? A
hydrodynamic interpretation
Jian Deng,∗ Shuhong Wang, and Lingxin Zhang
Department of Mechanics, Zhejiang University,
Hangzhou 310027, People’s Republic of China.
Xuerui Mao
Department of Mechanical, Material and Manufacturing Engineering,
the University of Nottingham, University Part, Nottingham NG7 2RD, UK.
Abstract
Flying fish have been observed jumping out of warm ocean waters worldwide. Before take-off,
the flying fish are seen to taxi on the water surface by rapidly beating their semi-submerged tail
fins, which process may help them airborne with enough speed to glide over a long distance. To
understand the underlying physical mechanisms, here, we study a flying fish, 0.25 m in length
and 0.191 kg in weight, considering both its underwater swimming and surface taxiing locomotion.
Its hydrodynamic characteristics are numerically studied by computational fluid dynamics (CFD).
Underwater, the fish is assumed to swim at a constant speed of 10 m s−1.
Different critical
frequencies are identified for various maximum deflected angles, ranging from θ0 = 10o to 30o, at
which the fish reaches cruising states, when the horizontal forces are balanced. The corresponding
minimum power required for cruising swimming is 350 W, obtained at a deflected angle of 10o and
a critical frequency of 145 Hz. In contrast, in the taxiing stage, the minimum power required for a
stead-state locomotion at 10 m s−1 is 36 W, occurring at a deflected angle of 15o and a frequency
of 50 Hz. We note that the power is significantly smaller than the swimming locomotion. Further,
by increasing the flapping power, we find that larger speeds can be achieved. In specific, when
the power is brought up to 350 W, it can reach a speed of 16.5 m s−1. Clearly, from the direct
comparison between the two locomotive modes, it is apparently evidenced that the flying fish can
be further accelerated by taxiing along the water surface.
∗ zjudengjian@zju.edu.cn(Corresponding author)
1
I.
INTRODUCTION
In recent years, there is a rising interest in new concepts of robots with hybrid and
multi-modal locomotion (Low et al. 2015), learning from their natural counterparts. For
example, the so-called amphibians, such as turtles and salamanders can swim in water and
walk on land, while swans can swim on water and fly effectively. More interestingly, there is
a family of marine fish, exocoetidae, in the order Beloniformes, class Actinopterygii, known
colloquially as flying fish (Breder Jr 1938, Nelson et al. 2016). They do not fly, in the sense
of flapping their wing-sized pectoral fins, but actually perform unpowered glide. Their re-
markable abilities could inspire innovative designs to improve the way man-made systems
operating in the environments consisting of multiple media. However, it is very challenging
to design such aerial-aquatic robots or Aquatic Micro Aerial Vehicles (AquaMAVs), mim-
icking real flying fish. The challenges lie in platform design, high power density propulsion
systems and control across the air-water interface, due to the impossibility to simultane-
ously optimize the performance in different locomotive modes (Gao and Techet 2011). It is
therefore necessary to identify the key design principles that make their mobility realizable
and effective by understanding the underlying physics of multi-modal locomotion (Low et al.
2015).
According to the statistic data provided by zoologists, adult flying fish vary in size from
the two-winged Parexocoetus brachypterus with a maximum recorded standard length (L)
of 125 mm to the four-winged Cypsilurus lineatus with a maximum length of 378 mm (Bruun
1935). It was reported that an adult four-winged flying fish, 0.3 m in length, can reach a
speed of up to 10 m s−1 (about 20-30 body lengths s−1) in water with its pectoral fins folded
tightly against its streamlined body (Davenport 1994). Upon piercing the sea surface, it
spreads its enlarged pectoral fins and gains additional thrust by beating its tail fins with
the lower lobes submerged at a frequency of 50-70 strokes s−1. When sufficiently high speed
has been attained, the tail is lifted clear of the water and the fish is airborne, gliding a few
meters above the surface at a peak speed of about 15 − 20 m s−1 (Davenport 1994). Field
observations have revealed that a flying fish can reach a gliding distance of 50 m and a peak
height up to 8 m, when the tail fin is held high and still (Hubbs 1933). The flying fish can
make several consecutive glides, with its tail propelling it up again each time it sinks back to
the surface. A total flight distance of 400 m can be achieved in 30 s by the repeated taxiing
2
before the fish submerges eventually into the water (Franzisket 1965).
Flying fish were thought to have evolved the remarkable gliding ability to escape preda-
tors (Davenport 1994), which is intuitively attractive since the predator may lose sight of
the flying fish when it bursts into air. However, some scientists believed that the periodic
flights of flying fish could also be part of an energy-saving strategy akin to some marine
mammals which repeatedly jump out of the water when cruising for long distance, yield-
ing additional benefits of being a more efficient mode of transportation (Rayner 1986). It
has been reported that animals use diverse strategies to reduce the energetically expensive
cost of locomotion, ranging from morphological to behavioural solutions (Schmidt-Nielsen
1972). Intermittent locomotion, as a specific behavioural strategy is widely taken by both
vertebrates and invertebrates to reduce the cost of movement (Kramer and McLaughlin
2001). For example, the intermittent locomotion, analogous to undulating flight, of a bird
involves gliding with flexed wings interspersed with active flapping, in which the potential
energy from gravity and altitude is translated into horizonal distance via gliding, resulting
in savings of mechanical power compared with continuous level flight (Gleiss et al. 2011).
Similarly, the aerial-aquatic locomotion of a flying fish can also be regarded as a strategy for
power energy saving, whilst in a more unique way. It is apparent that flying fish are likely
to experience less resistance in air than that in water assuming moving forward at the same
speed. Since the flying stage is unpowered, the flying fish should, first, be equipped with
highly modified pectoral fins providing sufficient lift during gliding, which has been proved
by previous experiments (Park and Choi 2010) as well as our numerical simulations (Deng
et al. 2019). Second, a large take-off speed is preferred for the flying fish to achieve the
horizontal distance of gliding flight as long as possible. To achieve the second goal, taxiing
on the water surface seems to be a reasonable choice, which as we have introduced above,
can probably accelerate the flying fish to a speed up to 15 − 20 m s−1 (Davenport 1994).
From the point of view of physiology, the swimming ability of a flying fish 0.3 m in length
swimming at a maximum cruising speed of 10 m s−1, is extraordinary. Wardle (1975) found
that the maximum ‘steady-state’, or cruising speed of fish could be predicted empirically by
Umax = Y L
2T ,
(1)
where Y is the stride length (the ratio between an one-tail-beating forward distance and
the body length L), and T is the time for one contraction of the swimming muscles (two
3
contraction time is equal to one tail beating cycle). Y varies from 0.6 to 0.81. Therefore,
for the flying fish of L = 0.3 m with a stride length of 0.8, and cruising at 10 m s−1:
10 = 0.8 × 0.3
2 × T
,
(2)
then T = 0.012 s is obtained, which means that the flying fish has to perform 83.3 con-
tractions s−1 or 41.7 tail beats s−1. Unfortunately, according to Wardle’s curve (Wardle
1975) for the relationship between L and T for various ambient temperatures, the minimum
contraction time is about 0.025 s for L = 0.3 m (at 20 oC). Therefore, the required T value
is about half the predicted from Wardle’s curve (Wardle 1975). It is thereby unlikely that
flying fish will be able to emerge and fly at temperatures lower than 20oC. Indeed, the tail
beating rates of about 50 beats s−1 have been recorded in warmer waters (in Caribbean
waters at about 25 oC), indicating that the contraction rate is within the capability of flying
fish (Hertel 1966).
Hydrodynamically, as an approximation, we can evaluate the drag force, then the power
required to achieve a ‘steady-state’ swim by using the dead-drag coefficient of a fish of
similar size, i.e., CD ≈ 0.02 (at Re = 2.54 × 106) (Blake 1983). The drag force FD and the
corresponding power Pr can be expressed as
FD = 1
2CDρU 2Aw,
(3)
and
Pr = FDU,
(4)
respectively. For a flying fish with L = 0.25 m, swimming at U =10 m s−1, excluding the
pectoral fins (which will be folded against the body during swimming), assuming the wing
area Aw = 0.023 m2 and ρ = 1000 kg m−3 (Gao and Techet 2011), we get FD = 23 N and
Pr = 230 W. As suggested by Gao and Techet (2011), this required power for the flying fish,
weighing 0.2 kg, relates to a muscle power density of 2300 W kg−1 (assuming 50% muscle
by weight). We note that this power density is far beyond the range of the existing artificial
actuators, from electromagnetic actuators such as DC motors (on the order of 100 W kg−1)
to pneumatic actuators and air muscles (weight ratios of up to 400 W kg−1).
It appears that the maximum swimming speed of 10 m s−1 for a flying fish is markedly
high from the perspectives of hydrodynamic resistance and power requirement. It is unlikely
that the flying fish will be able to swim faster, due to the quadratical increase of the drag
4
FIG. 1: Geometrical configurations of the three flying fish models: underwater swimming
models with (a) a rigid flapping tail and (b) a periodically morphing tail, and (c) the
taxiing model with a spreading pectoral fins. Note that (a) and (b) are shown at the time
instant of 1/4T when the tails reach their maximum deflected angles (θ0 = 20o), where T is
the period of one beating cycle.
force with the swimming speed. We therefore conjecture that taxiing on the water surface is a
necessary stage for the flying fish to reach an ideal take-off speed, though more solid evidence
is required, which is the central aim of the current study.
Moreover, despite the great
challenges associated with the design of aerial-aquatic robots, there is limited understanding
of the underlying physical mechanisms for their natural counterparts.
In this paper, we study a flying fish with both underwater swimming and water-surface
taxiing locomotion considered.
We aim to explain why a flying fish taxis on the water
surface before take-off from the perspective of hydrodynamics, with a particular concern on
direct comparisons in beating frequency and power requirement between the two locomotive
modes.
II.
PHYSICAL MODEL AND NUMERICAL METHOD
A.
Physical model
We consider two different locomotive modes of a flying fish, i.e., underwater swimming
and water surface taxiing, as shown in figure 1. The pectoral fins are folded for the swimming
5
mode (see figure 1(a) and (b)), while spread for the taxiing mode (see figure 1(c)). For both
modes, the pectoral fins are with straight leading edge, following the same geometry used
in our previous study (Deng et al. 2019). As a simplification, the fins are very thin with
simple rectangular cross sections, with a thickness of 1 mm, which can thereby be neglected
in the analysis, and the pelvic fins are removed.
The morphological parameters are chosen as follows: the standard length L = 0.25 m,
the wing area for the spread pectoral fins A = 0.024 m2, the wing span S = 0.47 m, the
wing aspect ratio AR = 9.2 (AR = S2/A), the average wing chord length C = 0.051 m.
The body mass is set to W = 0.191 kg, resulting in a wing loading of 78 N m−2, falling in
the range of wing loadings for six genera reported by Fish (1990). The root chord length
of the fish tail is D = 0.03 m, measuring from the pivoting point to the tail fork point.
We note that the currently adopted flying fish model follows closely the model B that we
chose previously for a gliding flight (Deng et al. 2019), while differing from the model B in
standard length, which was L = 0.2 m.
For the rigid flapping tail, it pitches periodically along a pivoting point as marked in
figure 1(a), while for the periodically morphing tail, it deforms laterally and periodically
with the amplitude of lateral displacement increasing linearly from the pivoting point to the
tail tips, as shown in figure 1(b). A maximum deflected angle θ0 is defined to quantize the
beating amplitude of the tail.
The angles of attack (α) are 0o and 5o for the swimming and taxiing locomotion, re-
spectively. For the taxiing locomotion as shown in figure 1(c), the tail is bent down to an
inclined angle of 30o with respect to the horizontal plane, therefore the lower lobe of the tail
fin is submerged in the initial flow field.
B.
Numerical method
The numerical simulations are carried out using the commercial CFD code STAR-CCM+
12.06.011 (CD-adapco 2017), which is based on the finite volume method. The governing
equations for the incompressible, viscous flow include a continuity equation and momentum
equation for each of the three dimensions. A segregated flow model is used to solve each
of the momentum equations in turn, one for each dimension.
The linkage between the
momentum and continuity equations is achieved with a predictor-corrector approach. A
6
FIG. 2: Computational domain showing the background mesh and overset mesh, and the
mesh refinement around the flying fish as well as that along the free surface.
hybrid second-order upwind/bounded-central scheme is used for the convection term, with
the upwind blending factor set to 0.15. For temporal discretization, a first-order implicit
scheme is used, with several inner iterations involved in each physical time step to converge
the solution for the given time step.
Volume of fluid (VOF) method is used to model the free surface for the surface taxxing
cases. The high-Resolution Interface Capturing (HRIC) scheme is applied to the convective
terms of the VOF transport equation, resulting in a scheme that is suited for sharp interface
tracking (Muzaferija and Peric 1999). To model the turbulence, the SST K-Omega Detached
Eddy model is used, which combines the features of SST K-Omega RANS model in the
boundary layers with a large eddy simulation (LES) in unsteady separated regions (Shur
et al. 2008).
To deal with the flapping or periodically morphing tail, an overset meshing technique is
adopted (Hadzic 2006). The computational domain includes a background mesh enclosing
the entire solution domain, containing the pectoral fins and the fish body, and one smaller
overset mesh (a cubic box) containing the tail fin, as shown in figure 2. We note that the
mesh shown in figure 2 is for the surface taxiing cases, while the swimming cases follow the
same strategy of mesh generation, except for the less requirement of cell number due to the
7
folded pectoral fins. For the cases of rigid flapping tail, the entire overset mesh moves with
the tail, while for the cases of morphing tail, the vertices in the overset mesh redistribute
in response to the movement of the fish tail. The cells in the background mesh covered
by the overset region are deactivated and do not take part in the simulations. These cells
can be anyway reactivated later on, if by means of movements they become active again.
At the boundaries between the background and overset regions, active (discretization) and
interpolation cells are present.
The solution is computed at the active cells, while it is
interpolated at the interpolation cells. Detailed implementation of the overset techniques
can be found in Ref.(Hadzic 2006) or the STAR-CCM+ manual (CD-adapco 2017).
C.
Computational setup
The computational domain is a cuboid, of which the dimensions of its outer boundary is
2 m × 4 m × 2 m at the x-y-z coordinates, as shown in figure 2, for both the swimming
and taxiing locomotion. In the streamwise, or y direction, inlet and outlet boundaries are 1
m and 3 m respectively away from the fish nose. The grid cells on the surfaces, particularly
around the fins, of the flying fish, and along the free surface are refined, as shown in figure
2. Five layers of cells are generated within the wall boundaries, with the first layer with the
height of about 0.0003 m.
The velocity magnitude at the inlet boundary is set to a constant value, for example
U∞=10 m s−1. The density and dynamic viscosity for water are ρ1 = 997.56 kg m−3 and
µ1 = 8.887×10−4 kg m−1 s−1, respectively. They are ρ2 = 1.18 kg m−3 and µ2 = 1.855×10−5
kg m−1 s−1 respectively for air. Therefore, the resulting Reynolds number in the water
Re1 = ρ1U∞L/µ1 = 2.8 × 106, taking the standard length as its length scale, and that in
the air Re2 = ρ2U∞C/µ2 = 3.2 × 104, taking the chord length as its length scale, which is a
relatively low Reynolds number in contrast to artificial aircrafts.
D.
Validation
To resolve the unsteadiness of the flow induced by flapping tail, first, we set the time
step size to be 1/(1000f), i.e., 1000 time steps per flapping cycle, which yields time-accurate
predictions for both mean and instantaneous values. Moreover, the time step size ∆t is
8
TABLE I: Results of validation through space refinement;
taxiing at U∞ = 10 m s−1, θ0 = 15o and f = 100 Hz
Cell number for
Cell number for Thrust force Power input Error for Error for
background mesh
overset mesh
(FT )(N)
(Pin)(W)
FT
Pin
6, 522, 198
476, 650
7.39
362
6.5%
5.8%
9, 448, 796
584, 526
7.28
355
4.9%
3.8%
25, 465, 732
885, 864
7.03
349
1.3%
2.0%
33, 267, 959
1, 006, 804
6.96
340
0.3%
−0.5%
49, 123, 779
1, 556, 511
6.93
344
0.1%
0.6%
63, 834, 110
1, 998, 974
6.94
342
0%
0%
adjusted during simulations to meet the Courant-Friedrichs-Lewy (CFL) condition, i.e.,
Co = ∆t|U|/∆x < 1. To obtain mesh independence results, we choose a typical surface
taxiing case to perform rigorous self-consistency tests. Different mesh resolutions are used,
from a coarse background mesh (6, 522, 198 cells) to a very fine mesh (63, 834, 110 cells),
with the overset mesh adjusted accordingly. In this case, the flying fish taxis at U∞ = 10
m s−1 and θ0 = 15o, and the tails beats at a frequency of f = 100 Hz. In table I, we
show the time-averaged thrust force FT = −Fy and the power input Pin, which is calculated
by integrating the inner product of distributed forces and moving velocities along the tail
surface. The relative errors are evaluated with respect to the finest mesh resolution. It
can be seen that our medium-mesh resolution provides satisfactory accuracy in space, as far
as thrust force and power input concerned. Indeed, for our test cases, differences between
the medium-mesh (33, 267, 959 cells) and the fine-mesh (63, 834, 110 cells) results are quite
small, with variations of less than 1% on both thrust force and power input. Therefore,
we use the medium mesh of 33, 267, 959 cells for all the following calculations of taxiing
locomotion.
For the underwater swimming cases, since the pectoral fins are folded (see
figure 1(a) and (b)), requiring less mesh requirement around them, and there is no free
surface to perform mesh refinement, the corresponding mesh numbers are greatly reduced,
which are 9, 777, 998 for the background mesh and 981, 935 for the overset mesh, by adopting
the same mesh generation strategy.
9
75
80
85
90
95
100
105
110
115
120
125
-15
-10
-5
0
5
10
15
20
25
30
F
T
(N)
Frequency (Hz)
30
o
Morphing tail
20
o
Morphing tail
15
o
Morphing tail
30
o
Drag force on body
20
o
Drag force on body
15
o
Drag force on body
80
90
100
110
120
130
140
150
160
170
180
190
200
-15
-10
-5
0
5
10
15
20
25
30
F
T
(N)
Frequency (Hz)
30
o
Rigid flapping tail
20
o
Rigid flapping tail
15
o
Rigid flapping tail
10
o
Rigid flapping tail
10
o
Drag force on body
(b)
(a)
FIG. 3: Thrust forces on the flying fish when it swims at a constant speed of 10 m s−1 for
different beating frequencies, propelled by (a) rigid flapping tail; (b) periodically morphing
tail.
III.
RESULTS AND DISCUSSION
A.
Underwater swimming locomotion
For the swimming locomotion, the pectoral fins are folded, as shown in figure 1 (a) and
(b). Two propulsive modes are considered, both swimming at a constant speed of U∞ = 10
m s−1. First, the tail performs a periodically pitching motion (the rigid flapping mode) with
a sinusoidal variation of the deflected angle with time. For the second one, the morphing
mode, the grid points on the tail oscillate laterally (in the x direction) and periodically,
10
FIG. 4: Wake topologies for the underwater swimming flying fish, visualized using
Q = 7.5(U∞/L)2, in which U∞ =10 m s−1 is the velocity magnitude at the inlet boundary,
or the cruising speed, and L=0.25 m is the standard length of the flying fish. Three typical
cases are chosen to present: (a) the rigid flapping tail at f = 145 Hz and θ0 = 10o, (b) the
rigid flapping tail at f = 118 Hz and θ0 = 30o and (c) the periodically morphing tail at
f = 98 Hz and θ0 = 30o. Blue (dark gray) and red (light gray) colors denote negative and
positive pressure values respectively for 32 contour levels between -20000 and 5000 Pa.
with the amplitude of oscillation increasing linearly from the pivoting point to the tail tips,
mimicking the undulating motion of a fish tail. Since we have defined a maximum deflected
angle θ0 (see figure 1 (a) and (b)), which can be regarded as a characteristic width for the
wake dynamics, it is possible to make a direct comparison between these two propulsive
modes.
Figure 3(a) shows the variations of thrust forces with beating frequency for the rigid
flapping mode, with four different θ0 considered. It is seen that the thrust force increases
along with the beating frequency for all cases. We find that, for each case, the thrust force
crosses the zero value line at a specific critical frequency, indicating the transition from a
deceleration state to an acceleration swimming state. At these critical points, the horizontal
forces are balanced, therefore the flying fish reaches a steady-state or cruising locomotion.
We should point out that the thrust force FT considered here is calculated by integrating
11
the distributed pressure and viscous forces along all parts, including the fish body, folded
pectoral fins and the tail.
It is observed that the critical frequencies for transition are
around 115 Hz for θ0 = 15o, 20o and 30o. However, for the small deflected angle, θ0 = 10o,
the transition occurs at a markedly higher frequency of f = 145 Hz. It suggests that the
fish should beat its tail at a higher frequency if the amplitude has been reduced. We can
define a non-dimensional frequency, or the Strouhal number as
St = 2fDsin(θ0)
U∞
,
(5)
in which D =0.03 m (defined in section II A), then the deceleration-to-acceleration transi-
tions occur in the range of St = 0.151 to 0.345. This range accords surprisingly well with the
cruise Strouhal numbers, lying within a narrow interval 0.2 < St < 0.4, for a wide range of
flying and swimming animals (Taylor et al. 2003). It is noted that our previously identified
Strouhal number, St = 0.225, of a pure pitching foil for drag-thrust transition also lies in
these ranges (Deng et al. 2015, 2016).
For the periodically morphing tail, the critical frequencies are 110 Hz, 100 Hz and 98 Hz
respectively for θ = 15o, 20o and 30o, or in the range of Strouhal number St = 0.152 to 0.33,
as shown in figure 3(b). It appears that the propulsive mode does not change significantly
the locomotive performance, which is mainly determined by the wake width. In figure 4, we
show the wake topologies represented by Q iso-surfaces. The Q-criterion (Jeong and Hussain
1995) defines a vortex as a spatial region where
Q = 1
2
[
|Ω|2 − |S|2]
> 0,
(6)
where S =
1
2[∇v + (∇v)T] is the rate of strain tensor, and Ω =
1
2[∇v − (∇v)T] is the
vorticity tensor. The positive value of Q means that the Euclidean norm of the vorticity
tensor dominates that of the rate of strain. In figure 4, we highlight the vortex cores using
iso-surfaces of Q = 7.5(U∞/L)2. Positive values of Q give prominence to regions of high swirl
in comparison to shear to represent coherent vortices. The wake width is mainly determined
by the flapping amplitude, as clearly seen in In figure 4. In figure 4 (a), when the deflected
angle is small, the vortical structures are confined to a very narrow lateral space, while
for the larger deflected angles, the flow wakes are wider, as seen in figure 4 (b) and (c).
Unlike traditional B´enard-von K´arm´an (BvK) vortex streets viewed in a two-dimensional
plane (Deng and Caulfield 2015), here the flow wakes produced by the forked tail (caudal
12
350
582
1034
3075
416
613
1542
10
15
20
25
30
0
500
1000
1500
2000
2500
3000
3500
Pin (W)
Maximum deflected angle
0
(
o
)
Power for rigid flapping tail
Power for morphing tail
required to achieve a cruising speed of 10m/s
FIG. 5: The power required for the flying fish swimming at a cruising (‘steady-state’)
speed of 10 m s−1. Note that the cruising state is achieved when a zero horizontal force is
obtained, i.e., the intersection points of the force curves and the dash lines shown in figure
3 (a) and (b).
fin) show strongly three-dimensionality. Nevertheless, the reversed BvK streets signaling
propulsive wakes can still by observed. As shown noticeably in figure 4 (b) and (c), the
vortices formed from one side of the tail fin shed to the other side of the wake, forming two
rows of vortex streets, which are connected by vortex filaments in the braid region.
We also present the drag forces in figure 3 (the dash-dotted lines), which are calculated by
integrating the pressure and viscous forces along the flying fish excluding the contributions
from its tail, or the propeller. For all cases, the drag forces are around -7.5 N. As suggested
by Maertens et al. (2015) that it is very challenging to measure the efficiency for a self-
propelled body in steady state, unless we can separate the propeller from the body. Here,
assuming that the propulsive force produced by the tail is balanced by the drag force in
steady state, we define the propulsive efficiency as
ηn = −FDU∞
Pin
,
(7)
in which the power input Pin is obtained by integrating the inner product of distributed
forces and moving velocities along the tail surface. For the steady states, or the cruising
points identified in figure 3, we present their respective power input, or the power required
to achieve the cruising states in figure 5. It is interesting to find that the required power
13
30
40
50
60
70
80
90
100
110
120
130
-2
0
2
4
6
8
10
12
from top to bottom: 15
20
25m/s
from top to bottom: 15
20
20.5
21
21.5m/s
from top to bottom: 15
16
16.5m/s
F
T
(N)
Frequency (Hz)
0
=10
o
surface taxiing
0
=15
o
surface taxiing
at a speed of 10m/s
FIG. 6: Horizontal forces on the flying fish when it taxis along the water surface for
different beating frequencies. The lines marked with symbols show that for the flying fish
taxiing at a speed of 10 m s−1, while the scattered symbols show that with various taxiing
speeds aiming to identify the cruising states for three specific frequencies.
increases with the maximum deflected angle, suggesting that the flying fish tends to consume
less energy by beating its tail at a higher frequency rather than engaging a large amplitude.
However, the minimum power input of 350 W shown in figure 5 requires a very high beating
frequency of up to 145 Hz, which is far beyond the observed data (50 Hz), as well as that from
the physiological constrain, as mentioned in section I. Nevertheless, from a hydrodynamic
point of view alone, this minimum power is achievable. Actually, a real fish has more freedom
of deformations than the model with a fixed incoming flow used in the current study, which
could help reduce the required power when the deflected angle is large. Therefore, a self-
propelled model might be more suitable (Deng and Caulfield 2018, 2016), which can be
considered in future studies.
B.
Surface taxiing locomotion
To understand the mechanism of further acceleration by taxiing on the water surface, here
we study a flying fish beating its semi-submerged tail, with a fixed angle of attack α = 5o
providing sufficient lift forces lifting up the fish from the water. The tail is bent down to
an inclined angle of 30o with respect to the horizontal plane, as shown in figure 2. The tail
14
FIG. 7: Water-air interfaces represented by the isosurface of volume fraction of water
being equal to 0.5, with the colors from blue (dark gray) to red (light gray) denoting the
depths varying from -0.05 m to 0.05 m. The different colors on the body and pectoral fins
represent the pressure distributions. Four typical cases are shown. They are (a) f = 50 Hz,
U∞ = 10m s−1, θ0 = 15o, (b) f = 60 Hz, U∞ = 10 m s−1, θ0 = 10o, (c) f = 80 Hz, U∞ =
16.5 m s−1, θ0 = 15o, and (d) f = 100 Hz, U∞ = 20.5 m s−1, θ0 = 15o.
performs periodically pitching motion, following the first propulsive mode employed by the
underwater swimming locomotion. Intuitively, it is easy to understand that the flying fish
will experience less resistance since most components, including the body and the pectoral
fins, are airborne.
In figure 6, we show the variations of thrust forces with the beating frequency with two
15
different θ0 considered. It is unsurprising to see that the critical frequencies for the flying
fish to reach a cruising state, at the same speed of U∞ = 10 m s−1, are lower than that for
the swimming locomotion (see figure 3 (a)). They are 50 Hz and 60 Hz for θ0 = 15o and
θ0 = 10o, respectively. It seems that these frequencies accord more well with the observed
data (50 Hz, as reported by Hertel (1966)).
Furthermore, we fix the maximum deflected angle at θ0 = 15o and increase the flow
incoming velocity, or equivalently the locomotive speed, at three typical frequencies of f = 80
Hz, 100 Hz and 120 Hz. It is clearly shown from figure 6 that the flying fish has the potential
to cruise faster as the beating frequency is increased. For example, when the flying fish beats
its tail at a frequency of 80 Hz, a cruising speed of 16.5 m s−1 can be achieved, and if the
tail beats at a frequency of 120 Hz, the cruising speed can reach up to 25 m s−1, which is
far beyond the existing data and out of the physical limit of the fish.
Another issue that we should note is the vertical force balance, i.e., between the lift
force and gravity. For a specific case, at θ0 = 15o, U∞ = 16.5 m s−1 and f = 80 Hz, we
find that the lift force provided by the pectoral fins is 2.08 N, or 0.21 kg in weight, which
is very close to its body mass as mentioned in section 1. The drag force on the pectoral
fins is 0.6 N, which is mainly induced by the lift, resulting in a lift-to-drag ratio of 3.47.
Nondimensionlized by 1/2ρ2U 2
∞A, we find that the corresponding lift coefficient is 0.539,
consistent with our previous study for a gliding flying fish at the same angle of attack of 5o
(Deng et al. 2019). We thereby believe that the lift force is sufficient to lift up the flying
fish undergoing a steady state taxiing. It is important to appreciate that the inclined tail fin
can actually provide a small portion of lift force by propelling itself again the water, which
is however not the major concern of the current study, and we leave it open to the future
investigations.
In figure 7, we show the water-air interfaces for four typical taxiing cases, exhibiting
unique wake patterns left on the glassy water surface, which have been widely observed
by zoologists (Howell 2014).
The clearly observed water splashes demonstrate sufficient
resolution of our numerical method for the free surface.
In figure 8 we present the powers required to achieve respectively the five cruising points
in figure 6. By making a direct comparison with figure 5, we see that the minimum power
required to reach a taxiing speed of 10 m s−1 is 36 W, which is one order of magnitude
smaller than that obtained for an underwater swimming flying fish. It is interesting to find
16
36
118
350
684
1204
40
50
60
70
80
90
100
110
120
130
0
200
400
600
800
1000
1200
1400
10m/s
10m/s
16.5m/s
20.5m/s
25m/s
Pin (W)
Frequency (Hz)
Power at different frequencies
required to achieve various cruising speeds
FIG. 8: The power required for the flying fish taxiing at various cruising (‘steady-state’)
speeds, corresponding to the five horizontal force balanced points in figure 6.
that by inputting the same amount of power of 350 W, which is the minimum requirement
for swimming locomotion, the taxiing fish is able to reach a cruising speed of up to 16.5
m s−1. Therefore, it is apparently evidenced that the flying fish can be further accelerated
before take-off by taxiing on the water surface.
Assuming that the flying fish’s work capacity is constrained by its muscle power density,
as we have mentioned above, and considering the currently studied flying fish with 0.191 kg
in weight, the 350 W power relates to a muscle power density of 3664 W kg−1 (assuming
50% muscle by weight). Although this value is considerably higher than the muscle power
density (2300 W kg−1) given by Gao and Techet (2011), we still believe that it is a reasonable
estimation for a real flying fish at this length scale.
Moreover, we present the wake topologies for a typical case of surface taxiing in figure 9,
exhibiting both the water surface distortion and the vortical structure of the flow. Besides
the trailing edge shedding vortices from the spread pectoral fins, the finer vortex filaments
induced by the interaction between the flow wake and the free surface are clearly observed.
IV.
CONCLUSIONS
Following our previous study on the gliding performance of flying fish, here, we are con-
cerned with their underwater swimming and water surface taxiing locomotion. Computa-
17
FIG. 9: Wake topologies for the surface taxiing flying fish, visualized using
Q = 7.5(U∞/L)2, in which U∞ =16.5 m s−1, and the flying fish beats its tail at the
frequency f = 80 Hz. Blue (dark gray) and red (light gray) colors denote negative (below
the still water surface) and positive (above the still water surface) depths respectively.
tional fluid dynamic (CFD) method is applied to investigate the hydrodynamic performance,
focusing on its horizontal force balance and the power required to achieve a cruising state.
We aim to answer the question ‘why does a flying fish taxi on sea surface before take-off’
from a hydrodynamic perspective.
First, for the underwater swimming locomotion, we consider two different propulsive
modes and four different maximum deflected angles, at a constant speed of 10 m s−1. The
results show that the critical frequencies are around 115 Hz for θ0 = 15o, 20o and 30o, while
it is 145 Hz for the small deflected angle of θ0 = 10o, at which point the minimum power
input of 350 W is recorded. We state that the required beating frequency, 145 Hz, for the
minimum power is far beyond the data provided by field observation, which is around 50
Hz. Nevertheless, we suggest that it is achievable from a pure hydrodynamic point of view.
Second, we study a flying fish beating its semi-submerged tail fin, with a fixed angle of
attack α = 5o, and a bent tail down to the water with an inclined angle of 30o with respect
to the horizontal plane. It is unsurprising to find that the critical frequencies for θ0 = 15o
and θ0 = 10o are 50 Hz and 60 Hz respectively, which are markedly lower than that of
the swimming locomotion. Also they are much closer to the observed frequency of 50 Hz
provided by previous zoologists. It sounds reasonable that the frequencies of surface taxiing
are more likely to be recorded in field observation thanks to the wave patterns left on the
18
glassy sea. It is exciting to find that by inputting the same amount of power of 350 W,
which is the minimum requirement for swimming locomotion, the taxiing flying fish is able
to reach a cruising speed of up to 16.5 m s−1. It is apparently evidenced that the flying fish
can be further accelerated before take-off by taxiing on the water surface.
We understand that natural animals are unique with many unrevealed capabilities. It
is a great challenge for artificial robots to wholly duplicate their locomotive mechanisms.
The current study only sheds a first light to the aerial-aquatic locomotion of flying fish.
We believe that there are more underlying physics lying behind flying fish waiting to be
discovered.
ACKNOWLEDGMENTS
This research has been supported by the National Natural Science Foundation of China
(Grant No: 11772299).
REFERENCES
Blake, R. W. (1983). Fish locomotion, CUP Archive.
Breder Jr, C. (1938). A contribution to the life histories of atlantic ocean flyingfishes, Bull. Bingham
Oceangr. Coll. 6: 1–126.
Bruun, A. F. (1935). Flying-fishes (exocoetidae) of the atlantic, Dana Rep. 6: 1–106.
CD-adapco, S. (2017). Star ccm+ user guide version 12.06, CD-Adapco: New York, NY, USA .
Davenport, J. (1994). How and why do flying fish fly?, Reviews in Fish Biology and Fisheries
4(2): 184–214.
Deng, J. and Caulfield, C. (2015). Three-dimensional transition after wake deflection behind a
flapping foil, Physical Review E 91(4): 043017.
Deng, J. and Caulfield, C. (2018). Horizontal locomotion of a vertically flapping oblate spheroid,
Journal of Fluid Mechanics 840: 688–708.
Deng, J. and Caulfield, C.-c. P. (2016). Dependence on aspect ratio of symmetry breaking for
oscillating foils: implications for flapping flight, Journal of Fluid Mechanics 787: 16–49.
19
Deng, J., Sun, L. and Shao, X. (2015). Dynamical features of the wake behind a pitching foil,
Physical Review E 92(6): 063013.
Deng, J., Sun, L., Teng, L., Pan, D. and Shao, X. (2016). The correlation between wake transition
and propulsive efficiency of a flapping foil: A numerical study, Physics of Fluids 28(9): 094101.
Deng, J., Zhang, L., Liu, Z. and Mao, X. (2019). Numerical prediction of aerodynamic performance
for a flying fish during gliding flight, Bioinspiration & Biomimetics 14(4): 046009.
Fish, F. (1990). Wing design and scaling of flying fish with regard to flight performance, Journal
of Zoology 221(3): 391–403.
Franzisket, L. (1965).
Beobachtungen und messungen am flug der fliegenden fische, Zool.
Jb.(Physiol.) 7(0): 235–240.
Gao, A. and Techet, A. H. (2011). Design considerations for a robotic flying fish, OCEANS 2011,
IEEE, pp. 1–8.
Gleiss, A. C., Jorgensen, S. J., Liebsch, N., Sala, J. E., Norman, B., Hays, G. C., Quintana,
F., Grundy, E., Campagna, C., Trites, A. W. et al. (2011). Convergent evolution in locomotory
patterns of flying and swimming animals, Nature communications 2: 352.
Hadzic, H. (2006). Development and application of finite volume method for the computation of
flows around moving bodies on unstructured, overlapping grids.
Hertel, H. (1966). Structure, form, movement, Reinhold.
Howell, S. N. (2014). The Amazing World of Flyingfish, Princeton University Press.
Hubbs, C. L. (1933). Species and hybrids of mollienesia, Aquarium 1(10): 263–268.
Jeong, J. and Hussain, F. (1995). On the identification of a vortex, Journal of fluid mechanics
285: 69–94.
Kramer, D. L. and McLaughlin, R. L. (2001). The behavioral ecology of intermittent locomotion,
American Zoologist 41(2): 137–153.
Low, K., Hu, T., Mohammed, S., Tangorra, J. and Kovac, M. (2015). Perspectives on biologically
inspired hybrid and multi-modal locomotion, Bioinspiration & biomimetics 10(2): 020301.
Maertens, A., Triantafyllou, M. and Yue, D. (2015). Efficiency of fish propulsion, Bioinspiration
& biomimetics 10(4): 046013.
Muzaferija, S. and Peric, M. (1999). Computation of free surface flows using interface tracking
and interface capturing methods,’chap. 2, Mahrenholtz, O. and Markewicz, M., Nonlinear Water
Wave Interaction, Computational Mechanics Publications .
20
Nelson, J. S., Grande, T. C. and Wilson, M. V. (2016). Fishes of the World, John Wiley & Sons.
Park, H. and Choi, H. (2010). Aerodynamic characteristics of flying fish in gliding flight, Journal
of Experimental Biology 213(19): 3269–3279.
Rayner, J. M. (1986). Pleuston: animals which move in water and air, Endeavour 10(2): 58–64.
Schmidt-Nielsen, K. (1972). Locomotion: energy cost of swimming, flying, and running, Science
177(4045): 222–228.
Shur, M. L., Spalart, P. R., Strelets, M. K. and Travin, A. K. (2008). A hybrid rans-les approach
with delayed-des and wall-modelled les capabilities, International Journal of Heat and Fluid Flow
29(6): 1638–1649.
Taylor, G. K., Nudds, R. L. and Thomas, A. L. (2003). Flying and swimming animals cruise at a
strouhal number tuned for high power efficiency, Nature 425(6959): 707.
Wardle, C. (1975). Limit of fish swimming speed, Nature 255(5511): 725.
21
| 2019 | Why does a flying fish taxi on sea surface before take-off? A hydrodynamic interpretation | 10.1101/765560 | [
"Deng Jian",
"Wang Shuhong",
"Zhang Lingxin",
"Mao Xuerui"
] | creative-commons |
Pushing the envelope: force balance in
fission yeast closed mitosis
Marcus A Begley 1, Christian Pagán Medina1, Parsa Zareiesfandabadi1,2,
Matthew B Rapp1, Mary Williard Elting1,3*
1Department of Physics, North Carolina State University, Raleigh, NC, 2Present address:
Department of Biology, Duke University, Durham, NC, 3Quantitative and Computational
Developmental Biology Cluster, North Carolina State University, Raleigh, NC
*Correspondence: mary.elting@ncsu.edu
SUMMARY
The fission yeast S. pombe divides via closed mitosis. In short, mitotic spindle elongation and
chromosome segregation transpire entirely within the complete nuclear envelope. Both the
spindle and nuclear envelope must undergo significant conformation changes and are subject to
varying external forces during this process. While the mechanical relationship between the two
mitotic structures has been explored previously1–4, much is still left to be discovered. Here, we
investigate this relationship by observing the behaviors of spindles and nuclei in live mitotic
fission yeasts following laser ablation. First, we characterize these dynamics in molecularly
typical S. pombe spindles, finding them to be stabilized by dense crosslinking, before
demonstrating that the compressive force acting on the spindle poles is higher in mitotic cells
with greater nuclear envelope tension and that spindle compression can be relieved by
lessening nuclear envelope tension. We finally examine the differences between the mitotic
apparatus in S. pombe and S. japonicus, an evolutionary relative of S. pombe that undergoes
semi-open mitosis, and show that S. japonicus mitotic spindles appear to both splay and bend
more easily than those of their S. pombe relatives. Altogether, these data suggest that fission
yeast spindle crosslinking may be tuned to support spindle extension and oppose nuclear
envelope tension.
RESULTS AND DISCUSSION
S. pombe mitotic spindles are highly crosslinked
The Schizosaccharomyces pombe mitotic spindle consists of a single bundle of 10-20
microtubules, held together along its length by microtubule-crosslinking proteins5–7. In the
spindle midzone, microtubules of alternating geometric polarity form a square lattice5. During
anaphase, motor proteins can crosslink and slide apart antiparallel microtubule neighbors,
creating extensile force within the spindle and driving spindle elongation6,8. In S. pombe, this
elongation is quite dramatic with spindles elongating from hundreds of nanometers in length to
roughly 10 μm over the course of mitosis9,10. Furthermore, these spindles are quite
spatiotemporally stereotyped, adhering to a strict protocol throughout mitosis11, suggesting that
fission yeasts can precisely tune their mitotic forces. The simplicity and standardization of the S.
pombe mitotic apparatus allows us to effectively probe the fundamental physical characteristics
of microtubule bundles subjected to different molecular perturbations.
We began by characterizing spindle structure in control S. pombe cells using laser
ablation of live cells12 (Figure 1A). Upon ablation, the two spindle halves collapse toward each
other in a motor driven process, bringing the two poles together and reforming the spindle8,12,13
(Figure 1A). Before a spindle reforms, the two half-spindles are detached from each other, and
microtubules within each half bundle can either stay tightly associated with each other or
become detached along their lengths, a process that we term “splaying”14.
Comparison of the probability of each of these two behaviors creates an indicator of the
degree of crosslinking between the microtubules in the ablated spindle14. In S. pombe, ablated
spindle halves splay only rarely (Figure 1B), suggesting a high degree of crosslinking,
consistent with previous work characterizing S. pombe spindle structure. Another striking feature
of the behavior of S. pombe spindle halves following severing by laser ablation is that their
overall length is maintained, with very little shortening (Figure 1A, C). We expect the spindle at
the point of severing (generally near the center of the spindle) to form an antiparallel
architecture5, and thus for approximately half of the microtubule ends in each bundle created by
laser ablation to have their plus-ends facing toward the site of ablation. A typical hallmark of the
response to laser ablation of microtubules in many cell types is depolymerization of “naked”
plus-ends created by microtubule severing15–18. Thus, it was initially surprising to us that such
behavior is so rarely observed in these spindles (Figure 1A, C). High densities of crosslinking
proteins have been shown to stabilize microtubule bundles against catastrophic
depolymerization19,20. Therefore, the fact that spindle halves mostly resist depolymerization
following ablation provides additional evidence that S. pombe spindles are highly crosslinked.
Indeed, such crosslinking is also consistent with the highly stereotyped organization of the
central spindle at this stage5. Altogether, these data paint a picture of the S. pombe mitotic
spindle as a tightly-bound microtubule bundle that is stable, even under significant physical
disturbance.
Figure 1: S. pombe mitotic spindles are highly crosslinked. (A) Typical example of laser ablation
of an S. pombe spindle expressing GFP-Atb2 (MWE2), showing post-ablation spindle collapse,
followed by the reattachment of ablated spindle halves at their plus-ends. Scale bars are 2 μm
and timestamps are in min:sec. (B) Ablated S. pombe spindle halves rarely splay apart. Error
bar indicates the square root of the number of splayed events, as an estimate on the variation in
this number assuming Poisson statistics. (C) S. pombe spindles do not depolymerize much,
following ablation. The y-axis is the change in spindle half length, compared to the first frame
after ablation. The line represents the average spindle half length change and the shaded
region represents average +/- standard error on the mean.
Ase1 crosslinking prevents microtubule depolymerization in S.
pombe
An important component of the S. pombe mitotic spindle is the passive microtubule-microtubule
crosslinking protein Ase1, which preferentially localizes to antiparallel microtubule overlaps in
the spindle midzone, particularly at anaphase10,21–23. Here, it functions to stabilize the spindle
midzone and is therefore implicated in the modulation of spindle dynamics. Namely, Ase1
supports bipolar spindle formation in S. pombe through its stabilization of microtubule bundles,
and even allows the formation of a bipolar spindle in the absence of kinesin-5 mediated
antiparallel microtubule sliding24. Intriguingly, Ase1 has also been shown, through the same
underlying mechanism of antiparallel microtubule stabilization, to slow motor-driven anaphase
spindle elongation in budding yeast7,25. Thus, we sought to investigate the role of Ase1 in the
highly crosslinked nature of S. pombe spindles.
To this end, we laser ablated mitotic spindles in Ase1-deletion (ase1Δ) S. pombe, similar
to the experiments described above in cells with normal Ase1 expression. Strikingly, rapid
ablated spindle half depolymerization was common in ase1Δ cells (Figure 3), consistent with the
role of Ase1 in stabilizing bundles19,24,25. Though this depolymerization prevented reliable
characterization of post-ablation spindle half splaying as for control S. pombe cells (Figure 1),
some splaying was visible of the remaining spindle halves as they depolymerized (Figure 2A).
These data are consistent overall with a central role for Ase1 in both crosslinking S. pombe
spindles and stabilizing their microtubules.
Figure 2: Ase1 crosslinking prevents microtubule depolymerization in S. pombe. (A) Typical
example of spindle ablation in ase1Δ S. pombe expressing GFP-Atb2 (MWE3), in which both
spindle halves depolymerize. Some spindle splaying is also visible during depolymerization
(0:49). Scale bars are 2 μm and timestamps are in min:sec. (B) ase1Δ S. pombe (orange)
halves tend to depolymerize more quickly and severely than those from S. pombe with
unperturbed ase1 (blue, same data as Figure 1C), as shown by traces of the average change in
spindle half length after ablation. Traces represent averages across all videos and shaded
regions represent average +/- standard error on the mean.
The fission yeast spindle and nuclear envelope are a mechanical pair
While the S. pombe anaphase spindle elongates, the nuclear envelope undergoes drastic
conformational changes, beginning as a spheroid before transitioning into a dumbbell-like
shape1,3. Simultaneously, the spindle elongates, but still remains an approximately linear
bundle1. The spindle and nuclear envelope are mechanically linked at the two spindle pole
bodies (SPBs), which nest into the nuclear envelope during mitosis in S. pombe26. To probe the
mechanical effect that nuclear envelope tension has on the spindle, we performed laser
ablations in cerulenin-treated S. pombe (Figure 3). Cerulenin is a fatty acid synthesis inhibitor
that prevents the addition of phospholipids to the nuclear envelope during nuclear envelope
expansion, and has previously been shown to alter S. pombe mitotic nuclear shape1,27. In S.
pombe, cerulenin-treatment has been shown to greatly increase nuclear envelope tension,
which in turn can often lead to bending in anaphase spindles1 (Figure 3A and B). We performed
two sets of laser ablation experiments in cerulenin-treated cells.
First, we selectively ablated mitotic spindles that appear curved, presumably due to the
increased surface tension causing the spindle to bow or bend (Figure 3A). Consistent with this
interpretation, the two spindle fragments straighten upon ablation. After severance, the incident
angle between the two halves is much greater than that of straight spindles. As with spindles in
control S. pombe, the two spindle poles collapse toward each other after ablation in
cerulenin-treated cells, however at a much greater rate (Figure 3C). Because the misaligned
configuration of the two ablated fragments would not be conducive to motor transport, compared
to the aligned conformation of ablated control spindles (Figure 1A), this suggests an increase in
compressive force acting on the spindle ends from the stretched nuclear envelope. Interestingly,
these spindles rarely repair and, after failing to reconnect, ablated spindle halves often appear
to depolymerize (Figure 3A, left spindle half).
Secondly, we perform experiments in which we ablate the nuclear envelope of
cerulenin-treated S. pombe and track the changes to the spindle and nucleoplasm (Figure 3B).
This experiment allows us to test whether the envelope tension is indeed the direct mechanical
cause of spindle bending. For these experiments, we use cells expressing both GFP-Atb2 (to
visualize tubulin) and GFP-NLS, which enables us to determine whether we successfully open
the nuclear envelope, allowing nucleoplasm to leak into the cytoplasm. Here, we also observe a
viscoelastic-like ablation response of the spindle. Usually within a minute following ablation, the
spindle straightens from a curved conformation to its relaxed, straight state (Figure 3B, D). This
implies a reduction in the compressive force acting on the spindle poles and provides additional
evidence of nuclear envelope tension as the source of this force. Furthermore, we typically
observe spindle relaxation in concurrence with nucleoplasm leakage, visualized by the
weakening intensity of GFP-NLS signal in the nucleus (Figure 3B). This correlation is apparent
through the nearly overlapping curves of average spindle curvature and average nuclear
(GFP-NLS) intensity over time, following ablation (Figure 3D). Note the delay, at the individual
cell level, following many of the nuclear envelope ablations before substantial spindle and
nuclear envelope relaxation (Figure 3E, inset). The duration of these two delays in spindle
collapse and in nucleoplasm leakage onset are highly correlated, suggesting a causative
relationship (Figure 3E). Because both the spindle and nuclear envelope tend to relax nearly
simultaneously, we hypothesize that the initial nuclear envelope ablation, while likely small in
area, triggers a subsequent catastrophic event in which the hole expands and nuclear envelope
tension plummets.
Figure 3: The fission yeast spindle and nuclear envelope are a mechanical pair. (A) Typical
example of spindle ablation in cerulenin-treated S. pombe cell expressing GFP-Atb2 (MWE2).
Dynamic changes to spindle structure seen here include spindle half straightening (0:00),
spindle collapse, and unrepaired spindle half depolymerization (left half, 1:10). Scale bars are 2
μm and timestamps are in min:sec. (B) Typical example of nuclear envelope ablation in
cerulenin-treated S. pombe cell expressing GFP-Atb2 and GFP-NLS (MWE48), showing
nucleoplasm leakage and spindle relaxation, following nuclear envelope rupture. Scale bars are
2 μm and timestamps are in min:sec. (C) Cerulenin-treated S. pombe spindles (gold) collapse
much faster and more severely than spindles in untreated S. pombe (blue). All cells strain
MWE2. Traces represent average change in pole separation, compared to the first frame after
ablation, and shaded regions represent average +/- standard error on the mean. (D) The
averages of NLS intensity (purple) and spindle curvature (gold) show similar trajectories,
following ablation. Traces represent averages across all videos and shaded regions represent
average +/- standard error on the mean. All cells strain MWE48. (E) The initiations of
nucleoplasm leakage and spindle relaxation, defined using best fit traces for each individual
video, are approximately simultaneous. (E - inset) Example of typical NLS intensity (purple) and
spindle curvature (gold) traces, following laser ablation of cerulenin-treated S. pombe nuclear
envelopes. Jagged traces represent raw data and smooth lines show fits to the raw data, which
have been normalized to the highest NLS intensity and spindle curvature values for that video.
S. japonicus spindles are less crosslinked than the S. pombe spindles
and do not have to oppose as much compressive force
We next took a comparative biology approach to examine how the trends we observed in S.
pombe compared to the fission yeast Schizosaccharomyces japonicus. In general, the S.
japonicus and S. pombe mitotic machinery share many commonalities. The spindle forms a
spindle microtubule bundle as it elongates throughout anaphase and spindle elongation
accompanies significant changes in nuclear envelope shape. However, there are notable
differences between the two systems that are especially relevant to our study. Importantly,
unlike in S. pombe, which maintains nuclear envelope closure throughout its lifecycle, partial
nuclear envelope breakdown can occur during chromosome segregation in S. japonicus, in a
process known as semi-open mitosis1,28,29. Because the nuclear envelope is incomplete for
much of anaphase, the mitotic spindle in S. japonicus elongates while experiencing a lower level
of nuclear envelope tension than in S. pombe.
Mitotic spindles of the two fission yeasts, like their nuclear envelopes, look generally
similar at the beginning and end of anaphase, but can diverge in their intermediate
conformations. Notably, the S. pombe spindle appears quite straight throughout mitosis,
whereas S. japonicus spindles often bend, especially in late anaphase1. This suggests a
possible divergence in spindle structure between the two evolutionary cousins. To explore
possible effects on the mechanical relationship between the spindle and nuclear envelope in
these two yeasts, we repeated many of the experiments described above in S. japonicus.
Figure 4: S. japonicus spindles are less crosslinked than S. pombe spindles and do not have to
oppose as much compressive force. (A) Typical example of spindle ablation in S. japonicus cell
expressing GFP-Atb2 (strain MWE92, Nup189-mCherry expressed but not imaged), showing
the splaying apart of microtubules at the plus ends of one of the ablated spindle halves (upper
right), but not the other (lower left). Scale bars are 2 μm and timestamps are in min:sec. (B)
Nearly half of ablated S. japonicus spindle halves splay apart at some point during spindle
repair (compare to Figure 1B). Error bar indicates the square root of the number of splayed
events, as an estimate on the variation in this number assuming Poisson statistics. (C) Rare
example of cerulenin-treated S. japonicus cell expressing GFP-Atb2 and Nup189-mCherry
(strain MWE92) showed extreme bending, with some individual microtubules mechanically
dissociating with the rest of the bundle as it bows excessively (2:20-3:30).
First, we laser ablated anaphase S. japonicus spindles and investigated the frequency
with which the microtubules in spindle halves splay apart following laser ablation (Figure 4A).
Strikingly, while splaying was rare in S. pombe ablated spindle halves (Figure 1B), the plus-ends
of at least some microtubules mechanically dissociated in nearly half of all ablated S. japonicus
spindle halves (Figure 4A, B), suggesting that microtubules in these spindles are less
tightly-bound to each other along their lengths than their S. pombe counterparts. Further
evidence that S. japonicus spindles may be less cohesive than S. pombe spindles comes from
observations of mitotic progression in cerulenin-treated cells. Here, cerulenin-treated S.
japonicus are imaged, without mechanical perturbations such as laser ablation, as spindles
attempt anaphase elongation. Rarely, but strikingly, we see constrained spindles bow, as they
elongate inside a more tense nuclear envelope (Figure 4C). We never observe such dramatic
bending in S. pombe, and their ability to undergo such bending suggests that S. japonicus
spindles may be less rigid than S. pombe spindles. Additionally, as the spindle shown in Figure
4C bends, individual microtubules appear to dissociate with the rest of the spindle in its
midsection, while remaining bound at the centrosomes (red arrow), a behavior that we never
observe in S. pombe spindles. This behavior suggests a possible mechanism for the apparently
decreased rigidity of S. japonicus spindles: they may either be less densely or robustly
crosslinked, allowing microtubules to detach from the main spindle bundle when subjected to
sufficient compressive force. While examples of extreme spindle bending are quite rare among
cerulenin-treated S. japonicus, occurring in only two of the observed cells, this phenomenon
was never observed in cerulenin-treated S. pombe. It is also important to note that, although this
was only observed in two cells, we only found three curved spindles in all of our imaging of
cerulenin-treated S. japonicus. We suppose that this lack of curved spindles is likely due to the
frequency at which S. japonicus nuclear envelopes locally break down, thereby relieving nuclear
envelope tension and allowing spindles to straighten, a noted difference between untreated S.
pombe and S. japonicus cells1.
In total, our results demonstrate significant mechanical coupling, during mitosis, between
the mitotic spindle and nuclear envelope in fission yeasts. The rapid spindle collapse following
spindle ablation in cerulenin-treated S. pombe suggests the presence of a compressive force on
the spindle poles, resulting from increased nuclear envelope tension. Likewise, the
simultaneous nuclear envelope rupture and spindle relaxation after nuclear envelope ablation
implies this compressive force can be lessened by relieving tension in the nuclear envelope.
The high degree of crosslinking in the S. pombe spindle likely allows it to oppose a typical level
of nuclear envelope tension to remain straight during normal mitosis, when nuclear envelope
expansion accommodates an elongating anaphase spindle. In contrast, S. japonicus spindles
appear less crosslinked, since their microtubules more easily splay apart from the main bundle
and the spindle as a whole bends more readily. These mechanical features are consistent with a
lower need to oppose nuclear envelope tension in semi-open mitosis. While previous work has
described differences in the nuclear envelopes of the two species during mitosis1,30,31, little is
known about the molecular-level structure and organization of the S. japonicus mitotic spindle.
Future work could provide further insight into how the mechanical properties of these two
functionally integrated structures, the mitotic spindle and the nuclear envelope, may have
co-evolved to support robust chromosome segregation and genetic integrity.
Materials and Methods
RESOURCE AVAILABILITY
Lead contact
Further information and requests for resources and reagents should be directed to and will be
fulfilled by the Lead Contact, Mary Williard Elting (mary.elting@ncsu.edu).
FISSION YEAST STRAINS AND CULTURE
All experiments were performed using cells from the two fission yeast species,
Schizosaccharomyces pombe and Schizosaccharomyces japonicus. For strain details, see Key
Resources table. S. pombe strains MWE2 and MWE3 are from Fred Chang lab stock (original
strains FC2861 and FC198423, respectively). MWE48 was created by crossing S. pombe strains
MWE2 and MWE40, which was from Gautam Dey lab stock (original strain GD20832). Cross was
performed by tetrad dissection using standard methods33. S. japonicus strain MWE92 was from
Snezhana Oliferenko lab stock (original strain SOJ3981). Both S. pombe and S. japonicus
strains were cultured at 25C on YE5S plates using standard techniques33. For imaging, liquid
cultures were grown in YE5S media at 25 °C with shaking by a rotating drum for 12-24 hours
before imaging. To ensure that cells were in growth phase for imaging, we measured OD595
with a target of 0.1-0.2. If cells had grown beyond this point, we diluted them and allowed them
to recover for ~1 hour before imaging.
As a method of increasing nuclear envelope tension in S. pombe and S. japonicus, cells
were treated with 1mM cerulenin, from stock solutions at 50 mM in DMSO, as previously
described1,27.
LIVE CELL IMAGING AND LASER ABLATION
Spinning disk confocal live imaging and laser ablation experiments were performed similar to
those described previously9,12,14. Live videos were captured using a Nikon Ti-E stand on an
Andor Dragonfly spinning disk confocal fluorescence microscope; spinning disk dichroic Chroma
ZT405/488/561/640rpc; 488 nm (50 mW) diode laser (240 ms exposures) with Borealis
attachment (Andor); emission filter Chroma Chroma ET525/50m; and an Andor iXon3 camera.
Imaging was performed with a 100x 1.45 Ph3 Nikon objective and a 1.5x magnifier (built-in to
the Dragonfly system). An Andor Micropoint attachment with galvo-controlled steering was used
for targeted laser ablation, delivering 20-30 3 ns pulses of 551 nm light at 20 Hz. Each cell was
imaged until either spindle repair had been completed or the spindle had failed to repair after ~5
min. For spindle ablation videos frames were collected every 3.5 s, while frames were collected
every 280 ms for nuclear envelope ablation videos. Andor Fusion software was used to manage
imaging and Andor IQ software was used to simultaneously manipulate the laser ablation
system.
Prior to imaging, samples were placed onto gelatin pads on microscope glass slides. For
gelatin pads, 125 mg gelatin was added to 500 μL EMM5S and heated, in a tabletop dry heat
bath at 90oC for at least 20 min. A small sample volume (~5 μL) of the gelatin mixture was
pipetted onto each slide, covered with a coverslip, and given a minimum of 30 min to solidify.
For each microscope slide, 1 mL volume of cells suspended in YE5S liquid growth media were
centrifuged (enough to see a pellet), using a tabletop centrifuge. Nearly all the supernatant was
decanted and the cells were resuspended in the remaining supernatant. Next, 2 μL of
resuspended cells were pipetted onto the center of the gelatin pad, which was immediately
covered with a cover slip. Finally, the coverslip is sealed using VALAP (1:1:1:
Vaseline:lanolin:paraffin). All samples, sealed between the gelatin pads and coverslips, were
imaged at room temperature (~22oC).
QUANTIFICATION AND STATISTICAL ANALYSIS
Image and video preparation and editing
To optimize the identification and tracking of spindle and nuclear envelope features,
modifications were made to fluorescence microscopy images and videos using FIJI34. First,
images and videos were cropped to show only cells of interest and extra frames were
eliminated. Typically, linear adjustments were made to the brightness and contrast of the
images, in order to track features more clearly. For measurements of NLS intensity, however,
pixel intensities were measured from unadjusted images. For immunofluorescence images, the
same brightness and contrast scaling was used for all images in each set.
Tracking of spindle features in ablation videos
All quantitative data regarding post-ablation spindle dynamics was collected via a tracking
program, home-written in Python. For each ablated spindle, the two spindle poles are tracked
following ablation, using this program, and the ‘line’ tool in FIJI is used to measure the length of
each spindle immediately prior to ablation. This data is then used to calculate the change in pole
separation (length) for each spindle over time, following ablation. Additionally, the positions of
the two new plus-ends of each ablated spindle half are tracked throughout the video. Our
tracking program includes a method for indicating whether or not spindle repair has occurred,
with the reconnection of the two ablated spindle halves, in each frame. The data for frames
collected before the reformation of a single spindle is used to compute time traces for the
change in spindle half length.
We define spindle half splaying as the lateral dissociation of microtubule ends from each
other in ablated spindle halves. Splay state is determined by eye in FIJI, using videos in which
the brightness and contrast have been linearly adjusted for clarity. If a spindle half appears
splayed in any frame of a video, that spindle half is counted as splaying. Otherwise, it is counted
as a spindle that does not splay. For videos in which splay state is not readily apparent
throughout, for one or both spindle halves, no splaying designation is made for the unclear
half/halves. In some cases, a spindle half depolymerizes within the first few minutes following
ablation and is therefore not included in our splay state analysis.
Quantification of spindle relaxation and nucleoplasm leakage
For all nuclear envelope ablation videos, data was collected on the time-evolution of spindle
curvature using a home-written Matlab program. The program takes microscope image files and
fits a quadratic function to a chosen object in the image. It then outputs curvature and length
data from that fitted curve. For videos this process was semi-automated to perform the fit
frame-by-frame.
Another program, home-written in Python was used to track the rate of nucleoplasm
leakage from the nucleus of each cell following nuclear envelope ablation. This program
requires the unadjusted video, spindle curvature data, and spindle length data as inputs. Using
these inputs, nuclear intensity is calculated for each frame as the average GFP intensity (after
background subtraction) of a 25-pixel square near the center of the nucleus. This same program
is then used to compute best-fit curves for both post-ablation spindle curvature and nuclear
intensity.
KEY RESOURCES TABLE
REAGENT of RESOURCE
SOURCE
IDENTIFIER
Chemicals, peptides, and other recombinant proteins
Cerulenin
Genesee Scientific
Cat# J64538.#0
Experimental models: Organisms/strains
S. pombe h+ GFP-atb2:kanMX ade6- leu1-32
ura4-D18
F. Chang (FC2861)
MWE2
S. pombe h? ase1::KanMX
leu1-32::SV40-GFP-atb2[LEU1] leu1-32
ura4-D18 his7+
F. Chang (FC198423)
MWE3
S. pombe h- pBIP1-NLS-GFP-NLS:leu1+ ade-
leu1-32 ura4D-18
G. Dey (GD20832)
MWE40
S. pombe pBIP1-NLS-GFP-NLS:leu1+
GFP-atb2:kanMX
This work (cross of
MWE2 and MWE40)
MWE48
S. japonicus GFP-Atb::ura4+
Nup189-mCherry::ura4+ urasj-D3
ade6sj-domE
S.Oliferenko
(SOJ3981)
MWE92
Software and algorithms
Jupyter Notebook
Project Jupyter
MATLAB
MathWorks
Fiji
ImageJ
ACKNOWLEDGEMENTS
We thank M. Betterton, G. Dey, F. Chang, A. Molînes, S. Oliferenko, and members of the Elting
lab for advice and helpful discussions, and we thank F. Chang, G. Dey, and S. Oliferenko for
strains. We thank the Weninger Lab (NCSU) and Wang Lab (NCSU) for sharing lab space and
equipment. We thank the Cellular and Molecular Imaging Facility (CMIF) at NCSU for
microscopy support. This work was supported by NIH 1R35GM138083 and NSF 2133276.
REFERENCES
1.
Yam, C., He, Y., Zhang, D., Chiam, K.-H., and Oliferenko, S. (2011). Divergent strategies for
controlling the nuclear membrane satisfy geometric constraints during nuclear division.
Curr. Biol. 21, 1314–1319. 10.1016/j.cub.2011.06.052.
2.
Mori, R., and Oliferenko, S. (2020). Cell Biology: An Open Solution for Closed Mitosis. Curr.
Biol. 30, R942–R944. 10.1016/j.cub.2020.06.067.
3.
Zhu, Q., Zheng, F., Liu, A.P., Qian, J., Fu, C., and Lin, Y. (2016). Shape Transformation of
the Nuclear Envelope during Closed Mitosis. Biophys. J. 111, 2309–2316.
10.1016/j.bpj.2016.10.004.
4.
Fang, C., Zheng, F., Yao, J., Wei, X., Fu, C., Shi, X., and Lin, Y. (2020). A model for bridging
microtubule dynamics with nuclear envelope shape evolution during closed mitosis. J.
Mech. Phys. Solids 144, 104116. 10.1016/j.jmps.2020.104116.
5.
Ding, R., McDonald, K.L., and McIntosh, J.R. (1993). Three-dimensional reconstruction and
analysis of mitotic spindles from the yeast, Schizosaccharomyces pombe. Journal of Cell
Biology 120, 141–151. 10.1083/jcb.120.1.141.
6.
Krüger, L.K., Gélin, M., Ji, L., Kikuti, C., Houdusse, A., Théry, M., Blanchoin, L., and Tran,
P.T. (2021). Kinesin-6 Klp9 orchestrates spindle elongation by regulating microtubule sliding
and growth. Elife 10, e67489. 10.7554/eLife.67489.
7.
Thomas, E.C., Ismael, A., and Moore, J.K. (2020). Ase1 domains dynamically slow
anaphase spindle elongation and recruit Bim1 to the midzone. Mol. Biol. Cell 31,
2733–2747. 10.1091/mbc.E20-07-0493-T.
8.
Tolić-Nørrelykke, I.M., Sacconi, L., Thon, G., and Pavone, F.S. (2004). Positioning and
elongation of the fission yeast spindle by microtubule-based pushing. Curr. Biol. 14,
1181–1186. 10.1016/j.cub.2004.06.029.
9.
Uzsoy, A.S.M., Zareiesfandabadi, P., Jennings, J., Kemper, A.F., and Elting, M.W. (2021).
Automated tracking of S. pombe spindle elongation dynamics. J. Microsc. 284, 83–94.
10.1111/jmi.13044.
10. Loïodice, I., Staub, J., Setty, T.G., Nguyen, N.-P.T., Paoletti, A., and Tran, P.T. (2005).
Ase1p organizes antiparallel microtubule arrays during interphase and mitosis in fission
yeast. Mol. Biol. Cell 16, 1756–1768. 10.1091/mbc.E04-10-0899.
11. Nabeshima, K., Nakagawa, T., Straight, A.F., Murray, A., Chikashige, Y., Yamashita, Y.M.,
Hiraoka, Y., and Yanagida, M. (1998). Dynamics of Centromeres during
Metaphase–Anaphase Transition in Fission Yeast: Dis1 Is Implicated in Force Balance in
Metaphase Bipolar Spindle. MBoC 9, 3211–3225. 10.1091/mbc.9.11.3211.
12. Zareiesfandabadi, P., and Elting, M.W. (2022). Force by minus-end motors Dhc1 and Klp2
collapses the S. pombe spindle after laser ablation. Biophys. J. 121, 263–276.
10.1016/j.bpj.2021.12.019.
13. Khodjakov, A., La Terra, S., and Chang, F. (2004). Laser microsurgery in fission yeast; role
of the mitotic spindle midzone in anaphase B. Curr. Biol. 14, 1330–1340.
10.1016/j.cub.2004.07.028.
14. Begley, M.A., Solon, A.L., Davis, E.M., Sherrill, M.G., Ohi, R., and Elting, M.W. (2021).
K-fiber bundles in the mitotic spindle are mechanically reinforced by Kif15. Mol. Biol. Cell
32, br11. 10.1091/mbc.E20-06-0426.
15. Maiato, H., Rieder, C.L., and Khodjakov, A. (2004). Kinetochore-driven formation of
kinetochore fibers contributes to spindle assembly during animal mitosis. J. Cell Biol. 167,
831–840. 10.1083/jcb.200407090.
16. Spurck, T.P., Stonington, O.G., Snyder, J. a., Pickett-Heaps, J.D., Bajer, A., and Mole-Bajer,
J. (1990). UV microbeam irradiations of the mitotic spindle. II. Spindle fiber dynamics and
force production. J. Cell Biol. 111, 1505–1518.
17. Brugués, J., Nuzzo, V., Mazur, E., and Needleman, D.J. (2012). Nucleation and transport
organize microtubules in metaphase spindles. Cell 149, 554–564.
10.1016/j.cell.2012.03.027.
18. Sheykhani, R., Baker, N., Gomez-Godinez, V., Liaw, L.-H., Shah, J., Berns, M.W., and
Forer, A. (2013). The role of actin and myosin in PtK2 spindle length changes induced by
laser microbeam irradiations across the spindle. Cytoskeleton 70, 241–259.
10.1002/cm.21104.
19. Stoppin-Mellet, V., Fache, V., Portran, D., Martiel, J.-L., and Vantard, M. (2013). MAP65
coordinate microtubule growth during bundle formation. PLoS One 8, e56808.
10.1371/journal.pone.0056808.
20. Ichihara, K., Kitazawa, H., Iguchi, Y., Hotani, H., and Itoh, T.J. (2001). Visualization of the
stop of microtubule depolymerization that occurs at the high-density region of
microtubule-associated protein 2 (MAP2). J. Mol. Biol. 312, 107–118.
10.1006/jmbi.2001.4934.
21. Janson, M.E., Loughlin, R., Loïodice, I., Fu, C., Brunner, D., Nédélec, F.J., and Tran, P.T.
(2007). Crosslinkers and motors organize dynamic microtubules to form stable bipolar
arrays in fission yeast. Cell 128, 357–368. 10.1016/j.cell.2006.12.030.
22. Yamashita, A., Sato, M., Fujita, A., Yamamoto, M., and Toda, T. (2005). The roles of fission
yeast ase1 in mitotic cell division, meiotic nuclear oscillation, and cytokinesis checkpoint
signaling. Mol. Biol. Cell 16, 1378–1395. 10.1091/mbc.E04-10-0859.
23. Bratman, S.V., and Chang, F. (2007). Stabilization of overlapping microtubules by fission
yeast CLASP. Dev. Cell 13, 812–827. 10.1016/j.devcel.2007.10.015.
24. Rincon, S.A., Lamson, A., Blackwell, R., Syrovatkina, V., Fraisier, V., Paoletti, A., Betterton,
M.D., and Tran, P.T. (2017). Kinesin-5-independent mitotic spindle assembly requires the
antiparallel microtubule crosslinker Ase1 in fission yeast. Nat. Commun. 8, 15286.
10.1038/ncomms15286.
25. Braun, M., Lansky, Z., Fink, G., Ruhnow, F., Diez, S., and Janson, M.E. (2011). Adaptive
braking by Ase1 prevents overlapping microtubules from sliding completely apart. Nat. Cell
Biol. 13, 1259–1264. 10.1038/ncb2323.
26. Ding, R., West, R.R., Morphew, D.M., Oakley, B.R., and McIntosh, J.R. (1997). The spindle
pole body of Schizosaccharomyces pombe enters and leaves the nuclear envelope as the
cell cycle proceeds. Mol. Biol. Cell 8, 1461–1479. 10.1091/mbc.8.8.1461.
27. Awaya, J., Ohno, T., Ohno, H., and Omura, S. (1975). Substitution of cellular fatty acids in
yeast cells by the antibiotic cerulenin and exogenous fatty acids. Biochim. Biophys. Acta
409, 267–273. 10.1016/0005-2760(75)90022-3.
28. Zhang, D., and Oliferenko, S. (2013). Remodeling the nuclear membrane during closed
mitosis. Curr. Opin. Cell Biol. 25, 142–148. 10.1016/j.ceb.2012.09.001.
29. Gu, Y., Yam, C., and Oliferenko, S. (2012). Divergence of mitotic strategies in fission
yeasts. Nucleus 3, 37–41. 10.4161/nucl.19514.
30. Makarova, M., and Oliferenko, S. (2016). Mixing and matching nuclear envelope
remodeling and spindle assembly strategies in the evolution of mitosis. Curr. Opin. Cell Biol.
41, 43–50. 10.1016/j.ceb.2016.03.016.
31. Makarova, M., Gu, Y., Chen, J.-S., Beckley, J.R., Gould, K.L., and Oliferenko, S. (2016).
Temporal Regulation of Lipin Activity Diverged to Account for Differences in Mitotic
Programs. Curr. Biol. 26, 237–243. 10.1016/j.cub.2015.11.061.
32. Dey, G., Culley, S., Curran, S., Schmidt, U., Henriques, R., Kukulski, W., and Baum, B.
(2020). Closed mitosis requires local disassembly of the nuclear envelope. Nature 585,
119–123. 10.1038/s41586-020-2648-3.
33. Forsburg, S.L., and Rhind, N. (2006). Basic methods for fission yeast. Yeast 23, 173–183.
10.1002/yea.1347.
34. Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V., Longair, M., Pietzsch, T.,
Preibisch, S., Rueden, C., Saalfeld, S., Schmid, B., et al. (2012). Fiji: an open-source
platform for biological-image analysis. Nat. Methods 9, 676–682. 10.1038/nmeth.2019.
| 2022 | Pushing the envelope: force balance in fission yeast closed mitosis | 10.1101/2022.12.28.522145 | [
"Begley Marcus A",
"Pagán Medina Christian",
"Zareiesfandabadi Parsa",
"Rapp Matthew B",
"Elting Mary Williard"
] | creative-commons |
Diminished miRNA activity is associated with aberrant cytoplasmic
intron retention in ALS pathogenesis
Marija Petric-Howe1,2, Hamish Crerar1,2, Jacob Neeves1,2, Giulia E. Tyzack1,2,
Rickie Patani1,2,# , Raphaëlle Luisier5,#
1The Francis Crick Institute, 1 Midland Road, London NW1 1AT, UK; 2Department of
Molecular Neuroscience, UCL Institute of Neurology, Queen Square, London, UK;
4Idiap Research Institute, Genomics and Health Informatics, Martigny, Switzerland;
#These authors contributed equally to this work.
Key words: Cytoplasmic intron retention, human stem cell model, amyotrophic
lateral sclerosis, miRNA
Correspondence
should
be
addressed
to
Raphaëlle
Luisier
(raphaelle.luisier@idiap.ch) and Rickie Patani (rickie.patani@ucl.ac.uk).
SUMMARY
Intron retention (IR) is now recognized as a dominant splicing event during motor
neuron (MN) development, however the role and regulation of intron-retaining
transcripts (IRTs) localized to the cytoplasm remain particularly understudied. By
resolving the spatiotemporal dynamics of IR underlying distinct stages of MN
lineage restriction, we identify a cytoplasmic group of IRTs that is not associated
with reduced expression of their own genes but instead with an upregulation of
predicted target genes of specific miRNAs, the motifs of which are enriched within
the intronic sequences of this group. Next, we show that ALS-causing VCP
mutations lead to a selective increase in IR of this particular class of introns. This in
turn temporally coincides with an increase in the expression level of predicted
target genes of these miRNAs, providing a potential mechanistic insight into ALS
pathogenesis. Altogether, we propose a novel role for the cytoplasmic intronic
sequences in regulating miRNA activity through miRNA sequestration, which
potentially contributes to ALS pathogenesis.
INTRODUCTION
Intron retention, a mode of alternative splicing whereby one or more introns are
retained within a mature polyadenylated mRNA, has been greatly understudied in
mammalian systems and for a long time mostly considered as a product of
inefficient or mis-splicing. With advances in detection strategies, IR became
recognised as a more widespread and regulated process than previously thought,
and the idea that IR could even functionally modulate cellular processes has come
into focus, with its role(s) in cellular physiology beginning to unfold (1, 2).
Neural cells exhibit a higher proportion of retained introns compared to
other cell types and there is an expanding body of evidence demonstrating a
2
functional role for intron retention (IR) both in neuronal development and
homeostasis (1, 3–5). Transcripts that exhibit IR often remain in the nucleus, mostly
considered to be a means of reducing the expression levels of transcripts not
required for cellular physiology at a particular stage (2, 6, 7). Some of these
transcripts would eventually be degraded by the nuclear exosome, while specific
signals could stimulate splicing of the retained intron in others, resulting in export
of the fully spliced mRNA into the cytoplasm and its subsequent translation (5).
Indeed, nuclear detention of intron-retaining transcripts (IRTs) provides a powerful
mechanism to hold gene expression in a suppressed but poised state that allows
rapid protein production if and when an appropriate stimulus is received (4, 5,
8–10).
Although the stable cytoplasmic localisation of intronic sequences in neurons
has been reported since 2013 (11), there has been limited investigation into the
possible role of cytoplasmic IRTs. This has presumably been overlooked in part due
to detection limitations, but also due to a notion that these transcripts would likely
contain premature translation termination codons (PTCs) and as such, be degraded
by nonsense mediated mRNA decay (NMD) (12). Whilst examples of IR coupled with
NMD have been found to downregulate gene expression, such as in granulocyte
development (13), these transcripts can encounter other fates in the cell (14).
Indeed, one of the few studies focussing on cytoplasmic IR in neurons showed an
‘addressing’
function
for
intronic
RNA
sequences,
determining
the
spatial
localization
of their host transcripts within cellular compartments such as
dendrites (15).
Another speculated function of IR has arisen following the
identification of miRNA binding motifs within the retained intronic sequences. This
offers an intriguing route through which miRNA-directed degradation pathways
might regulate abundance of IRTs; alternatively, the retained introns themselves
may serve as miRNA sinks, or even encode novel miRNAs termed mirtrons (16, 17).
3
Altogether, despite advances in our understanding of IR in neuronal cells, much
remains unanswered.
The
importance
of
investigating
the
roles
of
IR
has
been
further
corroborated by studies that demonstrate its relevance across a diverse range of
neurodegenerative diseases (18–21). One such example is amyotrophic lateral
sclerosis (ALS) (20), a rapidly progressive and incurable disease, which leads to
selective degeneration of motor neurons (MNs). ALS is characterised by protein
inclusions and axonal degeneration, and is often associated with RNA processing
defects.
ALS-causing
mutations
occur
in numerous genes encoding crucial
regulators
of
RNA-processing,
which
are
normally
expressed
throughout
development. Despite the growing number of causative gene mutations being
identified in ALS, the precise aetiology remains unknown and early molecular
pathogenic events remain poorly understood. We previously made the novel
discovery that aberrant IR is a widespread phenomenon in ALS (20), which was
corroborated by subsequent studies (21, 22). Moreover, we went on to demonstrate
aberrant cytoplasmic IR as a widespread molecular phenomenon in VCP-related
ALS
(23).
We
showed
that
ALS-related
aberrant
cytoplasmic
IRTs
have
conspicuously high affinity for RNA binding proteins (RBPs), including those that
are mislocalized in ALS and proposed that a subset of cytoplasmic intronic
sequences serve as ‘blueprints’ for the hallmark protein mislocalization events in
ALS (24, 25). This raises an exciting possibility that intronic RNA sequences play
additional significant roles beyond their recognized nuclear function. Nevertheless,
the
role
and
physiological
relevance
of
cytoplasmic
IR
during
neuronal
development and disease still remains largely unresolved.
Against this background we sought to characterise the spatiotemporal
dynamics of IRTs by re-analysing RNA-seq data from nuclear and cytoplasmic
fractions of patient-specific hiPSCs undergoing motor neurogenesis. We first show
4
that
retained
introns
exhibit
compartment-specific
features
including
their
dynamics, biological pathways, and molecular characteristics during this process.
We reveal a specific class of retained introns in the cytoplasm that is not associated
with gene expression changes but exhibits high miRNA binding potential, which is
functionally validated by identifying an altered expression profile of the predicted
miRNA target genes. We finally analyze this class of retained introns in stem cell
models of familial ALS and find evidence for a functional depletion of specific
miRNAs,
possibly
as
a
result
of
cytoplasmic
intronic
sequences-mediated
sequestration, which has potential implications for ALS pathogenesis and the
development of therapies in this devastating and incurable disease.
MATERIALS AND METHODS
Compliance with ethical standards
Informed consent was obtained from all patients and healthy controls in this study.
Experimental protocols were all carried out according to approved regulations and
guidelines by UCLH’s National Hospital for Neurology and Neurosurgery and UCL’s
Institute of Neurology joint research ethics committee (09/0272).
RNA-sequencing data
We
obtained
paired-end
polyA
stranded
RNAseq
libraries
prepared
from
fractionated nucleus and cytoplasm obtained from 6 distinct stages of motor
neuron differentiation from control and VCPmu samples (iPSC, and days 3, 7, 14, 21
and 35; Supplementary Table S1) from previously published study (GSE152983) (23).
We also obtained paired-end RNA sequencing reads derived from one independent
study on familial form of ALS caused by mutant SOD1 (n=5; 2 patient-derived
SOD1A4V and 3 isogenic control MN samples where the mutation has been
corrected; Hb9 FACS purified MNs, GSE54409 (26).
5
Transcript and Gene expression analysis
Kallisto (27) was used to (1) build a transcript index from the Gencode hg38 release
Homo sapiens transcriptome (-k 31), (2) pseudo-align the RNA-seq reads to the
transcriptome and (3) quantify transcript abundances (-b 100 -s 50—rf-stranded).
Subsequent analysis was performed with the R statistical package version 3.3.1
(2016) and Bioconductor libraries version 3.3 (R Core Team. R: A Language and
Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical
Computing; 2013). Kallisto outputs transcript abundance, and thus we calculated
the abundance of genes by summing up the estimated raw count of the constituent
isoforms to obtain a single value per gene. For a given sample, the histogram of log2
gene count is generally bimodal, with the modes corresponding to non-expressed
and expressed genes. Reliably expressed genes/transcripts for each condition
(VCPmu or control at days 0, 3, 7, 14, 22 and 35 in each fraction) were next identified
by fitting a two-component Gaussian mixture to the log2 estimated count
gene/transcript data with R package mclust (28) ; a pseudocount of 1 was added
before log2 transformation. A gene/transcript was considered to be reliably
expressed
in
a
given
condition
if
the
probability
of
it
belonging
to
the
non-expressed class was under 1% in each sample belonging to the condition.
18,834 genes and 102,047 transcripts were selected based on their detected
expression in at least one of the 24 conditions (i.e. 6 different timepoints of lineage
restriction for control and VCPmu in nuclear and cytoplasm). Next we quantile
normalized the columns of the count matrices with R package limma (29). For
differential gene expression analysis we ran Sleuth (30).
6
Splicing analysis
The identification of all classes of alternative splicing (AS) events in motor neuron
differentiation
was
performed
with
the Vertebrate Alternative Splicing and
Transcription Tools (VAST-TOOLS) toolset, which works synergistically with the
VastDB web server, a collection of species-specific alternative splicing library files
(31). Paired-end stranded RNA-seq reads were first aligned with VAST-TOOLS
against the Homo sapiens hg38, Hs2 assembly from VastDB with the scaffold
annotation Ensembl v88. This contains 74030 exon skipping events, 153119 intron
retention events, 474 microexon events, 20812 alternative 3’ UTR events, and 15804
alternative 5’ UTR events. We then merged files from identical samples but different
lanes together and then performed differential splicing analysis over time either for
the control or for the VCP mutant samples separately using the vast-tools diff
command which takes into account the different biological replicates. We then
imported the result tables into R. For an AS event to be considered differentially
regulated between two conditions, we required a minimum average ΔPSI (between
the paired replicates) of at least 15%
and that the transcript targeted by the
splicing event in question to be reliably expressed in all samples from the
conditions compared i.e enough read coverage in all samples of interest.
Gene ontology enrichment analysis
GO enrichment analysis was performed using classic Fisher test with topGO
Bioconductor package (32). Only GO terms containing at least 10 annotated genes
were considered. A p-value of 0.05 was used as the level of significance. On the
figures, top significant GO terms were manually selected by removing redundant
GO terms and terms which contain fewer than 5 significant genes.
7
Mapping and analysis of CLIP data
To identify RBPs that bind to retained introns, we examined iCLIP data for 21 RBPs
(33), and eCLIP data from K562 and HepG2 cells for 112 RBPs available from
ENCODE (34, 35). Before mapping the reads, adapter sequences were removed
using Cutadapt v1.9.dev1 and reads shorter than 18 nucleotides were dropped from
the analysis. Reads were mapped with STAR v2.4.0i (36) to UCSC hg19/GRCh37
genome assembly. The results were lifted to hg38 using liftOver (37). To quantify
binding to individual loci, only uniquely mapping reads were used.
Analysis of cis-acting features
MaxEntScan (38) was used to calculate maximum entropy scores for 9-bp 5’ splice
sites and 23-bp 3’ splice sites. Intron lengths and GC content were calculated using
the hg38 human genome assembly. The intronic enrichment for RBP binding site
was obtained by computing the proportion of crosslink events mapping to retained
intron compared to non-retained introns of the same genes, accounting for intron
length.
These were defined in relation to the acceptor and donor splice sites,
namely the last 30 nucleotides (nts) of exonic sequence upstream of the 5’ splice
site (R1), the first 30nts of intronic sequence downstream of the 5’ splice site (R2),
the 30nts in the middle of the intron (R3), the last 30nts of intronic sequence
upstream of the 3’ splice site (R4), and the first 30ntsof exonic sequence
downstream of the 3’ splice site (R5). These regions were defined based on the past
studies of the Nova RNA splicing map (39), which has been determined by the
positioning of conserved YCAY clusters as well as by the binding sites identified by
HITS-CLIP as reported in (40). The nucleotide-level evolutionary phastCons scores
for multiple alignments of 99 vertebrate genomes to the human genome were
obtained from UCSC (41, 42) and a median score was derived for each individual
8
intron and defined regions of interest. The RBP crosslink event enrichment scores
in each region of interest or in each group of intron was obtained by dividing the
fractions of introns in a given group over the fraction in the full list of introns that
exhibit at least one crosslink event for a given RBP in the defined region or across
the full intronic region.
Spatiotemporal taxonomisation of the retained introns
We performed singular value decomposition (SVD) on the PIR cytoplasmic versus
nuclear values of 94,457 introns in n = 48 cytoplasmic samples and n=47 nuclear
samples across the 5 distinct stages of motor neuron differentiation from healthy
controls and VCP mutants. We analysed 94,457 introns out of the 153,119 annotated
introns in VAST-TOOLS given their overlap with reliably expressed genes. We then
selected the components maximally capturing variance in PIR. To visualize the right
singular vectors
, we plotted the PIR on the vertical axis as a function of the
time corresponding to each sample on the horizontal axis and coloring all samples
corresponding to healthy controls with filled circles, and those corresponding to
VCP-mutants in empty circles. Next we identified introns whose PIR profiles
correlated (Pearson correlation between individual intron PIR profile and right
singular vectors) and contributed (projection of each individual intron PIR profile
onto right singular vectors) most strongly (either positively or negatively) with the
profile of the singular vectors. In order to identify representative introns for each
singular vector, events were ranked according to both projection and correlation
scores. The highest (most positive scores in both projection and correlation) and
lowest (most negative scores in both correlation and projection) motifs were
selected for each singular vector using K-mean clustering.
9
MiRNA expression analysis
Total RNA including small RNAs was extracted from “patterned” precursor motor
neurons of five control and four mutant cell lines using mirVana™ miRNA Isolation
Kit (ambion, life technologies). RNA quantification and its 260/280 ratio were
assessed using the nanodrop. Poly(A) tailing and reverse transcription of mature
miRNAs was performed using miRCURY LNA RT kit (QIAGEN), with 20 ng of total
RNA as input. Reverse transcribed cDNA was quantified using miRCURY LNA SYBR
green dye, specifically designed primers (appropriate miRCURY LNA miRNA PCR
assays) and QuantStudio 6 Flex Real-Time PCR System (Applied Biosystems).
Relative miRNA expression levels between control and mutant cells were quantified
using ΔΔCT Method, with U6 snRNA as a reference gene for normalisation. Data
was plotted using RStudio software.
RESULTS
Nuclear and cytoplasmic IR affect two functionally divergent mRNA subsets
We previously reported a transient IR programme early during human motor
neurogenesis using whole-cell RNA-sequencing data (20). To further examine the
spatiotemporal dynamics of IR during this process in healthy cells, we re-analysed
high-throughput poly(A) RNA-seq data derived from nuclear and cytoplasmic
fractions
of human induced pluripotent stem cells (hiPSCs; day 0), neural
precursors (NPCs; day 3 and day 7), ‘patterned’ ventral spinal motor neuron
precursors (pMNs; day 14), post-mitotic but electrophysiologically immature motor
neurons (MNs; day 22), and electrophysiologically active MNs (mMNs; day 35) (Fig.
1A and Supplementary Table 1; 47 nuclear and cytoplasmic samples from 6
time-points; 4 clones from 4 different healthy controls) (23, 43). Using the RNA-seq
pipeline VAST-TOOLS (31), we identified 4,189 nuclear and 1,542 cytoplasmic
10
significant alternative splicing (AS) events over time (Fig. 1B and Supplementary
Fig. 1A).
In line with our previous study (20), IR was the predominant mode of
splicing during neurodevelopment, accounting for 64% and 49% of the included AS
events in the nucleus (638 events) and the cytoplasm (541 events) respectively,
indicating that cytoplasmic IRTs are more abundant than previously recognized
(Fig. 1B). Further examining the distributions of percent intron retention (PIR)
during MN differentiation in the nucleus and the cytoplasm for 211,501 events
revealed that IR exhibits distinct dynamics in the two compartments (Fig. 1C). In
particular, the nuclear compartment exhibits the highest level of PIR at the hiPSC
stage, while the cytoplasmic compartment exhibits the highest level of PIR at
DIV=14, which is reminiscent of the early wave of IR we previously reported (20).
Notably the cytoplasmic increase in PIR early during differentiation is likely
explained by a change in the subcellular localisation of (some) IRTs rather than a
modulation of the splicing given the coincident stable level of PIR in the nucleus.
Genes related to RNA processing and splicing are among the most affected by IR
(13, 44–48) and we previously showed that IR early during MN development
specifically affects RNA processing related biological pathways (20). Here we find
that cytoplasmic (but not nuclear) IR affects essential genes concerned with mRNA
metabolism. In contrast we find that genes targeted by alternative exons (AltEx) are
enriched in similar biological pathways in the nucleus and the cytoplasm as shown
by Gene Ontology (GO) function analysis (Fig. 1D). These findings indicate that the
previously reported wave of IR during MN differentiation, using whole-cell
RNA-sequencing, likely reflected signals from cytoplasmic IRTs.
Prior studies reported specific features associated with retained introns
including higher GC content, lower intron length and enrichment in RBP binding
motifs compared to non-retained introns (1, 47, 49, 50). In line with these studies,
we find that the PIR negatively correlates with the intron length and positively
11
correlates with the GC content and the enrichment in crosslink events for 131 RBPs
for which CLIP data were available (33–35) both in nuclear and cytoplasmic
compartments (Figs. 1E, G, I and Supplementary Fig. 1F). Additionally we find that
i) retained introns are detected in genes containing fewer introns, and ii) the PIR
positively
correlates
with
the
intronic
sequence
conservation
score
(Supplementary Figs. 1B, D, F). Surprisingly, however, by specifically comparing the
cytoplasmic
retained
introns
(PIRNUCLEUS
>
20%
and
PIRCYTOPLASM
>
15%;
Supplementary Table S2) with the nuclearly detained retained introns (PIRNUCLEUS >
20% and PIRCYTOPLASM < 5%; Supplementary Table S3) we find that the retained
introns that localise to the cytoplasm are on average longer, have lower GC content
compared to their nuclear counterparts, have higher RBPs enrichment scores and
are more evolutionarily conserved (Figs. 1F, H, J and Supplementary Figs. 1C,E).
Altogether these results show that nuclear and cytoplasmic retained introns exhibit
distinct features including their dynamics during human motor neurogenesis, their
associated biological pathways, and their molecular characteristics.
A spatiotemporal taxonomy reveals cytoplasmic retained introns with distinct
RBP binding profiles
Having established that nuclear and cytoplasmic IR affect two functionally
divergent mRNA subsets, we next used singular value decomposition (SVD) analysis
to categorize 94,457 analysed introns into nine groups based on their PIR
spatiotemporal dynamics during MN differentiation (Fig. 2A and Supplementary
Tables S4-S12). Of these, three categories are nuclearly-detained retained introns
(termed N1-N3 hereafter) and the other six are cytoplasmic retained introns
(termed C1-C6 hereafter), that exhibit the following compartment-specific PIR
dynamics: stable nuclear (>15%)
and low cytoplasmic (<10%) PIR over time (N1),
12
steady reduction in the nuclear PIR over time and stable low cytoplasmic PIR (N2), a
transient increase in nuclear PIR and stable low cytoplasmic PIR (N3), steady
reduction in both nuclear and cytoplasmic PIR over time (C1), steady increase in
nuclear and cytoplasmic PIR over time (C2), increase in cytoplasmic PIR only in
terminal differentiation (C3), consistently high nuclear and cytoplasmic PIR over
time (C4), early transient increase in nuclear and cytoplasmic PIR (C5) and a late
transient increase in nuclear and cytoplasmic PIR (C6).
Next looking at the percentage of introns per genes targeted by IR in each
identified group revealed that IR during MN differentiation does not occur
stochastically, but appears to target a specific set of introns in each gene
(Supplementary Fig. 2A), indicating that additional layer(s) of specific regulation
must underlie the regulation of these 9 distinct IR programmes as previously
suggested (47). Previous studies suggest that cis-regulatory elements bound by
trans-acting factors such as RBPs are likely to play a crucial role in regulating IR
(50–52). Thus we next sought to test whether the 9 spatiotemporally distinct
classes
of
IR
we
identified
are
associated
with
different
combinations
of
trans-acting factors that could regulate them. We achieved this by using the
publicly available CLIP data to evaluate the crosslink events for 131 RBPs mapping to
5 regions we defined in relation to the acceptor and donor splice sites, namely the
last 30 nucleotides (nts) of exonic sequence upstream of the 5’ splice site (R1), the
first 30nts of intronic sequence downstream of the 5’ splice site (R2), the 30nts in
the middle of the intron (R3), the last 30nts of intronic sequence upstream of the 3’
splice site (R4), and the first 30nts of exonic sequence downstream of the 3’ splice
site (R5) (Fig. 2B, upper and Supplementary Tables S12-S17). First, looking at the
fractions of regions which are mapped by at least one crosslink event for each RBP,
we find that the R1 and R5 exonic regions, sequences of which are the most
evolutionarily conserved (Supplementary Fig. 2B), exhibit the highest frequency in
13
crosslink events across the 131 RBPs irrespective of the IR grouping, with the
exception of the C1 group (Fig. 2B, lower). These results, which are in line with
previous studies showing that the splicing machinery is more likely to form across
the exonic regions than across the introns for similarly long introns (>250 nts),
indicate that the 9 groups of introns bear similar chances of splicing complex
formation with respect to their R1 and R5 exonic regions (53, 54). This is further
supported by the finding that identically optimal splicing signals are detected
among the 9 groups of introns, with the exception of the C4 group (Supplementary
Fig. 2C) as failure in splice site recognition (50–52) or decreased expression levels
of splicing factors (47) have also been proposed to underlie IR.
Although RBPs exhibit similarly high frequency of binding to the R1 and R5
exonic regions across the majority of the 9 groups of introns, we indeed noticed
that the R2, R3 and R4 intronic regions display large variability in the percentages of
crosslink events across the different spatiotemporal IR dynamics (Fig. 2B, lower).
Next, looking at the enrichment of RBPs mapping to each of these regions further
revealed that the C3, C4 and C5 groups of cytoplasmic retained introns is indeed
specifically enriched in RBPs binding to the R2, R3 and R4 intronic regions as
opposed to the R1 and R5 regions which display as much RBP binding as the full set
of introns (Fig. 2C). One of the most enriched RBPs in the R2, R3 and R4 intronic
regions is UPF1, an RNA helicase required for nonsense-mediated mRNA decay
(NMD) in eukaryotes. UPF1 exhibits high binding occurrence across the five regions
of interest for the {C4, C5, C6} groups as opposed to the other groups for which
UPF1 is strictly enriched in R1 and R5 regions (Supplementary Fig. 2D). Different
combinations of RBPs have been shown to coordinately regulate functionally
coherent
“networks”
of
exons
and
introns
(55).
Thus,
using
unsupervised
hierarchical clustering of the 9 groups of introns based on their 131 RBP enrichment
scores profiles we finally showed that {C3, C4, C5} form a coherent regulated group
14
of introns in respect to all selected regions except the R1 exonic region (Fig. 2D).
Altogether, this analysis identifies a coherent regulated supergroup of retained
introns which exhibits specific elements in their intronic regions that are not
necessarily related to splicing efficiency, but rather may perform an additional role
in the regulation of mRNA metabolism.
Identification of a cytoplasmic group of IRTs with a high capacity for miRNA
sequestration
The finding that a coherent regulated supergroup of retained introns bears similar
chances of splicing complex formation with respect to their R1 and R5 exonic
regions but with a potential regulatory role in mRNA metabolism through RBP
intronic sequence binding, prompted us to look at the occurrence of RBP crosslink
events across the entire intronic region as opposed to the 5 predefined regions of
interest. We showed that while the {C4, C5, C6} groups have a somewhat higher
prevalence of crosslink events (Fig. 3A), the C5 group specifically has the highest
fraction of introns with at least one crosslink event across the 131 studied RBPs
compared to the full set of introns (Fig. 3B). Notably, the {C4, C5, C6} groups overall
have a lower intron length compared to {C1, C2, C3}, and thus this result cannot be
due to a bias in size. Of the 51 RBPs that exhibit >19% higher fraction of binding to
the C5 group compared to the full set (Fisher enrichment P-value < 0.01; Fig. 3C) at
least 9 are key regulators of mRNA transport such as UPF1 and IGF2BP1 (56–58), and
6 RBPs are involved in miRNA regulatory pathways such as DROSHA and PUM2
(59–61), UPF1 being involved in the two (62–64) (Figs. 3C-E). Noting that miRNA
regulators are avidly binding the C5’s intronic regions, we next looked at miRNA
motif enrichment within the introns using HOMER (65), which revealed significant
enrichment for 14 miRNA motifs (Supplementary Fig. 3A). C5 IR does not play a role
in gene expression regulation, as revealed by the analysis of fold-changes over time
15
of the genes containing the retained introns (Supplementary Fig. 3B), and may thus
serve other functions, particularly in the cytoplasm. Focusing on miR-4519,
miR-1976, miR-4716-5p, miR-485-5p and miR-4267, the top five miRNAs motifs
enriched in the C5 intronic sequences (Fig. 3F), we next sought to test whether
changes in the C5 PIR over time might relate to changes in gene expression of
specific miRNA predicted target genes, as a result of trapping/releasing these
miRNAs. To this end, we examined the gene expression profile of their predicted
target genes by combining two miRNA target prediction algorithms, TargetScan
(66) and miRanda (67). Strikingly we find that the predicted target genes of
miR-4519, miR-1976, and miR-4716-5p exhibit a reduction in expression from DIV=7
to DIV=14, which coincides with a reduction of the C5 introns PIR, while the
predicted target genes of miR-485-5p and miR-4267 do not exhibit such trend (Fig.
3G and Supplementary Figs. 5,6). This result indicates that the decrease in IR from
DIV=7 to DIV=14 correlates with an increase in miRNA activity, supporting the
hypothesis that cytoplasmic retained introns reduce miRNA activity potentially by
sequestering them, as previously shown for long non-coding RNA (lncRNA) (68).
VCP mutation-related transient accumulation of cytoplasmic IRTs correlates
with reduced miRNA activity.
We previously demonstrated aberrant cytoplasmic IRTs in ALS-related VCPmu
samples during MN differentiation that exhibit high predicted binding affinity for
RBPs (20, 23). We next sought to test whether VCP mutations affect any of the
cytoplasmic groups of introns in particular. Examining the two most prominent
classes of cytoplasmic IR dynamics during MN differentiation, as captured by right
singular vectors of the SVD analysis performed on the cytoplasmic PIR values,
confirmed prior findings that VCP mutations leads to exceptionally large IR
perturbations at DIV=14 (Supplementary Figs. 4A, B) (23). Further comparing the
16
PIR distributions between control and VCPmu samples in each of the six groups of
cytoplasmic retained introns revealed that VCP mutations specifically impact two
classes of events, namely C5, and to a much lesser extent C1 while the other groups
remain unchanged (Fig. 4A and Supplementary Fig. 4C). Most notably VCP-driven
changes in cytoplasmic IR are 1) unidirectional i.e. we only detect increases in IR in
VCPmu samples compared to control samples irrespective of PIR dynamics, and 2)
the VCP mutation specifically affects groups of introns in which the PIR exhibits a
large decrease from DIV=7 to DIV=14. This is in contrast to those groups of introns
where the PIR increases from DIV=7 to DIV=14, such as C2 and C6, where we find
similar increase in control and VCPmu samples (Supplementary Figs. 4C). These
results suggest that VCP mutations enhance the cytoplasmic stability of IRTs,
rather than affecting nuclear export, which would equally impact C1, C2, C5 and C6.
Having found that VCP mutations lead to large perturbations of the C5 group
of cytoplasmic IRTs at DIV=14, which we have shown above to associate with
decreased activity of specific miRNAs, we next sought to test whether the increase
in cytoplasmic IR of the C5 group in VCP mutants correlates with a reduction in
miRNA activity by looking at the changes in gene expression of their predicted
target genes between VCP and control samples. This analysis revealed that the
increase in IR in VCP mutant cultures correlates with a decrease in miR-4519,
miR-1976 and miR-4716-5p activities as predicted by the up-regulation of their
respective target genes at DIV=14 (Fig. 4B).
Additionally, these changes are not
explained by a change in the expression levels of these miRNAs which are not
significantly different between VCP and control cultures at this time point
(Supplementary Figure 4D). Notably, the predicted activities of miR-485-5p and
miR-4267, whose target gene expression profiles did not correlate with IR level in
control samples over time, also do not correlate with increase in C5 IR in VCP
17
samples, thus supporting the hypothesis that the activities of the same miRNAs
correlate to IR in both VCP and control samples over time (Supplementary Fig. 6).
Noting that we have studied a relatively rare form of familial ALS (fALS)
caused by gene mutations in VCP (selected as it exhibits the pathological hallmark
of TDP-43 nuclear-to-cytoplasmic mislocalisation), we next sought to understand
the generalizability of the association between increase in IR in ALS samples and
decrease in miRNA activity. To this end, we chose to study one of the most
common forms of fALS (SOD1), which in contrast does not exhibit the pathological
hallmark of TDP-43 nuclear-to-cytoplasmic mislocalisation. We first looked at the
PIR of the C5 group in SOD1 mutant hiPSC-derived MNs, revealing a statistically
significant increase in IR in SOD1 (Fig. 4C). Next, looking at the changes in gene
expression of the miRNA predicted target genes between SOD1 mutant MNs and
their isogenic controls, further showed a decrease in miR-4519, miR-1976 and
miR-4716-5p predicted activities (Fig. 4D). These findings further substantiate the
relevance of the correlation between increased IR and decreased miRNA activity.
Altogether these findings support the hypothesis that the cytoplasmic pool of C5
introns leads to a reduction in miRNA activity, potentially through direct binding
and sequestration, which may have important roles in ALS pathogenesis, and
indeed implications for new therapeutic strategies.
DISCUSSION
Neuronal biology relies on complex regulation of gene expression and mRNA
metabolism. Alternative splicing has been shown to play a key role in this process
and IR is now recognized as the dominant mode of splicing during MN development
(1, 20), including cytoplasmic IR, which we recently showed to affect >100
transcripts during neuronal development (23). Because nuclear IR has been the
focus of most previous studies, the regulation and role of cytoplasmic IRTs remain
18
unclear. The objective of this study was twofold: to deepen our understanding of
the role(s) of cytoplasmic IR in normal cellular physiology by resolving the
spatiotemporal dynamics of IR underlying distinct stages of MN lineage restriction,
and to decipher whether specific classes of IRTs become dysregulated in the
context of disease by systematically examining the influence of ALS-causing VCP
mutations on this process. In order to achieve this we re-analyzed nuclear and
cytoplasmic RNA-sequencing data from a time course of patient-specific iPSCs
differentiating into spinal MNs.
We first show that nuclear and cytoplasmic IR target distinct classes of
mRNA associated with particular dynamics, biological pathways and molecular
characteristics. Specifically, we find that the sequences of the retained introns that
localise to the cytoplasm are evolutionarily more conserved and exhibit a higher
capacity for RBP binding compared to the nuclearly detained introns. This argues
against the hypothesis that cytoplasmic intron-containing pre-mRNAs simply ‘leak’
from the nucleus (69), which is also further excluded by polyA selection during
library preparation, and suggests that 1) cytoplasmic localisation signals for these
IRTs are contained in the intronic sequences, and 2) cytoplasmic IRTs likely serve a
biological function that has yet to be discovered.
We next show that MN differentiation exhibits complex IR spatiotemporal
dynamics captured by 9 distinct IR programmes, 3 which are nuclearly detained
and 6 that localise to the cytoplasm. Given the time and cell compartment
specificity of these programmes, they are expected to associate with distinct
complex regulation. IR has been previously proposed to be the consequence of
globally inefficient splicing (47, 70), that could be linked to several mechanisms
including the occupancy of MeCP2 near the splice junction (71), the expression of
PRMT5 (7), and relatively weak splice sites (1). Here we find that the 9 groups of
introns exhibit similar 5’ and 3’ maximum entropy scores as well as similarly high
RBP binding in their exonic regions juxtaposed to the splice sites where the splicing
19
machinery is more likely to form (53, 54) as opposed to the intronic regions. These
findings
indicate
that
an
overall
change
in
splicing
efficiency
during
MN
differentiation is unlikely to be the dominant regulatory factor for most of these IR
programmes.
Furthermore,
IR
during
MN
differentiation
does
not
occur
stochastically, but appears to target a specific set of introns in each gene, and thus
an additional layer of specific regulation must underlie the regulation of these 9
distinct IR programmes as previously suggested (47). Indeed similar combinations of
trans-acting factors are detected across 4 regions juxtaposed to the splices sites
among 3 groups of cytoplasmic retained introns -{C4, C5, C6}- suggesting similar
regulation. Additionally these three groups of introns exhibit avid RBP binding
within their intronic regions juxtaposed to the splice sites when compared to the
full set of analysed introns, indicating a potential regulatory mechanism in mRNA
metabolism through intronic sequence binding for the {C4, C5, C6} groups of
introns. Notably the full list of retained introns for each group together with the
regional RBP enrichment is freely accessible as supplementary tables providing a
rich resource for researchers across the disciplines of genomics and basic
neuroscience.
Although the stable cytoplasmic localisation of intronic sequences in neurons
has been recognized since 2013 (11), their role has remained poorly understood. One
of the few studies focussing on cytoplasmic IR showed an ‘addressing’ function for
intronic RNA sequences in determining their spatial localization within cellular
compartments (15). Here we show that the avid RBP binding we previously observed
in ALS-related aberrant cytoplasmic retained introns (23) is indeed specifically
detected in one IR programme which exhibits a transient increase in the cytoplasm
during MN differentiation, namely the C5 group. Notably this group of cytoplasmic
intron, which is the most impacted by VCP mutations, exhibits the same PIR
dynamics of the group of introns we previously showed to be impacted by VCP
mutations using whole-cell RNA-sequencing (20). This suggests that VCP mutations
20
specifically affect the cytoplasmic stability of IRTs rather than leading to a
reduction in splicing efficiency. The absence of correlation between the C5 PIR
level and the gene expression dynamics during MN differentiation raises the
possibility of new roles for intronic RNA sequences beyond a function in gene
expression regulation, particularly in the cytoplasm. As previously proposed,
cytoplasmic retained introns may act as RNA regulators in the homeostatic control
of RBP localisation during development and disease (23), which may in turn lead to
loss of function. For example some splicing factors that avidly bind the C5 intronic
sequences may remain sequestered upon their nuclear export and cytoplasmic
localisation, contributing to a transient reduction in splicing efficiency during MN
differentiation. Another intriguing molecular characteristic of the C5 group of
introns is the enrichment for several miRNA motifs across the full length of the
intron, predicted activity for which negatively correlates with the PIR level of this
intron group during MN differentiation. Thus the presence of long intronic
sequences in the cytoplasm of neuronal cells may serve as a regulatory mechanism
for miRNA functionality through their sequestration and downstream up-regulation
of their target genes. Indeed previous studies speculated that stable intron-derived
RNA sequences (sisRNA) (72) act as molecular sinks to sequester miRNA (72) and/or
RBPs
(73) leading to reduction in their activities and future studies will test
whether sisRNA are derived from the cytoplasmic intronic sequences of the C5
group.
We previously demonstrated aberrant cytoplasmic IR in ALS-related VCPmu
samples during MN differentiation (23). Here we show that VCP mutations lead to
an aberrant PIR increase specifically of the C5 group of introns. Furthermore we
show that the aberrant increase in C5 PIR level at DIV=14 in ALS mutant cells
correlates with a decrease in the predicted activities of miR-1976, miR-4519 and
miR-4716-5p, motifs of which are enriched in the C5 intronic sequences. These
findings were further generalized to SOD1-related ALS hiPSC-derived mutant MNs,
21
supporting the hypothesis of a functional depletion of specific miRNAs as a result of
cytoplasmic intronic sequences-mediated sequestration in ALS cells. Notably a
reduction in miR-1976 activity, motifs of which are detected in 76% of the C5
intronic regions, is expected to occur in some sporadic ALS patients due to a
mutation (rs17162257) in its enhancer (74). Furthermore, several miRNAs, and their
target genes, are recognized to be involved in the occurrence and pathophysiology
of neurodegenerative diseases including ALS (75–77). Thus, here we propose that a
group of intronic sequences which accumulate in the cytoplasm of VCP mutant
cells, as previously shown (23), act as molecular sponges for miRNA, thus resulting
in elevated expression of their target genes. Future work will be required to
demonstrate the direct role of these cytoplasmic intronic sequences in regulating
miRNA activity through sequestration.
In conclusion we propose that cytoplasmic retained introns function as RNA
regulators in the homeostatic control of RBP localisation and miRNA activity during
MN
development
and
disease,
which
has
potential
implications
for
ALS
pathogenesis and the development of therapies for this devastating and incurable
disease.
22
FIGURES AND TABLES
Figure 1 | Nuclear and cytoplasmic IR affect two distinct mRNA subsets.
A. Schematic depicting the iPSC differentiation strategy for motor neurogenesis.
Arrows indicate sampling time-points in days when cells were fractionated into
nuclear and cytoplasmic compartments prior to (polyA) RNA-sequencing. Four iPSC
23
lines were obtained from four different healthy controls. Induced-pluripotent stem
cells (iPSC); neural precursors (NPC); “patterned” precursor motor neurons (ventral
spinal cord; pMN); post-mitotic but electrophysiologically inactive motor neurons
(MN);
electrophysiologically
active
MNs
(mMN).
B.
Pie
charts
representing
proportions
of
included
splicing
events
at
defined
stages
during
motor
neurogenesis in nuclear (left) and cytoplasmic (right) fractions. Total number of
events are indicated above the charts. Intron retention (IR); alternative exon (AltEx);
microexons (MIC); alternative 5′ and 3′ UTR (Alt5 and Alt3). C. Comparison of the
percent intron retention (PIR) during MN differentiation in nucleus (left) and
cytoplasm (right) for 21,161 events that exhibit >10% PIR in at least 3 out of 47
nuclear samples. D. Heatmap of the GO biological functions enriched among the
genes targeted by AltEx or IR in either the nucleus or the cytoplasm. P-values
obtained by Fisher enrichment test. E. Analysis of the relationship between the PIR
in the nucleus and the intron length. Retained introns are grouped in five
categories of increasing level of retention in the nucleus as indicated on the x-axis.
P-values obtained from analysis of variance comparing the full model of the logit of
maximum IR across all nuclear samples according to the five characteristics with
the reduced model removing the characteristic of interest. F. Comparison of intron
length between nuclear and cytoplasmic retained introns. Nuclear retained introns
are defined as intron exhibiting >20% IR in nuclear fraction and <5% IR in
cytoplasmic fraction. Cytoplasmic retained introns are defined as intron exhibiting
>20% IR in nuclear fraction and >15% IR in cytoplasmic fraction. P-values obtained
from Mann-Withney test.
G. Analysis of the relationship between the PIR in the
nucleus and the GC content in %. Data shown as in (E). H. Comparison of GC
content (%) between nuclear and cytoplasmic retained introns. I. Analysis of the
relationship between the PIR in the nucleus and the median enrichment for RBP
binding site compared to the non-retained introns of the same gene. Data shown as
in (E). J. Comparison of median enrichment for RBP binding sites between nuclear
24
and cytoplasmic retained introns. For C, E-J-J: Data shown as box plots in which
the centre line is the median, limits are the interquartile range and whiskers are the
minimum and maximum.
25
26
Figure 2 | A spatiotemporal taxonomy reveals cytoplasmic IRTs with distinct RBP
binding profiles. A. Comparison of the nuclear and cytoplasmic percent intron
retention (PIR) distributions for 9 groups of retained introns exhibiting distinct
spatio-temporal dynamics during MN differentiation as identified using SVD (see
Materials and Methods). N1, N2 and N3 contain introns primarily retained in the
nuclear compartment while the remaining 6 groups contain introns with significant
detection in the cytoplasm. Gold boxes = nucleus; blue boxes = cytoplasm. Grey
area indicates the range of PIR values for which an intron is considered
non-retained.
B. (Upper) Schematic depicting the selected splicing regulatory
regions juxtaposing the splice sites, namely the last 30 nucleotides (nts) of the
upstream exon (R1), the first 30nts of 5’ intron region (R2), 30nts in the middle of
the intron (R3), the last 30nts of 3’ intron region (R4), and the first 30nts of
downstream exon (R5). (Lower) Distributions of the percentage of regions in each
group of introns that are mapped by at least one crosslink event for each of the
available 131 RBPs. C. Distribution of the enrichments in crosslink events in each of
the selected regions R1, R2, R3, R4 and R5 for the available 131 RBPs across the 9
categories of introns. Enrichment is obtained by dividing the fraction of regions
from the group of interest with at least one crosslink event with the fraction of
regions from the complete set of introns (n=61872) with a crosslink event. D.
Heatmaps of the enrichment scores of the crosslinking events for 131 RBPs in the
R1, R2, R3, R4 and R5 regions for the 9 groups of introns hierarchically clustered
using Manhattan distance and Ward clustering. Data shown as box plots in which
the centre line is the median, limits are the interquartile range and whiskers are the
minimum and maximum.
27
Figure 3 | Identification of a cytoplasmic group of retained introns with a high
capacity for RBP and miRNA sequestration.
A. Analysis of the percentage of
nucleotides with cross-linking events for 131 RBPs across the entire retained intron
for all 9 categories. B. Analysis of the enrichment in binding sites for 131 RBPs
across the entire retained intron for all 9 categories. Enrichment is obtained by
28
dividing the fraction of retained introns in the category of interest with a CLIP
binding with the fraction of retained introns in the complete set of introns
(n=61872) with a crosslinking event.
C. Heatmap of the enrichment score of
crosslinking events in both the entire intron and each of the five regulatory regions
of the 270 retained introns of the C5 category for 51 RBPs that exhibit a difference
in the fraction of cross-linking events of more than 19% in the pool of C5 retained
introns compared to the full set of introns. The blue box highlights RBPs involved in
RNA transport and the gold box represents those involved in miRNA regulation. D,
E. Percentage of retained introns with crosslink events for two RBPs involved in
RNA transport (UPF1 and IGF2BP1), and two RBPs involved in miRNA regulatory
pathway (DROSHA and PUM2). F. Five top motifs enriched in the 270 retained
introns of the C5 category identified by HOMER (65). G. Distributions of the
changes in nuclear expression over time of the control samples for the TargetScan
(66) and miRDB (78) predicted target genes of miR-4519, miR-1976 and miR-4716-5p.
Fold-changes over time obtained by comparing the log2 expression level at time of
interest (
) with the expression level at previous stage (
).
0, 3, , 4, 2, 5}
dt = {
7 1
2
3
dt−1
Gold
shaded
area
indicates
the
time-point
where
the
largest
changes
in
cytoplasmic IR are observed over time for the control samples. For A,B & G: Data
shown as box plots in which the centre line is the median, limits are the
interquartile range and whiskers are the minimum and maximum.
29
Figure 4 | ALS-related transient accumulation in cytoplasmic retained introns
correlates with reduced miRNA activity. A. Comparison of the distributions of
nuclear and cytoplasmic percent intron retention (PIR) between control (colored
boxes) and VCPmu (white boxes) samples during MN differentiation for the C1 and C5
groups of cytoplasmic retained introns. P-values obtained with two-sided Welch
t-test.
B. Distributions of the changes in nuclear expression between VCPmu and
control samples at each time-point (right) for the TargetScan (66) and miRDB (78)
predicted target genes of miR-4519, miR-1976 and miR-4716-5p. Fold-changes
obtained
by
comparing
the
log2
expression
level
at
each
time
point
(
). Gold shaded area indicates the time-point where the largest
0, 3, , 4, 2, 5}
dt = {
7 1
2
3
changes in cytoplasmic IRTs are observed between the control and VCP mutant
samples. C. Boxplots displaying the distribution of percentage retention for the 270
30
introns of the C5 category in control MNs (white box), and SOD1mu MNs samples
(left, blue bar) (79, 80). Mutant samples exhibit a systematically higher proportion of
IR compared with controls. Linear mixed effects analysis of the relationship
between the PIR for the 270 introns and SOD1 mutation to account for idiosyncratic
variation due to patient differences: SOD1 mutation significantly increases IR by
about 5.6%
3 (standard errors;
(1) = 7.4, P = 6.5E-03). D. Distributions of the
±
χ 2
changes in expression between SOD1mu and control samples for the predicted
target genes of miR-4519, miR-1976, miR-4716-5p, miR-485-5p and miR-4267. For
A-D: Data shown as box plots in which the centre line is the median, limits are the
interquartile range and whiskers are the minimum and maximum.
31
Supplementary Figure 1 | A. Pie charts representing proportions of skipped splicing
events
at
defined
stages
during
motor
neurogenesis
in
nuclear
(left)
and
cytoplasmic (right) fractions. Total number of events are indicated above the
charts. Intron retention (IR); alternative exon (AltEx); microexons (MIC); alternative
5′ and 3′ UTR (Alt5 and Alt3). B, D. Analysis of the relationship between the percent
intron retention (PIR) in the nucleus and the number of introns per gene (A), and
32
the retained intron average conservation scores (D). Retained introns are grouped
in five categories of increasing level of retention in the nucleus as indicated on the
x-axis. Data shown as in Fig. 1D. P-values obtained from analysis of variance
comparing the full model of the logit of maximum IR across all nuclear samples
according to the five characteristics with the reduced model removing the
characteristic of interest. C, E. Comparison of number of introns per gene and the
conservation scores between nuclear and cytoplasmic retained introns. Nuclear
retained introns are defined as those exhibiting >20% IR in nuclear fraction and
<5% IR in cytoplasmic fraction. Cytoplasmic retained introns are defined as those
exhibiting >20% IR in nuclear fraction and >15% IR in cytoplasmic fraction. Data
shown as in Fig. 1D. P-values obtained from Mann-Withney test. F. Analysis of the
relationship between the percent intron retention (PIR) in the cytoplasm and the
intron length, the GC content in %, the number of introns per gene, the retained
intron average conservation scores and the median enrichment for RBP binding site
compared to the non-retained introns of the same gene. Retained introns are
grouped in five categories of increasing level of retention in the cytoplasm as
indicated on the x-axis. Data shown as box plots in which the centre line is the
median, limits are the interquartile range and whiskers are the minimum and
maximum.
33
Supplementary Figure 2 | A. Percentage of retained introns per gene for the genes
targeted by intron retention in each group. B. Distribution of the average
evolutionary sequence conservation scores in the in the last 30nts of the upstream
exon (R1), the first 30nts of 5’ intron region (R2), the 30nts in the middle of the
intron (R3), the last 30nts of 3’ intron region (R4), and the first 30nts of downstream
exon (R5) for 9 categories of introns for the 9 categories of introns. C. Distribution
34
of the maximum entropy scores for 9-bp 5′ splice sites and 23-bp 3′ splice sites for
the 9 categories of intron as obtained from MaxEntScan (38). D. Percentage of
introns
with
UPF1
regional
cross-linking
events
(left)
and
UPF1
regional
cross-linking enrichment (right) for each splicing regulatory regions R1, R2, R3, R4
and R5 in each group of introns. Dashed lines indicate the average percentage of all
61872 analysed introns with a CLIP binding (left) and the one-fold enrichment (right)
in the intronic regulatory regions (R2, R3, R4).
Supplementary Figure 3 | A. 14 motifs enriched in the 270 retained introns of the
C5 category identified by HOMER (65). B. Changes in gene expression over time in
35
the nucleus (gold boxes) and cytoplasm (blue boxes) for groups of genes containing
the 9 different categories of retained introns. Fold-changes obtained by comparing
the
log2
expression level at time of interest (
) with the
3, , 4, 2, 5}
di = {
7 1
2
3
expression level at iPSC stage (
). Data shown as box plots in which the centre line
d0
is the median, limits are the interquartile range and whiskers are the minimum and
maximum.
36
37
Supplementary Figure 4 | A, B. Singular value decomposition analysis of the PIR
cytoplasmic values of 94,457 introns in n = 48 cytoplasmic samples. Line plots
showing the PIR profiles of the first two singular vectors
and
, capturing 22%
v1
v2
and 9% of the variance in PIR respectively. Filled and empty data points indicate PIR
values for the control and VCPmu samples. C. Comparison of the distributions of
nuclear and cytoplasmic PIR between control (colored boxes) and VCPmu (white
boxes) samples during MN differentiation for the 6 groups of cytoplasmic retained
introns. P-values obtained with two-sided Welch t-test. Data shown as box plots in
which the centre line is the median, limits are the interquartile range and whiskers
are the minimum and maximum. D. MiRNA expression in “patterned” precursor
motor neuron cells (DIV=14) – Relative expression of miR-1976, miR-4519 and
miR-4716 in control (white) and mutant (gray) cells lines. Datapoints depicted as
black circles in the controls and black triangles in mutant cells.
38
39
Supplementary Figure 5 | Distributions of the changes in nuclear and cytoplasmic
expression over time of the control samples (left) and between VCPmu and control
samples at each time-point (right) for the 395 TargetScan (66) and miRDB (78)
predicted target genes of miR-4519, miR-1976, and miR-4716-5p. Fold-changes over
time obtained by comparing the log2 expression level at the time of interest (
) with the expression level at previous stage (
). Gold
0, 3, , 4, 2, 5}
dt = {
7 1
2
3
dt−1
shaded area indicates the time-point where the largest changes in cytoplasmic IR
are observed either over time for the control samples or between control and
mutant samples.
40
Supplementary Figure 6 | Same details and format as Supplementary Figure 5 but
for miR-485-5p and miR-4267.
41
SUPPLEMENTARY TABLES 1-18 can be accessed here.
Table S1 | Description of the iPSC lines and RNA sequencing samples used in this
study.
Table S2 | List of the 4490 nuclearly detained retained introns reported on Figs.
1F,H,J.
Table S3 | List of the 3633 cytoplasmic retained introns reported on Figs. 1F,H,J.
Tables S4-S12 | Lists of the 9 groups of retained introns (N1, N2 N3, C1, C2, C3, C4,
C5,
C6)
associated
with
the
distinct
spatiotemporal
dynamics
during
MN
differentiation reported on Fig. 2A.
Tables S12-S17 | Frequency and enrichment in 133 RBPs crosslink events in 5
defined regions of interest (R1, R2, R3, R4, and R5).
AUTHOR CONTRIBUTIONS
Conceptualization, R.L., R.P.; Formal Analysis, R.L.; Investigation, R.L., M.P-H., H.C.,
J.N., G.E.T.; Writing – Original Draft, M.P-H., R.L., R.P.; Writing – Review & Editing,
M.P-H., H.C., J.N., G.E.T., R.L., R.P; Resources, R.L., R.P.; Visualization, R.L.; Funding
Acquisition, R.L., R.P.; Supervision, R.L, R.P.
ACKNOWLEDGMENTS
The authors wish to thank the patients for fibroblast donations. We also thank Anob
M Chakrabarti for sharing BED files of aligned CLIP data. We are grateful for the
help and support provided by the Scientific Computing section and the DNA
Sequencing section of Research Support Division at OIST.
This work was
supported by the Idiap Research Institute and by the Francis Crick Institute which
42
receives its core funding from Cancer Research UK (FC010110), the UK Medical
Research Council (FC010110), and the Wellcome Trust (FC010110). D.M.T. is
supported by a Newton-Mosharafa scholarship. R.P. holds an MRC Senior Clinical
Fellowship [MR/S006591/1].
REFERENCES
1. Braunschweig,U., Barbosa-Morais,N.L., Pan,Q., Nachman,E.N., Alipanahi,B.,
Gonatopoulos-Pournatzis,T., Frey,B., Irimia,M. and Blencowe,B.J. (2014)
Widespread intron retention in mammals functionally tunes transcriptomes.
Genome Res., 24, 1774–1786.
2. Boutz,P.L., Bhutkar,A. and Sharp,P.A. (2015) Detained introns are a novel,
widespread class of post-transcriptionally spliced introns. Genes Dev., 29,
63–80.
3. Buckley,P.T., Lee,M.T., Sul,J.-Y., Miyashiro,K.Y., Bell,T.J., Fisher,S.A., Kim,J. and
Eberwine,J. (2011) Cytoplasmic intron sequence-retaining transcripts can be
dendritically targeted via ID element retrotransposons. Neuron, 69, 877–884.
4. Yap,K., Lim,Z.Q., Khandelia,P., Friedman,B. and Makeyev,E.V. (2012) Coordinated
regulation of neuronal mRNA steady-state levels through developmentally
controlled intron retention. Genes Dev., 26, 1209–1223.
5. Mauger,O., Lemoine,F. and Scheiffele,P. (2016) Targeted Intron Retention and
Excision for Rapid Gene Regulation in Response to Neuronal Activity. Neuron,
92, 1266–1278.
6. Xu,Q., Walker,D., Bernardo,A., Brodbeck,J., Balestra,M.E. and Huang,Y. (2008)
Intron-3 retention/splicing controls neuronal expression of apolipoprotein E
in the CNS. J. Neurosci., 28, 1452–1459.
7. Braun,C.J., Stanciu,M., Boutz,P.L., Patterson,J.C., Calligaris,D., Higuchi,F.,
Neupane,R., Fenoglio,S., Cahill,D.P., Wakimoto,H., et al. (2017) Coordinated
Splicing of Regulatory Detained Introns within Oncogenic Transcripts Creates
an Exploitable Vulnerability in Malignant Glioma. Cancer Cell, 32, 411–426.e11.
8. Ninomiya,K., Kataoka,N. and Hagiwara,M. (2011) Stress-responsive maturation of
Clk1/4 pre-mRNAs promotes phosphorylation of SR splicing factor. J. Cell Biol.,
195, 27–40.
43
9. Boothby,T.C., Zipper,R.S., van der Weele,C.M. and Wolniak,S.M. (2013) Removal of
retained introns regulates translation in the rapidly developing gametophyte of
Marsilea vestita. Dev. Cell, 24, 517–529.
10. Kalyna,M., Simpson,C.G., Syed,N.H., Lewandowska,D., Marquez,Y., Kusenda,B.,
Marshall,J., Fuller,J., Cardle,L., McNicol,J., et al. (2012) Alternative splicing and
nonsense-mediated decay modulate expression of important regulatory genes
in Arabidopsis. Nucleic Acids Res., 40, 2454–2469.
11. Khaladkar,M., Buckley,P.T., Lee,M.T., Francis,C., Eghbal,M.M., Chuong,T.,
Suresh,S., Kuhn,B., Eberwine,J. and Kim,J. (2013) Subcellular RNA sequencing
reveals broad presence of cytoplasmic intron-sequence retaining transcripts in
mouse and rat neurons. PLoS One, 8, e76194.
12. Jaillon,O., Bouhouche,K., Gout,J.-F., Aury,J.-M., Noel,B., Saudemont,B.,
Nowacki,M., Serrano,V., Porcel,B.M., Ségurens,B., et al. (2008) Translational
control of intron splicing in eukaryotes. Nature, 451, 359–362.
13. Wong,J.J.-L., Ritchie,W., Ebner,O.A., Selbach,M., Wong,J.W.H., Huang,Y., Gao,D.,
Pinello,N., Gonzalez,M., Baidya,K., et al. (2013) Orchestrated intron retention
regulates normal granulocyte differentiation. Cell, 154, 583–595.
14. Vanichkina,D.P., Schmitz,U., Wong,J.J.-L. and Rasko,J.E.J. (2018) Challenges in
defining the role of intron retention in normal biology and disease. Semin. Cell
Dev. Biol., 75, 40–49.
15. Sharangdhar,T., Sugimoto,Y., Heraud‐Farlow,J., Fernández‐Moya,S.M., Ehses,J.,
de los Mozos,I.R., Ule,J. and Kiebler,M.A. (2017) A retained intron in the 3′‐ UTR
of Calm3 mRNA mediates its Staufen2‐ and activity‐dependent localization to
neuronal dendrites. EMBO reports, 18, 1762–1774.
16. Schmitz,U., Pinello,N., Jia,F., Alasmari,S., Ritchie,W., Keightley,M.-C., Shini,S.,
Lieschke,G.J., Wong,J.J.-L. and Rasko,J.E.J. (2017) Intron retention enhances
gene regulatory complexity in vertebrates. Genome Biol., 18, 216.
17. Monteuuis,G., Wong,J.J.L., Bailey,C.G., Schmitz,U. and Rasko,J.E.J. (2019) The
changing paradigm of intron retention: regulation, ramifications and recipes.
Nucleic Acids Res., 47, 11497–11513.
18. Jangi,M., Fleet,C., Cullen,P., Gupta,S.V., Mekhoubad,S., Chiao,E., Allaire,N.,
Bennett,C.F., Rigo,F., Krainer,A.R., et al. (2017) SMN deficiency in severe models
of spinal muscular atrophy causes widespread intron retention and DNA
damage. Proc. Natl. Acad. Sci. U. S. A., 114, E2347–E2356.
44
19. Adusumalli,S., Ngian,Z.-K., Lin,W.-Q., Benoukraf,T. and Ong,C.-T. (2019)
Increased intron retention is a post-transcriptional signature associated with
progressive aging and Alzheimer’s disease. Aging Cell, 18, e12928.
20. Luisier,R., Tyzack,G.E., Hall,C.E., Mitchell,J.S., Devine,H., Taha,D.M., Malik,B.,
Meyer,I., Greensmith,L., Newcombe,J., et al. (2018) Intron retention and nuclear
loss of SFPQ are molecular hallmarks of ALS. Nat. Commun., 9, 2010.
21. Wang,Q., Conlon,E.G., Manley,J.L. and Rio,D.C. (2020) Widespread intron
retention impairs protein homeostasis in C9orf72 ALS brains. Genome Res., 30,
1705–1715.
22. Hogan,A.L., Grima,N., Fifita,J.A., McCann,E.P., Heng,B., Fat,S.C.M., Maharjan,R.,
Cain,A.K., Henden,L., Tarr,I., et al. (2020) SFPQ intron retention, reduced
expression and aggregate formation in central nervous system tissue are
pathological features of amyotrophic lateral sclerosis. Cold Spring Harbor
Laboratory, 10.1101/2020.09.22.309062.
23. Tyzack,G.E., Neeves,J., Klein,P., Crerar,H., Ziff,O., Taha,D.M., Luisier,R.,
Luscombe,N.M. and Patani,R. (2020) An aberrant cytoplasmic intron retention
programme is a blueprint for ALS-related RBP mislocalization.
10.1101/2020.07.20.211557.
24. Neumann,M., Sampathu,D.M., Kwong,L.K., Truax,A.C., Micsenyi,M.C., Chou,T.T.,
Bruce,J., Schuck,T., Grossman,M., Clark,C.M., et al. (2006) Ubiquitinated
TDP-43 in frontotemporal lobar degeneration and amyotrophic lateral
sclerosis. Science, 314, 130–133.
25. Tyzack,G.E., Luisier,R., Taha,D.M., Neeves,J., Modic,M., Mitchell,J.S., Meyer,I.,
Greensmith,L., Newcombe,J., Ule,J., et al. (2019) Widespread FUS
mislocalization is a molecular hallmark of amyotrophic lateral sclerosis. Brain,
142, 2572–2580.
26. Kiskinis,E., Sandoe,J., Williams,L.A., Boulting,G.L., Moccia,R., Wainger,B.J.,
Han,S., Peng,T., Thams,S., Mikkilineni,S., et al. (2014) Pathways disrupted in
human ALS motor neurons identified through genetic correction of mutant
SOD1. Cell Stem Cell, 14, 781–795.
27. Bray,N.L., Pimentel,H., Melsted,P. and Pachter,L. (2016) Near-optimal
probabilistic RNA-seq quantification. Nat. Biotechnol., 34, 525–527.
28. Fraley,C. and Raftery,A.E. mclust Version 4 for R: Normal Mixture Modeling for
Model-Based Clustering, Classification, and Density Estimation.
45
29. Boldstad,B.M., Irizarry,R.A., Astrand,M. and Speed,T.P. (2003) A Comparison of
Normalization Methods for High Density Oligonucleotide Array Data Based on
Bias and Variance. Bioinformatics, 19, 185–193.
30. Pimentel,H.J., Bray,N., Puente,S., Melsted,P. and Pachter,L. (2016) Differential
analysis of RNA-Seq incorporating quantification uncertainty. bioRxiv.
31. Irimia,M., Weatheritt,R.J., Ellis,J.D., Parikshak,N.N., Gonatopoulos-Pournatzis,T.,
Babor,M., Quesnel-Vallières,M., Tapial,J., Raj,B., O’Hanlon,D., et al. (2014) A
highly conserved program of neuronal microexons is misregulated in autistic
brains. Cell, 159, 1511–1523.
32. Alexa,A. and Rahnenfuhrer,J. (2016) topGO: Enrichment Analysis for Gene
Ontology.
33. Attig,J., Agostini,F., Gooding,C., Chakrabarti,A.M., Singh,A., Haberman,N.,
Zagalak,J.A., Emmett,W., Smith,C.W.J., Luscombe,N.M., et al. (2018) Heteromeric
RNP Assembly at LINEs Controls Lineage-Specific RNA Processing. Cell, 174,
1067–1081.e17.
34. Sloan,C.A., Chan,E.T., Davidson,J.M., Malladi,V.S., Strattan,J.S., Hitz,B.C.,
Gabdank,I., Narayanan,A.K., Ho,M., Lee,B.T., et al. (2016) ENCODE data at the
ENCODE portal. Nucleic Acids Res., 44, D726–32.
35. Van Nostrand,E.L., Freese,P., Pratt,G.A., Wang,X. and Wei,X. (2017) A large-scale
binding and functional map of human RNA binding proteins. bioRxiv.
36. Dobin,A., Davis,C.A., Schlesinger,F., Drenkow,J., Zaleski,C., Jha,S., Batut,P.,
Chaisson,M. and Gingeras,T.R. (2013) STAR: ultrafast universal RNA-seq aligner.
Bioinformatics, 29, 15–21.
37. Hinrichs,A.S., Karolchik,D., Baertsch,R., Barber,G.P., Bejerano,G., Clawson,H.,
Diekhans,M., Furey,T.S., Harte,R.A., Hsu,F., et al. (2006) The UCSC Genome
Browser Database: update 2006. Nucleic Acids Res., 34, D590–8.
38. Yeo,G. and Burge,C.B. (2004) Maximum entropy modeling of short sequence
motifs with applications to RNA splicing signals. J. Comput. Biol., 11, 377–394.
39. Ule,J., Stefani,G., Mele,A., Ruggiu,M., Wang,X., Taneri,B., Gaasterland,T.,
Blencowe,B.J. and Darnell,R.B. (2006) An RNA map predicting Nova-dependent
splicing regulation. Nature, 444, 580–586.
40. Cereda,M., Pozzoli,U., Rot,G., Juvan,P., Schweitzer,A., Clark,T. and Ule,J. (2014)
RNAmotifs: prediction of multivalent RNA motifs that control alternative
46
splicing. Genome Biol., 15, R20.
41. Pollard,K.S., Hubisz,M.J., Rosenbloom,K.R. and Siepel,A. (2010) Detection of
nonneutral substitution rates on mammalian phylogenies. Genome Res., 20,
110–121.
42. Siepel,A., Bejerano,G., Pedersen,J.S., Hinrichs,A.S., Hou,M., Rosenbloom,K.,
Clawson,H., Spieth,J., Hillier,L.W., Richards,S., et al. (2005) Evolutionarily
conserved elements in vertebrate, insect, worm, and yeast genomes. Genome
Res., 15, 1034–1050.
43. Hall,C.E., Yao,Z., Choi,M., Tyzack,G.E., Serio,A., Luisier,R., Harley,J., Preza,E.,
Arber,C., Crisp,S.J., et al. (2017) Progressive Motor Neuron Pathology and the
Role of Astrocytes in a Human Stem Cell Model of VCP-Related ALS. Cell Rep.,
19, 1739–1749.
44. Edwards,C.R., Ritchie,W., -L. Wong,J.J., Schmitz,U., Middleton,R., An,X.,
Mohandas,N., Rasko,J.E.J. and Blobel,G.A. (2016) A dynamic intron retention
program in the mammalian megakaryocyte and erythrocyte lineages. Blood, 127,
e24–e34.
45. Pimentel,H., Parra,M., Gee,S.L., Mohandas,N., Pachter,L. and Conboy,J.G. (2016)
A dynamic intron retention program enriched in RNA processing genes
regulates gene expression during terminal erythropoiesis. Nucleic Acids Res.,
44, 838–851.
46. Middleton,R., Gao,D., Thomas,A., Singh,B., Au,A., Wong,J.J.-L., Bomane,A.,
Cosson,B., Eyras,E., Rasko,J.E.J., et al. (2017) IRFinder: assessing the impact of
intron retention on mammalian gene expression. Genome Biol., 18, 51.
47. Ullrich,S. and Guigó,R. (2020) Dynamic changes in intron retention are tightly
associated with regulation of splicing factors and proliferative activity during
B-cell development. Nucleic Acids Res., 48, 1327–1340.
48. Dvinge,H. and Bradley,R.K. (2015) Widespread intron retention diversifies most
cancer transcriptomes. Genome Med., 7, 45.
49. Galante,P.A.F., Sakabe,N.J., Kirschbaum-Slager,N. and de Souza,S.J. (2004)
Detection and evaluation of intron retention events in the human
transcriptome. RNA, 10, 757–765.
50. Sakabe,N.J. and de Souza,S.J. (2007) Sequence features responsible for intron
retention in human. BMC Genomics, 8, 59.
47
51. Dirksen,W.P., Sun,Q. and Rottman,F.M. (1995) Multiple splicing signals control
alternative intron retention of bovine growth hormone pre-mRNA. J. Biol.
Chem., 270, 5346–5352.
52. Lejeune,F., Cavaloc,Y. and Stevenin,J. (2001) Alternative splicing of intron 3 of
the serine/arginine-rich protein 9G8 gene. Identification of flanking exonic
splicing enhancers and involvement of 9G8 as a trans-acting factor. J. Biol.
Chem., 276, 7850–7858.
53. Zhu,L., Zhang,Y., Zhang,W., Yang,S., Chen,J.-Q. and Tian,D. (2009) Patterns of
exon-intron architecture variation of genes in eukaryotic genomes. BMC
Genomics, 10, 47.
54. Robberson,B.L., Cote,G.J. and Berget,S.M. (1990) Exon definition may facilitate
splice site selection in RNAs with multiple exons. Mol. Cell. Biol., 10, 84–94.
55. Ule,J. and Blencowe,B.J. (2019) Alternative Splicing Regulatory Networks:
Functions, Mechanisms, and Evolution. Mol. Cell, 76, 329–345.
56. Graber,T.E., Freemantle,E., Anadolu,M.N., Hébert-Seropian,S., MacAdam,R.L.,
Shin,U., Hoang,H.-D., Alain,T., Lacaille,J.-C. and Sossin,W.S. (2017) UPF1 Governs
Synaptic Plasticity through Association with a STAU2 RNA Granule. J. Neurosci.,
37, 9116–9131.
57. Dalla Costa,I., Buchanan,C.N., Zdradzinski,M.D., Sahoo,P.K., Smith,T.P.,
Thames,E., Kar,A.N. and Twiss,J.L. (2020) The functional organization of axonal
mRNA transport and translation. Nat. Rev. Neurosci.,
10.1038/s41583-020-00407-7.
58. Das,S., Singer,R.H. and Yoon,Y.J. (2019) The travels of mRNAs in neurons: do they
know where they are going? Curr. Opin. Neurobiol., 57, 110–116.
59. Han,J., Lee,Y., Yeom,K.-H., Kim,Y.-K., Jin,H. and Kim,V.N. (2004) The
Drosha-DGCR8 complex in primary microRNA processing. Genes Dev., 18,
3016–3027.
60. Hu,R., Zhu,X., Chen,C., Xu,R., Li,Y. and Xu,W. (2018) RNA-binding protein PUM 2
suppresses osteosarcoma progression via partly and competitively binding to
STARD 13 3′ UTR with mi RNA s. Cell Prolif., 51, e12508.
61. Treiber,T., Treiber,N., Plessmann,U., Harlander,S., Daiß,J.-L., Eichner,N.,
Lehmann,G., Schall,K., Urlaub,H. and Meister,G. (2017) A Compendium of
RNA-Binding Proteins that Regulate MicroRNA Biogenesis. Mol. Cell, 66,
270–284.e13.
48
62. Elbarbary,R.A., Miyoshi,K., Hedaya,O., Myers,J.R. and Maquat,L.E. (2017) UPF1
helicase promotes TSN-mediated miRNA decay. Genes Dev., 31, 1483–1493.
63. Park,J., Seo,J.-W., Ahn,N., Park,S., Hwang,J. and Nam,J.-W. (2019)
UPF1/SMG7-dependent microRNA-mediated gene regulation. Nat. Commun.,
10, 4181.
64. Jin,H., Suh,M.R., Han,J., Yeom,K.-H., Lee,Y., Heo,I., Ha,M., Hyun,S. and Kim,V.N.
(2009) Human UPF1 participates in small RNA-induced mRNA downregulation.
Mol. Cell. Biol., 29, 5789–5799.
65. Heinz,S., Benner,C., Spann,N., Bertolino,E., Lin,Y.C., Laslo,P., Cheng,J.X.,
Murre,C., Singh,H. and Glass,C.K. (2010) Simple combinations of
lineage-determining transcription factors prime cis-regulatory elements
required for macrophage and B cell identities. Mol. Cell, 38, 576–589.
66. Agarwal,V., Bell,G.W., Nam,J.-W. and Bartel,D.P. (2015) Predicting effective
microRNA target sites in mammalian mRNAs. Elife, 4.
67. John,B., Enright,A.J., Aravin,A., Tuschl,T., Sander,C. and Marks,D.S. (2004) Human
MicroRNA targets. PLoS Biol., 2, e363.
68. Duss,O., Michel,E., Yulikov,M., Schubert,M., Jeschke,G. and Allain,F.H.-T. (2014)
Structural basis of the non-coding RNA RsmZ acting as a protein sponge.
Nature, 509, 588–592.
69. Yoshimoto,R., Kaida,D., Furuno,M. and Burroughs,A.M. (2017) Global analysis of
pre-mRNA subcellular localization following splicing inhibition by spliceostatin
A. RNA.
70. Ninomiya,K., Adachi,S., Natsume,T., Iwakiri,J., Terai,G., Asai,K. and Hirose,T.
LncRNA-dependent nuclear stress bodies promote intron retention through SR
protein phosphorylation. 10.1101/681056.
71. Wong,J.J.-L., Gao,D., Nguyen,T.V., Kwok,C.-T., van Geldermalsen,M.,
Middleton,R., Pinello,N., Thoeng,A., Nagarajah,R., Holst,J., et al. (2017) Intron
retention is regulated by altered MeCP2-mediated splicing factor recruitment.
Nat. Commun., 8, 15134.
72. Moss,W.N. and Steitz,J.A. (2013) Genome-wide analyses of Epstein-Barr virus
reveal conserved RNA structures and a novel stable intronic sequence RNA.
BMC Genomics, 14, 543.
73. Yin,Q.-F., Yang,L., Zhang,Y., Xiang,J.-F., Wu,Y.-W., Carmichael,G.G. and
49
Chen,L.-L. (2012) Long noncoding RNAs with snoRNA ends. Mol. Cell, 48,
219–230.
74. Xie,T., Deng,L., Mei,P., Zhou,Y., Wang,B., Zhang,J., Lin,J., Wei,Y., Zhang,X. and
Xu,R. (2014) A genome-wide association study combining pathway analysis for
typical sporadic amyotrophic lateral sclerosis in Chinese Han populations.
Neurobiol. Aging, 35, 1778.e9–1778.e23.
75. Wang,L. and Zhang,L. (2020) MicroRNAs in amyotrophic lateral sclerosis: from
pathogenetic involvement to diagnostic biomarker and therapeutic agent
development. Neurol. Sci., 41, 3569–3577.
76. Eyileten,C., Sharif,L., Wicik,Z., Jakubik,D., Jarosz-Popek,J., Soplinska,A.,
Postula,M., Czlonkowska,A., Kaplon-Cieslicka,A. and Mirowska-Guzel,D. (2021)
The Relation of the Brain-Derived Neurotrophic Factor with MicroRNAs in
Neurodegenerative Diseases and Ischemic Stroke. Mol. Neurobiol., 58, 329–347.
77. Zuo,X., Zhou,J., Li,Y., Wu,K., Chen,Z., Luo,Z., Zhang,X., Liang,Y., Esteban,M.A.,
Zhou,Y., et al. (2021) TDP-43 aggregation induced by oxidative stress causes
global mitochondrial imbalance in ALS. Nat. Struct. Mol. Biol.,
10.1038/s41594-020-00537-7.
78. Chen,Y. and Wang,X. (2020) miRDB: an online database for prediction of
functional microRNA targets. Nucleic Acids Res., 48, D127–D131.
79. Kiskinis,E., Sandoe,J., Williams,L.A., Boulting,G.L., Moccia,R., Wainger,B.J., Han,S.,
Peng,T., Thams,S., Mikkilineni,S., et al. (2014) Pathways disrupted in human ALS
motor neurons identified through genetic correction of mutant SOD1. Cell Stem
Cell, 14, 781–795.
80. Kapeli,K., Pratt,G.A., Vu,A.Q., Hutt,K.R., Martinez,F.J., Sundararaman,B., Batra,R.,
Freese,P., Lambert,N.J., Huelga,S.C., et al. (2016) Distinct and shared functions
of ALS-associated proteins TDP-43, FUS and TAF15 revealed by multisystem
analyses. Nat. Commun., 7, 12143.
50
| 2021 | Diminished miRNA activity is associated with aberrant cytoplasmic intron retention in ALS pathogenesis | 10.1101/2021.01.27.428555 | [
"Petric-Howe Marija",
"Crerar Hamish",
"Neeves Jacob",
"Tyzack Giulia E.",
"Patani Rickie",
"Luisier Raphaëlle"
] | creative-commons |
1
A Hyperactive Kunjin Virus NS3 Helicase Mutant Demonstrates Increased
1
Dissemination and Mortality in Mosquitoes
2
3
Kelly E. Du Pont,a Nicole R. Sexton,b,d Martin McCullagh,c Gregory D. Ebel,b,d and Brian
4
J. Geissb,d,e#
5
6
aDepartment of Chemistry, Colorado State University, Fort Collins, Colorado, USA
7
bArthropod-borne and Infectious Diseases Laboratory, Department of Microbiology,
8
Immunology and Pathology, Colorado State University, Fort Collins, Colorado, USA
9
cDepartment of Chemistry, Oklahoma State University, Stillwater, Oklahoma, USA
10
dDepartment of Microbiology, Immunology and Pathology, Colorado State University,
11
Fort Collins, Colorado, USA
12
eSchool of Biomedical Engineering, Colorado State University, Fort Collins, Colorado,
13
USA
14
15
Running Head: Hyperactive Viral Helicase Alters Dissemination and Mortality in
16
Mosquitoes
17
18
#Address correspondence to Brian J. Geiss, Brian.Geiss@colostate.edu.
19
K.E.D. and N.R.S. contributed equally to this work.
20
21
Abstract word count – 245 words
22
Importance word count – 110 words
23
2
ABSTRACT
24
The unwinding of double-stranded RNA intermediates is critical for replication and
25
packaging of flavivirus RNA genomes. This unwinding activity is achieved by the ATP-
26
dependent nonstructural protein 3 (NS3) helicase. In previous studies, we investigated
27
the mechanism of energy transduction between the ATP and RNA binding pockets
28
using molecular dynamics simulations and enzymatic characterization. Our data
29
corroborated the hypothesis that Motif V is a communication hub for this energy
30
transduction. More specifically, mutations T407A and S411A in Motif V exhibit a
31
hyperactive helicase phenotype leading to the regulation of translocation and unwinding
32
during replication. However, the effect of these mutations on viral infection in cell culture
33
and in vivo is not well understood. Here, we investigated the role of Motif V in viral
34
replication using T407A and S411A West Nile virus (Kunjin subtype) mutants in cell
35
culture and in vivo. We were able to recover S411A Kunjin but unable to recover T407A
36
Kunjin. Our results indicated that S411A Kunjin decreased viral infection, and increased
37
cytopathogenicity in cell culture as compared to WT Kunjin. Similarly, decreased
38
infection rates in surviving S411A-infected Culex quinquefasciatus mosquitoes were
39
observed, but S411A Kunjin infection resulted in increased mortality compared to WT
40
Kunjin. Additionally, S411A Kunjin increased viral dissemination and saliva positivity
41
rates in surviving mosquitoes compared to WT Kunjin. These data suggest that S411A
42
Kunjin increases pathogenesis in mosquitoes. Overall, these data indicate that NS3
43
Motif V may play a role in the pathogenesis, dissemination, and transmission efficiency
44
of Kunjin virus.
45
46
3
IMPORTANCE
47
Kunjin and West Nile viruses belong to the arthropod-borne flaviviruses, which can
48
result in severe symptoms including encephalitis, meningitis, and death. Flaviviruses
49
have expanded into new populations and emerged as novel pathogens repeatedly in
50
recent years demonstrating they remain a global threat. Currently, there are no
51
approved anti-viral therapeutics against either Kunjin or West Nile viruses. Thus, there
52
is a pressing need for understanding the pathogenesis of these viruses in humans. In
53
this study, we investigate the role of the Kunjin virus helicase on infection in cell culture
54
and in vivo. This work provides new insight into how flaviviruses control pathogenesis
55
and mosquito transmission through the nonstructural protein 3 helicase.
56
57
INTRODUCTION
58
Kunjin virus, a West Nile virus (WNV) subtype, causes encephalitis epidemics in horses
59
that are localized to Australia (1–4). Whereas, WNV has a much larger global impact
60
present in almost every major continent except for South America and Antarctica (4, 5)
61
and regularly results in encephalitis in humans as well as horses (6). Within the United
62
States alone, approximately 3 million people are thought to have been infected with
63
West Nile virus between 1999 and 2010 (7–9). Kunjin and WNV share a natural
64
transmission cycle between Culex mosquito vectors and bird reservoir hosts (2).
65
Humans and horses are considered dead-end hosts because they do not contribute to
66
viral perpetuation. In humans, around 80% of WNV infected individuals are
67
asymptomatic and the majority of symptomatic individuals experience a mild febrile
68
illness. However, approximately 1:150 infections result in severe symptoms including
69
4
meningitis and/or encephalitis, and ~9% of these cases are fatal (6, 10). Currently, there
70
are vaccines against WNV for horses, but not for humans; no vaccines are available for
71
Kunjin virus (5). Thus, there is a need for the development of vaccines and/or antiviral
72
therapies for Kunjin and WNV infections. Developing a fundamental understanding of
73
how Kunjin and WNV replicate within hosts, including the mosquito vector, is essential
74
to the development of interventional strategies.
75
76
Kunjin and WNV belong to the flavivirus genus within the Flaviviridae family. Flaviviridae
77
is a group of single-stranded positive-sense RNA viruses with genomes of
78
approximately 11 kb in length (11–13). Kunjin virus is a subtype of WNV with a
79
nucleotide and amino acid sequence identity of 82% and 93%, respectively (14–16).
80
However, in humans, Kunjin virus results in low morbidity compared with WNV making it
81
an excellent tool to study WNV replication with well-established molecular tools while
82
minimizing risk (17). Additionally, Kunjin virus is less cytopathic than WNV, allowing for
83
differences in virus-induced cell viability to be more easily visualized. Proteins and
84
processes involved in viral replication are conserved across the flavivirus genus
85
including for Kunjin, WNV, dengue, yellow fever, Japanese encephalitis, and Zika
86
viruses (12, 18). Initially, the viral RNA genome is translated into a single polyprotein
87
which is cleaved by host and viral proteases into three structural proteins (C, prM, and
88
E) and eight nonstructural proteins (NS1, NS2A, NS2B, NS3, NS4A, 2K, NS4B, and
89
NS5) (12, 18, 19). The viral NS replication proteins then generate a negative-sense anti-
90
genomic RNA that is in complex with the positive-sense genomic RNA, forming the
91
double-stranded RNA (dsRNA) intermediate complex (20, 21). The negative-sense anti-
92
5
genomic RNA serves as a template for positive-strand synthesis (20); therefore,
93
unwinding of the dsRNA intermediate is required for replication. Unwinding is achieved
94
by the C-terminal helicase domain of NS3 (22–24).
95
96
NS3 helicase domain is a multi-functional viral protein that houses three enzymatic
97
activities: RNA helicase, nucleoside triphosphatase (NTPase), and RNA
98
5’triphosphatase (RTPase) (25–28). NS3 helicase is a member of the superfamily 2
99
(SF2) helicases (29). The helicase domain consists of three subdomains (1, 2, and 3).
100
Subdomains 1 and 2 are RecA-like structures that are highly conserved across all SF2
101
helicases, while subdomain 3 is unique to the viral/DEAH-like group of SF2 helicases
102
(30). Additionally, there are eight structural motifs (Motifs I, Ia, II, III, IV, IVa, V, and VI)
103
that are highly conserved across all viral/DEAH-like subfamilies with the SF2 helicases
104
(29). These structural motifs are responsible for both substrate binding and enzymatic
105
function within the helicase. The helicase domain is responsible for translocation and
106
unwinding of the double-stranded RNA intermediate in an ATP-dependent manner
107
during viral replication (31). Previous studies further identified Motif V as potentially
108
critical for translocation and unwinding of the double-stranded RNA intermediate (32,
109
33). Motif V was described as a potential link between the ATP binding pocket and the
110
RNA binding cleft through strong correlation between residues within Motif V and both
111
binding pockets (32). The strongly correlated movements between ATP binding pocket
112
and RNA binding cleft residues in our simulations suggest a physical linkage between
113
the two sites that may be important for ATP driven helicase function. Additionally,
114
mutants T407A and S411A in Motif V increased unwinding activity and decreased viral
115
6
genome replication as compared to wild-type (WT), suggesting that the hydrogen bond
116
between these two residues in WT inhibits helicase unwinding activity in vitro and in
117
vivo (33). These data suggested that Motif V may serve as a molecular throttle on NS3
118
helicase function, but what effect these residues play on the larger viral replication cycle
119
was not clear.
120
121
To better understand the effects NS3 Motif V mutations have on flavivirus replication,
122
we sought to investigate the role of Motif V T407 and S411 residues on helicase
123
function in cell culture and in vivo by introducing alanine mutations in full-length
124
infectious Kunjin virus: T407A Kunjin and S411A Kunjin. Only the S411A Kunjin was
125
recovered and it resulted in reduced viral yields compared with wild-type (WT) Kunjin.
126
Additionally, S411A Kunjin showed increased cytopathic effect in comparison to WT
127
Kunjin in cell culture. Similarly, when WT or S411A Kunjin viruses were intrathoracically
128
injected into Culex quinquefasciatus mosquitoes, S411A Kunjin resulted in increased
129
mortality compared with WT Kunjin. Upon further investigation of mosquito infection,
130
S411A Kunjin viruses were found to disseminate and transmit more effectively than WT
131
Kunjin viruses, even though the overall infection rate was lower than WT Kunjin.
132
Overall, our data suggest that flaviviruses may use NS3 Motif V to help control
133
cytotoxicity induced by NS3 during infection and limit virus-induced mortality in mosquito
134
vectors.
135
136
RESULTS
137
7
S411A Kunjin virus increases cytopathic effect in cell culture. Previously, Motif V
138
residues, T407 and S411, were mutated to alanine to disrupt a hydrogen bond that
139
potentially stabilizes the Motif V secondary structure of NS3 helicase during viral
140
replication (Fig. 1). These mutations were shown to decrease viral genome replication in
141
a replicon-based system, while increasing helicase unwinding activity biochemically
142
(33). In the present study, we introduced these mutations into the full-length infectious
143
Kunjin virus to investigate the effects of these mutations on infectivity compared to WT
144
Kunjin both in cell culture and in mosquito infections. We utilized a novel mutagenesis
145
and a bacteria-free viral launch system to generate the T407A Kunjin and S411A Kunjin
146
viruses in Vero cells. The first generation of S411A Kunjin was recovered from infection
147
and the presence of the alanine mutation was verified with sequencing (Fig. 2). On the
148
other hand, we were unable to recover the T407A Kunjin despite repeated attempts,
149
which was consistent with our previously reported decrease in T407A viral genome
150
replication in replicon assays (33). Second generation stocks of WT Kunjin and S411A
151
Kunjin were generated and the viruses were titered for further experiments. We noted
152
the plaque morphology for both WT Kunjin and S411A Kunjin (Fig. 3). WT Kunjin
153
showed large, faint plaque sizes (Fig. 3A), while S411A Kunjin showed small, but
154
distinctly clear plaques (Fig. 3B), suggesting a potential decrease in viral cell-to-cell
155
spread and an increase in cytopathic effect for S411A Kunjin infected cells compared to
156
WT Kunjin. Since these results suggest that S411A Kunjin may be more toxic to cells
157
during infection, we further investigated the effect of the S411A Kunjin on cell viability.
158
159
8
S411A Kunjin reduces NADH and intracellular ATP levels leading to increased
160
cellular death. We utilized resazurin and CellTiter-Glo assays to quantify virus-induced
161
cell killing in HEK293T and Vero cells infected with either WT Kunjin or S411A Kunjin at
162
a multiplicity of infection (MOI) of five PFU/cell. Both of these assays estimate cell
163
viability through the measurement of metabolically active cells using fluorescence and
164
luminescence, respectively. In the resazurin assay, resazurin, a nonfluorescent dye,
165
converts to resorufin, a highly fluorescent dye, in response to the reducing environment
166
of heathy, growing cells (34–36). We measured the relative fluorescence units (RFU) of
167
resazurin in uninfected, WT Kunjin, or S411A Kunjin infected Vero and HEK293T cells
168
every 24 hours for six days (Fig. 4A and B). We also measured media as a negative
169
control to determine the baseline media fluorescence. The cell viability measurements
170
of uninfected Vero and HEK293T cells increased gradually over the duration of the
171
experiment suggesting that the cells are healthy and growing for the entirety of the
172
experiment. The cell viability measurements during the first 72 hours for WT Kunjin
173
infection in Vero and HEK293T cells were similar to that of uninfected cells. However,
174
cell viability measurements were lower in fluorescent signal compared to uninfected
175
cells. After 72 hours post infection (p.i.), cell viability measurements for WT Kunjin
176
infections continued to increase in fluorescence reaching 7.5 0.3 x105 RFU at 120
177
hours p.i. for Vero cells and 7.8 0.2 x105 RFU at 96 hours p.i. for HEK293T cells. After
178
which point, cell viability measurements decreased in fluorescence by 144 hours p.i.
179
suggesting that WT Kunjin induced cell toxicity is overtaking cellular replication. In the
180
case of S411A Kunjin infected Vero and HEK293T cells during the first 72 hours, cell
181
viability measurements demonstrated similar levels of fluorescence to that of uninfected
182
9
cells. Although the cell viability measured for S411A Kunjin was decreased compared to
183
uninfected cells. As the S411A Kunjin infection continued, cell viability measurements
184
significantly reduced in fluorescence between 96 and 144 hours p.i. ending with 5.3
185
0.3 x105 RFU for Vero cells and 5.7 0.2 x105 RFU for HEK293T cells. Together, these
186
data suggest that cells are relatively healthy in Kunjin infected cells for at least the first
187
72 hours in Vero and HEK293T cells; after which point population cell viability in S411A
188
Kunjin infected cells is negatively affected immediately in both cell lines, whereas a 24
189
hour and 48 hour delay are observed for decreased cell viability measurements with WT
190
Kunjin infection for HEK293T and Vero cells, respectively.
191
192
Another way to infer metabolically active cells or cell viability is through detection of
193
intracellular ATP levels. We utilized the CellTiter-Glo assay which uses the luciferase
194
reaction, an ATP-dependent reaction, to convert luciferin to oxyluciferin and several
195
byproducts including light (34). The byproduct, light, was measured in relative
196
luminescence units (RLU) for uninfected, WT Kunjin or S411A Kunjin infected Vero and
197
HEK293T cells every 24 hours for six days (Fig. 4C and D). Over the course of the
198
experiment, uninfected Vero cells progressively increased in luminescence from 5.5
199
0.3 x105 to 1.4 0.1 x106 RLU (Fig. 4C) suggesting that the uninfected cells were
200
healthy and metabolically active for the six-day experiment. However, cell viability
201
measurements of uninfected HEK293T cells increased linearly for the first 72 hours;
202
after which point, the cell viability measurements decreased and then leveled off at 1.7
203
0.07 x106 RLU (Fig. 4D), suggesting that uninfected HEK293T cells become less
204
metabolically active after 96 hours compared to the Vero cells. As for infection with WT
205
10
Kunjin, the cell viability measurements steadily increased for the first 72 hours for Vero
206
cells and for the first 48 hours for HEK293T cells similar to the observed cell viability
207
measurements of uninfected Vero and HEK293T cells. At 96 hours p.i. in Vero cells and
208
72 hours p.i. in HEK293T cells, cell viability measurements of WT Kunjin infected cells
209
decreased compared to uninfected cells. The population cell viability of WT Kunjin
210
infected cells continued to decrease reaching 6.5 3.0 x104 RLU in Vero cells and 4.0
211
2.0 x105 RLU in HEK293T cells at 144 hours. These data suggested that infection with
212
WT Kunjin negatively affected cell viability after 72 hours p.i. compared to uninfected
213
cell viability. On the other hand, cell viability measurements with S411A Kunjin infection
214
decreased after 24 hours p.i. in Vero cells and after 48 hours p.i. for HEK293T cells. For
215
the remainder of the experiment, the population cell viability continued to decrease in
216
S411A Kunjin infected Vero and HEK293T cells suggesting that both Vero and
217
HEK293T cells are extremely sensitive to S411A Kunjin and thus cell viability is
218
significantly reduced in the presence of the mutated virus. Together, these results
219
suggest that infection with S411A Kunjin in either Vero or HEK293T cells negatively
220
affected cell viability more quickly than infection with WT Kunjin.
221
222
S411A Kunjin results in decreased and delayed viral replication kinetics. The
223
results presented in the previous section indicated that S411A Kunjin induced increased
224
cellular death during infection. This prompted the question: how does increased cellular
225
death resulting from infection with S411A Kunjin affect replication kinetics of the virus?
226
Therefore, we performed a multi-step replication kinetics experiment with WT or S411A
227
Kunjin infected HEK293T cells at a MOI of 0.01 PFU/cell over a five day period. Every
228
11
12 hours viruses were collected and viral titers were determined via focus forming
229
assays (Fig. 5). At 12 hours post infection, the WT and S411A Kunjin viral titers were
230
not significantly different. At 24 hours p.i., S411A Kunjin remained in the lag phase while
231
WT Kunjin had entered the exponential replication phase, demonstrating delayed
232
replication with the S411A Kunjin infection. Over the last four days of infection, S411A
233
Kunjin maintained and expanded the initial delay in exponential replication and reached
234
an ~1 log lower peak viral titer compared to WT Kunjin. Overall, these data suggest that
235
S411A Kunjin does not replicate as efficiently as WT Kunjin. These results are
236
consistent with data reported by Du Pont et al., suggesting that the increased helicase
237
unwinding activity seen with the recombinant S411A NS3 helicase negatively affects
238
viral replication in fully infectious S411A Kunjin virus (33). Considering the observations
239
that S411A Kunjin resulted in decreased viral replication and increased cellular death,
240
we next investigated the effects of the S411A mutation on Kunjin infection in vivo.
241
242
S411A Kunjin results in increased mortality in mosquitoes compared to WT
243
Kunjin when IT injected but not when bloodfed. For the in vivo studies, we did not
244
have access to a colony of Cx. annulirostris mosquitoes, the primary vector for Kunjin
245
virus, but we had an established colony of Cx. quinquefasciatus that are infectable by
246
Kunjin virus. Cx. quinquefasciatus mosquitoes were bloodfed with defibrinated calf’s
247
blood diluted by half with titer equilibrated WT Kunjin, S411A Kunjin, or media alone as
248
a negative control. Similarly, female Cx. quinquefasciatus mosquitoes were subjected to
249
intrathoracic injection (IT) of 345 plaque forming units (PFU) per mosquito of WT Kunjin,
250
S411A Kunjin, or conditioned media. Mosquito mortality was recorded daily for 15 or 9
251
12
days, respectively. Overall, virus exposed mosquito mortality was low in both the
252
bloodfed and IT injected cohorts (Fig. 6), consistent with previous observations of Kunjin
253
virus in Cx. quinquefasciatus mosquitoes (37). When bloodfed, no difference was
254
observed in mortality rates for mosquitoes exposed to WT Kunjin vs. S411A Kunjin.
255
However, the small rate of mortality for virus exposed mosquitoes (~10%) was
256
significantly different from mosquitoes exposed to media alone (Fig. 6A). In contrast
257
with bloodfed data but consistent with cell culture and replication kinetics data, when
258
virus was introduced through IT injection, to bypass the midgut barrier, only S411A
259
Kunjin resulted in increased mortality (Fig. 6B). Together these data suggest that S411A
260
Kunjin is more lethal to mosquitoes than WT Kunjin once the virus has been able to
261
establish infections and/or transverse through the mosquito midgut barrier. This result
262
led us to further investigate the specifics of infection of Cx. quinquefasciatus by WT and
263
S411A Kunjin viruses.
264
265
S411A Kunjin has a lower infection rate but disseminates more efficiently than
266
WT Kunjin. Similar to the mortality experiments, Cx. quinquefasciatus mosquitoes were
267
infected with either WT Kunjin or S411A Kunjin by bloodmeal. Mosquito legs/wings,
268
saliva, and bodies were collected after 7 days and determined to be positive or negative
269
for infection by plaque assay. While ~58% of mosquitoes infected with WT Kunjin were
270
positive for the virus at day 7, only ~8% of mosquitoes infected with S411A Kunjin were
271
positive (Fig. 7A). Dissemination was inefficient for WT Kunjin with only 6% of
272
mosquitoes having positive titers in the legs and wings, demonstrating a strong barrier
273
to escape from the midgut. Similarly, less than 2% of infected mosquitoes resulted in
274
13
positive saliva samples (Fig. 7A). Despite low infection rates for mosquitoes infected
275
with S411A Kunjin, positive legs/wings and saliva were identified across multiple
276
replicate experiments, with nearly 50% of infected mosquitoes having disseminated
277
virus and 50% of those with disseminated virus having positive saliva. These data led to
278
the question: does S411A Kunjin allow for higher relative rates of dissemination?
279
280
To answer this question a second, much larger cohort of Cx. quinquefasciatus
281
mosquitoes were infected by bloodmeal with WT Kunjin or S411A Kunjin. Enough
282
mosquitoes were dissected to generate and estimated 30 infected mosquitoes per
283
condition: 60 exposed to WT Kunjin and 390 exposed to S411A Kunjin. Since
284
mosquitoes continue to die up to 14 days post bloodfeed, mosquitoes were collected at
285
14 days post blood meal instead of 7 days in an attempt to assure sufficient numbers of
286
S411A Kunjin infected mosquitoes. Again, WT Kunjin was observed to infect a larger
287
percent of exposed mosquitoes compared with S411A Kunjin (~30% vs. ~15%) (Fig.
288
7B,C), whereas, S411A Kunjin demonstrated higher rates of dissemination compared
289
with WT Kunjin (Fig. 7B,D). No legs/wings or saliva samples from WT Kunjin infected
290
mosquitoes were found to be positive at 14 days post blood meal (Fig. 7B,D,E). In
291
contrast and supporting these data from smaller cohorts collected at 7 days post blood
292
meal, 48% of S411A Kunjin infected mosquitoes had infected legs/wings and 61% of
293
mosquitoes with S411A Kunjin infected legs/wings resulted in positive saliva samples.
294
These data demonstrate that the S411A Kunjin was less capable of infecting Cx.
295
quinquefasciatus via blood meal compared with WT Kunjin. However, these data also
296
suggest that when S411A Kunjin was able to establish infection in Cx. quinquefasciatus
297
14
mosquitoes it is able to escape the midgut barrier more efficiently than WT Kunjin,
298
resulting in dissemination, infection of the salivary glands, and delivery to the saliva.
299
Finally, when considered in combination with the survival data, these data further
300
support that when S411A Kunjin was able to establish infection in Cx. quinquefasciatus
301
mosquitoes it is more lethal.
302
303
DISCUSSION
304
Previous work by our group has supported the hypothesis that Motif V in flavivirus NS3
305
helicase is a communication hub for translocation and unwinding of the dsRNA
306
intermediate during flavivirus replication (32, 33). More specifically, we found that NS3
307
Motif V residues T407 and S411 exhibit an increased helicase unwinding activity in
308
biochemical assays when mutated to alanine residues, while we observed a reduction in
309
replication of T407 and S411 mutant replicons. These previous results suggest that
310
T407 and S411 are responsible for regulating NS3 helicase function during flavivirus
311
replication. In this study we further investigated the role of T407 and S411 helicase
312
residues in the full-length infectious Kunjin virus in cell culture and in vivo experiments.
313
S411A Kunjin was successfully recovered and confirmed via sequencing (Fig. 2).
314
However, T407A Kunjin was not recovered which was consistent with the previous
315
results indicating ablated viral genome replication activity (33). We utilized WT Kunjin
316
and S411A Kunjin in several cell culture experiments including viral replication,
317
resazurin and CellTiter-Glo assays. Additionally, we compared WT Kunjin and S411A
318
Kunjin in several in vivo experiments including infection, dissemination and transmission
319
within Cx. quinquefasciatus mosquitoes. We observed that the S411A Kunjin reduced
320
15
cell viability during infection leading to increased cytopathic effect observed in the
321
plaque morphology and several metabolic assays in cell culture. Additionally, results
322
demonstrated a lower initial infection rate for S411A Kunjin within mosquitoes but once
323
infection is established efficient dissemination occurs compared with WT Kunjin
324
infections, potentially causing the observed increased mortality rates in mosquitoes.
325
Overall, our data suggest that the NS3 S411 in Motif V influences infection induced
326
cellular death and subsequent mortality in mosquito vectors.
327
328
Plaque morphology of viruses is a classical indicator of the effects of a mutation on viral
329
cytopathic effect in cells and spread between cells. We observed large and fuzzy
330
plaques with WT Kunjin, while S411A Kunjin plaques were small and clearly defined
331
(Fig. 3), suggesting that S411A Kunjin is more toxic to cells, but is not able to spread as
332
rapidly as WT Kunjin. Our previous work had indicated that the S411A mutation in a
333
replicon-based system reduced viral genome replication (33), so the small plaque size
334
was expected. However, the formation of clearer plaques was not. Therefore, we
335
performed a more quantitative investigation of S411A Kunjin effect on cell viability using
336
two assays (resazurin and CellTiter-Glo) that probed for different aspects of
337
metabolically active cells, NADH content and ATP content. The results from both
338
assays indicated that infection with S411A Kunjin results in a larger decrease in
339
metabolic activity compared to WT Kunjin within both HEK293T and Vero cells (Fig. 4).
340
Previously, studies have shown that reduced intracellular ATP levels leads to
341
proteasome inhibition that induces apoptosis leading to cellular death (38–43).
342
Therefore, our metabolic activity data is consistent with our plaque morphology data in
343
16
that infection with S411A Kunjin results reduced intracellular ATP levels and increased
344
cytopathic effect through increased cell death. S411A Kunjin exhibited delayed and
345
decreased viral replication kinetics compared to WT Kunjin (Fig. 5) suggesting that even
346
though the mutated Kunjin virus is more toxic to cells, it does not replicate as efficiently
347
as WT Kunjin. These data are consistent with previous studies reporting a decrease in
348
viral genome replication with S411A helicase replicon (33).
349
350
An interesting but different hypothesis is that hyperactive NS3 helicase affects cellular
351
mRNA. Studies on NS3 helicase function have focused primarily on its effect on
352
genome replication and packaging (44), but our finding that a NS3 hyperactive helicase
353
mutant increases cell death opens up the possibility that NS3 has roles in altering
354
cellular physiology as well. Previously observed results indicated that recombinant NS3
355
S411A helicase mutant had a higher helicase rate but did not have a significantly higher
356
ATPase rate (33), so it is unlikely that reduction of cell viability was due to decreased
357
ATP from NS3 ATP degradation. However, it is possible that increased cytotoxicity is
358
due to another effect of helicase activity on cellular physiology. The hyperactive NS3
359
helicase may be interacting with cellular RNAs leading to dysregulation of cellular
360
homeostasis. NS3 could bind to cellular mRNAs and unwind their secondary structures,
361
causing a disruption in RNA stability and recruitment of translational factors. This
362
unwinding of cellular mRNAs would result in an imbalance within the cell inducing
363
cellular apoptosis. We are currently exploring if NS3 effects cellular RNAs.
364
365
17
Observed reductions in cell viability led us to investigate the effect of S411A on infection
366
in mosquitoes. Generally, the longevity of mosquitoes infected with flaviviruses are
367
similar to that of uninfected mosquitoes (45, 46). During mosquito infection, flaviviruses
368
must overcome four barriers: 1) midgut infection barrier, 2) midgut escape barrier, 3)
369
salivary gland infection barrier, and 4) salivary gland escape barrier (47). For the first
370
barrier, the virus must successfully infect and replicate in the midgut epithelial cells (47,
371
48). Infection is dependent on the arbovirus-specific interactions with the midgut
372
epithelial receptors (49). If the virus cannot establish an infection in the midgut epithelial
373
cells, then the mosquito cannot be infected by the virus. If the virus can establish
374
infection in the midgut, then the next barrier is escaping the midgut by crossing the
375
basal lamina which surrounds the midgut epithelium (47). After escaping the midgut, the
376
virus can disseminate throughout the rest of the mosquito tissues. If the virus is able to
377
penetrate into the salivary gland, the virus must replicate and be deposited into the
378
apical cavities of acinar cells for the mosquito to transmit the virus to other hosts (47).
379
Not all mosquitoes will be able to transmit virus due to unknown reasons. Culex
380
mosquitoes in our study were bloodfed or submitted to intrathoracic injection (IT) with
381
either WT or S411A Kunjin. Mosquito mortality was recorded for 15 days for bloodfed
382
mosquitoes or 9 days for IT injected mosquitoes. Results indicated no significant
383
difference in mortality between mosquitoes bloodfed with either WT or S411A Kunjin
384
viruses. Mosquitoes that were intrathoracically injected with S411A Kunjin exhibited an
385
increase in mortality compared to WT Kunjin. Together, our data suggests that S411A
386
Kunjin viruses were inefficient at crossing the midgut infection barrier to establish
387
infection (Fig. 7). However, upon bypassing the midgut infection and midgut escape
388
18
barriers through IT injection S411A Kunjin was more lethal (Fig. 6B). The basis for the
389
observed increased mortality is not yet clear but could be due to increased cytopathic
390
effect in infected cells similar to what was observed in cell culture.
391
392
To further investigate the distribution of WT Kunjin and S411A Kunjin infection within the
393
Cx. quinquefasciatus mosquitoes, bodies, legs/wings, and saliva were collected after 7
394
or 14 days post-bloodfeed and analyzed for the presence of virus. 30 (day 14) to 50%
395
(day 7) of mosquito bodies were positive for WT Kunjin infection, whereas less than
396
15% of bodies were positive for S411A Kunjin on either collection day. These data
397
suggest that WT Kunjin was able to routinely establish infection within midgut epithelial
398
cells, while S411A Kunjin did so less effectively. However, when legs/wings and saliva
399
were analyzed, WT Kunjin was found at extremely low levels, while S411A Kunjin was
400
found in over half of infected mosquitoes suggesting that once S411A Kunjin was able
401
to cross the midgut escape barrier, it was able to replicate more efficiently in peripheral
402
tissues than WT Kunjin. Previous studies have suggested that arboviruses may require
403
apoptosis to escape the midgut and infect the salivary glands of Culex mosquitoes (48,
404
50–53). Thus, taking into account the cell culture results suggesting S411A Kunjin
405
induces increased cellular death, S411A Kunjin viruses may be able to exit the midgut
406
more effectively than WT Kunjin due to increased induction of apoptosis. Even though
407
S411A Kunjin has a lower initial infection rate, the mutant virus is more toxic to infected
408
cells, and thus, the mutant virus may be able to induce apoptosis and disseminate into
409
the rest of the body leading to a higher potential transmission rate with increased
410
salivary gland infection.
411
19
412
In conclusion, this study provides insight into how a hyperactive NS3 helicase mutant
413
virus contributes to Kunjin virus replication and the effect on cellular responses during
414
infection. S411A Kunjin negatively affects overall replication of the virus and increases
415
the cytopathic effect in cells potentially resulting in increased mosquito mortality.
416
Infection with S411A Kunjin results in less metabolic activity in cells and ultimately
417
cellular death. When considering the increased mortality of mosquitoes IT injected with
418
S411A Kunjin, it seems likely that cells within mosquitoes are undergoing similar
419
cytopathic effect as was observed in cell culture. Cellular death in mosquitoes could
420
allow S411A Kunjin to disseminate into the legs/wings and saliva more efficiently than
421
WT Kunjin and result in increased mosquito death. Virus-induced mortality is not ideal
422
for long-term maintenance of virus in mosquitoes, so flaviviruses appear to have
423
evolved mechanisms to reduce their helicase activity to reduce virus-induced cell killing.
424
Overall, these data indicate that NS3 helicase activity may have significant roles during
425
viral infection in cell culture and in vivo, and that NS3 Motif V may play a central role in
426
controlling virus-induced mortality in mosquito vectors to allow for efficient viral
427
transmission.
428
429
MATERIALS AND METHODS
430
Cell Culture and Viruses. HEK293T and Vero (African Green Monkey kidney
431
epithelial) cells were maintained in Hyclone Dulbecco’s modified Eagle medium
432
(DMEM) supplemented with 10% fetal bovine serum (FBS), 50 mM HEPES (pH 7.5),
433
5% penicillin/streptomycin and 5% L-Glutamine. All cells were grown in humidified
434
20
incubators at 37 C with 5% CO2. The West Nile virus (Kunjin subtype) infectious clone
435
was generously provided from Alexander Khromykh (University of Queensland) (54).
436
437
Virus Mutagenesis. To produce the T407A Kunjin and S411A Kunjin NS3 mutants
438
viruses, a novel bacteria-free virus launch system was used based on in vitro NEBuilder
439
assembly of PCR-amplified DNAs containing a eukaryotic Pol II promoter with PCR
440
fragments containing viral genome sequences and direct transfection of assembled
441
DNAs into Vero cells. Three PCR fragments were produced using the Q5 DNA
442
polymerase system (New England Biolabs) according to the manufacturer’s instructions
443
(54). PCR fragment #1 contained the cytomegalovirus (CMV) immediate early promoter
444
(612 bp) using pcDNA-3.1 as the PCR template. PCR fragment #2 (5867 bp) contained
445
the 5’ region of the Kunjin virus genome. PCR fragment #3 (5309 bp) contained the 3’
446
end of the Kunjin virus genome in addition to a hepatitis delta virus ribozyme. The
447
Kunjin virus infectious clone plasmid FLSDXHDVr was used as the PCR template for
448
fragments #2 and #3 (55). Primer sequences used to produce PCR fragments with
449
overlapping 5’ and 3’ ends for NEBuilder assembly were designed using the NEBuilder
450
Assembly tool (https://nebuilder.neb.com/) and are listed in Table 1.
451
452
The NS3 T407A and S411A mutations(33) were separately engineered into the
453
Fragment #2 reverse primer and Fragment #3 forward primers. PCR products were gel
454
extracted with the Qiagen Gel Extraction kit and quantified by UV spectrophotometry
455
and agarose gel electrophoresis. To assemble the WT Kunjin, T407A Kunjin, or S411A
456
Kunjin fragments, equal molar amounts of each fragment were mixed in a total DNA
457
21
mass of 200 ng for each virus in ultrapure water in a final volume of 15 µL. An equal
458
volume of New England Biolabs NEBuilder 2X Master Mix was added to the DNAs, and
459
the reaction was incubated at 50C for 4 hrs. The assembled DNAs were transfected
460
directly into Vero cells by adding 1 µL of JetPrime transfection reagent (PolyPlus) to the
461
assembly mixture, incubated at 22C for 15 minutes, and the transfection mixture was
462
added to 50% confluent Vero cells. DMEM media containing 10% fetal bovine serum
463
and 50 mM HEPES (pH 7.5) was changed 24 hours after transfection, and the cells
464
were incubated for 6 additional days and monitored for cytopathic effect. Media was
465
collected on day 6 as the P0 stock. Virus was amplified in a T75 flask seeded at 50%
466
confluency for 7 additional days, and clarified media was collected as the P1 stock.
467
Finally, the P1 stock was used to infect a T150 flask of 50% confluent Vero cells for 7
468
days, media was collected and clarified of cellular debris, and clarified media frozen at -
469
80C as the P2 stock. P2 stocks were quantified for infectivity via focus forming assay.
470
T407A Kunjin was unrecoverable from infections. The presence of the S411A Kunjin
471
was verified by extracting RNA from the P2 stock, reverse transcribing and PCR
472
amplifying the NS3 region of Kunjin virus using Kunjin NS3 sequence forward (5’-
473
ATGCACCAATATCCGACTTACA) and reverse (5’- TGGCCTCAGAATCTTCCTTTC)
474
primers, and the sequence of the PCR 794 bp amplicon determine by Sanger
475
sequencing.
476
477
Viral Infectivity. HEK293T cells were plated into 12-well plates at 20,000 cells/well and
478
allowed to adhere to the plates overnight. The next day, the cells were infected at a MOI
479
of 0.01 PFU/cell with either WT Kunjin or S411A Kunjin in triplicate under BSL2
480
22
conditions. Both intracellular and extracellular RNA samples were collected every 12
481
hours for five days. The extracellular RNA samples were processed through focus
482
forming assays to determine the viral titer at each time point. The growth curves were
483
plotting using matplotlib (56).
484
485
Resazurin Assay. HEK293T cells were plated into 96-well plates at 10,000 cells/well.
486
Additionally, DMEM with 10% FBS was plated into one row for each plate as a negative
487
control for resazurin. The following day, cells were either not infected or infected with
488
either WT or S411A Kunjin at a MOI of five PFU/cell. The DMEM media was not
489
infected. Every 24 hours over the course of six days, the cells as well as the negative
490
control were treated with resazurin (0.15 mg/mL). The treated plate was then incubated
491
for 1 hour at 37C with 5% CO2 before measuring the fluorescence at an excitation
492
wavelength of 560 nm and an emission wavelength of 590 nm on a Victor X5 multilabel
493
plate reader (Perkin Elmer).
494
495
CellTiter-Glo Assay. Vero and HEK293T cells were plated into 96-well plates at 10,000
496
cells/well. The following day, cells in each plate were either not infected or infected with
497
WT or S411A Kunjin at a MOI of five PFU/cell. Every 24 hours for the next six days,
498
cells were treated with 1X of CellTiter-Glo and incubated at room temperature for 10
499
minutes before measuring luminescence with an exposure time of 0.5 seconds on a
500
Victor X5 multilabel plate reader.
501
502
23
Mosquitoes. Cx. quinquefasciatus mosquito larvae(57), were propagated on a 1:1 mix
503
of powdered Tetra food and powdered rodent chow. Adult mosquitoes were kept on a
504
16:8 light:dark cycle at 28C with 70%-80% humidity. Water and sugar were provided
505
ad libitum and citrated sheep blood was provided to maintain the colony. Mosquito
506
infection experiments with Kunjin were performed exclusively on female mosquitoes and
507
under BSL3 conditions.
508
509
Infection of mosquitoes with Kunjin virus and analysis. Cx. quinquefasciatus
510
mosquitoes were either fed infectious bloodmeals or intrathoracically injected to
511
introduce Kunjin virus. Bloodfed mosquitoes were fed an infectious bloodmeal of
512
defibrillated calf blood diluted by half with 2.5 X 106 PFU/mL Kunjin virus, or media
513
alone as a negative control. Bloodmeals also contained 2 mM ATP. For IT injection
514
experiments, mosquitoes were injected with 138 nL WT or S411A Kunjin virus (~345
515
PFU/mosquito) using a Nanoject II (Drummond Scientific). Engorged female mosquitoes
516
were maintained for up to 15 days under conditions described above but in the BSL3
517
insectary and mortality rate counted daily. For infection, dissemination, and
518
transmission experiments after 7 or 14 days of incubation, mosquitos were cold
519
anesthetized and kept on ice while legs and wings were removed, mosquitoes were
520
salivated for 30 minute in a capillary tube filled with immersion oil, and bodies were
521
collected. Legs/wings and bodies were homogenized at 24Hz for 1 minute in 500 L
522
mosquito diluent with a stainless steel bead, and saliva samples were stored in 250 L
523
mosquito diluent as previously described (58). All mosquito samples were clarified by
524
24
centrifugation at 15,000 X g for 5 minute at 4C then determined to be positive or
525
negative by infection with undiluted samples by Vero cell plaque assays.
526
527
ACKNOWLEDGEMENTS. We would like to acknowledge the support of NIH grants
528
R01 AI132668 to BJG and R01 AI067380 to GDE. We would also like to acknowledge
529
the helpful discussion with Erin R. Lynch, MS.
530
531
REFERENCES
532
1.
Frost MJ, Zhang J, Edmonds JH, Prow NA, Gu X, Davis R, Hornitzky C, Arzey
533
KE, Finlaison D, Hick P, Read A, Hobson-Peters J, May FJ, Doggett SL, Haniotis
534
J, Russell RC, Hall RA, Khromykh AA, Kirkland PD. 2012. Characterization of
535
Virulent West Nile Virus Kunjin Strain, Australia, 2011. Emerg Infect Dis 18:792–
536
800.
537
2.
Gray TJ, Burrow JN, Markey PG, Whelan PI, Jackson J, Smith DW, Currie BJ.
538
2011. West nile virus (Kunjin subtype) disease in the northern territory of
539
Australia--a case of encephalitis and review of all reported cases. Am J Trop Med
540
Hyg 85:952–6.
541
3.
Solomon T. 2004. Flavivirus Encephalitis. N Engl J Med 351:370–378.
542
4.
Weaver SC, K RW. 2010. Present and Future Arboriral Threaths. Antiviral Res
543
85:1–36.
544
5.
World Health Organization. 2017. West Nile Virus Key Facts. Geneva,
545
Switzerland.
546
6.
Byas AD, Ebel GD. 2020. Comparative Pathology of West Nile Virus in Humans
547
25
and Non-Human Animals. Pathogens 9:48.
548
7.
Ciota AT. 2017. West Nile virus and its vectors. Curr Opin Insect Sci 22:28–36.
549
8.
Petersen LR, Carson PJ, Biggerstaff BJ, Custer B, Borchardt SM, Busch MP.
550
2013. Estimated cumulative incidence of West Nile virus infection in US adults,
551
1999-2010. Epidemiol Infect 141:591–5.
552
9.
Busch MP, Wright DJ, Custer B, Tobler LH, Stramer SL, Kleinman SH, Prince HE,
553
Bianco C, Foster G, Petersen LR, Nemo G, Glynn SA. 2006. West Nile virus
554
infections projected from blood donor screening data, United States, 2003. Emerg
555
Infect Dis 12:395–402.
556
10. Flynn LM, Coelen RJ, Mackenzie JS. 1989. Kunjin Virus Isolates of Australia Are
557
Genetically Homogeneous. J Gen Virol 70:2819–2824.
558
11. Grubaugh ND, Ebel GD. 2016. Dynamics of West Nile virus evolution in mosquito
559
vectors. Curr Opin Virol 21:132–138.
560
12. Lindenbach BD, Heinz-Jurgen T, Rice CM. 2007. Flaviviridae: The Viruses and
561
Their Replication, p. 1101–1152. In Knipe, DM, Howley, PM (eds.), Fields
562
Virology, 5th ed. Lippincott-Raven Publishers, Philadelphia.
563
13. Mastrangelo E, Milani M, Bollati M, Selisko B, Peyrane F, Pandini V, Sorrentino
564
G, Canard B, Konarev P V, Svergun DI, de Lamballerie X, Coutard B, Khromykh
565
A a, Bolognesi M. 2007. Crystal Structure and Activity of Kunjin Virus NS3
566
Helicase; Protease and Helicase Domain Assembly in the Full Length NS3
567
Protein. J Mol Biol 372:444–455.
568
14. Scherret JH, Poidinger M, Mackenzie JS, Broom AK, Deubel V, Lipkin WI, Briese
569
T, Gould EA, Hall RA. 2001. The Relationships between West Nile and Kunjin
570
26
Viruses. Emerg Infect Dis 7:697–705.
571
15. Hall RA, Scherret JH, Mackenzie JS. 2001. Kunjin virus: an Australian variant of
572
West Nile? Ann N Y Acad Sci 951:153–60.
573
16. Coia G, Parker MD, Speight G, Byrne ME, Westaway EG. 1988. Nucleotide and
574
Complete Amino Acid Sequences of Kunjin Virus: Definitive Gene Order and
575
Characteristics of the Virus-specified Proteins. J Gen Virol 69:1–21.
576
17. Liu WJ, Sedlak PL, Kondratieva N, Khromykh AA. 2002. Complementation
577
analysis of the flavivirus Kunjin NS3 and NS5 proteins defines the minimal regions
578
essential for formation of a replication complex and shows a requirement of NS3
579
in cis for virus assembly. J Virol 76:10766–75.
580
18. Neufeldt CJ, Cortese M, Acosta EG, Bartenschlager R. 2018. Rewiring cellular
581
networks by members of the Flaviviridae family. Nat Rev Microbiol 16:125–142.
582
19. Chambers TJ, Hahn CS, Galler R, Rice CM. 1990. Flavivirus genome
583
organization, expression, and replication. Annu Rev Microbiol 44:649–88.
584
20. Brand C, Bisaillon M, Geiss BJ. 2017. Organization of the Flavivirus RNA
585
replicase complex. Wiley Interdiscip Rev RNA 8:e1437.
586
21. Saeedi BJ, Geiss BJ. 2013. Regulation of flavivirus RNA synthesis and capping.
587
Wiley Interdiscip Rev RNA 4:723–35.
588
22. Luo D, Xu T, Watson RP, Scherer-Becker D, Sampath A, Jahnke W, Yeong SS,
589
Wang CH, Lim SP, Strongin A, Vasudevan SG, Lescar J. 2008. Insights into RNA
590
unwinding and ATP hydrolysis by the flavivirus NS3 protein. EMBO J 27:3209–19.
591
23. Mastrangelo E, Bolognesi M, Milani M. 2012. Flaviviral helicase: insights into the
592
mechanism of action of a motor protein. Biochem Biophys Res Commun 417:84–
593
27
7.
594
24. Pérez-Villa A, Darvas M, Bussi G. 2015. ATP dependent NS3 helicase interaction
595
with RNA: Insights from molecular simulations. Nucleic Acids Res 43:8725–8734.
596
25. Wang C-C, Huang Z-S, Chiang P-L, Chen C-T, Wu H-N. 2009. Analysis of the
597
nucleoside triphosphatase, RNA triphosphatase, and unwinding activities of the
598
helicase domain of dengue virus NS3 protein. FEBS Lett 583:691–6.
599
26. Warrener P, Tamura JK, Collett MS. 1993. RNA-stimulated NTPase activity
600
associated with yellow fever virus NS3 protein expressed in bacteria. J Virol
601
67:989–96.
602
27. Wengler G, Wengler G. 1991. The carboxy-terminal part of the NS 3 protein of the
603
West Nile flavivirus can be isolated as a soluble protein after proteolytic cleavage
604
and represents an RNA-stimulated NTPase. Virology 184:707–15.
605
28. Wengler G, Wengler G. 1993. The NS 3 Nonstructural Protein of Flaviviruses
606
Contains an RNA Triphosphatase Activity. Virology 197:265–273.
607
29. Fairman-Williams ME, Guenther UP, Jankowsky E. 2010. SF1 and SF2 helicases:
608
Family matters. Curr Opin Struct Biol 20:313–324.
609
30. Story RM, Steitz TA. 1992. Structure of the recA protein-ADP complex. Nature
610
355:374–6.
611
31. Swarbrick CMD, Basavannacharya C, Chan KWK, Chan S-A, Singh D, Wei N,
612
Phoo WW, Luo D, Lescar J, Vasudevan SG. 2017. NS3 helicase from dengue
613
virus specifically recognizes viral RNA sequence to ensure optimal replication.
614
Nucleic Acids Res 45:12904–12920.
615
32. Davidson RB, Hendrix J, Geiss BJ, McCullagh M. 2018. Allostery in the dengue
616
28
virus NS3 helicase: Insights into the NTPase cycle from molecular simulations.
617
PLoS Comput Biol 14:e1006103.
618
33. Du Pont KE, Davidson RB, McCullagh M, Geiss BJ. 2020. Motif V regulates
619
energy transduction between the flavivirus NS3 ATPase and RNA-binding cleft. J
620
Biol Chem 295:1551–1564.
621
34. Daniel F. Gilbert OF. 2017. Cell Viability Assays. Springer New York, New York,
622
NY.
623
35. O’Brien J, Wilson I, Orton T, Pognan F. 2000. Investigation of the Alamar Blue
624
(resazurin) fluorescent dye for the assessment of mammalian cell cytotoxicity. Eur
625
J Biochem 267:5421–5426.
626
36. Rampersad SN. 2012. Multiple Applications of Alamar Blue as an Indicator of
627
Metabolic Function and Cellular Health in Cell Viability Bioassays. Sensors
628
12:12347–12360.
629
37. Kay BH, Fanning ID, Carley JG. 1982. Vector competence of Culex pipiens
630
quinquefasciatus for Murray Valley encephalitis, Kunjin, and Ross River viruses
631
from Australia. Am J Trop Med Hyg 31:844–848.
632
38. Huang H, Zhang X, Li S, Liu N, Lian W, McDowell E, Zhou P, Zhao C, Guo H,
633
Zhang C, Yang C, Wen G, Dong X, Lu L, Ma N, Dong W, Dou QP, Wang X, Liu J.
634
2010. Physiological levels of ATP negatively regulate proteasome function. Cell
635
Res 20:1372–85.
636
39. Latta M, Künstle G, Leist M, Wendel A. 2000. Metabolic Depletion of Atp by
637
Fructose Inversely Controls Cd95- and Tumor Necrosis Factor Receptor 1–
638
Mediated Hepatic Apoptosis. J Exp Med 191:1975–1986.
639
29
40. Eguchi Y, Shimizu S, Tsujimoto Y. 1997. Intracellular ATP levels determine cell
640
death fate by apoptosis or necrosis. Cancer Res 57:1835–40.
641
41. Leist M, Single B, Castoldi AF, Kühnle S, Nicotera P. 1997. Intracellular
642
adenosine triphosphate (ATP) concentration: a switch in the decision between
643
apoptosis and necrosis. J Exp Med 185:1481–6.
644
42. Liu J, Zheng H, Tang M, Ryu Y-C, Wang X. 2008. A therapeutic dose of
645
doxorubicin activates ubiquitin-proteasome system-mediated proteolysis by acting
646
on both the ubiquitination apparatus and proteasome. Am J Physiol Heart Circ
647
Physiol 295:H2541-50.
648
43. Kwong JQ, Henning MS, Starkov AA, Manfredi G. 2007. The mitochondrial
649
respiratory chain is a modulator of apoptosis. J Cell Biol 179:1163–77.
650
44. Patkar CG, Kuhn RJ. 2008. Yellow Fever Virus NS3 Plays an Essential Role in
651
Virus Assembly Independent of Its Known Enzymatic Functions. J Virol 82:3342–
652
3352.
653
45. Samuel GH, Adelman ZN, Myles KM. 2018. Antiviral Immunity and Virus-
654
Mediated Antagonism in Disease Vector Mosquitoes. Trends Microbiol 26:447–
655
461.
656
46. Chamberlain RW, Sudia WD. 1961. Mechanism of Transmission of Viruses by
657
Mosquitoes. Annu Rev Entomol 6:371–390.
658
47. Rückert C, Ebel GD. 2018. How Do Virus-Mosquito Interactions Lead to Viral
659
Emergence? Trends Parasitol 34:310–321.
660
48. Franz A, Kantor A, Passarelli A, Clem R. 2015. Tissue Barriers to Arbovirus
661
Infection in Mosquitoes. Viruses 7:3741–3767.
662
30
49. Mercado-Curiel RF, Black WC, Muñoz M de L. 2008. A dengue receptor as
663
possible genetic marker of vector competence in Aedes aegypti. BMC Microbiol
664
8:118.
665
50. Wang H, Gort T, Boyle DL, Clem RJ. 2012. Effects of Manipulating Apoptosis on
666
Sindbis Virus Infection of Aedes aegypti Mosquitoes. J Virol 86:6546–6554.
667
51. Kelly EM, Moon DC, Bowers DF. 2012. Apoptosis in mosquito salivary glands:
668
Sindbis virus-associated and tissue homeostasis. J Gen Virol 93:2419–2424.
669
52. Liu Q, Clem RJ. 2011. Defining the core apoptosis pathway in the mosquito
670
disease vector Aedes aegypti: the roles of iap1, ark, dronc, and effector
671
caspases. Apoptosis 16:105–113.
672
53. O’Neill K, Olson BJSC, Huang N, Unis D, Clem RJ. 2015. Rapid selection against
673
arbovirus-induced apoptosis during infection of a mosquito vector. Proc Natl Acad
674
Sci U S A 112:E1152-61.
675
54. Khromykh AA, Westaway EG. 1994. Completion of Kunjin virus RNA sequence
676
and recovery of an infectious RNA transcribed from stably cloned full-length
677
cDNA. J Virol 68:4580–8.
678
55. Khromykh AA, Sedlak PL, Guyatt KJ, Hall RA, Westaway EG. 1999. Efficient
679
trans-complementation of the flavivirus kunjin NS5 protein but not of the NS1
680
protein requires its coexpression with other components of the viral replicase. J
681
Virol 73:10272–80.
682
56. Hunter JD. 2007. Matplotlib: A 2D Graphics Environment. Comput Sci Eng 9:90–
683
95.
684
57. Ciota AT, Chin PA, Kramer LD. 2013. The effect of hybridization of Culex pipiens
685
31
complex mosquitoes on transmission of West Nile virus. Parasit Vectors 6:305.
686
58. Weger-Lucarelli J, Duggal NK, Bullard-Feibelman K, Veselinovic M, Romo H,
687
Nguyen C, Rückert C, Brault AC, Bowen RA, Stenglein M, Geiss BJ, Ebel GD.
688
2017. Development and Characterization of Recombinant Virus Generated from a
689
New World Zika Virus Infectious Clone. J Virol 91:1–11.
690
691
692
FIGURE LEGENDS
693
FIG. 1. S411 and T407 interaction within Motif V in NS3 helicase. Residues T407
694
and S411 interact with each other through a hydrogen bond within Motif V.
695
696
FIG. 2. Verification of alanine mutation in S411A Kunjin virus via Sanger
697
sequencing. Results from Sanger sequencing verifies alanine mutation for position 411
698
through the presence of the alanine codon (highlighted in red box). The original serine
699
codon within the red box was TCT. Two nucleotides were changed to introduce the
700
alanine mutation. Refer to GenBank accession number (AY274504.1) for wild-type
701
Kunjin FLSDX.
702
703
FIG. 3. Plaque morphology suggests an increased cytopathic effect for S411A
704
Kunjin. Viral titers were obtained for WT and S411A Kunjin viruses and the plaque
705
morphology is shown for A) WT Kunjin and B) S411A Kunjin.
706
707
32
FIG. 4. S411A Kunjin decreases cell viability. WT Kunjin and S411A Kunjin infected
708
A) Vero cells and B) HEK293T cells were measured for cellular metabolism through
709
resazurin. Similarly, WT Kunjin and S411A Kunjin infected C) Vero cells and D)
710
HEK293T were measured for intracellular ATP levels through CellTiter-Glo. All
711
infections were performed at a MOI of five PFU/cell.
712
713
FIG. 5. S411A Kunjin decreases and delays viral replication kinetics. Replication
714
kinetics experiments were performed for WT and S411A Kunjin viruses. HEK293T cells
715
were infected at a MOI of 0.01 PFU/cell.
716
717
FIG. 6. S411A Kunjin viruses are more lethal to Cx. quinquefaciatus mosquitoes
718
than WT Kunjin. Female Cx. quinquefaciatus mosquitoes were exposed to WT (blue
719
circles) or S411A (red triangles) Kunjin virus through either A) infectious bloodmeals, or
720
B) by IT injection. Control mosquitoes were exposed to bloodmeals containing media or
721
injected with media alone. Mortality was recorded daily for 15 or 9 days respectively.
722
Survival curves compared by Logrank test for trend (P<0.0001 = ****, P<0.05 = *) A) n =
723
425/condition, B) n = 40/condition.
724
725
FIG. 7. The S411A Kunjin is less capable than WT Kunjin of infecting mosquitoes
726
but disseminates and transmits more efficiently once established. Engorged
727
female Cx. quinquefascitus mosquitoes exposed to infectious bloodmeals containing
728
either WT or S411A Kunjin virus were housed for A) 7 or B-E) 14 days post bloodfeed.
729
Mosquitoes were dissected and legs/wings, saliva and bodies were collected and tested
730
33
for the presence of Kunjin virus by plaque assay. Data is shown as A and B) percent of
731
total exposed infected, C) total negative and positive bodies, D) positive legs/wings from
732
total infected, or E) total positive saliva from total disseminated. A) n = 64/condition, B)
733
WT Kunjin n = 60, S411A Kunjin n = 390.
734
735
Table 1. NEBuilder Primers for T407A and S411A Kunjin Viruses. The mutant
736
Kunjin viruses were generated from three fragments: #1, #2, and #3. Primers for
737
fragments #2 and #3 contain the alanine mutation at either position 407 or 411
738
(highlighted in red). The product of Fragment #2 from the NEBuilder Assembly reaction
739
will contain the specified mutation.
740
741
FIGURES
742
FIG. 1:
743
744
FIG. 2:
745
746
T407
S411
Motif V
NS3h
ssRNA
ATP
34
FIG. 3:
747
748
FIG. 4:
749
750
FIG. 5:
751
A.
B.
WT Kunjin
S411A Kunjin
C.
B.
A.
D.
C.
35
752
FIG. 6:
753
754
A
0
1
2
3
4
5
6
7
8
9
90
95
100
105
Days
Percent survival
IT Injected
Control
WT Kunjin
S411A Kunjin
*
A.
B.
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
90
95
100
105
Days
Percent survival
Bloodfed
****
36
FIG. 7:
755
756
757
758
TABLES
759
Table 1:
760
Bodies
Legs/Wings
Saliva
Bodies
Legs/Wings
Saliva
0
25
50
75
Percent Infected
WT Kunjin
S411A Kunjin
WT Kunjin
S411A Kunjin
Positive vs. Negative Legs/Wings
of Infected
Dissemination
+
-
WT Kunjin
S411A Kunjin
Positive vs. Negative Bodies
Infection
+
-
WT Kunjin
S411A Kunjin
Positive vs. Negative Saliva
of Disseminated
Transmission
+
-
A.
B.
C.
Bodies
Legs/Wings
Saliva
Bodies
Legs/Wings
Saliva
0
25
50
75
Percent Infected
D.
E.
37
761
NEBuilder Primers
Primer Sequence (5’-overlap/spacer/ANNEAL-3’)
CMV Forward
atcggaatctGATTATTGACTAGTTATTAATAGTAATCAATTACG
CMV Reverse
gcgaactactCGGTTCACTAAACGAGCTC
5’ Kunjin Forward
tagtgaaccgAGTAGTTCGCCTGTGTGAG
5’ Kunjin (T407A) Reverse
atatatctgtGGCGACGACAAAGTCCCAATC
3’ Kunjin (T407A) Forward
tgtcgtcgccACAGATATATCTGAGATGGG
3’ Kunjin Reverse
gtcaataatcTTCCGATAGAGAATCGAG
5’ Kunjin Forward
tagtgaaccgAGTAGTTCGCCTGTGTGAG
5’ Kunjin (S411A) Reverse
ctcccatctcTGCTATATCTGTTGTGACGAC
3’ Kunjin (S411A) Forward
agatatagcaGAGATGGGAGCAAACTTTAAG
3’ Kunjin Reverse
gtcaataatcTTCCGATAGAGAATCGAG
Fragment #
#1: CMV
#2: 5’ T407A
Kunjin Virus
#3: 3’ T407A
Kunjin Virus
#2: 5’ S411A
Kunjin Virus
#3: 3’ S411A
Kunjin Virus
| 2020 | A Hyperactive Kunjin Virus NS3 Helicase Mutant Demonstrates Increased Dissemination and Mortality in Mosquitoes | 10.1101/2020.05.26.117580 | [
"Du Pont Kelly E.",
"Sexton Nicole R.",
"McCullagh Martin",
"Ebel Gregory D.",
"Geiss Brian J."
] | creative-commons |
1
Biological condensates form percolated networks with molecular motion properties
1
distinctly different from dilute solutions
2
3
Zeyu Shen1, Bowen Jia1, Yang Xu1, Jonas Wessén2, Tanmoy Pal2, Hue Sun Chan2,
4
Shengwang Du3,4,7, and Mingjie Zhang1,5,6,*
5
6
1Division of Life Science, Hong Kong University of Science and Technology, Clear
7
Water Bay, Kowloon, Hong Kong, China.
8
2Department of Biochemistry, University of Toronto, Toronto, Ontario, Canada
9
3Department of Physics, Hong Kong University of Science and Technology, Clear Water
10
Bay, Kowloon, Hong Kong, China
11
4Department of Chemical and Biological Engineering, Hong Kong University of Science
12
and Technology, Clear Water Bay, Kowloon, Hong Kong, China
13
5Greater Bay Biomedical Innocenter, Shenzhen Bay Laboratory, Shenzhen 518036, China
14
6School of Life Sciences, Southern University of Science and Technology, Shenzhen
15
518055, China
16
17
18
19
Running title: biological condensates form percolated networks
20
21
7Present address: Department of Physics, The University of Texas at Dallas, Richardson,
22
Texas 75080, USA
23
24
25
*Corresponding authors (Mingjie Zhang: zhangmj@sustech.edu.cn)
26
27
28
2
Abstract
29
Formation of membraneless organelles or biological condensates via phase separation hugely
30
expands cellular organelle repertoire. Biological condensates are dense and viscoelastic soft
31
matters instead of canonical dilute solutions. Unlike discoveries of numerous different
32
biological condensates to date, mechanistic understanding of biological condensates remains
33
scarce. In this study, we developed an adaptive single molecule imaging method that allows
34
simultaneous tracking of individual molecules and their motion trajectories in both condensed
35
and dilute phases of various biological condensates. The method enables quantitative
36
measurements of phase boundary, motion behavior and speed of molecules in both condensed
37
and dilute phases as well as the scale and speed of molecular exchanges between the two
38
phases. Surprisingly, molecules in the condensed phase do not undergo uniform Brownian
39
motion, but instead constantly switch between a confined state and a random motion state.
40
The confinement is consistent with formation of large molecular networks (i.e., percolation)
41
specifically in the condensed phase. Thus, molecules in biological condensates behave
42
distinctly different from those in dilute solutions. This finding is of fundamental importance
43
for understanding molecular mechanisms and cellular functions of biological condensates in
44
general.
45
3
Introduction
46
Phase separation-mediated formation of condensed macro-molecular assemblies is
47
being recognized as a general mechanism for cells to form a distinct class of cellular
48
organelles with diverse functions (Banani et al., 2017; Chen et al., 2020; Lyon et al., 2021;
49
Shin et al., 2017; Wu et al., 2020). Compared to the classical cellular organelles that are
50
demarcated by lipid membranes, organelles formed via phase separation either do not
51
associate with or are not enclosed by lipid membranes and such organelles are referred to as
52
biological condensates or membraneless organelles in the literature (we use biological
53
condensates throughout this paper). Formation of biological condensates greatly expands the
54
means of how a living cell compartmentalizes its molecular constituents for specific and
55
diverse functions. Since biological condensates are not enclosed by membranes, molecules
56
within a biological condensates are in dynamic exchange with the counterparts in dilute
57
solution without energy input, thus establishing a foundation for numerous unique properties
58
of biological condensates (e.g. how sharp concentration gradient between the condensed and
59
dilute phases is maintained; how molecules are selected to be included or excluded in the
60
condensates; means and rates to regulate condensate formation/dispersion; etc.) with respect
61
to the membrane-enclosed organelles. The concept of biological condensate formation and
62
function has gained extensive interests in recent years, but the field is still in its infancy and
63
sometimes under debate (McSwiggen et al., 2019; Mittag and Pappu, 2022; Musacchio,
64
2022).
65
Molecules within biological condensates can be massively concentrated. For example,
66
proteins can be concentrated by more than 10,000 folds upon chromatin condensate
67
formation (Gibson et al., 2019). In cell peripherals such as synapses in neurons, phase
68
separation can concentrate numerous proteins into postsynaptic densities by >1,000-fold
69
(Zeng et al., 2018). A fundamental task in biological phase separation research is to
70
understand how molecules in the condensed phase behave and function. The existing
71
biochemistry and biophysics theories that have been guiding our understandings of molecular
72
behaviours and their interactions in living cells in the past are mainly developed for
73
molecules in dilute solutions. A biological condensate formed via phase separation is more of
74
a condensed soft matter system, thus theories dealing with dilute solutions are not expected to
75
be generally adequate for condensed molecular systems. Due to extreme complexities of
76
molecular constituents (i.e., proteins, nucleic acids and lipids), molecular compositions (i.e.,
77
each functional biological condensate often contains hundreds or more different types of
78
4
molecules), and broad range of interaction modes (e.g., very large dynamic ranges of binding
79
affinities and molecular valency, different levels of cooperativities, etc.) of biological
80
condensates in cells, currently available theories in soft matter physics and polymer
81
chemistry, though very useful, are likely not sufficient to be directly adapted to characterize
82
biological condensates. Experimental methods currently used to study molecules in
83
biological condensates are largely qualitative and descriptive.
84
Here we develop an adaptive super-resolution imaging-based method that can
85
simultaneously and robustly monitor and quantify motion properties of individual molecules
86
in dilute and condensed phases of biological condensates formed in solution or on lipid
87
membranes. In addition to directly visualizing motion trajectories within and between phases,
88
this method affords direct measurements of diffusion parameters of each molecule in the
89
dilute and condensed phase. Unexpectedly, we observed that molecules in the condensed
90
phase spend a very large fraction of time in transient motion-frozen state. Such temporary
91
motion freeze exists in various biological condensates and is governed by specific and
92
multivalent interaction-mediated large molecular network formation in condensed phases.
93
The motion property changes due to formation biological condensates can fundamentally
94
alter action mechanisms and cellular functions of biomolecules.
95
96
5
Results
97
Localization-based super-resolution imaging of phase separation
98
In an earlier study, we introduced localization-based single molecule tracking
99
experiment to study motion properties of proteins in the condensed phase of in vitro
100
reconstituted active zone condensates formed on two-dimensional supported lipid bilayer
101
(SLB) (Wu et al., 2019). Here, we further developed the method into an assay that can
102
simultaneously track molecules in both condensed and dilute phases. We used the in vitro
103
reconstituted postsynaptic density (PSD) condensates formed on SLB (Zeng et al., 2018) to
104
demonstrate this method. Four major PSD proteins (PSD95, SHANK3, GKAP, Homer) and
105
Trx-tagged GCN4-His8-NR2B-CT tetramer (termed as NR2B in this article) were included in
106
our study (Figure 1A). These five proteins, via specific and multivalent interactions, form a
107
large molecular network capable of phase separation at physiological concentrations (Zeng et
108
al., 2018).
109
Since the densities of proteins are hugely different between condensed and dilute
110
phases, sparse labelling would lead to lack of information for molecules in the dilute phase
111
and dense labelling would cause extensive overlapping of single molecule signals in the
112
condensed phase during conventional fluorescence imaging experiments. To overcome this
113
dilemma, we utilized dSTORM imaging (van de Linde et al., 2011) to obtain a large number
114
of stochastically emitted single molecule tracks in both condensed and dilute phases by
115
labelling proteins with photo-switchable dyes (in this case by labelling NR2B with 1%
116
Alexa647). TIRF illumination mode was used to detect protein signals on SLB so that signals
117
from molecules not tethered to the membrane were minimized.
118
The PSD mixtures formed noncircular condensed phase on SLB with around or less
119
than 1 μm in size, but conventional TIRF images could only provide fuzzy phase boundaries
120
at this scale (Figure 1B, top; also see (Zeng et al., 2018)). The same area was then first photo-
121
bleached by a high laser intensity and then imaged with a moderate laser intensity optimized
122
for the fluorophore lifetime lasting for 3,000 frames with an exposure time of 30ms per frame,
123
resulting in a high-resolution image containing ~100,000 individual localizations (Figure 1B,
124
bottom). The overall phase boundary did not undergo obvious change during the imaging
125
process as the shapes and boundaries of the condensed droplets in the system remained
126
essentially the same (Figure 1B).
127
6
Due to the stochastic nature of fluorophore switch on and off, the reconstructed super
128
resolution image could be treated as static molecular distributions of labeled molecules in
129
both dilute and condensed phases. Based on this super resolution image, we could define
130
those areas that have higher localization densities as the condensed phase regions, and the
131
rest as the dilute phase regions (Figure 1C). Accurate phase boundaries could be clearly
132
visualized for each condensed region. Comparing the average localization densities in the
133
condensed and dilute phases, we could estimate the partition coefficient of ~61 for NR2B
134
(i.e., NR2B was enriched into the condense phase by ~61 folds). The calculated NR2B
135
enrichment derived from the super-resolution imaging study was close to the value obtained
136
by a bulk fluorescence imaging-based method shown in our previous study (Zeng et al.,
137
2018). We noted with interest that the distribution of NR2B in the condensed phase are not
138
homogeneous (Figure 1C), indicating formation of nanodomain-like clusters within the
139
condensed phase.
140
141
Simultaneous single molecule tracking in different phases
142
The localizations obtained from dSTORM images contained information about
143
distributions as well as mobilities of molecules in both phases. However, the diffusion mode
144
and densities of molecules are very different in condensed and dilute phases. A striking
145
feature is that molecules tend to experience transiently confined state in the condensed phase
146
(Supplemental movie 1). We developed an adaptive single molecule tracking algorithm that
147
could automatically and robustly define optimal search ranges for molecules in different
148
phases, and the method could effectively minimize the global assignment errors in tracking
149
molecules in both condensed and dilute phases (Figure 2A and Figure 2—figure supplement
150
1-4; see “Methods” for extended description of the algorithm). Briefly, phase boundaries
151
were determined by densities of localizations at the beginning. A default search range (500
152
nm) was used to assign all localizations into tracks in both condensed and dilute phases.
153
Diffusion coefficients of NR2B in both dilute and condensed phases were estimated for
154
determining optimized search range for different phases. All localizations were reassigned
155
with the optimized search range to obtain final tracks of NR2B in both condensed and dilute
156
phases.
157
After adaptively assigning all localizations into tracks, we could obtain single
158
molecule tracks of molecules in both condensed and dilute phases (Figure 2B). With this
159
7
method, we could directly record events of molecules entering into and escaping from the
160
condensed phase as well as switch motions of molecules converting between confined state
161
and mobile state in the condensed phase (Figure 2C). The number of NR2B molecules
162
entering into and escaping from the condensed phase were equal (Figure 2D), a finding that is
163
consistent with the bulk equilibrium state of the PSD condensate. Interestingly, NR2B
164
molecules within the condensed phase spent a large proportion of time in the confined state
165
and molecules could switch between confined state and mobile state (Figure 2C&E). No
166
confined state could be detected when only NR2B was tethered to SLB (Figure 3—figure
167
supplement 1A). The above result indicated that NR2B in the condensed phase did not
168
undergo homogeneous diffusion motions as one might expect. We also imaged motions of
169
NR2B in the PSD condensates formed in 3D solution at the single molecular resolution and
170
found that, in the condensed phase, each NR2B molecule spent a large proportion of time in
171
the confined state (Figure 3—figure supplement 1B).
172
The histogram of NR2B displacement tracks in the condensed PSD phase has a
173
dominant peak with very small displacements, corresponding to the large proportion of time
174
of NR2B in the confined state. A small and relatively flat shoulder tailing the main peak
175
represents the small proportion of time of NR2B in the mobile state with larger displacements
176
(Figure 3A1). Fitting the histogram with a single population of NR2B undergoing Brownian
177
motions could only cover the confined state peak of molecule but not the high-displacement
178
tail of this distribution (Figure 3—figure supplement 1C1). In contrast, in the dilute phase,
179
the overall displacements of NR2B are much larger and broader with no prominent peak with
180
very small displacements (Figure 3A2). Even so, the displacement histogram of NR2B in the
181
dilute PSD phase cannot be described by Brownian motions of single NR2B population
182
(Figure 3—figure supplement 1C2), suggesting a presence of multiple populations of pre-
183
percolated NR2B/PSD protein complexes even in the dilute phase. In contrast, in the control
184
system with only NR2B tethered to SLB (i.e., no addition of any other PSD proteins), the
185
histogram of NR2B displacements can be nicely fitted by a simple diffusion model (Figure
186
3B).
187
We next used the Hidden Markov Model (HMM) to fit the motions of NR2B in the
188
condensed phase with a two-state diffusion model (Das et al., 2009; Persson et al., 2013),
189
motivated by the clear observation of relative immobile as well as mobile NR2B molecules in
190
the imaging experiments (Figure 2E). The parameters included the diffusion coefficients of
191
the molecule in transiently confined and mobile states (Dc and Dm) and the switching
192
8
probabilities between the two states (Pmc and Pcm). Maximum likelihood estimation was used
193
to estimate the parameters iteratively, and the parameters converged quickly after several
194
thousand iterations of optimization (Figure 3C&D). The diffusion coefficients in the mobile
195
state and in the confined state were Dm=0.17 μm2/s and Dc=0.013 μm2/s, respectively. The
196
switching probabilities were Pmc=82.8% (mobile to confined states) per frame and Pcm=3.8%
197
(confined to mobile states) per frame, respectively. The confinement ratio (Pc), defined as the
198
percentage of time that a molecule spends in the confined state, could be calculated by these
199
two switching probabilities as: Pc=Pmc/(Pmc+Pcm). For NR2B in the PSD condensates,
200
Pc=95.6%. The mobile ratio, defined as the percentage of time that a molecule spends in the
201
mobile state, could be calculated as: Pm=1-Pc=4.4%. The diffusion coefficient of the molecule
202
in each phase could be extracted by fitting the MSD (mean square displacement) values as a
203
function of time. The diffusion coefficient of NR2B in the dilute phase of the PSD
204
condensate is ~0.47 μm2/s, which is very close to that of NR2B alone tethered to SLB (~0.61
205
μm2/s) (Figure 3E). The apparent diffusion coefficient of NR2B in the condensed phase
206
derived by fitting MSD vs time is ~0.017 μm2/s (Figure 3E). This fitted diffusion constant
207
should be considered as an apparent diffusion constant (Da) as it contains information of the
208
confined states and the mobile states of the molecules in the condensed phase. When the
209
confinement ratio is large, the apparent diffusion coefficient will be dominated by the
210
confined state and significantly differ from the diffusion coefficient in the mobile state. The
211
simulation results of molecular diffusions of condensates formed on 2D SLB with different
212
motion switch conditions based on free diffusion with motion switch model are consistent
213
with the experimental trend (Figure 3F).
214
215
A diffusion model for equilibrium liquid-liquid phase separation
216
We next developed a simple diffusion model to describe a phase separation system
217
based on parameters measured using our adaptive single molecule tracking method. Consider
218
a small region near the phase boundary (Figure 3G). The dilute phase contains sparse and
219
fast-moving molecules. The condensed phase contains dense and slow-moving molecules
220
which can be further categorized into either in mobile state or transiently confined state. We
221
did not observe any obvious hinderance against motions when molecules cross the phase
222
boundaries, thus the energy barrier at the interface between the condensed and dilute phases
223
is likely negligible (Brangwynne et al., 2011; Feric et al., 2016). We assume that the number
224
9
of molecules crossing the boundary from one side to the other are proportional to the
225
diffusion coefficient (D) and the molecular density (σ). The influx of molecules from the
226
dilute phase to the condensed phase can be written as Jdc = kσdDd, where k is a constant, σd is
227
molecule density in dilute phase, and Dd is the diffusion coefficient in dilute phase. Since a
228
portion of molecules in the condensed phase are transiently confined, the efflux of molecules
229
from the condensed phase to the dilute phase is Jcd = kPmσcDm, where Pm is the mobile ratio
230
of molecules in the condensed phase, σc is molecule density in condensed phase, Dm is the
231
diffusion coefficient in mobile state. The net molecule flux should be zero at the equilibrium
232
state, thus we have Jdc = Jcd or σdDd = PmσcDm. The enrichment fold (EF) of molecules in the
233
condensed phase over the dilute phase is defined as σc/σd. Thus:
234
EF = σc/σd = Dd/ PmDm
[1]
235
or
236
𝐸𝑛𝑟𝑖𝑐ℎ𝑚𝑒𝑛𝑡 𝑓𝑜𝑙𝑑 �
𝐷𝑖𝑓𝑓𝑢𝑠𝑖𝑜𝑛 𝑐𝑜𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑡 𝑖𝑛 𝑑𝑖𝑙𝑢𝑡𝑒 𝑝ℎ𝑎𝑠𝑒
𝐷𝑖𝑓𝑓𝑢𝑠𝑖𝑜𝑛 𝑐𝑜𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑡 𝑖𝑛 𝑐𝑜𝑛𝑑𝑒𝑛𝑠𝑒𝑑 𝑝ℎ𝑎𝑠𝑒 𝑜𝑓 𝑚𝑜𝑏𝑖𝑙𝑒 𝑚𝑜𝑙𝑒𝑐𝑢𝑙𝑒𝑠 � 𝑀𝑜𝑏𝑖𝑙𝑒 𝑟𝑎𝑡𝑖𝑜
This equation connects the variables at the microscopic level (diffusion coefficient and
237
mobile ratio) to the macroscopic level observables (enrichment fold) with a very simple
238
relationship. If the molecules have multiple diffusion states, we can readily extend the model,
239
viz.,
240
∑ 𝜎𝑖
𝛼𝐷𝑖
𝛼
𝑚
𝑖
� ∑ 𝜎𝑗
𝛽𝐷𝑗
𝛽
𝑛
𝑗
[2]
241
where molecules in phase α and β contains, respectively, m and n types of diffusion states.
242
When there are multiple (two or more) coexisting phases, the sum of the products of
243
molecule density and diffusion coefficient in different states should be equal for all the
244
coexisting phases. The above analysis can be easily extended to phase separations in 3D
245
solutions by simply replacing the present 2D molecule densities with corresponding 3D
246
molecular concentrations.
247
Taking NR2B in the PSD condensates formed on SLB as an example, the measured
248
diffusion parameters are: Dd =0.47 μm2/s, Dm=0.17 μm2/s, and Pm=4.4%. The theoretical
249
enrichment fold of NR2B in the condensed phase based on eq. 1 is Dd/PmDm=62.8, a value
250
that is very close to value (~61) derived from the experimentally observed localizations
251
(Figure 1B). The model further provides mechanistic insights into some unique properties of
252
molecules in the condensed phase. For example, although concentrations of molecules in the
253
10
condensed phase are much higher than those in the dilute phase, it is remarkable that the
254
diffusion coefficient for the mobile fraction of molecules in the condensed phase is not
255
dramatically different from that in the dilute phase (0.17 vs 0.47 μm2/s). The fraction of time
256
that a molecule spends in the mobile state vs the confined state dramatically influences the
257
macroscopic properties of the molecule in the condensed phase. For instance, NR2B spends
258
over 95% of the time in the confined state in the PSD condensate. Accordingly, the apparent
259
diffusion constant of NR2B in the condensed phase is small at 0.017 μm2/s instead of 0.17
260
μm2/s measured for the mobile NR2B fractions in the condensed phase. Additionally, the
261
enrichment fold (or factor of enrichment) of a molecule into the condensed phase upon phase
262
separation is also dominantly reflected by the fraction of time molecules spent in the mobile
263
state vs that in the confined state. As we will demonstrate below, the binding affinity between
264
molecules in the phase separation system and the molecular network complexity in the
265
condensed phase determine the fraction of the time that a molecule spends in the mobile state
266
vs confined state as well as the enrichment of molecules in the condensed phase.
267
We further validated the diffusion model with simulations of molecular motions from
268
homogeneously mixed state towards a phase separated equilibrium state. A 2D simulation
269
box with size of 15x30 μm2 and periodical boundary conditions was prepared (Figure 3—
270
figure supplement 1D). Diffusion coefficients for the simulated molecule in the condensed
271
phase (for mobile state only) and in the dilute phase, mobile ratio, and mobile state lifetime
272
(derive from switching probability) of the molecule in the mobile and confined states in the
273
condensed phase were defined as simulation input (Table 1). The Monte Carlo method was
274
used to simulate a total of 50,000 molecules diffusing in the box for 100 seconds for each
275
condition. The enrichment fold and the ratio of molecules exchange between condensed and
276
dilute phases (exchange ratio) along the simulation trajectory were determined. Every
277
simulated system eventually reached equilibrium. The enrichment fold of the molecule at the
278
equilibrium state under each condition matched the theoretical value calculated by eq. 1
279
(Table 1).
280
Taken together, we now have a diffusion model for the equilibrium state of a phase
281
separation system. This model explicitly connects a set of measurable microscopic molecular
282
motion properties with the observable macroscopic parameters of molecules in the system.
283
The experimental method and the theoretical model developed above are simple and robust to
284
be implemented for analyzing biomolecular phase separations in general.
285
11
286
Dynamic molecular networks in condensed phases
287
Molecules or molecular complexes in dilute solution obey the diffusion law. In sharp
288
contrast, NR2B molecules in the PSD condensates formed on SLB and in 3D solution spend a
289
very large proportion of time in the immobile/confined state as if the PSD condensates can
290
form some sort of very large and thus immobile molecular network (a process known as
291
percolation in associative polymers including biopolymers; (Choi et al., 2020; Harmon et al.,
292
2017; Winnik and Yekta, 1997)) capable of trapping NR2B tansiently. Since molecular
293
processes such as molecular interactions, chemical reactions, etc. require molecules to be able
294
to collide with each other, phase separation-mediated immobilization of biomolecules in
295
condensed phases can have huge implications on numerous fundamental properties of these
296
molecules (e.g., binding kinetics, catalytic speed and specificity of enzymes, spatial
297
distributions in cellular sub-compartments, etc.).
298
We next asked how such network-like structure might form and what factors
299
determine the network stability in the condensed phase of a biological condensate. We
300
hypothesized that a phase separated system driven by multivalent and strong inter-molecular
301
interactions, such as the PSD condensates studied above, would form highly stable and larger
302
molecular network in the condensed phase. Accordingly, molecules bound to the network
303
might be considered as immobile or confined by the network. In contrast, molecular networks
304
in the condensed phase formed by weak but also multivalent molecular interactions (e.g.,
305
Intrinsic Disordered Region (IDR)-mediated phase separations) would be more dynamic and
306
the size of the network would also be smaller (see Figure 4A for a scheme). One might
307
envision that the molecular networks formed in the condensed phase could locally break or
308
reform, as molecules within the network could still undergo binding and unbinding processes.
309
Thus, the molecular networks formed in condensed phase are dynamic. The fraction of time
310
that a molecule stays on the network is directly proportional to the binding affinity (i.e., the
311
off-rate of the molecule from the network, a value directly related to the dissociation constant
312
of the binding) and avidity (the available binding sites in the vicinity of the molecule, a
313
parameter related to valency of the molecular interactions in the system) between the
314
molecule and the network.
315
To validate this hypothesis, we created a one-component phase separation system, a
316
chimeric protein composed of the prion like domain (PrLD) of FUS connected with the SAM
317
12
domain of Shank3 (PrLD-SAMWT, Figure 4B). The PrLD of FUS is a well-characterized
318
IDR protein capable of phase separation by itself (Kato et al., 2012). The SAM domain of
319
Shank3 can specifically interact with each other in a head-to-tail manner forming large
320
polymers (Baron et al., 2006). Thus, the PrLD-SAM chimera contains a weak and multivalent
321
interaction domain and a specific and multivalent interaction domain within one protein. The
322
PrLD-SAM chimera could undergo phase separation at very low concentrations (see below),
323
so we “caged” the chimera with highly soluble maltose binding protein (MBP) at its N-
324
terminus and a small, highly soluble protein GB1 at its C-terminus (Figure 4B). The resulting
325
“caged” protein could be purified and concentrated to as high as 500 µM. Cleavage of the
326
caging tags of caged PrLD-SAM with HRV-3C protease induced phase separation of the
327
protein. Substitution of Met1718 of the Shank3 SAM domain with Glu dramatically weakens
328
the head-to-tail interaction of the domain and the mutant SAM domain (SAMME) has very
329
weak propensity of forming oligomers in solution ((Baron et al., 2006); data not shown). We
330
also created a “caged” PrLD-SAMME chimera to investigate the impact of weakening
331
specific interaction on the molecular network formation in the condensed phase (Figure 4B).
332
Lastly, we used the FUS PrLD only to investigate the property of the molecular network in
333
the condensed phase that is solely formed by a weak and multivalent IDR sequence (Figure
334
4B). Again, we “caged” FUS PrLD at its both termini with GB1, so that the caged PrLD can
335
be purified in its native form and concentrated to very high concentrations without phase
336
separation. We predicted that the molecular networks in the condensed phase formed by
337
PrLD-SAMWT would be the largest and most stable followed by PrLD-SAMME, and the
338
molecular network of the PrLD condensed phase should be the most dynamic.
339
We compared the threshold concentrations of the three proteins for phase separation
340
to occur. Phase separation of each protein was induced by mixing 1 µM HRV-3C protease
341
and immediately injecting the digestion mixture into a sealed, home-made chamber incubated
342
at 20℃. All caged proteins were completely digested within ~30 min after addition of HRV-
343
3C protease. However, the phase separation of PrLD or the PrLD-SAMME chimera took up
344
to 12 hours to occur (i.e., with very slow nucleation rates). Thus, we compared DIC images
345
of the three cage-cleaved proteins captured at 12 hours after addition of HRV-3C protease
346
(Figure 4C). The PrLD-SAMWT chimera underwent phase separation at concentration as low
347
as 1 μM. In contrast the threshold phase separation concentrations for PrLD-SAMME and
348
PrLD were much higher (~150 μM and ~250 μM, respectively). These results demonstrate
349
that specific and multivalent interactions act in concert in biomolecular condensates and that
350
13
the latter can dramatically lower the threshold concentration of phase separation (Espinosa et
351
al., 2020; Lin et al., 2022; Riback et al., 2020; Zeng et al., 2018; Zeng et al., 2016).
352
We then compared the motion properties of the three proteins in the condensed phase
353
use the adaptive single molecule tracking method developed above. A concentration
354
somewhat higher than the phase separation threshold was used for each protein (i.e., 50, 200,
355
300 μM for PrLD-SAMWT, PrLD-SAMME, and PrLD; respectively). Each protein was very
356
sparsely labeled with Alexa 555 (0.02% for PrLD-SAMWT, 0.005% for PrLD-SAMME, and
357
0.005% for PrLD) to obtain sparse but long lifetime (>1s) single molecule tracks. It is clear
358
that, in the condensed phase, the motion properties of PrLD-SAMWT are dramatically
359
different from those of PrLD-SAMME and PrLD (Figure 4D). Each PrLD-SAMWT
360
molecule spent most of their time (74.0±3.7%) in the transiently confined state, but these
361
molecules were able to switch between the confined and mobile states (Figure 4E). In
362
contrast, the mobilities of PrLD-SAMME or PrLD in the condensed phase were much higher.
363
PrLD-SAMME spent 98.1±1.8% of time in the mobile state and PrLD spent 99.6±0.4% in the
364
mobile state (Figure 4E). These findings indicated that PrLD-SAMWT molecules in the
365
condensed phase formed a very large but still dynamic molecular network due to the presence
366
of specific and multivalent SAM-SAM interaction. In contrast, the molecular networks in the
367
condensed phase formed solely by weak and multivalent interactions were much smaller and
368
more dynamic. Interestingly, the diffusion coefficients of PrLD-SAMWT and PrLD-SAMME
369
in their mobile state were very similar (0.18±0.05 μm2/s vs 0.19±0.02 μm2/s) (Figure 4E),
370
indicating that PrLD-SAMWT and PrLD-SAMME have similar molecular size in their
371
mobile state. The mobile state of both proteins likely corresponds to each molecule not bound
372
to the large molecular network in the condensed phase. The diffusion coefficient of the
373
mobile state of PrLD is 0.36±0.04 μm2/s (Figure 4E) and the molecular weight of PrLD is
374
about half of PrLD-SAMWT and PrLD-SAMME, again suggesting that the mobile state of
375
PrLD likely corresponds to the network unbound and monomeric form of PrLD. Taken
376
together, the above single molecule tracking study revealed that strong multivalent
377
interactions could lead to formation of large and more stable molecular networks in the
378
condensed phase, which could dramatically reduce the overall motions of the molecules in
379
the condensed phase. Most prominently, proteins that bind to large and stable molecular
380
networks no longer obey the free diffusion law found in dilute solutions. Instead, in such
381
condensed phase, proteins switch between immobile/confined state and free diffusion state,
382
corresponding respectively to the network-bound and free forms. In the condensed phase
383
14
formed by weakly interacting IDR sequences, the molecular network in the condensed phase
384
is very dynamic and much smaller resulting in IDR proteins displaying simple diffusion-like
385
behavior.
386
387
Fluorescence recovery after photobleaching (FRAP) in phase separation systems
388
FRAP assays are widely used to examine dynamic properties of phase separation
389
systems. Quantitative theory for analyzing FRAP results in phase separation systems based
390
on Flory-Huggins theory have also been developed for weak interaction systems (Hubatsch et
391
al., 2021). For heterogeneous condensed phase systems, traditional FRAP experiment might
392
not be a good way to test the liquid-like properties (McSwiggen et al., 2019). As we have
393
shown above, motion properties of molecules in the condensed phase are radically different
394
for the systems involving specific multivalent interactions compared to the system with only
395
weak interactions. Seeking a better understanding, we simulated FRAP properties of several
396
different phase separation systems pertinent to our experiments. A 2-dimensional phase
397
separation system mimicking phase separation on a flat surface such as lipid membranes was
398
constructed with three different sizes of round-shaped condensed phase (radius of 0.5, 1, and
399
2 μm) in a box with periodic boundary (Figure 5A). By monitoring the molecule positions for
400
100 seconds after photobleaching, we can simulate the FRAP curve of any region in the box.
401
We define a region with certain size to be bleached as region of interest (ROI).
402
We first asked whether the size of a ROI vs the size of a condensed phase may affect
403
FRAP results. For this simulation, the diffusion coefficients of the molecule were set at 0.01
404
μm2/s and 1 μm2/s for condensed and dilute phases, respectively. The mobile ratio was set at
405
100% (i.e., simulating FRAP curves for a phase separation system dominated by weak
406
multivalent interactions). This setting leads to 100 times enrichment of the molecule into the
407
condensed phase. The ROIs with a radius of 0.5 μm were selected at the center of the three
408
different sized droplets (ROIs 1~3, Figure 5A). The simulated results showed that the
409
apparent recovery curve for ROI3 was considerably slower than the first two ROIs (Figure
410
5B), as all molecules in ROI3 needed to diffuse from the dilute phase into the entirely
411
bleached condensed phase. This simulation indicates that one should select condensed phases
412
with similar sizes when comparing FRAP curves of related phase separation systems. When
413
possible, always select a small ROI within a large droplet for FRAP analysis.
414
15
We next simulated FRAP curves for the phase separation systems containing specific
415
interactions. We first set the mobile ratio of simulated molecules in condensed phase at 10%,
416
with average lifetime of the confined state being 10 seconds. To maintain the same apparent
417
diffusion coefficient and the same enrichment level in the condensed phase as those in the
418
system with no molecular confinement described above, the diffusion coefficient of the
419
molecule in the mobile state within the condensed phase was set as 0.1 μm2/s. We simulated
420
the FRAP curves for the same three ROIs as indicated in figure 5A. Although the FRAP
421
recovery speed for ROI3 was still slower than the other two ROIs, the difference became
422
smaller (Figure 5C). Further increasing the lifetime of the confined state to 100 seconds while
423
keeping other parameters unchanged led to linear-like slow recovery curves and the
424
differences among the three ROIs were further diminished (Figure 5D). In an extreme case
425
where the lifetime of the confined molecules were infinitely long (i.e., extremely strong
426
binding and resulting in behaviors similar to those of unrecoverable aggregates), all three
427
recover curves looked similar with a fast recovery speed to their maximal recovery level at 10%
428
(Figure 5E). These simulations suggest that the property of the molecular network in the
429
condensed phase can have dramatic impact on its FRAP recovery rate in the system. For
430
example, a slow recovery speed from a FRAP experiment does not necessarily mean that the
431
molecule in condensed phase moves slowly. Instead, it may mean that the molecule spends
432
most of the time bound to the molecular network. Once switched into mobile state, it can
433
diffuse quite freely and rapidly.
434
We validated our simulations by experimentally measuring FRAP curves of the three
435
phase separation systems shown in Figure 4B. A small ROI with identical radius within a
436
relatively large condensed droplet was selected for the FRAP experiments in each of the three
437
systems (Figure 5—figure supplement 1). The PrLD system showed the fastest recovery
438
speed, and the PrLD-SAMME system showed a slightly slower recover speed. The PrLD-
439
SAMWT system, with its large and stable molecular network in the condensed phase,
440
displayed a very slow and near linear recovery curve (Figure 5F). We further simulated the
441
FRAP curves of the three systems using the diffusion coefficients and confinement ratios
442
derived from the single molecule tracking data (Figure 4E). The simulated FRAP curves were
443
overall very similar to those obtained experimentally (Figure 5G vs 5F). The faster recovery
444
rate in the simulation curve vs the experimental curve for the PrLD-SAMWT system is likely
445
due to the overestimations of the mobile ratio of the protein in the tracking experiment. Taken
446
together, the above theoretical and experimental studies revealed that the FRAP curve of a
447
16
molecule in a phase separation system is heavily influenced by the proportion of time of the
448
molecule transiently trapped in the confined state, a parameter that is directly linked to the
449
specific multivalent interactions of the system. A low recovery rate in the FRAP assay does
450
not necessary mean that a large fraction of molecules in the condensed phase is permanently
451
immobile.
452
453
Discussion
454
In this study, we developed a method that can track single-molecule motions in both
455
condensed phase and dilute phase simultaneously by using photo-switchable dye labelled
456
proteins. To accommodate the heterogeneity of both the distributions and diffusion modes of
457
molecules in the condensed and dilute phase, an adaptive single-molecule tracking algorithm
458
was developed by setting the optimized search range for molecules with different diffusion
459
coefficients. The method is simple and highly robust. It can be deployed to track single-
460
molecule motions of phase separation systems with very broad dynamic ranges including
461
highly dynamic systems formed by IDR proteins or very stable (or highly percolated) systems
462
formed by strong and specific multivalent molecular assemblies such as PSDs. With
463
implementations of sparse labeling techniques such as HaloTag, our method can be applied to
464
track single-molecule motions of biological condensates in living cells.
465
The most important finding from our study is that molecules in the condensed phase
466
do not exhibit simple diffusion behaviors as those observed in dilute solutions. Instead,
467
molecules constantly switch between transient confined state and mobile state in the
468
condensed phase, a phenomenon that is most likely underpinned by phase separation-
469
mediated formation of large molecular networks. The size and dynamic properties of the
470
molecular network in the condensed phase are determined by the binding affinity (or
471
affinities) and valence of the interaction(s) of the molecule(s) in the phase separation system,
472
a finding that has been recently predicted by theoretical simulations treating biomolecules are
473
associative polymers (Choi et al., 2020; Harmon et al., 2017). The fraction of time and the
474
time duration that a molecule spends at the confined state is also determined by the binding
475
affinity of the molecule to and the dynamic property of the network. Surprisingly, the
476
diffusion coefficients of mobile-state molecules in the condensed phase are only slightly
477
lower than the counterpart molecules in the dilute phase. For biological condensates of
478
which their formations are largely driven by specific molecular interactions (it is our opinion
479
17
that most cellular condensates belong to such category; (Chen et al., 2020; Feng et al., 2021)),
480
molecules in the condensed phase spend most of their time in the confined state. Since the
481
fraction of time and the time duration that a molecule spends in the confined state vs those in
482
the mobile state are basic defining parameters for the functions of molecules in any reaction
483
systems (e.g., binding/unbinding rates, kinetics of enzyme catalysis, lifetime/dwelling time a
484
molecule in a molecular machinery, etc.), our finding reveals a fundamental aspect of
485
molecular properties created by biological condensates that is distinctly different from that in
486
dilute solutions. Our study implies further that, in theory, unlimited types of biological
487
condensates with very broad dynamic network properties may form using the existing
488
repertoires of proteins and nucleic acids via different combinations of binding affinities and
489
interaction valences. Thus, phase separation-mediated formation of biological condensates is
490
a very powerful means for cells to form numerous subcellular organelles with a continuum
491
dynamic and material properties ranging from very dynamic, dilute solution-like assemblies
492
to highly stable, solid-like systems. Confinements of molecules in cellular condensates have
493
been observed by single-molecule tracking experiments recently (Chong et al., 2022; Miné-
494
Hattab et al., 2021; Niewidok et al., 2018). However, since the molecular compositions of
495
cellular condensates cannot be easily defined, the mechanistic bases underlying the confined
496
state of molecules in these cellular condensates are difficult to be discerned. Single molecule
497
tracking experiments using the reconstituted and compositionally defined phase separation
498
systems in the current study allowed delineation of the mechanism underlying the unique
499
motion properties of molecules in the condensed state.
500
During our adaptive single molecule imaging process, we did not observe obvious
501
motion speed slow down or enhancement when molecules enter or leaving a condensed phase
502
in all condensate systems investigated in this work. Therefore, the tension (or depth of the
503
energy well) at the interface between the condensed phase and the dilute phase in each
504
condensate is not large enough to significantly impact molecular exchanges between the two
505
phases.
506
Since the interactions between molecules can be modulated via numerous cellular
507
processes such as posttranslational modifications, protein biogenesis/turnovers, epigenetic
508
modification, cellular milieu alterations, etc., the dynamic network properties and
509
consequently the functions of organelles formed via phase separation may be regulated in
510
ways that are distinctly different from those occurring in dilute solutions. Compared to the
511
rich knowledge to and quantitative theories for the dilute-solution systems, few satisfying
512
18
theoretical frameworks have been established for the condensed assemblies formed via phase
513
separation in cells (see (Mittag and Pappu, 2022) and refs therein), partly due to our poor
514
understandings of microscopic motion properties of molecules in the condensed phase. The
515
dramatic dynamic and material property differences of condensates formed by weakly
516
associative IDR proteins and by biomolecules with specific interactions indicate that
517
biological phase separation research has only touched the tip of the iceberg, given that the
518
vast majority of research only deals with IDR proteins.
519
520
521
Materials and Methods
522
Protein expression and purification
523
Constructs for expression of Trx-His-GCN4-NR2B, PSD-95 (UniProt: P78352-1),
524
Shank3 (UniProt: Q4ACU6), GKAP (UniProt: Q4ACU6-1), Homer3 (UniProt: Q9NSC5-1)
525
were described previously (Zeng et al., 2018). MBP-His8-GCN4-NR2B tetramer was created
526
by inserting GCN4-NR2B sequence into an in-house modified pET32a vector. MBP-His8-
527
GCN4-NR2B trimer and dimer were mutated from the tetramer version by changing the
528
hydrophobic residues in the GCN4 domain (Delano and Brünger, 1994). Constructs of GB1-
529
PrLD-GB1 contained FUS-PrLD (UniPort: P56959, segment: 1-212) with the protecting GB1
530
protein fused to the N- and C-terminal ends. MBP-PrLD-SAMWT-GB1 is a fusion protein
531
with the SAM domain of Shank3 (aa 1654-1730) fused to the C-terminal end of PrLD, and
532
the resulting chimeric protein was further protected by tagging its N-terminus with MBP and
533
C-terminus with GB1. MBP-PrLD-SAMME-GB1 is the same as MBP-PrLD-SAMWT
534
except that Met1718 in the SAM domain was replaced by Glu. An additional cysteine was
535
inserted at the N-terminus of PrLD for cysteine labeling. All constructs were confirmed by
536
DNA sequencing. Recombinant proteins were expressed in Escherichia coli BL21 (DE3)
537
cells in LB medium at 16 °C. Protein expressions were induced by adding 0.25 mM IPTG
538
when OD600 reached 0.6-0.8. His8-tag containing recombinant proteins were purified using
539
Ni2+-NTA agarose affinity column followed by size exclusion chromatography (Superdex
540
200 26/60 column from GE healthcare) in a final buffer containing 100 mM NaCl, 50 mM
541
Tris-HCl (pH 7.8), 1 mM DTT and 1 mM EDTA. The purified proteins (except of the NR2B
542
proteins for lipid binding and PrLD/PrLD-SAMME/PrLD-SAMWT) were then subject to tag
543
removal by HRV-3C or TEV protease at 4 °C overnight followed by another round of size
544
19
exclusion chromatography. All purified proteins were checked for free of nucleic acid
545
contamination.
546
547
Protein fluorescence labeling
548
His8-tagged NR2B proteins were labelled with Alexa 647 NHS ester (Thermo Fisher)
549
and PrLD-SAMWT/PrLD-SAMME/PrLD proteins labelled with Alexa 555 maleimide
550
(Thermo Fisher). Alexa 647 NHS ester was first dissolved in DMSO at a concentration of 10
551
mg/mL. Before labeling, all purified proteins were exchanged to a Tris-free buffer containing
552
100 mM NaHCO3 (pH 8.4), 100 mM NaCl, and 1 mM EDTA (plus 1 mM DTT for NR2B)
553
using a HiTrap desalting column. NR2B was concentrated to 20-50 μM and mixed with the
554
corresponding dye at a 1:1 molar ratio. Alexa 555 maleimide was dissolved in DMSO at a
555
concentration of 10 mg/mL and mixed with PrLD-SAMWT/PrLD-SAMME/PrLD (>100 μM,
556
without DTT) at a 1:1 molar ratio. The mixture was incubated at room temperature for about
557
1 hour and the reaction was terminated by adding 200 mM Tris-HCl (pH 8.2). The mixture
558
was next loaded to a HiTrap desalting column to separate the unreacted fluorophores and to
559
exchange proteins to buffers for following experiments. Efficiency of individual labelling
560
was measured by Nanodrop 2000 (ThermoFisher). Unlabeled protein was mixed with each
561
labelled protein to adjust the final labelling ratio needed for imaging experiments.
562
563
Fast protein liquid chromatography coupled with static light scattering (FPLC-SLS)
564
assay
565
The analysis was performed on an AKTA FPLC system (GE Healthcare) coupled
566
with a static light scattering detector (miniDawn,Wyatt) and a differential refractive index
567
detector (Optilab, Wyatt). Protein samples (concentrations for each reaction were indicated in
568
the figure legends) were filtered and loaded into a Superose 12 10/300 GL column pre-
569
equilibrated by a column buffer composed of 50 mM Tris, pH 8.2, 100 mM NaCl, 1 mM
570
EDTA, and 2 mM DTT. Data were analyzed with ASTRA6 (Wyatt).
571
572
Lipid preparation
573
20
POPC (Avanti lipids, Cat No:850457P), DGS-NTA-Ni2+ (Avanti lipids, Cat
574
No:790404P) and PEG-5000 PE (Avanti lipids, Cat No:880230P) were first solubilized in
575
chloroform to a stock concentration of 20 mg/mL, 10 mg/mL and 1 mg/mL, respectively.
576
Lipid mixture containing 98% POPC, 2% DGS-NTA-Ni2+ and 0.1% PEG-5000 PE was dried
577
under a stream of nitrogen gas followed by a vacuum pumping to evaporate chloroform
578
thoroughly. The dried lipids were then resuspended in PBS to a final concentration of 0.5
579
mg/mL. Multi-lamellar vesicle solution was next solubilized by 1% w/v sodium cholate and
580
loaded onto a desalting column. During the desalting process, sodium cholate will be diluted
581
allowing small uni-lamellar vesicles (SUVs) to form in the buffer containing 100 mM NaCl,
582
50 mM Tris-HCl (pH 7.8), and 1 mM TCEP (the 2D buffer).
583
584
Coating chambered cover glass with lipids
585
Chambered cover glass (Lab-tek) was washed with Hellmanex II (Hëlma Analytics)
586
overnight, thoroughly rinsed with MilliQ H2O. The chambered cover glass was then washed
587
with 5 M NaOH for 1 hr at 50 ℃ and then thoroughly rinsed with MilliQ H2O. The cleaned
588
coverslips were washed three times with the coating buffer (50 mM Tris, pH 8.2, 100 mM
589
NaCl, 1 mM TCEP). Typically, 150 μL SUVs were added to a cleaned chamber and
590
incubated for 1 hr at 42℃, the SUVs would fully collapse on glass and fuse to form
591
supported lipid bilayers (SLBs). Chambers with SLBs were then gently washed three times
592
each with 750 µl of coating buffer to remove extra SUVs before being blocked by the
593
clustering buffer (coating buffer plus 1 mg/ml of BSA) for 30 mins at room temperature.
594
595
Phase separation on SLB
596
The supported lipid bilayers contained 2% DGS lipid with Ni2+-NTA attached to its
597
head. We used GCN4-NR2B with an N-terminal thioredoxin (TRX)-His8 tag (referred to as
598
NR2B tetramer) to attach to SLBs via binding to DGS-NTA-Ni2+. The NR2B (4µM final
599
concentration) tetramer was added to a SLB-containing chamber. After 30 mins incubation at
600
room temperature, the chamber was washed with the clustering buffer for three times (each
601
time at 750 µl volume) to remove excessive NR2B tetramers. PSD-95, Shank3, GKAP and
602
Homer3 (each at 2 µM final concentration) were sequentially added into the system. Imaging
603
acquisition started at 15 mins after adding all components.
604
21
605
Phase separation in 3-dimensional solution in chamber
606
For PSD condensates, 10 µM of 5 PSD proteins were mixed and injected into a homemade
607
chamber and sealed immediately (Zeng et al., 2016). The mixtures were incubated for 15
608
mins before starting image acquisitions. For the PrLD/PrLD-SAMME/PrLD-SAMWT
609
systems, 300/200/50 µM of “caging” tag-containing protein was mixed with 1 µM of HRV-
610
3C protease, each mixture was injected into a homemade chamber and sealed immediately.
611
The samples were incubated at 20 ℃ for 12 hrs before image acquisitions.
612
613
dSTORM imaging
614
Freshly prepared imaging buffer (the 2D buffer plus 1% D-Glucose (Sigma G8270),
615
5.6 μg/mL glucose oxidase (Sigma G2133-50KU, from 100 x stock prepared in the coating
616
buffer), 40 μg/mL catalase (Sigma C9322-10G, from 100 x stock prepared in the coating
617
buffer) and 15 mM β-mercaptoethanol) was injected into an imaging glass chamber to replace
618
the original coating buffer. Imaging of each sample was completed within 30 min upon
619
addition of the imaging buffer.
620
dSTORM images for the condensates formed on SLBs were taken by a home-built
621
two-color super-resoution localization microscope based on a Nikon Ti-E inverted
622
microscope body (Zhao et al., 2015). Here only one channel was used to image samples
623
labelled with Alexa 647. A 100x objective lens (CFI Apo TIRFM 100x Oil, N.A. 1.49, Nikon)
624
was used to observe the fluorescence signals. An EMCCD (electron-multiplying charge-
625
coupled device, Andor, IXon-Ultra) was applied to collect the emission lights that passed
626
through a channel splitter. For each sample, 2000 frames of images with an exposure time of
627
30 ms/frame were captured from at least 6 different areas. The laser intensity was fixed at 1
628
kW/cm2 during the imaging and the microscope was at the TIRF mode. If single molecule
629
signal density of a sample was too high, a pre-image photobleaching with a strong laser
630
intensity (4.0 kW/cm2) was used to reduce the single molecule signal density. The TIRF raw
631
images were processed by Rohdea (Nanobioimaging Ltd., Hong Kong) to generate the
632
localization coordinates in each frame.
633
dSTORM images for 3D phase separation system were taken by a Zeiss Elyra7
634
microscope with a 63x oil objective lens. Samples were first bleached with a full power laser
635
22
(500 mW) and then imaged with 20% of the full power of the 488/561/641 nm lasers with the
636
HILO mode illumination. A TIRF-hp filter was used during imaging. For each sample, 4000
637
images were captured an exposure time of 30 ms/frame. Autofocus with the “definite focus”
638
strategy was performed at every 500 frames. Maximum point spread function size was set at
639
9 and signal-to-noise ratio was set at 5 when capturing single molecules with Zeiss Elyra7.
640
Samples were labeled with 0.005%~0.1% ratio of dyes depending on the signal density.
641
642
Adaptive single molecule tracking algorithm
643
The heterogeneity of molecule distributions and diffusion modes in the dilute and
644
condensed phases of liquid-liquid phase separation systems requires an adaptive single
645
molecule tracking algorithm to minimize the track assign error locally and globally.
646
Traditional single molecule tracking in high density systems (Jaqaman et al., 2008; Manley et
647
al., 2008; Tinevez et al., 2017) usually set a global search range (step limit) manually to
648
connect molecular tracks. A step limit that is too small will cause lots of missing connection
649
for molecules with fast diffusions; whereas a step limit that is too large will cause lots of false
650
positive connections for molecules with slow diffusions. Such errors cannot be fixed in post-
651
tracking data analysis. A biological condensate system typically contains a condensed phase
652
with slow diffusing molecules coexisting with a dilute phase with fast diffusing molecules,
653
and molecules in the condensed phase constantly switch between mobile state and confined
654
state. A single global maximum step limit without any prior knowledge might not suit for
655
tracking molecules in a phase separation system. A typical solution for motion switch is to
656
use the Hidden Markov Model (HMM) to fit a diffusion model that contains diffusion
657
coefficients (D) and switching probabilities (P) for different diffusion states (S) (Das et al.,
658
2009; Persson et al., 2013). Taking all above factors into consideration, we developed a new
659
algorithm by adaptively choosing maximum step limit for different diffusion states to link the
660
localizations into tracks and using HMM to fit a two-state diffusion model in the condensed
661
phase.
662
The track assignment errors can be divided into two parts, true negatives and false
663
positives. A true negative error is defined as an existing track was not linked, which leads to
664
miss of long-distance steps. This part of the error can be estimated by the Boltzmann
665
distribution if we assume that molecules undergo Brownian motion in the mobile state,
666
𝐸�� � 𝑒���
���
�
. A false positive error is defined as linking of a non-existing track. This part
667
23
of the error can be estimated with collision frequency of particles that have a diameter same
668
as the search range, 𝐸�� � √𝜋𝐷𝑡 � 𝜎𝑅. Thus, the estimated assignment error under a certain
669
search range R is:
670
𝐸 � 𝐸�� � 𝐸�� � 𝑒���
���
�
� √𝜋𝐷𝑡 � 𝜎𝑅
[3]
671
To find the minimized error of the search range R under certain diffusion coefficient D and
672
molecule density 𝜎, we just needed to find the point that
��
�� � 0 𝑎𝑛𝑑
���
��� � 0. Noted that the
673
root mean square displacement (RMSD) was √4𝐷𝑡, we replace 𝑋 �
�
���� �
�
√��� as the ratio
674
of search range to RMSD of the molecules. The first derivative
��
�� � 0 could be transformed
675
into 𝑋𝑒��� � √𝜋𝜎𝐷𝑡. The fluorophore density in the condensed phase and the dilute phase
676
were typically 0.20~0.40 /μm2 and 0.01~0.02 /μm2, respectively; diffusion coefficients were
677
0.02~0.2 μm2/s and 0.5~2 μm2/s, respectively. The solution of X under such conditions was
678
within a very small region of 2.5~3 (Figure 2—figure supplement 1A). We could roughly
679
estimate the diffusion coefficients by displacement distributions (Figure 3A&B) (Hansen et
680
al., 2018) starting with a default search range (500 nm). We then moved on to find an
681
optimized search range and use this optimized search range to complete the final track
682
assignments (illustrated in Figure 2A).
683
We next validated this optimized R (search range) with simulated molecular systems
684
undergoing homogeneous Brownian motions with different diffusion coefficients and
685
densities. For systems that are similar to our 2D PSD system showed the same converged
686
optimal X at ~2.5 (Figure 2—figure supplement 1B~E). We simulated many other conditions
687
with different molecular densities (N) and moving speeds (RMSD) in both dilute and
688
condensed phases, and all showed a similar optimal X of ~2.5 (Figure 1—figure supplement
689
2&3). For systems with molecules undergoing switching between confined state and mobile
690
state in the condensed phase (defined as fraction of mobile state or mobile ratio, Pm) and
691
with different lifetime in the mobile state tm, the optimal X value also converged to ~2.5
692
(Figure 1—figure supplement 4). These simulation results indicated that the optimal X value
693
was very similar under different conditions. Thus, we used a default X=2.5 to determine the
694
optimized R (search range) for different experiments.
695
To valid the algorithm experimentally, we prepared His6-tagged MBP fused with
696
GCN4 dimer, trimer, or tetramer, respectively. The three fusion proteins have a molecular
697
24
weights ratio of 2:3:4 measured by light scattering experiment (Figure 1—figure supplement
698
1F). We measured the diffusion coefficients of the three MBP proteins coated onto SLBs
699
with our adaptive single molecule tracking algorithm without any pre-set parameters (Figure
700
1—figure supplement 1G). The measured diffusion coefficients for the MBP-GCN4 dimer,
701
trimer, and tetramer are very close to 1/2:1/3:1/4, which is the expected theoretical value for
702
the three proteins on SLB. Thus, this experimental data showed that our developed algorithm
703
is robust in adaptively determine the diffusion coefficients without any prior knowledge.
704
705
Generating simulated homogenous and heterogenous molecular systems
706
Molecules distributed homogeneously both in condensed and dilute phases were
707
generated based on the Monte Carlo method to simulate localizations obtained in single
708
molecule tracking experiments. Diffusion coefficients D, molecule densities N were set for
709
different scenarios. Averaged lifetime was set to 3.5 frames based on our experimental
710
average track length and with a Poisson distribution. Number of total tracks were calculated
711
before simulation and every track started with a random frame following uniform distribution.
712
For one single track with diffusion coefficient D, the displacement in a short time interval t
713
(0.0001 s) will be:
714
𝑑𝑥 � 𝑑𝑦 � 𝑟𝑎𝑛𝑑𝑛 ∗ √2𝐷𝑡
[4]
715
where dx and dy are the displacements along horizontal and vertical axes, randn is a random
716
number with the standard normalized distribution. All data were saved into two versions, one
717
formatted including track information as ground truth and the other formatted by frames and
718
localizations in each frame without track information as simulated data for evaluate tracking
719
algorithm.
720
Molecules distributed heterogeneously in condensed phases were generated with
721
similar process with an additional set of switching probabilities between mobile and confined
722
states. Molecules in a confined state will be restricted to a certain position, thus the simulated
723
localizations will follow a Gaussian distribution based on the point spread function. A
724
confined molecule can switch to mobile state in next frame with a probability Pcm, and a
725
mobile molecule can switch to a confined state in next frame with a probability Pmc. Mobile
726
ratio (Pm) can be calculated from the switching probability Pm, which is defined as Pm= Pcm
727
25
/( Pcm + Pmc). The lifetime of the mobile state can be expressed as tm=1/Pmc. The data were
728
saved into two versions same as that described for the homogenous system.
729
Each scenario was simulated for 500-10,000 frames dependent on the molecule
730
density, which led to a similar total track number for each system. The script for the
731
simulations was in-house coded by MATLAB.
732
733
Evaluation of simulated data
734
For each condition, simulated data were fed to the algorithm to assign localizations
735
into tracks with different maximum step limit. According to equation 3, the ratio of maximum
736
step limit to root mean square displacement (R/RMSD, defined as X) is the key variable, so
737
we covered X from 1 to 5 with a step size of 0.5. Track assignment error composed of “True
738
Negative tracks” (TN, two localizations belong to a same track in ground truth but not
739
recognized as the same track) and “False Positive tracks” (FP, two localizations do not
740
belong to a same track in ground truth but recognized as a same track) and calculated as the
741
ratio of the absolute value of (the algorithm output - ground truth)/ground truth. Diffusion
742
coefficient error was calculated by fitting mean square displacement (for the homogenous
743
system) or fitting a two-state model (for the heterogenous system) and compared with the
744
original setting for each condition. The scripts for the evaluations were coded by MATLAB.
745
746
Brownian motion with motion switch model simulation
747
Consider a particle undergoing Brownian motion on a 2-dimensional surfacewith the
748
additional feature that it can switch between confined and mobile states at each time step with
749
certain switching probabilities. Specifically, during each time step, if the particle is in a
750
confined state, it will have a probability of Pcc=0.9 to remain in the confined state and
751
therefore a chance of Pcm=0.1 to switch to the mobile state. If the particle is in the mobile
752
state, it moves in a random direction with a random displacement (i.e., following the standard
753
normal distribution) and also have a chance of Pmc = Pcm*Pc/Pm to switch to confined state.
754
To maintain a steady-state confined-mobile balance, the total switch events from mobile state
755
to confined state and vice versa are identical, i.e., Pm*Pmc = Pc*Pcm. Since Pc+Pm=1, the
756
chance of switching from mobile state to confined state is Pmc = Pcm*(1-Pm)/Pm. n=500
757
26
particles and simulation time length T=100 sec are used in each simulation. Mobile ratio Pm
758
was studied from 0 to 0.9 with a step size of 0.1.
759
760
Equilibrium state phase simulation
761
The phase boundaries were constructed in accordance with experiments (Figure 1B).
762
The boundaries were not change during the simulation. An area of 15x30 μm2 with periodic
763
boundary conditions was used for the simulation (Figure 3—figure supplement 1D). At the
764
beginning of the simulation, a large number of molecules (n=50,000) were randomly
765
distributed in the condensed and dilute phases with initial enrichment fold of 60. All
766
molecules in dilute phase were treated as mobile during the simulation, the diffusion
767
coefficient was set at 0.6 μm2/s. The motions of molecules in condensed phase consisted of
768
those in the confined state and the mobile state. Molecules in the confined state were treated
769
as immobile with fixed positions. Molecules in the mobile state were undergoing Brownian
770
motion and the diffusion coefficient was set at 0.1 μm2/s. The switch between confined state
771
and mobile state was determined by the switching probability Pcm and Pmc. Pmc could be
772
directly calculated through the averaged dwell time of mobile state (0.1 second) as the
773
lifetime of the fluorophore (~1 second) was much longer than the molecule’s dwell time. Pcm
774
was calculated using the equilibrium condition between the mobile to immobile state by:
775
𝑃�� � 𝜂 � 𝑃�� � �1 � 𝜂�
776
where η is the mobile ratio (10%). The simulation time step was t 0.0001 second with the
777
total simulation time being 100 seconds. The script for the simulation was in-house coded by
778
MATLAB.
779
780
Fluorescence recovery after photobleaching (FRAP) assay
781
Proteins were labelled with Alexa Flour 555 (Thermo Fisher) at 1% for PrLD-
782
SAMME and PrLD-SAMWT) or 0.5% (for PrLD). FRAP assays were performed on a Zeiss
783
LSM 880 confocal microscope at room temperature. A region for bleaching (R1) with a
784
diameter of 2 μm was selected within a large, condensed droplet. A reference region (R2)
785
with the same size of R1 was selected in another large, condensed droplet as the system
786
control. R1 was bleached with 40/30/10 iterations with 100% 561 nm laser power and
787
27
followed by recording fluorescence intensity of the selected regions for 100 seconds in the
788
time-lapse mode with a 10-second gap between each point for the PrLD-SAMWT system and
789
5 seconds for PrLD-SAMME and PrLD systems. The fluorescence intensities were
790
normalized to 0% right after photobleaching and to 100% before photobleaching. Each data
791
point was calibrated by recorded fluctuation of the intensity of the reference region R2.
792
793
FRAP simulation
794
FRAP simulations were based on the simulation for the equilibrium state phase
795
separation. A 20x8 μm2 box with periodic boundary containing seven spherical condensed
796
droplets was used (Figure 5A). The radius and centre coordinates of condensed droplets were
797
0.5/1/1/2/1/1/0.5 μm and 1/3/5/10/15/17/19 μm from the left edge of the box, respectively.
798
Conditions 1/2/3 (see Figure 5A) had the same bleaching size with a diameter of 1 μm and
799
centered in the large/median/small droplets. The enrichment fold was set at 100 for all
800
simulations, the mobile ratio was set at 100% or 10%, and the confined state lifetime was set
801
at 10 or 100 seconds. For totally immobilized molecules, the lifetime of the confined state
802
was infinite and the switching probability between mobile and immobile was zero. All
803
simulations were carried out for 100 seconds with a time step of 0.0001 second. A total of
804
50,000 molecules were used for each simulation.
805
806
807
Acknowledgments: This work was supported by grants from the Minister of Science and
808
Technology of China (2019YFA0508402), National Natural Science Foundation of China
809
(82188101), Shenzhen Bay Laboratory (S201101002), RGC of Hong Kong (AoE-M09-12,
810
16104518 and 16101419), and a HFSP Research Grant (RGP0020/2019) to MZ. The research
811
effort in HSC’s group was supported by Canadian Institutes of Health Research grant PJT-
812
155930 and Natural Sciences and Engineering Research Council of Canada grant RGPIN-
813
2018-04351.
814
815
Author contribution: ZS and MZ conceived the idea and designed experiments; ZS, BJ, YX
816
performed experiments; ZS, JW, and TP performed simulations; ZS, HSC, SD and MZ
817
analyzed data; SD supervised imaging experiments, ZS and MZ wrote and revised the
818
manuscript and all authors provided input, MZ coordinated the study.
819
28
820
Competing interest claim: The authors declare no competing financial interests.
821
822
823
824
29
References
825
Banani, S.F., Lee, H.O., Hyman, A.A., and Rosen, M.K. (2017). Biomolecular condensates: organizers
826
of cellular biochemistry. Nature Reviews Molecular Cell Biology 18, 285‐298.
827
Baron, M.K., Boeckers, T.M., Vaida, B., Faham, S., Gingery, M., Sawaya, M.R., Salyer, D., Gundelfinger,
828
E.D., and Bowie, J.U. (2006). An architectural framework that may lie at the core of the postsynaptic
829
density. Science 311, 531‐535.
830
Brangwynne, C.P., Mitchison, T.J., and Hyman, A.A. (2011). Active liquid‐like behavior of nucleoli
831
determines their size and shape in Xenopus laevis oocytes. Proc Natl Acad Sci U S A 108, 4334‐4339.
832
Chen, X., Wu, X., Wu, H., and Zhang, M. (2020). Phase separation at the synapse. Nature
833
Neuroscience 23, 301‐310.
834
Choi, J.‐M., Hyman, A.A., and Pappu, R.V. (2020). Generalized models for bond percolation
835
transitions of associative polymers. Physical Review E 102, 042403.
836
Chong, S., Graham, T.G., Dugast‐Darzacq, C., Dailey, G.M., Darzacq, X., and Tjian, R. (2022). Tuning
837
levels of low‐complexity domain interactions to modulate endogenous oncogenic transcription.
838
Molecular Cell.
839
Das, R., Cairo, C.W., and Coombs, D. (2009). A hidden Markov model for single particle tracks
840
quantifies dynamic interactions between LFA‐1 and the actin cytoskeleton. PLoS Comput Biol 5,
841
e1000556.
842
Delano, W.L., and Brünger, A.T. (1994). Helix packing in proteins: prediction and energetic analysis of
843
dimeric, trimeric, and tetrameric GCN4 coiled coil structures. Proteins: Structure, Function, and
844
Bioinformatics 20, 105‐123.
845
Espinosa, J.R., Joseph, J.A., Sanchez‐Burgos, I., Garaizar, A., Frenkel, D., and Collepardo‐Guevara, R.
846
(2020). Liquid network connectivity regulates the stability and composition of biomolecular
847
condensates with many components. Proc Natl Acad Sci U S A 117, 13238‐13247.
848
Feng, Z., Jia, B., and Zhang, M. (2021). Liquid‐Liquid Phase Separation in Biology: Specific
849
Stoichiometric Molecular Interactions vs Promiscuous Interactions Mediated by Disordered
850
Sequences. Biochemistry 60, 2397‐2406.
851
Feric, M., Vaidya, N., Harmon, T.S., Mitrea, D.M., Zhu, L., Richardson, T.M., Kriwacki, R.W., Pappu,
852
R.V., and Brangwynne, C.P. (2016). Coexisting Liquid Phases Underlie Nucleolar Subcompartments.
853
Cell 165, 1686‐1697.
854
Gibson, B.A., Doolittle, L.K., Schneider, M.W.G., Jensen, L.E., Gamarra, N., Henry, L., Gerlich, D.W.,
855
Redding, S., and Rosen, M.K. (2019). Organization of Chromatin by Intrinsic and Regulated Phase
856
Separation. Cell 179, 470‐484 e421.
857
Hansen, A.S., Woringer, M., Grimm, J.B., Lavis, L.D., Tjian, R., and Darzacq, X. (2018). Robust model‐
858
based analysis of single‐particle tracking experiments with Spot‐On. Elife 7, e33125.
859
Harmon, T.S., Holehouse, A.S., Rosen, M.K., and Pappu, R.V. (2017). Intrinsically disordered linkers
860
determine the interplay between phase separation and gelation in multivalent proteins. Elife 6.
861
Hubatsch, L., Jawerth, L.M., Love, C., Bauermann, J., Tang, T.Y.D., Bo, S., Hyman, A.A., and Weber,
862
C.A. (2021). Quantitative theory for the diffusive dynamics of liquid condensates. eLife 10, e68620.
863
Jaqaman, K., Loerke, D., Mettlen, M., Kuwata, H., Grinstein, S., Schmid, S.L., and Danuser, G. (2008).
864
Robust single‐particle tracking in live‐cell time‐lapse sequences. Nature Methods 5, 695‐702.
865
Kato, M., Han, Tina W., Xie, S., Shi, K., Du, X., Wu, Leeju C., Mirzaei, H., Goldsmith, Elizabeth J.,
866
Longgood, J., Pei, J., et al. (2012). Cell‐free Formation of RNA Granules: Low Complexity Sequence
867
Domains Form Dynamic Fibers within Hydrogels. Cell 149, 753‐767.
868
Lin, Y.H., Wu, H., Jia, B., Zhang, M., and Chan, H.S. (2022). Assembly of model postsynaptic densities
869
involves interactions auxiliary to stoichiometric binding. Biophys J 121, 157‐171.
870
Lyon, A.S., Peeples, W.B., and Rosen, M.K. (2021). A framework for understanding the functions of
871
biomolecular condensates across scales. Nat Rev Mol Cell Biol 22, 215‐235.
872
30
Manley, S., Gillette, J.M., Patterson, G.H., Shroff, H., Hess, H.F., Betzig, E., and Lippincott‐Schwartz, J.
873
(2008). High‐density mapping of single‐molecule trajectories with photoactivated localization
874
microscopy. Nature Methods 5, 155‐157.
875
McSwiggen, D.T., Mir, M., Darzacq, X., and Tjian, R. (2019). Evaluating phase separation in live cells:
876
diagnosis, caveats, and functional consequences. Genes & development 33, 1619‐1634.
877
Miné‐Hattab, J., Heltberg, M., Villemeur, M., Guedj, C., Mora, T., Walczak, A.M., Dahan, M., and
878
Taddei, A. (2021). Single molecule microscopy reveals key physical features of repair foci in living
879
cells. Elife 10.
880
Mittag, T., and Pappu, R.V. (2022). A conceptual framework for understanding phase separation and
881
addressing open questions and challenges. Mol Cell.
882
Musacchio, A. (2022). On the role of phase separation in the biogenesis of membraneless
883
compartments. Embo j 41, e109952.
884
Niewidok, B., Igaev, M., Pereira da Graca, A., Strassner, A., Lenzen, C., Richter, C.P., Piehler, J., Kurre,
885
R., and Brandt, R. (2018). Single‐molecule imaging reveals dynamic biphasic partition of RNA‐binding
886
proteins in stress granules. J Cell Biol 217, 1303‐1318.
887
Persson, F., Lindén, M., Unoson, C., and Elf, J. (2013). Extracting intracellular diffusive states and
888
transition rates from single‐molecule tracking data. Nature Methods 10, 265‐269.
889
Riback, J.A., Zhu, L., Ferrolino, M.C., Tolbert, M., Mitrea, D.M., Sanders, D.W., Wei, M.T., Kriwacki,
890
R.W., and Brangwynne, C.P. (2020). Composition‐dependent thermodynamics of intracellular phase
891
separation. Nature 581, 209‐214.
892
Shin, Y., Berry, J., Pannucci, N., Haataja, M.P., Toettcher, J.E., and Brangwynne, C.P. (2017).
893
Spatiotemporal Control of Intracellular Phase Transitions Using Light‐Activated optoDroplets. Cell
894
168, 159‐171.e114.
895
Tinevez, J.‐Y., Perry, N., Schindelin, J., Hoopes, G.M., Reynolds, G.D., Laplantine, E., Bednarek, S.Y.,
896
Shorte, S.L., and Eliceiri, K.W. (2017). TrackMate: An open and extensible platform for single‐particle
897
tracking. Methods 115, 80‐90.
898
van de Linde, S., Löschberger, A., Klein, T., Heidbreder, M., Wolter, S., Heilemann, M., and Sauer, M.
899
(2011). Direct stochastic optical reconstruction microscopy with standard fluorescent probes. Nature
900
Protocols 6, 991‐1009.
901
Winnik, M.A., and Yekta, A. (1997). Associative polymers in aqueous solution. Current Opinion in
902
Colloid & Interface Science 2, 424‐436.
903
Wu, X., Cai, Q., Feng, Z., and Zhang, M. (2020). Liquid‐Liquid Phase Separation in Neuronal
904
Development and Synaptic Signaling. Dev Cell 55, 18‐29.
905
Wu, X., Cai, Q., Shen, Z., Chen, X., Zeng, M., Du, S., and Zhang, M. (2019). RIM and RIM‐BP Form
906
Presynaptic Active‐Zone‐like Condensates via Phase Separation. Mol Cell 73, 971‐984 e975.
907
Zeng, M., Chen, X., Guan, D., Xu, J., Wu, H., Tong, P., and Zhang, M. (2018). Reconstituted
908
Postsynaptic Density as a Molecular Platform for Understanding Synapse Formation and Plasticity.
909
Cell 174, 1172‐1187 e1116.
910
Zeng, M., Shang, Y., Araki, Y., Guo, T., Huganir, R.L., and Zhang, M. (2016). Phase Transition in
911
Postsynaptic Densities Underlies Formation of Synaptic Complexes and Synaptic Plasticity. Cell 166,
912
1163‐1175 e1112.
913
Zhao, T., Wang, Y., Zhai, Y., Qu, X., Cheng, A., Du, S., and Loy, M.M. (2015). A user‐friendly two‐color
914
super‐resolution localization microscope. Opt Express 23, 1879‐1887.
915
916
917
31
Figures and Legends
918
919
920
Figure 1: Single molecule imaging of phase separation on supported lipid bilayers
921
(A) Schematic diagram showing phase separation of PSD protein assembly on SLBs (Zeng et
922
al., 2018).
923
(B) Upper panel: a TIRF image of Alexa647 labelled His8-NR2B tetramer clustered within
924
the PSD condensate on SLB. Lower panel: Stacking of 4000 frames of dSTORM images of
925
Alexa 647 labelled His8-NR2B within the same PSD condensates as shown in the TIRF
926
image above. Black dots represent localizations recognized during the imaging. Scale bar: 2
927
μm.
928
(C) Phase boundary of the PSD condensates determined by localization densities. The
929
boundaries are shown in blue lines. Localizations are color-coded according to their local
930
densities from low (blue) to high (red). A zoom in view of a typical condensed patch on SLB,
931
showing heterogeneous distributions and nano-cluster-like structures of molecules within the
932
condensed phase. Scale bar of the original image: 2 μm, scale bar for the zoom in view: 500
933
nm.
934
935
936
937
32
938
Figure 2: Development of an adaptive single molecule tracking algorithm for imaging
939
single molecules in the condensed and dilute phases simultaneously.
940
(A) Flow chart of the adaptive single molecule tracking algorithm.
941
(B) Assignments of motion tracks of NR2B in both condensed and dilute phases in the PSD
942
condensates formed on SLB. Each track is color coded from red to black representing from
943
the beginning to the end of the track. The boundaries of the condensates are marked by blue
944
lines. Scale bar: 2 μm.
945
(C) Representative tracks showing typical NR2B motions that contains different state
946
including exchange events of molecules between condensed and dilute phases (track 1&2),
947
and both transiently confined and freely diffusing states in the condensed phase (track 3).
948
33
(D) Percentages of NR2B molecules exchange from the dilute phase into the condensed
949
phase (“go-in”) and exchange from the condensed phase out to the dilute phase (“go-out”)
950
were counted in four sessions dSTORM imaging experiments. No significant difference
951
between go-in ratio and go-out ratio was detected. P=0.378 using paired t test.
952
(E) Raw image data superimposed with phase boundary (blue line), molecule localization
953
(green cross) and track steps (red line) shown a typical trajectory of a molecule that
954
experience multiple motion switchs between confined and mobile states in the condensed
955
phase. Scale bar: 200 nm.
956
957
34
958
35
Figure 3: Dynamic parameters and a diffusion model for an equilibrium state phase
959
separation system.
960
(A) Displacement distribution of tracks in (A1) condensed phase and (A2) dilute phase. Bin
961
size of histogram is 5 nm.
962
(B) Displacement distribution of tracks of NR2B along tethered to SLB. Red Curve is the
963
fitting with a simple 2D Brownian motion distribution use non-linear least squares method
964
using MATLAB, R2 = 0.91, RMSE = 42.5. Bin size of histogram is 5 nm.
965
(C&D) Fitting of the dynamic parameters of NR2B in the PSD condensates formed on SLBs
966
with Hidden Markov Model assuming that NR2B is undergoing a two-state motion model (i.e.
967
a transient confined state and a mobile state) in the condensed phase. The parameters are
968
diffusion coefficient of the confined state in condensed phase (Dc) and diffusion coefficient
969
of the mobile state (Dm), switching probability from the confined state to the mobile state
970
(Pcm) and the reversed switching probability (Pmc).
971
(E) Determination of the diffusion coefficients of NR2B in dilute phase (blue) and condensed
972
phase (red) by fitting the MSDs against time with a linear regression. The figure also includes
973
the curve and fitting of NR2B alone tethered to SLB (green). The insert shows a y-axis zoom-
974
in view of NR2B in the condensed phase. The number of trajectories used in the fittings were
975
2443 for the condensed phase, 13 for the dilute phase, and 248 for NR2B alone on SLB.
976
(F) Simulation of molecule diffusion on 2D surface with the Brownian motion with motion
977
switch model under different mobile ratio (Pm), diffusion coefficient (D) was fitted with the
978
MSD curve and normalized to mobile ratio = 100% scenario.
979
(G) Schematic diagram showing molecular motions between condensed phase and dilute
980
phase under the equilibrium state. Black and red dots represent molecules adopting confined
981
and mobile states, respectively, in the condensed phase. Green dots represent molecules in the
982
dilute phase. The lengths of the arrows are to indicate different mobilities of indicated
983
molecules.
984
36
985
Figure 4: Immobilization of molecules by the large and dynamic molecular network in
986
the condensed phase of phase separated systems.
987
(A) Schematic diagrams showing large and stable molecular networks in the condensed phase
988
formed by specific and multivalent interactions (left, blue) and small and dynamic molecular
989
37
networks in the condensed phase formed by weak, nonspecific but multivalent interactions
990
(right, gray). Red edge highlights the large dynamic network.
991
(B) Schematic diagram showing composition of three designed and “caged” single protein
992
phase separation systems with different interaction properties. PrLD, prion-like domain of
993
FUS; SAMWT, WT SAM domain from Shank3; SAMME, the M1718E mutant of Shank3
994
SAM domain; MBP, maltose binding protein as a caging tag; GB1, the B1 domain
995
of Streptococcal protein G as another caging tag. The HRV-3C cleavage sites (“3C cut site”)
996
of the proteins are also indicated.
997
(C) DIC images showing phase separations of the three designed proteins at different
998
concentrations after removal of the caging tags by HRV-3C protease cleavage. Scale bar: 20
999
μm.
1000
(D) Representative tracks showing different motion properties of the three designed proteins
1001
in condensed phase. Scale bar: 2 μm.
1002
(E) Comparison of diffusion coefficient in mobile state (left) and mobile ratio in condensed
1003
phase (right) of the three designed proteins. N=12, data are expressed as mean ±SD with
1004
**** P<0.0001, * P<0.0332 by t-test.
1005
38
1006
Figure 5: Simulations of FRAP experiments and comparison with the experimental
1007
FRAP results.
1008
(A) Schematic diagram of the phase separation system for the FRAP simulations. The
1009
simulation region is a 20 μm x8 μm box with periodic boundary. Three ROIs (1,2,3) with a
1010
fixed diameter of 0.5 μm and positioned at the center of three different sized droplets (2, 1
1011
and 0.5 μm in diameters, respectively) were selected for photobleaching. Green lines indicate
1012
the phase boundaries of the droplets. Black dots represent bleached molecules that can
1013
exchange with unbleached molecules in red. scale bar: 1 μm.
1014
39
(B) FRAP curves of the three ROIs under the condition that all molecules in the condensed
1015
phase are mobile, with an enrichment fold of 100, and diffusion coefficients in condensed and
1016
dilute phase of 0.01 μm2/s and 1 μm2/s, respectively. Data are expressed as mean ±SD from
1017
10 repeats of simulations.
1018
(C) FRAP curves of the three ROIs under the condition that only 10% of molecules in
1019
condensed phase are mobile and diffusion coefficients of the molecule in the condensed and
1020
dilute phase of 0.1 μm2/s and 1 μm2/s, respectively. The lifetime of the molecule in the
1021
confined state was set at 10 seconds.
1022
(D) Same as in C except that the lifetime of the molecule in the confined state was set at 100
1023
seconds.
1024
(E) Same as in C except that the molecule in the confined state were treated as permanently
1025
immobilized.
1026
(F) Experimental FRAP curves of PrLD, PrLD-SAMME, and PrLD-SAMWT condensates. In
1027
each case, a photobleaching region with the size of 1.95 μm in diameter was selected inside a
1028
large droplet (see Figure 5—figure supplement 1). The zoom-in panel is an expanded view of
1029
the FRAP curve of PrLD-SAMWT. Data are expressed as mean ±SD, with recovery
1030
experiments performed on 10 different droplets.
1031
(G) Simulated FRAP curves of the three designed proteins in the condensed phase using the
1032
parameters derived from the experiments described in Figure 4. The region selected for
1033
photobleaching is with a diameter of 2 μm and located in condensed phase with infinite size.
1034
Data are expressed as mean ±SD from three simulations.
1035
1036
40
Supplemental Figures and Legends
1037
1038
41
Figure 1—figure supplement 1: Evaluation of the adaptive single molecule tracking
1039
algorithm by simulation and by experiments.
1040
(A) Optimization point demonstration. Blue line showing the curve of left part of the
1041
equation 𝑋𝑒��� � √𝜋𝜎𝐷𝑡, green line showing the right part range of the equation in typical
1042
scenarios of phase separation. The red dashed line showing the optimized point range in
1043
different scenarios.
1044
(B&C) Simulated track assignment errors of molecules in homogeneous condensed phase
1045
with a slow diffusion (D~0.1 μm2/s, panel B) or in dilute phase with fast diffusions (D~1.0
1046
μm2/s, panel C) under different molecule densities (N) and maximum step limit/RMSD
1047
(R/RMSD) ratios. The red box highlighted the situation matching our experimental data for
1048
the PSD system.
1049
(D&E) Simulated diffusion coefficient errors of molecules in homogeneous condensed phase
1050
with slow diffusions (D~0.1 μm2/s, panel D) or in dilute phase with fast diffusions (D~1.0
1051
μm2/s, panel E) under different molecule densities and R/RMSD values. The red box
1052
highlighted the situation matching our experimental data.
1053
(F) FPLC-coupled with static light scattering analysis showing the column behavior and
1054
measured molecular weight of the purified MBP-His6-GCN4-Dimer, Trimer, and Tetramer.
1055
(G) Diffusion coefficient of homogeneous solutions of MBP-His6-GCN4-Dimer, Trimer, and
1056
Tetramer derived by our adaptive single molecule tracking algorithm. The diffusion
1057
coefficients were derived by fit MSD as a function of time and shown as mean ± SD with n
1058
equals of 9 samples for each protein.
1059
42
1060
Figure 1 —figure supplement 2: Simulations of track assignment errors of phase
1061
separations with molecules in the condensed phase undergoing homogenous free
1062
diffusions.
1063
Simulated track assignment errors vs maximum step limit/RMSD ratios under different phase
1064
separation conditions. Different colours were used to distinguish different average molecule
1065
density (N) in every frame. Each panel used different root mean square displacement (RMSD)
1066
in consecutive frames: (A) RMSD=50 nm, (B) RMSD=150 nm, (C) RMSD=200 nm, (D)
1067
RMSD=300 nm, (E) RMSD=400 nm, (F) RMSD=500 nm.
1068
43
1069
44
1070
Figure 1—figure supplement 3: Simulations of diffusion coefficient errors of phase
1071
separations with molecules in the condensed phase undergoing homogenous free
1072
diffusions.
1073
Simulated results of diffusion coefficient errors vs maximum step limit/RMSD ratios under
1074
different conditions (molecular densities and RMSD values as in Figure 1 — figure
1075
45
supplement 2). (A) RMSD=50 nm, (B) RMSD=150 nm, (C) RMSD=200 nm, (D)
1076
RMSD=300 nm, (E) RMSD=400 nm, (F) RMSD=500 nm.
1077
46
1078
Figure 1 —figure supplement 4: Simulations of track assignment errors of phase
1079
separations with molecules in the condensed phase containing both confined and mobile
1080
states.
1081
Simulated results of track assignment errors vs maximum step limit/RMSD ratios under
1082
different conditions. Different line colours were used to distinguish different mobile fraction
1083
(Pm). RMSD=100 nm for all panels, but molecule density (N) and dwell time (tm) is
1084
different for each condition. (A) N=100, tm=0.1 s, (B) N=150, tm=0.1 s, (C) N=200, tm=0.1
1085
s, (D) N=150, tm=0.2 s.
1086
47
1087
48
Figure 3—figure supplement 1: Typical tracks of NR2B only tethered to SLB or NR2B
1088
in 3D PSD condensates.
1089
(A) Representative tracks of NR2B only tethered to SLB, showing that the molecules
1090
undergo homogeneous diffusions on the membrane surface. Scale bar: 2 μm.
1091
(B) Representative tracks showing that NR2B molecules in the 3D PSD condensates formed
1092
by PSD-95, GKAP, Shank3, and Homer switch between confined state and mobile state.
1093
Scale bar: 2 μm.
1094
(C) Best fit of displacement distribution in condensed (C1) and dilute (C2) phase with a
1095
simple diffusion model. Red Curve is the fitting with a simple 2D Brownian motion
1096
distribution use non-linear least squares method using MATLAB. (C1) R2 = 0.97, RMSE =
1097
213.9. (C1) R2 = 0.76, RMSE = 4.11. Bin size of histogram is 5 nm.
1098
(D) Schematic of our phase equilibrium simulation. The simulation region was a 15x30 μm2
1099
2-dimensional box with periodic boundary conditions. Green and purple filled/empty dots
1100
indicate that molecules crossing a boundary of the simulation box will re-enter the box at a
1101
symmetric position through the opposing boundary. Red and black filled/empty dots
1102
represent molecules that switch their motion states.
1103
1104
49
1105
Figure 5 — figure supplement 1: Representative confocal images showing FRAP
1106
experiments of the condensed droplets formed by PrLD, SAMME, or SAMWT. The
1107
region selected for photobleaching is with the size of 20 pixels or 1.95 μm in diameter.
1108
Photobleaching started at time point 0. Scale bar: 2 μm.
1109
1110
1111
50
Supplementary table
1112
Table 1: Simulation of phase separations with different input of diffusion parameters.
1113
Monte-Carlo method-based simulations of molecule diffusions SLB with experimental phase
1114
boundaries. A total of 50,000 molecules were included in each simulation and these
1115
molecules were randomly distributed at the beginning of simulations. Each simulation lasted
1116
for 100 seconds and was repeated 10 times. Output enrichment folds results were presented
1117
as mean value of last 10 seconds ± SD.
1118
1119
Input parameters
Theoretical
results
Output
results
Dm(μm2/s) Dd(μm2/s)
Mobile
ratio
Mobile state
lifetime(s)
Enrichment
folds
Enrichment
folds
0.2
0.6
0.05
0.1
60
58.9±0.7
0.1
0.6
0.1
0.1
60
58.6±0.9
0.01
0.6
1
-
60
57.0±0.8
0.1
0.6
0.1
0.5
60
57.9±0.7
0.1
0.6
0.05
0.1
120
116.7±1.9
0.2
0.6
0.1
0.1
30
29.5±0.3
1120
1121
51
Supplemental movie
1122
Movie 1: Raw image superimposed with phase boundary and tracks in the NR2B+PSD
1123
phase separation system on 2D SLB.
1124
Raw image superimposed with phase boundary (blue line) and tracks (steps length >5) in
1125
NR2B+PSD phase separation system on 2D SLB. Red lines represent tracks in the condensed
1126
phase, green lines represent tracks in the dilute phase, and yellow lines represent tracks cross
1127
phase boundaries. Molecules can switch between the confined and diluted states and can
1128
directly observe molecules diffuse cross the phase boundaries. Movie 1 is played in real time.
1129
Scale bar: 500 nm.
1130
1131
| 2022 | Biological condensates form percolated networks with molecular motion properties distinctly different from dilute solutions | 10.1101/2022.07.20.500769 | [
"Shen Zeyu",
"Jia Bowen",
"Xu Yang",
"Wessén Jonas",
"Pal Tanmoy",
"Chan Hue Sun",
"Du Shengwang",
"Zhang Mingjie"
] | creative-commons |
1
Title: Micronutrient supplements with iron promote disruptive protozoan and fungal
1
communities in the developing infant gut
2
3
Ana Popovic1,2, Celine Bourdon3,4, Pauline W. Wang5,6, David S. Guttman5,6, Sajid Soofi7,
4
Zulfiqar A. Bhutta4,7, Robert H. J. Bandsma3,4, John Parkinson1,2,8,* and Lisa G. Pell4
5
6
1Program in Molecular Medicine, Hospital for Sick Children Research Institute
7
2Department of Biochemistry, University of Toronto, Toronto, Ontario, Canada
8
3Division of Gastroenterology, Hepatology and Nutrition, Hospital for Sick Children, Toronto,
9
Ontario, Canada
10
4Centre for Global Child Health, Hospital for Sick Children, Toronto, Ontario, Canada
11
5Department of Cell & Systems Biology, University of Toronto, Toronto, Ontario, Canada
12
6Centre for the Analysis of Genome Evolution & Function, University of Toronto, Toronto,
13
Ontario, Canada
14
7Center of Excellence in Women and Child Health, the Aga Khan University, Karachi, Pakistan
15
8Department of Molecular Genetics, University of Toronto, Toronto, Ontario, Canada
16
17
*To whom correspondence should be addressed:
18
john.parkinson@utoronto.ca
19
20
Keywords: Eukaryotic microbiota; Parasites; Malnutrition; Micronutrient Supplementation;
21
Microbiome
22
2
Abstract
23
24
Supplementation with micronutrients, including vitamins, iron and zinc, is a key strategy to
25
alleviate child malnutrition. However, adverse events resulting in gastrointestinal disorders,
26
largely associated with iron, has resulted in ongoing debate over their administration. To better
27
understand their impact on gut microbiota, we analysed the bacterial, protozoal, fungal and
28
helminth communities of stool samples collected from children that had previously been recruited
29
to a cluster randomized controlled trial of micronutrient supplementation in Pakistan. We show
30
that while bacterial diversity was reduced in supplemented children, vitamins and iron may
31
promote colonization with distinct protozoa and mucormycetes, whereas the addition of zinc
32
ameliorates this effect. In addition to supplements, residence in a rural versus urban setting is an
33
important determinant of eukaryotic composition. We suggest that the risks and benefits of such
34
interventions may be mediated in part through eukaryotic communities, in a manner dependent on
35
setting.
36
37
38
3
Introduction
39
Malnutrition is a global health crisis with 149 million children stunted and 45 million children
40
wasted under the age of five years1,2. With increased vulnerability to infection, undernourished
41
children are at elevated risk of death, not least from diarrheal diseases3,4. Previous studies have
42
demonstrated the role of gut microbiota in malnutrition, with microbiome immaturity (bacterial
43
communities that are underdeveloped with respect to age) representing a key factor in disease
44
development5,6. Beyond bacterial communities, parasites such as hookworm, Cryptosporidium and
45
Entamoeba have also been associated with severe diarrheal disease and intestinal malabsorption7,8.
46
However, much less is known regarding the role of other, potentially commensal, eukaryotic gut
47
microbes in undernutrition. Of particular interest is their ability to interact with and alter bacterial
48
communities. For example, indole-producing gut bacteria were found to confer protection against
49
Cryptosporidium infection, while deworming treatments targeting helminth endemic communities
50
reduced abundance of protective Clostridiales9,10. Mouse studies further showed that helminths
51
and protozoa influence bacterial communities by modulating the host immune system9,11,12. While
52
the number of published gut microbiome studies have increased rapidly over the last decade, few
53
have explored the composition of eukaryotic gut communities and their potential interactions with
54
bacteria. Previously, we applied 18S rRNA and internal transcribed spacer (ITS) sequence surveys
55
to systematically characterize eukaryotic microbiota in severely malnourished Malawian children,
56
and identified a high prevalence of protozoa, including commensals and pathobionts13. We
57
furthermore associated Blastocystis colonization with increased gut bacterial diversity.
58
59
Global health programs targeting vulnerable child populations include the use of micronutrient
60
supplements, consisting of vitamins as well as essential minerals zinc and iron, that have been
61
4
demonstrated to improve growth and reduce morbidity14-16. Such supplements are thought to
62
address deficiencies that can impair immune responses to infectious pathogens and impact gut
63
bacterial communities17-20. While beneficial, supplementation, especially with iron, may also
64
promote unintended pathogen growth, particularly where the host is unable to restrict
65
micronutrient bioavailability21. For example, it has been shown that surplus iron promotes the
66
growth of enteropathogens and induces intestinal inflammation in infants22,23. Furthermore, while
67
known to reduce the duration of childhood diarrheal episodes, zinc supplementation has been
68
associated with increased duration of Entamoeba histolytica infections24,25.
69
70
In an attempt to understand the impact of micronutrient supplementation on the complex
71
interactions between eukaryotic and bacterial microbiota in the maturing infant gut and health, we
72
performed 18S rRNA and 16S rRNA amplicon surveys on stool samples obtained at 12 and 24
73
months of age from 80 children, previously recruited as part of a cluster randomized trial conducted
74
in Pakistan. The trial was designed to investigate the impact of micronutrient powders (MNP)
75
containing vitamins and iron with or without zinc on growth and morbidity, and has shown an
76
excess of significant diarrheal and dysenteric episodes among children receiving MNPs26.
77
Microbial profiles were analysed in the context of supplementation, nutritional status, age and
78
place of residence (i.e., urban or rural) to reveal a complex landscape of associations with microbial
79
diversity, as well as specific taxa.
80
81
Results
82
Description of cohort
83
5
A total of 80 children (160 paired stool samples at 12 and 24 months of age) from all three
84
supplementation arms in the parent cRCT26 (control (n=24), MNP (n=29), and MNP with zinc
85
(n=27)) conducted in Sindh, Pakistan were selected based on sample availability for inclusion in
86
this study (Supplementary Fig. 1). The cohort includes children from both urban (Bilal colony)
87
and rural (Matiari district) study sites (Fig. 1a). Children were stratified by weight-for-length z-
88
scores (WLZ) at 24 months into a reference WLZ (WLZ >-1) or undernourished (WLZ < -2) group.
89
Subject characteristics are summarized in Table 1. The WLZ growth trajectories of the children
90
selected as the reference WLZ group approximately tracked the upper 50th percentile of the
91
original cohort, while the undernourished group started around the lower 50th percentile and
92
gradually dropped over time ending at the bottom 80th percentile of the cohort (Fig. 1b). This drop
93
in the WLZ of the undernourished children was driven by poor weight gain (Supplementary Fig.
94
2).
95
96
The developing infant gut is colonized by complex eukaryotic communities
97
We applied 18S rRNA amplicon sequencing to profile the eukaryotic communities in all 160 stool
98
samples. We generated a total of 11,639,233 paired 18S rRNA amplicon sequence reads (median
99
70,642) of which 4,386,494 could be classified as a eukaryotic microbe (median 22,932;
100
Supplementary Table 1). From these we identified a total of 859 eukaryotic OTUs (median 66;
101
Supplementary Table 1), which included 438 protozoan, three helminth and 418 fungal OTUs (Fig.
102
2a). Fungi, dominated by Mucoromycota and Ascomycota, accounted for 71% of all reads. The
103
most abundant were species in the Candida-Lodderomyces clade, Saccharomyces, and taxa
104
increasingly associated with rare but fatal infections known as mucormycoses: Rhizomucor,
105
Actinomucor
and
Lichtheimia.
Alveolates
accounted
for
25%
of
reads,
with
106
6
Gregarina/Gregarinasina and Cryptosporidium as the most abundant (5% and 3%, respectively).
107
Remaining reads were classified to numerous taxa, including known gut parasites such as
108
Enterocytozoon bieneusi, Pentatrichomonas hominis and the tapeworm Hymenolepis nana, as well
109
as uncharacterized alveolates, Amoebozoa and Cercozoa (Supplementary Table 2).
110
111
Protozoa were highly prevalent, with 89% of children colonized by at least one protozoan organism
112
by 12 months of age, and 92% by 24 months of age (Fig. 2b). Carriage of multiple species was
113
common in both the reference WLZ and undernourished groups, with on average 18 and 19 OTUs
114
per child at each time point, and a maximum of 91. A high detection of gregarines, typically
115
considered parasites of invertebrates, has not previously been reported in the human gut. In our
116
cohort, gregarine sequences accounted for nearly 230,000 reads and were identified in 69% and
117
71% of children at 12 and 24 months of age (Fig. 2b).
118
119
Micronutrient supplementation without zinc is associated with increased carriage of
120
protozoa and mucormycetes
121
Protozoan microbiota were significantly associated with place of residence, micronutrient
122
supplementation and/or nutritional status, but not age. Children residing in the rural study site had
123
increased protozoan richness (number of OTUs) compared to those from the urban setting (β = 11,
124
CI [5.3 – 16.6], p < 0.001) (Fig. 2c). Differences were attributed to higher carriage of
125
predominantly alveolate taxa, particularly Cryptosporidium (Fisher’s exact, CI [2-11], p < 0.01,
126
OR 4.9), species known to cause enteric symptoms (Fig. 2d). When stratifying by age group, only
127
Cryptosporidium and two OTUs classified as unknown Conoidasida, with 93% sequence identity
128
to Cryptosporidium, reached statistical significance at 24 months, with 2.4 and 9.6-fold higher
129
7
carriage, respectively, in children from rural settings (Fisher’s exact, CI [2.5-29], p < 0.05, OR
130
8.1; CI [2-670], p < 0.05, OR 15.2).
131
132
While we observed trends in increased fungal and protozoan richness in the undernourished cohort
133
(Fig. 2c), only the tapeworm Hymenolepis nana was detected with overall significantly higher
134
frequency in undernourished children (Fisher’s exact, CI [2-23], p < 0.05, OR 6.2) (Fig. 2d). At
135
12 months, detections were only 2% and 3% in reference WLZ and undernourished children,
136
respectively. However, by 24 months, carriage increased to 8% in reference WLZ and 43% in the
137
undernourished group (ns after multiple testing correction). We also observed trends of increased
138
carriage of Cryptosporidium and Cryptosporida (coccidians), represented by 46 OTUs in total, in
139
undernourished children (74% versus 65% at 12 months and 71% versus 61% at 24 months; ns)
140
(Fig. 2b). Furthermore, undernourished children receiving MNP with zinc had significantly fewer
141
protozoan OTUs relative to undernourished children in the control and MNP arms (β = -15.19, CI
142
[-29.27 – -1.12], p < 0.05), suggesting a possible inhibitory effect by the metal (Fig. 2c,
143
Supplementary Fig. 3).
144
145
Analysis of compositional differences among samples revealed four distinct clusters of protozoan
146
communities (Fig. 2e). The overall compositional variance was significantly explained by place of
147
residence (adonis, R2 0.02, p < 0.05) and micronutrient supplementation (adonis, R2 0.09, p <
148
0.001), where protozoan communities in children supplemented with MNP differed significantly
149
from those in control and MNP with zinc arms (MNP-CTL, R2 0.05, p < 0.01; MNP-MNP with
150
zinc R2 0.04, p < 0.01). Cluster 1, in particular, was enriched in MNP samples, X2 (6, N = 114) =
151
38.5, p < 0.001 (Fig. 2f). Key drivers of the diversity included Tritrichomonas, detected almost
152
8
exclusively in samples found in clusters 1 and 3 (correlation coefficient R2 0.21, p = 0.001), and
153
an OTU assigned to an unknown alveolate found predominantly in clusters 1 and 2 (R2 0.17, p =
154
0.001). These organisms were highly prevalent in both age groups, at 42% and 45%
155
(Tritrichomonas) and 20% and 21% (unknown alveolate). Fungal richness and phylogenetic
156
composition were not associated with any of the variables studied here.
157
158
We identified significantly higher carriages of seven phylogenetically distinct protozoa and six
159
fungi in children receiving MNPs without zinc, relative to those that were given zinc (six protozoa
160
and six fungi relative to the control group; Fig. 2d). Indeed, we noted a trend where MNP with
161
zinc reduced carriage of microbial eukaryotes to or below that observed in the control samples.
162
For example, Gregarina and an uncharacterized alveolate, which contributed to the previously
163
observed differences in beta diversity (Fig. 2e), were detected with 1.8 and 3.8-fold higher
164
frequency in the MNP group, with no differences between samples from the control and MNP with
165
zinc groups. Similarly, the carriages of three mucormycete genera (Rhizomucor, Actinomucor and
166
Mucor) were 1.3, 1.5 and 1.8-fold higher, respectively, in the MNP group compared to the control,
167
with no significant differences between the control and MNP with zinc groups. Toxoplasma was
168
the only genus with significantly reduced carriage in children receiving MNP with zinc; however,
169
we observed non-significant reductions in other organisms such as Cercomonas and Mucor (2 and
170
1.4-fold, respectively) suggesting possible species-specific effects. Despite previous reports of the
171
impact of zinc on helminths24, we did not detect significant differences in the carriage of the
172
tapeworm Hymenolepis nana among treatment arms.
173
174
Micronutrient supplements are associated with specific bacterial communities
175
9
Using 16S rRNA amplicon sequencing, we also profiled the stool bacterial microbiota. From the
176
13,984,120 sequenced reads (median 92,628), we identified 1108 bacterial OTUs across all 160
177
samples (median 50; Supplementary Table 3). Actinobacteria and Firmicutes were found to
178
dominate with just two OTUs (both assigned to Bifidobacterium) accounting for over 50% of all
179
reads (Fig. 3a, Supplementary Table 4). Age was the primary determinant of bacterial richness (β
180
= 43.65, CI [31.98 – 55.31], p < 0.001) and evenness (β = 0.80, CI [0.59 – 1.02], p < 0.001) (Fig.
181
3b, Supplementary Fig. 4,) as well as patterns of taxonomic composition as measured by Bray-
182
Curtis and weighted Unifrac dissimilarities (Fig. 3c; adonis, R2 0.06, p < 0.001; R2 0.05, p < 0.001).
183
Regression of dissimilarities in each child over time using partial correspondence analysis
184
indicated that 56% of Bray-Curtis and 59% of weighted Unifrac changes may be attributed to age.
185
By correlating the abundances of bacterial taxa with the first two axes of the Bray-Curtis
186
ordination, we identified the candidate drivers of community differences as the two dominant
187
Bifidobacterium species, with opposite abundance patterns perhaps suggesting succession of one
188
species by the other.
189
190
Consistent with a previous study27, bacterial richness was reduced in undernourished children (β
191
= -29.19, CI [-52.99 – -5.39], p < 0.05), while a significant interaction between nutritional status
192
and place of residence indicated that bacterial evenness was reduced in undernourished children
193
from the urban setting (β = 1.03, CI [0.11 – 1.95], p < 0.05) (Fig. 3b, Supplementary Fig. 4b). We
194
detected no significant association between nutritional status and locality and bacterial beta
195
diversities in this cohort.
196
197
10
Treatment with MNPs was associated with an overall increased abundance of Actinobacteria in
198
children at 12 months compared to the control group and those receiving MNP with zinc (β =
199
36020, CI [7239 – 64802], p < 0.05), but reduced abundance in the MNP group at 24 months (β =
200
-52670, CI [-93373 – -11966], p < 0.05) (Fig. 3d). Firmicutes were reduced in the presence of zinc
201
in both age groups (β = -261976, CI [-476591 – -47362], p < 0.05), with a non-significant reduction
202
in those supplemented without zinc (β = -206413, CI [-416049 – 3221], p = 0.055).
203
Supplementation tended to reduce overall bacterial richness with an effect that reached
204
significance in the MNP group (β = -14.66, CI [-29.01 – -0.31], p < 0.05) (Fig. 3b) and influenced
205
taxonomic composition as measured by weighted Unifrac (adonis, R2 0.03, p < 0.01) but not Bray-
206
Curtis dissimilarities. Specifically, phylogenetic variance differed among groups (p < 0.001), with
207
significantly smaller differences among 12 month old children receiving MNP and MNP with zinc
208
(Tukey posthoc, p < 0.01) (Fig. 3e, Supplementary Fig. 4c). This may suggest that micronutrients
209
support or restrict the growth of select taxa. Through differential abundance analysis, we identified
210
14 taxa with reduced abundances in both supplemented groups at 12 months compared to controls,
211
including over 10-fold reductions in Anaerostipes, Anaerosalibacter and Clostridium XI (Fig. 3f).
212
Two additional Anaerostipes OTUs were reduced in supplemented groups at both ages, with six
213
OTUs reduced at 24 months only. MNP with zinc was associated with changes in an additional 46
214
taxa, and 29 taxa were altered in MNP samples. These included a seven-fold increase in
215
Escherichia-Shigella abundance in 12 month old MNP-supplemented children, increases in
216
several Lactobacilli and a 1.3-fold reduction in one Bifidobacterium OTU (Fig. 3g). These data
217
reveal that micronutrient supplementation may impact bacterial communities during early
218
development.
219
220
11
MNPs may destabilize microbial interactions in undernourished infants.
221
Microbial interaction networks were constructed to define significant taxonomic co-occurrences
222
(Fig. 4). We found that interactions, calculated as edges per node, increased with age irrespective
223
of treatment, nutritional status or place of residence, which reflects the development of more
224
complex communities as the child matures (Fig. 4a). The greatest change, with a 2.5-fold increase,
225
was noted in children in the MNP arm, which had the fewest taxon interactions at 12 months but
226
achieved parity with the control and MNP with zinc groups by 24 months. Cross-kingdom
227
interactions between bacteria and eukaryotes represented 20% to 30% of all interactions at 12
228
months, falling to between 15% and 24% by 24 months of age (Fig. 4b).
229
230
When split by nutritional status, we observed important differences in the networks of 12 month
231
old undernourished infants supplemented with micronutrients compared to the control and
232
reference WLZ groups (Fig. 4c,d). Within control groups, the microbial networks of
233
undernourished infants and those within the reference WLZ group had similar levels of
234
connectivity, with non-significant differences in degree distribution and betweenness centrality
235
scores. While children in the reference WLZ group receiving either supplement were associated
236
with small but significant reductions in microbiota betweenness (Wilcoxon rank sum, p < 0.05 and
237
p < 0.01), greater reductions were observed in supplemented undernourished children (Wilcoxon
238
rank sum, p < 0.001). Since betweenness provides a measure of the degree of coordination within
239
a network, these findings suggest that micronutrient supplementation, with or without zinc, results
240
in microbial communities that are less organized at 12 months of age. This is further illustrated by
241
the network visualizations (Fig. 4e), where, in addition to changes in network density, we also
242
identified shifts in taxa with the highest betweenness values (which can be interpreted as those
243
12
taxa most likely to mediate important coordinating roles within the communities). For example,
244
within the control group, Clostridia, two species of Mucoromycota and the ciliate Bromeliothrix
245
occupy central roles in the network of reference WLZ infants, while in undernourished infants
246
these central roles are held by Trichosporon, Longamoeba and Prevotella. In supplemented
247
reference WLZ groups, Bacilli exhibit the highest betweenness values in the absence of zinc, while
248
these are replaced by Proteobacteria in the communities from infants receiving MNP with zinc.
249
However, within undernourished infants receiving either supplement, microbial networks appear
250
largely fragmented (Fig. 4e), with dramatically lower degree distributions and betweenness
251
compared to the control group suggesting that early treatment with micronutrient powders may
252
destabilize a fragile microbial community. Comparison of microbial networks by location of
253
residence further showed an increased density of interactions within each rural group (control or
254
supplemented) compared to the urban groups (Supplementary Fig. 5). Low subject numbers
255
precluded us from successfully generating networks at 24 months, where numbers of microbial
256
taxa are greater.
257
258
Complex cross-kingdom interrelationships over time are more influenced by place of
259
residence than early supplementation
260
Based on our findings, we hypothesized that direct effects of supplementation and place of
261
residence on microbial communities at 12 months could translate to indirect influence on later
262
microbial profiles. We further hypothesized that early exposure to eukaryotes (before or at 12
263
months of age) would change the course of bacterial microbiome maturation. To explore the
264
complex direct and indirect interrelationships among these factors, we generated an integrated
265
model using partial least squares (PLS) path modelling (Fig. 5, Supplementary Table 5). First,
266
13
place of residence had strong direct and indirect influences on eukaryotic and bacterial profiles at
267
both 12 and 24 months. The greatest direct effects were on eukaryotic composition (12mo, path
268
coefficient 0.52 ± 0.09, p < 0.0001; 24 mo, path coefficient 0.48± 0.1, p < 0.0001). Consistent with
269
our findings above, children from the rural community had increased levels of several alveolates
270
including Cryptosporidium at 12 and 24 months (12 months, 0.40 loading; 24 months 0.69
271
loading). While there was no significant direct effect on bacteria at 12 months (path coefficient
272
0.17 ± 0.12, p = 0.15), the locality indirectly influenced bacterial composition via eukaryotes
273
(indirect path coefficient of 0.14 with a total effect of 0.31 at 12 months). Children from the rural
274
community loaded positively for several Clostridium OTUs at both ages, and sustained higher
275
levels of Lactobacillus at 24 months. Micronutrient supplementation appeared to influence the
276
composition of eukaryotes and bacteria in an opposing manner to place of residence at 12 months
277
(path coefficient -0.27 ± 0.11, p = 0.014; path coefficient -0.27 ± 0.09, p = 0.0058), with possible
278
carryover effects to microbial compositions at 24 months (indirect effects of -0.11). Also consistent
279
with our findings, Mucor and Euglyphida correlated with supplementation at 12 months (-0.35 and
280
-0.34 cross-loadings, respectively).
281
282
Eukaryotic profiles at 12 months of age were significantly associated with a shift in bacterial
283
profiles at 12 months suggesting possible cross-kingdom interactions (Fig. 5, arrow 1; path
284
coefficient 0.27 ± 0.12, p = 0.033). These bacteria, in turn, exhibited a significant influence on
285
eukaryotic composition at 24 months (Fig. 5, arrow 4; path coefficient 0.21 ± 0.095, p = 0.033).
286
Differences in path coefficients were also tested in a stratified analysis of reference WLZ and
287
undernourished children but none reached statistical significance in our cohort. While the pathway
288
coefficients identified above were found to be statistically significant, due to large standard errors
289
14
likely resulting from heterogeneity and small sample size, we were unable to validate this support
290
using more robust bootstrapping procedures (Supplementary Table 5). Nevertheless, given the
291
consistency of these relationships with our earlier findings, this model provides additional support
292
for the indirect association of MNP supplementation and bacterial communities mediated through
293
the promotion of specific eukaryotic microbes.
294
295
Discussion
296
Malnutrition, both undernutrition and obesity, are associated with altered bacterial compositions,
297
where in the former, underdeveloped bacterial communities have the capacity to induce weight
298
loss6,28. Here, we have shown that the gut microbiota of both undernourished children and those
299
within a healthy weight range include a diverse group of protozoa, helminths and fungi, each with
300
the capacity to impact host health. We have also shown that supplementation with MNPs, a
301
strategy used to improve growth and alleviate micronutrient deficiencies14,16, has the capacity to
302
influence the development of the microbiome in these susceptible populations.
303
304
Consistent with previous studies, we found that bacterial communities became more complex
305
during growth. Eukaryotic communities, however, were not significantly impacted by age, but
306
instead were associated with micronutrient supplementation and place of residence. Only the
307
tapeworm H. nana was identified at significantly higher levels in undernourished children. While
308
H. nana infection is usually asymptomatic, high egg burdens in children have previously been
309
associated with diarrhea, abdominal pain and weight loss29, with exacerbated morbidity in children
310
<5 years30. We associated rural habitation with significantly more diverse protozoan communities,
311
and in particular increased prevalence of Cryptosporidium. An important cause of infant mortality
312
15
and childhood malnutrition, Cryptosporidium infection is attributed to unsafe drinking water and
313
inadequate sanitation often associated with rural settings26,31. While approximately half of all
314
children enrolled in the trial had access to piped drinking water (41% and 52% in the urban Bilal
315
colony and rural Matiari sites respectively), only 4% of children in the Matiari district had access
316
to underground sewage, compared to 95% in the Bilal Colony26, consistent with a lack of waste
317
water sanitation resulting in higher parasite carriage. While the large multicenter GEMS study
318
reported Cryptosporidium as a leading cause of death in 12 to 23 month old children with moderate
319
to severe diarrhea in developing countries32, we found a high prevalence of this parasite in absence
320
of diarrhea (80% and 83% at 12 and 24 months in the Matiari district, and 60% and 33% in the
321
Bilal urban colony). As our detection is based on 18S rRNA amplicon sequencing, we may have
322
detected a broader group of species of variable pathogenic potential compared to the GEMS study,
323
which applied a specific oocyst antigen immunoassay. Alternatively, our findings may indicate a
324
high prevalence of asymptomatic infections, with symptomatic infections resulting from additional
325
unknown factors7,33. The prevalence of Cryptosporidium in our cohort was also higher than
326
previously reported in non-diarrheal stools, using oocyst antigen testing, in the neighbouring
327
Naushero Feroze District (5.1% between 12 and 21 months of age), where Cryptosporidium
328
contributed to 8.8 diarrheal episodes per 100 child years34,35. This same study associated
329
asymptomatic enteropathogen infection, including Cryptosporidium and Giardia, across eight
330
countries with elevated inflammation and intestinal permeability, factors thought to increase risk
331
of stunting and impact the effectiveness of nutritional interventions in low-resource settings35.
332
333
A major focus of our study was to estimate the effect of micronutrient supplementation on the gut
334
microbiota. We found that children receiving supplements without zinc were associated with
335
16
distinct eukaryotic communities, featuring an increased prevalence of multiple protozoan and
336
fungal taxa; however, the addition of zinc to these supplements alleviated these increases, while
337
significantly reducing the prevalence of Toxoplasma and overall protozoan richness. These
338
findings are consistent with a previous report which suggested that zinc has a parasite-specific
339
protective effect against infection and ensuing diarrhea24. Fungal diversity was not impacted by
340
age, supplementation, place of residence or nutritional status. However, the predominance of
341
Mucoromycota, particularly in children receiving MNPs without zinc, is of concern, as these
342
organisms are responsible for rare but lethal invasive fungal infections that have previously been
343
reported in low birth weight infants and malnourished children36. Although incidence of infections
344
is rising globally, rates of mucormycoses are particularly high in Asia37. Notably, a recent spike in
345
infections, also termed ‘Black fungus’, in thousands of active and recovered Covid-19 patients in
346
India, was attributed to treatment with corticosteroids to control inflammation, in conjunction with
347
a high prevalence of diabetes38.
348
349
It has been well established that iron supplementation can promote the virulence of particular fungi
350
and parasites39,40. Several studies have shown that iron alone or in combination with other
351
micronutrients worsens existing infections, lengthens the duration and severity of diarrhea and
352
increases mortality rates in children22,26,39. Consequently, sequestration of free iron by host
353
proteins such as lactoferrin is a key defense mechanism to limit growth of pathogens including
354
Mucorales41. Iron deficiency has furthermore been suggested as protective against malaria
355
infection42,43, and provision of supplements containing iron in endemic regions has been cautioned
356
against due to increased malaria-related hospitalization and mortality of children39. While
357
deficiency in zinc has been associated with impaired immune function and susceptibility to
358
17
enteroinfections44, supplementation in the context of enteric pathogens was shown to have
359
parasite-specific outcomes. Provision of zinc alone can increase the incidence of Ascaris
360
lumbricoides and duration of Entamoeba histolytica infections, but it has also been shown to
361
reduce the duration of associated diarrheal episodes as well as lower the prevalence of Giardia
362
lamblia infections24. Interestingly, asymptomatic Giardia infections in children in Tanzania were
363
associated with reduced rates of diarrhea and fever, an effect which was lost in children receiving
364
vitamin and mineral supplements, including both iron and zinc45. Our data suggest that while iron,
365
vitamins, or both, may promote growth and survival of commensal and potentially pathogenic
366
eukaryotes, resulting in a shift in eukaryotic community structure, the addition of zinc may reduce
367
the ability of at least some eukaryotic microbes to infect and persist. The findings of reduced
368
bacterial diversity in 12 month old infants receiving micronutrient supplements, together with
369
elevated levels of Escherichia-Shigella and reduced beneficial Bifidobacteria, are also consistent
370
with previous reports, where reductions in beneficial Bifidobacterium and Lactobacilli and
371
increased enterobacteria in infants receiving iron-containing micronutrients were linked to
372
elevated risk of inflammation and diarrhea22,23,46. The original cRCT trial associated Aeromonas
373
infection with increased diarrhea in MNP supplemented groups26. We did not detect this bacterium
374
in our data, possibly due to exclusion of diarrheal samples.
375
376
The impact of micronutrient supplementation also extended to the structure of the microbial
377
communities. Microbial networks, representing significant correlations in the co-occurrence of
378
bacteria and eukaryotes, revealed higher network connectivity in the control groups, with the
379
networks generated from the undernourished infants receiving both types of supplements,
380
revealing a more fragmented structure. This fragmentation suggests a destabilization of species-
381
18
interactions within the developing gut microbiota in undernourished infants. Possibly contributing
382
to this destabilization is the presence of specific eukaryotic microbes, as evidenced by higher
383
proportions of eukaryotic-bacterial interactions in healthy infants receiving either supplement,
384
and/or the expansion of pathogenic bacteria. These microbes may interfere with the maturation of
385
commensal bacteria through predation, competition for resources and/or modulation of host
386
immunity. In undernourished infants, the cumulative effect of increases in pathogenic organisms
387
on community structure may be more pronounced than in infants within a healthy weight.
388
Enteropathogens Giardia lamblia and enteroaggregative Escherichia coli, for example, were
389
shown to have a greater impact on growth in protein-deficient mice during co-infection, an effect
390
which was dependent on the resident gut bacteria47. Taken together, our data showing increased
391
carriage of eukaryotic microbes and increased abundance of Escherichia-Shigella in children
392
supplemented with micronutrients, as well as a potential loss of organization in microbial
393
interactions in supplemented undernourished children, may offer at least a partial explanation for
394
previous reports of increased duration and severity of diarrhea as well as increased intestinal
395
inflammation in children supplemented with micronutrient powders26.
396
397
Due to the relatively small numbers of samples, we were unable to generate separate networks for
398
the three treatment arms for 24 month old children. We note that supplementation had ceased six
399
months prior, consequently the acute effects of these supplements may have dissipated. Small
400
sample sizes also preclude us from further segregating microbial networks by place of residence.
401
Micronutrient interventions may impact undernourished children differently in the context of a
402
high Cryptosporidium burden, for example. The notable absence of Giardia, a parasite typically
403
prevalent in this demographic, is likely due to mismatches to the 18S rRNA sequencing primers13.
404
19
Nevertheless, parasite diagnostic data from the trial did identify Giardia in 37 infants at 12 months,
405
and Cryptosporidium in seven, but noted no significant increases in either of the supplemented
406
groups26. Prevalence was nearly two-fold higher at the rural site, consistent with our findings for
407
Cryptosporidium, emphasizing the need for location-specific investigations of the effects of
408
micronutrient supplements. In addition to potential intraspecies variation, our detection of high
409
sequence diversity in Cryptosporidium OTUs specifically, and eukaryotic taxa in general, may be
410
exaggerated by a high proportion of non-overlapping amplicon reads, a consequence we have
411
attempted to minimize through manual curation. Regardless, we report that eukaryotic microbiota
412
are abundant members of the gut microbiome even in infancy, and given the known role of
413
parasitic pathogens in diarrheal disease and the association of fungi with obesity and inflammatory
414
bowel disease48,49, their role in malnutrition should be further studied.
415
416
Although not supported by robust bootstrapping, our integrated model of microbial relationships
417
and influencing external factors was able to recapitulate a number of key earlier findings, including
418
the impact of locality and micronutrients on gut eukaryotes. Furthermore, the prediction from our
419
model that complex cross-kingdom interactions may influence gut bacterial composition, provides
420
a valuable framework to dissect the direct and indirect effects of eukaryotic infections or nutritional
421
interventions on the maturing gut microbiome. Given the current debate over the use of MNP
422
supplementation and its role in gastrointestinal disorders, such a framework is expected to play a
423
key role in identifying scenarios where MNP supplementation may require more cautious thinking.
424
425
20
Conclusion
426
This study demonstrates that micronutrient powders impact the infant microbiota, with potentially
427
destabilizing effects driven through the promotion of specific organisms during early stages of
428
microbiome development. These findings are of relevance to micronutrient supplementation
429
strategies, especially those targeting vulnerable children in low resource settings.
430
431
Methods
432
Study design and subject selection
433
Study participants were selected from a multicenter clustered randomized controlled trial
434
(ClinicalTrials.gov identifier NCT00705445) that investigated the effects of micronutrient
435
supplementation with or without zinc among 2746 children from either an urban (Bilal colony,
436
squatter settlement within Karachi) or rural (Matiari district, 200 km from Karachi) site in Sindh,
437
Pakistan26. In the trial, daily supplementation with micronutrient powders (MNP) containing
438
vitamins A, C, D, folic acid and microencapsulated iron, with or without zinc spanned 6 to 18
439
months of age, with prospective follow-up until 24 months for the collection of health and
440
demographic information and stool samples26. Eighty children were selected for microbiome
441
profiling according to the following criteria (Supplementary Fig. 1): 1) having stool samples
442
collected at 12 and 24 months of age available and archived at -80oC; 2) having at 24 months a
443
weight-for-length z-score (WLZ) < -2 below the median (undernourished) or > -1 (reference WLZ)
444
based on WHO 2006 growth references (www.who.int/childgrowth); 3) no record of antibiotic
445
administration within 14 days of stool sample collection; and, 4) no reported diarrhea within seven
446
days of stool collection. Subjects within the reference group were further selected based on fewest
447
WLZ scores < -1 at other time points, to represent as healthy as possible a comparator group.
448
21
Participant characteristics were summarized as medians with interquartile ranges (IQRs) or means
449
± standard deviations (SD) if continuous variables, and percentages if categorical.
450
451
DNA extraction and amplicon sequencing
452
DNA was extracted from 100-200 mg of stool using the E.Z.N.ATM Stool kit (Omega Bio-Tek
453
Inc, GA, USA) according to the manufacturer’s protocol. Mechanical disruption of cells was
454
carried out with the MP Bio FastPrep-24 for 5 cycles of 1 min at 5.5 M/s. 16S variable region 4
455
(V4) amplifications were carried out using the KAPA2G Robust HotStart ReadyMix (KAPA
456
Biosystems) and barcoded primers 515F and 806R50. The cycling conditions were 95°C for 3 min,
457
22 cycles of 95°C for 15 s, 50°C for 15 s and 72°C for 15 s, followed by a 5 min 72°C extension.
458
Libraries were purified using Ampure XP beads and sequenced using MiSeq V2 (150bp x 2)
459
chemistry (Illumina, San Diego, CA). 18S V4+V5 amplification was achieved using the iProof
460
DNA polymerase (Bio-Rad Laboratories, Hercules, CA) with primers V4-1 and V4-4 as
461
previously described13. Briefly, the cycling conditions used were 94°C for 3 min, 30 cycles of
462
94°C for 45 s, 56°C for 1 min and 72°C for 1 min, followed by a 10 min 72°C extension. Barcodes
463
were ligated and libraries were sequenced using MiSeq V3 (300bp x 2) chemistry (Illumina, San
464
Diego, CA). Sequencing was performed at the Centre for the Analysis of Genome Evolution and
465
Function (Toronto, Canada).
466
467
Sequence data analysis
468
16S data were quality filtered and processed using VSEARCH v2.10.451 and the UNOISE pipeline
469
in USEARCH v11.0.66752,53. Filtered sequences were clustered to 99% sequence identity, and the
470
22
resulting operational taxonomic units (OTUs) were classified with a minimum confidence of 0.8
471
using the SINTAX54 algorithm and the Ribosomal Database Project version 1655.
472
473
18S data were quality filtered using Trimmomatic v0.3656 and read pairs with minimum 200
474
nucleotide length were merged using VSEARCH, or artificially joined using a linker of 50
475
ambiguous nucleotides (N50) using USEARCH. Resultant amplicon sequences were clustered to
476
97% sequence identity using the UCLUST52 algorithm, and taxonomically classified using SINA
477
v1.2.1157 with a minimum 90% sequence similarity threshold. Unclassified sequences were
478
submitted for classification using SINTAX and the SILVA v132 non-redundant reference
479
database58, and those still unclassified were compared to the NCBI non-redundant nucleotide
480
database59 (downloaded Nov 28, 2017) by BLAST60 using a 90% cutoff for both sequence identity
481
and query coverage. Phylogenetic tree construction for both 16S and 18S OTUs was performed
482
using the FastTree61 algorithm and visualized using the Iroki viewer62, with taxon prevalence
483
values calculated at a minimum threshold of 5 reads.
484
485
Microbial diversity and differential abundance analyses
486
Microbiota richness (number of OTUs) and evenness (Shannon Diversity Index, H) were
487
calculated using Phyloseq 1.20.063. Rarefaction curves were generated at 100 read intervals to a
488
maximum of 5,000 or 50,000 for eukaryotes and bacteria, respectively. Values were averaged and
489
standard errors calculated by the grouping variable. As intra-class correlation was low, we
490
implemented generalized linear models (GLMs) using richness and evenness values averaged from
491
100 independent rarefactions at read depths of 25,000 (bacteria) and 1,000 (protozoa and fungi).
492
To identify a final model that best explains diversity, we performed stepwise model selection using
493
23
AIC with MASS64 with the following explanatory variables: age, nutritional status,
494
supplementation and urban versus rural site.
495
496
Differences in bacterial composition, based on Bray-Curtis and weighted Unifrac dissimilarity
497
scores, were calculated with Phyloseq and vegan65 using DESeq2-normalized counts prefiltered
498
for taxa represented by a minimum of 5 reads in at least 5% of the samples. The contribution of
499
age to beta diversity was calculated using the capscale function, and the remaining variables were
500
tested for significance in age-stratified samples using adonis. The compositional variance within
501
groups, measured as distances to centroids, was evaluated using the betadisper function, and
502
pairwise differences were delineated using a post hoc Tukey test. All adonis and betadisper tests
503
were carried out with 9999 permutations. We applied non-metric dimensional scaling (NMDS) to
504
ordinate samples based on their compositional dissimilarity. The envfit function was used to
505
identify taxa significantly correlated with the first two ordination axes (candidate drivers of
506
community differences), indicated by arrows in the direction of cosines and scaled by the root
507
square of the correlation. Protozoan and fungal beta diversities were evaluated at 1000 read depth
508
using Principal Coordinate Analysis of unweighted Unifrac scores, and significance was tested as
509
above. Differential taxon abundance was tested with DESeq2 1.22.266 in samples containing a
510
minimum of 1000 reads, using data internally transformed with the median of ratios method.
511
512
Fisher’s Exact or pairwise test from the rstatix package was used to evaluate differences in
513
eukaryote carriage among participant groups, using a minimum 5 read detection threshold per
514
OTU and grouping OTUs to the genus level or the lowest assigned taxonomic level. Benjamini-
515
Hochberg correction was applied for multiple testing.
516
24
517
Microbial interaction networks
518
Bacterial and eukaryotic datasets were rarefied to 25,000 and 1,000 reads, respectively, and
519
eukaryotes were agglomerated to genera or the lowest assigned taxonomic level. Microbial
520
interaction networks, including both microbial datasets simultaneously, were generated using
521
SpiecEasi67 with the neighbour selection (MB) method, nlambda 100 and lambda.min.ratio 1e-02,
522
and visualized using igraph68.
523
524
Partial least squares path analysis
525
To explore the complex system of direct and indirect relationships between micronutrient
526
supplementation, place of residence and the multivariate matrices of bacteria and eukaryotes over
527
time, we conducted partial least squares (PLS) path analysis using the plspm package in R69.
528
Microbial read counts were center-log transformed after pre-filtering for taxa with more than
529
0.01% abundance across all samples. The analysis was set to collapse the high dimensional
530
microbial community matrices into latent PLS-scores representing community patterns of 1)
531
eukaryotes at 12 months, 2) eukaryotes at 24 months, 3) bacteria at 12 months and 4) bacteria at
532
24 months. The analysis estimates the relationships between factors based on cross correlations,
533
e.g. how eukaryotes detected at 12 months load into a community pattern summarized by a latent
534
PLS-score (i.e. “Eukaryotes, 12 mo”) in a manner that optimises the cross-correlation with the
535
other variables (i.e. supplementation, place of residence and other community patterns). Path
536
coefficients indicate the strength of the internodal relationship and can be conceptually understood
537
as correlation coefficients. Bootstrapping procedures were followed for validation and differences
538
in path coefficients were also tested between nutritional groups.
539
25
540
All microbial data and statistical analyses were carried out with R version 4.0.270.
541
542
Ethics Approval
543
The protocol for the cRCT trial was approved by the Ethics Review Committee of Aga Khan
544
University (752-Peds/ERC-07). This sub-study protocol was approved by research ethics board at
545
The Hospital for Sick Children, Toronto (REB No. 1000054244), the ethics review committee at
546
Aga Khan University, Karachi, Pakistan (4840-Ped-ERC-17), and the National Bioethics
547
Committee Pakistan (4-87/NBC-277/17/1191).
548
549
Data availability
550
Raw sequence data have been deposited to the NCBI Sequence Read Archive with the BioProject
551
identifier PRJNA717317.
552
553
Code availability
554
R code for analyses is available on GitHub (https://github.com/ParkinsonLab/gut-eukaryotes-
555
malnutrition-and-micronutrient-supplementation).
556
557
26
References
558
1
Global Nutrition Report: Action on equity to end malnutrition., (Bristol, UK, 2020).
559
2
(UNICEF), U. N. C. s. F., Organization, W. H. & Bank, I. B. f. R. a. D. T. W. Levels and trends in child
560
malnutrition: key findings of the 2021 edition of the joint child malnutrition estimates., (Geneva,
561
2021).
562
3
Black, R. E. et al. Maternal and child undernutrition and overweight in low-income and middle-
563
income countries. Lancet 382, 427-451, doi:10.1016/S0140-6736(13)60937-X (2013).
564
4
Liu, L. et al. Global, regional, and national causes of under-5 mortality in 2000-15: an updated
565
systematic analysis with implications for the Sustainable Development Goals. Lancet 388, 3027-
566
3035, doi:10.1016/S0140-6736(16)31593-8 (2016).
567
5
Subramanian, S. et al. Persistent gut microbiota immaturity in malnourished Bangladeshi
568
children. Nature 510, 417-421, doi:10.1038/nature13421 (2014).
569
6
Smith, M. I. et al. Gut microbiomes of Malawian twin pairs discordant for kwashiorkor. Science
570
339, 548-554, doi:10.1126/science.1229000 (2013).
571
7
Desai, N. T., Sarkar, R. & Kang, G. Cryptosporidiosis: An under-recognized public health problem.
572
Trop Parasitol 2, 91-98, doi:10.4103/2229-5070.105173 (2012).
573
8
Mondal, D. et al. Contribution of enteric infection, altered intestinal barrier function, and
574
maternal malnutrition to infant malnutrition in Bangladesh. Clin Infect Dis 54, 185-192,
575
doi:10.1093/cid/cir807 (2012).
576
9
Ramanan, D. et al. Helminth infection promotes colonization resistance via type 2 immunity.
577
Science 352, 608-612, doi:10.1126/science.aaf3229 (2016).
578
10
Chappell, C. L. et al. Fecal Indole as a Biomarker of Susceptibility to Cryptosporidium Infection.
579
Infect Immun 84, 2299-2306, doi:10.1128/IAI.00336-16 (2016).
580
11
Chudnovskiy, A. et al. Host-Protozoan Interactions Protect from Mucosal Infections through
581
Activation of the Inflammasome. Cell 167, 444-456 e414, doi:10.1016/j.cell.2016.08.076 (2016).
582
12
Reynolds, L. A. et al. Commensal-pathogen interactions in the intestinal tract: lactobacilli
583
promote infection with, and are promoted by, helminth parasites. Gut Microbes 5, 522-532,
584
doi:10.4161/gmic.32155 (2014).
585
13
Popovic, A. et al. Design and application of a novel two-amplicon approach for defining
586
eukaryotic microbiota. Microbiome 6, 228, doi:10.1186/s40168-018-0612-3 (2018).
587
14
Tam, E., Keats, E. C., Rind, F., Das, J. K. & Bhutta, A. Z. A. Micronutrient Supplementation and
588
Fortification Interventions on Health and Development Outcomes among Children Under-Five in
589
Low- and Middle-Income Countries: A Systematic Review and Meta-Analysis. Nutrients 12,
590
doi:10.3390/nu12020289 (2020).
591
15
Keats, E. C. et al. Effective interventions to address maternal and child malnutrition: an update
592
of the evidence. Lancet Child Adolesc Health 5, 367-384, doi:10.1016/S2352-4642(20)30274-1
593
(2021).
594
16
(UNICEF), U. N. C. s. F. Nutrition, for Every Child: UNICEF Nutrition Strategy 2020–2030.,
595
(UNICEF, New York, 2020).
596
17
Mayo-Wilson, E. et al. Zinc supplementation for preventing mortality, morbidity, and growth
597
failure in children aged 6 months to 12 years of age. Cochrane Database Syst Rev, CD009384,
598
doi:10.1002/14651858.CD009384.pub2 (2014).
599
18
Imdad, A., Mayo-Wilson, E., Herzer, K. & Bhutta, Z. A. Vitamin A supplementation for preventing
600
morbidity and mortality in children from six months to five years of age. Cochrane Database Syst
601
Rev 3, CD008524, doi:10.1002/14651858.CD008524.pub3 (2017).
602
19
Papier, K. et al. Childhood malnutrition and parasitic helminth interactions. Clin Infect Dis 59,
603
234-243, doi:10.1093/cid/ciu211 (2014).
604
27
20
Hibberd, M. C. et al. The effects of micronutrient deficiencies on bacterial species from the
605
human gut microbiota. Sci Transl Med 9, doi:10.1126/scitranslmed.aal4069 (2017).
606
21
Becker, K. W. & Skaar, E. P. Metal limitation and toxicity at the interface between host and
607
pathogen. FEMS Microbiol Rev 38, 1235-1249, doi:10.1111/1574-6976.12087
608
10.1111/1574-6976.12087. Epub 2014 Sep 29. (2014).
609
22
Jaeggi, T. et al. Iron fortification adversely affects the gut microbiome, increases pathogen
610
abundance and induces intestinal inflammation in Kenyan infants. Gut 64, 731-742,
611
doi:10.1136/gutjnl-2014-307720 (2015).
612
23
Paganini, D. et al. Iron-containing micronutrient powders modify the effect of oral antibiotics on
613
the infant gut microbiome and increase post-antibiotic diarrhoea risk: a controlled study in
614
Kenya. Gut 68, 645-653, doi:10.1136/gutjnl-2018-317399 (2019).
615
24
Long, K. Z. et al. Effect of vitamin A and zinc supplementation on gastrointestinal parasitic
616
infections among Mexican children. Pediatrics 120, e846-855, doi:10.1542/peds.2006-2187
617
(2007).
618
25
Richard, S. A. et al. Zinc and iron supplementation and malaria, diarrhea, and respiratory
619
infections in children in the Peruvian Amazon. Am J Trop Med Hyg 75, 126-132,
620
doi:10.4269/ajtmh.2006.75.1.0750126 (2006).
621
26
Soofi, S. et al. Effect of provision of daily zinc and iron with several micronutrients on growth
622
and morbidity among young children in Pakistan: a cluster-randomised trial. The Lancet 382, 29-
623
40, doi:10.1016/S0140-6736(13)60437-7 (2013).
624
27
Mendez-Salazar, E. O., Ortiz-Lopez, M. G., Granados-Silvestre, M. L. A., Palacios-Gonzalez, B. &
625
Menjivar, M. Altered Gut Microbiota and Compositional Changes in Firmicutes and
626
Proteobacteria in Mexican Undernourished and Obese Children. Front Microbiol 9, 2494,
627
doi:10.3389/fmicb.2018.02494 (2018).
628
28
Ridaura, V. K. et al. Gut microbiota from twins discordant for obesity modulate metabolism in
629
mice. Science 341, 1241214, doi:10.1126/science.1241214 (2013).
630
29
Ryan, E. T., Hill, D. R., Solomon, T., Endy, T. P. & Aronson, N. Hunter's tropical medicine and
631
emerging infectious diseases. (2020).
632
30
Soares Magalhaes, R. J. et al. Extending helminth control beyond STH and schistosomiasis: the
633
case of human hymenolepiasis. PLoS Negl Trop Dis 7, e2321, doi:10.1371/journal.pntd.0002321
634
(2013).
635
31
Daniels, M. E., Smith, W. A. & Jenkins, M. W. Estimating Cryptosporidium and Giardia disease
636
burdens for children drinking untreated groundwater in a rural population in India. PLoS Negl
637
Trop Dis 12, e0006231, doi:10.1371/journal.pntd.0006231 (2018).
638
32
Kotloff, K. L. et al. Burden and aetiology of diarrhoeal disease in infants and young children in
639
developing countries (the Global Enteric Multicenter Study, GEMS): a prospective, case-control
640
study. Lancet 382, 209-222, doi:10.1016/S0140-6736(13)60844-2 (2013).
641
33
Ajjampur, S. S. et al. Symptomatic and asymptomatic Cryptosporidium infections in children in a
642
semi-urban slum community in southern India. Am J Trop Med Hyg 83, 1110-1115,
643
doi:10.4269/ajtmh.2010.09-0644 (2010).
644
34
Platts-Mills, J. A. et al. Use of quantitative molecular diagnostic methods to assess the aetiology,
645
burden, and clinical characteristics of diarrhoea in children in low-resource settings: a reanalysis
646
of the MAL-ED cohort study. Lancet Glob Health 6, e1309-e1318, doi:10.1016/S2214-
647
109X(18)30349-8 (2018).
648
35
Kosek, M. N. & Investigators, M.-E. N. Causal Pathways from Enteropathogens to Environmental
649
Enteropathy: Findings from the MAL-ED Birth Cohort Study. EBioMedicine 18, 109-117,
650
doi:10.1016/j.ebiom.2017.02.024 (2017).
651
28
36
Francis, J. R., Villanueva, P., Bryant, P. & Blyth, C. C. Mucormycosis in Children: Review and
652
Recommendations for Management. J Pediatric Infect Dis Soc 7, 159-164,
653
doi:10.1093/jpids/pix107 (2018).
654
37
Prakash, H. & Chakrabarti, A. Global Epidemiology of Mucormycosis. J Fungi (Basel) 5,
655
doi:10.3390/jof5010026 (2019).
656
38
Raut, A. & Huy, N. T. Rising incidence of mucormycosis in patients with COVID-19: another
657
challenge for India amidst the second wave? Lancet Respir Med, doi:10.1016/S2213-
658
2600(21)00265-4 (2021).
659
39
Sazawal, S. et al. Effects of routine prophylactic supplementation with iron and folic acid on
660
admission to hospital and mortality in preschool children in a high malaria transmission setting:
661
community-based, randomised, placebo-controlled trial. Lancet 367, 133-143,
662
doi:10.1016/S0140-6736(06)67962-2 (2006).
663
40
Symeonidis, A. S. The role of iron and iron chelators in zygomycosis. Clin Microbiol Infect 15
664
Suppl 5, 26-32, doi:10.1111/j.1469-0691.2009.02976.x (2009).
665
41
Ibrahim, A. S., Spellberg, B., Walsh, T. J. & Kontoyiannis, D. P. Pathogenesis of mucormycosis.
666
Clin Infect Dis 54 Suppl 1, S16-22, doi:10.1093/cid/cir865 (2012).
667
42
Gwamaka, M. et al. Iron deficiency protects against severe Plasmodium falciparum malaria and
668
death in young children. Clin Infect Dis 54, 1137-1144, doi:10.1093/cid/cis010 (2012).
669
43
Jonker, F. A. et al. Iron status predicts malaria risk in Malawian preschool children. PLoS One 7,
670
e42670, doi:10.1371/journal.pone.0042670 (2012).
671
44
Shankar, A. H. & Prasad, A. S. Zinc and immune function: the biological basis of altered
672
resistance to infection. Am J Clin Nutr 68, 447S-463S, doi:10.1093/ajcn/68.2.447S (1998).
673
45
Veenemans, J. et al. Protection against diarrhea associated with Giardia intestinalis Is lost with
674
multi-nutrient supplementation: a study in Tanzanian children. PLoS Negl Trop Dis 5, e1158,
675
doi:10.1371/journal.pntd.0001158 (2011).
676
46
Paganini, D. & Zimmermann, M. B. The effects of iron fortification and supplementation on the
677
gut microbiome and diarrhea in infants and children: a review. Am J Clin Nutr 106, 1688S-1693S,
678
doi:10.3945/ajcn.117.156067 (2017).
679
47
Bartelt, L. A. et al. Cross-modulation of pathogen-specific pathways enhances malnutrition
680
during enteric co-infection with Giardia lamblia and enteroaggregative Escherichia coli. PLoS
681
Pathog 13, e1006471, doi:10.1371/journal.ppat.1006471 (2017).
682
48
Sokol, H. et al. Fungal microbiota dysbiosis in IBD. Gut 66, 1039-1048, doi:10.1136/gutjnl-2015-
683
310746 (2017).
684
49
Mar Rodriguez, M. et al. Obesity changes the human gut mycobiome. Sci Rep 5, 14600,
685
doi:10.1038/srep14600 (2015).
686
50
Caporaso, J. G. et al. Global patterns of 16S rRNA diversity at a depth of millions of sequences
687
per sample. Proc Natl Acad Sci U S A 108 Suppl 1, 4516-4522, doi:10.1073/pnas.1000080107
688
(2011).
689
51
Rognes, T., Flouri, T., Nichols, B., Quince, C. & Mahe, F. VSEARCH: a versatile open source tool
690
for metagenomics. PeerJ 4, e2584, doi:10.7717/peerj.2584 (2016).
691
52
Edgar, R. C. Search and clustering orders of magnitude faster than BLAST. Bioinformatics 26,
692
2460-2461, doi:10.1093/bioinformatics/btq461 (2010).
693
53
Edgar, R. C. UNOISE2: improved error-correction for Illumina 16S and ITS amplicon sequencing.
694
bioRxiv (2016).
695
54
Edgar, R. C. SINTAX: a simple non-Bayesian taxonomy classifier for 16S and ITS sequences.
696
bioRxiv (2016).
697
55
Cole, J. R. et al. Ribosomal Database Project: data and tools for high throughput rRNA analysis.
698
Nucleic Acids Res 42, D633-642, doi:10.1093/nar/gkt1244 (2014).
699
29
56
Bolger, A. M., Lohse, M. & Usadel, B. Trimmomatic: a flexible trimmer for Illumina sequence
700
data. Bioinformatics 30, 2114-2120, doi:10.1093/bioinformatics/btu170 (2014).
701
57
Pruesse, E., Peplies, J. & Glockner, F. O. SINA: accurate high-throughput multiple sequence
702
alignment of ribosomal RNA genes. Bioinformatics 28, 1823-1829,
703
doi:10.1093/bioinformatics/bts252 (2012).
704
58
Quast, C. et al. The SILVA ribosomal RNA gene database project: improved data processing and
705
web-based tools. Nucleic Acids Res 41, D590-596, doi:10.1093/nar/gks1219 (2013).
706
59
Coordinators, N. R. Database Resources of the National Center for Biotechnology Information.
707
Nucleic Acids Res 45, D12-D17, doi:10.1093/nar/gkw1071 (2017).
708
60
Altschul, S. F., Gish, W., Miller, W., Myers, E. W. & Lipman, D. J. Basic local alignment search
709
tool. J Mol Biol 215, 403-410, doi:10.1016/S0022-2836(05)80360-2 (1990).
710
61
Price, M. N., Dehal, P. S. & Arkin, A. P. FastTree: computing large minimum evolution trees with
711
profiles instead of a distance matrix. Mol Biol Evol 26, 1641-1650, doi:10.1093/molbev/msp077
712
(2009).
713
62
Moore, R. M., Harrison, A. O., McAllister, S. M., Polson, S. W. & Wommack, K. E. Iroki: automatic
714
customization and visualization of phylogenetic trees. PeerJ 8, e8584, doi:10.7717/peerj.8584
715
(2020).
716
63
McMurdie, P. J. & Holmes, S. phyloseq: an R package for reproducible interactive analysis and
717
graphics of microbiome census data. PLoS One 8, e61217, doi:10.1371/journal.pone.0061217
718
(2013).
719
64
Venables, W. N., Ripley, B. D. & Venables, W. N. Modern applied statistics with S. 4th edn,
720
(Springer, 2002).
721
65
Oksanen, J. B., F. Guillaume; Friendly, Michael; Kindt, Roeland; Legendre, Pierre; McGlinn, Dan;
722
Minchin, Peter R. ; O'Hara, R. B.; Simpson, Gavin L.; Solymos, Peter; Henry M.; Stevens, H.;
723
Szoecs, Eduard; Wagner, Helene. vegan: Community Ecology Package, <https://CRAN.R-
724
project.org/package=vegan> (2017).
725
66
Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for
726
RNA-seq data with DESeq2. Genome Biol 15, 550, doi:10.1186/s13059-014-0550-8 (2014).
727
67
Kurtz, Z. D. et al. Sparse and compositionally robust inference of microbial ecological networks.
728
PLoS Comput Biol 11, e1004226, doi:10.1371/journal.pcbi.1004226 (2015).
729
68
Csardi, G. & Nepusz, T. The igraph software package for complex network research. InterJournal,
730
Complex Systems 1695 (2006).
731
69
Sanchez, G. (Berkeley, 2013).
732
70
Team, R. C. R: A language and environment for statistical computing., <https://www.R-
733
project.org/> (2020).
734
735
30
Acknowledgements
736
We thank Imran Ahmed (Aga Khan University, Karachi, Pakistan) and Didar Alam (Aga Khan
737
University) for assistance with organizing stool sample shipments from Pakistan to Canada and
738
providing secure access to clinical data. We also appreciate the helpful advice and insights from
739
Amel Taibi and Elena Comelli in addressing challenges encountered during extraction of sample
740
DNA. This work was supported by a HSBC Bank Canada Catalyst Research Grant from the
741
Hospital for Sick Children awarded to CB, RB, DG, JP and LGP; the Canadian Institute for Health
742
Research grant PJT-152921 to JP; Restracomp scholarship administered by the Research Training
743
Centre (Hospital for Sick Children) and a graduate scholarship from the Government of Ontario
744
to AP. Computing resources were provided by the SciNet High Performance Computing (HPC)
745
Consortium; SciNet is funded by the Canada Foundation for Innovation under the auspices of
746
Compute Canada, the Government of Ontario, Ontario Research Fund - Research Excellence, and
747
the University of Toronto.
748
749
750
Author contributions
751
L.G.P., Z.A.B., J.P. and R.H.J.B. conceived and designed the study. S.S. and Z.A.B. participated
752
in original collection of clinical samples. A.P. isolated DNA and processed the sequencing data.
753
P.W.W and D.S.G. aided in design of amplicon generation. A.P. and C.B. analyzed the data and
754
wrote the paper and all authors reviewed and/or edited the paper.
755
756
Competing interests
757
The authors declare no competing interests.
758
759
31
Tables
760
761
Table 1. Participant characteristics. Categorical values are presented as n (%), continuous
762
variables show the mean and 95% confidence intervals. Premature birth was defined as
763
gestational age < 37 months. Initiation of breastfeeding was reported for the period prior to
764
recruitment into the study.
765
766
Undernourished Reference WLZ
Total
(n=31)
(n=49)
(n=80)
Rural site, n(%)
25 (80.6%)
28 (57.1%)
53 (66.2%)
Treatment arm, n(%)
Control
10 (32.3%)
14 (28.6%)
24 (30.0%)
MNP
14 (45.2%)
15 (30.6%)
29 (36.2%)
MNP with zinc
7 (22.6%)
20 (40.8%)
27 (33.8%)
Female, n(%)
14 (45.2%)
30 (61.2%)
44 (55.0%)
Premature birth, n(%)
6 (19.4%)
10 (20.4%)
16 (20.0%)
Initiated breastfeeding, n (%)
31 (100.0%)
47 (95.9%)
78 (97.5%)
Anthropometry, 12 mo
Weight, Kg
6.6 (6.3, 7.0)
8.5 (8.2, 8.8)
7.8 (7.5, 8.1)
Length, cm
69.2 (67.7, 70.7)
71.2 (70.4, 72.1)
70.6 (69.9, 71.4)
Weight-for-length, z-score
-2.4 (-3.1, -1.7)
-0.0 (-0.3, 0.2)
-0.7 (-1.1, -0.4)
Anthropometry, 24 mo
Weight, Kg
8.0 (7.6, 8.3)
10.5 (10.2, 10.9)
9.5 (9.2, 9.9)
Length, cm
78.9 (77.3, 80.4)
80.5 (79.5, 81.4)
79.8 (79.0, 80.7)
Weight-for-length, z-score
-2.9 (-3.2, -2.7)
0.2 (-0.1, 0.4)
-1.0 (-1.4, -0.7)
767
32
Figures and figure legends
768
769
770
Fig. 1. Participant characteristics. (a) Level of childhood undernutrition in Pakistan and the
771
surrounding regions. Latest country data was retrieved from www.who.int/data/gho/indicator-
772
metadata-registry/imr-details/27 on Feb 1, 2021. Urban and rural places of residence of the
773
participants are indicated. (b) Weight-for-length z-scores of children recruited into clinical trial
774
NCT00705445 during the first 24 months of life. Median and quantile values are shown, with
775
medians for participants profiled in current study indicated by red (undernourished) and black
776
(reference WLZ) lines.
777
778
33
779
Fig. 2. Eukaryotic communities in the gut are diverse and impacted by micronutrient
780
supplementation and place of residence. (a) Phylogenetic tree representing eukaryotic taxa
781
detected in children. Branches are coloured by phylum and bars represent the prevalences of OTUs
782
in the cohort. Named organisms represent those detected in more than 5% of samples with a
783
minimum of 100 reads. (b) Prevalences of protozoan (left), and specifically gregarine (middle) or
784
coccidian (right) OTUs detected in children at 12 and 24 months of age. Prevalences are subdivided
785
by nutritional group in bottom graphs, where shaded regions denote binned numbers of OTUs
786
identified per sample. (c) Rarefaction curves comparing the mean protozoan and fungal species
787
34
richness by age group, micronutrient supplementation, nutritional status and place of residence
788
(site). Shaded regions represent standard error. Dashed lines denote the read depth at which
789
significance was tested. (d) Carriage of eukaryotic taxa significantly associated with micronutrient
790
supplementation, place of residence (site) or nutritional status. Results from Fisher’s pairwise tests
791
among supplementation groups are indicated to the right. *p < 0.05, **p < 0.01, ***p < 0.001. (e)
792
Principal coordinate analysis of sample dissimilarities (n=106) based on protozoan composition,
793
calculated using unweighted Unifrac scores. Samples are coloured by supplementation arm, and
794
arrows indicate the direction of cosines of taxa significantly correlated with the first two principal
795
components. Arrow lengths are scaled by the root square (r2) of the correlation. Identified clusters
796
are numbered 1 though 4. (f) Proportions of samples from the respective supplementation arms
797
within each protozoan community cluster.
798
799
35
800
Fig. 3. Bacterial microbiota change with age and supplementation. (a) Relative abundances of
801
bacterial phyla in 12 (top) and 24 (bottom) month old children based on 16S data. Samples are
802
sorted by the proportion of Firmicutes along the horizontal axis. (b) Rarefaction curves comparing
803
mean species richness by age group, micronutrient supplementation, nutritional status and place
804
of residence (study site). Shaded regions represent standard errors and the dotted lines denote the
805
read depth at which significance was tested. (c) Non-metric multidimensional scaling of bacterial
806
compositions in samples based on Bray-Curtis dissimilarities. Samples are coloured by age and
807
ellipses represent 95% confidence intervals. Arrows indicate the direction of cosines of the top 10
808
bacterial OTUs significantly correlated with the ordination axes, and are scaled by their strength
809
of correlation (r2). (d) Mean DESeq2-transformed abundance of Actinobacteria and Firmicutes
810
grouped by nutritional status and treatment. (e) Compositional variance among samples grouped
811
by supplementation arm and age measured as distances to centroid, based on NMDS of weighted
812
Unifrac dissimilarity scores. *p < 0.05, **p < 0.01, ***p < 0.001 (f) Venn diagram showing the
813
numbers of bacterial taxa with significantly increased or decreased abundance, as indicated by
814
arrows, in supplemented groups relative to the control group. The pairs of numbers within brackets
815
refer to taxa at 12 and 24 months of age respectively, and select taxa are listed in boxes. (g)
816
36
Normalized
abundance
of
Escherichia-Shigella
and
Bifidobacterium
OTUs
across
817
supplementation arms at 12 months.
818
819
820
37
821
Fig. 4. Supplementation influences microbial interactions. (a) Density of microbial interactions,
822
calculated as significant correlations among microbiota (edges) normalized by the numbers of taxa
823
(nodes), by nutritional status, supplementation arm and place of residence (site). Lighter and darker
824
hues represent samples from 12 and 24 months respectively. (b) Proportions of significant
825
microbial interactions occurring cross-kingdom, within indicated sample groups. (c) Degree
826
distribution and (d) betweenness centrality scores of microbial networks in 12 month old children
827
grouped by nutritional status and supplementation arm. (e) Graphic representations of
828
38
aforementioned networks representing predicted microbial interactions in 12 month old children,
829
grouped by nutritional status and micronutrient treatment. Nodes represent bacterial OTUs
830
(yellow) and protozoan and fungal genera (red and grey, respectively), scaled by betweenness
831
centrality scores. Edges represent significant positive (grey) and negative (blue) correlations
832
among microbiota. Taxa with no predicted interactions have been removed. Numbers of samples
833
used to generate each network are indicated within brackets.
834
835
836
39
837
Fig. 5. Graphic representation of the cross-associations among demographic variables,
838
micronutrient supplementation and microbiota over time. Interconnected arrows indicate the tested
839
cross-correlated paths between nodes of: place of residence (site), supplementation, and the
840
composite measures of bacterial and eukaryotic OTUs detected at 12 and 24 months, collapsed as
841
latent PLS-scores. Negative correlations are indicated in pink and positive in blue. Arrow thickness
842
is weighted by the effect size of the direct path coefficients as indicated in Supplementary Table
843
5. Significance of direct paths, *p < 0.05, **p < 0.01, ***p < 0.0001. OTUs that loaded positively
844
(>0.4) or negatively (<-0.4) within each PLS-score are listed within boxes. PLS, partial least
845
square; OTUs, operational taxonomic units.
846
847
848
| 2021 | Micronutrient supplements with iron promote disruptive protozoan and fungal communities in the developing infant gut | 10.1101/2021.07.06.451346 | [
"Popovic Ana",
"Bourdon Celine",
"Wang Pauline W.",
"Guttman David S.",
"Soofi Sajid",
"Bhutta Zulfiqar A.",
"Bandsma Robert H. J.",
"Parkinson John",
"Pell Lisa G."
] | creative-commons |
Selective whole genome amplification as a tool to enrich specimens with low Treponema pallidum genomic DNA copies for whole genome
1
sequencing
2
3
Charles M. Thurlow,a# Sandeep J. Joseph,a Lilia Ganova-Raeva,b Samantha S. Katz,a Lara Pereira,a Cheng Chen,a Alyssa Debra,a Kendra
4
Vilfort,a Kimberly Workowski,a,c Stephanie E. Cohen,d Hilary Reno,e,f Yongcheng Sun,a Mark Burroughs,g Mili Sheth,g Kai-Hua Chi,a
5
Damien Danavall,a Susan S. Philip,c Weiping Cao,a Ellen N. Kersh,a and Allan Pillaya#
6
7
aDivision of STD Prevention, Centers for Disease Control and Prevention, Atlanta, Georgia, USA
8
bDivision of Viral Hepatitis, Centers for Disease Control and Prevention, Atlanta, Georgia, USA
9
cDepartment of Medicine, Emory University, Atlanta, Georgia, USA
10
dSan Francisco Department of Public Health, San Francisco, California, USA
11
eSt. Louis County Sexual Health Clinic, St. Louis, Missouri, USA
12
fDivision of Infectious Diseases, Washington University, St. Louis, Missouri, USA
13
gDivision of Scientific Resources, Centers for Disease Control and Prevention, Atlanta, Georgia, USA
14
15
Running title: Sequencing of T. pallidum from Clinical Specimens
16
17
#Address correspondence to Dr. Charles M. Thurlow, cthurlow@cdc.gov and Dr. Allan Pillay, apillay@cdc.gov.
18
19
Abstract.
20
Downstream next generation sequencing of the syphilis spirochete Treponema pallidum subspecies pallidum (T. pallidum) is hindered by
21
low bacterial loads and the overwhelming presence of background metagenomic DNA in clinical specimens. In this study, we investigated
22
selective whole genome amplification (SWGA) utilizing Multiple Displacement Amplification (MDA) in conjunction with custom
23
oligonucleotides with an increased specificity for the T. pallidum genome, and the capture and removal of CpG-methylated host DNA followed by
24
MDA as enrichment methods to improve the yields of T. pallidum DNA in rabbit propagated isolates and lesion specimens from patients with
25
primary and secondary syphilis. Sequencing was performed using the Illumina MiSeq v2 500 cycle or NovaSeq 6000 SP platform. These two
26
enrichment methods led to 93-98% genome coverage at 5 reads/site in 5 clinical specimens from the United States and rabbit propagated isolates,
27
containing >14 T. pallidum genomic copies/µl input for SWGA and >129 genomic copies/µl for CpG methylation capture with MDA. Variant
28
analysis using sequencing data derived from SWGA-enriched specimens, showed that all 5 clinical strains had the A2058G mutation associated
29
with azithromycin resistance. SWGA is a robust method that allows direct whole genome sequencing (WGS) of specimens containing very low
30
numbers of T. pallidum, which have been challenging until now.
31
Importance
32
Syphilis is a sexually transmitted, disseminated acute and chronic infection caused by the bacterial pathogen Treponema pallidum
33
subspecies pallidum. Primary syphilis typically presents as single or multiple mucocutaneous lesions, and if left untreated, can progress through
34
multiple stages with varied clinical manifestations. Molecular studies rely on direct amplification of DNA sequences from clinical specimens;
35
however, this can be impacted by inadequate samples due to disease progression or timing of patients seeking clinical care. While genotyping has
36
provided important data on circulating strains over the past two decades, whole genome sequencing data is needed to better understand strain
37
diversity, perform evolutionary tracing, and monitor antimicrobial resistance markers. The significance of our research is the development of a
38
SWGA DNA enrichment method that expands the range of clinical specimens that can be directly sequenced to include samples with low numbers
39
of T. pallidum.
40
Introduction
41
Syphilis rates have been steadily increasing in the United States with 38,992 cases (11.9 per 100,000 people) of primary and secondary
42
syphilis and 1,870 cases (48.5 per 100,000 live births) of congenital syphilis reported to the CDC during 2019 (1). This represents a 167.2%
43
increase in primary and secondary syphilis rates since 2010 and a 291.1% increase in congenital syphilis reported since 2015. While syphilis rates
44
have been on the rise in the U.S., the genetic diversity of the bacterial pathogen Treponema pallidum subspecies pallidum (hereafter referred to as
45
T. pallidum), in this setting, is not well understood due to the lack of recently sequenced whole genomes from clinical specimens. Strain diversity
46
has been gleaned from molecular epidemiology studies, which are based on 3 to 4 genetic loci, but may not be representative of the entire T.
47
pallidum genome (2-5).
48
Molecular studies have relied primarily on T. pallidum strains propagated in rabbits or DNA amplified directly from clinical specimens,
49
because T. pallidum cannot be grown on routine laboratory media. However, advances have been made with in vitro tissue culture and the
50
propagation of T. pallidum in rabbits from cryopreserved genital lesion specimens, which may make routine culture directly from clinical
51
specimens a possibility in the near future (6-7). Despite these advances, the methods are still time-consuming and impractical for laboratory
52
diagnosis and molecular epidemiological studies of syphilis.
53
Metagenomic shotgun sequencing approaches have made significant advances in recent years with sequence data being used for pathogen
54
detection, in silico or whole genome typing, and antimicrobial resistance marker detection, in addition to phylogenetic analyses (8-10). However,
55
direct whole genome sequencing (WGS) of T. pallidum from clinical specimens and rabbit isolates can be problematic due to bacterial genomic
56
DNA being outweighed by either human or rabbit DNA. Several DNA enrichment methods have been described for T. pallidum including RNA
57
bait capture techniques, methyl-directed enrichment using the restriction nuclease DpnI, and pooled whole genome amplification, which have
58
generated T. pallidum specific WGS data from over 700 metagenomic samples; however, specimens with low numbers of T. pallidum remains
59
challenging (11-16). Therefore, additional approaches that would enable sequencing of samples with low bacterial loads are needed.
60
Azithromycin has been used as an alternative to penicillin for treating early syphilis in the US; however, macrolide-resistant T. pallidum
61
strains, associated with two mutations (A2508G, A2509G) in the 23S rRNA genes, have been reported in many states (17-18). While macrolides
62
are no longer recommended for treatment of syphilis in the US, periodic monitoring is useful to determine the prevalence of resistant strains (19).
63
In this study, we describe a robust DNA enrichment method based on selective whole genome amplification (SWGA) using multiple
64
displacement amplification (MDA) and custom primers that enables WGS of clinical specimens with very low genomic copies of T. pallidum and
65
use of the sequence data for macrolide mutation analysis. We also investigated an alternative method that uses CpG methylated capture of host
66
DNA followed by MDA with random oligonucleotide primers.
67
Results
68
Real-time qPCR on clinical specimens and spiked samples. The T. pallidum PCR results for all clinical specimens are shown in Table 1. Out of
69
the 11 Emory specimens processed using the standard extraction protocol, only one specimen exceeded 100 genomic copies/µl based on polA
70
qPCR (Table 1). The remaining 10 specimens had an average copy number <1 copy/µl of DNA extract. These 11 specimens had an average
71
standardized RNP cycle threshold (RNPCt) value of 30.71 ± 0.13, and the lowest Ct value (highest concentration of RNP) was 25.22. Based on this
72
data, an RNPCt value of 25.22 was targeted as the cut-off for the spiked samples below.
73
NEBNext microbiome enrichment with MDA. The serially diluted spiked samples enriched with the NEB Microbiome Enrichment Kit with
74
subsequent REPLIg Single Cell MDA (hereafter referred to as NEB+MDA) showed a marked increase in polA copy number by qPCR (Fig. 1,
75
Table S1). The non-diluted samples indicated an average polA copy number of 6.67 x 106 ± 2.74 x 105 per µl of enriched DNA, which was 603.02
76
times greater than the input copy number. The 10-fold diluted samples also indicated increases in polA copy numbers, with an average of 7.85 x
77
105 ± 3.79 x 104, 1.28 x 105 ± 1.27 x 104, 8.66 x 103 ± 2.54 x 103, and 964 ± 574.23 copies/µl from 1:10 -1:10,000 dilution, respectively (Table S1).
78
This was a 482 – 995.09 times enrichment when compared to the input copy number. Upon comparing the average RNPCt of each dilution in the
79
series, the enriched samples indicated 29.28 ± 1.07, 31.15 ± 0.46, 30.25 ± 0.56, 31.08 ± 0.59, 31.42 ± 0.45 for the neat – 1:10,000 dilution,
80
respectively (Table S1). The RNPCt value of each enriched sample in the dilution series were insignificantly different from one another, with an
81
average RNPCt = 30.64 ± 0.33 for all dilutions in the series (P = 0.22).
82
After enriching with NEB+MDA, the average DNA percent for the neat to 1:10,000 dilutions indicated a range of 2.33% ± 0.10- 3.91 x
83
10-4 % ± 2.50 x 10-4% of the total DNA belonging to T. pallidum, respectively (Fig. 2). Further, this form of enrichment generated up to a 26.12-
84
fold increase in the percent of T. pallidum DNA, and an average of 16.27-fold ± 1.92-fold increase, amongst all enriched replicates when
85
compared to the unenriched input. All samples enriched by NEB+MDA were significantly different in their percent T. pallidum DNA when
86
compared to their respective inputs (P < 0.01). Apart from enriched samples from the 1:100 and 1:1,000 diluted polA inputs, we observed that by
87
increasing the polA input copy number 10-fold resulted in a significant increase in the total DNA belonging to T. pallidum post-enrichment (P =
88
0.06 and P < 0.05, respectively).
89
Genome sequencing data derived from samples enriched by NEB+MDA showed 0.01 – 10.52% of the quality-controlled reads binned as
90
T. pallidum, along with a mean mapping read depth to T. pallidum Nichols reference genome (NC_021490.2) ranging from 0.05 – 501.75. An
91
average percent coverage of 99.99%, 99.99%, and 97.29% across the Nichols reference genome with at least 5 reads mapped per site (5X) for the
92
neat, 1:10, and 1:100 diluted samples, respectively, was observed among the NEB+MDA enriched samples (Fig.3A; Table S1 and Fig. S1). The
93
coverage estimates indicated low deviations from this average in all replicates, with 2.92 x 10-4 % - 1.38% standard error between all replicates for
94
the neat – 1:100 diluted samples. At the same time, for a higher coverage of at least 10 reads mapped per nucleotide (10X), the 1:100 diluted
95
samples had an average percentage coverage of 84.14% while neat and 1:10 dilution samples were covered at 99.99% and 99.99% across the
96
reference genome, respectively. A sharp decline in coverage was observed in the 1:1,000 diluted samples, with a break down in replication at an
97
average coverage of 27.08% ± 18.62 % for the 1:1,000 dilution and 4.80% ± 0.63% for the 1:10,000 diluted samples at 5X read depth. With the
98
QC criteria for efficiency set at ≥90% at ≥5X read depth, samples sequenced post NEB+MDA enrichment had a limit of detection (LoD) of 129
99
polA copies/µl of extract (Fig. 3A; Table S1 and Fig. S1).
100
Post NEB+MDA enrichment of isolate CDC-SF003, we observed 2.39 x 106 ± 1.35 x 105 polA copies/µl of DNA extract. Further, 1.06%
101
of the total DNA belonged to T. pallidum post enrichment and 3.29% of the host removed quality-controlled sequencing reads were classified as T.
102
pallidum. Sequencing indicated a 98.60% coverage across the T. pallidum SS14 reference genome (NC_021508.1) at 5X read depth with a mean
103
mapping depth of 46.43 (Fig. 4; Table 2 and Fig. S2).
104
SWGA Enrichment of T. pallidum Nichols. A total of 12 primer sets were tested by SWGA using Equiphi29 MDA (Table S2-S3). The 1:100
105
diluted Nichols DNA sample (~129 copies/µl) was used to evaluate each of the 12 primers since it was comparable to specimen EUHM-004,
106
which had 106.7 polA copy/µl (Table 1; Table S4). Each of the primer sets indicated a 6.86 – 1.16 x 105 times enrichment when compared to the
107
input Nichols copy number (Fig. 5A). Further, we observed a >10,000-fold increase in polA copy number in samples enriched with 7 of the 12
108
primer sets (SWGA Pal 2, 4, 5, 9, 10, 11, and 12). SWGA Pal 9 and Pal 11 gave the highest enrichment at 1.13 x 105, and 1.16 x 105 times,
109
respectively (Table S4). The difference observed between Pal 9 and Pal 11 in the T. pallidum polA copy number and relative percent DNA
110
belonging to T. pallidum was insignificant; however, Pal 11 was selected for testing the SWGA limit of detection (P > 0.1; Fig. 5 and Table S4).
111
To determine the SWGA Pal 11 primer set’s LoD and enrichment for T. pallidum, SWGA was performed in triplicate on the 10-fold
112
dilution series. The ~1.11x104 copies/µl (neat) sample was eliminated from the dilution series, as this was ~100-fold increase in T. pallidum copy
113
number when compared to the clinical specimens tested. We observed a marked increase in polA copy number in every dilution in the series post
114
enrichment (Fig. 1; Table S1). The polA copy number ranged from 1.11 x 106 ± 6.68 x 105 for the 1:10,000 dilution to 2.04 x 107 ± 1.20 x 107 in
115
the 1:10 dilution (Table S1). When compared to the input polA copy number, this was a 2.01 x 104-fold, 1.19 x 105-fold, 3.53 x 105-fold, and 5.53
116
x 105-fold increase in the enriched samples, from 1:10 -1:10,000 dilution, respectively. Upon comparing the average RNPCt of each dilution in the
117
series, the SWGA enriched samples indicated a 29.36 ± 0.37 - 28.65 ± 0.16 for the 1:10 -1:10,000 dilution, respectively (Table S1). The average
118
RNPCt at each 10-fold increase in polA concentration were insignificantly different from one another (P > 0.1); however, by increasing the polA
119
input 100-fold, we observed a significant decrease in RNP concentration (P < 0.03).
120
After enriching with SWGA, we observed that dilutions ranging from 1:10 to 1:10,000 held 27.93% ± 1.57% - 3.29% ± 1.93% of the total
121
DNA belonging to T. pallidum, respectively (Fig. 2). This reflected up to a 1.63 x 105-fold increase in the relative T. pallidum and an average of
122
2.43 x 104-fold ± 1.05 x 104-fold increase amongst all replicate SWGA enriched samples when compared to the unenriched samples. All samples
123
were significantly increased in their relative T. pallidum DNA when compared to their respective inputs (P < 0.0001). While there was observed
124
deviations in the percent DNA between replicates, the 1:10,000 diluted replicates still yielded a 28.40-fold ± 17.71-fold increase in DNA
125
belonging to T. pallidum post SWGA when compared to the non-enriched neat dilution (P < 0.0001).
126
Genome sequencing data derived from the SWGA enriched Nichols samples showed 0.98%-78.05% of the quality-controlled reads binned
127
as T. pallidum, along with a mean mapping read depth to T. pallidum Nichols reference genome ranging from 65.82 – 4.89 x 103. An average
128
percent coverage of 98.67% ± 0.005%, 98.62% ± 0.003%, and 96.15% ± 0.082% across the Nichols genome at 5X read depth was observed
129
among the SWGA enriched 10-fold dilution series samples for the 1:10, 1:100 and 1:1,000 diluted samples, respectively (Fig. 3B; Table S1 and
130
Fig. S3). Further, coverage indicated low deviations from this average in all replicates, with a 0.0002% - 1.72% standard error between all
131
replicates for the 1:10 – 1:1,000 diluted samples. We did observe a sharp decline in coverage from the 1:1,000 to 1:10,000 dilution with an average
132
coverage of 38.46% ± 2.50% for the 1:10,000 diluted replicates a 5X read depth (Fig. 3B; Table S1 and Fig. S3).
133
Upon comparing the percent T. pallidum DNA derived from both enrichment methods, we observed that SWGA consistently produced
134
higher relative T. pallidum DNA in all samples (Fig. 2). We observed that the 10-fold dilutions enriched with SWGA exhibited an average of
135
94.08-fold - 1.41 x 104-fold increase in relative T. pallidum DNA in the 1:10-1:10,000 diluted samples when compared to the dilutions enriched by
136
NEB+MDA. All dilutions of each enrichment were significantly different from one another (P < 0.01), apart from the 1:10,000 and 1:1,000 diluted
137
samples enriched by SWGA and the neat diluted samples enriched by NEB+MDA (P > 0.07).
138
Comparing the sequencing data derived from the 1:10 and 1:100 diluted Nichols samples enriched using the NEB+MDA and SWGA, all
139
samples exhibited >95% coverage at 5X read depth (Fig. 3; Table S1 and Fig. S1, S3). There was a decline in coverage observed in the 1:1,000
140
diluted samples enriched by NEB+MDA, with an average coverage of 27.08% ± 24.80% at 5X read depth. This drop was not observed in the
141
1:1,000 diluted samples enriched by SWGA, which still held >95% coverage at 5X read depth. The 1:10,000 diluted samples enriched NEB+MDA
142
and SWGA exhibited <95% coverage at 5X read depth.
143
Enrichment of Clinical Strains.
144
Due to the increased sequencing coverage derived from the SWGA enriched Nichols strain, SWGA was chosen for enriching clinical
145
specimens with low numbers of T. pallidum (Fig. 3, Table S1). SWGA on clinical specimen EUHM-004 gave an average polA of 6.37 x 106 ± 2.24
146
x 105copies/µl with 5.56% of the total DNA belonging to T. pallidum (Table 2). Next generation sequencing using the MiSeq v2 (500 cycle)
147
platform revealed 95.13% coverage across the T. pallidum genome at 5X read depth (Fig. 4; Table 2 and Fig. S2). After large-scale DNA
148
extraction, we observed 31.5 ± 0.5, 122 ± 1.15, and 103 ± 6.55 polA copies/µl for specimens EUHM-012 – EUHM-014, respectively (Table 1).
149
For specimen EUHM-012, we observed an average polA of 2.14 x 106 ± 2.82 x 104 copies/µl with 1.72% of the total DNA belonging to T.
150
pallidum post-enrichment by SWGA (Table 2). Sequencing indicated a 93.98% coverage across the T. pallidum genome at 5X read depth (Fig. 4;
151
Table 2 and Fig. S2).
152
When compared to EUHM-012, EUHM-013 had a higher polA copy number at 5.16 x 106 ± 2.20 x 105 copies/µl with 15.48% of the total
153
DNA belonging to T. pallidum (Table 2). The sequencing data correlated with the qPCR data, indicating a 98.56% coverage across the T. pallidum
154
genome at 5X read depth (Fig. 4; Table 2 and Fig. S2). We also observed EUHM-014 held an increased polA copy number post-SWGA, with 2.57
155
x 106 ± 2.21 x 105copies/µl and 4.72% of the total DNA belonging to T. pallidum (Table 2). Upon sequencing, we observed 98.49% coverage
156
across the T. pallidum genome at 5X read depth (Fig. 4; Table 2 and Fig. S2). The polA copy number for specimen STLC-001 was 7.42 x 106 ±
157
7.20 x 105 copies/µl with 8.34% of the total DNA belonging to T. pallidum (Table 2). The sequencing coverage was 95.94% at 5X read depth
158
where 38.91% of the quality-controlled reads binned as T. pallidum along with a mean depth read coverage of 1,133.43X (Fig. 4; Table 2 and Fig.
159
S2).
160
Phylogenetic Analysis and Characterization of Genotypic Macrolide Resistance.
161
To analyze whether genomes generated from the 7 clinical specimens or isolates clustered to any of the two deep-branching monophyletic
162
T. pallidum lineages, Nichols-like and Street-14(SS14)-like, a whole genome phylogenetic tree was constructed using the genomes derived from
163
the clinical specimens/isolates along with 126 high quality published T. pallidum genome sequences as of May 2021 (12-15, 20-22; see Table S5,
164
methods in supplemental materials). Phylogenetic analysis revealed the presence of two dominant lineages, of which most strains belonged to the
165
SS14-like lineage. We identified a total of four monophyletic clades within this phylogenetic tree with ≥ 30 bootstrap support (Fig. 6). Three of the
166
clinically derived genomes from Atlanta, EUHM-004 (2019) EUHM-012 (2019), and EUHM-014 (2020), belonged to Nichols-like lineage (clade
167
1; n=12; Fig. 6). Interestingly, the other nine Nichols-like genomes in clade 1 were recent clinically derived genomes from Cuba (n=2; 2015-
168
2016), Australia (n=1; 2014), France (n=2; 2012-2013) and UK (n=3; 2016), and were distinct from the original Nichols strain isolated in 1912
169
and sent to different North American labs as in vivo derived clones, suggesting that we might not yet fully understand the current diversity of this
170
lineage. The three clinical specimens from Atlanta (EUHM-004, EUHM-012, and EUHM-014) and three clinically derived genomes from UK
171
isolated in 2016 (NL14, NL19 and NL17) carried the 23S rRNA A2058G mutation that confers macrolide resistance, suggesting a recent
172
acquisition of this antibiotic resistance variant in the Nichols-like lineage.
173
Even though previous phylogenomic analyses indicated that SS14-lineage showed a polyphyletic structure, our phylogenetic analysis with
174
a greater number of genomes showed the presence of 3 monophyletic clades (Clades 2, 3 and 4)(12, 14; Fig. 6). Clades 2 and 4 contained genomes
175
clustered within the previously reported SS14Ω-A sub cluster, which also contained two clades corresponding to the clades 2 and 4 detected in this
176
study, and contained genomes derived from Europe and North America; while clade 3 was similar to sub cluster SS14Ω-B and composed of
177
Chinese and North American derived T. pallidum genomes. The rabbit-derived clinical isolate, CDC-SF003 (San Francisco, U.S; 2017) sequenced
178
in this study, clustered within clade 2; while EUHM-013 (Atlanta, U.S; 2020) and STLC-001 (St. Louis, U.S; 2020) genomes clustered within
179
clade 4. Sequence analysis showed that all 3 strains carried the A2058G AMR variant for macrolide resistance. Macrolide resistance strains were
180
widespread among the SS14-lineage with higher proportion among the genomes in clades 2 and 3 compared to clade 4 genomes. The A2058G
181
point mutation identified in 4 patient specimens and isolate CDC-SF003 was verified by real-time PCR testing of genomic DNA and SWGA-
182
enriched samples (data not shown). There was inadequate sample for the fifth specimen to confirm the mutation by real-time PCR testing.
183
All the Nichols-like genomes derived from the NEB+MDA and SWGA 10-fold dilution series that contained T. pallidum reads mapped to
184
≥90% of the genome with at least 5X read depth formed a tight monophyletic clade (bootstrap support of 88/100) and clustered with the lab-
185
derived Nichols-Houston-J genome (bootstrap support of 100/100), indicating that genomes generated from both methods are adequate to capture
186
genetic variants required to perform a high resolution phylogenetic analysis (Fig. S4).
187
Discussion
188
WGS of T. pallidum is often challenging due to low bacterial loads or the difficulty of obtaining adequate samples for testing. In this
189
study, we sought to develop a method for performing WGS from rabbit propagated isolates and clinical specimens containing lower T. pallidum
190
numbers, leading us to investigate CpG capture and SWGA.
191
CpG capture has been successfully used for enriching bacterial genomic DNA in metagenomic samples (23-24), but this method has not
192
been used for T. pallidum. During our testing, we observed increases in polA copy numbers and relative T. pallidum percent DNA in the neat to
193
1:1,000 dilutions enriched by NEB+MDA when compared to the non-enriched inputs. Further, the results of the percent T. pallidum observed in
194
the enriched 1:10,000 diluted samples correlated with the decrease in overall coverage across the Nichols genome. Even though we observed an
195
increase in both polA copy number and relative percent T. pallidum DNA for the enriched diluted 1:1,000 diluted samples, we still only gained
196
~50% genomic coverage. This could be due to the remnant human DNA that was not initially captured prior to MDA, or the loss of T. pallidum
197
DNA during the enrichment. While there was no significant difference in the relative human RNP copy number from dilution to dilution, there is a
198
minimum T. pallidum copy number input required to outweigh the remnant human DNA during the metagenomic shotgun sequencing. Taking the
199
above into consideration, we observed that >129 polA copies/µl can generate >95% coverage at 5X read depth from the Nichols strain post
200
NEB+MDA. The results observed post NEB+MDA enrichment of clinical isolate CDC-SF003 correlated with the Nichols limit of detection
201
validation, with >98% coverage at 5X read depth across the T. pallidum genome. In silico variant analysis correlated with real-time PCR detection
202
of the mutations associated with macrolide resistance in clinical isolate CDC-SF003. Further, phylogenetics revealed that this strain belonged to
203
the SS14 lineage, which correlated with its enhanced CDC typing method (ECDCT) strain type, 4d9f, as previously reported (7). While this
204
enrichment method yielded good results with isolates, most clinical specimens collected in this study had lower than 100 polA DNA copies/µl of
205
T. pallidum leading us to consider an alternative method.
206
SWGA has been shown to be successful with other bacterial pathogens in metagenomic samples; however, it has not been investigated
207
with T. pallidum (25-27). We observed that samples enriched by SWGA using multiple primer sets exhibited a 10,000-fold increase in polA copy
208
number, with Pal 9 and 11 producing the highest relative percent T. pallidum DNA at 29% and 31%, respectively. While we chose to work with
209
Pal 11 as the optimal set, Pal 9 could also be a good alternative for enriching syphilis specimens. Further testing using Pal 11 showed that the limit
210
of detection was increased when compared to the T. pallidum enrichment obtained with NEB+MDA, with significant increases in both polA copy
211
number and percent T. pallidum across the 10-fold dilution series. Coverage across the T. pallidum genome exceeded 95% at 5X read depth for all
212
diluted samples, apart from the 1:10,000 diluted samples. Interestingly, we observed that increasing the input 100-fold resulted in a significant
213
decrease in the presence of RNP post-enrichment. Our data shows that >14 T. pallidum polA copies/µl can generate at least 95% coverage at 5X
214
read depth with the Nichols strain, which translated well to the clinical specimens tested. While there was a decrease in coverage in one of the
215
clinical specimens at 94.44% with 5X read depth when compared to the 98.62% coverage at 5X read depth observed in the 1:100 diluted Nichols
216
isolates, this could be primarily due to the improved capabilities of the NovaSeq 6000 when compared to the MiSeq v2 (500 cycle) platform used
217
to sequence this clinical specimen. Another possible reason for the variation in coverage could be due to the lower T. pallidum input copy number
218
in the clinical specimens.
219
The genomes derived directly from the 5 clinical specimens using SWGA were phylogenetically associated with the representative
220
lineages (either Nichols-like or SS14-like) and also provided high levels of within lineage strain resolution, which is ideal for effective tracking of
221
various strains circulating within a geographical area and outbreak investigations. In addition, the NGS methods described here can be used for
222
macrolide resistance marker detection. As observed with NEB+MDA enrichment, in silico azithromycin mutation detection performed on the
223
SWGA enriched specimens matched the results obtained with a real-time PCR, indicating that all clinical specimens contained the A2058G
224
mutation. SWGA-based enrichment also enabled sequencing of specimens within the range of detection limits for real-time PCR assays,
225
suggesting that our NGS workflow can be adapted for T. pallidum detection in metagenomic samples.
226
In terms of expense, both methods are cost-effective for enriching T. pallidum genomic DNA, and while SWGA is cheaper than
227
NEB+MDA, sequencing reagents are the true limiting factor for WGS. With the recent advancements in large-scale sequencing platforms, overall
228
sequencing costs can be further reduced. While NovaSeq 6000 has a much higher potential for multiplex sequencing, our data shows compatibility
229
of these enrichments for both NovaSeq 6000 and MiSeq platforms.
230
While we successfully enriched T. pallidum whole genomes in clinical specimens, the success of SWGA is limited by the constraint on
231
primer size, which may reduce the selectivity for the target genome. Phi29 functions best between 30-35°C, and ramp-down incubations have been
232
shown as an effective means of utilizing larger primers with increased melting temperatures (26-29). To help alleviate the constraints on primer
233
size, we utilized a thermostable phi29 mutant which has a much higher optimal temperature at 45°C (30) compared to the 30-35°C functional
234
range of the phi29 polymerase (26-27). This higher optimal temperature permits the use of longer oligonucleotides to be used in the SWGA
235
reaction, potentially increasing the selectivity for the T. pallidum genome. The phi29 mutant has also shown to be more efficient, with a 3-hour
236
exhaustion time when compared to the 8-16 hours required for the wild-type phi29 (30).
237
Our results show that SWGA is more sensitive, less cumbersome, and a faster method for enriching clinical specimens when compared to
238
NEB+MDA, allowing for WGS of metagenomic samples with very low numbers of T. pallidum. In addition, the sequencing data generated is of
239
sufficient quality to enable phylogenetic analyses and detection of mutations associated with azithromycin resistance. While the NEB+MDA was
240
unsuitable for the clinical specimens in this study, our data suggests that it can be used for samples exceeding 129 genomic copies/µl.
241
Materials and Methods.
242
Specimen collection, T. pallidum strains used for WGS, and real-time qPCR. Specimens used in this study were collected from men
243
presenting with lesions of primary or secondary syphilis to the Emory Infectious Diseases Clinic, Emory University Hospital Midtown (EUHM) in
244
Atlanta, GA and St Louis County STD Clinic (STLC) in St. Louis, MO (Table 1). Patients were diagnosed with syphilis based on clinical
245
presentation and serology testing. Fourteen swab specimens were collected in Aptima Multitest storage medium (Hologic, Inc., Marlborough, MA)
246
at Emory Infectious Diseases Clinic and 1 specimen at St. Louis County STD Clinic (Table 1). All specimens were stored at -80°C until shipment
247
on dry ice to the CDC. The T. pallidum Nichols reference strain was used for initial optimization and validation of the two enrichment methods. A
248
recent rabbit propagated isolate, CDC-SF003, was also included for testing (Table 1; 7). Prior to study commencement, local IRB approvals were
249
obtained from, Emory University, and St. Louis County Department of Public Health, and the project was approved at CDC
250
DNA was extracted from specimens and rabbit testis extracts using the QIAamp DNA Mini Kit (Qiagen, Germantown, MD) following the
251
manufacturer’s recommendations. Large-scale DNA extraction of three specimens was carried out on 1.5 ml of the Aptima stored specimen using
252
the QIAamp DNA Mini Kit following the manufacturer’s recommendations for upscaling with slight modifications (Table 1). Proteinase K was
253
added at 0.1X total sample volume, and AL Buffer and absolute ethanol were added at 1X total sample volume. Each sample was processed
254
through a single column, washed following the manufacturer’s recommendations, and eluted in 100 µl AE Buffer (Qiagen). Following DNA
255
extraction, each sample was tested by a real-time quantitative duplex PCR (qPCR) targeting the polA gene of T. pallidum and human RNase P
256
gene (RNP) using a Rotor-Gene 6000 instrument (Qiagen) as previously described with modifications (7; see additional methods in supplemental
257
materials).
258
Enrichment of T. pallidum by capture of CpG methylated host DNA and multiple displacement amplification (MDA). Initially, DNA
259
concentration of extracts from clinical specimens and rabbit propagated strains were measured using the Qubit dsDNA HS assay (Thermo Fisher
260
Scientific, Waltham, MA). Capture and removal of CpG methylated host DNA from samples were carried out using the NEBNext Microbiome
261
DNA Enrichment Kit following the manufacturer’s recommendations with modifications (New England Biolabs, Ipswich, MA). For all samples
262
tested, 250 ng of DNA was subjected to two rounds of bead capture using the NEBNext Microbiome DNA Enrichment Kit and enriched
263
treponemal genomic DNA was purified using AMPure XP beads (Beckman Coulter, Indianapolis, IN). Enriched DNA samples were stored at -
264
20°C until MDA was performed. MDA was carried out using the REPLI-g Single Cell Kit following the manufacturer’s recommendations with
265
slight modifications (Qiagen). Each MDA reaction was incubated at 30°C for 16 hr. Following amplification, the polymerase was inactivated at
266
65°C for 10 min, samples were purified with AMPure XP beads, and eluted with 100 µl 1X AE Buffer (Qiagen). For each enrichment using the
267
REPLI-g Single Cell Kit, non-template controls were included to confirm the absence of T. pallidum.
268
A 10-fold dilution series on the Nichols strain was used to determine the limit of detection (LoD) for enrichment (see supplemental
269
materials) with NEB+MDA followed by sequencing on an Illumina NovaSeq 6000. After DNA extraction, each dilution in the series was enriched
270
by NEB+MDA, genomic copy numbers estimated by polA qPCR, and sequencing performed in triplicate. Enriched samples were diluted 1:10
271
prior to measuring RNP amplification. The LoD was set at the minimal genome copy number required to generate a ≥5X read depth with ≥95%
272
genome coverage compared to the reference genome.
273
Selective whole genome amplification (SWGA) primer design, validation, and enrichment. Primers with an increased affinity to T. pallidum
274
were identified using the swga Toolkit as previously described with slight modifications (https://www.github.com/eclarke/swga; 26; see
275
supplemental materials). Eight primer sets (SWGA Pal 1-8), including 4 additional primer sets (SWGA Pal 9-12) generated by combining primers
276
in the initial set (Table S1), were chosen for SWGA using the EquiPhi29 DNA Polymerase (Thermo Fisher Scientific, Waltham, MA). To account
277
for the 3’-5’ exonuclease activity of the phi29 polymerase, all SWGA primers were generated with phosphorothioate bonds between the last two
278
nucleotides at the 3’ end (Table S1). Each of the 12 primer sets were tested in triplicate against the spiked sample diluted to an estimated 100 T.
279
pallidum polA copies/µl (see supplemental materials).
280
Prior to SWGA enrichment, samples were denatured for 5 min at 95°C after adding 2.5 µl of DNA to 2.5 µl reaction buffer, containing
281
custom primers, then placed immediately on ice until the Equiphi29 master mix, prepared as per manufacturer’s recommendations, was added
282
(Thermo Fisher Scientific, Waltham, MA). MDA was carried out following the manufacturer’s recommendations with modifications (Thermo
283
Fisher Scientific; 30). The reaction contained EquiPhi29 master mix, with EquiPhi29 Reaction Buffer at a final concentration of 1X, each primer
284
at a final concentration of 4 µM, and nuclease-free H2O was added to a final reaction volume of 20 µl. Reaction tubes were gently mixed by pulse
285
vortexing and incubated at 45°C for 3 hr. MDA was stopped by inactivating the DNA polymerase at 65°C for 15 min. All reactions were purified
286
using AMPure XP beads and eluted in 100 µl AE buffer (Qiagen). Non-template controls were included to confirm the absence of contaminate T.
287
pallidum DNA.
288
Relative percent T. pallidum in each sample was calculated as shown in Figure S1. SWGA Pal 11 was chosen for testing the LoD for
289
downstream genome sequencing post-SWGA enrichment using the 10-fold dilution series, excluding the undiluted (neat) spiked sample. All
290
enriched samples were validated by polA real-time qPCR in triplicate.
291
Sequencing and genome analysis of T. pallidum strains. Libraries were prepared using the NEBNext Ultra DNA Library Preparation Kit for
292
NovaSeq and NEBNext Ultra II FS DNA Library Preparation Kit for MiSeq sequencing following the manufacturer’s recommendations (New
293
England Biolabs, Ipswich, MA). For the validation experiments, sequencing was carried out on the Nichols reference strain using the Illumina
294
NovaSeq 6000 platform following the manufacturer’s recommendations (Illumina, San Diego, CA). Sequencing of isolate CDC-SF003 and swab
295
specimens were carried out using the MiSeq v2 (500 cycle) platform following the manufacturer’s recommendations (Illumina, San Diego, CA).
296
Post sequencing, reads were deduplicated, trimmed, and down selected for T. pallidum (supplemental materials). All down selected T.
297
pallidum reads were mapped to the T. pallidum reference genomes, and de novo assembled. Phylogenetic analyses were performed as described in
298
the supplemental materials. Apart from the genomes sequenced in this study, 122 high quality (with at least 5x read depth covering > 90% of the
299
genome) T. pallidum genomes deposited in the NCBI’s Sequencing Read Archive (SRA) under the BioProject number PRJEB20795 and
300
PRJNA508872 were also included (12, 14). The publicly available raw sequencing data were re-analyzed to determine the quality as described in
301
the supplemental materials. A second phylogenetic tree was also reconstructed by including all the genomes sequenced from the 10-fold dilution
302
series for both NEB+MDA and SWGA enriched samples. Genomic sequencing data from samples included in the phylogenetic analyses covered
303
at least 90% of the reference genome with 5X read depth. Variant calls for the A2058G and A2059G macrolide resistance mutations were
304
validated using a real-time PCR assay as previously described (31).
305
Statistical analyses. Statistical analyses were performed in R (R Foundation for Statistical Computing, Vienna, Austria) using the R companion
306
software RStudio (Rstudio, Boston, MA). Statistical significance was determined by analysis of variance (ANOVA) and Tukey post hoc multiple
307
comparisons tests. T. pallidum percent DNA were normalized through Log10 conversions. Quantitative data are presented as means ± standard
308
error. Differences were considered statistically significant if a P < 0.05.
309
Data availability. All sequencing data associated with this study were submitted to the National Center for Biotechnology Information’s sequence
310
read archive (SRA) under the BioProject accession ID PRJNA744275.
311
Acknowledgments
312
We thank Teressa Burns at the Emory University Hospital Midtown; Tamara Jones from the St Louis County STD Clinic; Yetty Fakile,
313
Kevin Pettus, and Jack Cartee at CDC’s Division of STD Prevention; The veterinary staff at the CDC’s Comparative Medicine Branch; Mark Itsko
314
at CDC’s Division of Bacterial Diseases; and Nikhat Sulaiman and Justin Lee at CDC’s Division of Scientific Resources for their assistance,
315
consults, and support throughout this study. This work was made possible through CDC’s Division of STD Prevention with support from the
316
Advanced Molecular Detection (AMD) program.
317
Author Contributions
318
Allan Pillay and Ellen N. Kersh conceived the study. Allan Pillay, Charles M. Thurlow, Cheng Chen, and Lilia Ganova-Raeva designed
319
the study. Charles M. Thurlow and Allan Pillay designed the enrichment protocols. Charles M. Thurlow designed the SWGA specific custom
320
primer sets used during this study and performed all enrichment experiments. Charles M. Thurlow, Allan Pillay, Samantha S. Katz, Lara Pereira,
321
Alyssa Debra, Kendra Vilfort, Yongcheng Sun, Kai-Hua Chi, and Damien Danavall performed the laboratory experiments and assisted with
322
specimen collection. Kimberly Workowski, Stephanie E. Cohen, Hilary Reno, and Susan S. Philip collected clinical specimens and patient data.
323
Mark Burroughs, Mili Sheth, and Charles M. Thurlow performed Illumina sequencing. Sandeep J. Joseph performed the bioinformatic analyses of
324
the genomic data, phylogenetic analysis and contributed to the generation of tables and figures. Charles M. Thurlow and Sandeep J. Joseph
325
performed data analysis. Charles M. Thurlow wrote and prepared the manuscript with oversight by Allan Pillay and contributions from Sandeep J.
326
Joseph and Weiping Cao, which was reviewed by all authors for revisions.
327
Disclaimer
328
The findings and conclusions in this report are those of the authors and do not necessarily represent the official position of the Centers for
329
Disease Control and Prevention. We declare that there are no competing interests.
330
References
331
1.
CDC. 2020. Sexually Transmitted Disease Surveillance 2019. US Department of Health and Human Services, Atlanta
332
2.
Marra C, Sahi S, Tantalo L, Godornes C, Reid T, Behets F, Rompalo A, Klausner JD, Yin Y, Mulcahy F, Golden MR, Centurion-Lara A,
333
Lukehart SA. 2010. Enhanced molecular typing of treponema pallidum: geographical distribution of strain types and association with
334
neurosyphilis. J Infect Dis 202:1380-8.
335
3.
Pillay A, Lee M-K, Slezak T, Katz SS, Sun Y, Chi K-H, Morshed M, Philip S, Ballard RC, Chen CY. 2019. Increased Discrimination of
336
Treponema pallidum Strains by Subtyping With a 4-Component System Incorporating a Mononucleotide Tandem Repeat in rpsA.
337
Sexually Transmitted Diseases 46:e42-e45.
338
4.
Katz K, Pillay A, Ahrens K, Kohn R, Hermanstyne K, Bernstein K, Ballard R, Klausner J. 2010. Molecular Epidemiology of Syphilis—
339
San Francisco, 2004–2007. Sexually Transmitted Diseases 37:660-3.
340
5.
Grillová L, Bawa T, Mikalová L, Gayet-Ageron A, Nieselt K, Strouhal M, Sednaoui P, Ferry T, Cavassini M, Lautenschlager S, Dutly F,
341
Pla-Díaz M, Krützen M, González-Candelas F, Bagheri HC, Šmajs D, Arora N, Bosshard PP. 2018. Molecular characterization of
342
Treponema pallidum subsp. pallidum in Switzerland and France with a new multilocus sequence typing scheme. PLoS One 13:e0200773.
343
6.
Edmondson DG, Wormser GP, Norris SJ. 2020. In Vitro Susceptibility of Treponema pallidum pallidum to Doxycycline. Antimicrobial
344
Agents and Chemotherapy 64:e00979-20.
345
7.
Pereira LE, Katz SS, Sun Y, Mills P, Taylor W, Atkins P, Thurlow CM, Chi K-H, Danavall D, Cook N, Ahmed T, Debra A, Philip S,
346
Cohen S, Workowski KA, Kersh E, Fakile Y, Chen CY, Pillay A. 2020. Successful isolation of Treponema pallidum strains from patients’
347
cryopreserved ulcer exudate using the rabbit model. PLOS ONE 15:e0227769.
348
8.
Bachmann NL, Rockett RJ, Timms VJ, Sintchenko V. 2018. Advances in Clinical Sample Preparation for Identification and
349
Characterization of Bacterial Pathogens Using Metagenomics. Frontiers in Public Health 6.
350
9.
Thorburn F, Bennett S, Modha S, Murdoch D, Gunson R, Murcia PR. 2015. The use of next generation sequencing in the diagnosis and
351
typing of respiratory infections. Journal of Clinical Virology 69:96-100.
352
10.
Lefterova MI, Suarez CJ, Banaei N, Pinsky BA. 2015. Next-Generation Sequencing for Infectious Disease Diagnosis and Management: A
353
Report of the Association for Molecular Pathology. The Journal of Molecular Diagnostics 17:623-634.
354
11.
Beale MA, Marks M, Cole MJ, Lee M-K, Pitt R, Ruis C, Balla E, Crucitti T, Ewens M, Fernández-Naval C, Grankvist A, Guiver M,
355
Kenyon CR, Khairulin R, Kularatne R, Arando M, Molini BJ, Obukhov A, Page EE, Petrovay F, Rietmeijer C, Rowley D, Shokoples S,
356
Smit E, Sweeney EL, Taiaroa G, Vera JH, Wennerås C, Whiley DM, Williamson DA, Hughes G, Naidu P, Unemo M, Krajden M,
357
Lukehart SA, Morshed MG, Fifer H, Thomson NR. 2021. Contemporary syphilis is characterised by rapid global spread of pandemic
358
Treponema pallidum lineages. medRxiv doi:10.1101/2021.03.25.21250180:2021.03.25.21250180.
359
12.
Beale MA, Marks M, Sahi SK, Tantalo LC, Nori AV, French P, Lukehart SA, Marra CM, Thomson NR. 2019. Genomic epidemiology of
360
syphilis reveals independent emergence of macrolide resistance across multiple circulating lineages. Nature Communications 10:3255.
361
13.
Pinto M, Borges V, Antelo M, Pinheiro M, Nunes A, Azevedo J, Borrego MJ, Mendonça J, Carpinteiro D, Vieira L, Gomes JP. 2016.
362
Genome-scale analysis of the non-cultivable Treponema pallidum reveals extensive within-patient genetic variation. Nature Microbiology
363
2:16190.
364
14.
Grillová L, Oppelt J, Mikalová L, Nováková M, Giacani L, Niesnerová A, Noda AA, Mechaly AE, Pospíšilová P, Čejková D, Grange PA,
365
Dupin N, Strnadel R, Chen M, Denham I, Arora N, Picardeau M, Weston C, Forsyth RA, Šmajs D. 2019. Directly Sequenced Genomes of
366
Contemporary Strains of Syphilis Reveal Recombination-Driven Diversity in Genes Encoding Predicted Surface-Exposed Antigens.
367
Frontiers in microbiology 10:1691-1691.
368
15.
Arora N, Schuenemann VJ, Jäger G, Peltzer A, Seitz A, Herbig A, Strouhal M, Grillová L, Sánchez-Busó L, Kühnert D, Bos KI, Davis
369
LR, Mikalová L, Bruisten S, Komericki P, French P, Grant PR, Pando MA, Vaulet LG, Fermepin MR, Martinez A, Centurion Lara A,
370
Giacani L, Norris SJ, Šmajs D, Bosshard PP, González-Candelas F, Nieselt K, Krause J, Bagheri HC. 2016. Origin of modern syphilis and
371
emergence of a pandemic Treponema pallidum cluster. Nature Microbiology 2:16245.
372
16.
Chen W, Šmajs D, Hu Y, Ke W, Pospíšilová P, Hawley KL, Caimano MJ, Radolf JD, Sena A, Tucker JD, Yang B, Juliano JJ, Zheng H,
373
Parr JB. 2021. Analysis of Treponema pallidum Strains From China Using Improved Methods for Whole-Genome Sequencing From
374
Primary Syphilis Chancres. The Journal of Infectious Diseases 223:848-853.
375
17.
Lukehart SA, Godornes C, Molini BJ, Sonnett P, Hopkins S, Mulcahy F, Engelman J, Mitchell SJ, Rompalo AM, Marra CM, Klausner
376
JD. 2004. Macrolide resistance in Treponema pallidum in the United States and Ireland. N Engl J Med 351:154-8.
377
18.
Chi KH, Danavall D, Taleo F, Pillay A, Ye T, Nachamkin E, Kool JL, Fegan D, Asiedu K, Vestergaard LS, Ballard RC, Chen CY. 2015.
378
Molecular differentiation of Treponema pallidum subspecies in skin ulceration clinically suspected as yaws in Vanuatu using real-time
379
multiplex PCR and serological methods. Am J Trop Med Hyg 92:134-8.
380
19.
Workowski KA, Bolan GA. 2015. Sexually transmitted diseases treatment guidelines, 2015. MMWR Recomm Rep 64:1-137.
381
20.
Čejková D, Zobaníková M, Chen L, Pospíšilová P, Strouhal M, Qin X, Mikalová L, Norris SJ, Muzny DM, Gibbs RA, Fulton LL,
382
Sodergren E, Weinstock GM, Šmajs D. 2012. Whole Genome Sequences of Three Treponema pallidum ssp. pertenue Strains: Yaws and
383
Syphilis Treponemes Differ in Less than 0.2% of the Genome Sequence. PLOS Neglected Tropical Diseases 6:e1471.
384
21.
Pětrošová H, Pospíšilová P, Strouhal M, Čejková D, Zobaníková M, Mikalová L, Sodergren E, Weinstock GM, Šmajs D. 2013.
385
Resequencing of Treponema pallidum ssp. pallidum Strains Nichols and SS14: Correction of Sequencing Errors Resulted in Increased
386
Separation of Syphilis Treponeme Subclusters. PLOS ONE 8:e74319.
387
22.
Sun J, Meng Z, Wu K, Liu B, Zhang S, Liu Y, Wang Y, Zheng H, Huang J, Zhou P. 2016. Tracing the origin of Treponema pallidum in
388
China using next-generation sequencing. Oncotarget 7.
389
23.
Feehery GR, Yigit E, Oyola SO, Langhorst BW, Schmidt VT, Stewart FJ, Dimalanta ET, Amaral-Zettler LA, Davis T, Quail MA, Pradhan
390
S. 2013. A method for selectively enriching microbial DNA from contaminating vertebrate host DNA. PLoS One 8:e76096.
391
24.
Thoendel M, Jeraldo PR, Greenwood-Quaintance KE, Yao JZ, Chia N, Hanssen AD, Abdel MP, Patel R. 2016. Comparison of microbial
392
DNA enrichment tools for metagenomic whole genome sequencing. Journal of Microbiological Methods 127:141-145.
393
25.
Leichty AR, Brisson D. 2014. Selective Whole Genome Amplification for Resequencing Target Microbial Species from Complex Natural
394
Samples. Genetics 198:473-481.
395
26.
Clarke EL, Sundararaman SA, Seifert SN, Bushman FD, Hahn BH, Brisson D. 2017. swga: a primer design toolkit for selective whole
396
genome amplification. Bioinformatics 33:2071-2077.
397
27.
Itsko M, Retchless AC, Joseph SJ, Norris Turner A, Bazan JA, Sadji AY, Ouédraogo-Traoré R, Wang X. 2020. Full Molecular Typing of
398
Neisseria meningitidis Directly from Clinical Specimens for Outbreak Investigation. Journal of Clinical Microbiology 58:e01780-20.
399
28.
Sundararaman SA, Plenderleith LJ, Liu W, Loy DE, Learn GH, Li Y, Shaw KS, Ayouba A, Peeters M, Speede S, Shaw GM, Bushman
400
FD, Brisson D, Rayner JC, Sharp PM, Hahn BH. 2016. Genomes of cryptic chimpanzee Plasmodium species reveal key evolutionary
401
events leading to human malaria. Nature Communications 7:11078.
402
29.
Cowell AA-O, Loy DE, Sundararaman SA, Valdivia H, Fisch K, Lescano AG, Baldeviano GC, Durand S, Gerbasi V, Sutherland CJ,
403
Nolder D, Vinetz JM, Hahn BH, Winzeler EA. 2017. Selective Whole-Genome Amplification Is a Robust Method That Enables Scalable
404
Whole-Genome Sequencing of Plasmodium vivax from Unprocessed Clinical Samples. mBio 8:e02257-16.
405
30.
Povilaitis T, Alzbutas G, Sukackaite R, Siurkus J, Skirgaila R. 2016. In vitro evolution of phi29 DNA polymerase using isothermal
406
compartmentalized self replication technique. Protein Eng Des Sel 29:617-628.
407
31.
Chen C-Y, Chi K-H, Pillay A, Nachamkin E, Su JR, Ballard RC. 2013. Detection of the A2058G and A2059G 23S rRNA Gene Point
408
Mutations Associated with Azithromycin Resistance in Treponema pallidum by Use of a TaqMan Real-Time Multiplex PCR Assay.
409
Journal of Clinical Microbiology 51:908-913.
410
411
Tables
412
413
Table 1. Clinical and laboratory data for specimens and clinical isolate CDC-SF003.
Sample/
Isolate ID
Collection
Year
Source
Gender
Sexual Status
Syphilis Stage
Site of
Lesion
Antibody
Titer (Assay)
qPCR
(T. pallidum polA in
DNA Extract)
Extraction
Method
Reference
CDC-SF003
2017
San Francisco
Male
MSM
Primary
Penis
1:4 (VDRL)
9, 680 copies/µl
Standard
Pereira et al., 2020
EUHM-001
2019
Atlanta
Male
MSM
Secondary
Neck
1:128 (RPR)
< 1 copy/µl
Standard
This study
EUHM-002
2019
Atlanta
Male
MSM
Secondary
Perianal
1:256 (RPR)
< 1 copy/µl
Standard
This study
EUHM-003
2019
Atlanta
Male
MSM
Secondary
Penis
1:32 (RPR)
< 1 copy/µl
Standard
This study
EUHM-004
2019
Atlanta
Male
MSM
Primary
Penis
1:4 (RPR)
106.7 ± 6.5 copies/µl
Standard
This study
EUHM-005
2019
Atlanta
Male
MSM
Secondary
Penis
1:64 (RPR)
< 1 copy/µl
Standard
This study
EUHM-006
2019
Atlanta
Male
MSM
Primary
Penis
1:16 (RPR)
< 1 copy/µl
Standard
This study
EUHM-007
2019
Atlanta
Male
MSM
Secondary
Hand
1:64 (RPR)
< 1 copy/µl
Standard
This study
EUHM-008
2019
Atlanta
Male
MSM
Secondary
Scrotum
1:64 (RPR)
0.9 ± 0.1 copy/µl
Standard
This study
EUHM-009
2019
Atlanta
Male
MSM
Secondary
Scrotum
1:64 (RPR)
< 1 copy/µl
Standard
This study
EUHM-010
2019
Atlanta
Male
MSM
Secondary
Scrotum
1:128 (RPR)
< 1 copy/µl
Standard
This study
EUHM-011
2019
Atlanta
Male
MSM
Primary
Penis
1:32 (RPR)
< 1 copy/µl
Standard
This study
EUHM-012
2019
Atlanta
Male
MSM
Primary
Penis
1:8 (RPR)
31.5 ± 0.5 copies/µl
Large
Scale
This study
EUHM-013
2020
Atlanta
Male
MSM
Secondary
Penis
1:64 (RPR)
122 ± 1.2 copies/µl
Large
Scale
This study
EUHM-014
2020
Atlanta
Male
MSM
Secondary
NA*
1:16 (RPR)
103 ± 6.7 copies/µl
Large
Scale
This study
STLC-001
2020
St. Louis
Male
MSW
Primary
Penis
NR** (RPR)
28.8 ± 3.1 copies/µl
Standard
This study
* Not available
** Non-reactive
Table 2. Sequencing percent coverage for the Nichols isolates, clinical isolate CDC-SF003, and clinical specimens across the T. pallidum reference genome.
Sample
Enrichment
method*
Clonal
complex
T.pallidum polA post
enrichment genome
copies/µl
Raw read
pairs
Non-host read
pairs
Total read
pairs after QC
Read pairs
classified as T.
pallidum
Percent of total
read pairs
classified as T.
pallidum
Mean
read
depth
Percent
genome
covered
≥1X
Percent
genome
covered
≥5X
Percent
genome
covered
≥10X
Nichols_CDC
non-enriched
Nichols-like
NA***
4,053,500
3,645,649
3,588,414
70,299
1.96
6.33
86.26
60.30
22.28
Nichols_CDC**
SWGA
Nichols-like
11,565,333 ± 1,294,672
3,701,303
3,692,932
3,648,044
3,414,111
93.59
751.17
98.39
98.24
98.16
CDC-SF003
NEB + MDA
SS14-like
2,394,930 ± 135,210
5,798,777
3,988,173
3,949,036
129,998
3.29
46.44
98.87
98.60
98.01
EUHM-004
SWGA
Nichols-like
6,367,089.5 ±240,811.5
6,102,826
4,440,618
4,280,401
1,403,645
32.79
370.39
96.99
95.13
92.67
EUHM-012
SWGA
Nichols-like
2,140,753 ± 28,192
10,350,274
5,870,287
5,716,082
2,793,693
48.87
639.86
96.34
93.98
91.89
EUHM-013
SWGA
SS14-like
5,159,716 ± 220,318.5
11,975,324
11,966,460
11,838,431
8,308,234
70.18
2,503.96
98.72
98.56
98.37
EUHM-014
SWGA
Nichols-like
2,573,508 ± 221,900.5
11,250,518
9,266,926
9,059,022
2,355,426
26.00
930.87
98.79
98.49
98.04
STLC-001
SWGA
SS14-like
7,420,534 ± 719,765
11,293,960
7,770,834
7,721,767
3,004,631
38.91
1,133.43
98.32
95.94
94.10
*All sequencing was performed using Illumina’s MiSeq v2 (500 cycle) platform
** Enrichment performed on 1,000 copies/µl T. pallidum polA input
*** Not available
Figures
414
Fig 1. T. pallidum polA copies/µl for the 10-fold dilution series spiked samples enriched by the NEBNext
415
Microbiome Enrichment Kit with REPLIg Single Cell MDA (NEB+MDA) or SWGA. The input T.
416
pallidum polA copies/µl for each dilution is displayed as Non-Enriched. The y-axis has been log10 scaled
417
for depiction of the Non-Enriched dilution series. Error bars represent standard error among three
418
replicate enriched T. pallidum samples.
419
Fig 2. Relative percent T. pallidum Nichols DNA for Non-Enriched, NEBNext Microbiome Enrichment
420
Kit with REPLIg Single Cell MDA (NEB+MDA), and SWGA enriched samples. Percent T. pallidum
421
DNA was calculated based on the input DNA concentration and polA copies/µl (Non-Enriched), and the
422
DNA concentration and polA copies/µl for the Nichols -spiked samples post-enrichment (NEB+MDA or
423
SWGA). The y-axis has been log10 scaled for depiction of the Non-Enriched dilution series. Error bars
424
represent standard error among three replicate samples.
425
Fig 3. Percent coverage of sequencing reads of enriched T. pallidum Nichols spiked samples. (A)
426
Sequencing reads of samples enriched using the NEB Microbiome Enrichment Kit and REPLIg Single
427
Cell MDA (NEB+MDA). (B) Sequencing reads of samples enriched using SWGA. All samples were
428
sequenced using the Illumina NovaSeq 6000 platform. Error bars represent standard error between the
429
mapped reads derived from three replicate enriched Nichols samples.
430
Fig 4. Percent coverage of isolates and clinical specimens. All samples were sequenced using the Illumina
431
MiSeq v2 (500 cycle) platform. Percent of T. pallidum reads are derived from down selected T. pallidum
432
reads. Prefiltered reads for Nichols-CDC were mapped to the Nichols reference genome (NC_000919.1).
433
The prefiltered reads in all clinical isolates and specimens were mapped against the SS14 reference
434
genome (NC_021508.1).
435
Fig 5. SWGA primer set validation. (A) T. pallidum polA copies/µl for the Nichols mock sample (1:100
436
diluted) enriched with each SWGA primer set. (B) Relative percent T. pallidum DNA for the Nichols
437
spiked sample (1:100 dilution) enriched with each SWGA primer set. Percent T. pallidum DNA was
438
calculated based on the input DNA concentration and polA copies/µl for the Nichols mock samples post-
439
SWGA enrichment. The y-axis has been log10 scaled for depiction of the relative percent T. pallidum
440
post-enrichment with each primer set. Error bars represent standard error among three replicate Nichols
441
samples.
442
Fig 6. Maximum likelihood global phylogenetic tree of the clinical isolate/specimen genome sequenced in
443
this study along with publicly available T. pallidum genomes. The two major lineages, Nichols-like and
444
SS14-like are highlighted along with presence of genotypic mutation responsible for macrolide resistance
445
and country of origin.
446
447
0
10
20
30
40
50
60
70
80
90
100
Non-Diluted
1:10
1:100
1:1,000
1:10,000
Percent Coverage of T. pallidum Nichols
Genome
Input T. pallidum Dilution Factor
1X
5X
10X
0
10
20
30
40
50
60
70
80
90
100
1:10
1:100
1:1,000
1:10,000
Percent Coverage of T. pallidum Nichols
Genome
Input T. pallidum Dilution Factor
1X
5X
10X
0
10
20
30
40
50
60
70
80
90
100
T. pallidum
Nichols
CDC-SF003
EUHM-004
EUHM-012
EUHM-013
EUHM-014
STLC-001
Percent Coverage of T. pallidum Genome
T. pallidum Strain
1X
5X
10X
Percent T.
pallidum reads
Percent
T. pallidum
Reads
0E+00
2E+06
4E+06
6E+06
8E+06
1E+07
1E+07
1E+07
2E+07
2E+07
Final T. pallidum polA Copies/µL Extract
SWGA Primer Set
A.
0.001
0.010
0.100
1.000
10.000
100.000
T.pallidum DNA Ratio (Percent)
SWGA Primer Set
B.
s oN NRREOT SENET EEES RRR AR RR RR NS FN RR RE FOTN TERE OURO FREER REO eRe RD Nichols Houston _E
Cnn ene eee eee eee eee eee eee eee eee eee e eens Nichols Houston O
) SESRSEOH | BREEOGe dU eeNGEES fs Oeedend ss SoeGREe Vt OSeGuRe s HeGEEES Ce banGeES (| necmuEED s SBeeeeE D eeaueERs ¢ HEaRUEES f eoeREEES + CoeRSENG s SeaaeERG | eouneEus s ouL Nichols Houston J
eee eee eee eee eee eee eee eee eee eee eee eee eee eee eee ee eee ee eee ee eee eee eens erieage ty tube2
Lineage
a Nichols
WM ssi4
Genotypic macrolide
Resistance
7 Resistant
] Sensitive
Uncertain
Lee eee eee ee eee eee eee e eee e eee e eee e eee e eee e eee e eee e esse eeeaeeeenneeeaees CW
erent ttt tttttertteterereeseses NEV Clade 1
CW86
Lene e eee e cece eee eee e eee e ete ee eee eee eee eeeeeeeteeeeeeteneeeeseneeeteaes EUHM-012
bee e eee eee eee e ee eee eee e ee eee e eee n ee eee ee een e eee e ee ee ee een e eee e cece neste eee cence eee e eee e ence eect e eee n cece eee e anette geet ence eae e cence eee eeeeeeeeen essen eeeeeeeeeeeeeeeesteneeeeeeeeuneeeneeeegueeeeees Chicago _ Population
wee Nichols | v2
DAL-
ee ee ee ee eee eee eee ee eee ees Nicho aNich
BA
Country
a Austria
a China
Czech Republic
Finland
Mexico
Netherlands
Portugal
Switerland
on CW30 Clade 2
Leeeeee W074B
-- PT _SIF1261
reese UW259B
---» UW304B
-- UW526B
- NL
Cc
~N
USA
France
Australia
-C
- UW852B cuba
- UWW824B
-- PT _SIF1140
-NLT4
- CW84
wee eee eee eee Reference
VY)
Oo
c
=
oO
@
This_ Study
NCBI
SHE-V Clade 3
beeen PT SIF1063
.-» NLT6
---- UW492B
--- STLC-001
-- UW148B
--+ UW473B
--+- PT SIFO857
-- UW195B
: UW244B Clade 4
Lineag
23S mt
Countr
eee ewww eee e eee eeeeennuwnnes Nichols_Houston_E
ee ee ee eee Nichols Houston_O
wee eee eee eee weet eee eee eee Nichols Houston_J
eee eryeago tube2
CASE URONTUAM SE SONOS ROMs BORA e eS EUHM-014
Lineage
a Nichols
WM ssi4
Genotypic macrolide
Resistance
7 Resistant
] Sensitive
Uncertain
W
CII Abd Clade 1
NZ
©
4
®
Co
ome
i
6
Sues ETRE eRe Rees HARE EUHM-012
eee es eeage Popuralton
eee a ESTE 6 TR Tew bE CDC-
ee Nichols_v2
CTKS s PARA STA 6 CODEC TO Ss ESR 8 ws DAL-1
ee ee ees Seattle Nichols
ee eee ees Nic ols-CDC
BAL73
Country
a Austria
a China
Czech Republic
Finland
Mexico
wee eee eee PT SIF0O877 3
Netherlands
Portugal
Switerland
wo CW30 Clade 2
beens 074B
-- PT SIF1261
sone UW259B
--+- UW304B
-- UW526B
- NL10
Cc
~N
USA
France
beens W.
-- UW383B
- CW56
- UW852B
- UW824B
-- PT_SIF1140
-NLT1
Australia
Cuba
V~)
Oo
c
=
oO
@
- CW84
seen | GEMEREEE SO GEE Reference
V
vee eee ee eeeeeees PT STF0908
we BIESIF 1002 This_ Study
NCBI
ST SHE-V Clade 3
cseRRNOS PRE Gra y
ween eee eee Philadephia-1
....+-+» PT SIF1063
--- NLT
. UW244B Clade 4
Lesa UW102B
1.00E+00
1.00E+01
1.00E+02
1.00E+03
1.00E+04
1.00E+05
1.00E+06
1.00E+07
1.00E+08
Non-Diluted
1:10
1:100
1:1,000
1:10,000
Final T. pallidum polA Copies/µL Extract
Input T. pallidum Dilution Factor
Non-Enriched
NEB+MDA
SWGA
0.00001
0.00010
0.00100
0.01000
0.10000
1.00000
10.00000
100.00000
Non-Diluted
1:10
1:100
1:1,000
1:10,000
T. pallidum DNA Ratio (Percent)
Input T. pallidum Dilution Factor
Non-Enriched
NEB+MDA
SWGA
| 2021 | Selective whole genome amplification as a tool to enrich specimens with low genomic DNA copies for whole genome sequencing | 10.1101/2021.07.09.451864 | [
"Thurlow Charles M.",
"Joseph Sandeep J.",
"Ganova-Raeva Lilia",
"Katz Samantha S.",
"Pereira Lara",
"Chen Cheng",
"Debra Alyssa",
"Vilfort Kendra",
"Workowski Kimberly",
"Cohen Stephanie E.",
"Reno Hilary",
"Sun Yongcheng",
"Burroughs Mark",
"Sheth Mili",
"Chi Kai-Hua",
"Danavall Dami... | creative-commons |
Page 1 sur 22
Aquatic long-term persistence of Francisella tularensis ssp. holarctica is
1
driven by water temperature and transition to a viable but non-culturable
2
state
3
Camille D. Brunet1, Julien Peyroux1,2, Léa Pondérand3,4, Stéphanie Bouillot3, Thomas Girard4,
4
Éric Faudry3, Max Maurin1,4, Yvan Caspar3,4*
5
6
1 Univ. Grenoble Alpes, CNRS, Grenoble INP, TIMC, 38000 Grenoble, France.
7
2 Laboratoire d’Informatique de Grenoble, Bâtiment IMAG, 38401 Saint Martin d’Hères
8
3 Univ. Grenoble Alpes, CEA, CNRS, IBS, 38000 Grenoble, France
9
4 Centre National de Référence des Francisella, CHU Grenoble Alpes, 38000 Grenoble, France.
10
11
12
* Corresponding author:
13
Dr Yvan Caspar, CHU Grenoble Alpes, Institut de Biologie et de Pathologie, CS10217, 38043,
14
15 Grenoble, Cedex 9, France, ycaspar@chu-grenoble.fr
15
16
Page 2 sur 22
Abstract
17
Francisella tularensis is a highly virulent bacterium causing tularemia zoonosis. An increasing
18
proportion of infections occur through contaminated hydro-telluric sources, especially for the
19
subspecies holarctica (Fth). Although this bacterium has been detected in several aquatic
20
environments, the mechanisms of its long-term persistence in water are not yet elucidated. We
21
evaluated the culturability and the viability of a virulent Fth strain in independent microcosms
22
filled with nutrient-poor water. At 37°C, the bacteria remained culturable for only one week,
23
while culturability was extended to 6 weeks at 18°C and up to 11 weeks at 4°C. However, while
24
the viability of the bacteria declined similarly to culturability at 37°C, the viability of the
25
bacteria remained stable overtime at 18°C and 4°C for more than 24 months, long after loss of
26
culturability. We identified water temperature as one of the major factors driving the aquatic
27
survival of Fth through a transition of the whole Fth population in a viable but non-culturable
28
(VBNC) state. Low temperature of water (≤18°C) favors the persistence of the bacteria in a
29
VBNC state, while a temperature above 30°C kills culturable and VBNC Fth bacteria. These
30
findings provide new insights into the environmental cycle of Francisella tularensis that
31
suggest that the yet unidentified primary reservoir of the subspecies holarctica may be the
32
aquatic environment itself in which the bacteria could persist for months or years without the
33
need for a host.
34
35
Keywords
36
Francisella tularensis, tularemia, viable but non-culturable, dormancy, water microbiology
37
38
Page 3 sur 22
Introduction
39
Francisella tularensis is a Gram-negative bacterium causing the zoonosis tularemia. It is a
40
highly virulent human pathogen classified in category A of potential agents of biological threat
41
by the US Centers for Disease Control and Prevention [1]. Two subspecies are associated with
42
human tularemia: F. tularensis ssp. tularensis (Ftt) (type A strains), only present in North
43
America; and F. tularensis ssp. holarctica (Fth) (type B strains), spread all over the Northern
44
Hemisphere, with a few strains identified in the last decade in Australia [1,2].
45
Terrestrial and aquatic lifecycles of F. tularensis have been described but remain not fully
46
characterized despite many decades of research [3]. Especially, the survival of the bacteria in
47
hydro-telluric environments is still under active investigation [4,5]. The terrestrial animal
48
reservoir of F. tularensis is large, but lagomorphs and small rodents are considered primary
49
sources of human infections. Recent data corroborate that the aquatic lifecycle of the subspecies
50
Fth may be predominant over the terrestrial lifecycle, in particular for the persistence of the
51
disease in the environment, as initially suggested by Jellison [5,6]. This aquatic cycle involves
52
mainly mosquitoes, mosquito larvae, and aquatic rodents [3]. In Northern Europe, mosquitos
53
can transmit Fth after larva contamination in water and consequently be responsible for large
54
outbreaks [7–10]. Cases of tularemia related to water have also been described after an aquatic
55
activity (e.g., swimming or canyoning) [11,12] or through drinking or using contaminated water
56
[13,14].
57
Some studies suggested the potential persistence of this bacterium in aquatic environments over
58
long periods. Genomic studies have confirmed that diverse clones of Fth survive for a
59
prolonged period and that a single clone may be responsible for human or animal cases of
60
tularemia over several decades (up to 70 years) [15–17]. Multiple independent respiratory
61
infections with Fth strains acquired from the environment over a short period were observed
62
during an outbreak in Sweden in 2010 and in France in 2018, arguing in favor of environmental
63
Page 4 sur 22
changes acting as the trigger of these outbreaks [17,18]. Analysis of exposition factors
64
suggested environmental contamination, presumably through aerosols originating from an
65
unidentified environmental reservoir [5]. Low temperature and salinity have been identified to
66
impact the duration of culturability of Fth. It has been described that this bacterium can remain
67
culturable up to 70 days at 8°C [19,20], ten days in fresh water at room temperature, 21 days in
68
seawater, and 45 days in brackish water [21]. Recently while studying biofilm formation of F.
69
tularensis in aquatic environments, Golovliov et al. identified that Fth remained culturable and
70
infectious in a mice model after 24 weeks of incubation at 4°C in low nutrient water containing
71
9 g/L of NaCl. They suggested that this improved survival at low temperature in freshwater
72
may be a critical mechanism to help the bacteria overwinter and survive between host-
73
associated replication events [4]. In such situations the bacteria may choose to switch to a
74
dormancy state that reduces competition with actively growing cells. Among potential
75
persistence and/or quiescence mechanisms identified in bacteria, bacterial switch to a viable
76
but non-culturable (VBNC) state that has been poorly studied in virulent F. tularensis strains
77
[20,22]. Initially described in 1982 for Escherichia coli and Vibrio cholerae [23], the VBNC
78
state corresponds to bacteria that lose their ability to grow, may change their shape and lose
79
their virulent traits, although remaining still alive. The VBNC state is induced during a stress
80
such as nutrient starvation, physicochemical changes of the environment, or thermal shock. It
81
has already been identified that the virulence-attenuated live vaccine strain (LVS) of Fth is able
82
to survive in a VBNC state at least 140 days at 8°C [20]. Survival of a fluorescent Fth strain in
83
a VBNC state up to 38 days has also been described in the control conditions of a co-culture
84
experiment with protozoan using a gfp-modified Fth strain [22].
85
Consequently, our goal was to investigate the role of water temperature and salinity on the
86
persistence of a virulent human strain of Fth in water and explore the possibility of a transition
87
of Fth into a VBNC state triggered by these factors.
88
Page 5 sur 22
89
Material and methods
90
Bacterial strains and preparation of aquatic microcosms
91
All culture assays were performed in a BSL3 laboratory. We used the fully virulent Fth biovar
92
I clinical strain CHUGA-Ft6 (genome accession: VJBK00000000) [15]. This strain was grown
93
on Polyvitex-enriched chocolate agar plates (PVX, BioMérieux, Marcy l’Etoile, France)
94
incubated at 37°C in a 5% CO2-enriched atmosphere. The F. tularensis collection of French
95
National Reference Center for Francisella is approved by the Agence Nationale de Sécurité du
96
Médicament et des produits de santé (France) (ANSM, authorization number ADE-103892019-
97
7).
98
Six independent aquatic microcosms were defined, consisting of 6 aliquots of the same
99
environmental water sample from the Rhône-Alpes region in France (send for analysis in the
100
water laboratory Abiolab-Asposan, Monbonnot-Saint-Martin, France; Table S1). Microcosms
101
were incubated at 4°C, 18°C, and 37°C, and supplemented with either 0 or 10 g/L of NaCl.
102
Each condition was tested in biological triplicate. Bacterial suspensions were prepared in PBS
103
and adjusted to 109 CFU.ml-1, and 25 mL were added to 225 mL of environmental water
104
previously sterilized using a 0,22µm filter.
105
Monitoring of culturability and viability of bacteria in water
106
The culturability and viability of bacteria in the six environmental models were monitored each
107
week. The culturability was measured by CFU counts after plating 100µL of serials dilutions
108
of each microcosm and on PVX agar plates after 48h incubation at 37°C. The viability of the
109
bacteria was determined using qPCR after PMAxx™ Dye treatment (Biotium San Francisco,
110
US) allowing specific DNA amplification of viable bacteria only. In brief, 1 mL of bacteria
111
suspension was added to 250 µL of enhancer for Gram-negative Bacteria (Biotium San
112
Page 6 sur 22
Francisco, US). PMAxx™ Dye was dissolved in H2O at 5mM and added to a bacterial solution
113
at a final concentration of 25µM. After 10 min of incubation in the dark, samples were exposed
114
30 min to light with GloPlateTM Blue LED Illuminator (Biotum, San Francisco, US). Bacterial-
115
PMA suspensions were centrifugated at 11,000g for 10 min, and DNA was extracted using
116
NucleoSpin Blood Kit (Macherey Nagel, Hoerdt, France) according to manufacturers’
117
recommendations. At each sampling point, DNA extraction of 1 mL of bacterial suspension
118
without PMA treatment was realized in parallel to determine amplification of total DNA present
119
in the samples. At each sampling point, control with dead bacteria for PMAxx™ Dye was
120
realized using a 1 mL suspension of bacteria previously lysed. Each qPCR reaction contained
121
10µL of Master Mix EvaGreen 2X (Biotium, San Francisco, US), 1µM of each primer, 5µL of
122
DNA template, and 1µL of sterile H2O. The 23S ribosomal RNA gene was amplified using the
123
following
forward
(5’-CATACGAACAAGTAGGACGG-3’)
and
reverse
(5’-
124
GCAAGCGGTTTCAGATTCTA-3’). The qPCR was performed using a LightCycler 480
125
instrument (Roche, Meylan, France) and SYBRGreen channel, with the following protocol:
126
initial denaturation at 95°C for 5 min, followed by 40 cycles of 95°C for 5s and 60°C for 30s.
127
Melting curve analysis was performed from 57°C to 99°C. A negative control (H2O) was
128
included in each qPCR run. The viability was evaluated by the Cycle threshold (Ct) of DNA
129
amplification of living Bacteria. Statistical analyses were performed by student t-test.
130
To compare the shape and the length of VBNC and culturable bacteria, pictures of bacteria
131
incubated in water at 18°C without NaCl one hour for culturable bacteria, and 6 months for
132
VBNC bacteria labeled with Syto9 were analyzed by Image J software and Microbe J plugging
133
[24]. Bacterial morphology was described by parameters: area (0.1-1.2µm); length (0.2-1.4
134
µm); width (0-1.4 µm); circularity (0.3-max µm). Parameters were calculated for 1055 bacteria
135
in each sample.
136
Bacterial viability of VBNC cells after temperature change of microcosm
137
Page 7 sur 22
Several months after the loss of culturability of Fth in water, 5 mL of microcosm at 4°C were
138
transferred at 18°C, 30°C and 37°C. 5 mL of microcosm at 18°C were transferred at 4°C, 30°C
139
and 37°C. After 7 and 14 days, the viability was evaluated by qPCR after PMAxx™ Dye
140
treatment. Statistical analyses were performed using R (version 4.0.3) for the comparison of
141
multiple groups by one-way ANOVA. False discovery rate (FDR) correction was applied for
142
pairwise t-tests.
143
Biofilm quantification
144
Biofilm quantification was performed on the bacterial suspensions evolved in nutrient-poor
145
water in T75 culture flasks, one year at 4°C, six months at 18°C, or four months at 37°C after
146
inoculation. Negative control consisted in fresh culturable bacteria incubated for one hour in
147
water. Each condition was tested in a biological duplicate. After incubation, the culture medium
148
was aspirated, and flasks were washed three times with PBS. 5 mL of Crystal violet (0,2% w/v)
149
were added, and flasks were re-incubated for 15 minutes. Crystal violet was washed three times
150
with PBS, and biofilm was solubilized by 1 mL of ethanol 95%. 200µL of this biofilm was
151
added to microtiter plates in three wells, and biofilm was quantified by measuring absorbance
152
at 570nm. Statistical analyses were performed by student t-test.
153
154
Results
155
Culturability of F. tularensis ssp. holarctica extends to 11 weeks at low temperature
156
In low nutrient-containing water, at 37°C, the culturability of the virulent clinical strain of Fth
157
biovar I decreased from 108 to 0 CFU/mL in 8 days (Figure 1a). However, the culturability of
158
bacteria extended dramatically when reducing the temperature of water microcosms. At 18°C,
159
bacteria decreased from 108 to 0 CFU/mL in 6 weeks (Figure 1b). At 4°C, culturability of
160
bacteria declined even more slowly with a complete absence of growing colonies only 11 weeks
161
Page 8 sur 22
after inoculation of the water sample (Figure 1c). At 37°C, NaCl concentration enrichment of
162
the microcosm at 10g/L conferred only a slight transient survival advantage to the bacteria
163
(Figure 1a). No significant differences were observed at 18°C or 4°C.
164
F. tularensis ssp. holarctica switched to VBNC state at low temperature in nutrient-poor water
165
The virulent clinical strain of Fth biovar I did not survive for more than eight days at 37°C in
166
nutrient-poor water. In this microcosm, viability was correlated with culturability. qPCR-PMA
167
Ct value increased from 12.1±0.5 to 25.8 ±0.2 after eight days in nutrient-poor water without
168
NaCl showing a strong reduction of viable bacteria in this microcosm. In comparison, for each
169
condition the Ct value of the controls with dead bacteria, i.e., lysed and PMAxx™ Dye treated
170
bacteria, was 22.7±3.9. On the opposite, while the culturability declined, almost all bacteria
171
remained alive during the eight-week study at 18°C and the 14 weeks study at 4°C (Figures 1e
172
and 1f). The Ct value of viable bacteria stayed stable at 12.5 ±1.3 for all four conditions during
173
the whole experiment and for more than two weeks after the loss of culturability (Water at 4°C
174
without NaCl, Ct range: 11.8-13.8. Water at 4°C with NaCl, Ct range: 11-13.5. Water at 18°C
175
without NaCl, Ct range: 12-13.6. Water at 18°C with NaCl, Ct range: 10.4-13.7). Two replicates
176
were kept in the water for 24 months and tested again. Interestingly, the Ct of PMA-qPCR
177
remained unchanged (13.2 and 14.1). Thus, we observed that roughly the full initial bacterial
178
inoculum switched at low temperature to viable but non-culturable state corresponding to the
179
definition of transition into VBNC state. It is interesting to note that the temperatures of 4°C
180
and 18°C differentially affected culturability but not viability. Viability of the bacteria in the
181
microcosms at 4 and 18°C was confirmed by the Live/Dead® BacLight™ assay (Figure S1).
182
The addition of 10g/L NaCl conferred a slight transient survival advantage to the bacteria at
183
37°C. On the fourth day, there was a 4-log difference (p-value = 0.0005) between the two
184
conditions but qPCR-PMA Ct value also increased from 11.4±1.1 to 22.6 ±0.4 showing that all
185
the bacteria were dead in eight days in both conditions (Figure 1d). The addition of salt to the
186
Page 9 sur 22
microcosm did not significantly affect the culturability and viability of the bacteria at 4°C and
187
18°C (p-value > 0,05 for each time points, Figures 1b,c,e,f).
188
To visualize bacterial morphology after transition into VBNC state, fresh bacteria suspended
189
one hour in water and VBNC bacteria sampled five months after the loss of culturability were
190
labeled with an anti- F. tularensis LPS antibody and observed with oil immersion objective
191
100X. After the loss of culturability the anti-LPS antibody was still able to bind to the LPS of
192
Fth strain and microscopic examination suggested a reduced length of VBNC bacteria (Figure
193
S2). Modification of the size of the bacteria was confirmed by Syto9 staining and image analysis
194
that showed that VBNC bacteria were smaller than culturable Fth with respectively an area of
195
0.31 ±0.19 µm² and 0.47 ±0.27 µm²; a length of 0.62 ±0.22 µm and 0.76 ±0.26 µm; a perimeter
196
of 1.85±0.63 µm and 2.29±0.75 µm (p-value <0.0001 for each parameters). However,
197
circularity was not statistically different (0.97±0.03 for both culturable and VBNC Fth samples;
198
p-value = 0.47) (Figure S3).
199
High water temperatures inactivated F. tularensis ssp. holarctica VBNC bacteria
200
After their transition into the VBNC state, the viability of the bacteria was still dependent on
201
the temperature of the water. Several months after the loss of culturability, when VBNC bacteria
202
were moved from 4°C to 18°C and vice versa, the temperature change did not influence the
203
viability of the bacteria during after 14 days of incubation (4°C to 18°C: Ct value from 13.8±0.8
204
to 14.3±2.5; 18°C to 4°C: Ct value from 12.7±1.1 to 15.3 ±3.7). However, when the temperature
205
was shifted to 30°C, the viability of VBNC bacteria significantly declined in 14 days (4°C to
206
30°C: Ct value from 13.8±0.8 to 20±1.5; 18°C to 30°C: Ct value from 12.7±1.1 to 20±6) (p-
207
value < 0.05). Moreover, when placed at 37°C, viability of VBNC bacteria declined in 14 days
208
under the threshold corresponding to dead bacteria only (4°C to 37°C: Ct value from 13.8±0.8
209
to 23.8±2.5; 18°C to 37°C: Ct value from 12.7±1.1 to 26.3±3.3) (p-value <0.05) (Figure 2).
210
Page 10 sur 22
Virulent F. tularensis ssp. holarctica strain was able to form biofilm in water
211
Optical density at 570 nm of the water flasks containing fresh bacteria after crystal violet
212
staining was 0.14±0.01 while optical density of the flasks containing the Fth VBNC bacteria
213
were increased twofold: 0.27±0.01 for VBNC bacteria after one year at 4°C (p-value = 0.019)
214
and 0.3±0.01 (p-value < 0.0001) for VBNC bacteria after six months at 18°C. On the opposite,
215
optical density of the flasks containing dead Fth bacteria (4 months at 37°C) was 0.15±0.03
216
showing no significant biofilm production compared to fresh bacteria (p-value = 0.7) (Figure
217
3). Microscopic observation of the stained flasks showed small Gram-negative coccobacilli
218
embedded and surrounded by a structure resembling a biofilm (Figure S4).
219
220
Discussion
221
Although the presence and potential survival of Fth in the aquatic environment have been
222
identified in several studies [25–29], the mechanisms of its persistence and its precise
223
environmental reservoir remain unclear. According to current descriptions of the aquatic cycle
224
of Fth, aquatic environments may be initially contaminated by F. tularensis through dead
225
animals or excrements of infected animals [3]. However, how these bacteria can persist for
226
weeks or even years within these environments remains to be elucidated. Following recent work
227
showing extended culturability of Fth at 4°C in water, we hypothesized that in environmental
228
water, this bacterium might also survive in a dormancy form such as the VBNC state. This
229
hypothesis would help the bacteria to survive in hostile environments, as described for several
230
other Gram-negative bacteria, thus limiting nutrient starvation and competition with other
231
microorganisms [20,22,30,31].
232
We observed that a clinical strain of Fth remained culturable for more than 11 weeks of
233
incubation in nutrient-poor water at 4°C; more than one month at 18°C but only one week at
234
Page 11 sur 22
37°C, consistent with previously published data on the culturability of Fth FSC200 and LVS
235
[4,19–21]. However, we show here that culturability is not representative of the viability of Fth
236
strains since the bacterium may switch to the VBNC state under conditions that remain to be
237
fully characterized. Indeed, our results showed prolonged survival in nutrient-poor water at 4°C
238
and 18°C of a virulent Fth biovar I strain long after the bacterium had lost its ability to grow on
239
an agar plate. Our main approach assessing bacterial survival is based on qPCR amplification
240
of DNA from bacteria preincubated with PMAxx™ Dye widely used to detect and determine
241
the viability of human pathogens [32]. As PMAxx™ Dye does not pass through intact bacterial
242
membranes, it cannot bind to the DNA of living bacteria although binding to the DNA of dead
243
bacteria and extracellular DNA is possible. While the amount of DNA from living bacteria
244
decreased similarly to culturability at 37°C, it remained remarkably stable over time at 4°C and
245
18°C during the whole experiment matching the definition of bacterial switch into a VBNC
246
state as the majority of initial bacteria remained viable despite the loss of culturability and
247
results were confirmed by Live/Dead® BacLight™ assay [33]. Morphological analysis showed
248
that VBNC Fth bacteria are smaller as they have a reduced length, perimeter and area compared
249
to the culturable forms.
250
Like the seeds of plants, VBNC forms allow preserving the genetic heritage of bacteria in
251
unfavorable conditions [34]. Fth bacteria could then remain viable for a very long time as
252
VBNC bacteria in aquatic environments without the need for a host. When more favorable
253
conditions return, VBNC bacteria revert to their vegetative state, usually recovering their
254
culturability and virulence. Reversion after switch into VBNC state remains to be demonstrated
255
for Fth in further studies.
256
Bacteria evolve to a VBNC state to withstand environmentally induced stresses. In our
257
experiments, incubation of Fth in water at 37°C was the most deleterious environmental
258
condition. It did not induce a transition to the VBNC state since bacterial mortality correlated
259
Page 12 sur 22
with loss of culturability. Therefore, it appears that conditions that are too harmful to Fth and
260
associated with a loss of their culturability in one week do not allow the development of VBNC
261
bacteria. The most favorable conditions for Fth survival are close to environmental conditions
262
in tularemia endemic areas, i.e., areas of water temperatures ranging from 4 to 20°C between
263
winter and summer periods [4,28]. Importantly after the switch into VBNC state, Fth viability
264
was still dependent on the temperature of the water. Over 30°C, the viability of VBNC bacteria
265
declined, and was completely abolished after seven days at 37°C. Thus, the temperature tipping
266
point no longer supporting the transition of Fth to the VBNC state is between 18°C and 30°C.
267
Our results support the aquatic environmental distribution of Fth in Northern regions where
268
water temperature may not often exceed the temperature limit killing bacteria in a VBNC state.
269
The inter-tropical region, with higher water temperature, could therefore represents a physical
270
limitation to spreading towards the southern hemisphere. The seasonality could also have a
271
significant role in maintaining this environmental reservoir since the bacterial persistence is
272
better in freshwater.
273
One other mode of persistence of bacteria in aquatic environments is biofilm formation, as
274
observed for many bacteria like Legionella pneumophila [35]. Experimental studies have
275
shown that environmental species of Francisella can form biofilms in vitro [36]. F. novicida
276
starts biofilm formation after two hours and can be evidenced by crystal violet staining after
277
24h [36]. In our study, we observed thin biofilm formation at the bottom of the flasks after six
278
or 12 months of incubation of the microcosms at 18°C or 4°C but no biofilm formation after
279
four months at 37°C. The biofilm was very fragile and therefore difficult to manipulate for
280
observation. Biofilm formation of Fth strains may be a slow process requiring the viability of
281
the bacteria for more than one week. The absence of biofilm formation of F. tularensis strains
282
observed in the study of Golovliov et al. may be related to experimental conditions in axenic
283
media not mimicking the natural aquatic environment [4]. Our study is closer to environmental
284
Page 13 sur 22
conditions, although it did not contain other competitive bacteria or predatory microorganism,
285
because the Fth strain was incubated in a large volume of filtered French lake water. In our
286
experiment, VBNC bacteria seemed to be embedded in a biofilm matrix, as previously shown
287
for other pathogens (e.g., Legionella pneumophila and Listeria monocytogenes) [30,37]. In
288
2016, Flemming et al. described biofilms as a “reservoir of VBNC bacteria,” especially in the
289
starvation zones of the biofilm [35]. The biofilm and VBNC states play an essential role in the
290
persistence of bacteria. Both allow the bacteria to survive in hostile environments while many
291
pathogens lose their virulence properties after their switch into a VBNC state [31].
292
The persistence of Fth in aquatic environments in a VBNC state questions our capacity to detect
293
and fight this bacterium in this specific reservoir. VBNC state may be a way of long-term
294
bacterial persistence of Fth that cannot be detected by conventional culture-based techniques.
295
The VBNC formation process likely explains that detection of F. tularensis in the aquatic
296
environment has been obtained by species-specific molecular methods but very rarely by
297
culture techniques [5]. In case of accidental or intentional dispersal of F. tularensis, the
298
bacterium may thus survive for many months in water environments although undetected by
299
culture methods. Identification of reactivation factors from the VBNC state into a more virulent
300
and culturable state will have to be addressed in further experiments. It would help prevent and
301
control waterborne sporadic and outbreak tularemia cases.
302
The impact of temperature may also have some effects on tularemia diagnosis in humans. This
303
bacterium is usually grown at 37°C from clinical samples in only 10% of tularemia patients. In
304
the light of this work, this temperature may not be optimal for isolating this bacterium from
305
patients, animal samples, or environmental samples or even for bacterial counts after growth on
306
an agar medium. Potential switch into VBNC state in vivo in infected tissues (especially lymph
307
nodes) could also partly explain the failure to isolate this bacterium and may impact therapeutic
308
outcome as VBNC bacteria usually exhibit increased antibiotic resistance because of a reduced
309
Page 14 sur 22
metabolism while biofilms also increase resistance to antibiotics [31,38]. These findings may
310
have implications in treatment failures observed in 20 to 30% of patients, especially when the
311
diagnosis is delayed, which could allow the bacteria to switch to a VBNC state in vivo.
312
Finally, all these data about the environmental survival of Fth in water at low temperature brings
313
new important features allowing updating the aquatic cycle of this bacterium and proposing
314
new hypotheses (Figure 4). Indeed, many other Francisella species are aquatic bacteria, making
315
several parts of this aquatic cycle questionable [5]. What if Fth had rather evolved to adapt to
316
an aquatic niche yet poorly characterized so far while becoming infectious for various mammal
317
species, which are usually dead ends for the bacteria because it often kills its hosts [3]? Indeed,
318
this bacterium has been identified in many aquatic areas, including the sea water, rivers, ponds
319
[25–28,39,40], which might suggest that the primary reservoir of the subspecies holarctica
320
could rather be the aquatic environment itself. This hypothesis could explain why a specific
321
reservoir within the environment has not been identified despite decades of research. Some
322
aquatic environments may thus represent the largest reservoir of Fth with the implication of
323
aquatic rodents to maintain the cycle through bacterial inoculum amplification. Animals and
324
humans may thus be infected directly from this environmental reservoir through water
325
consumption or aerosols, explaining some sporadic human respiratory contaminations after
326
outdoor activities. Aquatic environments could act as a primary source of human and animal
327
infections or mosquito larvae contamination after reactivation of the VBNC state into virulent
328
bacteria upon particular environmental conditions. Further studies will be necessary to
329
determine: 1/ if Fth VBNC bacteria are also virulent and able to infect animals and mosquito
330
larvae who are at the interface between the aquatic and the terrestrial cycle; 2/ to identify
331
reactivation factors from the VBNC state towards the culturable and virulent state; 3/ to study
332
interaction of VBNC Fth bacteria within biofilms and with amoeba.
333
Page 15 sur 22
In conclusion, our study demonstrated the extended persistence of a virulent strain of Fth in
334
water up to 24 months through the formation of VBNC bacteria and thin biofilms. Water
335
temperature appears as a major factor for bacterial survival in aquatic environments. It affects
336
the culturability of the bacteria, the switch toward the VBNC state, and the viability of VBNC
337
cells. Our findings reinforce the hypothesis of a long-term environmental aquatic reservoir of
338
this pathogen.
339
340
Page 16 sur 22
Funding
341
This work and the doctorate allocation of Camille D. Brunet are funded by the Agence
342
Innovation Defense, Direction Générale de l’Armement, France, [grant number Tulamibe
343
ANR-17-ASTR-0024].
344
Acknowledgment
345
We thank the company Abiolab Asposan for the chemical analysis of the water. We thank
346
Ludovic Sansoni for his help on the figure of aquatic cycle.
347
Declaration of interest statement
348
The authors declare no conflicts of interest
349
Page 17 sur 22
References
350
[1] A. Sjostedt, Tularemia: History, Epidemiology, Pathogen Physiology, and Clinical
351
Manifestations, Ann. N. Y. Acad. Sci. 2007;110:1–29.
352
[2] J. Jackson, A. McGregor, L. Cooley, et al. Francisella tularensis Subspecies holarctica,
353
Tasmania, Australia, 2011, Emerg. Infect. Dis. 2012;18:1484–1486.
354
[3] M. Maurin and M. Gyuranecz, Tularaemia: clinical aspects in Europe, Lancet Infect. Dis.
355
2016;16:113–124.
356
[4] I. Golovliov, S. Bäckman, M. Granberg, et al. Long-term survival of virulent tularemia
357
pathogens outside a host in conditions that mimic natural aquatic environments, Appl.
358
Environ. Microbiol. 2021;027:13-20.
359
[5] A. Hennebique, S. Boisset and M. Maurin, Tularemia as a waterborne disease: a review,
360
Emerg. Microbes Infect. 2019;8:1027–1042.
361
[6] W.L. Jellison, Tularemia in North America, 1930-1974, (1974).
362
[7] S. Bäckman, J. Näslund, M. Forsman, et al. Transmission of tularemia from a water source
363
by transstadial maintenance in a mosquito vector, Sci. Rep. 2015;5:7793.
364
[8] O. Wik, Large tularaemia outbreak in Varmland, central Sweden, 2006, Euro Surveill.
365
Bull. Eur. Sur Mal. Transm. Eur. Commun. Dis. Bull. 2006;11
366
[9] H. Eliasson and E. Bäck, Tularaemia in an emergent area in Sweden: An analysis of 234
367
cases in five years, Scand. J. Infect. Dis. 2007;39:880–889.
368
[10] U. Jounio, M. Renko and M. Uhari, An outbreak of holarctica-type tularemia in pediatric
369
patients, Pediatr. Infect. Dis. J. 2010;29:160–162.
370
[11] E. Peker, A. Ayaydin and N. Duran, Familial tularaemia, Indian J. Med. Microbiol.
371
2009;27:272.
372
[12] B. Guerpillon, A. Boibieux, C. Guenne, et al. Keep an Ear Out for Francisella tularensis:
373
Otomastoiditis Cases after Canyoneering, Front. Med. 2016;3.
374
[13] A. Karadenizli, M. Forsman, H. Şimşek, et al. Genomic analyses of Francisella tularensis
375
strains confirm disease transmission from drinking water sources, Turkey, 2008, 2009 and
376
2012, Euro Surveill. Bull. Eur. Sur Mal. Transm. Eur. Commun. Dis. Bull. 2015;20.
377
[14] T. Kantardjiev, I. Ivanov, T. Velinov, et al., Tularemia outbreak, Bulgaria, 1997-2005,
378
Emerg. Infect. Dis. 2006;12:678–680.
379
[15] M. Kevin, G. Girault, Y. Caspar, et al. Phylogeography and Genetic Diversity of
380
Francisella tularensis subsp. holarctica in France (1947–2018), Front. Microbiol.
381
2020;11.
382
[16] C. Dwibedi, D. Birdsell, A. Lärkeryd, et al., Long-range dispersal moved Francisella
383
tularensis into Western Europe from the East, Microb. Genomics 2016;2:100.
384
[17] A. Johansson, A. Lärkeryd, M. Widerström, et al. An Outbreak of Respiratory Tularemia
385
Caused by Diverse Clones of Francisella tularensis, Clin. Infect. Dis. Off. Publ. Infect.
386
Dis. Soc. Am. 2014;59:1546–1553.
387
[18] A. Mailles, R. Ollivier, F. Benezit, et al. [A tularemia outbreak in France in 2018
388
suggesting a change in the epidemiology of the disease] Médecine Mal. Infect. 2019;49:4.
389
French.
390
[19] S.E. Gilbert and L.J. Rose, Survival and persistence of nonspore-forming biothreat agents
391
in water: Persistence of nonspore-forming biothreat agents in water, Lett. Appl. Microbiol.
392
2012;55:189–194.
393
[20] M. Forsman, E.W. Henningson, E. Larsson, et al. Francisella tularensis does not manifest
394
virulence in viable but non-culturable state, FEMS Microbiol. Ecol. 2000;31:217–224.
395
[21] Z.L. Berrada and S.R. Telford III, Survival of Francisella tularensis Type A in brackish-
396
water, Arch. Microbiol. 2011;193:223–226.
397
Page 18 sur 22
[22] J. Thelaus, A. Andersson, P. Mathisen, et al. Influence of nutrient status and grazing
398
pressure on the fate of Francisella tularensis in lake water: Role of nutrient and protozoa
399
for the fate of F. tularensis, FEMS Microbiol. Ecol. 2009;67:69–80.
400
[23] H.S. Xu, N. Roberts, F.L. Singleton, et al. Survival and viability of nonculturable
401
Escherichia coli and Vibrio cholerae in the estuarine and marine environment, Microb.
402
Ecol. 1982;8:313–323.
403
[24] A. Ducret, E.M. Quardokus and Y.V. Brun, MicrobeJ, a tool for high throughput bacterial
404
cell detection and quantitative analysis, Nat. Microbiol. 2016;1:16077.
405
[25] I. Janse, R.Q.J. Van der Plaats, A.M. de Roda Husman et al. Environmental Surveillance
406
of Zoonotic Francisella tularensis in the Netherlands, Front. Cell. Infect. Microbiol.
407
2018;8:140.
408
[26] T. Broman, J. Thelaus, A.-C. Andersson, et al. Molecular Detection of Persistent
409
Francisella tularensis Subspecies holarctica in Natural Waters, Int. J. Microbiol. 2011:1–
410
10.
411
[27] S. Duodu, P. Larsson, A. Sjödin, et al. The Distribution of Francisella-like Bacteria
412
Associated with Coastal Waters in Norway, Microb. Ecol. 2012;64:370–377.
413
[28] C.D. Brunet, A. Hennebique, J. Peyroux, et al. Presence of Francisella tularensis subsp.
414
holarctica DNA in the Aquatic Environment in France, Microorganisms 2021;9:1398.
415
[29] Z.L. Berrada and S.R. Telford, Diversity of Francisella Species in Environmental Samples
416
from Martha’s Vineyard, Massachusetts, Microb. Ecol. 59 (2010), pp. 277–283.
417
[30] L. Li, N. Mendis, H. Trigui, et al. The importance of the viable but non-culturable state in
418
human bacterial pathogens, Front. Microbiol. 2014;5
419
[31] J.D. Oliver, Recent findings on the viable but nonculturable state in pathogenic bacteria,
420
FEMS Microbiol. Rev. 2010;34:415–425.
421
[32] L. Soto-Muñoz, N. Teixidó, J. Usall, et al. Development of PMA real-time PCR method
422
to quantify viable cells of Pantoea agglomerans CPA-2, an antagonist to control the major
423
postharvest diseases on oranges, Int. J. Food Microbiol. 2014;180:49–55.
424
[33] M. Ayrapetyan and J.D. Oliver, The viable but non-culturable state and its relevance in
425
food safety, Curr. Opin. Food Sci. 2016;8:127–133.
426
[34] J.T. Lennon and S.E. Jones, Microbial seed banks: the ecological and evolutionary
427
implications of dormancy, Nat. Rev. Microbiol. 2011;9:119–130.
428
[35] H.-C. Flemming, J. Wingender, U. Szewzyk, et al. Biofilms: an emergent form of bacterial
429
life, Nat. Rev. Microbiol. 2016;14:563–575.
430
[36] C. Siebert, C. Villers, G. Pavlou, et al., Francisella novicida and F. philomiragia biofilm
431
features conditionning fitness in spring water and in presence of antibiotics, PLOS ONE
432
2020;15:0228591.
433
[37] J. Wingender and H.-C. Flemming, Biofilms in drinking water and their role as reservoir
434
for pathogens, Int. J. Hyg. Environ. Health. 2011;204:417–423.
435
[38] S.J. Kassinger and M.L. van Hoek, Genetic Determinants of Antibiotic Resistance in
436
Francisella, Front. Microbiol. 2021;12:644855.
437
[39] P. Kaysser, E. Seibold, K. Mätz-Rensing, et al. Re-emergence of tularemia in Germany:
438
Presence of Francisella tularensis in different rodent species in endemic areas, BMC
439
Infect. Dis. 2008;8:157.
440
[40] J. Hightower, I.T. Kracalik, N. Vydayko et al. Historical distribution and host-vector
441
diversity of Francisella tularensis, the causative agent of tularemia, in Ukraine, Parasit.
442
Vectors 2014;7:453.
443
444
Page 19 sur 22
Figures
445
446
Figure 1: Culturability and viability of F. tularensis ssp. holarctica in nutrient-poor water.
447
Culturability (1a-c) and viability (1d-f) of Fth in nutrient-poor water at respectively 37°C (1a,d),
448
18°C (1b,e) and 4°C (1c,f). Culturability was measured by CFU counts after serial dilutions and
449
spreading on chocolate agar plates. Viability was evaluated by amplification of DNA after
450
PMAxx™ Dye treatment. Black circle: nutrient-poor water with 0 g/L NaCl; black square:
451
nutrient-poor water with 10 g/L NaCl.; Dotted line: mean Ct of all the controls performed on
452
dead populations at each sampling points. The results are expressed as the average of three
453
biological replicates. Data were analyzed by student t-test. * p value <0,05 between samples
454
with and without NaCl.
455
Page 20 sur 22
456
Figure 2: Viability of VBNC F. tularensis ssp. holarctica in water after a temperature change.
457
Several months after the loss of culturability of Fth in water, 5 mL of microcosm at 4°C were
458
transferred at 18°C, 30°C and 37°C (3a) and 5 mL of microcosm at 18°C were transferred at
459
4°C, 30°C and 37°C (3b). After 7 and 14 days, the viability was evaluated by qPCR after
460
PMAxx™ Dye treatment. Dash bars: Ct at day 0, black bars: Ct at day 7, white bars: Ct at day
461
14, dotted line: Ct of the control corresponding to average of a dead population. The results are
462
expressed as the average of three biological replicates. Data were analyzed by one-way
463
ANOVA with pairwise t-tests using FDR correction. * p-value < 0.05.
464
465
Page 21 sur 22
466
Figure 3: Quantitative measurement of biofilm formation of F. tularensis ssp. holarctica in
467
VBNC state. Fth bacteria were incubated in nutrient-poor water for one hour for cultivable
468
bacteria, one year at 4°C and six months at 18°C for VBNC and for four months at 37°C for
469
non-persistent bacteria. Biofilm biomass was estimated by absorbance at 570 nm of crystal
470
violet assay. The results are expressed as the average of three biological replicates. Data were
471
analyzed by student t-test, * p-value <0.05 ** p-value < 0.01.
472
473
Page 22 sur 22
474
Figure 4: The hidden aquatic reservoir of F. tularensis ssp. holarctica?
475
We updated the current knowledge about the aquatic cycle of Fth according to the results of
476
this work and proposed hypotheses that emerged from our observations. We showed that the
477
survival of Fth in aquatic environments is driven by water temperature and transition into a
478
VBNC state. While Fth culturability is prolonged in water at low temperatures (4-18°C), these
479
low temperatures actually also allow the survival of the bacteria for months or years after
480
transition into a VBNC state. On the opposite high temperatures (> 30°C) are associated to
481
complete loss of culturability and loss of viability of the bacteria, even if the bacteria has already
482
switched into the VBNC state at lower temperatures. Thus, mammals or accidentally human
483
may be contaminated from this long-term aquatic reservoir by water drinking, direct contact or
484
by inhalation of contaminated droplets that could explain several respiratory tularemia cases
485
related to environmental exposure only. When infected, wild animals can amplify the bacterial
486
inoculum within the same aquatic environment or disperse the bacteria in other environments
487
with their carcasses and feces and may contaminate other animals or exceptionally humans as
488
described in the terrestrial cycle of the bacteria.
489
| 2022 | Aquatic long-term persistence of ssp. is driven by water temperature and transition to a viable but non-culturable state | 10.1101/2022.02.18.480867 | [
"Brunet Camille D.",
"Peyroux Julien",
"Pondérand Léa",
"Bouillot Stéphanie",
"Girard Thomas",
"Faudry Éric",
"Maurin Max",
"Caspar Yvan"
] | creative-commons |
1
Title
1
Genome-wide macroevolutionary signatures of key innovations in
2
butterflies colonizing new host plants
3
4
Authors
5
Rémi Allio1*, Benoit Nabholz1, Stefan Wanke2, Guillaume Chomicki3, Oscar A. Pérez-
6
Escobar4, Adam M. Cotton5, Anne-Laure Clamens6, Gaël J. Kergoat6, Felix A.H. Sperling7 &
7
Fabien L. Condamine1,7*
8
9
Affiliations
10
1Institut des Sciences de l’Evolution de Montpellier (Université de Montpellier | CNRS | IRD
11
| EPHE), Place Eugène Bataillon, 34095 Montpellier, France. 2Institut für Botanik,
12
Technische Universität Dresden, Zellescher Weg 20b, 01062, Dresden, Germany.
13
3Department of Bioscience, Durham University, Stockton Rd, Durham DH1 3LE, UK. 4Royal
14
Botanic Gardens, Kew, TW9 3AB, Surrey, UK. 586/2 Moo 5, Tambon Nong Kwai, Hang
15
Dong, Chiang Mai, Thailand. 6CBGP, INRAE, CIRAD, IRD, Montpellier SupAgro, Univ.
16
Montpellier, Montpellier, France. 7University of Alberta, Department of Biological Sciences,
17
Edmonton T6G 2E9, AB, Canada.
18
19
Correspondence
20
Rémi Allio: rem.allio@yahoo.fr
21
Fabien L. Condamine: fabien.condamine@gmail.com
22
23
2
The exuberant proliferation of herbivorous insects is attributed to their associations
24
with plants. Despite abundant studies on insect-plant interactions, we do not know
25
whether host-plant shifts have impacted both genomic adaptation and species
26
diversification over geological times. We show that the antagonistic insect-plant
27
interaction between swallowtail butterflies and the highly toxic birthworts began 55
28
million years ago in Beringia, followed by several major ancient host-plant shifts. This
29
evolutionary framework provides a unique opportunity for repeated tests of genomic
30
signatures of macroevolutionary changes and estimation of diversification rates across
31
their phylogeny. We find that host-plant shifts in butterflies are associated with both
32
genome-wide adaptive molecular evolution (more genes under positive selection) and
33
repeated bursts of speciation rates, contributing to an increase in global diversification
34
through time. Our study links ecological changes, genome-wide adaptations and
35
macroevolutionary consequences, lending support to the importance of ecological
36
interactions as evolutionary drivers over long time periods.
37
3
Plants and phytophagous insects constitute most of the documented species of terrestrial
38
organisms. To explain their staggering diversity, Ehrlich and Raven1 proposed a model in
39
which a continual arms race of attacks by herbivorous insects and new defences by their host
40
plants is linked to species diversification via the creation of new adaptive zones, later termed
41
the ‘escape-and-radiate’ model2. Study of insect-plant interactions has progressed
42
tremendously since then through focus on chemistry3, phylogenetics4,5, and genomics6–9.
43
Divergence of key gene families7–10 and high speciation rates11–13 have been identified after
44
host-plant shifts, with one example linking duplication of key genes to the ability to feed on
45
new plants and increase diversification7. However, a major knowledge gap lies in our
46
understanding of the evolutionary linkages and drivers of host-plant shifts, genome-wide
47
signatures of adaptations, and processes of species diversification14.
48
Here we address this gap with an emblematic group that was instrumental in Ehrlich
49
& Raven’s model - the swallowtail butterflies (Lepidoptera: Papilionidae). First, we created
50
an extensive phylogenetic dataset including 7 genetic markers for 71% of swallowtail species
51
diversity (408 of ~570 described species, Methods). Second, we compiled host-plant
52
preferences for each swallowtail species in the dataset. Their caterpillars feed on diverse
53
flowering-plant families, and a third of swallowtail species are specialized on the flowering
54
plant family Aristolochiaceae (birthworts), which is one of the most toxic plant groups and
55
carcinogenic to many organisms15,16. Phylogenetic estimates of ancestral host-plant
56
preferences indicate that Aristolochiaceae were either the foodplant of ancestral
57
Papilionidae17 or were colonized twice18, suggesting an ancient and highly conserved
58
association with Aristolochiaceae throughout swallowtail evolution. Using a robust and
59
newly reconstructed time-calibrated phylogeny (Supplementary Figs. 1-3), we have traced the
60
evolutionary history of food-plant use and infer that the family Aristolochiaceae was the
61
ancestral host for Papilionidae (Fig. 1; relative probabilities = 0.915, 0.789, and 0.787 with
62
three models, Supplementary Figs. 4, 5). We further show that the genus Aristolochia was the
63
ancestral host-plant, as almost all Aristolochiaceae-associated swallowtails feed on
64
Aristolochia (Supplementary Fig. 6). Across the swallowtail phylogeny, we recover only 14
65
host-plant shifts at the family level (14 nodes out of 407; Supplementary Figs. 4, 5),
66
suggesting strong evolutionary host-plant conservatism.
67
With the ancestor of swallowtails feeding on birthworts, evidence for synchronous
68
temporal and geographical origins further links the genus Aristolochia and the family
69
Papilionidae and supports the ‘escape and radiate’ model. Reconstructions of co-phylogenetic
70
history for other insect-plant antagonistic interactions have shown either synchronous
71
4
diversification5 or herbivore diversification lagging behind that of their host plants4,19. We
72
assembled a molecular dataset for ~45% of the species diversity of Aristolochiaceae (247 of
73
~550 described species; Methods) and reconstructed their phylogeny (Supplementary Fig. 7).
74
Divergence time estimates indicate highly synchronous radiation by Papilionidae (55.4
75
million years ago [Ma], 95% credibility intervals: 47.8-71.0 Ma) and Aristolochia (55.5 Ma,
76
95% credibility intervals: 39.2-72.8 Ma) since the early Eocene (Fig. 2; Supplementary Figs.
77
3, 8, 9). This result is robust to known biases in inferring divergence times, with slightly older
78
ages inferred for both groups when using more conservative priors on clade ages
79
(Supplementary Fig. 9). Such temporal congruence between Aristolochia and Papilionidae
80
raises the question of whether both clades had similar geographical origins and dispersal
81
routes. To characterize the macroevolutionary patterns of the Aristolochia/Papilionidae arms-
82
race in space, we assembled two datasets of current geographic distributions for all species
83
included in the phylogenies of both Aristolochiaceae and Papilionidae. We reconstructed the
84
historical biogeography of both groups, taking into account palaeogeographical events
85
throughout the Cenozoic (Methods). The results show that both Papilionidae and Aristolochia
86
were ancestrally co-distributed throughout a region including West Nearctic, East Palearctic,
87
and Central America in the early Eocene, when Asia and North America were connected by
88
the Bering land bridge (Fig. 2, Supplementary Figs. 10, 11). This extraordinary combination
89
of close temporal and spatial congruence provides strong evidence that Papilionidae and
90
Aristolochia diversified concurrently through time and space until several swallowtail
91
lineages shifted to new host-plant families in the middle Eocene.
92
Our ancestral state estimates and biogeographic analyses are consistent with a
93
sustained arms race between Aristolochia and Papilionidae in the past 55 million years.
94
According to the escape-and-radiate model, a host-plant shift should confer higher rates of
95
species diversification for herbivores through the acquisition of novel resources to radiate
96
into1,2 and/or the lack of competitors (Aristolochiaceae-feeder swallowtails have almost no
97
competitors20). We tested the hypothesis that increases of diversification rates occurred in
98
swallowtail lineages that shifted to new host-plants. Applying a suite of birth-death models
99
(Methods), we find evidence for (1) upshifts of diversification at host-plant shifts with trait-
100
dependent birth-death models (Fig. 3a; Supplementary Figs. 12, 13, Supplementary Table 1),
101
and (2) host-plant shifts contributing to a global increase through time with time-dependent
102
birth-death models (Fig. 3b; Supplementary Figs. 14-16). Surprisingly, we do not observe the
103
classical slowdown of diversification recovered in most phylogenies, often attributed to
104
ecological limits and niche filling processes21. This sustained and increasing diversification
105
5
during the Cenozoic may be explained by ecological opportunities not decreasing, due to a
106
steady increase in host breadth for Papilionidae with new host-plant families colonized
107
through time (Supplementary Fig. 17). Opening up new niches would allow continuous
108
increase in diversification rates through time in a dynamic biotic environment, lending
109
support to the primary role of ecological interactions in clade diversification over long
110
timescales.
111
Key innovations are often considered to underlie ecological opportunities and/or
112
evolutionary success22, particularly in the case of chemically mediated interactions between
113
butterflies and their host-plants7. Studies on Papilionidae have provided strong examples of
114
specific changes in key genes that confer new abilities to feed on toxic plants and allow host-
115
plant shifts23,24. Adaptations of swallowtails to their hosts have particularly been assessed
116
through the study of cytochrome P450 monooxygenases (P450s), which have a major role in
117
detoxifying secondary plant compounds. New P450s appear to arise in swallowtails that
118
colonize new hosts to bypass toxic defences, providing survival and diversification on some
119
but not all plants9,23,25. This supports the hypothesis that insect-plant interactions contributed
120
to P450-gene family diversification, with P450s being key innovations that explain the
121
evolutionary and ecological success of phytophagous insects8,9,24,26–28. However, host-plant
122
shifts not only alter single genes but may also influence unlinked genes29. Moreover, host-
123
plant shifts can accompany changes of abiotic environment, which may in turn require further
124
adaptation (new predators and/or competitors). But the macroevolutionary and genomic
125
consequences of the evolutionary dynamics of host-plant shifts have not yet been
126
demonstrated.
127
Relying on a genomic dataset comprising 45 genomes covering all swallowtail
128
genera30–33, we asked whether there are any genomic signatures of positive selection caused
129
by host-plant shifts within swallowtails. We performed a comparative genomic survey of
130
molecular evolution to test whether there is a contrasting pattern of molecular adaptation
131
between swallowtail lineages that shifted to new host plants compared to non-shifting
132
lineages (Methods). We selected 14 phylogenetic branches representing a host-plant shift and
133
14 phylogenetic branches with no change as negative controls34,35 (Fig. 4a). For a fair
134
molecular comparison, each branch selected as a negative control was chosen to be as close
135
as possible to a test branch representing a host-plant shift (i.e. sister groups, Supplementary
136
Fig. 18). Among branches with host-plant shifts, 5 branches also had a shift in climate
137
preference (represented by distributional changes from tropical to temperate conditions).
138
Using a maximum-likelihood method, we estimated the ratio of non-synonymous
139
6
substitutions (dN) other synonymous substitutions (dS) in all branches where a host-plant
140
shift was identified relative to branches with no host-plant shift36,37 (Methods). The dN/dS
141
analyses on branches with host-plant shifts (combined or not with environmental shifts)
142
showed more genome-wide molecular adaptations (i.e. more genes under positive selection,
143
dN/dS > 1) in lineages shifting to a new plant family, although the difference was marginally
144
non-significant (Fig. 4b, P = 0.0501 / 0.0345 for the two datasets, respectively, Wilcoxon
145
rank-sum test, see Methods for the definition of the datasets). However, dN/dS analyses on
146
branches with environmental shifts indicated a balanced number of genes under positive
147
selection (Fig. 4c, P = 0.336 / 0.834 for the two datasets, respectively, Wilcoxon rank-sum
148
test), suggesting a lower impact of environmental shifts than host-plant shifts. We then
149
performed dN/dS analyses for branches with host-plant shifts only (not followed by
150
environmental shifts) and found that swallowtail lineages shifting to a new host-plant family
151
had significantly more genes under positive selection (4.41% / 3.64% of genes under positive
152
selection for the two datasets, respectively) than non-shifting lineages (3.02% / 2.33% of
153
genes under positive selection for the two datasets, respectively, Fig. 4d, P = 0.0071 / 0.0152
154
for the two datasets, respectively, Wilcoxon rank-sum test). We checked individually the
155
gene alignments and performed sensitivity analyses that showed our results are not driven
156
either by an excess of misaligned regions, nor missing data and GC-content variations among
157
species (Methods; Supplementary Figs. 19-25). Surprisingly, the dual changes in climate and
158
host-plant preferences did not spur molecular adaptation across swallowtail lineages (P = 1 /
159
0.517 for the two datasets, respectively, Wilcoxon rank-sum test) and even less than host-
160
plant shifts only (P = 0.0327 / 0.147 for the two datasets, respectively, Wilcoxon rank-sum
161
test; Fig. 3d). Although these genome-wide comparisons rely on a few branches (5 out of 14
162
which significantly differ from others, tested with 1000 random comparisons), no plausible
163
hypothesis can explain this result that would require more in-depth work.
164
We further studied the functional categories of positively selected genes by using
165
gene ontology (GO) analyses (PANTHER and EggNOG; Methods). Applied to the high-
166
quality genomes of Papilio xuthus31 and Heliconius melpomene38, we found that ~70% of the
167
genes are associated with a gene function, which suggests a gap of knowledge in insect gene
168
function database. Among the annotated genes, we found that genes under positive selection
169
along branches with host shifts did not contain over- or under-represented functional GO
170
categories: 252 out of 1213 GO categories represented by genes under positive selection (P >
171
0.05, Fisher’s exact test after false discovery rate correction; Supplementary Table 2). These
172
results support the hypothesis that genome-wide signatures of adaptations are associated with
173
7
host-plant shifts, and encourage extending the long-held hypothesis that only changes in a
174
single candidate family gene are enough to act as a key innovation for adaptation to new
175
resources7,10. Despite a weak signal, it is striking that host-plant shifts left stronger genome-
176
wide signatures than were associated with changing climate preferences. This result further
177
suggests that the success of phytophagous insects involved deeper adaptation to biotic
178
interactions than for shifts in the abiotic environment.
179
Establishing linkages between ecological adaptations, genomic changes, and species
180
diversification over geological timescales remains a tremendous challenge14 with, for
181
instance, important limitations due to the lack of knowledge in functional gene annotations in
182
insects. However, the successful development of powerful analytical tools in conjunction
183
with the increasing availability of insect genomes and improvements in genomic analyses39
184
allow detecting more genes than the known genes involved in detoxification pathways
185
playing a role in long-term relationships between plants and insects. This opens new research
186
avenues for finding the functionality of genes involved in the adaptation and diversification
187
of phytophagous insects. We hope that our study will help movement in that direction, and
188
that it will provide interesting perspectives for future investigations of other model groups.
189
Over a half century ago, Ehrlich and Raven1 proposed that insect-plant interactions
190
driven by diffuse co-evolution over long evolutionary periods can be a major source of
191
terrestrial biodiversity. Applied to a widely appreciated case in the insect-plant interactions
192
theory, our study reveals that genome-wide adaptive processes and corresponding
193
macroevolutionary consequences are more pervasive than previously recognized in the
194
diversification of herbivorous insects. Close relationships between insects and their larval
195
host plants involve more adaptations than in just the gene families in detoxification pathways
196
that were detected through antagonist interactions39, and show genomically wide-ranging co-
197
evolutionary consequences29,40. Hence, genome-wide macroevolutionary consequences of
198
key adaptations in new insect-plant interactions may be a general feature of the co-
199
evolutionary interactions that have generated Earth’s diversity.
200
201
8
References
202
1.
Ehrlich, P. R. & Raven, P. H. Butterflies and plants: a study in coevolution. Evolution
203
18, 586–608 (1964).
204
2.
Thompson, J. N. Concepts of coevolution. Trends Ecol. Evol. 4, 179–183 (1989).
205
3.
Berenbaum, M. R. & Feeny, P. P. Chemical mediation of host-plant specialization: the
206
papilionid paradigm. in Specialization, Speciation, and Radiation: The Evolutionary
207
Biology of Herbivorous Insects (ed. Tilmon, K.) 2–19 (University of California Press,
208
2008).
209
4.
Winter, S., Friedman, A. L. L., Astrin, J. J., Gottsberger, B. & Letsch, H. Timing and
210
host plant associations in the evolution of the weevil tribe Apionini (Apioninae,
211
Brentidae, Curculionoidea, Coleoptera) indicate an ancient co-diversification pattern of
212
beetles and flowering plants. Mol. Phylogenet. Evol. 107, 179–190 (2017).
213
5.
Kergoat, G. J. et al. Opposite macroevolutionary responses to environmental changes
214
in grasses and insects during the Neogene grassland expansion. Nat. Commun. 9, 5089
215
(2018).
216
6.
Wheat, C. W. et al. The genetic basis of a plant–insect coevolutionary key innovation.
217
Proc. Natl. Acad. Sci. 104, 20427–20431 (2007).
218
7.
Edger, P. P. et al. The butterfly plant arms-race escalated by gene and genome
219
duplications. Proc. Natl. Acad. Sci. U. S. A. 112, 8362–8366 (2015).
220
8.
Calla, B. et al. Cytochrome P450 diversification and hostplant utilization patterns in
221
specialist and generalist moths: Birth, death and adaptation. Mol. Ecol. 26, 6021–6035
222
(2017).
223
9.
Nallu, S. et al. The molecular genetic basis of herbivory between butterflies and their
224
host plants. Nat. Ecol. Evol. 2, 1418–1427 (2018).
225
10.
Karageorgi, M. et al. Genome editing retraces the evolution of toxin resistance in the
226
monarch butterfly. Nature 574, 409–412 (2019).
227
11.
Sahoo, R. K., Warren, A. D., Collins, S. C. & Kodandaramaiah, U. Hostplant change
228
and paleoclimatic events explain diversification shifts in skipper butterflies (Family:
229
Hesperiidae). BMC Evol. Biol. 17, 174 (2017).
230
12.
Condamine, F. L., Rolland, J., Höhna, S., Sperling, F. A. H. & Sanmartín, I. Testing
231
the role of the red queen and court jester as drivers of the macroevolution of apollo
232
butterflies. Syst. Biol. 67, 940–964 (2018).
233
13.
Letsch, H. et al. Climate and host-plant associations shaped the evolution of
234
ceutorhynch weevils throughout the Cenozoic. Evolution 72, 1815–1828 (2018).
235
9
14.
Hua, X. & Bromham, L. Darwinism for the genomic age: connecting mutation to
236
diversification. Front. Genet. 8, 12 (2017).
237
15.
Schmeiser, H. H., Stiborovà, M. & Arlt, V. M. Chemical and molecular basis of the
238
carcinogenicity of Aristolochia plants. Curr. Opin. Drug Discov. Devel. 12, 141–148
239
(2009).
240
16.
Poon, S. L. et al. Genome-wide mutational signatures of aristolochic acid and its
241
application as a screening tool. Sci. Transl. Med. 5, 197ra101 (2013).
242
17.
Condamine, F. L., Sperling, F. A. H., Wahlberg, N., Rasplus, J.-Y. & Kergoat, G. J.
243
What causes latitudinal gradients in species diversity? Evolutionary processes and
244
ecological constraints on swallowtail biodiversity. Ecol. Lett. 15, 267–277 (2012).
245
18.
Simonsen, T. J. et al. Phylogenetics and divergence times of Papilioninae
246
(Lepidoptera) with special reference to the enigmatic genera Teinopalpus and
247
Meandrusa. Cladistics 27, 113–137 (2011).
248
19.
McKenna, D. D., Sequeira, A. S., Marvaldi, A. E. & Farrell, B. D. Temporal lags and
249
overlap in the diversification of weevils and flowering plants. Proc. Natl. Acad. Sci. U.
250
S. A. 106, 7083–7088 (2009).
251
20.
Nishida, R. Sequestration of Defensive Substances from Plants by Lepidoptera. Annu.
252
Rev. Entomol. 47, 57–92 (2002).
253
21.
Moen, D. & Morlon, H. Why does diversification slow down? Trends Ecol. Evol. 29,
254
190–197 (2014).
255
22.
Losos, J. B. Adaptive radiation, ecological opportunity, and evolutionary determinism.
256
Am. Nat. 175, 623–639 (2010).
257
23.
Cohen, M. B., Schuler, M. A. & Berenbaum, M. R. A host-inducible cytochrome P-
258
450 from a host-specific caterpillar: molecular cloning and evolution. Proc. Natl.
259
Acad. Sci. U. S. A. 89, 10920–10924 (1992).
260
24.
Berenbaum, M. R., Favret, C. & Schuler, M. A. On defining ‘Key Innovations’ in an
261
adaptive radiation: Cytochrome P450S and Papilionidae. Am. Nat. 148, S139–S155
262
(1996).
263
25.
Li, W., Schuler, M. A. & Berenbaum, M. R. Diversification of furanocoumarin-
264
metabolizing cytochrome P450 monooxygenases in two papilionids: Specificity and
265
substrate encounter rate. Proc. Natl. Acad. Sci. U. S. A. 100 Suppl, 14593–14598
266
(2003).
267
26.
Schuler, M. A. P450s in plant–insect interactions. Biochim. Biophys. Acta - Proteins
268
Proteomics 1814, 36–45 (2011).
269
10
27.
Cheng, T. et al. Genomic adaptation to polyphagy and insecticides in a major East
270
Asian noctuid pest. Nat. Ecol. Evol. 1, 1747–1756 (2017).
271
28.
Rane, R. V et al. Detoxifying enzyme complements and host use phenotypes in 160
272
insect species. Curr. Opin. Insect Sci. 31, 131–138 (2019).
273
29.
Thompson, J. N., Wehling, W. & Podolsky, R. Evolutionary genetics of host use in
274
swallowtail butterflies. Nature 344, 148–150 (1990).
275
30.
Cong, Q., Borek, D., Otwinowski, Z. & Grishin, N. V. Tiger swallowtail genome
276
reveals mechanisms for speciation and caterpillar chemical defense. Cell Rep. 10, 910–
277
919 (2015).
278
31.
Li, X. et al. Outbred genome sequencing and CRISPR/Cas9 gene editing in butterflies.
279
Nat. Commun. 6, 8212 (2015).
280
32.
Nishikawa, H. et al. A genetic mechanism for female-limited Batesian mimicry in
281
Papilio butterfly. Nat. Genet. 47, 405–409 (2015).
282
33.
Allio, R. et al. Whole genome shotgun phylogenomics resolves the pattern and timing
283
of swallowtail butterfly evolution. Syst. Biol. 69, 38–60 (2020).
284
34.
Thomas, G. W. C. & Hahn, M. W. Determining the null model for detecting adaptive
285
convergence from genomic data: A case study using echolocating mammals. Mol. Biol.
286
Evol. 32, 1232–1236 (2015).
287
35.
Zou, Z. & Zhang, J. No genome-wide protein sequence convergence for echolocation.
288
Mol. Biol. Evol. 32, 1237–1241 (2015).
289
36.
Kimura, M. The Neutral Theory of Molecular Evolution. (Cambridge University Press,
290
1983).
291
37.
Yang, Z. Computational Molecular Evolution. (Oxford University Press, 2006).
292
38.
Dasmahapatra, K. K. et al. Butterfly genome reveals promiscuous exchange of
293
mimicry adaptations among species. Nature 487, 94–98 (2012).
294
39.
Thomas, G. W. C. et al. Gene content evolution in the arthropods. Genome Biol. 21, 15
295
(2020).
296
40.
de Medeiros, B. A. S. & Farrell, B. D. Evaluating species interactions as a driver of
297
phytophagous insect divergence. bioRxiv 842153 (2019). doi:10.1101/842153
298
299
11
Acknowledgements This project has received funding from the Marie Curie International
300
Outgoing Fellow under the European Union’s Seventh Framework Programme (project
301
BIOMME, agreement No. 627684), a PICS grant from the CNRS (project PASTA), an
302
“Investissement d’Avenir” grant from the Agence Nationale de la Recherche (project
303
CASMA, CEBA, ref. ANR-10-LABX-25-01), and the European Research Council (ERC)
304
under the European Union’s Horizon 2020 research and innovation programme (project
305
GAIA, agreement No. 851188) to F.L.C.; a Natural Sciences and Engineering Research
306
Council of Canada (NSERC) Discovery Grant (RGPIN-2018-04920) to F.A.H.S.; and a
307
German Research Foundation grant (WA 2461/9-1) to S.W. We are grateful to Sophie Dang,
308
Troy Locke, and Corey Davis at the Molecular Biology Service Unit of the University of
309
Alberta for their help, assistance, and advice on next-generation sequencing. The analyses
310
benefited from the Montpellier Bioinformatics Biodiversity (MBB) platform services.
311
Finally, we are grateful to Seth Bybee, Frédéric Delsuc, Claude dePamphilis, Krushnamegh
312
Kunte, Conrad Labandeira, Harald Letsch, Sören Nylin, Timothy O’Hara, Susanne Renner
313
and Chris Wheat for helpful comments and discussions on earlier drafts of the study.
314
315
Author contributions F.L.C. and F.A.H.S. designed and conceived the research. R.A. and
316
F.L.C. assembled the phylogenetic data for swallowtail butterflies. S.W., O.A.P.E., G.C.,
317
F.L.C and R.A. assembled the phylogenetic data for birthworts. R.A. and F.L.C. analysed the
318
phylogenetic data. R.A. and F.L.C. performed the ancestral states estimations. F.L.C.
319
performed the diversification analyses. A.-L.C. and F.L.C. generated the genomic data. R.A.
320
and B.N. assembled and analysed the genomic data. All authors contributed to the
321
interpretation and discussion of results. R.A. and F.L.C. drafted the paper with substantial
322
input from all authors.
323
324
Competing interests The authors declare no competing interests.
325
326
12
Figures
327
328
329
330
Fig. 1. Evolution of host-plant association through time shows strong host-plant
331
conservatism across swallowtail butterflies. Phylogenetic relationships of swallowtail
332
butterflies, with coloured branches mapping the evolution of host-plant association, as
333
inferred by a maximum-likelihood model (Supplementary Figs. 4, 6). Additional analyses
334
with two other maximum-likelihood and Bayesian models inferred the same host-plant
335
associations across the phylogeny (Supplementary Fig. 5). Lue. = Luehdorfiini, Zerynth. =
336
Zerynthiini, and T. = Teinopalpini.
337
338
13
339
340
Fig. 2. Synchronous temporal and geographic origin for swallowtails and birthworts.
341
Bayesian molecular divergence times with exponential priors estimate an early Eocene origin
342
(~55 Ma) for both swallowtails and Aristolochia (alternatively, analyses with uniform prior
343
estimated an origin around 67 Ma for swallowtails and 64 Ma for Aristolochia,
344
Supplementary Figs. 3, 8, 9). Biogeographical maximum-likelihood models infer an ancestral
345
area of origin comprising West Nearctic, East Palearctic and Central America for both
346
swallowtails and birthworts (Supplementary Figs. 10, 11). K = Cretaceous, P = Palaeocene, E
347
= Eocene, O = Oligocene, M = Miocene, Pl = Pliocene, and P = Pleistocene. Ma = million
348
years ago.
349
350
14
351
352
Fig. 3. Host-plant shifts lead to repeated bursts in diversification rates and a sustained
353
overall increase in diversification through time. a, Diversification tends to be higher for
354
clades shifting to new host plants, as estimated by trait-dependent diversification models.
355
Boxplots represent Bayesian estimates of net diversification rates for clades feeding on
356
particular host plants (see also Supplementary Fig. 12). b, A global increase in diversification
357
is recovered with birth-death models estimating time-dependent diversification (see also
358
Supplementary Figs. 14, 15). Taking into account rate heterogeneity by estimating host-plant
359
and clade-specific diversification indicates positive gains of net diversification after shifting
360
to new host plants (see also Supplementary Fig. 13). K = Cretaceous, Paleoc. = Palaeocene,
361
Oligoc. = Oligocene, Pl = Pliocene, P = Pleistocene, Ma = million years ago.
362
363
15
364
365
16
Fig. 4. Host-plant shifts promote higher molecular adaptations. a, Genus-level
366
phylogenomic tree displaying branches with and without host-plant shifts, on which genome-
367
wide analyses of molecular evolution are performed. b, Number of genes under positive
368
selection (dN/dS > 1) for swallowtail lineages shifting to new host-plant families (green) or
369
not (grey). c, Number of genes under positive selection for swallowtail lineages undergoing
370
climate shifts (orange) or not (grey). d, Number of genes under positive selection for
371
swallowtail lineages shifting to new host plants (green), shifting both host plant and climate
372
(blue) or not (grey). This demonstrates genome-wide signatures of adaptations in swallowtail
373
lineages shifting to new host-plant families. Genes under positive selection did not contain
374
over- or under-represented functional GO categories (Supplementary Table 2). n.s. = not
375
significant (P > 0.05), * = P ≤ 0.05, ** = P ≤ 0.01.
376
377
17
Methods
378
Time-calibrated phylogeny of Papilionidae. We assembled a supermatrix dataset with
379
available data extracted from GenBank as of May 2017 (most of which has been generated by
380
our research group), using five mitochondrial genes (COI, COII, ND1, ND5 and rRNA 16S)
381
and two nuclear markers (EF-1a and Wg) for 408 Papilionidae species (~71% of the total
382
species diversity) and 20 outgroup species. We aligned the DNA sequences for each gene
383
using MAFFT 7.11041 with default settings (E-INS-i algorithm), and the alignments were
384
checked for codon stops and eventually refined by eye with Mesquite 3.1 (available at:
385
www.mesquiteproject.org). The best-fit partitioning schemes and substitution models for
386
phylogenetic analyses were determined with PartitionFinder 2.1.142 using the greedy search
387
algorithm and the Bayesian Information Criterion. All gene alignments were concatenated in
388
a supermatrix, which is available in Figshare (see Data availability).
389
Phylogenetic relationships were estimated with both maximum likelihood (ML) and
390
Bayesian inference. ML analyses were carried out with IQ-TREE 1.6.843. We set the best-fit
391
partitioning scheme and used ModelFinder to determine the best-fit substitution model for
392
each partition44 and then estimated model parameters separately for every partition45 such that
393
all partitions shared the same set of branch lengths, but we allowed each partition to have its
394
own evolution rate. We performed 1,000 ultrafast bootstrap replicates to investigate nodal
395
support across the topology, considering values > 95 as strongly supported nodes46.
396
Estimating phylogenetic relationships for such a dataset is computationally intensive
397
with Bayesian inference. The ML tree inferred with IQ-TREE was used as a starting tree for
398
Bayesian inference as implemented in MrBayes 3.2.647. Rather than using a single
399
substitution model per molecular partition, we sampled across the entire substitution-model
400
space48 using reversible-jump Markov Chain Monte Carlo (rj-MCMC). Two independent
401
analyses with one cold chain and seven heated chains, each run for 50 million generations,
402
sampled every 5,000 generations. Convergence and performance of Bayesian runs were
403
evaluated using Tracer 1.7.149, the average deviation of split frequencies (ADSF) between
404
runs, the effective sample size (ESS) and the potential scale reduction factor (PSRF) values
405
for each parameter. A 50% majority-rule consensus tree was built after conservatively
406
discarding 25% of sampled trees as burn-in. Node support was evaluated with posterior
407
probability considering values > 0.95 as strong support50. All analyses were performed on the
408
CIPRES
Science
Gateway
computer
cluster51,
using
BEAGLE52.
409
Dating inferences were performed using Bayesian relaxed-clock methods accounting
410
18
for rate variation across lineages53. MCMC analyses implemented in BEAST 1.8.454 were
411
employed to approximate the posterior distribution of rates and divergences times and infer
412
their credibility intervals. Estimation of divergence times relied on constraining clade ages
413
through fossil calibrations. Swallowtail fossils are scarce, but five can unambiguously be
414
attributed to the family. The oldest fossil occurrences of Papilionidae are the fossils
415
†Praepapilio colorado and †Praepapilio gracilis55, both from the Green River Formation
416
(Colorado, USA). The Green River Formation encompasses a 5 million-years period between
417
~48.5 and 53.5 Ma, which falls within the Ypresian (47.8-56 Ma) in the early Eocene56.
418
These fossils can be phylogenetically placed at the crown of the family as they share
419
synapomorphies with all extant subfamilies57,58, and have proven to be reliable calibration
420
points for the crown group12,17,33. Two other fossils belong to Parnassiinae, whose systematic
421
position was assessed using phylogenetic analyses based on both morphological and
422
molecular data in a total-evidence approach12. The first is †Thaites ruminiana59, a
423
compression fossil from limestone in the Niveau du gypse d’Aix Formation of France
424
(Bouches-du-Rhône, Aix-en-Provence, France) within the Chattian (23.03–28.1 Ma) of the
425
late Oligocene60,61. †Thaites is sister to Parnassiini, and occasionally sister to Luehdorfiini +
426
Zerynthiini12. Thus we constrained the crown age of Parnassiinae with a uniform distribution
427
bounded by a minimum age of 23.03 Ma. The second is †Doritites bosniaskii62, an
428
exoskeleton and compression fossil from Italy (Tuscany) from the Messinian (5.33–7.25 Ma,
429
late Miocene)61. †Doritites is sister to Archon (Luehdorfiini12), in agreement with
430
Carpenter63. The crown of Luehdorfiini was thus constrained for divergence time estimation
431
using a uniform distribution bounded with 5.33 Ma. Absolute ages of geological formations
432
were taken from the latest update of the geological time scale.
433
We used a conservative approach to applying calibration priors with the selected
434
fossil constraints by setting uniform priors bounded with a minimum age equal to the
435
youngest age of the geological formation where each fossil was found. All uniform
436
calibration priors were set with an upper bound equal to the estimated age of angiosperms
437
(150 Ma64), which is more than three times older than the oldest Papilionidae fossil. This
438
upper age is intentionally set as ancient to allow exploration of potentially old ages for the
439
clade. Since the fossil record of butterflies is incomplete and biased65, caution is needed in
440
using these fossil calibrations (effect shown in burying beetles66).
441
After enforcing the fossil calibrations, we set the following settings and priors: a
442
partitioned dataset (after the best-fitting PartitionFinder scheme) was analysed using the
443
uncorrelated lognormal distribution clock model, with the mean set to a uniform prior
444
19
between 0 and 1, and an exponential prior (lambda = 0.333) for the standard deviation. The
445
branching process prior was set to a birth–death67 process, using the following uniform
446
priors: the birth–death mean growth rate ranged between 0 and 10 with a starting value at 0.1,
447
and the birth–death relative death rate ranged between 0 and 1 (starting value = 0.5). We
448
performed four independent BEAST analyses for 100 million generations, sampled every
449
10,000th, resulting in 10,000 samples in the posterior distribution of which the first 2500
450
samples were discarded as burn-in. All analyses were performed on the CIPRES Science
451
Gateway computer cluster51, using BEAGLE52. Convergence and performance of each
452
MCMC run were evaluated using Tracer 1.7.149 and the ESS for each parameter. We
453
combined the four runs using LogCombiner 1.8.454. A maximum-clade credibility (MCC)
454
tree was reconstructed, with median ages and 95% credibility intervals (CI). The BEAST
455
files generated for this study are available in Figshare (see Data availability).
456
457
Estimating ancestral host-plant association. We inferred the temporal evolution of host-
458
plant association up to the ancestral host plant(s) at the root of Papilionidae using three
459
approaches: the ML implementation of the Markov k-state (Mk) model68, the ML Dispersal-
460
Extinction-Cladogenesis (DEC) model69, and the Bayesian approach in BayesTraits70. These
461
approaches require a time-calibrated tree and a matrix of character states (current host-plant
462
preference) for each species in the tree. An extensive bibliographic survey was conducted to
463
obtain primary larval host-plants at the family level1,71–74. The host associations of species
464
were categorized using the following twelve character states: (1) Annonaceae, (2) Apiaceae,
465
(3) Aristolochiaceae, (4) Crassulaceae or Saxifragaceae (core Saxifragales), (5) Fabaceae, (6)
466
Hernandiaceae, (7) Lauraceae; (8) Magnoliaceae, (9) Papaveraceae, (10) Rosaceae, (11)
467
Rutaceae, and (12) Zygophyllaceae. The host-plant matrix of Papilionidae is available in
468
Figshare (see Data availability).
469
Ancestral states for host-plant association were first reconstructed using the Mk
470
model (one rate for all transitions between states) allowing any host shift to be equally
471
probable. The Mk model does not allow multiple states for a species. The few species that use
472
multiple host families were thus scored with the most frequent host association. The Mk
473
model was performed with Mesquite 3.1 (available at: www.mesquiteproject.org). To
474
estimate the support of any one character state over another, the most likely state was selected
475
according to a decision threshold, such that if the log likelihoods between two states differ by
476
two log-likelihood units, the one with lower likelihood is rejected68.
477
20
The DEC model was also used to reconstruct ancestral host-plant states69,75. As the
478
Mk model, we assumed that host-plant shifts occurred at equivalent probabilities between
479
plant families and through time, which may not be true given that the host-plant families of
480
Papilionidae did not originate at the same time (e.g. Aristolochiaceae originated around
481
108.07 Ma [95% credibility intervals: 81.01-132.66 Ma]76, and Annonaceae originated about
482
98.94 Ma [95% credibility intervals: 84.78-113.70 Ma]76). We used the estimated molecular
483
ages of the different host-plant groups to constrain our inferences of ancestral host plants a
484
posteriori. We preferred such an approach compared to a more constrained one in which the
485
DEC model is informed with a matrix of host-plant appearances based on their estimated ages
486
by implementing matrices of presence/absence of the character states through time
487
(equivalent to the time-stratified palaeogeographic model, see below for inference of
488
biogeographical history).
489
Finally, the Bayesian approach implemented in BayesTraits 3.0.170 was performed to
490
provide a cross-validation of ML analyses. This approach automatically detects shifts in rates
491
of evolution for multistate data using rj-MCMC. Numbers of parameters and priors were set
492
by default. We ran the rj-MCMC for 10 million generations and sampled states and
493
parameters every 1,000 generations (burn-in of 10,000 generations). We specifically
494
estimated ancestral states at 21 nodes as well as at the root of Papilionidae. For this analysis,
495
we used a set of 100 trees randomly taken from the dating analysis to probe the robustness of
496
our ancestral state estimation across topological uncertainty.
497
The results of these inferences determined the host-plant family(ies) that was (were)
498
the most likely ancestral host(s) at the origin of Papilionidae, indicating (i) which plant
499
phylogeny to reconstruct for studying the macroevolution of the arms race, and (ii) the
500
evolution of ancestral host-plant association along the phylogeny to identify the tree branches
501
where shifts occurred and test for genome-wide changes.
502
The Mk and BayesTraits models always inferred with high support (relative
503
probability = 0.915 and 0.789, respectively) that Aristolochiaceae is the ancestral host plant at
504
the crown of Papilionidae. With the unconstrained DEC model, we found that the ancestral
505
host-plant preference for Papilionidae was always composed of Aristolochiaceae, but also
506
included another family (either Fabaceae, Hernandiaceae or Zygophyllaceae, which are only
507
fed upon by Baronia, Lamproptera and Hypermnestra, respectively). As the sister lineage to
508
all other Papilionidae, Baronia is the only species that feeds on Fabaceae. More precisely,
509
only one species of Fabaceae is consumed: Vachellia cochliacantha (formerly Acacia
510
cochliacantha; recent changes in Acacia taxonomy77). However, Vachellia diverged from its
511
21
sister clade in the Eocene, approximately 50 Ma, and diversified in the Miocene between 13
512
and 17 Ma78, which substantially postdate the origin of Papilionidae. Therefore this result
513
suggests that Aristolochiaceae family represents the most likely candidate as the ancestral
514
host-plant of Papilionidae. Hernandiaceae are consumed by Lamproptera (occasionally by
515
Papilio homerus, Graphium codrus, G. doson and G. empedovana73). More precisely, the
516
host plants of Lamproptera belong to the genus Illigera. This plant genus diverged from its
517
sister genus 48 Ma76 and started diversifying 27 Ma79. The derived phylogenetic position of
518
Lamproptera and the age of its use as a host plant make it very unlikely that Hernandiaceae
519
could constitute the ancestral host plant for Papilionidae. Similarly, the family
520
Zygophyllaceae is consumed by Hypermnestra, most specifically it feeds on the genus
521
Zygophyllum in Central Asia. The genus Zygophyllum is not monophyletic, but Asian
522
Zygophyllum appeared 19.6 Ma80. Applying the same rationale, we are able to discard
523
Zygophyllaceae as a candidate ancestral host plant for Papilionidae. To further refine our
524
ancestral host-plant estimates, we built a presence-absence matrix of plant families based on
525
clade origins estimated in molecular dating studies. Thereby, the age of the different plants
526
can be used to constrain the inference of ancestral host plants. Under such a constrained
527
model, Aristolochiaceae is always recovered as the most likely ancestral host-plant for
528
Papilionidae. It is also interesting that almost all Aristolochiaceae feeders have Aristolochia
529
as host plants, and tests to determine which genus of Aristolochiaceae was originally
530
consumed by Papilionidae showed that it was Aristolochia.
531
532
Time-calibrated phylogeny of the ancestral host: the Aristolochiaceae. Estimation of
533
ancestral host-plant relationships revealed that the family Aristolochiaceae was the ancestral
534
host for Papilionidae. We refer to Aristolochiaceae in its traditional circumscription including
535
the genera Asarum, Saruma, Thottea and Aristolochia. The Angiosperm Phylogeny Group81
536
proposes that Aristolochiaceae also includes the holoparasitic genera Hydnora and
537
Prosopanche (Hydnoraceae), as well as the monotypic family Lactoridaceae from the Juan
538
Fernandez Islands of Chile (Lactoris fernandeziana). The conclusion of APG81 is based on an
539
online survey82 rather than on primary data and this is why we disagree with their
540
argumentation as well as the resulting conclusion of APG given available resilient primary
541
molecular phylogenomic data. However, arguments based on morphology and anatomy83–86,
542
genetics87–92, molecular divergence time76,92, and conservation considerations (Tod Stuessy,
543
pers. comm. with S.W., July 2019) favour splitting them into four families: Aristolochiaceae
544
(Aristolochia and Thottea), Asaraceae (Asarum and Saruma), Hydnoraceae (Hydnora and
545
22
Prosopanche), and Lactoridaceae (Lactoris), collectively called the perianth-bearing
546
Piperales. Therefore we extracted and assembled a supermatrix dataset with available data
547
from GenBank for the perianth-bearing Piperales and its sister lineage, the perianth-less
548
Piperales including Saururaceae and Piperaceae (as of May 2017, most of which has been
549
generated by our research group). We obtained four chloroplast genes (matK, rbcl, trnL, trnL-
550
trnF) and one nuclear marker (ITS) for 247 species of perianth-bearing Piperales (~45% of
551
the total species diversity93) and six outgroups from perianth-less Piperales. We could not
552
include the two genera Hydnora and Prosopanche (Hydnoraceae) because available genetic
553
data do not overlap those of perianth-bearing Piperales87,91,94,95. We applied the same
554
analytical procedure that we did for Papilionidae. DNA sequences for each gene were aligned
555
using MAFFT 7.11041 with default settings (E-INS-i algorithm and Q-INS-I to take into
556
account secondary structure). Resulting alignments were checked for codon stops and
557
eventually refined by eye with Mesquite 3.1 (available at: www.mesquiteproject.org). The
558
best-fit partitioning schemes and substitution models for phylogenetic analyses were
559
determined with PartitionFinder 2.1.142. All gene alignments were concatenated into a
560
supermatrix; the final dataset is available in Figshare (see Data availability).
561
Phylogenetic relationships were estimated with Bayesian inference as implemented in
562
MrBayes 3.2.647. Rather than using a single substitution model per molecular partition, we
563
sampled across the entire substitution-model space48 using rj-MCMC. Two independent
564
analyses with one cold chain and seven heated chains, each were run for 50 million
565
generations, sampled every 5,000 generations. Convergence and performance of Bayesian
566
runs were evaluated using Tracer 1.7.149 and the ESS, ADSF and PSRF criteria. Once
567
convergence was achieved, a 50% majority-rule consensus tree was built after discarding
568
25% of the sampled trees as burn-in.
569
Bayesian relaxed-clock methods were used that accounted for rate variation across
570
lineages53. MCMC analyses implemented in BEAST 1.8.454 were employed to approximate
571
the posterior distribution of rates and divergences times and infer their credibility intervals.
572
Estimation of divergence times relied on constraining clade ages through fossil calibrations.
573
Three unambiguous fossils from perianth-bearing Piperales (Aristolochiaceae sensu lato), and
574
one corresponding to the family Saururaceae were used. First, we relied on the fossil record
575
of the monotypic family Lactoridaceae (Lactoris fernandeziana)87,92, a shrub endemic to
576
cloud forest of the Juan Fernández Islands archipelago of Chile. The oldest pollen fossil for
577
the group is †Lactoripollenites africanus96,97 from the Turonian/Campanian (72.1-89.8 Ma)
578
of the Orange Basin in South Africa. This fossil confers a minimum age of 72.1 Ma for the
579
23
stem node of Lactoris fernandeziana. Second, the oldest and only pollen record of the
580
Aristolochiaceae was recently described from Late Cretaceous sediments of Siberia:
581
†Aristolochiacidites viluiensis98 from the Timerdyakh Formation of the latest Campanian to
582
earliest Maastrichtian (66-72.1 Ma) in the Vilui Basin (Russia). Because inaperturate pollen
583
grains in combination with this unique exine configuration and fitting size can be observed in
584
extant members of Aristolochiaceae, this fossil provides a minimum age of 66 Ma for the
585
family. The third fossil belongs to the genus Aristolochia and described as †Aristolochia
586
austriaca99 from the Pannonian (late Miocene) in the Hollabrunn-Mistelbach Formation
587
(Austria). Based on a thorough morphological leaf comparison, this fossil is assigned to a
588
species group including Aristolochia baetica and Aristolochia rotunda, which then confers a
589
minimum age of 7.25 Ma for the clade. Finally, we used the fossil †Saururus tuckerae100
590
from the Princeton Chert of Princeton in British Columbia (Canada), which is part of the
591
Princeton Group, Allenby Formation dated with stable isotopes to the middle Eocene101. This
592
fossil has been phylogenetically placed as sister to extant Saururus species101, hence
593
providing a minimum age of 44.3 Ma for the stem node of Saururus. Absolute ages of
594
geological formations were taken from the latest update of the geological time scale.
595
We set the following settings and priors: a partitioned dataset (after the best-fitting
596
PartitionFinder scheme) was analysed using the uncorrelated lognormal distribution clock
597
model, with the mean set to a uniform prior between 0 and 1, and an exponential prior
598
(lambda = 0.333) for the standard deviation. The branching process prior was set to a birth–
599
death 67 process, using the following uniform priors: the birth–death mean growth rate ranged
600
between 0 and 10 with a starting value at 0.1, and the birth–death relative death rate ranged
601
between 0 and 1 (starting value = 0.5). We performed four independent BEAST analyses for
602
100 million generations, sampled every 10,000th, resulting in 10,000 samples in the posterior
603
distribution of which the first 2500 samples were discarded as burn-in. All analyses were
604
performed on the CIPRES Science Gateway computer cluster51, using BEAGLE52.
605
Convergence and performance of each MCMC run were evaluated using Tracer 1.7.149 and
606
the ESS for each parameter. We combined the four runs using LogCombiner 1.8.454. The
607
MCC tree was reconstructed with median age and 95% CI. The BEAST files generated for
608
this study are available in Figshare (see Data availability).
609
610
Dual biogeographic history of Papilionidae and Aristolochiaceae. We estimated the
611
ancestral area of origin and geographic range evolution for both clades using the ML
612
approach of DEC model69 as implemented in the C++ version102,103 that is available at:
613
24
https://github.com/champost/DECX. To infer the biogeographic history of a clade, DEC
614
requires a time-calibrated tree, the current distribution of each species for a set of geographic
615
areas, and a time-stratified geographic model that is represented by connectivity matrices for
616
specified time intervals spanning the entire evolutionary history of the group.
617
The geographic distribution for each species in Papilionidae72–74 and Aristolochiaceae
618
was categorized as present or absent in each of the following areas: (1) West Nearctic [WN],
619
(2) East Nearctic [EN], (3) Central America [CA], (4) South America [SA], (5) West
620
Palearctic [WP], (6) East Palearctic [EP], (7) Madagascar [MD], (8) Indonesia and Wallacea
621
[WA], (9) India [IN], (10) Africa [AF], and (11) Australasia [AU]. The resulting matrices of
622
species distribution for the two groups are available in Figshare (see Data availability).
623
A time-stratified geographic model was built using connectivity matrices that take
624
into account paleogeographic changes through time, with time slices indicating the possibility
625
or not for a species to access a new area103. Based on palaeogeographical reconstructions104–
626
106, we created a connectivity matrix for each geological epoch that represented a period
627
bounded by major changes in tectonic and climatic conditions thought to have affected the
628
distribution of organisms. The following geological epochs were selected: (i) 0 to 5.33 Ma
629
(Pliocene to present), (ii) 5.33 to 23.03 Ma (Miocene), (iii) 23.03 to 33.9 Ma (Oligocene), (iv)
630
33.9 to 56 Ma (Eocene), and (v) 56 Ma to the origin of the clade (Palaeocene to Late
631
Cretaceous). For each of these five time intervals, we specified constraints on area
632
connectivity by coding 0 if any two areas are not connected or 1 if they are connected in a
633
given time interval. We assumed a conservative dispersal matrix with equal dispersal rates
634
between areas through time107.
635
636
Impact of host-plant shifts on swallowtail diversification. We tested the effect of host-plant
637
association on diversification by estimating speciation and extinction rates with five methods
638
to cross-test hypotheses and corroborate results. Analyses were performed on 100 dated trees
639
randomly sampled from the Bayesian dating analyses to take into account the uncertainty in
640
age estimates. We used the following approaches: (i) ML-based trait-dependent
641
diversification108,109; (ii) ML-based time-dependent diversification110; (iii) Bayesian analysis
642
of macroevolutionary mixture111; (iv) Bayesian branch-specific diversification rates112; and
643
(v) Bayesian episodic birth-death model113. It is worth mentioning that each method differs at
644
several points in their estimation of speciation and extinction rates. For instance, trait-
645
dependent birth-death models estimate constant speciation and extinction rates 109, whereas
646
time-dependent birth-death models estimate clade-specific speciation and extinction rates and
647
25
their variation through time110,112. Therefore, we expect some differences in the values of
648
estimated diversification rates that are inherent to each approach. Our diversification analyses
649
should be seen as complementary to the inferred diversification trend rather than
650
corroborating the values and magnitude of speciation and extinction rates.
651
Firstly, we computed the probability of obtaining a clade as large as size n, given the
652
crown age of origin, the overall net diversification rate of the family, and an extinction rate as
653
a fraction of speciation rate following the approach in Condamine et al.17 relying on the
654
method of moments114. We used the R-package LASER 2.3115 to estimate the net
655
diversification rates of Papilionidae and six clades shifting to new host plants with the bd.ms
656
function (providing crown age and total species diversity). Then we used the crown.limits
657
function to estimate the mean expected clade size for each clade shifting to new host plants
658
given clades’ crown age and overall net diversification rates, and we finally computed the
659
probability to observe such clade size using the crown.p function. All rate estimates were
660
calculated with three ε values (ε=0/0.5/0.9), knowing that the extinction rate in swallowtails
661
is usually low17 (supported by the results of this study).
662
First, we relied on the state-dependent speciation and extinction (SSE) model, in
663
which speciation and extinction rates are associated with phenotypic evolution of a trait along
664
a phylogeny108. In particular, we used the Multiple State Speciation Extinction model
665
(MuSSE109) implemented in the R-package diversitree 0.9–10116, which allows multiple
666
character states to be studied. Larval host-plant data were taken from previous works1,12,17,72–
667
74,117. The following 10 host-plant character states and corresponding ratios of sampled
668
species in the tree of all known species for each character (sampling fractions) were used: 1 =
669
Aristolochiaceae (110/152), 2 = Annonaceae (69/138), 3 = Lauraceae (33/39), 4 = Apiaceae
670
(9/10), 5 = Rutaceae (119/163), 6 = Crassulaceae (19/19), 7 = Papaveraceae (44/44), 8 =
671
Fabaceae (1/1), 9 = Zygophyllaceae (2/2), and 10 = Magnoliaceae (2/2). Data at a lower
672
taxonomic level than plant family were not used because of the large number of multiple
673
associations exhibited by genera that could alter the phylogenetic signal. We assigned a
674
single state to each species by selecting the foodplant with the maximum number of
675
collections for each species. We did not employ multiple states per species, which represents
676
a lesser problem because (i) few swallowtail species feed on multiple plant families, (ii)
677
current shared-state models can only model two states, and (iii) the addition of multi-plant
678
states to the MuSSE analysis would have greatly increased the number of parameters. We
679
performed both ML and Bayesian MCMC analyses (10,000 steps) performed using an
680
exponential (1/(2 × net diversification rate)) prior with starting parameter values obtained
681
26
from the best-fitting ML model and resulting speciation, extinction and transition rates. After
682
a burnin of 500 steps, we estimated posterior density distribution for speciation, extinction
683
and transition rates. There have been concerns about the power of SSE models to infer
684
diversification dynamics from a distribution of species traits118–120, hence other birth-death
685
models were used to corroborate the results obtained with SSE models.
686
To provide an independent assessment of the relationship between diversification
687
rates and host specificity, we used the ML approach of Morlon et al.110 implemented in the R-
688
package RPANDA 1.3121. This is a birth–death method in which speciation and/or extinction
689
rates may change continuously through time. This method has the advantage of not assuming
690
constant extinction rate over time (unlike BAMM111), and allows clades to have declining
691
diversity since extinction can exceed speciation, meaning that diversification rates can be
692
negative110. For each clade that shifted to a new host family, we designed and fitted six
693
diversification models: (i) a Yule model, where speciation is constant and extinction is null;
694
(ii) a constant birth-death model, where speciation and extinction rates are constant; (iii) a
695
variable speciation rate model without extinction; (iv) a variable speciation rate model with
696
constant extinction; (v) a rate-constant speciation and variable extinction rate model; and (vi)
697
a model in which both speciation and extinction rates vary. Models were compared by
698
computing the ML estimate of each model and the resulting Akaike information criterion
699
corrected by sample size (AICc) We then plotted rates through time with the best fit model
700
for each clade, and the rates for the family as a whole for comparison purpose.
701
We also performed models that allow diversification rates to vary among clades across the
702
whole phylogeny. BAMM 2.5111,122 was used to explore for differential diversification
703
dynamic regimes among clades differing in their host-plant feeding. BAMM can
704
automatically detect rate shifts and sample distinct evolutionary dynamics that explain the
705
diversification dynamics of a clade without a priori hypotheses on how many and where
706
these shifts might occur. Evolutionary dynamics can involve time-variable diversification
707
rates; in BAMM, speciation is allowed to vary exponentially through time while extinction is
708
maintained constant: subclades in a tree may diversify faster (or slower) than others. This
709
Bayesian approach can be useful in detecting shifts of diversification potentially associated
710
with key innovations123. BAMM analyses were run with four MCMC for 10 million
711
generations, sampling every 10,000th and with three different values (1, 5 and 10) of the
712
compound Poisson prior (CPP) to ensure the posterior is independent of the prior124. We
713
accounted for non-random incomplete taxon sampling using the implemented analytical
714
correction; we set a sampling fraction per genus based on the known species diversity of each
715
27
genus. Mixing and convergence among runs (ESS > 200 after 15% burn-in) were assessed
716
with the R-package BAMMtools 2.1125 to estimate (i) the mean global rates of diversification
717
through time, (ii) the estimated number of rate shifts evaluating alternative diversification
718
models comparing priors and posterior probabilities, and (iii) the clade-specific rates through
719
time when a distinct macroevolutionary regime is identified.
720
BAMM has been criticized for incorrectly modelling rate-shifts on extinct lineages,
721
that is, unobserved (extinct or unsampled) lineages inherit the ancestral diversification
722
process and cannot experience subsequent diversification-rate shifts124,126. To solve this, we
723
used a novel Bayesian approach implemented in RevBayes 1.0.10127 that models rate shifts
724
consistently on extinct lineages by using the SSE framework 112,124. Although there is no
725
information of rate shifts for unobserved/extinct lineages in a phylogeny including extant
726
species only, these types of events must be accounted for in computing the likelihood. The
727
number of rate categories is fixed in the analysis but RevBayes allows any number to be
728
specified, thus allowing direct comparison of different macroevolutionary regimes.
729
Finally, we evaluated the impact of abrupt changes in diversification using the
730
Bayesian episodic birth-death model of CoMET113 implemented in the R-package TESS
731
2.1128. These models allow detection of discrete changes in speciation and extinction rates
732
concurrently affecting all lineages in a tree, and estimate changes in diversification rates at
733
discrete points in time, but can also infer mass extinction events (sampling events in which
734
the extant diversity is reduced by a fraction129). Speciation and extinction rates can change at
735
those points but remain constant within time intervals. In addition, TESS uses independent
736
CPPs to simultaneously detect mass extinction events and discrete changes in speciation and
737
extinction rates, while TreePar estimates the magnitude and timing of speciation and
738
extinction changes independently to the occurrence of mass extinctions (i.e. the three
739
parameters cannot be estimated simultaneously due to parameter identifiability issues129). We
740
performed two independent analyses allowing and disallowing mass extinction events. Bayes
741
factor comparisons were used to assess model fit between models with varying number and
742
time of changes in speciation/extinction rates and mass extinctions.
743
744
Detecting genome-wide adaptations during host-plant shifts. We analysed genomic
745
sequence data in swallowtails that have independently shifted to new ecological (biological)
746
traits. Similar approaches have been conducted on mammals130,131 and birds132, but have been
747
rarely implemented on arthropod groups and, to our knowledge, this is the first time over
748
such a long geological time scale. Here we estimated swallowtail molecular evolution with
749
28
whole genome data and compared selection regimes on protein-coding genes along
750
independent branches with or without host-plant shift and/or environmental shift.
751
For these analyses, we studied 45 whole genomes33 covering all 32 genera of the
752
family Papilionidae: 41 of which were previously generated by our research group added to
753
four genomes already available30–32. In summary, raw reads (Sequence Read Archive:
754
SRR8954507-SRR8954549) were cleaned using Trimmomatic 0.33133, and assembled into
755
contigs and scaffolds with SOAPdenovo-63mer 2.04134 to obtain whole genome assemblies
756
(30x average read depth33). All coding DNA sequences (CDS) were retrieved from the high-
757
quality annotated genome of Papilio xuthus31. To annotate the sequences of all our genomes,
758
a BLAST search using all available CDS of Papilio xuthus was performed at the amino-acid
759
level (using tblastn). For each species the recovered genes were aligned one by one with
760
Papilio xuthus using TranslatorX135. This method performs alignment at the amino-acid level
761
and preserves the open reading frame. All sites showing intraspecific variation were set to N,
762
to conservatively avoid false informative sites. Any contamination was removed using CroCo
763
0.1136 and orthologous proteins were identified with OrthoFinder 2.2.0137. Finally, CDS
764
alignments were strongly cleaned from misaligned sequences (gene by gene) using
765
HMMCleaner 1.8138. A last cleaning step was performed using trimAl 1.2rev59139, which is
766
designed to trim alignments for large-scale phylogenomic analyses. The resulting dataset
767
comprised 6,621 genes in at least four sampled species (median of 32% of missing data),
768
which was used to reconstruct a robust phylogenomic tree of Papilionidae33 (Supplementary
769
Fig. 18).
770
We used this genomic dataset of 45 for all consisting on all genera in which the
771
resulting genus-level swallowtail phylogenomic tree33 accurately represents the evolutionary
772
associations with host plants as estimated using the ancestral-state analyses applied to the
773
species-level phylogeny17 (Fig. 1, Supplementary Figs. 4, 5). We thus transferred the
774
inference of ancestral host-plant shifts on the phylogenomic tree and selected the branches
775
representing a host-plant shift and branches with a shift of climate preference (in general
776
from tropical to temperate conditions; Supplementary Fig. 10). We also selected branches
777
with no change as negative controls34. To test the impact of these different changes on the
778
genomes, two datasets were created, Dataset 1 and 2. Dataset 1 consists of 1,533 genes
779
selected from the 6,621-gene dataset for each focal branch using three criteria: (1) the dataset
780
is composed only of orthologous protein-coding genes (OrthoFinder 2.2137) , (2) the species
781
needed to accurately define the branch were available (i.e. crown node of the clade), and (3)
782
for each branch, one species per tribe was available, and therefore include a different number
783
29
of genes per branch. Dataset 2 comprises 520 genes necessary to define all focal branches
784
leading to less selected genes but the same genes for all branches. As a result, 14 branches are
785
selected to measure the impact of a host-plant shift and 14 branches are selected as controls
786
(Supplementary Fig. 18). Within these 28 branches, some branches represent environmental
787
shifts (from tropical to temperate climate). The genomic dataset is available in Figshare (see
788
Data availability).
789
We studied the ratio (ω) of nonsynonymous/synonymous substitution rate (dN/dS) to
790
find genes under positive selection37,140. The dN/dS ratio is traditionally used to estimate
791
selective pressure from protein-coding sequences. If host-plant shifts have no effect on the
792
selection of a given gene, we expect a dN/dS = 1 and the selective regime is considered
793
neutral. However, if host-plant shifts result in positive selection on coding genes, the ratio
794
increases such that dN/dS > 1. Finally, it is possible that host-plant shifts lead to purifying
795
selection, thus reducing the number of non-synonymous substitutions and resulting in dN/dS
796
< 1. Here we focused on the adaptation of Papilionidae to host plant shifts, i.e. outgroups are
797
not studied. We tested if branches representing inferred host-plant shifts along the phylogeny
798
of swallowtails have more genes with dN/dS > 1, representing adaptation, than branches
799
representing host-plant conservatism. The branch-site models allow ω to vary both among
800
sites in the protein and across branches on the tree and aim to detect positive selection
801
affecting a few sites along particular lineages. The approach described by Zhang et al.141 is
802
chosen to determine genome-wide selection regimes as performed with two maximum-
803
likelihood models: (1) a null model assuming two site classes, one with dN/dS < 1 and one
804
with dN/dS = 1; and (2) an alternative model adding a third site class with dN/dS > 1. The fit
805
for including positive selection is tested using a likelihood ratio test comparing the null model
806
with the alternative model with one degree of freedom37,142. If the alternative model is better
807
suited to host-shift branches, it is more likely the gene was under positive selection during the
808
host-plant shifts. For each gene, dN/dS is estimated with both the null and alternative models
809
using CodeML implemented in PAML 4143. To test the robustness of the estimations, we used
810
a false discovery rate test to control false positives144. Finally, we reported the number of
811
genes under positive selection on the total gene number for each focal branch. The number of
812
genes under positive selection was compared between branches representing host-plant shifts,
813
environmental shifts, both plant and environmental shifts or no shifts using the non-
814
parametric Wilcoxon signed-rank test145.
815
816
30
Sensitivity analyses. We performed several control analyses to ensure that the signal of more
817
genes under positive selection in host-plant shifts branches is not artefactual. Specifically, we
818
focused on missing data and GC content variation among genes known to bias dN/dS
819
estimations. Missing data are prone to introducing misaligned regions that could create false
820
positives in branch-site likelihood method for detecting positive selection146–148. Variations in
821
GC content are known to impact the estimation of dN/dS mainly through the process of GC-
822
biased gene conversion (gBGC149–151).
823
The number of missing data (‘N’ and ‘-’) sites and GC content at the third codon
824
position (GC3) were computed using a home-made C++ program created with BIO++
825
library152. Mean GC content and missing data was calculated per gene and for each branch.
826
For a given branch, mean GC3 and missing data were computed for the species of a clade for
827
which the branch is the root. All statistics and graphical representations were performed using
828
the R-packages tidyverse153 and cowplot154. We found that genes under positive selection
829
(PSgenes, nDataset1 = 142, nDataset2 = 407) have significantly more missing data and GC3 than
830
genes not under positive selection (NSgenes, nDataset1 = 378, nDataset2 = 1126; P = 0.001 / 0.02
831
for the two datasets, respectively, Mann-Whitney test; Supplementary Fig. 19). This result
832
confirms that branch-site likelihood methods for detecting positive selection are sensitive to
833
missing data, probably because of misaligned sites146,147, and that GC content that may be
834
influenced by gBGC149.
835
Missing data was, however, heterogeneously distributed among species, ranging from
836
less than 1% in Papilio xuthus to 45% in Hypermnestra helios (Supplementary Fig. 20). The
837
difference in missing data between branches with (n = 14, mean missing Dataset1 = 13.4%,
838
mean missingDataset2 = 14.1%) or without host-plant shifts (n = 14, mean missingDataset1 =
839
12.8%, mean missingDataset2 = 12.7%) is not significant (P = 0.83 / 1.00 for the two datasets,
840
respectively, Mann-Whitney test; Supplementary Fig. 21). Additionally, there is no
841
correlation between the number of genes under positive selection and the amount of missing
842
data (P = 0.33 / 0.20 for the two datasets, respectively, Spearman’s correlation test;
843
Supplementary Fig. 22). For GC3, we also found variation between species ranging from
844
37% in Parnassius smintheus to 44% in Papilio antimachus (Supplementary Fig. 23).
845
Similarly to missing data, we found no significant difference between plant-shift and no
846
plant-shift branches (P = 0.63 / 0.63 for the two datasets, Mann-Whitney test; Supplementary
847
Fig. 24) and there is no correlation between the number of genes under positive selection and
848
GC3 (P = 0.20 / 0.1362 for the two datasets, respectively, Spearman’s correlation test;
849
Supplementary Fig. 25).
850
31
Despite the known fact that false positives can increase with the amount of missing
851
data, our control analyses indicate that variations in missing data and GC content do not drive
852
the signal that more genes are under positive selection in branches that have undergone a
853
host-plant shift. Additionally to these controls, we checked by eyes all the gene alignments at
854
the amino-acid level for genes under positive selection in branches with and without host-
855
plant shifts using SeaView 4155. Misaligned regions, which could lead to biased dN/dS
856
ratios156, were not significantly more detected for genes under positive selection in branches
857
with host-plant shifts. In some cases we found ourselves in complicated situations to
858
discriminate between false and true positive selected genes.
859
Overall, given the our alignment checks and sensitivity analyses, we do not see any
860
reason for biased dN/dS ratios in genes along branches with or without host-plant shifts.
861
False positive and false negative genes can be present in the two categories of branches but,
862
in any cases, the general pattern observed is likely to remain conserved.
863
864
Gene ontology. To annotate proteins of our alignment, we used the two different approaches
865
implemented in PANTHER 14157 (available at: http://pantherdb.org/) and EggNOG 5.0158,159
866
(available at: http://eggnog5.embl.de/#/app/home). We used the HMM Scoring tool to assign
867
PANTHER family (library version 14.1157) to the protein of Papilio xuthus (assembly
868
Pxut_1.0); similar results were obtained using another high-quality annotated genome (from
869
Heliconius melpomene) as reference (assembly ASM31383v2). We performed the statistical
870
overrepresentation test implemented on the PANTHER online website, relying on the GO
871
categories in the PANTHER GO-Slim annotation dataset including Molecular function,
872
Biological process, and Cellular component. Firstly, we tested if positively selected genes
873
have over- or under-represented functional GO categories as compared to the whole set of
874
genes (option “PANTHER Generic Mapping”). Secondly, we tested if positively selected
875
genes involving a host-plant shift along the 14 branches have over- or under-represented
876
functional categories. These statistical comparisons were performed with the Fisher’s exact
877
test using the false discovery rate correction to control for false positives. Independently, we
878
used the eggNOG-mapper v2158 (https://github.com/eggnogdb/eggnog-mapper) and the
879
associated Lepidoptera database (LepNOG, including the genomes of Bombyx mori, Danaus
880
plexippus and Heliconius melpomene159) to annotate the proteins of our dataset. EggNOG
881
uses precomputed orthologous groups and phylogenies from the database to transfer
882
functional information from fine-grained orthologs only. We used the diamond method as
883
32
recommended158. Finally, we reported the known functions of proteins that were only
884
positively selected when there was a host-plant shift in the phylogeny.
885
886
Data availability
887
All data, including supermatrix datasets (for phylogenetic analyses), phylogenetic trees, host-
888
plant preferences, species geographic distributions, gene alignments (for dN/dS analyses) and
889
bioinformatic scripts, that are necessary for repeating the analyses described here have been
890
made
available
through
the
Figshare
digital
data
repository
891
(https://figshare.com/s/1ce98308a3c012514857).
892
893
References
894
1.
Ehrlich, P. R. & Raven, P. H. Butterflies and plants: a study in coevolution. Evolution
895
18, 586–608 (1964).
896
2.
Thompson, J. N. Concepts of coevolution. Trends Ecol. Evol. 4, 179–183 (1989).
897
3.
Berenbaum, M. R. & Feeny, P. P. Chemical mediation of host-plant specialization: the
898
papilionid paradigm. in Specialization, Speciation, and Radiation: The Evolutionary
899
Biology of Herbivorous Insects (ed. Tilmon, K.) 2–19 (University of California Press,
900
2008).
901
4.
Winter, S., Friedman, A. L. L., Astrin, J. J., Gottsberger, B. & Letsch, H. Timing and
902
host plant associations in the evolution of the weevil tribe Apionini (Apioninae,
903
Brentidae, Curculionoidea, Coleoptera) indicate an ancient co-diversification pattern of
904
beetles and flowering plants. Mol. Phylogenet. Evol. 107, 179–190 (2017).
905
5.
Kergoat, G. J. et al. Opposite macroevolutionary responses to environmental changes
906
in grasses and insects during the Neogene grassland expansion. Nat. Commun. 9, 5089
907
(2018).
908
6.
Wheat, C. W. et al. The genetic basis of a plant–insect coevolutionary key innovation.
909
Proc. Natl. Acad. Sci. 104, 20427–20431 (2007).
910
7.
Edger, P. P. et al. The butterfly plant arms-race escalated by gene and genome
911
duplications. Proc. Natl. Acad. Sci. U. S. A. 112, 8362–8366 (2015).
912
8.
Calla, B. et al. Cytochrome P450 diversification and hostplant utilization patterns in
913
specialist and generalist moths: Birth, death and adaptation. Mol. Ecol. 26, 6021–6035
914
(2017).
915
9.
Nallu, S. et al. The molecular genetic basis of herbivory between butterflies and their
916
host plants. Nat. Ecol. Evol. 2, 1418–1427 (2018).
917
33
10.
Karageorgi, M. et al. Genome editing retraces the evolution of toxin resistance in the
918
monarch butterfly. Nature 574, 409–412 (2019).
919
11.
Sahoo, R. K., Warren, A. D., Collins, S. C. & Kodandaramaiah, U. Hostplant change
920
and paleoclimatic events explain diversification shifts in skipper butterflies (Family:
921
Hesperiidae). BMC Evol. Biol. 17, 174 (2017).
922
12.
Condamine, F. L., Rolland, J., Höhna, S., Sperling, F. A. H. & Sanmartín, I. Testing
923
the role of the red queen and court jester as drivers of the macroevolution of apollo
924
butterflies. Syst. Biol. 67, 940–964 (2018).
925
13.
Letsch, H. et al. Climate and host-plant associations shaped the evolution of
926
ceutorhynch weevils throughout the Cenozoic. Evolution 72, 1815–1828 (2018).
927
14.
Hua, X. & Bromham, L. Darwinism for the genomic age: connecting mutation to
928
diversification. Front. Genet. 8, 12 (2017).
929
15.
Schmeiser, H. H., Stiborovà, M. & Arlt, V. M. Chemical and molecular basis of the
930
carcinogenicity of Aristolochia plants. Curr. Opin. Drug Discov. Devel. 12, 141–148
931
(2009).
932
16.
Poon, S. L. et al. Genome-wide mutational signatures of aristolochic acid and its
933
application as a screening tool. Sci. Transl. Med. 5, 197ra101 (2013).
934
17.
Condamine, F. L., Sperling, F. A. H., Wahlberg, N., Rasplus, J.-Y. & Kergoat, G. J.
935
What causes latitudinal gradients in species diversity? Evolutionary processes and
936
ecological constraints on swallowtail biodiversity. Ecol. Lett. 15, 267–277 (2012).
937
18.
Simonsen, T. J. et al. Phylogenetics and divergence times of Papilioninae
938
(Lepidoptera) with special reference to the enigmatic genera Teinopalpus and
939
Meandrusa. Cladistics 27, 113–137 (2011).
940
19.
McKenna, D. D., Sequeira, A. S., Marvaldi, A. E. & Farrell, B. D. Temporal lags and
941
overlap in the diversification of weevils and flowering plants. Proc. Natl. Acad. Sci. U.
942
S. A. 106, 7083–7088 (2009).
943
20.
Nishida, R. Sequestration of Defensive Substances from Plants by Lepidoptera. Annu.
944
Rev. Entomol. 47, 57–92 (2002).
945
21.
Moen, D. & Morlon, H. Why does diversification slow down? Trends Ecol. Evol. 29,
946
190–197 (2014).
947
22.
Losos, J. B. Adaptive radiation, ecological opportunity, and evolutionary determinism.
948
Am. Nat. 175, 623–639 (2010).
949
23.
Cohen, M. B., Schuler, M. A. & Berenbaum, M. R. A host-inducible cytochrome P-
950
450 from a host-specific caterpillar: molecular cloning and evolution. Proc. Natl.
951
34
Acad. Sci. U. S. A. 89, 10920–10924 (1992).
952
24.
Berenbaum, M. R., Favret, C. & Schuler, M. A. On defining ‘Key Innovations’ in an
953
adaptive radiation: Cytochrome P450S and Papilionidae. Am. Nat. 148, S139–S155
954
(1996).
955
25.
Li, W., Schuler, M. A. & Berenbaum, M. R. Diversification of furanocoumarin-
956
metabolizing cytochrome P450 monooxygenases in two papilionids: Specificity and
957
substrate encounter rate. Proc. Natl. Acad. Sci. U. S. A. 100 Suppl, 14593–14598
958
(2003).
959
26.
Schuler, M. A. P450s in plant–insect interactions. Biochim. Biophys. Acta - Proteins
960
Proteomics 1814, 36–45 (2011).
961
27.
Cheng, T. et al. Genomic adaptation to polyphagy and insecticides in a major East
962
Asian noctuid pest. Nat. Ecol. Evol. 1, 1747–1756 (2017).
963
28.
Rane, R. V et al. Detoxifying enzyme complements and host use phenotypes in 160
964
insect species. Curr. Opin. Insect Sci. 31, 131–138 (2019).
965
29.
Thompson, J. N., Wehling, W. & Podolsky, R. Evolutionary genetics of host use in
966
swallowtail butterflies. Nature 344, 148–150 (1990).
967
30.
Cong, Q., Borek, D., Otwinowski, Z. & Grishin, N. V. Tiger swallowtail genome
968
reveals mechanisms for speciation and caterpillar chemical defense. Cell Rep. 10, 910–
969
919 (2015).
970
31.
Li, X. et al. Outbred genome sequencing and CRISPR/Cas9 gene editing in butterflies.
971
Nat. Commun. 6, 8212 (2015).
972
32.
Nishikawa, H. et al. A genetic mechanism for female-limited Batesian mimicry in
973
Papilio butterfly. Nat. Genet. 47, 405–409 (2015).
974
33.
Allio, R. et al. Whole genome shotgun phylogenomics resolves the pattern and timing
975
of swallowtail butterfly evolution. Syst. Biol. 69, 38–60 (2020).
976
34.
Thomas, G. W. C. & Hahn, M. W. Determining the null model for detecting adaptive
977
convergence from genomic data: A case study using echolocating mammals. Mol. Biol.
978
Evol. 32, 1232–1236 (2015).
979
35.
Zou, Z. & Zhang, J. No genome-wide protein sequence convergence for echolocation.
980
Mol. Biol. Evol. 32, 1237–1241 (2015).
981
36.
Kimura, M. The Neutral Theory of Molecular Evolution. (Cambridge University Press,
982
1983).
983
37.
Yang, Z. Computational Molecular Evolution. (Oxford University Press, 2006).
984
38.
Dasmahapatra, K. K. et al. Butterfly genome reveals promiscuous exchange of
985
35
mimicry adaptations among species. Nature 487, 94–98 (2012).
986
39.
Thomas, G. W. C. et al. Gene content evolution in the arthropods. Genome Biol. 21, 15
987
(2020).
988
40.
de Medeiros, B. A. S. & Farrell, B. D. Evaluating species interactions as a driver of
989
phytophagous insect divergence. bioRxiv 842153 (2019). doi:10.1101/842153
990
41.
Katoh, K. & Standley, D. M. MAFFT multiple sequence alignment oftware version 7:
991
Improvements in performance and usability. Mol. Biol. Evol. 30, 772–780 (2013).
992
42.
Lanfear, R., Frandsen, P. B., Wright, A. M., Senfeld, T. & Calcott, B. PartitionFinder
993
2: New methods for selecting partitioned models of evolution for molecular and
994
morphological phylogenetic analyses. Mol. Biol. Evol. 34, 772–773 (2016).
995
43.
Nguyen, L.-T., Schmidt, H. A., von Haeseler, A. & Minh, B. Q. IQ-TREE: A fast and
996
effective stochastic algorithm for estimating maximum-likelihood phylogenies. Mol.
997
Biol. Evol. 32, 268–274 (2015).
998
44.
Kalyaanamoorthy, S., Minh, B. Q., Wong, T. K. F., von Haeseler, A. & Jermiin, L. S.
999
ModelFinder: fast model selection for accurate phylogenetic estimates. Nat. Methods
1000
14, 587–589 (2017).
1001
45.
Chernomor, O., von Haeseler, A. & Minh, B. Q. Terrace aware data structure for
1002
phylogenomic inference from supermatrices. Syst. Biol. 65, 997–1008 (2016).
1003
46.
Minh, B. Q., Nguyen, M. A. T. & von Haeseler, A. Ultrafast approximation for
1004
phylogenetic bootstrap. Mol. Biol. Evol. 30, 1188–1195 (2013).
1005
47.
Ronquist, F. et al. MrBayes 3.2: Efficient bayesian phylogenetic inference and model
1006
choice across a large model space. Syst. Biol. 61, 539–542 (2012).
1007
48.
Huelsenbeck, J. P., Larget, B. & Alfaro, M. E. Bayesian phylogenetic model selection
1008
using reversible jump Markov Chain Monte Carlo. Mol. Biol. Evol. 21, 1123–1133
1009
(2004).
1010
49.
Rambaut, A., Drummond, A. J., Xie, D., Baele, G. & Suchard, M. A. Posterior
1011
summarization in bayesian phylogenetics using Tracer 1.7. Syst. Biol. 67, 901–904
1012
(2018).
1013
50.
Douady, C. J., Delsuc, F., Boucher, Y., Doolittle, W. F. & Douzery, E. J. P.
1014
Comparison of bayesian and maximum likelihood bootstrap measures of phylogenetic
1015
reliability. Mol. Biol. Evol. 20, 248–254 (2003).
1016
51.
Miller, M. A. et al. A RESTful API for access to phylogenetic tools via the CIPRES
1017
Science Gateway. Evol. Bioinforma. 11, EBO.S21501 (2015).
1018
52.
Ayres, D. L. et al. BEAGLE: An application programming interface and high-
1019
36
performance computing library for statistical phylogenetics. Syst. Biol. 61, 170–173
1020
(2012).
1021
53.
Drummond, A. J., Ho, S. Y. W., Phillips, M. J. & Rambaut, A. Relaxed phylogenetics
1022
and dating with confidence. PLoS Biol. 4, e88 (2006).
1023
54.
Drummond, A. J., Suchard, M. A., Xie, D. & Rambaut, A. Bayesian phylogenetics
1024
with BEAUti and the BEAST 1.7. Mol. Biol. Evol. 29, 1969–1973 (2012).
1025
55.
Durden, C. J. & Rose, H. Butterflies from the middle Eocene: the earliest occurrence
1026
of fossil Papilionoidea (Lepidoptera). (Prarce-Sellards Ser. Tax. Mem. Mus., 1978).
1027
56.
Smith, M. E., Singer, B. & Carroll, A. 40Ar/39Ar geochronology of the Eocene Green
1028
River Formation, Wyoming. Geol. Soc. Am. Bull. 115, 549–565 (2003).
1029
57.
de Jong, R. Estimating time and space in the evolution of the Lepidoptera. Tijdschr.
1030
voor Entomol. 150, 319–346 (2007).
1031
58.
de Jong, R. Fossil butterflies, calibration points and the molecular clock (Lepidoptera:
1032
Papilionoidea). Zootaxa 4270, 1–63 (2017).
1033
59.
Scudder, S. H. Fossil butterflies. Mem. Am. Assoc. Adv. Sci. 1, 1–99 (1875).
1034
60.
Rasnitsyn, A. P. & Zherikhin, V. V. Appendix: Alphabetic list of selected insect fossil
1035
sites. in History of Insects 437–446 (Kluwer Academic Publishers, 2002).
1036
doi:10.1007/0-306-47577-4_4
1037
61.
Sohn, J., Labandeira, C., Davis, D. & Mitter, C. An annotated catalog of fossil and
1038
subfossil Lepidoptera (Insecta: Holometabola) of the world. Zootaxa 3286, 1–132
1039
(2012).
1040
62.
Rebel, H. Doritites bosniaskii. Sitzungsberichte der akademie der wissenschaften.
1041
Mathematischen-Naturwissenschaftliche classe. Abteilung 1 Mineral. Biol. Erdkd. 1,
1042
734–741 (1898).
1043
63.
Carpenter, F. Treatise on Invertebrate Paleontology: Arthropoda 4. Superclass
1044
Hexapoda. Geological Society of America (1992).
1045
64.
Magallón, S., Gómez‐Acevedo, S., Sánchez‐Reyes, L. L. & Hernández‐Hernández, T.
1046
A metacalibrated time‐tree documents the early rise of flowering plant phylogenetic
1047
diversity. New Phytol. 207, 437–453 (2015).
1048
65.
Sohn, J.-C., Labandeira, C. C. & Davis, D. R. The fossil record and taphonomy of
1049
butterflies and moths (Insecta, Lepidoptera): implications for evolutionary diversity
1050
and divergence-time estimates. BMC Evol. Biol. 15, 12 (2015).
1051
66.
Toussaint, E. F. A. & Condamine, F. L. To what extent do new fossil discoveries
1052
37
change our understanding of clade evolution? A cautionary tale from burying beetles
1053
(Coleoptera: Nicrophorus ). Biol. J. Linn. Soc. 117, 686–704 (2016).
1054
67.
Gernhard, T. The conditioned reconstructed process. J. Theor. Biol. 253, 769–778
1055
(2008).
1056
68.
Lewis, P. O. A likelihood approach to estimating phylogeny from discrete
1057
morphological character data. Syst. Biol. 50, 913–925 (2001).
1058
69.
Ree, R. H. & Smith, S. A. Maximum likelihood inference of geographic range
1059
evolution by dispersal, local extinction, and cladogenesis. Syst. Biol. 57, 4–14 (2008).
1060
70.
Pagel, M. & Meade, A. Bayesian analysis of correlated evolution of discrete characters
1061
by reversible-jump Markov chain Monte Carlo. Am. Nat. 167, 808–25 (2006).
1062
71.
Igarashi, S. The classification of the Papilionidae mainly based on the morphology of
1063
their immature stages. Lepid. Sci. 34, 41–96 (1984).
1064
72.
Collins, N. M. & Morris, M. Threatened swallowtail butterflies of the world: the IUCN
1065
red data book. (IUCN, 1985).
1066
73.
Tyler, H. A., Brown, K. S. & Wilson, K. H. Swallowtail Butterflies of the Americas: A
1067
Study
in
Biological
Dynamics,
Ecological
Diversity,
Biosystematics,
and
1068
Conservation. (Scientific Publishers, 1994).
1069
74.
Scriber, J., Tsubaki, Y. & Lederhouse, R. Swallowtail Butterflies: Their Ecology and
1070
Evolutionary Biology. (Scientific Publishers, 1995).
1071
75.
Ree, R. H., Moore, B. R., Webb, C. O. & Donoghue, M. J. A likelihood framework for
1072
inferring the evolution of geographic range on phylogenetic trees. Evolution 59, 2299–
1073
2311 (2005).
1074
76.
Massoni, J., Couvreur, T. L. & Sauquet, H. Five major shifts of diversification through
1075
the long evolutionary history of Magnoliidae (Angiosperms). BMC Evol. Biol. 15, 49
1076
(2015).
1077
77.
Kyalangalilwa, B., Boatwright, J. S., Daru, B. H., Maurin, O. & van der Bank, M.
1078
Phylogenetic position and revised classification of Acacia s.l. (Fabaceae:
1079
Mimosoideae) in Africa, including new combinations in Vachellia and Senegalia. Bot.
1080
J. Linn. Soc. 172, 500–523 (2013).
1081
78.
Miller, J. T., Murphy, D. J., Ho, S. Y. W., Cantrill, D. J. & Seigler, D. Comparative
1082
dating of Acacia: combining fossils and multiple phylogenies to infer ages of clades
1083
with poor fossil records. Aust. J. Bot. 61, 436–445 (2013).
1084
79.
Michalak, I., Zhang, L.-B. & Renner, S. S. Trans-Atlantic, trans-Pacific and trans-
1085
Indian Ocean dispersal in the small Gondwanan Laurales family Hernandiaceae. J.
1086
38
Biogeogr. 37, 1214–1226 (2010).
1087
80.
Wu, S.-D. et al. Evolution of asian interior arid-zone biota: Evidence from the
1088
diversification of asian Zygophyllum (Zygophyllaceae). PLoS One 10, e0138697
1089
(2015).
1090
81.
Chase, M. W. et al. An update of the Angiosperm Phylogeny Group classification for
1091
the orders and families of flowering plants: APG IV. Bot. J. Linn. Soc. 181, 1–20
1092
(2016).
1093
82.
Christenhusz, M. J. M., Vorontsova, M. S., Fay, M. F. & Chase, M. W. Results from
1094
an online survey of family delimitation in angiosperms and ferns: recommendations to
1095
the Angiosperm Phylogeny Group for thorny problems in plant classification. Bot. J.
1096
Linn. Soc. 178, 501–528 (2015).
1097
83.
González, F. & Rudall, P. The questionable affinities of Lactoris: Evidence from
1098
branching pattern, inflorescence morphology, and stipule development. Am. J. Bot. 88,
1099
2143–2150 (2001).
1100
84.
Gonzáles, F., Rudall, P. J. & Furness, C. A. Microsporogenesis and systematics of
1101
Aristolochiaceae. Bot. J. Linn. Soc. 137, 221–242 (2001).
1102
85.
Isnard, S. et al. Growth form evolution in Piperales and its relevance for understanding
1103
angiosperm diversification: An integrative approach combining plant architecture,
1104
anatomy, and biomechanics. Int. J. Plant Sci. 173, 610–639 (2012).
1105
86.
Wagner, S. T. et al. Major trends in stem anatomy and growth forms in the perianth-
1106
bearing Piperales, with special focus on Aristolochia. Ann. Bot. 113, 1139–1154
1107
(2014).
1108
87.
Nickrent, D. L. et al. Molecular data place Hydnoraceae with Aristolochiaceae. Am. J.
1109
Bot. 89, 1809–1817 (2002).
1110
88.
Kelly, L. M. & González, F. Phylogenetic relationships in Aristolochiaceae. Syst. Bot.
1111
28, 236–249 (2003).
1112
89.
Neinhuis, C., Wanke, S., Hilu, K. W., Müller, K. & Borsch, T. Phylogeny of
1113
Aristolochiaceae based on parsimony, likelihood, and Bayesian analyses of trnL-trnF
1114
sequences. Plant Syst. Evol. 250, 7–26 (2005).
1115
90.
Wanke, S. et al. Evolution of Piperales—matK gene and trnK intron sequence data
1116
reveal lineage specific resolution contrast. Mol. Phylogenet. Evol. 42, 477–497 (2007).
1117
91.
Naumann, J. et al. Single-copy nuclear genes place haustorial Hydnoraceae within
1118
piperales and reveal a cretaceous origin of multiple parasitic angiosperm lineages.
1119
PLoS One 8, e79204 (2013).
1120
39
92.
Salomo, K. et al. The emergence of earliest angiosperms may be earlier than fossil
1121
evidence indicates. Syst. Bot. 42, 607–619 (2017).
1122
93.
Christenhusz, M. J. M. & Byng, J. W. The number of known plants species in the
1123
world and its annual increase. Phytotaxa 261, 201–217 (2016).
1124
94.
Naumann, J. et al. Detecting and characterizing the highly divergent plastid genome of
1125
the nonphotosynthetic parasitic plant Hydnora visseri (Hydnoraceae). Genome Biol.
1126
Evol. 8, 345–363 (2016).
1127
95.
Jost, M., Naumann, J., Rocamundi, N., Cocucci, A. A. & Wanke, S. The first plastid
1128
genome of the Holoparasitic genus Prosopanche (Hydnoraceae). Plants 9, 306 (2020).
1129
96.
Zavada, M. S. & Benson, J. M. First fossil evidence for the primitive angiosperm
1130
family Lactoricidae. Am. J. Bot. 74, 1590–1594 (1987).
1131
97.
Gamerro, J. C. & Barreda, V. New fossil record of Lactoridaceae in southern South
1132
America: A palaeobiogeographical approach. Bot. J. Linn. Soc. 158, 41–50 (2008).
1133
98.
Hofmann, C.-C. & Zetter, R. Upper Cretaceous sulcate pollen from the Timerdyakh
1134
Formation, Vilui Basin (Siberia). Grana 49, 170–193 (2010).
1135
99.
Meller, B. The first fossil Aristolochia (Aristolochiaceae, Piperales) leaves from
1136
Austria. Palaeontol. Electron. 17, 1–17 (2014).
1137
100. Smith, S. Y. & Stockey, R. A. Establishing a fossil record for the perianthless
1138
Piperales: Saururus tuckerae sp. nov. (Saururaceae) from the Middle Eocene Princeton
1139
Chert. Am. J. Bot. 94, 1642–1657 (2007).
1140
101. Massoni, J., Doyle, J. & Sauquet, H. Fossil calibration of Magnoliidae, an ancient
1141
lineage of angiosperms. Palaeontol. Electron. 18, 1–25 (2015).
1142
102. Smith, S. A. Taking into account phylogenetic and divergence-time uncertainty in a
1143
parametric biogeographical analysis of the Northern Hemisphere plant clade
1144
Caprifolieae. J. Biogeogr. 36, 2324–2337 (2009).
1145
103. Beeravolu, C. R. & Condamine, F. L. An extended maximum likelihood inference of
1146
geographic range evolution by dispersal, local extinction and cladogenesis. bioRxiv
1147
038695 (2016). doi:10.1101/038695
1148
104. Scotese, C. R. A continental drift flipbook. J. Geol. 112, 729–741 (2004).
1149
105. Blakey, R. C. Gondwana paleogeography from assembly to breakup — A 500 m.y.
1150
odyssey. Geol. Soc. Am. Spec. Pap. 441, 1–28 (2008).
1151
106. Seton, M. et al. Global continental and ocean basin reconstructions since 200 Ma.
1152
Earth-Science Rev. 113, 212–270 (2012).
1153
107. Chacón, J. & Renner, S. S. Assessing model sensitivity in ancestral area reconstruction
1154
40
using Lagrange: A case study using the Colchicaceae family. J. Biogeogr. 41, 1414–
1155
1427 (2014).
1156
108. Maddison, W. P., Midford, P. E. & Otto, S. P. Estimating a binary character’s effect on
1157
speciation and extinction. Syst. Biol. 56, 701–710 (2007).
1158
109. FitzJohn, R. G., Maddison, W. P. & Otto, S. P. Estimating trait-dependent speciation
1159
and extinction rates from incompletely resolved phylogenies. Syst. Biol. 58, 595–611
1160
(2009).
1161
110. Morlon, H., Parsons, T. L. & Plotkin, J. B. Reconciling molecular phylogenies with the
1162
fossil record. Proc. Natl. Acad. Sci. U. S. A. 108, 16327–16332 (2011).
1163
111. Rabosky, D. L. et al. Rates of speciation and morphological evolution are correlated
1164
across the largest vertebrate radiation. Nat. Commun. 4, 1958 (2013).
1165
112. Höhna, S. et al. A Bayesian approach for estimating branch-specific speciation and
1166
extinction rates. bioRxiv 555805 (2019). doi:10.1101/555805
1167
113. May, M. R., Höhna, S. & Moore, B. R. A Bayesian approach for detecting the impact
1168
of mass-extinction events on molecular phylogenies when rates of lineage
1169
diversification may vary. Methods Ecol. Evol. 7, 947–959 (2016).
1170
114. Magallon, S. & Sanderson, M. J. Absolute diversification rates in angiosperm clades.
1171
Evolution 55, 1762–1780 (2001).
1172
115. Rabosky, D. L. Likelihood methods for detecting temporal shifts in diversification
1173
rates. Evolution 60, 1152–1164 (2006).
1174
116. FitzJohn, R. G. Diversitree: Comparative phylogenetic analyses of diversification in R.
1175
Methods Ecol. Evol. 3, 1084–1092 (2012).
1176
117. Scriber, J. M. Host-plant suitability. in Chemical Ecology of Insects (eds. Bell, W. J. &
1177
Cardé, R. T.) 159–202 (Springer US, 1984).
1178
118. Davis, M. P., Midford, P. E. & Maddison, W. Exploring power and parameter
1179
estimation of the BiSSE method for analyzing species diversification. BMC Evol. Biol.
1180
13, 38 (2013).
1181
119. Maddison, W. P. & FitzJohn, R. G. The unsolved challenge to phylogenetic correlation
1182
tests for categorical characters. Syst. Biol. 64, 127–136 (2015).
1183
120. Rabosky, D. L. & Goldberg, E. E. Model inadequacy and mistaken inferences of trait-
1184
dependent speciation. Syst. Biol. 64, 340–355 (2015).
1185
121. Morlon, H. et al. RPANDA: An R package for macroevolutionary analyses on
1186
phylogenetic trees. Methods Ecol. Evol. 7, 589–597 (2016).
1187
122. Rabosky, D. L., Donnellan, S. C., Grundler, M. & Lovette, I. J. Analysis and
1188
41
visualization of complex macroevolutionary dynamics: An example from australian
1189
scincid lizards. Syst. Biol. 63, 610–627 (2014).
1190
123. Rabosky, D. L. Automatic detection of key innovations, rate shifts, and diversity-
1191
dependence on phylogenetic trees. PLoS One 9, e89543 (2014).
1192
124. Moore, B. R., Höhna, S., May, M. R., Rannala, B. & Huelsenbeck, J. P. Critically
1193
evaluating the theory and performance of Bayesian analysis of macroevolutionary
1194
mixtures. Proc. Natl. Acad. Sci. U. S. A. 113, 9569–9574 (2016).
1195
125. Rabosky, D. L. et al. BAMMtools: An R package for the analysis of evolutionary
1196
dynamics on phylogenetic trees. Methods Ecol. Evol. 5, 701–707 (2014).
1197
126. Rabosky, D. L., Mitchell, J. S. & Chang, J. Is BAMM flawed? Theoretical and
1198
practical concerns in the analysis of multi-rate diversification models. Syst. Biol. 66,
1199
477–498 (2017).
1200
127. Höhna, S. et al. RevBayes: Bayesian phylogenetic inference using graphical models
1201
and an interactive model-specification language. Syst. Biol. 65, 726–736 (2016).
1202
128. Höhna, S., May, M. R. & Moore, B. R. TESS: An R package for efficiently simulating
1203
phylogenetic trees and performing Bayesian inference of lineage diversification rates.
1204
Bioinformatics 32, 789–791 (2016).
1205
129. Stadler, T. Mammalian phylogeny reveals recent diversification rate shifts. Proc. Natl.
1206
Acad. Sci. U. S. A. 108, 6187–6192 (2011).
1207
130. Partha, R. et al. Subterranean mammals show convergent regression in ocular genes
1208
and enhancers, along with adaptation to tunneling. Elife 6, e25884 (2017).
1209
131. Wu, J., Yonezawa, T. & Kishino, H. Rates of molecular evolution suggest natural
1210
history of life history traits and a Post-K-Pg nocturnal bottleneck of placentals. Curr.
1211
Biol. 27, 3025–3033 (2017).
1212
132. Zhang, G. et al. Comparative genomics reveals insights into avian genome evolution
1213
and adaptation. Science 346, 1311–1320 (2014).
1214
133. Bolger, A. M., Lohse, M. & Usadel, B. Trimmomatic: A flexible trimmer for Illumina
1215
sequence data. Bioinformatics 30, 2114–2120 (2014).
1216
134. Luo, R. et al. SOAPdenovo2: An empirically improved memory-efficient short-read de
1217
novo assembler. Gigascience 1, 18 (2012).
1218
135. Abascal, F., Zardoya, R. & Telford, M. J. TranslatorX: Multiple alignment of
1219
nucleotide sequences guided by amino acid translations. Nucleic Acids Res. 38, W7–
1220
W13 (2010).
1221
136. Simion, P. et al. A software tool ‘CroCo’ detects pervasive cross-species
1222
42
contamination in next generation sequencing data. BMC Biol. 16, 28 (2018).
1223
137. Emms, D. M. & Kelly, S. OrthoFinder: Phylogenetic orthology inference for
1224
comparative genomics. Genome Biol. 20, 238 (2019).
1225
138. Di Franco, A., Poujol, R., Baurain, D. & Philippe, H. Evaluating the usefulness of
1226
alignment filtering methods to reduce the impact of errors on evolutionary inferences.
1227
BMC Evol. Biol. 19, 21 (2019).
1228
139. Capella-Gutierrez, S., Silla-Martinez, J. M. & Gabaldon, T. trimAl: A tool for
1229
automated alignment trimming in large-scale phylogenetic analyses. Bioinformatics
1230
25, 1972–1973 (2009).
1231
140. Yang, Z. & Nielsen, R. Estimating synonymous and nonsynonymous substitution rates
1232
under realistic evolutionary models. Mol. Biol. Evol. 17, 32–43 (2000).
1233
141. Zhang, J., Nielsen, R. & Yang, Z. Evaluation of an improved branch-site likelihood
1234
method for detecting positive selection at the molecular level. Mol. Biol. Evol. 22,
1235
2472–2479 (2005).
1236
142. Yang, Z. Likelihood ratio tests for detecting positive selection and application to
1237
primate lysozyme evolution. Mol. Biol. Evol. 15, 568–573 (1998).
1238
143. Yang, Z. PAML 4: Phylogenetic analysis by maximum likelihood. Mol. Biol. Evol. 24,
1239
1586–1591 (2007).
1240
144. Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate: A practical and
1241
powerful approach to multiple testing. J. R. Stat. Soc. 57, 289–300 (1995).
1242
145. Bauer, D. F. Constructing confidence sets using rank statistics. J. Am. Stat. Assoc. 67,
1243
687–690 (1972).
1244
146. Mallick, S., Gnerre, S., Muller, P. & Reich, D. The difficulty of avoiding false
1245
positives in genome scans for natural selection. Genome Res. 19, 922–933 (2009).
1246
147. Fletcher, W. & Yang, Z. The effect of insertions, deletions, and alignment errors on the
1247
branch-site test of positive selection. Mol. Biol. Evol. 27, 2257–2267 (2010).
1248
148. Jordan, G. & Goldman, N. The effects of alignment error and alignment filtering on
1249
the sitewise detection of positive selection. Mol. Biol. Evol. 29, 1125–1139 (2012).
1250
149. Duret, L. & Galtier, N. Biased gene conversion and the evolution of mammalian
1251
genomic landscapes. Annu. Rev. Genomics Hum. Genet. 10, 285–311 (2009).
1252
150. Galtier, N. & Duret, L. Adaptation or biased gene conversion? Extending the null
1253
hypothesis of molecular evolution. Trends Genet. 23, 273–277 (2007).
1254
151. Ratnakumar, A. et al. Detecting positive selection within genomes: The problem of
1255
biased gene conversion. Philos. Trans. R. Soc. B Biol. Sci. 365, 2571–2580 (2010).
1256
43
152. Guéguen, L. et al. Bio++: Efficient extensible libraries and tools for computational
1257
molecular evolution. Mol. Biol. Evol. 30, 1745–1750 (2013).
1258
153. Wickham, H. & Grolemund, G. R for Data Science: Import, Tidy, Transform,
1259
Visualize, and Model Data. (O’Reilly Media, Inc. Canada, 2016).
1260
154. Wilke, C. O. cowplot: Streamlined plot theme and plot annotations for ‘ggplot2’.
1261
CRAN Repos https://cran.r-project.org/web/packages/cowplot/in (2016).
1262
155. Gouy, M., Guindon, S. & Gascuel, O. SeaView version 4: A multiplatform graphical
1263
user interface for sequence alignment and phylogenetic tree building. Mol. Biol. Evol.
1264
27, 221–224 (2010).
1265
156. Redelings, B. Erasing errors due to alignment ambiguity when estimating positive
1266
selection. Mol. Biol. Evol. 31, 1979–1993 (2014).
1267
157. Mi, H., Muruganujan, A., Ebert, D., Huang, X. & Thomas, P. D. PANTHER version
1268
14: More genomes, a new PANTHER GO-slim and improvements in enrichment
1269
analysis tools. Nucleic Acids Res. 47, D419–D426 (2019).
1270
158. Huerta-Cepas, J. et al. Fast genome-wide functional annotation through orthology
1271
assignment by eggNOG-mapper. Mol. Biol. Evol. 34, 2115–2122 (2017).
1272
159. Huerta-Cepas, J. et al. eggNOG 5.0: A hierarchical, functionally and phylogenetically
1273
annotated orthology resource based on 5090 organisms and 2502 viruses. Nucleic
1274
Acids Res. 47, D309–D314 (2019).
1275
1276
44
Supplementary Figures
1277
Supplementary Figure 1. Phylogenetic relationships of 408 swallowtail butterfly species
1278
(Papilionidae). Left phylogeny is inferred with the maximum-likelihood approach
1279
implemented with IQ-TREE, and right phylogeny is inferred with the Bayesian approach
1280
implemented with MrBayes. Both phylogenies show similar relationships except for the
1281
placement of the genus Teinopalpus, found as sister to Papilionini + Troidini with IQ-TREE
1282
and sister to Meandrusa (Papilionini) with MrBayes. Node support is indicated by ultrafast
1283
bootstrap and posterior probabilities on the maximum-likelihood and Bayesian phylogenies,
1284
respectively, with values of 95% and 0.95 considered as indicative of strong node support.
1285
Supplementary Figure 2. Node support (ultrafast bootstrap) of the maximum-likelihood
1286
phylogeny. The histogram shows the distribution of node support for all Papilionidae, and
1287
indicates a high overall tree resolution with ~80% of nodes having ultrafast bootstrap values
1288
≥ 95%.
1289
Supplementary Figure 3. Bayesian estimates of divergence times for swallowtail butterflies.
1290
The first inference was performed with exponential priors on fossil calibrations, while the
1291
second inference was carried out with uniform priors. The analysis based on exponential
1292
priors estimated a crown age for the family at 55.4 Ma (95% CI: 47.8-71.0 Ma), while the
1293
analysis based on uniform priors estimated the origin at 67.2 Ma (95% CI: 47.8-112 Ma).
1294
Supplementary Figure 4. Estimation of ancestral host-plant preferences for the two
1295
molecular dated trees with the Dispersal-Extinction-Cladogenesis (DEC) model. The results
1296
show that the family Aristolochiaceae is recovered as the ancestral feeding habit of the
1297
Papilionidae. K = Cretaceous, Pl = Pliocene, P = Pleistocene.
1298
Supplementary Figure 5. Estimation of ancestral host-plant preferences with the maximum-
1299
likelihood model of Markov 1-parameter (Mk) and the Bayesian approach of BayesTraits.
1300
The results are represented by pie charts indicating the relative probability for each state
1301
inferred at a given node. The results consistently show that (1) the family Aristolochiaceae is
1302
recovered as the ancestral feeding habit of the Papilionidae, and (2) the host-plant shifts are
1303
recovered at the same nodes, except at the root of Papilionini and at the root of Iphiclides +
1304
Lamproptera (due to the fact the the Mk model can include only 10 states).
1305
Supplementary Figure 6. Estimation of ancestral host-plant preferences for the
1306
Aristolochiaceae feeders with the Dispersal-Extinction-Cladogenesis (DEC) model. The
1307
results show that the genus Aristolochia is the primary Aristolochiaceae host plant while
1308
being also recovered as the ancestral feeding habit of the Papilionidae.
1309
45
Supplementary Figure 7. Phylogenetic relationships within the Aristolochiaceae (perianth-
1310
bearing Piperales) for 247 species. The phylogeny is inferred with the Bayesian approach of
1311
MrBayes. Node support is indicated by posterior probabilities, with values > 0.95 considered
1312
as strong node support.
1313
Supplementary Figure 8. Bayesian estimates of divergence times for Aristolochiaceae. The
1314
first inference was performed with exponential priors on fossil calibrations and 150 Ma as
1315
maximum age. The second inference was performed with exponential priors on fossil
1316
calibrations and 221 Ma as maximum age. The third inference was performed with uniform
1317
priors on fossil calibrations and 150 Ma as maximum age. The fourth inference was
1318
performed with uniform priors on fossil calibrations and 221 Ma as maximum age. The origin
1319
of the genus Aristolochia is estimated at 55.5 Ma (95% CI: 39.2-72.8 Ma) in the first
1320
analysis, at 58.8 Ma (95% CI: 42.5-76.2 Ma) in the second analysis, at 60.7 Ma (95% CI:
1321
43.9-80.5 Ma) in the third analysis, and at 64.8 Ma (95% CI: 47.3-83.1 Ma) in the fourth
1322
analysis.
1323
Supplementary Figure 9. Median node ages and 95% credibility intervals (CI) for the two
1324
dating analyses of Papilionidae and the four dating analyses of Aristolochiaceae. The 95% CI
1325
overlap substantially between the two groups regardless of the dating analysis. J = Jurassic, Pl
1326
= Pliocene, P = Pleistocene.
1327
Supplementary Figure 10. Estimation of the historical biogeography for the two molecular
1328
dated trees of Papilionidae with the Dispersal-Extinction-Cladogenesis (DEC) model. For
1329
each tree, two DEC analyses were performed: one with time-stratified palaeogeographic
1330
constraints, and one without such constraints. The swallowtail butterflies originated in a
1331
northern region centred around the Bering land bridge. K = Cretaceous, Pl = Pliocene, P =
1332
Pleistocene.
1333
Supplementary Figure 11. Estimation of the historical biogeography for the four molecular
1334
dated trees of Aristolochiaceae with the Dispersal-Extinction-Cladogenesis (DEC) model. For
1335
each tree, two DEC analyses were performed: one with time-stratified palaeogeographic
1336
constraints, and one without such constraints. The genus Aristolochia originated in a northern
1337
region centred around the Bering land bridge. J = Jurassic, K = Cretaceous, Pl = Pliocene, P =
1338
Pleistocene.
1339
Supplementary Figure 12. Trait-dependent diversification of Papilionidae linked to their
1340
host plant. a, Bayesian inferences made with the full MuSSE model showed that speciation
1341
rates vary according to the host-plant trait. b, Boxplots showing the increase of
1342
diversification rates following host-plant shifts from the ancestral state (Aristolochiaceae).
1343
46
Only the species-poor swallowtail lineages feeding on Fabaceae, Zygophyllaceae and
1344
Magnoliaceae show decrease of diversification rates.
1345
Supplementary Figure 13. Time-dependent diversification of Papilionidae after shifting to
1346
new host plants. Diversification is inferred with the RPANDA models, and the best-fit model
1347
is plotted showing rates through time for each clade. A model with increasing diversification
1348
over time best fits the Aristolochiaceae feeders. A model with a slowdown of diversification
1349
through time explained the diversification of Annonaceae feeders, Lauraceae feeders, and
1350
Papaveraceae feeders. A model with constant rates through time best fits the diversification
1351
of Apiaceae feeders, Crassulaceae feeders, and Rutaceae feeders. K = Cretaceous, Pl =
1352
Pliocene, P = Pleistocene.
1353
Supplementary Figure 14. Bayesian analysis of clade-specific and time-dependent
1354
diversification of Papilionidae obtained with BAMM. a, Phylorate plot showing that global
1355
diversification rates increase through time in Papilionidae with no significant rate shifts
1356
detected by BAMM (the inset plot indicates the posterior probability for the estimated
1357
number of shifts). b, Rate-through-time plots for selected swallowtail lineages feeding on
1358
distinct host-plant families. The results also show an overall diversification increase through
1359
time for each group of swallowtails. P = Palaeocene, E = Eocene, O = Oligocene, M =
1360
Miocene.
1361
Supplementary Figure 15. Bayesian analysis of branch-specific and time-dependent
1362
diversification of Papilionidae obtained with RevBayes. The median rates of diversification
1363
are plotted along each branch of the phylogeny, which shows a global increase of
1364
diversification rates through time in Papilionidae. Contrary to BAMM, this approach detected
1365
shifts in diversification rates in particular within the genera Parnassius and Papilio that have
1366
both shifted to new host-plant families. P = Palaeocene, E = Eocene, O = Oligocene, M =
1367
Miocene.
1368
Supplementary Figure 16. Bayesian analysis of episodic diversification of Papilionidae
1369
obtained with CoMET. The four plots represent speciation, extinction, net diversification, and
1370
relative extinction rates through time for the whole family. The result indicates a global
1371
increase of diversification rates over time, notably starting ~40 Ma. P = Palaeocene, E =
1372
Eocene, O = Oligocene, M = Miocene.
1373
Supplementary Figure 17. Number of host plants consumed through time by Papilionidae.
1374
Using the estimation of ancestral host-plant preferences (Supplementary Fig. 4), we plotted
1375
the time at which a new host-plant family was colonised. This result shows that the
1376
swallowtail butterflies have a steady increase in the number of host families consumed over
1377
47
time. This ecological diversification can be paralleled with the global increase in
1378
diversification rates estimated by birth-death models (Supplementary Figs. 13-16). K =
1379
Cretaceous, Pl = Pliocene, P = Pleistocene.
1380
Supplementary Figure 18. Genus-level phylogenomic tree of Papilionidae showing the 14
1381
selected branches with host-plant shifts and the 14 selected branches without host-plant shifts
1382
(control branches). The selection of these branches is based on the estimation of ancestral
1383
state models using the species-level phylogenies and current host-plant preferences
1384
(Supplementary Figs. 4, 5).
1385
Supplementary Figure 19. Violin plots of the percentage of missing data (“N” or “-”) and
1386
proportion of GC at third codon position (GC3) in alignment were positive selection have
1387
been detected (“Yes”) and positive selection have not been detected (“No”). Panels a and b
1388
are dataset 1 with 520 genes, and panels c and d are dataset 2 with 1533 genes.
1389
Supplementary Figure 20. The percentage of missing data (“N” or “-”) per genes across
1390
species computed for dataset 1 and dataset 2.
1391
Supplementary Figure 21. The percentage of missing data (“N” or “-”) per branch for the
1392
branches with (“Yes”, n = 14) and without (“No”, n = 14) host-plant shift. For a given
1393
branch, the percentage of missing data is the mean value of the species of a clade for which
1394
the branch is the root.
1395
Supplementary Figure 22. Relationship between the percentage of missing data (“N” or “-”)
1396
and the number of positively selected genes per branch. For a given branch, the percentage of
1397
missing data is the mean value of the species of a clade for which the branch is the root.
1398
Supplementary Figure 23. The percentage of GC at third codon position (GC3) per gene
1399
across species computed for dataset 1 and dataset 2.
1400
Supplementary Figure 24. The percentage of GC at third codon position (GC3) per branch
1401
for the branches with (“Yes”, n = 14) and without (“No”, n = 14) host-plant shift. For a given
1402
branch, the percentage of GC3 is the mean value of the species of a clade for which the
1403
branch is the root.
1404
Supplementary Figure 25. Relationship between the percentage of GC at third codon
1405
position (GC3) and the number of positively selected genes per branch. For a given branch,
1406
the percentage of GC3 is the mean value of the species of a clade for which the branch is the
1407
root.
1408
Supplementary Table 1. Results from analyses of diversification rates performed with
1409
LASER. For clades shifting to new host plants, net diversification rates are estimated based
1410
on their crown age and extant species diversity using the method of moments. Net
1411
48
diversification rates for shifting clades are higher than the global rates of the family,
1412
suggesting that shifting to a new host plant confer higher rates of species diversification.
1413
Estimates of expected clade size based on the global diversification rates and crown age of
1414
shifting clades show that four clades diversified significantly faster than background
1415
diversification rates of non-shifting clades.
1416
Supplementary Table 2. Information on orthogroups of dataset 2 (1,533 genes). The
1417
columns 2 to 6 indicate whether the genes are under positive selection and along which
1418
branch (column ‘Branch ID’ see Supplementary Fig. 18 for the annotated tree with branch
1419
numbers). The column ‘Papilio xuthus seq ID’ is the GenBank accession number for the
1420
corresponding sequences in Papilio xuthus. The column ‘PANTHER family:subfamily
1421
accession’ is family and subfamily accessions, and the column ‘PANTHER family name’ list
1422
the names for gene families based on PANTHER classification (see http://pantherdb.org/ for
1423
more information). Finally, ‘HMM e-value score’ is the Hidden Markov model e-value score,
1424
as reported by HMMER (Eddy 2011) performed through the online PANTHER scoring tool
1425
ftp://ftp.pantherdb.org/hmm_scoring/current_release.
Following
PANTHER
1426
recommendation, we have not considered e-values above 10-11 as significant.
1427
1428
12
Figures
327
328
329
330
Fig. 1. Evolution of host-plant association through time shows strong host-plant
331
conservatism across swallowtail butterflies. Phylogenetic relationships of swallowtail
332
butterflies, with coloured branches mapping the evolution of host-plant association, as
333
inferred by a maximum-likelihood model (Supplementary Figs. 4, 6). Additional analyses
334
with two other maximum-likelihood and Bayesian models inferred the same host-plant
335
associations across the phylogeny (Supplementary Fig. 5). Lue. = Luehdorfiini, Zerynth. =
336
Zerynthiini, and T. = Teinopalpini.
337
338
13
339
340
Fig. 2. Synchronous temporal and geographic origin for swallowtails and birthworts.
341
Bayesian molecular divergence times with exponential priors estimate an early Eocene origin
342
(~55 Ma) for both swallowtails and Aristolochia (alternatively, analyses with uniform prior
343
estimated an origin around 67 Ma for swallowtails and 64 Ma for Aristolochia,
344
Supplementary Figs. 3, 8, 9). Biogeographical maximum-likelihood models infer an ancestral
345
area of origin comprising West Nearctic, East Palearctic and Central America for both
346
swallowtails and birthworts (Supplementary Figs. 10, 11). K = Cretaceous, P = Palaeocene, E
347
= Eocene, O = Oligocene, M = Miocene, Pl = Pliocene, and P = Pleistocene. Ma = million
348
years ago.
349
350
14
351
352
Fig. 3. Host-plant shifts lead to repeated bursts in diversification rates and a sustained
353
overall increase in diversification through time. a, Diversification tends to be higher for
354
clades shifting to new host plants, as estimated by trait-dependent diversification models.
355
Boxplots represent Bayesian estimates of net diversification rates for clades feeding on
356
particular host plants (see also Supplementary Fig. 12). b, A global increase in diversification
357
is recovered with birth-death models estimating time-dependent diversification (see also
358
Supplementary Figs. 14, 15). Taking into account rate heterogeneity by estimating host-plant
359
and clade-specific diversification indicates positive gains of net diversification after shifting
360
to new host plants (see also Supplementary Fig. 13). K = Cretaceous, Paleoc. = Palaeocene,
361
Oligoc. = Oligocene, Pl = Pliocene, P = Pleistocene, Ma = million years ago.
362
363
15
364
365
16
Fig. 4. Host-plant shifts promote higher molecular adaptations. a, Genus-level
366
phylogenomic tree displaying branches with and without host-plant shifts, on which genome-
367
wide analyses of molecular evolution are performed. b, Number of genes under positive
368
selection (dN/dS > 1) for swallowtail lineages shifting to new host-plant families (green) or
369
not (grey). c, Number of genes under positive selection for swallowtail lineages undergoing
370
climate shifts (orange) or not (grey). d, Number of genes under positive selection for
371
swallowtail lineages shifting to new host plants (green), shifting both host plant and climate
372
(blue) or not (grey). This demonstrates genome-wide signatures of adaptations in swallowtail
373
lineages shifting to new host-plant families. Genes under positive selection did not contain
374
over- or under-represented functional GO categories (Supplementary Table 2). n.s. = not
375
significant (P > 0.05), * = P ≤ 0.05, ** = P ≤ 0.01.
376
377
| 2020 | Genome-wide macroevolutionary signatures of key innovations in butterflies colonizing new host plants | 10.1101/2020.07.08.193086 | [
"Allio Rémi",
"Nabholz Benoit",
"Wanke Stefan",
"Chomicki Guillaume",
"Pérez-Escobar Oscar A.",
"Cotton Adam M.",
"Clamens Anne-Laure",
"Kergoat Gaël J.",
"Sperling Felix A.H.",
"Condamine Fabien L."
] | creative-commons |
Unifying single-cell annotations based on the Cell Ontology
Sheng Wang1,2,*, Angela Oliveira Pisco3,*,#, Aaron McGeever3, Maria Brbic4, Marinka Zitnik4,
Spyros Darmanis3, Jure Leskovec3,4, Jim Karkanias3, Russ B. Altman1,2,3, #
1Department of Bioengineering, Stanford University, Stanford, CA 94305, USA.
2Department of Genetics, Stanford University, Stanford, CA 94305, USA.
3Chan Zuckerberg Biohub, San Francisco, CA 94158, USA
4Department of Computer Science, Stanford University, Stanford, CA 94305, USA.
*These authors contributed equally to this work
#Email:angela.pisco@czbiohub.org; russ.altman@stanford.edu
Abstract
Single cell technologies have rapidly generated an unprecedented amount of data that
enables us to understand biological systems at single-cell resolution. However, joint
analysis of datasets generated by independent labs remains challenging due to a lack
of consistent terminology to describe cell types. Here, we present OnClass, an
algorithm and accompanying software for automatically classifying cells into cell types
part of the controlled vocabulary that forms the Cell Ontology. A key advantage of
OnClass is its capability to classify cells into cell types not present in the training data
because it uses the Cell Ontology graph to infer cell type relationships. Furthermore,
OnClass can be used to identify marker genes for all the cell ontology categories,
independently of whether the cells types are present or absent in the training data,
suggesting that OnClass can be used not only as an annotation tool for single cell
datasets but also as an algorithm to identify marker genes specific to each term of the
Cell Ontology, offering the possibility of refining the Cell Ontology using a data-centric
approach.
1
Introduction
Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool to generate
comprehensive organismal atlases encompassing a wide range of organs and tissues1–10. One
of the most important tasks in single-cell analysis is cell type annotation because all
downstream analysis heavily rely on such information. This process that aims at characterizing
and labeling groups of cells according to their gene expression is currently very inefficient due to
the intense need for manual curation by a panel of tissue experts for each tissue and organ11–17.
Recent
efforts
in
scRNA-seq
have produced an unprecedented large compendium of
expert-curated cell type annotations, paving the way for scientists to better understand cellular
diversity3,18. However, utilizing these cell type annotations is challenging due to the inconsistent
terminology used to describe cell types collected by independent groups. This inconsistency will
likely increase as more groups generate new datasets and more cell types and states are
characterized, thus substantially preventing reproducible annotations and joint analysis of
multiple datasets.
The Cell Ontology offers a controlled vocabulary for cell types and has been proposed as the
basis for consistently annotating large-scale single-cell atlases19–23. A natural approach to
addressing the inconsistent vocabulary challenge is then to build computational methods that
automatically assign cells from different datasets to categories in the Cell Ontology. Ideally,
these methods should be fully automated such that the results can be quickly updated as the
Cell Ontology evolves.
However, assigning cells to terms (i.e., cell types) in the Cell Ontology has at least three
challenges. First, although the Cell Ontology contains valuable hierarchical relationships among
cell types, most of these cell type terms are not associated with marker genes which are crucial
for cell type annotation. Second, even though supervised learning approaches can be used to
predict Cell Ontology terms that have curated annotations, they are unable to classify cells to
unseen terms (i.e., terms that do not have any annotated cells in the training data). This issue
largely prevents us from fully understanding cellular diversity as more than 95% of cell types in
the Cell Ontology are unseen even in the largest datasets . Throughout this paper, we refer to
“unseen Cell Ontology terms” to describe cell types from the Cell Ontology that do not have
any annotated cells in the training data. In contrast, we use “seen Cell Ontology terms” to
denote cell types with some annotated cells in the training data. Third, as the Cell Ontology is
not developed specifically for scRNA-seq, it likely misses new cell types and cell states and so
certain cell type relationships might be inaccurate. Collectively, these challenges hinder
progress towards comprehensive cell type annotation and cellular diversity understanding.
We
developed
Ontology-based
Single
Cell
Classification
(OnClass)
to
address
these
challenges. OnClass is able to automatically classify cells to any cell type as long as its
corresponding term is captured in the Cell Ontology, even if this cell type does not have
2
annotated cells in the training data. To achieve this, OnClass first infers similarities among all
cell types according to their distances in the Cell Ontology graph. It then leverages these cell
type similarities to classify cells into unseen Cell Ontology terms based on the annotated cells of
other seen Cell Ontology terms. OnClass can thus classify cells into any Cell Ontology term and
consider even the hardest case when a term has no cell annotations in the training data.
OnClass is the first method that can classify cells into a specific cell type (rather than into a
generic unassigned category as previous work did11,12) even when the training set does not have
any annotated cells for such cell type. Furthermore, by projecting single cell transcriptomes and
the Cell Ontology into the same low-dimensional space, OnClass advances other important
applications, such as marker genes identification.
We evaluated OnClass on the Tabula Muris Senis dataset18, representing the existing largest
effort of cell type characterization. We found that our method outperformed existing methods at
annotating both seen and unseen Cell Ontology terms. We further demonstrated the ability of
OnClass to transfer annotations to 26 other single-cell datasets and assign cells to the correct
cell type even for cell types that were not part of the training data. Finally, we showed OnClass
was able to accurately identify marker genes for seen Cell Ontology terms as well as unseen
Cell Ontology terms. These OnClass referred marker genes achieved comparable performance
to curated marker genes on cell type annotation, paving the way for creating an organism-wide
molecular representation of cellular diversity.
Results
Overview of OnClass
The Cell Ontology is a controlled vocabulary that organized 2331 cell types anatomically derived
into a hierarchy based on the “is_a” relation. In OnClass, we first constructed a graph of cell
types based on the hierarchical “is_a” relationship in the Cell Ontology and embedded these cell
types into a low-dimensional space where similar cell types were close to each other24,25
(Supplementary Note 1, Supplementary Fig. 1). Single cell transcriptomes were then
projected into this low-dimensional space by finding a nonlinear transformation that projected
each annotated cell to the region of its cell type. Unannotated cells were also projected into this
low-dimensional space using the same nonlinear transformation and annotated to the cell type
corresponds to the region in which it lies. Importantly, such a procedure enables us to classify
cells to unseen Cell Ontology terms based on their regions in the low-dimensional space. In
addition to cell type annotation, OnClass used this low-dimensional space for other applications,
including marker genes identification (Fig. 1a). OnClass is Python-based open source package
available at https://github.com/wangshenguiuc/OnClass. Our implementation can take as input a
wide range of formats of the input gene expression matrix. It is able to consider any cell type
similarity between the hierarchical structure of the Cell ontology used in this paper. Moreover,
3
we provide a pre-trained model that is trained on TMS and can predict cell types for millions of
cells in a few minutes on a modern laptop.
Cell type embeddings reflect cell type similarity
Since OnClass annotated cells even to previously unseen Cell Ontology terms according to the
annotated cells from other Cell Ontology terms, its performance greatly relied on the quality of
cell type embeddings. High-quality cell type embeddings should place cell types with similar
gene expression profiles closely in the low-dimensional space, and can thus be used as good
features for classification. Therefore, we first verified the merit of our approach by comparing
three
types
of
cell
type
similarities:
the
Cell
Ontology
structure-based
similarity,
the
embedding-based similarity, and the gene expression-based similarity (Methods). We first
observed that the embedding-based similarity was strongly correlated with the Cell Ontology
structure-based similarity (Fig. 1b). For example, the average embedding-based similarity of
direct neighbors in the Cell Ontology graph was 0.86, which was 42% and 183% higher than the
average embedding-based similarity of two-hop neighbors and three-hop neighbors. For cell
types that are more than four-hop away in the Cell Ontology, the average embedding-based
similarity was less than 0.01. Next, we examined whether cell types with similar embeddings
would have similar gene expression profiles by comparing the embedding-based similarity and
the gene expression-based similarity. Using a collection of annotated cells as the benchmark,
we observed strong correlations between these two types of similarities. For instance, the
correlation between the gene expression-based similarity and the embedding-based similarity
was 0.70 (p-value < 1e-10) in pancreas and 0.77 (p-value < 1e-11) in kidney (Fig. 1c,d). The
strong correlation between these two types of similarities demonstrated the high-quality of cell
type embeddings and further suggested the possibility to annotate unseen Cell Ontology terms
by using the knowledge from other similar and seen Cell Ontology terms. Unfortunately, none of
the existing cell type annotation methods integrates with the Cell Ontology. OnClass’s ability to
annotate cells with any cell type in the Cell Ontology led us to consider whether we could
improve cell type annotation on large and diverse collections of scRNA-seq datasets.
Improved cell type annotation using Cell Ontology
We ran OnClass on the Tabula Muris Senis (TMS) dataset18. To investigate the effect of unseen
Cell Ontology terms, we split cells into test and training across different proportions of seen Cell
Ontology terms in the test set. Overall, we observed that OnClass led to a substantial
improvement in comparison to existing approaches (Fig. 2a-d). We first examined the ability of
OnClass to identify cells belonging to a given Cell Ontology term. We observed that OnClass
significantly outperformed all existing approaches in terms of AUROC on all proportions of seen
Cell Ontology terms (Fig. 2a). Even when only half of Cell Ontology terms were observed in the
training data, OnClass still achieved an AUROC of 0.87, while AUROCs of existing methods
were all below 0.72. Next, we investigated whether OnClass could accurately predict the Cell
Ontology term for a given cell. In a simpler setting where we combined all unseen Cell Ontology
4
terms as a generic “unseen” class, OnClass outperformed existing methods in terms of Cohen’s
Kappa statistic (i.e., balanced accuracy) from 10% to 90% of seen Cell Ontology terms (Fig.
2b). We found that the improvement of OnClass was more prominent with the increasing
proportion of unseen Cell Ontology terms. We next evaluated a more challenging setting where
unseen Cell Ontology terms were no longer combined and a prediction was deemed as correct
only if the cell was assigned to the specific correct term, even if it is an unseen Cell Ontology
term. By using Accuracy@3 and Accuracy@5 to quantify the performance, we observed
significant improvement of OnClass in comparison to existing methods (Fig. 2c,d). For example,
when 30% of Cell Ontology terms were unseen in the training data, OnClass obtained 0.45
Accuracy@3 and 0.55 Accuracy@5, while none of the existing approaches obtained accuracy
greater than 0.3. Again, the improvement of OnClass was larger with more unseen Cell
Ontology terms, indicating the advantage of using the Cell Ontology to transfer annotations from
seen
Cell Ontology terms to unseen Cell Ontology terms. To further demonstrate the
importance of using the Cell Ontology, we found that the performance of OnClass substantially
decreased
by
adding
random
noise
to
nodes
(Supplementary
Fig.
2a)
and
edges
(Supplementary Fig. 2b) in the Cell Ontology. Notably, even though TMS had one of the most
diverse and largest numbers of cell types, it still only covered less than 5% of all cell types in the
Cell Ontology. Therefore, we anticipate that OnClass will be even more useful as more single
cell RNA-seq datasets become available that contain transcriptomes for unobserved cell types
in TMS.
Annotating unseen Cell Ontology terms
We then examined the performance of OnClass in the more challenging case of annotating
unseen Cell Ontology terms, which cannot be accomplished by any existing methods. Although
recent efforts were able to classify cells into a generic “unknown” type11,12, they could neither
break down this new type into detailed cell types nor attach it to a specific cell type term. To
enable better comparison between OnClass and these approaches, we decided to extend
existing approaches by classifying “unknown” type cells to the nearest cell type in the Cell
Ontology (Methods). We studied the performance of OnClass by using an increasing number of
seen Cell Ontology terms as the training data and all cells in the test data belonged to the
remaining unseen Cell Ontology terms. We observed significant improvement with OnClass
across different proportions of seen Cell Ontology terms. For instance, when using 60% of seen
Cell Ontology terms as the training data (Fig. 3a), OnClass obtained an AUROC of 0.73. Even
when only using 20% of cell types as the training data, OnClass still obtained an AUROC of
0.68. On a randomly selected set of 9 new unseen terms, OnClass was able to accurately
classify 81% of cells (Fig. 3b-d). On a larger set of 21 unseen terms, OnClass still accurately
classified 58% of cells (Fig. 3h-j). We showed the comparison of OnClass annotation and
ground truth annotation in Fig. 3b-j. We found that OnClass was able to accurately classify a
majority of cell types, including rare cell types. For those cells that were not accurately
annotated, we found that the term assigned by OnClass was indeed biologically related to the
ground truth Cell Ontology term. When evaluating OnClass for each tissue separately, we also
5
observed good AUROCs from 0.84 to 0.93 on 21 tissues, with an average AUROC of 0.87
(Supplementary Figs. 3-22). As expert annotation can be imperfect and mostly limited to
familiar cell types, OnClass can correct these false positives and broaden expert knowledge.
We next examined the robustness and applicability of OnClass by using it to annotate diverse
datasets across animals, technologies, and organs. In particular, we used all cells in TMS to
train OnClass and then classified 105,476 cells collected from 26 single-cell datasets
(26-datasets) representing 9 technologies and 11 studies (Methods). We observed an average
AUROC of 0.75 for these 26 datasets. Among all 10 cell types, OnClass obtained an AUROC
greater than 0.8 for 5 of them (Fig. 4a). For B cell and macrophage that have annotated cells in
TMS, OnClass obtained AUROCs of 0.99 and 0.97, respectively (Fig. 4b, c). More importantly,
for cell types with no annotated cells in TMS, OnClass still achieved relatively high AUROCs
(0.85 for CD14+ monocytes cell, 0.85 for CD56+ natural killer cell, and 0.81 for regulatory T cell),
indicating its ability to accurately annotate and discover new cell types (Fig. 4d-f). Furthermore,
the predicted cell type annotations can be used as features to cluster cells from different
datasets. We used the predicted cell type annotations to combine these 26 datasets following
the same procedure as previous work26. We observed good performance by using OnClass,
where cells were clustered based on cell types rather than artifacts related to platforms (Fig.
4g). We further quantified the performance using the silhouette coefficient27 and observed a
significant improvement in comparison to the state-of-the-art data integration approach
Scanorama26 (Fig. 4h), indicating OnClass’s robustness to annotating cells from different
batches and datasets.
Identifying marker genes for unseen Cell Ontology terms
Given the accurate annotation of both seen and unseen Cell Ontology terms, we were then
interested in using OnClass to identify marker genes for the Cell Ontology terms. Marker genes
are the key to expert curation but the existing knowledge is incomplete and limited extensively
studied cell types. Here, we used OnClass to identify marker genes for both seen and unseen
Cell Ontology terms in TMS (Fig. 5a). OnClass was able to identify the correct marker genes for
64% of seen Cell Ontology terms within the top 10 candidate genes in the predicted marker
gene list. More importantly, since OnClass did not require any annotated cells to identify marker
genes, it was able to find marker genes for unseen Cell Ontology terms as well. For example,
OnClass identified the correct marker genes for 39% of unseen Cell Ontology terms within the
top
10
candidate
genes.
We
incorporated
these
OnClass-referred
maker
genes
(Supplementary Table 1) and functions enriched with these marker genes (Supplementary
Table 2) into our provisional Cell Ontology, in the hope of facilitating future expert curation. This
data is easily accessible through our portal http://onclass.ds.czbiohub.org. Although these
marker genes are by no means a completely accurate representation of cell type features, they
are
the
first
attempt
at
creating a comprehensive knowledge base of marker genes
representative of the entire cellular diversity.
6
Finally, we sought to examine whether OnClass-referred marker genes could be used to
accurately annotate cells. We first used all FACS cells in TMS to identify marker genes and then
used these marker genes to annotate droplet cells in TMS. We found that the performance of
using OnClass-referred marker genes was substantially better than using curated marker genes
for Cell Ontology terms with more than 500 individual cells that were annotated in such
category. For example, OnClass-referred marker genes achieved 0.98 AUROC, whereas
curated marker genes achieved 0.90 AUROC for Cell Ontology terms with more than 500 and
less than 1500 cells (Fig. 5b). For rare cell types, the performance of OnClass-referred marker
genes was comparable to curated marker genes (Fig. 5b). Furthermore, for those Cell Ontology
terms that have no curated marker genes, OnClass-referred marker genes also achieved
accurate cell type annotation performance (Fig. 5c). We found that the performance of OnClass
depends on the number of annotated cells and so as more data becomes available, we
anticipate substantial improvement at the level of identifying robust and accurate marker genes.
To assess the robustness of these marker genes, we next used these TMS-derived marker
genes to classify 26-datasets. Among all the 10 cell types, 8 of them achieved AUROCs larger
than 0.7 and 4 of them achieved AUROCs larger than 0.8 (Fig. 5d-i). Even for unseen Cell
Ontology terms, OnClass still obtained a desirable performance (Fig. 5g-i). Notably, when
comparing the performance with a supervised classifier, we found that using marker genes
could achieve better results on several cell types (e.g., CD14+ monocyte cells) (Fig. 4a, Fig.
5a). Although supervised models are more expressive, they are also prone to overfitting. In
contrast, marker genes are not only interpretable but also more robust to noise, thus enabling
accurate annotation of new cells.
Discussion
Ever since the emergence of scRNA-Seq, cell type annotation is a key step in single-cell data
analyses. As more cell types are discovered and expected to be discovered, recent efforts have
focused on classifying cells into existing labels or a generic unseen cell type11,12. Despite
encouraging results based on these approaches, these methods fail to provide meaningful
information specific to the cell types that are not part of the training sets. In contrast, our method
takes an important step forward by mapping each cell to the Cell Ontology, leading to accurate
annotations of cells with unseen Cell Ontology terms, which cannot be achieved by any existing
methods. Conceptually and methodologically, this is substantially different from existing
methods in the sense that our method explicitly leverages hierarchical cell-to-cell relationships
to directly classify cells into any cell type within the Cell Ontology.
While our method leverages the Cell Ontology to classify unseen cell types, it is inspired by
recent progress in single cell dataset integration approaches26,28. In the state-of-the-art single
cell integration frameworks, datasets from different technologies are aligned in the same
low-dimensional space by using mutual nearest neighbors as anchors to connect them. Indeed,
our method can be considered to be aligning the Cell Ontology to the gene expression matrix by
using known annotations as anchors. The key novelty of our method comes from effectively
7
embedding cell types based on the hierarchical structure of the Cell Ontology and dividing the
low-dimensional space into regions to enable assignment of cells to unseen Cell Ontology
terms. We therefore expect our method to be of broad use to the community of cell biologists
and computational biologists who are dwelling with the hard problem of identifying the cell
populations present in each dataset.
8
Fig. 1. a, Flow chart of OnClass. The Cell Ontology is used to embed cell types into a
low-dimensional space. OnClass then partitions this low-dimensional space into multiple
regions, each corresponding to a cell type. Cells are then projected into this space by reducing
the dimensionality of the gene expression matrix. These boundaries can then be used to
annotate cell type, identify marker genes and integrate datasets. b, Violin plot showing the
correspondence between the location of each cell type’s nearest neighbor in the Cell Ontology
and the embedding similarity. The nearest neighbor of each cell type is calculated by using the
cosine distance between cell type embeddings. c,d Scatter plots showing the correlations
between the embedding-based cell type similarity and the gene expression-based cell type
similarity in pancreas (c) and kidney (d).
Fig. 2. a-d Bar plots comparing OnClass and existing methods in terms of AUROC (a), Cohen’s
Kappa (b), Accuracy@3 (c) and Accuracy@5 (d). x-axis shows the proportion of seen Cell
Ontology terms in the test data.
Fig. 3. a, Bar plot comparing OnClass and existing methods for different proportions of seen
Cell Ontology terms. x-axis shows the proportion of seen Cell Ontology terms in the training
data and y-axis shows the AUROC. b,c,e,f,h,i, 2-D UMAP showing the predicted Cell Ontology
terms of OnClass (b, e, h) and ground truth labels (c, f, i) for 9 unseen Cell Ontology terms (b,
c), 11 unseen Cell Ontology terms (e, f), and 21 unseen Cell Ontology terms (h, i). The same
color between OnClass predicted labels and ground truth labels means correct annotation.
d,g,j, Sankey diagrams of the resulting mapping between predicted labels (left) to ground truth
labels (right) for 9 unseen Cell Ontology terms (d), 11 unseen Cell Ontology terms (g), and 21
unseen Cell Ontology terms (j).
Fig. 4. a, Bar plot showing the AUROC of OnClass on 9 cell types, including 2 present in TMS
(green) and 7 not (yellow). b-f AUROC plots of OnClass’s prediction for five cell types: B cell (b),
macrophage (c), CD14+ monocyte cell (d), CD56+ NK cell (e) and regulatory T cell (f). g, 2-D
UMAP showing the 26 datasets and 6 cell types. h, Box plot showing the comparison between
OnClass and Scanorama in terms of silhouette coefficient.
Fig. 5. a, Plot showing the proportion of cell types out of the ones present (green) or not (yellow)
in TMS for which OnClass can identify the marker genes in the top k genes out of 23,437 genes.
k is shown in the x-axis and corresponds to the position in the marker gene list sorted by
p-value. b, Boxplot showing the cell type annotation performance of using OnClass-referred
marker genes (red) and curated marker genes (blue) in terms of AUROC. x-axis shows the
number of cells per Cell Ontology term. c, Boxplot showing the cell type annotation performance
of using OnClass-referred marker genes in terms of AUROC. Only Cell Ontology terms that
have no curated marker genes are shown here. x-axis shows the number of cells per Cell
Ontology term. d, Bar plot showing the AUROC of OnClass for 10 cell types, including 2 present
in TMS (green) and 8 not (yellow). e-i AUROC plots of OnClass’s prediction for five cell types:
macrophage (e), B cell (f), CD14+ monocyte cell (g), CD56+ NK cell (h), and regulatory T cell (i).
9
Supplementary Fig. 1. Flowchart of the cell type embedding process. The Cell Ontology graph
is constructed based on the “is_a” relation in the Cell Ontology. Random walk with restart is
performed on the graph, restarting from each node. An equilibrium distribution is calculated for
restarting from each node. These distributions are then concatenated and then reduced into a
low-dimensional space.
Supplementary Fig. 2a,b, Plot of adding random noise on nodes (a) and edges (b) into the Cell
Ontology graph. X-axis the ratio of random noise and y-axis is the AUROC.
Supplementary Figs. 3-23, Plot of the Cell Ontology of cell types that have annotated cells in
aorta, BAT, brain myeloid, brain non-myeloid, diaphragm, GAT, heart, kidney, large intestine,
limb muscle, liver, lung, mammary gland, MAT, pancreas, SCAT, skin, spleen, thymus, tongue,
and trachea. AUROC of each cell type is shown in rings.
10
Methods
scRNA-seq datasets
We used the compendium of single cell transcriptomic data from the Tabular Muris Senis18. Cell
type annotations in Tabula Muris Senis were curated by domain experts and all cell type
annotations present in the dataset were manually mapped to the Cell Ontology vocabulary. We
next obtained 26 scRNA-seq datasets from 11 different studies6–10,29–35. We used the processed
collection from Scanorama26, where low-quality cells were excluded. There were 5,216 genes
across all 26 datasets and a total of 105,476 cells, with each dataset containing between 90 and
18,018 cells. Since these datasets did not provide cell type annotations that were mapped to the
Cell Ontology vocabulary, we manually mapped cell types in these datasets to Cell Ontology
terms (Supplementary Table 3). After the mapping, there were 10 different Cell Ontology terms
in these 26 datasets. We denoted these datasets as “26-datasets” in this paper.
The Cell Ontology
We
downloaded
the
Cell
Ontology
from
The
OBO
Foundry
(http://www.obofoundry.org/ontology/cl.html)20. We used the “is_a” relation in the Cell Ontology
to construct an undirected graph of cell types. There were in total of 2331 nodes in the
constructed graph, corresponding to 2331 different cell types. All edges in this graph have the
same weight.
Embedding the Cell Ontology into the low-dimensional space
OnClass computed a compressed, low-dimensional representation of each cell type based on
the constructed cell type graph. We used clusDCA24,25, which had been proposed to embed the
Gene Ontology, to embed the Cell Ontology. clusDCA first computed a propagated cell type
graph by applying the random walk with restart36,37 to the cell type graph. It then obtained the
low-dimensional representation of each cell type by using the singular value decomposition
(SVD)38 to reduce the dimensionality of this propagated cell type graph. As suggested by
clusDCA, we set the dimensionality of SVD to 1000 and the restart probability of the random
walk with restart to 0.8. A detailed description of embedding cell types can be found in the
Supplement (Supplementary Fig. 1, Supplementary Note).
Cell type annotation
OnClass used a bilinear neural network model to predict the Cell Ontology term for a novel cell.
Let M be an m by n matrix of input gene expression data, where m was the number of cells and
n was the number of genes. Let Y be an m by c label matrix, where c was the total number of
11
Cell Ontology terms in the Cell Ontology. Yij=1 if cell i was annotated to Cell Ontology term j,
otherwise Yij=0. Note that c was often much larger than the number of seen Cell Ontology terms
in the training data, as the majority of Cell Ontology terms were unseen in the training data. The
corresponding columns of unseen Cell Ontology terms were all zeros in the label matrix. Let X
be a c by q matrix of the low-dimensional representations of cell types, where q was the
dimension of cell type embedding dimensionality. X was the output of clusDCA and fixed during
optimization. OnClass optimized the following cross-entropy loss:
,
Σ
Y log(exp(Relu(Relu(M W )W )X ) / Σ
exp(Relu(Relu(M W )W )X ))
L = Σm
i=1
c
j=1
ij
i
1
2
j
T
c
k=1
i
1
2
k
T
where
and
were the parameters that needed to be estimated.
W 1 ∈ R
n✖h
W 2 ∈ R
h✖q
elu
R
was the rectifier function for nonlinear transformation39. h was the number of hidden dimensions
and set to 500. We observed that the performance of OnClass was stable for h between 200
and 2000. OnClass used ADAM 40 to optimize this objective function.
After the optimization, the Cell Ontology term of a new cell with expression vector z could then
be predicted as:
,
xp(Relu(Relu(zW )W )X ) / Σ
exp(Relu(Relu(zW )W )X )
pj = e
1
2
j
T
c
k=1
1
2
k
T
where
was the probability that this cell belonged to Cell Ontology term j.
pj
p , p , ..,
}
P = { 1 2 .
pc
was the probability distribution that this cell belonged to each Cell Ontology term, including both
seen Cell Ontology terms and unseen Cell Ontology terms. As a result, OnClass could
automatically assign cells to any term in the Cell Ontology, even if it does not have any
annotated cells in the training data.
Cell type embeddings reflect cell type similarity
We calculated three types of cell type similarities: the Cell Ontology structure-based similarity,
the embedding-based similarity and the gene expression-based similarity. The Cell Ontology
structure-based similarity was calculated as the shortest distance between two cell types in the
Cell Ontology graph. The embedding-based similarity was the cosine similarity between
low-dimensional representations of two cell types. We used the gene expression of all FACS
cells in TMS to calculate the gene expression-based similarity. The calculation was performed
per organ. For each organ, we first identified two sets of cells belonging to two given cell types.
We then calculated the mean of pairwise cosine similarities between gene expression of these
two sets of cells and used it as the gene expression-based cell type similarity.
Evaluation of cell type annotation
We evaluated across different proportions of seen Cell Ontology terms in the test set ranging
from 100% to 10%, where 10% indicates that 10% of Cell Ontology terms in the test set have at
least some annotated cells in the training data. For a proportion k percentage, we first randomly
12
selected k percentage of Cell Ontology terms as seen Cell Ontology terms and remaining Cell
Ontology terms as unseen Cell Ontology terms. All cells belonging to these unseen Cell
Ontology terms were used as the test set. For the seen cell types, we random split their cells
into five equal size folds, where one-fold was used as the training set and the remaining
four-folds were used as the test set. We created a five-fold of test and training here according to
the initial annotation process in Tabula Muris Senis, where about 20% of cells (3-month mice)
were annotated first and then extended to the remaining 80%. The test data thus contained all
cells in each of the unseen Cell Ontology terms and 80% of cells in each seen Cell Ontology
terms. We performed cross-validation by repeating this procedure 5 times for each proportion.
To evaluate the case where all Cell Ontology terms in the test set are unseen (Fig. 3a), we
compared the performance across different proportions of seen Cell Ontology terms in the
training set. For a given proportion k percentage, we randomly selected k percentage of cell
types as seen Cell Ontology terms and the remaining as unseen Cell Ontology terms. All cells
belonging to the seen (unseen) Cell Ontology terms were used as the training (test) set. We
performed cross-validation by repeating this procedure 5 times for each proportion.
We evaluated our method and comparison approaches on four metrics, including the area under
the receiver operating characteristic curve (AUROC), Accuracy@3, Accuracy@5, and Cohen’s
kappa statistic41. As we were evaluating a large number of classes (i.e., more than 80 cell
types), it was important to address the bias from class imbalance during evaluation. Therefore,
we used the macro-average AUROC rather than the micro-average AUROC to summarize
results across different Cell Ontology terms. Macro-average AUROC calculates the areas under
the curves for each class independently and then takes the average. Cohen’s kappa statistic
can handle well both multi-class and imbalanced class problems and has been widely used as
an alternative to accuracy11. A large cohen’s kappa statistic indicates better performance, while
1 indicates perfect classification. Accuracy@3 (Accuracy@5) is a widely used ranking metric,
which assesses the correctness of the top 3(5) predicted Cell Ontology terms in comparison to
only examining the top 1 Cell Ontology term in Cohen’s kappa statistic. A prediction would be
deemed as correct if any of the top 3 (5 for Accuracy@5) predicted Cell Ontology terms is the
correct Cell Ontology term.
Comparison approaches
We compared our method with four existing methods ACTINN, singleCellNet (sCN), one-vs-rest
logistic regression (LR), and DOC. ACTINN and sCN are two of the best approaches in cell type
annotation according to a recent study15. ACTINN used a three-layer neural network to predict
the
cell
type13.
We
used
the
implementation
of
ACTINN
from
the
authors
(https://github.com/mafeiyang/ACTINN) and ran it on TMS. We used the default parameters for
ACTINN since these parameters were used in their paper to annotate cells in the Tabula Muris3,
an earlier version and subset of our dataset. sCN used gene pairs as features and random
forest as the classifier to predict the cell type11. Notably, sCN was able to classify cells into an
13
unknown
cell
type.
We
obtained
the
implementation
of
singleCellNet
from
(https://github.com/pcahan1/singleCellNet). We found that the implementation of sCN was not
scaled to large datasets like TMS and it was not able to cross-validate rare cell types with less
than 50 cells. We reimplemented part of sCN to enable its annotation for rare cell types and
made the code available as part of our package. To make it scalable to TMS, we ran it on the
dimensionality reduced gene expression matrix instead of the original gene expression matrix.
LR was the standard machine learning classifier for multi-class classification on large-scale
datasets. We used the one-vs-rest logistic regression instead of the multinomial logistic
regression in order to obtain a probability cutoff of 0.5 to determine the unknown cell type. DOC
was an advanced machine learning method for classifying unseen text documents, which was a
natural solution to our problem and could be directly applied here42. The key idea of DOC was to
find a data-driven probability cutoff for the unknown class rather than using a fixed probability
cutoff of 0.5 as LR did. However, DOC was also not able to classify cells into the specific cell
type. As the original DOC codebase was developed for word sequences classification and could
not directly take gene expression as input, we reimplemented and replaced its underlying
convolutional neural network classifier with a multinomial logistic regression.
Although sCN, DOC and LR were able to classify cells into a “unknown” cell type, they were not
able to classify these cells into the specific cell type. To enable a fair comparison, we further
proposed to extend these three approaches by classifying cells belong to the unknown cell type
to a specific cell type. In particular, when a cell was annotated as the unknown cell type, we first
found the seen cell type that had the largest confidence score for this cell. We then annotated
the cell to the nearest unseen cell type of this seen cell type based on the Cell Ontology graph.
We denoted these extended approaches as sCN (extended), LR (extended), and DOC
(extended) for sCN, LR, and DOC, respectively.
Transfer annotations to 26-datasets
To transfer annotations from TMS to 26-datasets, we first used Scanorama to correct batch
effects among TMS and 26 datasets. Scanorama took the gene expression matrix of these 27
datasets as input, it then calculated the corrected gene expression of these 27 datasets. We
then ran OnClass on all cells in TMS and predicted the Cell Ontology term for each cell in the
26-datasets. To combine these 26-datasets, we used the output probability distribution of each
cell by OnClass as the feature for each cell. We visualized these cells by projecting these
features using UMAP43. We used silhouette coefficients to evaluate the clustering accuracy for
both our method and Scanorama27.
Marker genes identification
We used differential gene expression analysis to identify marker genes for each Cell Ontology
term. In particular, we first ran OnClass on all FACS cells in TMS and then predicted the
probability of these cells belonging to each Cell Ontology term in the Cell Ontology. For each
14
Cell Ontology term, we took the 50 cells with the highest probability as the positively annotated
group and the 50 cells with the lowest probability as the negatively annotated group. We then
used the t-test to test whether an individual gene was significantly overexpressed in the
positively annotated group then the negatively annotated group. We performed this one-sided
independent t-test for each gene and then ranked genes according to the resulted P-values.
This rank list was the predicted marker gene list. Curated marker genes of 69 Cell Ontology
terms were collected from literature by experts (Supplementary Table 4). 28 cell types in TMS
are in these 69 Cell Ontology terms and thus had curated marker genes. To classify a new cell
according to marker genes, we used the sum of the expression of marker genes of each Cell
Ontology term as the predicted score for that Cell Ontology term. A larger score indicated that
the cell more likely belonged to this Cell Ontology term.
Statistical analysis
We used the scipy.stats44 Python package implementation of the one-sided independent t-test,
Pearson correlation statistics, Spearman correlation statistics, and associated P-values used in
this study. We used the scikit-learn Python package implementation of one-vs-rest logistic
regression, silhouette coefficients, AUROC, and cohen’s kappa statistics used in this study45.
Data availability
All datasets used in this study are available at https://figshare.com/projects/OnClass/70637,
including gene expression data, pre-trained model, cell type embeddings, and the Cell
Ontology.
A
detailed
description
of
these
datasets
can
be
found
at
https://onclass.readthedocs.io/.
Code availability
OnClass codes are available at https://github.com/wangshenguiuc/OnClass. The OnClass
server can be found at http://onclass.ds.czbiohub.org/.
Competing interests
R.B.A. declares the following competing interests: stock or other ownership (Personalis,
23andme, Youscript); consulting or advisory role (United Health, Second Genome, Karius, UK
Biobank, Swiss Personalized Health Network).
Acknowledgments
The authors would like to thank the developers and maintainers of the Cell Ontology for
insightful discussions. This work is supported by the Chan-Zuckerberg Biohub, NIH GM102365,
LM005652, and TR002515.
15
16
References
1.
Klein, A. M. et al. Droplet barcoding for single-cell transcriptomics applied to
embryonic stem cells. Cell 161, 1187–1201 (2015).
2.
Guo, G. et al. Resolution of cell fate decisions revealed by single-cell gene
expression analysis from zygote to blastocyst. Dev. Cell 18, 675–685 (2010).
3.
Tabula Muris Consortium et al. Single-cell transcriptomics of 20 mouse organs
creates a Tabula Muris. Nature 562, 367–372 (2018).
4.
Han, X. et al. Mapping the Mouse Cell Atlas by Microwell-Seq. Cell 173, 1307
(2018).
5.
Tang, F. et al. mRNA-Seq whole-transcriptome analysis of a single cell. Nat.
Methods 6, 377–382 (2009).
6.
Grün, D. et al. De Novo Prediction of Stem Cell Identity using Single-Cell
Transcriptome Data. Cell Stem Cell 19, 266–277 (2016).
7.
Muraro, M. J. et al. A Single-Cell Transcriptome Atlas of the Human Pancreas. Cell
Syst 3, 385–394.e3 (2016).
8.
Baron, M. et al. A Single-Cell Transcriptomic Map of the Human and Mouse
Pancreas Reveals Inter- and Intra-cell Population Structure. Cell Systems vol. 3
346–360.e4 (2016).
9.
Zheng, G. X. Y. et al. Massively parallel digital transcriptional profiling of single
cells. Nat. Commun. 8, 14049 (2017).
10. Davie, K. et al. A Single-Cell Transcriptome Atlas of the Aging Drosophila Brain.
17
Cell 174, 982–998.e20 (2018).
11. Tan, Y. & Cahan, P. SingleCellNet: A Computational Tool to Classify Single Cell
RNA-Seq Data Across Platforms and Across Species. Cell Syst 9, 207–213.e2
(2019).
12. Pliner, H. A., Shendure, J. & Trapnell, C. Supervised classification enables rapid
annotation of cell atlases. Nat. Methods 16, 983–986 (2019).
13. Ma, F. & Pellegrini, M. ACTINN: Automated Identification of Cell Types in Single
Cell RNA Sequencing. Bioinformatics (2019) doi:10.1093/bioinformatics/btz592.
14. Hou, R., Denisenko, E. & Forrest, A. R. R. scMatch: a single-cell gene expression
profile annotation tool using reference datasets. Bioinformatics (2019)
doi:10.1093/bioinformatics/btz292.
15. Abdelaal, T. et al. A comparison of automatic cell identification methods for
single-cell RNA sequencing data. Genome Biol. 20, 194 (2019).
16. Lopez, R., Regier, J., Cole, M. B., Jordan, M. I. & Yosef, N. Deep generative
modeling for single-cell transcriptomics. Nat. Methods 15, 1053–1058 (2018).
17. Zhang, A. W. et al. Probabilistic cell-type assignment of single-cell RNA-seq for
tumor microenvironment profiling. Nat. Methods 16, 1007–1015 (2019).
18. The Tabula Muris consortium et al. A Single Cell Transcriptomic Atlas
Characterizes Aging Tissues in the Mouse. bioRxiv 661728 (2019)
doi:10.1101/661728.
19. Bard, J., Rhee, S. Y. & Ashburner, M. An ontology for cell types. Genome Biol. 6,
R21 (2005).
18
20. Smith, B. et al. The OBO Foundry: coordinated evolution of ontologies to support
biomedical data integration. Nat. Biotechnol. 25, 1251–1255 (2007).
21. Diehl, A. D. et al. The Cell Ontology 2016: enhanced content, modularization, and
ontology interoperability. J. Biomed. Semantics 7, 44 (2016).
22. Malladi, V. S. et al. Ontology application and use at the ENCODE DCC. Database
2015, (2015).
23. Lizio, M. et al. Gateways to the FANTOM5 promoter level mammalian expression
atlas. Genome Biol. 16, 22 (2015).
24. Wang, S., Cho, H., Zhai, C., Berger, B. & Peng, J. Exploiting ontology graph for
predicting sparsely annotated gene function. Bioinformatics 31, i357–64 (2015).
25. Cho, H., Berger, B. & Peng, J. Compact Integration of Multi-Network Topology for
Functional Analysis of Genes. Cell Syst 3, 540–548.e5 (2016).
26. Hie, B., Bryson, B. & Berger, B. Efficient integration of heterogeneous single-cell
transcriptomes using Scanorama. Nat. Biotechnol. 37, 685–691 (2019).
27. Rousseeuw, P. J. Silhouettes: A graphical aid to the interpretation and validation of
cluster analysis. J. Comput. Appl. Math. 20, 53–65 (1987).
28. Stuart, T. et al. Comprehensive Integration of Single-Cell Data. Cell 177,
1888–1902.e21 (2019).
29. Segerstolpe, Å. et al. Single-Cell Transcriptome Profiling of Human Pancreatic
Islets in Health and Type 2 Diabetes. Cell Metab. 24, 593–607 (2016).
30. Lawlor, N. et al. Single-cell transcriptomes identify human islet cell signatures and
reveal cell-type–specific expression changes in type 2 diabetes. Genome Res. 27,
19
208–222 (2017).
31. Nestorowa, S. et al. A single-cell resolution map of mouse hematopoietic stem and
progenitor cell differentiation. Blood 128, e20–31 (2016).
32. Paul, F. et al. Transcriptional Heterogeneity and Lineage Commitment in Myeloid
Progenitors. Cell 164, 325 (2016).
33. Shalek, A. K. et al. Single-cell RNA-seq reveals dynamic paracrine control of
cellular variation. Nature 510, 363–369 (2014).
34. Gierahn, T. M. et al. Seq-Well: portable, low-cost RNA sequencing of single cells at
high throughput. Nat. Methods 14, 395–398 (2017).
35. Kang, H. M. et al. Multiplexed droplet single-cell RNA-sequencing using natural
genetic variation. Nat. Biotechnol. 36, 89–94 (2018).
36. Pearson, K. The Problem of the Random Walk. Nature 72, 342–342 (1905).
37. Wang, S. et al. Typing tumors using pathways selected by somatic evolution. Nat.
Commun. 9, 4159 (2018).
38. Halko, N., Martinsson, P. & Tropp, J. Finding Structure with Randomness:
Probabilistic Algorithms for Constructing Approximate Matrix Decompositions.
SIAM Rev. 53, 217–288 (2011).
39. Hahnloser, R. H. R. & Seung, H. S. Permitted and Forbidden Sets in Symmetric
Threshold-Linear Networks. in Advances in Neural Information Processing Systems
13 (eds. Leen, T. K., Dietterich, T. G. & Tresp, V.) 217–223 (MIT Press, 2001).
40. Kingma, D. P. & Ba, J. Adam: A Method for Stochastic Optimization. arXiv [cs.LG]
(2014).
20
41. McHugh, M. L. Interrater reliability: the kappa statistic. Biochem. Med. 22, 276–282
(2012).
42. Shu, L., Xu, H. & Liu, B. DOC: Deep Open Classification of Text Documents. arXiv
[cs.CL] (2017).
43. Becht, E. et al. Dimensionality reduction for visualizing single-cell data using
UMAP. Nat. Biotechnol. (2018) doi:10.1038/nbt.4314.
44. Jones, E., Oliphant, T., Peterson, P. & Others. SciPy: Open source scientific tools
for Python. (2001).
45. Kramer, O. Scikit-Learn. in Machine Learning for Evolution Strategies (ed. Kramer,
O.) 45–53 (Springer International Publishing, 2016).
21
0.6
0.7
0.8
0.9
1.0
Gene expression similarity
−0.5
−0.3
−0.1
0.1
0.3
0.5
0.7
0.9
Low-dimensional representation
similarity
Pancreas
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Gene expression similarity
−0.4
−0.2
0.0
0.2
0.4
0.6
0.8
Low-dimensional representation
similarity
Kidney
“is-a”
relation
Cell type annotation
a
b
c
Cell Ontology
Step 1
Embed the Cell Ontology
Unannotated cells
Cell type embedding
General terms
Specific terms
Step 2
Partition
low-dimensional
space
Annotated cells
Gene expression of single cells
Step 3
Project single cells
Marker genes identifcation
1) CD4+
2) CD25+
...
Data integration
d
1
2
3
4
>4
Shortest distance in the Cell Ontology
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
Embedding similarity
Unseen Cell Ontology term
a
b
c
d
100% 90%
80%
70%
60%
50%
40%
30%
20%
10%
Proportion of seen Cell Ontology terms in the test set
0.0
0.2
0.4
0.6
0.8
1.0
Accuracy@3
100% 90%
80%
70%
60%
50%
40%
30%
20%
10%
Proportion of seen Cell Ontology terms in the test set
0.0
0.2
0.4
0.6
0.8
1.0
Accuracy@5
100% 90%
80%
70%
60%
50%
40%
30%
20%
10%
Proportion of seen Cell Ontology terms in the test set
0.0
0.2
0.4
0.6
0.8
1.0
Cohen’s kappa
100% 90%
80%
70%
60%
50%
40%
30%
20%
10%
Proportion of seen Cell Ontology terms in the test set
0.5
0.6
0.7
0.8
0.9
1.0
AUROC
OnClass
ACTINN
sCN
LR
d
e
f
g
h
j
i
90%
80%
70%
60%
50%
40%
30%
20%
10%
Proportion of seen Cell Ontology terms in the training set
0.4
0.5
0.6
0.7
0.8
0.9
AUROC
OnClass
sCN (extended)
DOC (extended)
LR (extended)
a
Aortic endothelial
Basal cells
Brush cell of epithelium proper of large intestine
CD8+ alpha-beta T
Ciliated columnar cell of tracheobronchial tree
Club cells
DN4 thymocyte
Epithelial cell
of large intestine
Epithelial cells
Fibroblast
Fibroblast of lung
Glial cells
Kidney collecting duct epithelial
Leukocyte
Lung endothelial
Mesenchymal stem cells
Monocyte
Pancreatic PP cells
Proerythroblast
Regular ventricular
cardiac myocyte
Respiratory basal cells
UMAP 1
Ground truth
9 unseen terms
UMAP 2
UMAP 1
OnClass
9 unseen terms
UMAP 2
UMAP 1
OnClass
11 unseen terms
UMAP 2
UMAP 1
Ground truth
11 unseen terms
UMAP 2
UMAP 1
OnClass
21 unseen terms
UMAP 2
UMAP 1
Ground truth
21 unseen terms
UMAP 2
UMAP 1
b
c
OnClass
Scanorama
−0.4
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
Silhouette coefficient
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
True Positive Rate
B cell
AUROC = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
True Positive Rate
Macrophage
AUROC = 0.97
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
True Positive Rate
CD14+ monocyte
AUROC = 0.85
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
CD56+ NK
AUROC = 0.85
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
True Positive Rate
Regulatory T
AUROC = 0.81
B cell
Macrophage
CD14+ monocyte
CD56+ NK
Regulatory T
Memory T
PBMC
Cytotoxic T
CD4+
helper T
0.5
0.6
0.7
0.8
0.9
1.0
AUROC
Cell types in TMS
Cell types not in TMS
a
b
c
d
e
f
g
h
UMAP 1
UMAP 2
True Positive Rate
HSCs
Jurkat + 293T
Macrophages
Neurons
PBMCs
Pancreatic islets
HSCs
Jurkat + 293T
Macrophages
Neurons
PBMCs
Pancreatic islets
Macrophage
B cell
CD14+ monocyte
CD56+ NK
Regulatory T
Memory T
PBMC
HSC
CD4+ helper T
Cytotoxic T
0.5
0.6
0.7
0.8
0.9
1.0
AUROC
Cell types in TMS
Cell types not in TMS
0
250
500
750
1000
Position in the marker gene list
0%
20%
40%
60%
80%
Proportion of cell types
Cell types in TMS (n=28)
Cell types not in TMS (n=41)
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
True Positive Rate
B cell
AUROC = 0.88
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
True Positive Rate
CD14+ monocyte
AUROC = 0.96
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
True Positive Rate
CD56+ NK
AUROC = 0.89
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
True Positive Rate
Regulatory T
AUROC = 0.79
<500
500-1500
>1500
Number of cells per term
0.6
0.7
0.8
0.9
1.0
AUROC
Cell Ontology terms without
curated marker genes
0.0
0.2
0.4
0.6
0.8
1.0
False Positive Rate
0.0
0.2
0.4
0.6
0.8
1.0
Macrophage
AUROC = 0.93
True Positive Rate
a
d
f
h
i
g
<500
500-1500 >1500
Number of cells per term
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
AUROC
OnClass-referred marker genes
Curated marker genes
Cell Ontology terms with
curated marker genes
b
e
c
| 2020 | Unifying single-cell annotations based on the Cell Ontology | 10.1101/810234 | [
"Wang Sheng",
"Pisco Angela Oliveira",
"McGeever Aaron",
"Brbic Maria",
"Zitnik Marinka",
"Darmanis Spyros",
"Leskovec Jure",
"Karkanias Jim",
"Altman Russ B."
] | creative-commons |
Increased Ca2+ signaling through CaV1.2 induces tendon hypertrophy with increased collagen
fibrillogenesis and biomechanical properties
Haiyin Li1, 2, Antonion Korcari1, 3, David Ciufo1, 2, Christopher L. Mendias4, Scott A. Rodeo5,
Mark R. Buckley1, 3, Alayna E. Loiselle1, 2, Geoffrey S. Pitt6, Chike Cao1, 2
1Center for Musculoskeletal Research, 2Department of Orthopeadics, 3Department of Biomedical Engineering,
University of Rochester Medical Center, Rochester, NY; 4Arizona Bone, Joint and Sports Medicine Center, Phoenix,
AZ; 5Sports Medicine and Shoulder Service, Hospital for Special Surgery, New York, NY; 6Cardiovascular
Research Institute, Weill Cornell Medicine, New York, NY,
Abstract
Tendons are tension-bearing tissues transmitting force from muscle to bone for body
movement. This mechanical loading is essential for tendon development, homeostasis, and healing
after injury. While Ca2+ signaling has been studied extensively for its roles in mechanotransduction,
regulating muscle, bone and cartilage development and homeostasis, knowledge about Ca2+
signaling and the source of Ca2+ signals in tendon fibroblast biology are largely unknown. Here,
we investigated the function of Ca2+ signaling through CaV1.2 voltage-gated Ca2+ channel in
tendon formation. Using a reporter mouse, we found that CaV1.2 is highly expressed in tendon
during development and downregulated in adult homeostasis. To assess its function, we generated
ScxCre;CaV1.2TS mice that express a gain-of-function mutant CaV1.2 channel (CaV1.2TS) in tendon.
We found that tendons in the mutant mice were approximately 2/3 larger and had more tendon
fibroblasts, but the cell density of the mutant mice decreased by around 22%. TEM analyses
demonstrated increased collagen fibrillogenesis in the hypertrophic tendon. Biomechanical testing
revealed that the hypertrophic Achilles tendons display higher peak load and stiffness, with no
changes in peak stress and elastic modulus. Proteomics analysis reveals no significant difference
in the abundance of major extracellular matrix (ECM) type I and III collagens, but mutant mice
had about 2-fold increase in other ECM proteins such as tenascin C, tenomodulin, periostin, type
XIV and type VIII collagens, around 11-fold increase in the growth factor of TGF-β family
myostatin, and significant elevation of matrix remodeling proteins including Mmp14, Mmp2 and
cathepsin K. Taken together, these data highlight roles for increased Ca2+ signaling through CaV1.2
on regulating expression of myostatin growth factor and ECM proteins for tendon collagen
fibrillogenesis during tendon formation.
Introduction
Tendons are tension-bearing tissues transmitting force from muscle to bone for body
movement. This specialized function of tendons is supported by tendon specific extracellular
matrix (ECM) structure with hierarchically organized collagen fibrils into fibers, fascicles, and
then tendon (Silver, Freeman, & Seehra, 2003). Collagen fibrils are composed primarily of type I
collagen, a triple helical polypeptide chains encoded by the genes Col1a1 and Col1a2. Other
tendon components are also present in the ECM important for tendon fibrillogenesis, such as type
III, V, VI, XII and XIV collagens, small leucine rich proteoglycans (e.g., decorin, aggrecan,
biglycan, and fibromodulin), and glycoproteins (e.g., tenascin C, tenomodulin, fibronectin, elastin,
and collagen oligomeric matrix protein)(Mouw, Ou, & Weaver, 2014).
Within the organized collagen fibrils reside tendon cells which are constantly subjected to
mechanical stimulation in vivo, including shear stress, tensile loading, and compressive force. Both
tendon fibroblasts, the major cells in tendon, and the tendon stem or progenitor cells (TSPCs) sense
and respond to a variety of mechanical loads (C. Zhang, Zhu, Zhou, Thampatty, & Wang, 2019; J. Zhang
& Wang, 2013), which are then converted to cellular responses and biological signals, a process
called mechanotransduction (Dunn & Olmedo, 2016; Ingber, 2006). Normal physiological loads are
required for tendon fibroblast differentiation, ECM synthesis and organization during tendon
development and adult tendon homeostasis (Bhole et al., 2009; Henderson & Carter, 2002; Kalson et
al., 2011; Nowlan, Murphy, & Prendergast, 2007; Subramanian, Kanzaki, Galloway, & Schilling, 2018). In
contrast, underloading or overloading results in decreased synthesis of tendon ECM proteins, ECM
degeneration or aberrant TSPC differentiation (Wang, Guo, & Li, 2012). However, the molecular
mechanisms of tendon mechanobiology remains unclear.
Ca2+, a ubiquitous intracellular signal, controls many cellular functions including muscle
contraction, immune cell activation, gene transcription and cell proliferation (Clapham, 2007). A
transient increase of intracellular Ca2+ concentration ([Ca2+]i) has been reported as one of the
earliest responses upon mechanical stimulation, activating various physiological and pathological
functions in tendon, ligaments, bone, cartilage, and muscle development, suggesting increased
[Ca2+]i plays a key role in mechanotransduction [reviewed in (Wall & Banes, 2005; Wall et al., 2016)].
In tendons, the Snedeker group recently showed with Ca2+ imaging that a transient [Ca2+]i increase
was observed by mechanical stimulation of the rat tail tendon fascicle ex vivo or isolated rat and
human tenocytes; Piezo1 loss-of-function, gain-of-function, and pharmacological approaches
identified this stretch-activated ion channel as the mechano-sensor in tendon cell
mechanotransduction, regulating tendon tissue stiffness (Passini et al., 2021). However, ion influx
through Piezo1 is rapid, transient, and not specific to Ca2+ (Coste et al., 2010). Other source of Ca2+
entry may be required for the mechanical stimulated Ca2+ response in tendon cells. It could be
functionally coupled with Piezo1 activation upon mechanical stimulation for opening and induces
substantial Ca2+ signal for tendon cell mechanotransduction. For example, the CaV1.2 channel has
been shown to mediate mechanosensitive Ca2+ influx in intestinal smooth muscle cells (Lyford et
al., 2002) and is broadly expressed (Pitt, Matsui, & Cao, 2021).
CaV1.2 belongs to the family of L-type voltage-gated Ca2+ channels (L-VGCCs), mediating
Ca2+ influx into the cell upon membrane depolarization. CaV1.2 channel is composed of α1C, β and
α2δ subunits (Catterall, 2000), among which the α1C subunit is the ion conducting pore while the
β and α2δ are auxiliary subunits that modulate channel properties (Serysheva, Ludtke, Baker, Chiu,
& Hamilton, 2002). CaV1.2 is highly voltage-dependent, which is characterized by an activation
threshold at a membrane potential around -20 mV (Catterall, Perez-Reyes, Snutch, & Striessnig,
2005). Most studies on CaV1.2 focused on its function in excitable cells such as cardiomyocytes
and neurons, where action potentials facilitate activation of this voltage-gated channel. However,
few studies focused on CaV1.2 in non-excitable cells, which have more restricted changes in
membrane potential. Interestingly, studies of Timothy syndrome (TS), a multiorgan disorder (e.g.,
cardiac arrhythmias, autism, syndactyly and craniofacial abnormalities), caused by a de novo
G406R mutation in the CaV1.2 pore forming α1C subunit encoded by CACNA1C (Splawski et al.,
2005; Splawski et al., 2004), revealed critical but previously unappreciated roles for CaV1.2 in
many non-excitable cells. For example, digital and craniofacial abnormalities in TS patients
suggest roles for CaV1.2 in development and morphogenesis, and that aberrant G406R mutant
channel (CaV1.2TS ) activity adversely affects canonical developmental signals. Consistent with
these hypotheses, we observed robust CaV1.2 endogenous expression in osteoblast progenitors
during craniofacial and limb development using a CaV1.2 lacZ reporter mouse line (C. Cao et al.,
2017; Kapil V. Ramachandran et al., 2013). In addition, by driving a CaV1.2TS transgene with
Prx1Cre, Col1a1Cre, Col2aCre or Sp7Cre, we demonstrated that CaV1.2TS mutant channels
promote bone formation via increased osteoblast differentiation and decreased osteoclast function.
This also prevented estrogen deficiency-induced bone loss, highlighting the unexpected role for
CaV1.2 in non-excitable tissue (Cao et al., 2019; C. Cao et al., 2017; Kapil V. Ramachandran et
al., 2013). The G406R mutation impairs CaV1.2 channel inactivation (closing) leading to more
Ca2+ ions to flow into the cytoplasm. Thus, the consequences of CaV1.2TS mutant channel
expression reported above result from a gain-of-function effect. However, whether CaV1.2 confers
analogous effects during tendon development is not known. To address this, we tested whether the
L-VGCC CaV1.2 is expressed in non-excitable tendon tissue, and whether an increase of Ca2+
signaling through CaV1.2TS mutant channels in tendon affects tendon formation.
Results
CaV1.2 is expressed in tendon fibroblasts during mouse tendon development and early
postnatal growth. To determine whether the Cav1.2 channel contributes to tendon formation
during development, postnatal growth and homeostasis, we first elucidated CaV1.2 channel
expression in tendons at different stages by using a CaV1.2 reporter mouse line (CaV1.2+/lacZ,
B6.129P2-Cacna1ctm1Dgen/J) in which the bacterial lacZ gene encoding β-galactosidase fused to a
nuclear localization signal was knocked into Cacna1c labeling nuclei of cells that express CaV1.2
(C. Cao et al., 2017; Kapil V. Ramachandran et al., 2013). We performed whole mount X-gal staining
of the forelimbs and hindlimbs, followed by frozen sectioning and histological analysis. We found
substantial X-gal staining in the developing digital tendons from E13.5 (Fig. 1A), and increased
intensity of staining at late embryonic stages (Fig. 1B). X-gal staining in digital tendons was
confirmed by histological analysis on frozen sections of the developing digits (Fig. 1C).
Furthermore, we found CaV1.2 expression persisted through early postnatal stages (~P3) in the
Achilles tendons (Fig. 1D) and patellar tendons (Fig. 1F). Notably, X-gal staining was exclusively
localized in the nucleus, which is distinct from any non-specifical staining resulting from
endogenous β-galactosidase activity, and thus provides an accurate picture of endogenous CaV1.2
expression. However, in adult tendons, CaV1.2 expression was dramatically downregulated. For
example, adult Achilles tendon demonstrated restricted CaV1.2 expression mostly seen in the
myotendinous junction with very sparse expression in the tendon substance (Fig. 1E). In contrast,
adult patellar tendon retains strong CaV1.2 expression throughout the tendon tissue, but in
relatively fewer cells compared with patellar tendons in early postnatal stage (Fig. 1G). In
summary, the dynamic expression of CaV1.2 during tendon development, postnatal growth and
adult homeostasis stage suggests that CaV1.2 and its mediated Ca2+ signaling may play a critical
role for tendon formation.
Expression of CaV1.2TS mutant channel in ScxCre+ lineage cells enhances tendon formation.
To investigate the role of CaV1.2 on tendon formation in vivo, we exploited the conditional
transgenic mouse line carrying the CaV1.2TS mutant cDNA in the Rosa26 locus (S. P. Paşca et al.,
2011). We generated ScxCre;CaV1.2TS mice to allow for CaV1.2TS expression under the Scleraxis
(Scx) promoter, which regulates transcription during tendon development, and is the earliest
known marker of tendon progenitors (Schweitzer et al., 2001). Macroscopic observation of
ScxCre;CaV1.2TS mutant mice revealed hypertrophic tendons in all tendons examined, including
Achilles tendon, plantaris tendons, patellar tendons, tail tendons, tendons in the forelimbs and back
tendons, compared to tendons in control mice at one month of age (Fig. 2). Histologic analyses
further confirmed tendon hypertrophy in ScxCre;CaV1.2TS mutant mice (Fig. 3). Fast green and
hematoxylin staining revealed that the cross-section area (CSA) was increased by 61% in patellar
tendons, 70% in plantaris tendons, and 74% in Achilles tendons from 1-month-old
ScxCre;CaV1.2TS mutant mice compared with those in control mice. Consistently, cell numbers in
patellar tendons, plantaris tendons and Achilles tendons were increased by 21%, 32% and 38% in
the mutant mice, respectively, indicating increased cell proliferation in the mutant tendons.
However, ScxCre;CaV1.2TS mutant mice had decreased cell density (cell number/CSA) by 22% in
all three types of tendons. This suggests that CaV1.2TS-expressing tendon fibroblasts are more
functionally active. Moreover, we found that ScxCre;CaV1.2TS mice had thicker tail tendon
fascicles, which have a broader size distribution than those in the control mice (Supplementary
Fig. 1). Notably, there was no change in the number of fascicles of both ventral and dorsal tail
tendons between genotypes, suggesting that CaV1.2TS mutant channels affect tendon fascicle
growth but not fascicle determination.
Expression of CaV1.2TS mutant channel in ScxCre+ lineage cells alters tendon collagen fibril
size distribution. To further investigate the effects of the CaV1.2TS mutant channel on tendon
collagen fibrillogenesis, we performed ultrastructural analyses using transmission electron
microscopy (TEM) of Achilles tendons from ScxCre;CaV1.2TS mutant mice and littermate controls
at 1 month of age. In the mutant Achilles tendons, the collagen fibrils displayed normal circular
cross-sectional profiles, similar to the control fibrils (Fig. 4A and B). However, the collagen fibril
density in the mutant tendons was increased by 68% without a significant change in interfibrillar
spacing (Fig. 4C and D), indicating increased collagen fibrillogenesis in ScxCre;CaV1.2TS mutant
mice. Furthermore, the mutant Achilles tendons were packed with more small- to-middle size
collagen fibrils, resulting in the change of the repartition of collagen fibrils in ScxCre;CaV1.2TS
mutant mice (Fig. 4E). In addition, a leftward shift of the fibril size distribution of the mutant
tendon in the cumulative fraction analysis further supported that the collagen fibrils in mutant mice
are smaller than those in the control mice (p <0.01 by Kolmogorov-Smirnov test) (Fig. 4F). Taken
together, these data suggest increased Ca2+ signaling through the CaV1.2TS mutant channel
increases collagen fibril assembly, which contributes to tendon hypertrophy in ScxCre;CaV1.2TS
mutant mice.
CaV1.2TS alters tendon biomechanical properties. Since cellular arrangement and fibril packing
are important determinants of biomechanical properties (Heather L. Ansorge et al., 2009; Dunkman
et al., 2013; Thorpe, Udeze, Birch, Clegg, & Screen, 2012), increased tendon growth and the change of
collagen fibril size distribution in ScxCre;CaV1.2TS mutant tendons may alter their mechanical
properties. Therefore, we performed the biomechanical property test in mature mutant Achilles
tendon in comparison with the control ones. Cross-sectional area (CSA), peak load, peak stress,
stiffness, and elastic modulus were measured. Mutant Achilles tendons displayed a significantly
larger CSA than the control Achilles tendons (Fig. 5A), exhibited ~1.50-fold increase in peak load,
and a ~1.52-fold increase in stiffness (tensile/displacement) in the force-displacement response
(Fig. 5B and C), indicating functional gain in structure properties of mutant Achilles tendons.
However, the material properties including the peak stress (the peak load per unit area), and the
elastic modulus (a measurement of the stiffness of an isotropic elastic material per unit area), did
not show significant changes between genotypes (Fig. 5D and E). This suggests that the increase
in structural stiffness in ScxCre;CaV1.2TS mutant tendons is due to the increased tendon mass.
CaV1.2TS alterations in the proteome of tendons. To define the molecular consequences of
CaV1.2TS expression, we quantified tissue-wide protein changes using mass spectrometry-based
proteomic analysis on ScxCre;CaV1.2TS versus control mice at 1 month of age. We observed 89
upregulated proteins (>1.5 fold-change) and 102 downregulated proteins (<-1.5 fold-change) in
CaV1.2TS-expressing mice compared with those in control mice (Fig. 6A and B, Supplementary
Fig. 2). The six proteins identified with the largest increase in expression in ScxCre;CaV1.2TS
mutant tendons were Mstn (myostatin, a member of the TGFβ-superfamily), Pavlb (a high affinity
Ca2+-binding protein similar to calmodulin in structure and function), Tnn (Tenascin-N), Cthrc1
(collagen triple helix repeat-containing protein 1), Angptl1 and Antptl2 (angiopoietin-like
proteins). In contrast, the proteins with the largest decrease were Chad (chondroadherin, a cartilage
matrix protein), Zmym4 (Zinc Finger MYM-type containing 4), Angptl7, Htra4 (high temperature
requirement factor A4) and Omd (Osteomodulin) (Fig. 6C). Moreover, Gene Ontology (GO)
enrichment analyses were performed to classify the putative functions of the differentially
upregulated and downregulated protein sets in ScxCre;CaV1.2TS mice in comparison with those of
the control mice. In GO terms of cellular component, we found that these differentially expressed
proteins are related to ECM, Proteinaceous ECM, extracellular region, extracellular exosome, and
extracellular space (Fig. 6D). Furthermore, GO analysis in term of biological process showed that
many of these proteins were involved in ECM organization (increase of Col8a1, Mmp14, Mmp2,
and Postn, decrease of Abi3bp, Ccdc80, Col15a1, Col24a1, Fbln1, Fbln2, Lgals3, Prdx4, Tnxb,
Vit and Vtn), collagen fibril organization (increase of Col14a1, decrease of Comp, Fmod, and
Tnxb), collagen catabolic process (increase of Ctsk, Mmp14, and Mmp2), response to mechanical
stimulus (increase of Mmp14, Mmp2, Postn, Tnc, and decrease of Thbs1 and Dcn) as shown in
Fig. 6E. We didn’t observe significant differences in the abundance of type I collagen and type
III collagen between genotypes.
To validate these findings, we performed real-time quantitative PCR (RT-qPCR) of
selected markers related to tendon formation including: Mstn, Tnc, Tnmd, Mmp14, Scx and Col1a1,
and found these gene mRNA expression was consistent with their expression at protein level (Fig.
7). For example, the growth factor Mstn, which is a positive regulator for tendon formation
(Christopher L. Mendias, Konstantin I. Bakhurin, & John A. Faulkner, 2008) and had the greatest protein
increase (~11.4-fold) in ScxCre;CaV1.2TS tendons, displayed a ~35.7-fold upregulation in mRNA
expression (Fig. 7A). The expression of Tnc, Tnmd and Mmp14 was increased around 2.9-, 2.5-
and 4.9-fold, irrespectively, in CaV1.2G406R-expressing Achilles(Fig. 7B-D). In contrast, Scx and
Col1a1 didn’t show significant differences between genotypes at mRNA level (Fig. 7E and F).
Taken together, these data suggest that CaV1.2TS mutant channels promoted tendon formation by
upregulating expression of Mstn and the less abundant ECM proteins for tendon collagen fibril
organization and ECM turnover.
Discussion
The current study identified the dynamic expression of endogenous CaV1.2, a voltage-
dependent Ca2+ channel in tendon fibroblasts during tendon development, growth and homeostasis
by utilizing a conclusive CaV1.2 lacZ reporter mouse line. This discovery prompted us to examine
whether this voltage-dependent Ca2+ channel functions in tendon formation. Using a transgenic
mouse model (ScxCre;CaV1.2TS), we demonstrated that expression of the gain-of-function G406R
mutant channel Cav1.2TS specifically in ScxCre+ tendon fibroblasts dramatically promotes tendon
formation. Biomechanical testing showed that the enlarged tendons in ScxCre;CaV1.2TS mutant
mice display a dramatic increase of their structural properties including stiffness and peak load,
but have similar material properties, such as peak stress and elastic modulus, compared to WT
tendons. Therefore, the increased tendon stiffness and peak load could be owing to the proportional
increase of tendon thickness (measured by CSA). Notably, these changes of tendon biomechanical
properties in ScxCre;CaV1.2TS mutant mice are comparable to those in the adaptation of human
tendons after years of long-term training with larger tendon CSA and increased tendon stiffness,
but no differences in material properties based on the meta-analysis of tendon property changes
with training (Wiesinger, Kosters, Muller, & Seynnes, 2015), suggesting the role of Ca2+ signaling via
CaV1.2 may be linked with tendon loading and mechanotransduction. Taken together, our data
provides the first evidence that modulating Ca2+ signaling through CaV1.2 in tendon fibroblasts in
vivo affects tendon formation, highlighting additional unexpected roles of CaV1.2 channels in non-
excitable tissues that we previously reported (Cao et al., 2019; C. Cao et al., 2017; Kapil V.
Ramachandran et al., 2013).
The tendon hypertrophy in ScxCre;CaV1.2TS mutant mice is likely due to both increased tendon
fibroblast proliferation and ECM collagen fibril formation. This speculation is supported by our
findings that the mutant tendons had significantly more cells but reduced cellular density (Fig.3).
This specific spatial organization of tendon cells can result from more increased collagen
fibrillogenesis and fibril assembly relative to the increase in tendon cell proliferation. This is
supported by our TEM findings that ScxCre;CaV1.2TS mutant tendons generated more collagen
fibrils with smaller diameter versus WT tendons (Fig. 4). During tendon development, short and
small-diameter fibril intermediates are initially assembled, which serve as the building blocks to
form large and long collagen fibrils in late stage of tendon growth (Nurminskaya & Birk, 1998). The
accumulated small-to-medium size fibrils in the mutant tendons indicated an increased collagen
fibrillogenesis and active fibril formation in response to upregulated Ca2+ signaling in
ScxCre;CaV1.2TS mice. Subsequently, we performed proteomic analysis to define the tissue-wide
protein change and understand the molecular mechanisms responsible for the altered collagen
fibrillogenesis. It is known that tendon collagen fibrillogenesis can be regulated in several ways,
including 1) the synthesis of fibril-forming collagen (predominantly type I collagen with varying
amounts of type II, III and V in tendon), 2) the abundance of specific fibril-associated
proteoglycans and glycoproteins (such as decorin, biglycan, lumican, fibromodulin and COMP),
or 3) the abundance of the fibril-associated collagen with interrupted triple helics (FACIT) (such
as type IX, XII, XIV, XVI, XIX, XX, XXI, XXII, and XXVI collagens) (Kadler, Baldock, Bella, &
Boot-Handford, 2007; Mouw et al., 2014; Nurminskaya & Birk, 1998). A deficiency of decorin, lumican,
fibromodulin, COMP and type XVI collagen in vivo has been shown to result in larger and
disorganized collagen fibrils (H. L. Ansorge et al., 2009; Chakravarti et al., 1998; Danielson et al., 1997;
Piróg et al., 2010; Svensson et al., 1999). Interesting, our proteomic analysis showed that in
ScxCre;CaV1.2TS mutant tendons, the major fibril-forming collagens (Col1a1 and Col1a2) was not
differentially expressed versus control tendons. However, the fibril-associated FACIT collagen
(Col14a1) was significantly upregulated, while proteoglycans including decorin and fibromodulin
and glycoprotein COMP were downregulated in ScxCre;CaV1.2TS tendons. Thus, while a
combination of these fibril-associated macromolecules with variable amount contributed to the
tendon collagen fibrillogenesis and growth, type XIV collagen may be the dominant factor
affecting collagen fibrillogenesis and assembly in ScxCre;CaV1.2TS tendons. Furthermore, collagen
fibrillogenesis requires activities of matrix metalloproteinases (MMPs) and their corresponding
tissue inhibitors (TIMPs) to convert procollagen into collagen by removing the N- and C-pro-
peptides, and alter the surface of fibril intermediates and/or interfibrillar matrix (Jones et al., 2006;
Mouw et al., 2014). Notably, we found that Mmp2 and Mmp14 were both dramatically upregulated
while Timp3 was downregulated in ScxCre;CaV1.2TS tendons by proteomics analysis (Fig. 6E and
6G). It is known that Mmp2 is initially produced as latent pro-Mmp2, which requires the
membrane type (MT) MMPs such as Mmp14 for cleavage and activation (Deryugina et al., 2001;
Strongin et al., 1995). It has been shown that in an in vitro system, knockdown of Mmp14 inhibited
proMmp2 activation (Wilkinson et al., 2012). Moreover, Mmp14 has been shown to promote new
formation of collagen fibers, the high order of tendon structure; the tendons of mutant mice lacking
Mmp14 have fewer collagen fibers than normal mice (Taylor et al., 2015). Whereas all active MMPs
can be inhibited by TIMPs, Timp3 is the main TIMP which inhibits activity against some of the
ADAMS and ADAMTS metalloproteinases (Del Buono, Oliva, Osti, & Maffulli, 2013; Mochizuki &
Okada, 2007). Taken together, decrease in Timp3 expression along with the elevated expression of
Mmp2, Mmp14 and Type XIV FACIT collagen in ScxCre;CaV1.2TS mutant tendon will facilitate
procollagen maturation, collagen fibril assembly and remodeling, all of which contribute to the
active collagen fibrillogenesis and the higher structure fiber growth, ultimately tendon hypertrophy
upon increased Ca2+ signaling.
Our proteomic analysis also identified a dramatic increase of myostatin in ScxCre;CaV1.2TS
mutant tendons. Myostatin, also called growth/differentiation factor-8 (GDF-8), is the growth
factor of the transforming growth factor-β (TGF-β) superfamily. Myostatin is mostly known as a
negative regulator of muscle growth, and the loss of myostatin function is associated with
hypermuscular phenotypes in mice and cattle (Alexandra C. McPherron, Lawler, & Lee, 1997; A. C.
McPherron & Lee, 1997). In contrast, myostatin was found to be a positive regulator for tendon
formation as myostatin-deficient mice have small (a decrease in fibroblast number) and brittle
tendons (a higher peak stress, a lower peak strain and increased stiffness) (C. L. Mendias, K. I.
Bakhurin, & J. A. Faulkner, 2008). Thus, upregulation of myostatin in ScxCre;CaV1.2TS mutant tendon
may promote tendon hypertrophy in the mutant mice upon increased Ca2+ signaling. However,
whether myostatin signaling mediates tendon growth in ScxCre;CaV1.2TS mice requires further
investigation. Conditional knockout of Mstn alleles in CaV1.2TS-expressing tendon fibroblasts
would be necessary to exclude the possibility that other myostatin-independent signaling pathways
may also contribute to CaV1.2TS-induced tendon formation. Furthermore, previous studies have
shown that the p38 mitogen-activate protein kinase (MAPK) and Smad2/3 signaling cascades (Lee
& McPherron, 2001; Philip, Lu, & Gao, 2005) in tendon fibroblasts were activated in response to
myostatin treatment, which are required for the increased cell proliferation and gene expression
including Scx, Col1a1, and Tnmd (C. L. Mendias et al., 2008). Consistently, we observed increased
expression of Tnmd both at protein and mRNA levels in ScxCre;CaV1.2TS mice. However, we
didn’t detect a significant change of Scx and Col1a1 expression in response to upregulated
myostatin in CaV1.2TS-expressing tendons. This discrepancy may be due to the maximal dose of
myostatin normally used in in vitro cell culture while in in vivo system, myostatin may be
dominantly maintained in an inactive form. Nevertheless, upregulation of Tnmd and Tnc, known
as the downstream targets of transcription factor Scx and regulated by pSmad2/3 pathway (Berthet
et al., 2013; Shukunami, Takimoto, Oro, & Hiraki, 2006), didn’t depend on a corresponding increase of
Scx expression in ScxCre;CaV1.2TS tendons.
Although the mechanism by which CaV1.2 activation occurs in non-excitable tendon
fibroblasts has yet to be explored, our finding that CaV1.2 is dynamically expressed during tendon
development, growth and adult homeostasis suggests a role of Ca2+ signaling in tendon fibroblasts
spatiotemporally. It also implies the potential mechanisms to activate the voltage-gated Ca2+
channel. In adults, CaV1.2 expression is dramatically decreased compared to that during tendon
development and early postnatal growth. CaV1.2 expression in adult Achilles tendon is restricted
to the myotendinous junction (Fig. 1), a site where forces generated by myofibrils are transmitted
across the cell membrane to act on tendon (Tidball & Lin, 1989). Notably, at this interface between
an excitable muscle and non-excitable tendon, topological action potential may occur via gap
junction (Ori et al., 2022), which results in non-excitable tendon fibroblasts in myotendinous
junctions electrically excitable and then activate this voltage-dependent Ca2+ channel. Tendon
fibroblasts in myotendinous junction could function as a signal initiator, which once stimulated by
muscle contraction, will diffuse signal factors down to the neighboring fibroblasts. However,
during stages before the stable myotendinous junction forms in tendons, usually at 1 month old in
rats (Curzi, Ambrogini, Falcieri, & Burattini, 2013), or in patellar tendons (ligaments precisely) without
the myotendinous junction structure, other mechanisms to activate CaV1.2 voltage-dependent Ca2+
channel may also exist, by which the depolarizing drive does not require action potentials.
Recently, the mechanosensitive channel Piezo1 has been reported to be expressed in tendon tissue
and to sense the mechanical loading in tenocytes; knocking out Piezo1 in cultured tenocytes greatly
decreases shear stress-induced Ca2+ signals (Passini et al., 2021). Given the fact that Piezo1 and its
homolog Piezo2 conduct a rapid and transient cation influx (non-selective to Ca2+) (Coste et al.,
2010), activation of Piezo1/2 by mechanical stimuli may provide an inward depolarizing current,
which in turn activates the voltage-gated CaV1.2 channels co-expressed in tendon fibroblasts to
amplify the Ca2+ signal. However, whether CaV1.2 is required for mechanotransduction in tendon
has yet to be determined. Nevertheless, this spontaneous CaV1.2 activation without action potential
may be inefficient which is compensated by higher expression of CaV1.2 channels to initiate the
Ca2+ signaling for mechanotransduction. It is the case that we observed more robust CaV1.2
expression in all tendons during early tendon growth before the formation of a stable myotendinous
junction as well as in adult patellar tendons without myotendinous junction.
In summary, our data identified a novel role of CaV1.2 voltage-dependent Ca2+ channel in non-
excitable tissue tendon and demonstrated that increased Ca2+ influx through CaV1.2 promotes
tendon formation predominantly by regulating tendon collagen fibrillogenesis. This was achieved
through increased expression of the growth factor myostatin and a combination of differentially
expressed fibril associated FACIT type XIV collagen, MMPs, TIMPs and other ECM proteins.
These biochemical responses following increased Ca2+ signaling may cooperate in the adaptation
of tendon in response to mechanical loading or tendon healing after tendon injuries.
Pharmacologically, CaV1.2 agonists, such as BayK-8644 or FPL 64176 can mimic the effect of
CaV1.2TS to increased Ca2+ influx across the plasma membrane. Therefore, our data in this study
highlights a potential therapeutic strategy to target CaV1.2 channel and promote tendon formation
and healing after injuries.
Methods
Mice: All animal studies were approved by the University of Rochester Committee for Animal
Resources. CaV1.2+/LacZ and CaV1.2TS mice have been described previously(Chike Cao et al., 2017;
Sergiu P Paşca et al., 2011; Kapil V Ramachandran et al., 2013). CaV1.2+/lacZ mouse line carries a lacZ
reporter with a nuclear localization signal under the promoter of Cacna1c, the gene encoding
CaV1.2. CaV1.2TS mouse line carries a rat G406R TS-causing mutant CaV1.2 cDNA which was
knocked into the Rosa26 locus with an upstream floxed stop codon to control the transgene
expression by the Cre-loxP system. Homozygous floxed CaV1.2TS mice were crossed with the
transgenic ScxCre mice (provided by R. Schweitzer) to induce CaV1.2TS expression and modulate
the Ca2+ signals in tendon during development and growth. All tendon analyses were performed
on mice at 1 month of age unless otherwise specified. Mutant mice or littermate controls of both
male and female mice were analyzed unless otherwise specified.
X-gal staining and histology: It has been descripted previously (Chike Cao et al., 2017). Briefly,
visualization of lacZ expression was done by X-gal 5-bromo-4-chloro-3-indolyl-β-D-
galactopyranoside) staining in whole mount embryos or limbs. For whole mount X-gal staining,
embryos, forelimbs or hindlimbs were fixed in ice-cold fixation solution (2% paraformaldehyde
and 0.5% glutaradehyde in 1× PBS) for 20 minutes (for embryos), 1 hour (for postnatal stage), or
2 hours (for adult stage), washed with 1x PBS for 3 times, each 10 minutes, processed with X-gal
staining solution (5 mM potassium ferrocyanide, 5 mM potassium ferricyanide, 1 mg/ml X-gal, 2
mM MgCl2, 0.1% sodium deoxycholate, and 0.2% IGEPAL CA-630) in dark for 24 ~ 72 hours at
37°C. For histological analysis, whole mount X-gal-stained tissues were further decalcified with
14% EDTA at 4°C, 30% sucrose, and snap-frozen embedded with OCT compound (Sakura
Finetek). Frozen sections (10 µm thickness) were counterstained with Nuclear Fast Red. For
histology on fresh tendon tissue, tendons were isolated and immediately processed into 30%
sucrose for 1 hour at room temperature, snap-frozen embedded with OCT compound. Samples
were cross-sectioned at 10 µm thickness. Sections were air-dried for 1 hour at room temperature,
fixed with 2% paraformaldehyde and 0.5% glutaradehyde in 1× PBS for 10 minutes, followed by
Hematoxylin and Fast Green staining with standard protocols.
RNA extraction and RT-qPCR: For total RNA isolated from tendon tissue, miRNeasy Mini Kit
(Qiagen) was used. Achilles tendon and plantaris tendon were carefully dissected from control and
ScxCre;CaV1.2TS mutant mice, both sexes at 1 month of age. Tendons from each animal
represented as one biological replicate without pooling tissues from different animals. Tendons
were homogenized in QiAzol lysis reagent (Qiagen) using Biomasher II disposable micro tissue
homogenizer and total RNA was purified following the kit instruction. Total RNA (500 ng) was
reverse-transcribed to cDNA using cDNA reverse transcription kit (Applied Biosystems, Thermo
Fisher Scientific) and qPCR with SYBR green Supermix (Bio-Rad). Relative expression was
calculated using the 2-ΔΔCt methods by first normalization to Gapdh (ΔCt) and second
normalization to control samples (ΔΔCt). The primers used for tested genes were listed in Table
S1, with Mstn, Scx, Tnmd, and Gapdh primers were previously described (Christopher L. Mendias
et al., 2008).
Collagen transmission electron microscopy (TEM): Achilles tendons from control and
ScxCre;CaV1.2TS mutant mice at 1 month of age were used for TEM analysis. First, mouse
hindlimbs were fixed in 1 x PBS containing 1.5% glutaraldehyde/1.5% formaldehyde (Electron
Microscopy Sciences, Cat#: 15950), 0.05% tannic acid at 4 °C for overnight with gentle agitation.
Achilles tendons were dissected out and post-fixed in 1% OsO4. After washing with 1 x PBS and
dehydration in a graded series of ethanol, tendon samples were rinsed in propylene oxide,
infiltrated in Spurrs epoxy and polymerized at 70 °C for overnight. Ultrathin sections at 80 nm
were used for imaging using a FEI G20 TEM by the core service at MicroImaging Center, Shriners
Hospital for Children, Portland. ImageJ was used for the measurement of collagen fibril CSA.
Mechanical properties testing: Uniaxial displacement-controlled stretching at 1% strain per
second until failure descripted previously (Korcari, Buckley, & Loiselle, 2022) was applied to Achilles
tendons isolated from ScxCre;CaV1.2TS mutant or littermate controls at 1 month of age for both
sexes. Achilles tendon preparation followed the previous reported description with some
modification (Sarver et al., 2017). Briefly, Achilles tendons (without plantaris tendons) were
dissected out with one end attaching to the calcaneus bone and the other end with muscle. Tendons
were wrapped in 1 x PBS-soaked kimwrap and stored at -20 °C until use. Prior to mechanical tests,
tendons were thawed at room temperature, submerged in 1 x PBS, cleaned away of muscle to
prevent slipping, placed with both ends between two layers of sandpaper, glued with cranoacrylate
(Superglue, LOCTITE), and secured with a compression clamps. Each Achilles tendon was first
quantified by its gauge length and CSA from 3 evenly spaced width and depth measurements from
high-resolution digital photographs of both top and side views of the tendon (Olympus BX51,
Olympus). Mechanical property testing was performed in a bath containing 1 x PBS at room
temperature. A uniaxial displacement-controlled stretching of 1% strain per second was applied
until failure occurred. Load and displacement were recorded, and the failure of each mechanically
tested tendon was confirmed which often occurred at tendon mid-substance. Tendon peak load
was taken as the maximum load prior tendon’s failure, while tendon stiffness was specified by the
slope of the linear region from the load-displacement curve. Tendon tensile stress was defined as
the recorded load divided by tendon CSA, while tendon tensile strain as the displacement divided
by the gauge length. Tendon elastic modulus was calculated by the slope of the linear region from
the plotted tendon tensile stress-strain curve. Tendon structural properties (stiffness, and peak load)
and material properties (peak stress and elastic modulus) were determined from each Achilles
tendon.
Proteomics and data analysis: Mass spectrometry (MS) proteomic analysis was performed at the
Mass Spectrometry Resource Laboratory, University of Rochester. Achilles tendons (combined
with plantaris tendon) were isolated from ScxCre;CaV1.2TS mutant or littermate control mouse at
1 month of age. Tendons from each animal represented as one biological replicate without pooling
tissues from different animals. Trypsin (Thermo Scientific) was used to digest the tendon proteins
followed by disulfide bond reduction with addition of 5 mM of Bond-Breaker TCEP solution
(Thermo Scientific) and by alkylation of reduced cysteines with the addition of 10 mM of
iodoacetamine. LC-MS/MS analysis was performed using a Q Exactive Plus mass spectrometer
(Thermo Scientific). Raw MS data files were analyzed with PEAKS to identify protein
composition. Searches were performed against the Uniprot mouse proteomes database
(UP000000589). Search results were adjusted to 1% false discovery rate (FDR), filtering out
peptides which had a p-value greater than 0.01. Z-scores were calculated from the normalized
abundance of each protein to create heatmaps via GraphPad. Only proteins with significantly
different abundance (p < 0.05) were used in the heatmaps. In addition, DAVID bioinformatics
Resources (https://david.ncifcrf.gov/tools.jsp) was used for GO term enrichment analysis for
proteins exhibiting 1.5-fold higher or 1.5-fold lower levels of expression in abundance and FDR
p-value <0.05.
Statistics: Statistical analyses were performed using GraphPad Prism 9.0 or OriginPro 8. Two-
tailed unpaired t tests were used to compare between mutant and control groups. Fold changes
were calculated by dividing the value of the mutant group by the value of the control group.
Increasing or decreasing changes were calculated by dividing the value of difference between the
mutant group and the control group by the value of the control group and then multiplying 100.
Kolmogorov-Smirnov test (http://www.physics.csbsju.edu/stats/KS-test.n.plot_form.html) was
performed to determine if the size of collagen fibril differs significantly between groups. * p <
0.05. ** p < 0.01.
Acknowledgments: We thank Douglas keene (MicrroImaging Center, Shriners Hospital for
Children) for providing the TEM core service, Dr. Sina Ghaemmaghami (Mass Spectrometry
Resource Lab, University of Rochester) for providing Mass spectrometry service and Dr. Edward
Schwartz (Center for Musculoskeletal Research, University of Rochester) for his critical reading
and helpful suggestions. This work was supported by NIH NIAMS R21AR075214 and
P30AR69655.
References
Ansorge, H. L., Meng, X., Zhang, G., Veit, G., Sun, M., Klement, J. F., . . . Birk, D. E. (2009). Type XIV Collagen
Regulates
Fibrillogenesis*.
Journal
of
Biological
Chemistry,
284(13),
8427-8438.
doi:https://doi.org/10.1074/jbc.M805582200
Ansorge, H. L., Meng, X., Zhang, G., Veit, G., Sun, M., Klement, J. F., . . . Birk, D. E. (2009). Type XIV Collagen
Regulates Fibrillogenesis: PREMATURE COLLAGEN FIBRIL GROWTH AND TISSUE DYSFUNCTION IN
NULL MICE. J Biol Chem, 284(13), 8427-8438. doi:10.1074/jbc.M805582200
Berthet, E., Chen, C., Butcher, K., Schneider, R. A., Alliston, T., & Amirtharajah, M. (2013). Smad3 binds
Scleraxis and Mohawk and regulates tendon matrix organization. J Orthop Res, 31(9), 1475-1483.
doi:10.1002/jor.22382
Bhole, A. P., Flynn, B. P., Liles, M., Saeidi, N., Dimarzio, C. A., & Ruberti, J. W. (2009). Mechanical strain
enhances survivability of collagen micronetworks in the presence of collagenase: implications for
load-bearing matrix growth and stability. Philos Trans A Math Phys Eng Sci, 367(1902), 3339-3362.
doi:10.1098/rsta.2009.0093
Cao, C., Oswald, A. B., Fabella, B. A., Ren, Y., Rodriguiz, R., Trainor, G., . . . Pitt, G. S. (2019). The CaV1.2 L-
type calcium channel regulates bone homeostasis in the middle and inner ear. Bone, 125, 160-
168. doi:10.1016/j.bone.2019.05.024
Cao, C., Ren, Y., Barnett, A. S., Mirando, A. J., Rouse, D., Mun, S. H., . . . Pitt, G. S. (2017). Increased Ca2+
signaling through CaV1.2 promotes bone formation and prevents estrogen deficiency-induced
bone loss. JCI Insight, 2(22), e95512. doi:10.1172/jci.insight.95512
Cao, C., Ren, Y., Barnett, A. S., Mirando, A. J., Rouse, D., Mun, S. H., . . . Pitt, G. S. (2017). Increased Ca2+
signaling through CaV1.2 promotes bone formation and prevents estrogen deficiency-induced
bone loss. JCI Insight, 2(22). doi:10.1172/jci.insight.95512
Catterall, W. A. (2000). Structure and regulation of voltage-gated Ca2+ channels. Annu Rev Cell Dev Biol,
16, 521-555. doi:10.1146/annurev.cellbio.16.1.521
Catterall, W. A., Perez-Reyes, E., Snutch, T. P., & Striessnig, J. (2005). International Union of Pharmacology.
XLVIII. Nomenclature and structure-function relationships of voltage-gated calcium channels.
Pharmacol Rev, 57(4), 411-425. doi:10.1124/pr.57.4.5
Chakravarti, S., Magnuson, T., Lass, J. H., Jepsen, K. J., LaMantia, C., & Carroll, H. (1998). Lumican regulates
collagen fibril assembly: skin fragility and corneal opacity in the absence of lumican. J Cell Biol,
141(5), 1277-1286. doi:10.1083/jcb.141.5.1277
Clapham, D. E. (2007). Calcium signaling. Cell, 131(6), 1047-1058. doi:10.1016/j.cell.2007.11.028
Coste, B., Mathur, J., Schmidt, M., Earley, T. J., Ranade, S., Petrus, M. J., . . . Patapoutian, A. (2010). Piezo1
and Piezo2 are essential components of distinct mechanically activated cation channels. Science,
330(6000), 55-60. doi:10.1126/science.1193270
Curzi, D., Ambrogini, P., Falcieri, E., & Burattini, S. (2013). Morphogenesis of rat myotendinous junction.
Muscles Ligaments Tendons J, 3(4), 275-280.
Danielson, K. G., Baribault, H., Holmes, D. F., Graham, H., Kadler, K. E., & Iozzo, R. V. (1997). Targeted
disruption of decorin leads to abnormal collagen fibril morphology and skin fragility. J Cell Biol,
136(3), 729-743. doi:10.1083/jcb.136.3.729
Del Buono, A., Oliva, F., Osti, L., & Maffulli, N. (2013). Metalloproteases and tendinopathy. Muscles
Ligaments Tendons J, 3(1), 51-57. doi:10.11138/mltj/2013.3.1.051
Deryugina, E. I., Ratnikov, B., Monosov, E., Postnova, T. I., DiScipio, R., Smith, J. W., & Strongin, A. Y. (2001).
MT1-MMP initiates activation of pro-MMP-2 and integrin alphavbeta3 promotes maturation of
MMP-2 in breast carcinoma cells. Exp Cell Res, 263(2), 209-223. doi:10.1006/excr.2000.5118
Dunkman, A. A., Buckley, M. R., Mienaltowski, M. J., Adams, S. M., Thomas, S. J., Satchell, L., . . . Iozzo, R.
V. (2013). Decorin expression is important for age-related changes in tendon structure and
mechanical
properties.
Matrix
Biology,
32(1),
3-13.
Retrieved
from
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3615887/pdf/nihms-426371.pdf
Dunn, S. L., & Olmedo, M. L. (2016). Mechanotransduction: Relevance to Physical Therapist Practice—
Understanding Our Ability to Affect Genetic Expression Through Mechanical Forces. Physical
Therapy, 96(5), 712-721. doi:10.2522/ptj.20150073
Henderson, J. H., & Carter, D. R. (2002). Mechanical induction in limb morphogenesis: the role of growth-
generated
strains
and
pressures.
Bone,
31(6),
645-653.
Retrieved
from
https://www.ncbi.nlm.nih.gov/pubmed/12531557
Ingber, D. E. (2006). Cellular mechanotransduction: putting all the pieces together again. FASEB J, 20(7),
811-827. doi:10.1096/fj.05-5424rev
Jones, G. C., Corps, A. N., Pennington, C. J., Clark, I. M., Edwards, D. R., Bradley, M. M., . . . Riley, G. P.
(2006). Expression profiling of metalloproteinases and tissue inhibitors of metalloproteinases in
normal and degenerate human achilles tendon. Arthritis & Rheumatism, 54(3), 832-842.
doi:https://doi.org/10.1002/art.21672
Kadler, K. E., Baldock, C., Bella, J., & Boot-Handford, R. P. (2007). Collagens at a glance. Journal of Cell
Science, 120(12), 1955-1958. doi:10.1242/jcs.03453
Kalson, N. S., Holmes, D. F., Herchenhan, A., Lu, Y., Starborg, T., & Kadler, K. E. (2011). Slow stretching that
mimics embryonic growth rate stimulates structural and mechanical development of tendon-like
tissue in vitro. Dev Dyn, 240(11), 2520-2528. doi:10.1002/dvdy.22760
Korcari, A., Buckley, M. R., & Loiselle, A. E. (2022). Characterization of scar tissue biomechanics during
adult murine flexor tendon healing. Journal of the Mechanical Behavior of Biomedical Materials,
130, 105192. doi:https://doi.org/10.1016/j.jmbbm.2022.105192
Lee, S. J., & McPherron, A. C. (2001). Regulation of myostatin activity and muscle growth. Proc Natl Acad
Sci U S A, 98(16), 9306-9311. doi:10.1073/pnas.151270098
Lyford, G. L., Strege, P. R., Shepard, A., Ou, Y., Ermilov, L., Miller, S. M., . . . Farrugia, G. (2002). alpha(1C)
(Ca(V)1.2) L-type calcium channel mediates mechanosensitive calcium regulation. Am J Physiol
Cell Physiol, 283(3), C1001-1008. doi:10.1152/ajpcell.00140.2002
McPherron, A. C., Lawler, A. M., & Lee, S.-J. (1997). Regulation of skeletal muscle mass in mice by a new
TGF-p superfamily member. Nature, 387(6628), 83-90. doi:10.1038/387083a0
McPherron, A. C., & Lee, S. J. (1997). Double muscling in cattle due to mutations in the myostatin gene.
Proc Natl Acad Sci U S A, 94(23), 12457-12461. doi:10.1073/pnas.94.23.12457
Mendias, C. L., Bakhurin, K. I., & Faulkner, J. A. (2008). Tendons of myostatin-deficient mice are small,
brittle, and hypocellular. Proceedings of the National Academy of Sciences, 105(1), 388-393.
doi:doi:10.1073/pnas.0707069105
Mendias, C. L., Bakhurin, K. I., & Faulkner, J. A. (2008). Tendons of myostatin-deficient mice are small,
brittle, and hypocellular. Proc Natl Acad Sci U S A, 105(1), 388-393. doi:10.1073/pnas.0707069105
Mochizuki, S., & Okada, Y. (2007). ADAMs in cancer cell proliferation and progression. Cancer Sci, 98(5),
621-628. doi:10.1111/j.1349-7006.2007.00434.x
Mouw, J. K., Ou, G., & Weaver, V. M. (2014). Extracellular matrix assembly: a multiscale deconstruction.
Nature Reviews Molecular Cell Biology, 15(12), 771-785. doi:10.1038/nrm3902
Nowlan, N. C., Murphy, P., & Prendergast, P. J. (2007). Mechanobiology of embryonic limb development.
Ann N Y Acad Sci, 1101, 389-411. doi:10.1196/annals.1389.003
Nurminskaya, M. V., & Birk, D. E. (1998). Differential expression of genes associated with collagen fibril
growth in the chicken tendon: identification of structural and regulatory genes by subtractive
hybridization. Arch Biochem Biophys, 350(1), 1-9. doi:10.1006/abbi.1997.0498
Ori, H., Ramirez, M. D., Hayward, R. F., Tian, H., Ortiz, G., & Cohen, A. E. (2022). Observation of topological
action
potentials
in
engineered
tissues.
bioRxiv,
2022.2003.2016.484369.
doi:10.1101/2022.03.16.484369
Paşca, S. P., Portmann, T., Voineagu, I., Yazawa, M., Shcheglovitov, A., Paşca, A. M., . . . Nishino, S. (2011).
Using iPSC-derived neurons to uncover cellular phenotypes associated with Timothy syndrome.
Nat Med, 17(12), 1657-1662.
Paşca, S. P., Portmann, T., Voineagu, I., Yazawa, M., Shcheglovitov, A., Paşca, A. M., . . . Dolmetsch, R. E.
(2011). Using iPSC-derived neurons to uncover cellular phenotypes associated with Timothy
syndrome. Nat Med, 17(12), 1657-1662. doi:10.1038/nm.2576
Passini, F. S., Jaeger, P. K., Saab, A. S., Hanlon, S., Chittim, N. A., Arlt, M. J., . . . Snedeker, J. G. (2021).
Shear-stress sensing by PIEZO1 regulates tendon stiffness in rodents and influences jumping
performance in humans. Nat Biomed Eng, 5(12), 1457-1471. doi:10.1038/s41551-021-00716-x
Philip, B., Lu, Z., & Gao, Y. (2005). Regulation of GDF-8 signaling by the p38 MAPK. Cell Signal, 17(3), 365-
375. doi:10.1016/j.cellsig.2004.08.003
Piróg, K. A., Jaka, O., Katakura, Y., Meadows, R. S., Kadler, K. E., Boot-Handford, R. P., & Briggs, M. D. (2010).
A mouse model offers novel insights into the myopathy and tendinopathy often associated with
pseudoachondroplasia and multiple epiphyseal dysplasia. Hum Mol Genet, 19(1), 52-64.
doi:10.1093/hmg/ddp466
Pitt, G. S., Matsui, M., & Cao, C. (2021). Voltage-Gated Calcium Channels in Nonexcitable Tissues. Annu
Rev Physiol, 83, 183-203. doi:10.1146/annurev-physiol-031620-091043
Ramachandran, K. V., Hennessey, J. A., Barnett, A. S., Yin, X., Stadt, H. A., Foster, E., . . . Kirby, M. L. (2013).
Calcium influx through L-type Ca V 1.2 Ca 2+ channels regulates mandibular development. The
Journal of clinical investigation, 123(4), 1638-1646.
Ramachandran, K. V., Hennessey, J. A., Barnett, A. S., Yin, X., Stadt, H. A., Foster, E., . . . Pitt, G. S. (2013).
Calcium influx through L-type CaV1.2 Ca2+ channels regulates mandibular development. The
Journal of clinical investigation, 123(4), 1638-1646. doi:10.1172/JCI66903
Sarver, D. C., Kharaz, Y. A., Sugg, K. B., Gumucio, J. P., Comerford, E., & Mendias, C. L. (2017). Sex
differences in tendon structure and function. J Orthop Res, 35(10), 2117-2126.
doi:10.1002/jor.23516
Schweitzer, R., Chyung, J. H., Murtaugh, L. C., Brent, A. E., Rosen, V., Olson, E. N., . . . Tabin, C. J. (2001).
Analysis of the tendon cell fate using Scleraxis, a specific marker for tendons and ligaments.
Development,
128(19),
3855-3866.
Retrieved
from
https://www.ncbi.nlm.nih.gov/pubmed/11585810
Serysheva, II, Ludtke, S. J., Baker, M. R., Chiu, W., & Hamilton, S. L. (2002). Structure of the voltage-gated
L-type Ca2+ channel by electron cryomicroscopy. Proc Natl Acad Sci U S A, 99(16), 10370-10375.
doi:10.1073/pnas.162363499
Shukunami, C., Takimoto, A., Oro, M., & Hiraki, Y. (2006). Scleraxis positively regulates the expression of
tenomodulin, a differentiation marker of tenocytes. Developmental biology, 298(1), 234-247.
Silver, F. H., Freeman, J. W., & Seehra, G. P. (2003). Collagen self-assembly and the development of tendon
mechanical properties. J Biomech, 36(10), 1529-1553. doi:10.1016/s0021-9290(03)00135-0
Splawski, I., Timothy, K. W., Decher, N., Kumar, P., Sachse, F. B., Beggs, A. H., . . . Keating, M. T. (2005).
Severe arrhythmia disorder caused by cardiac L-type calcium channel mutations. Proc Natl Acad
Sci U S A, 102(23), 8089-8096; discussion 8086-8088. doi:10.1073/pnas.0502506102
Splawski, I., Timothy, K. W., Sharpe, L. M., Decher, N., Kumar, P., Bloise, R., . . . Keating, M. T. (2004).
Ca(V)1.2 calcium channel dysfunction causes a multisystem disorder including arrhythmia and
autism. Cell, 119(1), 19-31. doi:10.1016/j.cell.2004.09.011
Strongin, A. Y., Collier, I., Bannikov, G., Marmer, B. L., Grant, G. A., & Goldberg, G. I. (1995). Mechanism of
cell surface activation of 72-kDa type IV collagenase. Isolation of the activated form of the
membrane metalloprotease. J Biol Chem, 270(10), 5331-5338. doi:10.1074/jbc.270.10.5331
Subramanian, A., Kanzaki, L. F., Galloway, J. L., & Schilling, T. F. (2018). Mechanical force regulates tendon
extracellular matrix organization and tenocyte morphogenesis through TGFbeta signaling. eLife,
7, e38069. doi:10.7554/eLife.38069
Svensson, L., Aszódi, A., Reinholt, F. P., Fässler, R., Heinegård, D., & Oldberg, A. (1999). Fibromodulin-null
mice have abnormal collagen fibrils, tissue organization, and altered lumican deposition in tendon.
J Biol Chem, 274(14), 9636-9647. doi:10.1074/jbc.274.14.9636
Taylor, S. H., Yeung, C. Y., Kalson, N. S., Lu, Y., Zigrino, P., Starborg, T., . . . Kadler, K. E. (2015). Matrix
metalloproteinase 14 is required for fibrous tissue expansion. eLife, 4, e09345.
doi:10.7554/eLife.09345
Thorpe, C. T., Udeze, C. P., Birch, H. L., Clegg, P. D., & Screen, H. R. (2012). Specialization of tendon
mechanical properties results from interfascicular differences. J R Soc Interface, 9(76), 3108-3117.
doi:10.1098/rsif.2012.0362
Tidball, J. G., & Lin, C. (1989). Structural changes at the myogenic cell surface during the formation of
myotendinous junctions. Cell Tissue Res, 257(1), 77-84. doi:10.1007/BF00221636
Wall, M. E., & Banes, A. J. (2005). Early responses to mechanical load in tendon: role for calcium signaling,
gap junctions and intercellular communication. J Musculoskelet Neuronal Interact, 5(1), 70-84.
Wall, M. E., Dyment, N. A., Bodle, J., Volmer, J., Loboa, E., Cederlund, A., . . . Banes, A. J. (2016). Cell
Signaling in Tenocytes: Response to Load and Ligands in Health and Disease. Adv Exp Med Biol,
920, 79-95. doi:10.1007/978-3-319-33943-6_7
Wang, J. H., Guo, Q., & Li, B. (2012). Tendon biomechanics and mechanobiology--a minireview of basic
concepts
and
recent
advancements.
J
Hand
Ther,
25(2),
133-140;
quiz
141.
doi:10.1016/j.jht.2011.07.004
Wiesinger, H. P., Kosters, A., Muller, E., & Seynnes, O. R. (2015). Effects of Increased Loading on In Vivo
Tendon Properties: A Systematic Review. Med Sci Sports Exerc, 47(9), 1885-1895.
doi:10.1249/MSS.0000000000000603
Wilkinson, J. M., Davidson, R. K., Swingler, T. E., Jones, E. R., Corps, A. N., Johnston, P., . . . Clark, I. M.
(2012). MMP-14 and MMP-2 are key metalloproteases in Dupuytren's disease fibroblast-
mediated contraction. Biochimica et Biophysica Acta (BBA) - Molecular Basis of Disease, 1822(6),
897-905. doi:https://doi.org/10.1016/j.bbadis.2012.02.001
Zhang, C., Zhu, J., Zhou, Y., Thampatty, B. P., & Wang, J. H. (2019). Tendon Stem/Progenitor Cells and Their
Interactions with Extracellular Matrix and Mechanical Loading. Stem Cells Int, 2019, 3674647.
doi:10.1155/2019/3674647
Zhang, J., & Wang, J. H. (2013). The effects of mechanical loading on tendons--an in vivo and in vitro model
study. PloS one, 8(8), e71740. doi:10.1371/journal.pone.0071740
Figure 1. CaV1.2 is expressed in tendon fibroblasts during embryonic development and early
postnatal growth in mice. (A&B) Representative images of whole-mount X-gal stained forelimbs
from CaV1.2+/LacZ embryos harvested at E13.5 (A) and at E16.5 (B) are shown to illustrate
expression of the transgene in the skeletal elements. (C-G) Fast red counterstaining was performed
on frozen sections of whole-mount X-gal stained CaV1.2+/LacZ forelimb digits at E14.5 (C), Achilles
tendon at P3 (D) and P45 (E), and patellar tendon P3 (F) and P45 (G); boxed regions were obtained
at high power. The nuclear localization of blue color resulting from X-gal staining illustrates the
specific transgene expression in tendon fibroblasts.
Figure 2. Enhanced tenogenesis in ScxCre; CaV1.2TS mice. Gross anatomy was performed on
tendons from 1 month-old Cre-; CaV1.2TS and ScxCre;CaV1.2TS mice, and representative images
of: plantaris tendons (arrows) and Achilles tendons (*) in hindlimbs (A), patellar tendons (B), tail
tendons (C), tendons in forelimbs (D) and back tendons (E) are shown. Note the white tissues are
tendons indicated by * or arrows, and appear larger in ScxCre;CaV1.2TS (TS) vs. Cre-; CaV1.2TS
control (Ctrl) mice.
Figure 3. Increased tendon hypertrophy in ScxCre;CaV1.2TS mice. (A, B, C and D)
Representative images of fast green and hematoxylin stained tendon cross sections from littermate
control (Cre-;CaV1.2TS) and ScxCre; CaV1.2TS mice at 1 month of age. (E) Histomorphometry was
performed on these sections to quantify tendon cross section area (CSA), number of cells in whole
tendon cross section area, and tendon cell density (number of cells/CSA), and the data for each
tendon are presented with the mean for the group ± SD (n >= 3, * p < 0.05, ** p < 0.01, ns, not
significant). Statistical analysis was performed by 2-tailed unpaired t test, and a p value less than
0.05 was considered significant. PT: patellar tendon, PL: plantaris tendon, Ach: Achilles tendon.
Figure 4. Altered collagen fibril size distribution of Achilles tendons in ScxCre;CaV1.2TS mice.
(A and B) Representative transmission electron microscopy (TEM) images of Achilles tendon
collagen fibrils from littermate control (Cre-;CaV1.2TS) and ScxCre;CaV1.2TS mice at 1 month of
age are shown to illustrate the smaller fibrils in ScxCre;CaV1.2TS mice as illustrated by the
difference in size of the largest fibril in each group (yellow highlight). (C and D)
Histomorphometry was performed on these TEM sections to quantify fibril density (number of
fibrils/CSA of fibrils) (C) and the collagen interfibrillar spacing (D); the data are presented for
each tendon with the mean ± SD (n >= 3, **** p < 0.0001). (D) Histograms showing the altered
frequencies of collagen fibril CSAs from ScxCre;CaV1.2TS mice compared to control mice. (E)
Cumulative fraction analysis of collagen fibril CSAs, showing the left-shift of the size of collagen
fibrils in ScxCre;CaV1.2TS mice. Kolmogorov-Smirnov test shows that the tendon fibrils are
significantly smaller than those in control mice, n >= 3, p < 0.001. CSA: cross section area. Ctrl:
Cre-;CaV1.2TS control mice; TS: ScxCre;CaV1.2TS mice.
Figure 5. Specific biomechanical alteration of Achilles tendons in ScxCre;CaV1.2TS mice.
Uniaxial displacement-controlled stretching test was performed on Achilles tendons from control
and ScxCre;CaV1.2TS of both male and female mice at 1 month of age to quantify: peak load (A),
stiffness (B), elastic modulus (D) and peak stress (E) mice. The data are presented for each tendon
with the mean ± SD (n >= 19, ** p < 0.01, ns: not significant via t-test).
Figure 6. Altered proteomics of Achilles tendons in ScxCre;CaV1.2TS mice. Proteomic analysis
was performed on Achilles tendon from control (Cre-;CaV1.2TS) and ScxCre;CaV1.2TS mice at 1
month old. (A) Heatmap of all significantly different proteins between control and
ScxCre;CaV1.2TS mice. (B) Volcano plot showing upregulated (red) and downregulated (blue)
protein expression in ScxCre;CaV1.2TS mice compared to control mice. (C) Heatmap showing the
proteins with largest increase and decrease in expression in ScxCre;CaV1.2TS mice compared to
control mice. (D) The GO terms of upregulated and downregulated differentially expressed
proteins in ScxCre;CaV1.2TS mice with DAVID analysis. (E) Normalized abundance of the proteins
with significant increase and decrease in expression in ScxCre;CaV1.2TS mice compared to control
mice. Values are mean ± SD. * p < 0.05, ** p < 0.01 , *** p < 0.001. n = 4 for each group.
Identification of proteins, > 1.5-fold or <-1.5-fold change in abundance and FDR p-value <0.05
was considered significant.
Figure 7. Gene expression analysis in Achilles/plantaris tendons of control and
ScxCre;CaV1.2TS mice. (A-F) Quantitative analysis of RT-qPCR for Mstn, Tnc, Tnmd, Mmp14,
Scx, and Col1a1 expression. Target gene expression values were normalized to the stable
housekeeping gene Gapdh, and then to relative control expression levels. Values are mean ± SD,
n = 3 for each group. Difference between groups were tested using 2-tailed unpaired t test (* p <
0.05, ** p < 0.01).
Table S1. PCR primer sequences
| 2023 | Increased Ca signaling through Ca1.2 induces tendon hypertrophy with increased collagen fibrillogenesis and biomechanical properties | 10.1101/2023.01.24.525119 | [
"Li Haiyin",
"Korcari Antonion",
"Ciufo David",
"Mendias Christopher L.",
"Rodeo Scott A.",
"Buckley Mark R.",
"Loiselle Alayna E.",
"Pitt Geoffrey S.",
"Cao Chike"
] | null |
1
Locus cœruleus noradrenergic neurons phase-lock to prefrontal cortical and hippocampal infra-slow
rhythms which synchronize with behavioral events
Liyang Xiang1,2, Antoine Harel1, Ralitsa Todorova1, HongYing Gao1, Susan J. Sara1,3 & Sidney I. Wiener1
1Center for Interdisciplinary Research in Biology (CIRB), College de France, CNRS, INSERM, PSL
Research University, Paris, France
2Zhejiang Key Laboratory of Neuroelectronics and Brain Computer Interface Technology, Hangzhou,
China
3Department of Child and Adolescent Psychiatry, New York University Medical School, New York, NY,
USA
Abstract
The locus cœruleus (LC) is the primary source of noradrenergic projections to the forebrain, and, in
prefrontal cortex, is implicated in decision-making and executive function. LC neurons phase-lock to
cortical slow wave oscillations during sleep. Such slow rhythms are rarely reported in awake states, despite
their interest since they correspond to the time scale of behavior. Thus, we investigated LC neuronal
synchrony with infra-slow rhythms in awake rats performing an attentional set-shifting task. Local field
potential (LFPs) oscillation cycles in prefrontal cortex and hippocampus on the order of 0.4 Hz phase-locked
to task events at crucial maze locations. Indeed, successive cycles of the infra-slow rhythm showed different
wavelengths, and thus these are not periodic oscillations. Simultaneously recorded infra-slow rhythms in
prefrontal cortex and hippocampus showed different cycle durations as well. Most LC neurons (including
optogenetically identified noradrenergic neurons) recorded here were phase-locked to these infra-slow
rhythms, as were hippocampal and prefrontal units recorded on the LFP probes. The infra-slow oscillations
also phase-modulated gamma amplitude, linking these rhythms at the time scale of behavior to those
coordinating neuronal synchrony. Noradrenaline, released by LC neurons in concert with the infra-slow
rhythm, would facilitate synchronizing or resetting those brain networks, underlying behavioral adaptation.
Introduction
The brain coordinates activity among interconnected
regions via coherent oscillatory cycles of excitation
and inhibition (Womelsdorf, et al., 2007). This can
facilitate communication among selected subsets of
neurons, groups of neurons, and brain regions.
Sensory stimuli or behavioral events can reset the
phase of these oscillations (Canovier, 2016; Voloh &
Womelsdorf, 2016), linking activity of multiple
neurons to process information in concert. However,
the principal brain rhythms studied in behaving
animals are at the time scale of cell neurophysiological
processes, which are much faster (on the order of tens
and hundreds of milliseconds) than real life behavioral
events, which typically occur at second and supra-
second time scales. The brain has several mechanisms
linking these two time scales, some of which involve
the hippocampus (e.g., Banquet, et al., 2021) and
associated networks, including the prefrontal cortex.
Little is known about brain rhythms that operate in this
crucial behavioral time scale during awake behavior.
The brain is indeed capable of generating rhythms on
2
the order of 0.1-1.0 Hz, although these have been
principally characterized during sleep (Steriade,
1993). Furthermore, during sleep or under anesthesia,
rat noradrenergic locus cœruleus (LC) and prefrontal
cortical (Pfc) neurons are phase-locked to slow
rhythms (Lestienne, et al., 1997; Eschenko, et al.,
2012). LC stimulation exerts powerful influence on
neurophysiological activity in Pfc and hippocampus
(Hip; Berridge and Foote, 1991). LC actions in
prefrontal cortex are implicated in vigilance, decision-
making, and executive function, while in Hip they are
associated with learning and processing contextual
information (e.g., Wagatsuma, et al., 2018; Sara, 2009
for review). Since oscillations can coordinate activity
in brain networks, we reasoned that there might also
be rhythmicity on this behavioral time scale in awake
animals, and investigated this possibility in rats
performing a task engaging Pfc, Hip and LC (Oberto,
et al., 2022; Xiang, et al., 2019). Such coordinated
activity could provide a possible link between
neuromodulation and oscillatory coordination of brain
areas on the time scale of behavior.
Results
LC
neuron
phase-locking
to
prefrontal
and
hippocampal infra-slow rhythms
Rats equipped for chronic recordings alternated
between visual and spatial discrimination tasks in an
automated
T-maze
with
return
arms.
Visual
inspections of hippocampal (Hip) and prefrontal
cortical (Pfc) local field potentials (LFPs) revealed
infra-slow rhythms (Fig. 1A). These were rendered
more salient by filtering the signal in a 0.1-1.0 Hz
window (Fig. 1B). We applied an amplitude threshold
to examine data from those periods when the infra-
slow rhythm amplitude was elevated (Fig. 1C), and
observed that LC neurons were phase-locked to Pfc, as
well as Hip infra-slow LFP rhythms (n=21 out of 37,
and 18 out of 37, respectively; Rayleigh test, p<0.05;
Fig. 1D; for histology, see Fig. 2 of Xiang, et al.,
2019).
The modal preferred infra-slow phase among these
neurons was 0.35*π radians for Hip infra-slow and
0.15*π radians for Pfc infra-slow (Fig. 3; p<0.05,
Rayleigh test). In one animal, noradrenergic LC
neurons
were
identified
optogenetically
(see
Methods), and most were phase locked to the infra-
slow rhythms (n= 8 out of 11 for LFPs in both Pfc and
Hip; Rayleigh test, p<0.05). (Since, apart from this
high
incidence,
the
LC
neurons
identified
optogenetically as noradrenergic had responses
similar to the others, they are all described together.)
Another rat had a high incidence of infra-slow phase-
locking by LC neurons (10/12 and 9/12 respectively
for Hip and Pfc), while in the other two rats with LC
recordings this was rare (1/8 and 1/8; 2/6 and 0/6). The
reason for this variation is not clear, but could be
related to the recordings having sampled different
subpopulations of LC neurons (Chandler, et al., 2014).
Figure 1. Calculation of LC spike phase relative to Hip or
Pfc LFP. A) Unfiltered signal with theta oscillations
dominating. B) The signal from A band-pass filtered at 0.1-
1.0 Hz. Red dots indicate LC neuron action potentials in all
panels. C) The amplitude of the signal in B was z-scored.
Low amplitude oscillations were excluded from analyses
according to an (arbitrary) criterion of z≤0 (excluded zones
are demarcated by the dotted rectangles). D) Phase of the
filtered signal in B. Note that the LC spikes generally occur
at phases between 0 and π/2 radians in this example. The
discontinuities near 138.5 and 143 s correspond to excluded
data, where phase could not be computed reliably.
Prefrontal and hippocampal infra-slow rhythms
synchronized to maze events
The infra-slow rhythms were phase-locked to
positions on the maze (Fig. 2; Supp. Fig. 1B). To
quantify this phase-locking, the mean (± SEM) phase
of the rhythm was plotted in peri-event time color plots
(see Fig. 4 for examples) over all trials in 57 sessions
from eight rats. ‘Regular phase-locking’ describes
those periods when the SEM range was less than
0.75*π radians (see Fig. 4, middle column).
3
Figure 2. A) The automated behavioral task. When the trained rat crosses the central arm photodetector (VC onset PD), this
triggers one of the two cue screens behind the reward arms to be lit in pseudo-random sequence. Crossing the appropriate
reward delivery PD triggers a drop of sweetened water to arrive at the corresponding reward site. Crossing the VC OFF
PD’s on the return arms triggers the lit screen to be turned off. These three photodetector events are used to synchronize
activity in subsequent Figures. B) Distribution of mean phase (left) and p-values of phase-locking (right; Rayleigh test) for
Pfc infra-slow oscillations in pooled data from multiple sessions (top), and in an example session (bottom).
Figure 3. A) Distribution of preferred phases (i.e., phases of
resultant vectors) for all LC neurons with spiking significant
phase-locking to Hip and Pfc infra-slow rhythms. Radius
values are numbers of neuron. B, C) Spike phase-locking to
infra-slow rhythms from two example LC neurons. Radius
values are spike counts. Red arrows represent resultant
vectors.
4
Figure 4. An example of simultaneous recordings of Pfc and Hip LFP infra-slow oscillations phase-locked to principal maze
events, the PD crossings (at time zero). Each row of the color plots corresponds to a single trial and the phase of the infra-
slow LFP is color-coded. Black rings correspond to the PD crossing prior to (left) or after (right) the event at zero for each
plot. Note that the time scales vary among the events in order to display prior and subsequent PD’s. The traces below show
mean (± SEM) phase. In the middle column, the blue vertical bars and blue double-headed arrow illustrate the calculation
of the range of regular phase-locking (defined here as the period with the arbitrary criterion of SEM range<0.75*π radians;
pink double-headed arrows). Here, desynchronization (zones with large SEM ranges) and discontinuities in the mean phase
result from inter-trial variability in speed and distance from the synchronization point. (PD - photodetector crossing). This
is from the same session as the recording in Fig. 3B.
Infra-slow rhythms were phase-locked to the reward
arm photodetector crossing (Rwd) in 51 of the
recording sessions for Hip, and 46 sessions for Pfc (see
Table 1). The other maze events had fewer incidences
of regular phase-locking (Pfc return arm photodetector
crossing, or Rtn: 11; Pfc central arm visual cue onset
PD, or VC: 18; Hip Rtn: 18; Hip VC: 20). The mean
phases at the respective PD crossings (when
SEM≤0.75*π radians there) were 0.70*π and 0.24*π
radians for Pfc and Hip Rtn, 0.25*π and 0.22*π radians
for Pfc and Hip Rwd, and 0.19*π and 0.01*π radians
for Pfc and Hip VC. The root-mean-square differences
between Pfc and Hip mean phase (calculated pairwise
by session) at the respective PD crossings were
0.14*π, 0.13*π and 0.12*π rad. The regular phase-
locking could last from less than one to over 2.5
successive rhythmic cycles (Fig. 3, Supp. Figs. 1 and
2, see Table 1) and could continue from one event to
the next (Fig. 4, Supp. Fig. 1). For PL Rwd and Hip
Rwd, 30 and 37 sessions had durations of regular
phase-locking lasting one or more cycles, respectively.
These permitted quantification of the temporal
duration of the cycles, which ranged from 2.0 to 2.6 s,
the equivalent of 0.4 to 0.5 Hz. In the six cases of Rwd
PD phase-locking which had a second complete cycle,
the mean of the first was 2.3 s, while the second was
lower, 2.0 s (pairwise t-test, p=0.0009, df=5). Thus,
these are not regular periodic oscillations, but rather
are consistent with phase-locking to task events. Pfc
and Hip infra-slow rhythms sometimes resembled one
another (e.g., Fig. 2). To compare them, sessions were
classified as having Pfc and Hip regular phase-locking
in the following ranges of cycles (see Table 1). In 17
of the 57 sessions, these numbers of cycles were
different between Pfc and Hip for VC, Rwd and/or Rtn
(e.g., Supp. Fig. 2). This indicates that it is unlikely
that Pfc and Hip infra-slow rhythms are related by
volume conduction.
The infra-slow rhythms were regularly phase-locked
to two (in 24 sessions), or even all three (in 8 sessions)
5
different task events. Thus, they were not linked to any
specific task-related behavior. To test whether infra-
slow rhythms were triggered by rapid head
movements, regression analysis compared the onset of
regular phase-locking and times of peak acceleration,
or deceleration around the Rwd PD crossings, and
were
not
significant
(R²=0.034,
p=0.49
and
R²=0.0056, p=0.80 respectively; df=15; see Supp. Fig.
3). Additionally, spatial distributions of speed and
acceleration do not resemble the phase maps (Supp.
Fig. 3). Moreover, in Xiang et al. (2019) we showed
that LC neurons fire more during accelerations.
Indeed, the periods with the greatest increase in LC
activity were not those most frequent for the start of
regular phase-locking (i.e., reset) of the infra-slow
rhythm; rather phase-locking occurred most frequent
ly to Rwd PD crossing (see above), where no consist-
Pfc
Rtn
Pfc
Rwd
Pfc
VC
Hip
Rtn
Hip
Rwd
Hip
VC
<1 cycle (n)
4
16
10
13
14
9
1 to 1.49 cycles
(n)
6
20
6
3
28
8
1.5 to 1.99
cycles (n)
1
6
2
2
7
3
2 to 2.49 cycles
(n)
0
4
0
0
1
0
2.5 to 3 cycles
(n)
0
0
0
0
1
0
Mean cycle
period (s)
2.48 2.22 2.05 2.62 2.45 2.32
Mean
frequency (Hz)
0.40 0.45 0.49 0.38 0.41 0.43
Table 1. Characterization of periods in sessions
with regular phase-locking to infra-slow LFP
oscillations. Note that for the six cases of two or
more cycles, only data from the first cycle were
counted for mean cycle period and frequency.
Cycles are only counted in the period from the
previous trial event to the next one, even though
infra-slow rhythms could continue before or after
(cf., Fig. 4, Supp. Fig. 1).
ent accelerations occurred (see Supp. Figs. 3 and 4).
These results indicate it is unlikely that Pfc and Hip
infra-slow rhythms are due to a biomechanical artifact,
for example from locomotion or head rocking.
Coordination of neuronal activity across time scales
In the four sessions where Hip and Pfc neurons could
be discriminated from the LFP electrodes, most were
also modulated by infra-slow rhythms (Pfc LFP
modulated 6/12 Pfc units and 8/11 Hip units; Hip LFP
modulated 8/12 Pfc units and 8/11 Hip units; Rayleigh
test p<0.05). The LC neurons had relatively consistent
phases with respect to the two infra-slow rhythms
(Fig. 3). LC neurons could be phase-locked to
oscillations in the delta frequency range (1-4 Hz) in
Pfc (n=15/37) and Hip (11/37) as well as theta (5-10
Hz; 7/37 and 5/37 respectively) for Pfc and Hip (Supp.
Table 1). While phase-locking of LC neurons to
gamma (40-80 Hz) was rare (n=2 for both structures’
LFPs), the infra-slow rhythm did modulate the
amplitude of their gamma oscillations at 35-45 Hz
(Fig. 5).
Figure 5. Example of infra-slow modulation of gamma
rhythm LFP in Pfc (top) and Hip (bottom).
6
Discussion
LFP oscillation cycles on the order of 0.4 Hz in
prefrontal cortex and hippocampus were phase-locked
to task events at crucial points on the maze. Successive
cycles had different cycle lengths, indicating that these
are not periodic oscillations. Simultaneous recordings
in prefrontal cortex and hippocampus could have
different cycle lengths as well. Over half of the LC
neurons recorded here were phase-locked to these
infra-slow prefrontal cortical and hippocampal LFPs,
including optogenetically identified noradrenergic
neurons. Hippocampal and prefrontal units were also
phase-locked to the infra-slow oscillations.
This is consistent with previous work showing
neuronal activity adapting to the time scale of
behavioral events. For example, in behavioral tasks
with delays, several brain structures show “time cell”
activity: neurons with sequential “tiling” activity
lasting on the order of several seconds. These periods
can expand or contract depending upon the duration of
task-imposed intervals (MacDonald, et al., 2011). We
speculate that this infra-slow rhythm may originate in
the hippocampal-prefrontal system since neuro-
physiological activity there tracks time intervals on the
order of several seconds based upon regularities in
temporal structure of behavioral or environmental
events.
Steriade, et al. (1993) observed infra-slow (0.3-1.0
Hz) rhythms in neocortical activity in anesthetized and
naturally sleeping cats. Eschenko et al (2012) showed
that LC neuronal activity in sleeping rats is
synchronized with the sleep slow wave cycle (1 Hz)
and is out of phase with Pfc neuronal activity.
Similarly, in rats under ketamine anesthesia, there is a
negative correlation between activity of LC NE
neurons and prefrontal neurons, when neuron
activation oscillates at ~1 Hz (Sara and Hervé-
Minvielle 1995; Lestienne, et al. 1997). While these
slow cycles of UP-DOWN state transitions are not
generally observed in awake animals, this does
demonstrate that these structures can coordinate their
activity at this time scale. Thus the LC could also be
associated with the Pfc-Hip in the origin, maintenance
and communication of behaviorally relevant infra-
slow rhythms in the brain. Further work is required to
elucidate the respective roles of these structures in
these processes.
In the awake state, there is evidence for infra-slow
neural processing although this was not observed as
rhythms per se. Molter, et al. (2012) observed a 0.7 Hz
modulation of the power of theta rhythm recorded in
rat Hip. This 0.7 Hz modulated Hip neuronal activity
during sleep, as well as during behavior in a maze, a
running wheel, and an open field. Positions on a
figure-8 maze corresponded to specific phases of this
modulatory rhythm, similar to the infra-slow rhythm
recorded here. (Their filter settings excluded 0.7 Hz
rhythms and thus this could not be directly measured
in that work.) In Molter et al. (2012), the 0.7 Hz
modulation of the power of the theta slow modulation
was locked at π radians to junction points in the maze
(their Figure 7B), where accelerations might be
expected. However, they found no overall correlation
between phase and acceleration.
Villette, et al. (2015) used calcium imaging to observe
CA1 pyramidal cells in head fixed mice moving in the
dark on a non-motorized treadmill. They found that
different neurons fired sequentially in cycles at the
same time scale as the infra-slow oscillations observed
here. Furthermore, the cycles could occur singly, or
consecutively in groups of two or three. The authors
interpreted this as representing an intrinsic metric for
representing distance walked. This resembles time cell
activity (Pastalkova et al., 2008; MacDonald et al.,
2011) evoked above, where the length of the cycle
extends to the time scale of the ongoing task (Kraus,
et al., 2013; Ravassard, et al., 2013). The 2 to 5 s
durations of the cycles in the Villette, et al. (2015)
study may represent a default value since their task had
no temporal structure. This is on the order of the time
scale of the infra-slow rhythm recorded here, and the
variable numbers of cycles they observed might
flexibly adapt to the positions of task-relevant events
to lead to the results found here.
The present observations of phase-locking of LC
neurons to infra-slow rhythms in hippocampus could
ostensibly be due to independent synchrony of the
infra-slow rhythms and the LC neurons to task events.
However, the LC neurons showed phase preferences
in the infra-slow rhythms in data pooled over multiple
task events. We did not observe any simple relation
between infra-slow rhythms and motor events (e.g., as
we showed for LC neurons with acceleration or
deceleration by Xiang, et al., 2019) since regular
phase-locking could start before (Supp. Fig. 1) or after
the same task events in different sessions (not shown)
and continue over periods including a variety of
associated behaviors.
7
The phase-locking of LC neurons to infra-slow
rhythms in Hip and Pfc, as well as to oscillations in the
delta, theta and gamma frequency bands could reveals
coordinated neuronal processing within a unified
temporal framework. The scale of this corresponded to
the temporal and spatial regularities characterizing the
current behavioral patterns. Cross-frequency coupling
could serve as a mechanism to link processing at
different time scales. This could facilitate both
‘Communication through coherence’ (CTC, Bosman
et al., 2012; Fries, 2005) and ‘Binding by synchrony’
(Eckhorn, et al., 1990; Engel, et al., 1999; Buehlmann
and Deco, 2010). Thus, infra-slow rhythms would
serve as a scaffold to link the time scales of dynamics
of neuronal processes to those of behavior and
cognitive processes. Noradrenaline, released by LC
neurons in concert with the infra-slow rhythm, would
participate in synchronizing or resetting those brain
networks underlying behavioral adaptation to these
events (Bouret & Sara, 2005; Sara & Bouret, 2012).
Materials and Methods
All experiments were carried out in accordance with
local (Comité d’éthique en matière d’expérimentation
animale no. 59), institutional (Scientific Committee of
the animal facilities of the Collège de France) and
international (US National Institutes of Health
guidelines; Declaration of Helsinki) standards, legal
regulations
(Certificat
no.
B751756),
and
European/national requirements (European Directive
2010/63/EU; French Ministère de l’Enseignement
Supérieur et de la Recherche 2016061613071167)
regarding the use and care of animals. The data here
are from experiments described in Xiang, et al. (2019)
and further details can be found there.
Animals
Four male Long-Evans rats (Janvier Labs, Le Genest-
Saint Isle France; weight, 280–400 g) were maintained
on a 12 h:12 h light-dark cycle (lights on at 7 A.M.).
The rats were handled on each workday. To motivate
animals for behavioral training on the T maze, food
was restricted to 14 g of rat chow daily (the normal
daily requirement) while water was partially restricted
except for a 10–30 min period daily to maintain body
weight at 85% of normal values according to age. Rats
were rehydrated during weekends.
The automated T maze with return arms
The behavoral task took place in an elevated
automated T-maze (see Fig. 1) consisting of a start
area, a central arm, two reward arms and two return
arms which connected the reward arms to the start
area. Small wells at the end of each reward arm
delivered liquid reward (30 µl of 0.25% saccharin
solution in water) via solenoid valves controlled by a
CED Power1401 system (Cambridge Electronic
Design, Cambridge, UK) with a custom-written script.
Visual cues (VCs) were displayed on video monitors
positioned behind, and parallel to the two reward arms.
The VCs were either lit or dim uniform fields.
Photodetectors detected task events and triggered cues
and rewards via the CED Spike2 script. The sequence
of left/right illumination of screens was programmed
according to a pseudorandom sequence.
Viral vector preparation and injection
The Canine Adenoviral vector (CAV2-PRS-ChR2-
mCherry) was produced at the University of Bristol
using previously described methods (Li, et al., 2016).
This CAV2 viral vector expresses channelrhopsin-2
(ChR2) under the control of PRSx8 (synthetic
dopamine beta-hydroxylase promoter), which restricts
the expression of the transgene to noradrenergic (NA)
neurons (Figure VI.1a in Hwang, et al., 2001; also see
Hickey, et al., 2014) In one rat (R328), 4 months
before the electrode implant surgery, CAV2-PRS-
ChR2-mCherry was injected into the right LC while
the rat was anesthetized with sodium pentobarbital (40
mg⁄kg, with 5 mg sodium pentobarbital as a
supplement every hour) intraperitoneally. The site
corresponding to LC position was marked on the
exposed skull for injection in right LC (AP ~3.9 mm
relative to lambda, ML ~1.2 mm), and a trephine was
made (~2 mm diameter). A micropipette (calibrated in
1 µl intervals, Corning Pyrex) with a tip diameter of
20 µm was connected to a Hamilton syringe, and
backfilled with 1 µl of the diluted viral vectors.
Microinjections of 0.33 µl were made into the LC (AP
-3.8~-4 mm relative to lambda, ML 1.1-1.2 mm, with
a 15° rostral tilt) at three sites dorsoventrally (5.2, 5.5,
5.7 mm below the brain surface). The pipette was left
at each depth for an additional 3-5 min before moving
down to the next site. When the injection was finished,
the trephine hole was covered with sterilized wax and
the scalp was sutured. The rat was observed until
recovery and was then singly housed.
Electrode and optrode implants
Following VC task pre-training, at least one day
before surgery, rats were returned to ad libitum water
and food. General surgical preparation is described in
8
the
previous
section.
Moveable
tungsten
microelectrodes
(insulated
with
epoxylite®,
impedance = 2-4 MΩ, FHC Inc, USA) were used for
LC recordings. A single microelectrode, or two or
three such electrodes glued together was implanted at
AP -3.8-4 mm relative to lambda, and ML 1.1-1.2 mm,
with a 15° rostral tilt. A stainless steel wire (Teflon
coated, diameter=178 µm, A-M systems Inc)
implanted in the midbrain area about 1-2 mm anterior
to the LC electrode tip served as a fixed LC reference
electrode, permitting differential recording. The rat
with the virus injection (R328) was implanted with an
optrode made of a tungsten microelectrode (insulated
with epoxylite, impedance = 2-4 MΩ, FHC Inc, USA)
glued to a 200 µm optic fiber implant with a ferrule
(0.37 numerical aperture, hard polymer clad, silica
core, multimode, Thorlabs), with tip distances 1 mm
apart (the electrode was deeper). The optic fiber
implant and optic fiber cables were constructed at the
NeuroFabLab (CPN, Ste. Anne Hospital, Paris. Two
screws (diameter = 1 mm, Phymep, Paris) with wire
leads were placed in the skull above the cerebellum to
serve as ground. LC electrodes were progressively
lowered under electrophysiological control until
characteristic LC spikes were identified (located ~ 5-6
mm below the cerebellar surface, see Xiang, et al.,
2019 for details). For the virus-injected rat, LC spikes
could also be identified by responses to laser
stimulations
(described
below).
Following
implantation, the microelectrode was fixed to a micro-
drive allowing for adjustments along the dorsal-
ventral axis. The headstage was fixed to the skull with
dental cement, and surrounded by wire mesh for
protection and shielding. After the surgery, animals
were returned to their home cages for at least one-
week recovery with ad libitum water and food and
regular observation.
Electrophysiological recordings
Rats were then returned to dietary restriction. The
movable electrodes were gradually advanced until a
well-discriminated LC unit was encountered and then
all channels were recorded simultaneously while the
rat performed in the T maze. If no cells could be
discriminated, the electrodes were advanced and there
was at least a 2 h delay before the next recording
session.
For daily online monitoring of LC spikes, pre-
amplified signals were filtered between 300-3000 Hz
for verification on the computer screen (Lynx-8,
Neuralynx, Bozeman, MT, USA) and also transmitted
to an audio monitor (audio analyzer, FHC). For
recordings, brain signals were pre-amplified at unity
gain (Preamp32, Noted Bt, Pecs, Hungary) and then
led through a flexible cable to amplifiers (x500, Lynx-
8, Neuralynx) and filters (0.1-9 kHz, Lynx-8,
Neuralynx). Brain signals were digitized at ~20 kHz
using CED Power1401 converter and Spike2 data
acquisition software. The LC unit activity was
identified by: 1) spike waveform durations ≥0.6 ms;
2) low average firing rate (1-2 Hz) during quiet
immobility; 3) brief responses to unexpected acoustic
stimuli followed by prolonged (around 1 s) inhibition;
4) for the virus-injected rat (R328), LC units were
verified by responses to laser stimulation. A laser
driver (Laserglow Technologies, Canada, wavelength
473 nm) was controlled by signals from a stimulator
(Grass Technologies, USA, Model SD9). Light
intensity from the tip of optic fiber was measured by a
power meter (Thorlabs, Germany, Model PM100D). If
unit firing was entrained to the pulses with an
increased rate (to at least twice the baseline firing rate)
averaged over all the stimulations, they were
considered to be noradrenergic LC units.
A light emitting diode (LED) was mounted on the
cable that was plugged into the headstage. This was
detected by a video camera mounted above the T-maze
and transmitted to the data acquisition system at a
sampling rate of ~30 Hz for the purpose of position
tracking.
Tissue processing
After all recording experiments, electrolytic lesions
(40 µA, 10 s cathodal current) were made at the tip of
the electrodes. Brain slices were cut coronally at a
thickness of 40 µm with a freezing microtome and
were collected in cold 0.1 M PB for Nissl staining.
Recordings at sites with reconstructed electrode
positions outside LC proper were excluded from
analysis. For fluorescent immunohistochemistry,
sections were then incubated in primary antibodies
overnight at 4°C in darkness with both chicken anti-
tyrosine hydroxylase (TH) antibody (1:500, Abcam)
and mouse anti-mCherry antibody (1:200, Ozyme) in
PBS containing 0.1% Triton X-100 and 3% NGS.
After three 5 min rinses in PBS, sections were then
incubated with secondary antibodies in PBS
containing 3% NGS for 1h at RT in darkness.
Secondary antibodies used in this study were Alexa
Fluor 488 goat anti-chicken IgG (1:3000, Life
9
Technologies) and Alexa Fluor 546 goat anti-mouse
IgG (1:3000, Life Technologies).
Signal processing, spike sorting and data analyses
For off-line spike detection of LC activity in three of
the rats, the wide-band signals were converted and
digitally high-pass filtered (nonlinear median-based
filter). Waveforms with amplitudes passing a
threshold were extracted, and then subjected to
principal component analysis (PCA). All of these
processes were performed with NDManager (Hazan,
et al., 2006). Spikes were sorted with a semi-automatic
cluster cutting procedure combining KlustaKwik (KD
Harris,
http://klustakwik.sourceforge.net)
and
Klusters (Hazan, et al., 2006). Spikes with durations
less than 0.6 ms were rejected. In one rat (R311) the
LC signal was filtered from 300-3000 Hz during
recording, and the spike sorting was performed with
Spike2 software (which employs a waveform template
matching algorithm). Most data analyses were
performed using Matlab (R2010a) with the statistical
toolbox FMAToolbox (developed by M. Zugaro,
http://fmatoolbox.sourceforge.net)
and
scripts
developed in the laboratory as well as some statistical
analyses performed with Microsoft© Excel©. The
latter application’s calculated the regressions (passing
through the origin) between the onset of regular phase-
locking and times of peak acceleration or deceleration,
and
p-values
were
taken
from
https://www.socscistatistics.com. This analysis was
performed only for Reward Arm PD synchronized
data when the prior Central Arm PD mean phase data
had no regular phase-locking (to avoid confounds and
have sufficiently large data set).
Acknowledgements
Thanks to Professor Anthony E. Pickering for
providing the virus and related advice. Thanks to Dr.
Michaël Zugaro for helpful suggestions and help with
analyses and computing, Drs. A Sirota and X
Leinkugel for helpful discussions, and France
Maloumian for help with figures. L.X. was supported
by a fellowship from the China Scholarship Council
(CSC).
The
Labex
Memolife
and
Fondation
Bettencourt Schueller provided support.
Competing interests
The authors declare that they have no competing
interests.
References
Berridge CW, Foote SL. (1991) Effects of locus
coeruleus activation on electroencephalographic
activity in neocortex and hippocampus. J Neurosci.
11(10):3135-45. doi: 10.1523/JNEUROSCI.11-10-
03135.1991
Bosman CA, Schoffelen J-M, Brunet N, Oostenveld
R, Bastos AM, Womelsdorf T, Rubehn B, Stieglitz T,
De Weerd P, Fries P. (2012) Attentional stimulus
selection through selective synchronization between
monkey visual areas. Neuron. 75:875–888. doi:
10.1016/j.neuron.2012.06.037.
Bouret S, Sara SJ. (2005) Network reset: a simplified
overarching theory of locus coeruleus noradrenaline
function. Trends Neurosci. 28(11):574-82. doi:
10.1016/j.tins.2005.09.002
Buehlmann A, Deco G. (2010) Optimal information
transfer in the cortex through synchronization. PLoS
Comput Biol 6. doi: 10.1371/journal.pcbi.1000934
Canavier CC. (2015) Phase-resetting as a tool of
information
transmission. Curr
Opin
Neurobiol.
31:206-213. doi:10.1016/j.conb.2014.12.003
Chandler DJ, Gao WJ, Waterhouse BD. (2014)
Heterogeneous organization of the locus coeruleus
projections to prefrontal and motor cortices. Proc Natl
Acad Sci (U. S. A.) 111(18):6816-21. doi:
10.1073/pnas.1320827111.
Eckhorn R, Reitboeck HJ, Arndt M, Dicke P (1990)
Feature linking via synchronization among distributed
assemblies: Simulations of results from cat visual
cortex.
Neural
Comput
2:293–307.
https://doi.org/10.1162/neco.1990.2.3.293
Engel AK, Fries P, König P, Brecht M, Singer W
(1999) Temporal binding, binocular rivalry, and
consciousness. Conscious Cogn 8:128–51.
Eschenko, O, Magri C, Panzeri S, Sara, SJ (2012)
Noradrenergic neurons of the cocus coeruleus are
phase-locked to cortical Up-Down States during sleep.
Cerebral
Cortex
22(2):426-43.
doi:10.1093/cercor/bhr121
Fries, P. A. (2005) A mechanism for cognitive
dynamics: neuronal communication through neuronal
coherence.
Trends
Cogn
Sci
9:474–80.
doi:
10.1016/j.tics.2005.08.011.
Hazan L, Zugaro M, Buzsáki G (2006) Klusters,
NeuroScope, NDManager: A free software suite for
10
neurophysiological data processing and visualization.
J.
Neurosci.
Meth.
155,
207–216.
doi:
10.1016/j.jneumeth.2006.01.017.
Hickey L, Li Y, Fyson SJ, Watson TC, Perrins R,
Hewinson J, Teschemacher AG, Furue H, Lumb BM,
Pickering AE (2014) Optoactivation of locus ceruleus
neurons evokes bidirectional changes in thermal
nociception in rats. J Neurosci 34: 4148–4160.
DOI: 10.1523/JNEUROSCI.4835-13.2014
Hwang DY, Carlezon WA, Isacson O, Kim KS (2001)
A high-efficiency synthetic promoter that drives
transgene expression selectively in noradrenergic
neurons. Human Gene Ther 12: 1731–1740. doi:
10.1089/104303401750476230.
Kraus BJ, Robinson RJ, 2nd, White JA, Eichenbaum
H, Hasselmo ME (2013). Hippocampal “time cells”:
time versus path integration. Neuron 78: 1090–1101.
doi: 10.1016/j.neuron.2013.04.015.
Lestienne R, Hervé A, Robinson D, Brios L, Sara
SJ (1997) Slow oscillations as a probe of the dynamics
of the locus coeruleus-frontal cortex interaction in
anesthetized rats. J. Physiology (Paris) 91, 273-284.
doi: 10.1016/s0928-4257(97)82407-2.
Li, Y. Hickey L, Perrins R, Werlen E, Patel AA,
Hirschberg S, Jones MW, Salinas S, Kremer EJ,
Pickering
AE
(2016)
Retrograde
optogenetic
characterization of the pontospinal module of the
locus coeruleus with a canine adenoviral vector. Brain
Res. 1641(Pt. B): 274–290. doi:
10.1523/JNEUROSCI.4835-13.2014
MacDonald, C., Lepage, K., Eden, U., Eichenbaum,
H. (2011) Hippocampal “time cells” bridge the gap in
memory for discontiguous events. Neuron 71(4):737–
749. doi: 10.1016/j.neuron.2011.07.012.
Molter C, O’Neill J, Yamaguchi Y, Hirase H,
Leinekugel X (2012) Rhythmic modulation of theta
oscillations supports encoding of spatial and
behavioral information in the rat hippocampus.
Neuron
75:889–
903. doi: 10.1016/j.neuron.2012.06.036
Oberto VJ, Boucly CJ, Gao H, Todorova R, Zugaro
MB, Wiener SI. (2022) Distributed cell assemblies
spanning prefrontal cortex and striatum. Curr Biol.
32(1):1-13.e6. doi: 10.1016/j.cub.2021.10.007
Pastalkova E, Itskov V, Amarasingham A, and
Buzsáki G (2008). Internally generated cell assembly
sequences
in
the
rat
hippocampus.
Science
321(5894):1322-7. doi: 10.1126/science.1159775.
Ravassard P, Kees A, Willers, B, Ho D, Aharoni D,
Cushman J, Aghajan ZM, Mehta MR (2013)
Multisensory control of hippocampal spatiotemporal
selectivity.
Science
340:1342–1346.
doi: 10.1126/science.1232655
Sara SJ (2009) The locus coeruleus and noradrenergic
modulation of cognition. Nat Rev Neurosci.
10(3):211-23. doi: 10.1038/nrn2573.
Sara SJ, Bouret S. (2012) Orienting and reorienting:
the locus coeruleus mediates cognition through
arousal. Neuron 76(1):130-41. doi:
10.1016/j.neuron.2012.09.011.
Sara SJ, Hervé-Minvielle A (1995) Inhibitory
influence of frontal cortex on locus coeruleus neurons
PNAS. 92:6032-6036. doi: 10.1073/pnas.92.13.6032.
Steriade M, Contreras D, Curro Dossi R, Nunez A
(1993) The slow (<1 Hz) oscillation in reticular
thalamic and thalamocortical neurons: scenario of
sleep rhythm generation in interacting thalamic and
neocortical networks. J Neurosci. 13:3284-3299. doi:
10.1523/JNEUROSCI.13-08-03284.1993.
Villette
V, Malvache
A, Tressard
T, Dupuy
N, Cossart R (2015) Internally recurring hippocampal
sequences as a population template of spatiotemporal
information.
Neuron. 88(2):357-66.
doi: 10.1016/j.neuron.2015.09.052
Voloh B, Womelsdorf T (2016) A role of phase-
resetting in coordinating large scale neural networks
during attention and goal-directed behavior. Front
Syst Neurosci 10:18. doi: 10.3389/fnsys.2016.00018
Wagatsuma A, Okuyama T, Sun C, Smith LM, Abe K,
Tonegawa S. (2018) Locus coeruleus input to
hippocampal CA3 drives single-trial learning of a
novel
context.
Proc
Natl
Acad
Sci
(USA)
115(2):E310-E316. doi: 10.1073/pnas.1714082115.
Womelsdorf T, Schoffelen JM, Oostenveld R, Singe,
W, Desimone R, Engel AK, Fries P. (2007)
Modulation of neuronal interactions through neuronal
synchronization. Science 316:
1609–1612.
doi: 10.1126/science.1139597
11
Xiang L, Harel A, Gao H, Pickering AE, Sara SJ,
Wiener SI. (2019) Behavioral correlates of activity of
optogenetically
identified
locus
cœruleus
noradrenergic neurons in rats performing T-maze
tasks. Sci Rep. 9(1):1361. doi: 10.1038/s41598-018-
37227-w.
12
Supplementary Figure 1. The infra-slow cycles could continue through sequential phases from one task
event to the following one. Plots are in same format as in Figure 4. A) Note that regular phase-locking
in the left column (VC) starts prior to zero just after the Rtn PD (black dots prior to zero) and continues
up to Rwd PD (black dots after zero). The other columns are comparable. B) Mean infra-slow phase
distribution for this session (as in Fig. 2) also shows the continuity of the phase distribution on maze.
Arrows show photodetector positions. (Note that colors indicating mean phase are different at the
respective photodetector types). (Note that the color codes in the scales of A and B are not the same.)
13
Supplementary Figure 2. Infra-slow rhythms recorded simultaneously in Hip and Pfc can have
different cycle lengths. Note that regular phase-locking to Rwd extends for about 2.5 infra-
slow cycles in Pfc, but only about 1 cycle for Hip. The durations of the cycles are about 2.1 s
(corresponding to 0.48 Hz) for the first Pfc Rwd cycle, 1.9 s for the second (0.53 Hz), and 3.1
s for Hip Rwd (0.32 Hz). (Same format as Figure 4.)
14
Supplementary Figure 3. Example of lack of a clear relation between speed, acceleration and
infra-slow phase. Acceleration increases before central and return arm PD crossings,with
speed increasing after. But, the phase is π radians for the former and -0.2*π radians for the
latter.
15
Supplementary Figure 4. Top) Example of LC neuron activity in relation to task events
for the cell recorded in the session of Figure 4 (reproduced here). Red dots to the left of
zero correspond to the previous task event, and those to the right indicate the timing of
the subsequent event. Note that the regular phase-locking of infra-slow oscillations at
the Reward arm PD continues in the interval (0 s, 2 s), while the LC neuron is inactive.
16
# phase-
locked
neurons
resultant
vector
length (%)
Hip Slow
21
15.2±1.4
Pfc Slow
18
15.0±1.5
Hip Delta
11
8.2±0.8
Pfc Delta
15
9.7±0.7
Hip Theta
5
5.6±0.5
Pfc Theta
7
8.1±1.0
Hip
Gamma
2
6.9±0.3
Pfc
Gamma
2
7.5±1.0
Supplementary Table 1. Tallies of LC neurons phase-locked to LFP oscillations (Rayleigh test,
p<0.05) in several frequency bands. The total population was 37 neurons. Resultant vector length
were calculated only from neurons with significant phase-locking. From top to bottom, the bands
correspond to 0.1-1 Hz, 1-4 Hz, 5-10 Hz, and 40-80 Hz.
| 2022 | Locus cœruleus noradrenergic neurons phase-lock to prefrontal cortical and hippocampal infra-slow rhythms which synchronize with behavioral events | 10.1101/2022.05.12.491630 | [
"Xiang Liyang",
"Harel Antoine",
"Todorova Ralitsa",
"Gao HongYing",
"Sara Susan J.",
"Wiener Sidney I."
] | null |
Brain Capillary Pericytes are Metabolic Sentinels that
Control Blood Flow through KATP Channel Activity
Ashwini Hariharan1, Colin D. Robertson2, Daniela C.G. Garcia1 & Thomas A. Longden1,#
1Department of Physiology, School of Medicine, University of Maryland, Baltimore, MD, USA
2Department of Pharmacology, School of Medicine, University of Maryland, Baltimore, MD, USA
#Correspondence: Thomas A. Longden, Department of Physiology, School of Medicine, University of Maryland, 655
West Redwood Street, 505 Howard Hall, Baltimore MD 21201 USA
Telephone: (410) 706-1956
Email: thomas.longden@som.umaryland.edu
Website: www.longdenlab.org
Keywords:
Pericytes, endothelial cells, capillaries, neurovascular coupling, functional
hyperemia, KATP channels, KIR channels, cerebral blood flow, glucose, energy,
metabolism
SUMMARY
Capillary pericytes and their processes cover ~90% of the total length of the brains
capillary bed. Despite their abundance, little is known of pericyte function, and their
contributions to the control of brain hemodynamics remain unclear. Here, we report that
deep capillary pericytes possess a mechanistic ‘energy switch’ that, when activated by a
decrease in glucose, elicits robust KATP channel activation to increase blood flow and
protect energy substrate availability. We demonstrate that pharmacological activation of
KATP channels profoundly hyperpolarizes capillary pericytes and leads to dilation of
upstream penetrating arterioles and arteriole-proximate capillaries covered with
contractile pericytes, leading to an increase in local capillary blood flow. Stimulation of a
single capillary pericyte with a KATP channel agonist is sufficient to evoke this response,
which is mediated via KIR channel-dependent retrograde propagation of hyperpolarizing
electrical signals. Genetic inactivation of pericyte KATP channels via expression of a
dominant-negative version of KIR6.1 eliminates these effects. Critically, we show that
lowering extracellular glucose below 1 mM evokes dramatic KATP channel-mediated
pericyte hyperpolarization. Inhibiting glucose uptake by blocking GLUT1 transporters in
vivo also activates this energy switch to increase pericyte KATP channel activity, dilate
arterioles and increase blood flow. Together, our findings recast capillary pericytes as
metabolic sentinels that respond to local energy deficits by robustly increasing blood flow
to protect metabolic substrate delivery to neurons and prevent energetic shortfalls.
Pericyte KATP channels tune blood flow to local metabolism
2
INTRODUCTION
Blood flow provides the oxygen and glucose that are critical for the metabolic processes that
underpin brain function and health. Thus, precise control of cerebral hemodynamics is essential
to meet the moment-to-moment energetic needs of neurons and glia. The brains vascular system
is fed by pial arteries, which originate at the circle of Willis and course over the surface of the
brain before branching orthogonally to give way to penetrating arterioles (PAs) that dive into the
parenchyma1. PAs, in turn, branch to give rise to a tortuous capillary network that is covered by a
diverse population of pericytes. At the first point of the PA-to-capillary transition, mural cells
termed ‘pre-capillary sphincters’ are found which exert dynamic control of blood flow into the
capillary bed by virtue of their α-smooth muscle actin (SMA) expression2. Adjacent to this, the
initial 3-4 branches of the capillary network are collectively referred to as a ‘transitional segment’3
due to their coverage by contractile pericytes that express α-SMA and these cells are capable of
rapidly regulating the diameter and therefore blood flow of the underlying vessel4–6. Immediately
downstream of the α-SMA terminus are mesh pericytes, and deeper in the capillary bed (from
approximately the 5th branch and above), the abluminal surface of the capillaries is adorned by
the processes and cell bodies of thin-strand pericytes4,7. The latter extend long, narrow processes
which stretch in some cases for hundreds of microns along the walls of local capillaries, coming
into close apposition with the arborizations of neighboring thin-strand pericytes8. They also form
‘peg-socket’ junctions with the underlying ECs, which are thought to be the sites of gap junction
coupling, permitting the ready exchange of molecules and charge between these cells9–11.
Pericytes contribute to multiple physiological processes including regulation of blood brain barrier
permeability and modulation of endothelial cell (EC) gene expression12,13. As they are ideally
positioned to mediate communication between the blood and brain parenchyma, it has been
suggested that these cells play a critical role in control of hemodynamics. A growing body of
evidence indicates that contractile pericytes of the transitional segment play a key role in rapidly
regulating the diameter of the underlying capillaries2,5,14–18. However, the mechanisms for blood
flow control by thin-strand pericytes have not been defined. Emerging evidence suggests that
subtle contractile processes in these cells may regulate capillary diameter and therefore local
capillary blood flow6. Here, we show that the electrical activity of thin-strand pericytes alone is
sufficient for robust, remote blood flow control by these cells via communication with the
underlying endothelium. We recently surveyed the molecular expression of ion channels and G
protein-coupled receptors (GPCRs) in thin-strand pericytes of the brain19. The vascular form of
the ATP-sensitive potassium (K+; KATP) channel, composed of inward rectifier K+ (KIR) 6.1 and
sulfonylurea receptor (SUR) 2 subunits is the most highly expressed ion channel subtype in these
cells and accounts for almost half of their relative expression of all ion channel genes20,21. As KATP
channels are found in a range of tissues where they play a major role in coupling metabolism to
membrane electrical activity22,23, we hypothesized that they may play a similar role in brain
pericytes, linking local metabolic substrate availability to membrane hyperpolarization and,
ultimately, blood flow control.
We demonstrate here that deep capillary pericytes control local blood flow via KATP channel-
mediated electrical signaling. Our results indicate that pericyte KATP channels are the molecular
cornerstone of an ‘energy switch’ mechanism, wherein a fall in glucose availability below a key
Pericyte KATP channels tune blood flow to local metabolism
3
threshold evokes a KATP channel-mediated blood flow increase to replenish energy substrate
delivery to neurons and glia. Our data thus recast thin-strand pericytes as metabolic sentinels that
dynamically modulate blood flow to ensure that the energy substrates required to support ongoing
neuronal function are continually provided.
RESULTS
KATP channel activation increases arteriolar diameter and capillary blood flow in vivo
To examine the role of KATP channels in the control of brain blood flow, we began by visualizing
the vascular network of a volume of cortex through a cranial window preparation in mice
anesthetized with urethane and alpha-chloralose (Fig. 1A). We identified pial arteries on the
brains surface and their arising PAs branching perpendicularly into the tissue by their morphology
in relation to nearby veins, and imaged these PAs and their daughter capillaries down to at least
the 5th branch of the capillary bed (Fig. 1A). In vivo, these arteries constrict partially in response
to intravascular pressure24, establishing a baseline of myogenic tone from which diameter can be
bidirectionally modulated to adjust blood flow. Under our conditions, we found that PAs had 46.36
± 2.84% tone (n = 8 arterioles from 5 mice), calculated by comparing baseline arteriolar diameter
to passive diameter in the absence of extracellular calcium (Ca2+) and the presence of the voltage-
dependent Ca2+ channel blocker diltiazem (200 µM) (Supplementary Fig. 1A). We also
measured the tone of the 1st to 4th order branches of the capillaries of the transitional segment
under the same conditions, covered with the cell bodies and processes of contractile pericytes.
These branches collectively averaged 39.35 ± 2.18 % tone at baseline, which was not different to
the tone of PAs and did not differ by branch order (n = 35 capillaries from 5 mice, Supplementary
Fig. 1A). To examine the influence of KATP channels on capillary and arteriole diameter and blood
flow, we assessed the effects of pharmacologically modulating these channels through
superfusion of agents over the cranial surface. Strikingly, we found that application of 10 µM
pinacidil, a selective KATP channel opener, produced near-maximal dilation of both PAs and
transitional segment capillaries (Fig. 1B-H), indicating that KATP channels can exert a strong
influence on the vasculature. In turn, these substantial increases in diameter translated into
profound elevation of capillary blood flow, measured as red blood cell (RBC) flux using a high-
frequency line scanning approach (Fig. 1I, J). To determine whether KATP channel activity
contributes to basal blood flow, we applied the KATP channel blocker glibenclamide (10 µM) to the
cranial surface. We observed no change in the diameter of the PA or 1st – 4th order capillaries to
this maneuver, and no change in capillary blood flow (Supplementary Fig. 2), suggesting that
vascular KATP channel activity in the brain is minimal under our resting conditions.
Pericyte KATP channels tune blood flow to local metabolism
4
Figure 1. KATP channel activation increases arteriole diameter and capillary blood flow in vivo. (A) In vivo imaging set-up. Left: A cranial
window was made over the somatosensory cortex and imaged using two-photon laser scanning microscopy. Right: Imaging field of the
cortical vasculature containing FITC-dextran showing a pial vein and artery, a penetrating parenchymal arteriole (PA) and downstream
capillaries. (B) A PA with a pre-capillary sphincter and its downstream 1st and 2nd order capillaries. Top: Baseline diameter of the PA
(red line). Bottom: Dilation of the PA after application of 10 μM pinacidil (red line: baseline diameter). The capillaries in view also dilated
to this maneuver. (C) A PA and its downstream 1st-4th order capillaries. Left: Baseline diameters of 1st-4th order capillaries indicated by
respective colored lines. Right: The same 1st-4th order capillaries, which dilated after application of 10 μM pinacidil. (D-H) Summary
data, analyzed using paired Student's t-test, showing application of 10 μM pinacidil produced a significant dilation of the (D) PA (n = 7
vessels, 7 mice, *P = 0.01, t6 = 3.697), (E) 1st order capillary (n = 9 vessels, 7 mice, *P = 0.017, t8 = 2.987), (F) 2nd order capillary (n =
11 vessels, 6 mice, **P = 0.0029, t10 = 3.921), (G) 3rd order capillary (n = 11 vessels, 6 mice, **P = 0.0011, t10 = 4.513) and (H) 4th order
capillary (n = 4 vessels, 4 mice, *P = 0.0326, t3 = 3.772). (I) Line scanning strategy used to measure blood flow in higher order capillaries.
Top-right: Kymograph taken at baseline displaying RBCs passing through the line-scanned capillary as dark shadows against the green
fluorescence of FITC-containing plasma. Bottom-right: Kymograph of the same capillary post-pinacidil application showing a dramatic
increase in RBC flux. (J) Summary of RBC flux responses showing significant hyperemia to 10 μM pinacidil (n = 6 vessels, 3 mice, *P
= 0.0188, t5 = 3.424, paired Student's t-test).
Pericyte KATP channels tune blood flow to local metabolism
5
Pericytes transmit KATP channel-mediated electrical signals via the endothelium to exert
remote control over the diameter of upstream arterioles
The expression of KATP channels is relatively lower in SMCs and in arteriolar and capillary
endothelial cells (ECs) of the brain compared to thin-strand pericytes19–21,25, and pharmacological
maneuvers designed to activate these channels in isolated PAs do not lead to dilation26. Given
that systems-level KATP channel activation evoked profound vasodilation and blood flow
increases, we reasoned that the high expression of KATP channels in deep capillary thin-strand
pericytes could be the primary source of hyperpolarizing signals that may then be relayed to
upstream PAs to drive their vasodilation. To test this possibility, we maneuvered a pipette
connected to a pressure-ejection system into the brain and positioned it next to a DsRed-positive
thin-strand pericyte in Cspg4-DsRed mice (Fig. 2A-C). On average, targeted pericytes were 268.5
± 25.9 µm from the upstream arteriole imaging site (n = 8 experiments, 8 mice). Consistent with
our hypothesis, activation of KATP channels in a single pericyte by local pressure ejection of 10
µM pinacidil onto the pericyte cell body (Fig. 2C) evoked a rapid and substantial upstream
arteriolar dilation (Fig. 2D,E,G and Supplementary Movie 1) which was accompanied by an
increase in underlying capillary blood flow (Fig. 2F,H).
Pericytes are intricately associated with adjacent ECs via peg-socket processes which are
thought to be the sites of gap junction coupling between these two cell types11,19,27. Accordingly,
we reasoned that signals originating in pericytes may be transmitted upstream via connected
underlying ECs. We previously identified an EC-mediated regenerative electrical signaling
mechanism dependent on inward-rectifier K+ (KIR2.1) channels that transmits dilatory signals from
deep within the capillary bed to upstream PAs28. Interestingly, blocking KIR2.1 channels by the
application of 100 µM barium (Ba2+) to the cortical surface prior to pinacidil ejection on a pericyte
abolished this increase in arteriolar diameter and capillary blood flow (Fig. 2E,I,J), suggesting
that KATP channel-initiated hyperpolarization modulates electrical signaling through the capillary
bed to produce its effects. Consistent with pericytes being the locus of pinacidil-evoked
vasodilatory drive, the ejection of this agent onto a segment of capillary lacking a pericyte soma
had no effect on arteriolar diameter or local blood flow (Fig. 2 E,K,L). As expected, diameter and
blood flow were also unchanged when pericytes were stimulated with vehicle (artificial
cerebrospinal fluid (aCSF) containing 0.3 mg/mL TRITC-dextran) (Supplementary Fig. 3).
Importantly, direct stimulation of the PA with 10 µM pinacidil also had no effect on diameter
(Supplementary Fig. 4), which aligns with previous observations of a lack of response of these
arterioles to KATP agonists26 and buttresses the conclusion that pericyte KATP channels exert
remote control of upstream PA diameter.
Pericyte KATP channels tune blood flow to local metabolism
6
Figure 2. Capillary pericytes exert remote control over upstream PA diameter. (A) Cartoon illustrating the experimental strategy,
showing an ejection pipette positioned next to a pericyte. (B) Z-projections of 3D volume acquisitions outlining the experimental
strategy. Left: Vasculature containing FITC-dextran and a pipette with TRITC-dextran positioned within the cortex. Right: A PA
and its downstream capillary network showing an ejection pipette containing TRITC-dextran with 10 μM pinacidil positioned next
to a DsRed+ pericyte on an 8th order capillary. (C) Depiction of the evolution (left to right) of TRITC diffusion (red) after pressure
ejection of 10 μM pinacidil onto a DsRed+ pericyte. The brevity and low pressure of the ejection conditions (10 psi, 30 ms) ensured
that the drug remained local. (D) Focal stimulation of capillary pericytes with 10 μM pinacidil dilates the connected upstream PA.
Left: PA and 1st order capillary diameter at baseline indicated by magenta lines. Right: Peak dilation of the same PA and 1st order
capillary after pinacidil-ejection on the downstream pericyte. (E) Representative time courses showing PA dilation to direct
stimulation of a pericyte with pinacidil (orange, top), but no change in PA diameter when pinacidil was applied in the presence of
the KIR channel blocker Ba2+ (purple, middle) or when pinacidil was ejected onto a segment of capillary without a pericyte cell body
(blue, bottom). (F) 1-s kymograph segments showing raw RBC flux of a >5th order capillary at baseline, and hyperemia after
pinacidil was ejected onto the overlying pericyte. (G) Summary of PA diameter changes after pinacidil-ejection on a downstream
pericyte (n = 14 paired measurements, 13 mice, ***P = 0.0007, t13 = 4.439, paired Student's t-test). (H) Summary capillary RBC
flux responses to pinacidil applied directly to a pericyte (n = 8 paired measurements, 4 mice, **P = 0.0045, t7 = 4.108, paired
Student's t-test). (I) Summary data showing PA diameter after pinacidil-stimulation of a pericyte in the presence of Ba2+ (n = 6
paired measurements, 6 mice, P = 0.6981, t5 = 0.411, paired Student's t-test). (J) Summary blood flow data showing RBC flux
before and after pinacidil-stimulation of a pericyte in the presence of Ba2+ (n = 5 paired measurements, 5 mice, P = 0.4613, t4 =
0.814, paired Student's t-test). (K) Summary data showing PA diameter changes on stimulation of a capillary segment without a
pericyte cell body with pinacidil (n = 6 paired measurements, 6 mice, P = 0.2162, t5 = 1.415, paired Student's t-test). (L) Summary
of RBC flux responses before and after stimulation of a capillary segment without pericytes with pinacidil (n = 5 paired
measurements, 5 mice, P = 0.394, t4 = 0.9543, paired Student's t-test).
Pericyte KATP channels tune blood flow to local metabolism
7
Expression of a dominant-negative mutant of the vascular KATP channel eliminates
pericyte-mediated dilations and hyperemia
To unequivocally confirm the central role of pericyte KATP channels in control of blood flow and
upstream PA diameter to pinacidil, we deployed mice that express a dominant-negative form of
the KIR6.1 subunit in which a Gly-Phe-Gly motif of the K+ selectivity filter is mutated to a non-
functional alanine triplet (KIR6.1AAA), which in turn eliminates KATP currents29,30. Expression of
KIR6.1AAA was controlled by tamoxifen-inducible Cre-recombinase under the Cspg4 promoter to
selectively suppress KATP channel activity in pericytes and SMCs. In this line, a floxed region
containing the sequence for enhanced green fluorescent protein (eGFP) upstream of a stop codon
is expressed under basal conditions, precluding expression of the downstream KIR6.1AAA
sequence without Cre-recombinase activity. When recombination is induced, eGFP along with
the stop codon are excised, permitting KIR6.1AAA expression (Fig. 3A). Accordingly, induction of
Cre activity in Cspg4-Cre-KIR6.1AAA mice by 4-hydroxy tamoxifen (4-OHT) eliminated eGFP
expression in capillary pericytes, while eGFP expression was retained in adjacent ECs (Fig.
3B,C), indicating successful cell type-selective expression of the KIR6.1AAA construct. To then
reveal pericytes with inactive KATP channels, we applied NeuroTrace 500/525 (NT500/525)31, to
the cranial surface which specifically stained thin-strand pericytes (Fig. 3C). Pressure-ejecting
pinacidil onto thus identified eGFP-negative, NT500/525-positive pericytes did not produce an
increase in upstream PA dilation or local capillary blood flow in Cspg4-Cre-KIR6.1AAA mice (Fig.
3C,D,I,J), indicating that functional KATP channels in pericytes are essential for these responses.
However, Cre control (KIR6.1AAA mice given 4-OHT) and vehicle control (Cspg4-Cre-KIR6.1AAA
mice given a 90:10% mixture of corn oil:ethanol) groups still demonstrated significant PA dilation
(11-13%) and capillary RBC flux still increased (32-38%) to these maneuvers (Fig. 3D,E-H). Thus,
pericytes are the primary site of KATP-mediated upstream arteriolar dilation and local capillary
hyperemia in vivo.
An energy-sensing switch couples decreases in local energy substrate availability to
membrane hyperpolarization via KATP channel activity
Having established that pericyte KATP channels can exert a profound influence over PA diameter
and local blood flow, we next turned our attention to the mechanisms through which KATP channels
may be engaged. In other tissues, KATP channels play a critical role in coupling metabolism to
membrane electrical activity, and are sensitive to the local level of glucose32. We thus
hypothesized that pericytes might sense fluctuations in glucose levels in the brain and respond to
decreases in glucose availability with KATP channel-mediated electrical signals.
Glucose concentration in bulk cerebrospinal fluid is ~4 mM33, whereas parenchymal glucose has
been measured in the range of 0.25-2.5 mM across a range of studies34–42. Accordingly, we
wondered whether subtle changes in local glucose concentration in this range would influence
the degree of KATP channel activity and thus modulate pericyte membrane potential (Vm). To
explore the relationship between glucose and pericyte Vm, we applied a series of decreasing
glucose concentrations to isolated capillaries from Cspg4-DsRed mice with intact thin-strand
pericytes, and measured Vm using microelectrode impalements (Fig. 4A). Across all conditions
of replete glucose (4 mM), pericyte Vm averaged -36 mV (22 cells, 10 mice; Fig. 4B,E,F,I).
Pericyte KATP channels tune blood flow to local metabolism
8
Figure 3. Capillary pericytes are the locus of KATP channel-mediated control of blood flow. (A) Pericyte KATP channels were
genetically inactivated by crossing mice possessing a modified KIR6.1 subunit (KIR6.1AAA) with Cspg4-Cre mice. Cre control mice
(KIR6.1AAA+, Cre -) and KIR6.1AAA+ mice (KIR6.1AAA+, Cre +) were given 4-hydroxytamoxifen (4-OHT), whereas vehicle control mice
(KIR6.1AAA+, Cre +) were given vehicle. (B) Successful inactivation of the KIR6.1 subunit was evidenced by elimination of eGFP
signal in pericytes. Left: Representative Z-projection from a vehicle control mouse. Right: Representative Z-projection from a
tamoxifen-induced Cspg4-Cre-KIR6.1AAA mouse showing fewer eGFP+ cells. (C) Experimental strategy to identify and target
inactivated KATP channels in pericytes. Left: A Cspg4-Cre-KIR6.1AAA mouse, with eGFP+ endothelial cells, and pericytes lacking
eGFP signal, indicating successful KIR6.1AAA induction. Right: The location of eGFP-negative pericytes was determined using the
in vivo pericyte-specific dye Neurotrace (NT) 500/525. Inset: A pipette containing FITC and 10 µM pinacidil positioned next to an
eGFP-, NT 500/525+ pericyte. (D) Example traces of PA diameter showing dilation to downstream ejection of pinacidil onto a
capillary pericyte in a Cre-control mouse (pink) and a lack of response in Cspg4-Cre-KIR6.1AAA mice (brown). (E-J) Summary data
of changes in PA diameter and blood flow to focal application of pinacidil onto a capillary pericyte across different experimental
groups. (E) PA diameter changes in Cre-control mice (n = 5 paired measurements, 5 mice, **P = 0.0026, t4 = 6.684). (F) RBC flux
changes in Cre-control mice (n = 5 paired measurements, 5 mice, ***P = 0.0006, t4 = 9.908). (G) PA diameter changes in vehicle
control mice (n = 3 paired measurements, 3 mice, *P = 0.0157, t2 = 7.883). (H) RBC flux changes in vehicle control mice (n = 3
paired measurements, 3 mice, **P = 0.0023, t2 = 20.78). (I) PA diameter changes in Cspg4-Cre-KIR6.1AAA mice (n = 10 paired
measurements, 10 mice, P = 0.3054, t9 = 1.087). (J) RBC flux changes in Cspg4-Cre-KIR6.1AAA mice (n = 10 paired measurements,
10 mice, P = 0.7249, t9 = 0.3631). All data were analyzed using paired Student's t-test.
Pericyte KATP channels tune blood flow to local metabolism
9
Figure 4. Lowering glucose activates KATP channels to hyperpolarize pericytes. (A) Overview of cell isolation and impalement.
Left to right: Pericytes were isolated by dissecting and mincing cortical tissue from a Cspg4-DsRed mouse. Minced pieces were
sequentially digested, homogenized and filtered to yield capillary fragments with DsRed-positive pericytes. (B-D) Example traces
of Vm measurements at baseline (B), with 10 μM pinacidil (C), and with 10 μM pinacidil in the presence of 10 μM glibenclamide
(D). (E) Summary data showing pinacidil hyperpolarizes pericyte Vm, and glibenclamide blocks this effect (baseline (10 cells, 5
mice) vs. pinacidil (9 cells, 4 mice): ***P = 0.002, t49 = 4.453; pinacidil vs. pinacidil + glibenclamide (6 cells, 4 mice): ****P < 0.0001,
t49 = 5.278, One-way ANOVA with Sidak's multiple comparison test). (F-H) Example traces of Vm measurements with 4 mM bath
glucose (F), with 0 bath glucose (G) and under 0 glucose conditions with the addition of 10 μM glibenclamide (H). (I) Summary
data showing that lowering glucose below 1 mM hyperpolarizes the pericyte membrane, and the effects of 0 glucose were blocked
by glibenclamide (4 mM glucose (12 cells, 5 mice) vs. 2 mM glucose (14 cells, 5 mice): P > 0.9999, = 0.1204; 4 mM glucose vs.
1 mM glucose (9 cells, 4 mice): P = 0.8784, t97 = 1.28; 4 mM glucose vs. 750 μM glucose (20 cells, 4 mice): ***P = 0.001, t97 =
4.04; 4 mM glucose vs. 250 μM glucose (16 cells, 4 mice): ***P = 0.0005, t97 = 4.193; 4 mM glucose vs. 0 glucose (9 cells, 5
mice): ***P = 0.0007, t97 = 4.078; 0 glucose vs. 0 glucose + 10 μM glibenclamide (10 cells, 4 mice): ****P < 0.0001, t97 = 5.22;
One-way ANOVA with Sidak's multiple comparison test). (J) Concentration-response curve showing pericyte membrane potential
hyperpolarizes abruptly in response to lowering glucose concentration.
Pericyte KATP channels tune blood flow to local metabolism
10
Under these conditions, activation of KATP channels with a saturating concentration of pinacidil (10
µM) hyperpolarized Vm by ~23 mV, an effect that was blocked by the co-application of 10 µM
glibenclamide (Fig. 4B-E). Strikingly, complete removal of glucose also strongly hyperpolarized
the membrane, to -52 mV (Fig 4G,I), an effect that was prevented by inclusion of 10 µM
glibenclamide in the bath (Fig. 4 H,I). Varying glucose within the physiological range measured
in the parenchyma (2 mM, 1 mM, 750 µM and 250 µM) revealed the presence of a threshold
around 1 mM (EC50: 934 µM; Fig. 4J), below which a dramatic increase in KATP channel activity
occurs that parallels that seen with 0 glucose, which we refer to as an ‘energy switch’ (Fig. 4I,J
and Supplementary Fig. 5). Together, these data indicate that pericytes monitor small
fluctuations of glucose within the physiological range, and if the concentration falls below a critical
threshold KATP channel activity is robustly increased to evoke substantial membrane
hyperpolarization.
GLUT1 block activates the pericyte energy switch in vivo and triggers profound arteriolar
dilation to increase local blood flow
The endothelium plays a major role in glucose import into the brain, predominately via highly-
expressed GLUT1 transporters (Fig. 5A), and pericytes also express the gene encoding GLUT1
and to a lesser extent the genes for GLUT3 and GLUT420,21. Given this central role, we
hypothesized that blocking GLUT1 would be sufficient to activate the pericyte energy switch and
generate KATP channel activity to hyperpolarize pericyte Vm. This, in turn, should influence
electrical signaling through the capillary network and drive an increase in arteriolar diameter and
blood flow.
In line with the predictions of our hypothesis, blocking glucose entry using the selective GLUT1
inhibitor BAY-876 (1 µM) hyperpolarized the pericyte membrane to -51 mV, as seen with
concentrations of glucose below 1 mM (Fig. 4I), and this effect was completely inhibited by
glibenclamide (Fig. 5B-D). Based on the known Vm-diameter relationship of PA smooth muscle,
a ~15-mV hyperpolarization is predicted to dilate PAs by approximately 50% (see ref 43).
Accordingly, we tested the effect of 1 µM BAY-876 on PA and capillary diameter, and capillary
blood flow when applied directly to the cranial surface in vivo. Strikingly, this maneuver produced
a 48% increase in PA diameter (Fig. 5E,G), in line with our predictions, and profoundly dilated
1st-4th order capillaries (Fig. 5F,H-K) while also almost doubling capillary blood flow (Fig. 5L,M).
Pre-incubation with glibenclamide (10 µM) eliminated the BAY-876–evoked increase in capillary
RBC flux (Fig. 5N,O) and significantly decreased the dilatory effect of BAY-876 at the level of the
PA (68% reduction) and in 1st-4th order capillaries (Fig. 5P). Together, these data indicate that a
reduction in glucose delivery to the pericyte interior triggers KATP channel-mediated electrical
signaling, which in turn is transmitted upstream to the PA to drive dilation and an increase blood
flow.
Pericyte KATP channels tune blood flow to local metabolism
11
DISCUSSION
Taken together, our data reveal that KATP channels in capillary thin-strand pericytes couple
changes in energy substrate levels to alterations of local brain blood flow. Our data support a
model in which pericyte KATP channels initiate robust hyperpolarization in response to a decrease
in local glucose below a critical threshold, which can be transferred over long distances through
Figure 5. Glucose levels control KATP
channel activity and blood flow in vivo. (A)
Staining with an anti-GLUT1 antibody
indicating the high density of this transporter
in brain capillaries. (B-C) Example traces of
membrane potential measurements under 1
μM BAY-876 (B) and 1 μM BAY-876 in the
presence of 10 μM glibenclamide (C). (D)
Summary data showing BAY-876 (19 cells,
6 mice) hyperpolarizes pericyte membrane
potential and this effect is blocked by
glibenclamide (10 cells, 6 mice: ***P =
0.0003, t27 = 4.192, unpaired Student's t-
test). (E) Effects of GLUT1 inhibitor BAY-
876 (1 μM) on PA diameter. Left: PA
diameter indicated by white line at baseline.
Right: Dilation of the same PA after
application of BAY-876 to the cranial
surface. (F) BAY-876 also dilates 1st-4th
order
capillaries.
Left:
A
Z-projection
showing diameters of 1st-4th order at
baseline, indicated by colored lines. Right:
Dilation of the same capillaries after BAY-
876 application. (G-K) Summary data
analyzed using paired Student's t-test,
showing dilation across all vessels with
BAY-876. (G) PA diameter (n = 17 vessels,
5 mice, ****P < 0.0001, t16 = 10.01). (H) 1st
order capillary diameter (n = 9 vessels, 5
mice, **P = 0.0029, t8 = 4.22). (I) 2nd order
capillary diameter (n = 16 vessels, 5 mice,
****P < 0.0001, t15 = 11.16). (J) 3rd order
capillary diameter (n = 16 vessels, 5 mice,
****P < 0.0001, t15 = 8.399) and (K) 4th order
capillary diameter (n = 17 vessels, 5 mice,
****P
<
0.0001,
t16
=
7.665).
(L)
Representative
1-s
segments
of
raw
kymographs demonstrating hyperemia to
BAY-876. Top: Baseline RBC flux. Bottom:
RBC flux measured in the same capillary
after BAY-876 application. (M) Summary
RBC flux data before and after BAY-876
application (n = 11 paired measurements, 4
mice, **P = 0.004, t10 = 3.71, paired
Student's t-test). (N) The blood flow
response to BAY-876 is mediated by KATP
channel activation. Top: RBC flux at
baseline. Bottom: RBC flux measured from
the same capillary showing no change in
blood flow after the application of BAY-876
in the presence of KATP channel blocker
glibenclamide (10 μM). (O) Summary RBC
flux data from >5th order capillaries when
BAY-876 was applied in the presence of
glibenclamide
glibenclamide (n = 24 paired measurements, 5 mice, P = 0.2324, t23 = 1.226, paired Student's t-test). (P) Summary data showing
significantly decreased dilatory responses to BAY-876 in the presence of glibenclamide across all vessel orders (n = 5 mice per
group; PA: **P = 0.0013, t163 = 3.734; 1st order capillary: *P = 0.0189, t163 = 2.936; 2nd order capillary: **P = 0.0079, t163 = 3.212;
3rd order capillary: ***P = 0.0009, t163 = 3.825; 4th order capillary: ***P = 0.0003, t163 = 4.14; one-way ANOVA with Sidak's multiple
comparison test).
Pericyte KATP channels tune blood flow to local metabolism
12
engagement of capillary electrical signaling, eliciting relaxation of remote arteriolar SMCs, leading
to vasodilation and an increase in blood flow into the capillary bed (Fig. 6).
A pericyte energy switch: membrane potential is steeply influenced by local glucose
concentration
The brain relies primarily on glucose and oxygen to fuel its energy requirements. The central
pathway for glucose entry into the brain is via the GLUT1 transporter, which is abundantly
expressed in blood brain barrier ECs21,44. The cell bodies and processes of pericytes that decorate
the vascular wall are embedded in the basement membrane that surrounds capillary ECs, and
this intimate association allows for the extension and receipt of projections known as peg-socket
junctions11 which bring the membranes of these two cell types into very close proximity and likely
facilitates the formation of gap junctions10,19,27. Given that gap junctions permit transfer of
molecules up to 1000 Da, combined with observations of cell-cell transfer of fluorescently-
conjugated glucose analogues45,46, it is reasonable to posit that glucose (~180 Da) taken up into
the EC cytoplasm may be transferred directly to pericytes via this avenue, the rate of which will
depend ultimately on the degree of coupling between these cell types. Pericytes also express
several GLUT-encoding genes (Slc2a1 > Slc2a4 > Slc2a3, which translate to GLUTs 1, 4, and 3,
respectively20,21), suggesting that they may also be capable of taking up glucose directly from their
surroundings. Collectively, these molecular features likely equip capillary pericytes to sense and
closely monitor glucose levels in their locale.
As a result, we hypothesized that pericytes may be capable of responding to changes in local
glucose availability through metabolically-evoked K+ channel activity and blood flow modulation,
by virtue of their robust KATP channel expression. Thus, to directly ascertain whether pericyte
electrical behavior is influenced by local energy availability, we sought to determine the
relationship between glucose concentration and pericyte Vm in granular detail, and specifically
focused on the contribution of KATP channels in this context. Accordingly, we tested the effects of
lowering glucose from 4 mM (the concentration typically found in bulk CSF) to 1 mM and below
(which aligns with measurements several independent groups have made of parenchymal
glucose concentrations34–42). We found that complete removal of glucose produced a striking 16-
mV hyperpolarization, mediated by KATP channel activation. Moreover, almost identical responses
were seen for glucose concentrations up to 0.75 mM and in circumstances in which we blocked
glucose import via GLUT1. In contrast, 2 mM glucose had little influence on Vm, which remained
close to the ‘resting’ value we obtained in 4 mM glucose (-36 mV). At 1 mM glucose, pericyte Vm
was slightly more hyperpolarized (-40 mV) but this was not significantly different than higher
glucose concentrations. These data indicate that pericytes are steeply sensitive to local changes
in this key energy substrate, and are consistent with the existence of a glucose concentration
threshold below which robust activation of K+ efflux through KATP is elicited. This ‘all-or-none’ effect
of energy substrate abundance on membrane potential—reminiscent of flipping a switch—may
be triggered by changes in glucose affecting the production of ATP in the pericyte, leading to a
new set point for the intracellular ATP:ADP ratio. Accompanying this could be an amplification
mechanism such as the engagement of capillary KIR channels, which are directly activated by
membrane hyperpolarization relieving voltage-dependent block of the channel pore by
Pericyte KATP channels tune blood flow to local metabolism
13
polyamines47. KIR channel activation in turn may boost KATP-initiated hyperpolarization and
combined, these factors could translate a change in intracellular metabolism into a binary
response, driving Vm towards EK and facilitating potent hyperemic responses to small changes in
external glucose availability. Alternatively, or perhaps in conjunction, other energy-sensing
molecules such as adenosine monophosphate-activated protein kinase (AMPK) may be engaged
by glucose deficits to phosphorylate KV channels (which have been reported in cultured retinal
pericytes but await confirmation in native cells48) and increase their activity49. As cECs have also
recently been shown to possess KATP channels, albeit at lower current density25, we cannot
presently fully rule out the possibility of their contribution to these KATP channel-mediated effects
on Vm, although our imaging data are consistent with pericytes playing the major role. Further
experiments are needed to explore these possibilities in detail.
What might be the circumstances, physiological and pathological, that engage this mechanism?
One possibility is that local fluctuations in glucose that occur during concerted neuronal activity50,51
continually adjust the electrical input of pericytes to the capillary endothelium, resulting in fine-
tuning of local blood flow to ensure that neuronal metabolism is protected on a moment-to-
moment time scale. Given that KATP channels do not appear to contribute to functional hyperemia
to a diffuse visual stimulus52, it may be that strong stimuli driving robust network activity and
rapidly ramping energy demands are required to engage this mechanism under physiological
conditions. It is also possible that the pericyte energy switch is reserved for pathological conditions
such as hypoglycemia, a common occurrence in diabetic individuals, where it might serve as an
emergency failsafe that has evolved to protect brain energy supply by increasing blood flow. In
support of these ideas, as parenchymal glucose approaches 0, blood flow has been observed to
increase by up to 57% (ref 35), and insulin-induced hypoglycemia increases blood flow by 42%
in adults53.
Pericyte KATP channel-meditated electrical signals are transmitted through multiple branch
orders to control blood flow
Thin-strand pericytes are found deep within the capillary bed, starting around the 5th order
branches and above. An elegant recent study deploying optogenetic tools in pericytes has shown
that these cells are capable of exerting slow constrictions of their underlying capillaries6, yet it
appears that they do not dilate during functional hyperemia16,54. To rapidly control blood flow, thin-
strand pericytes could modulate ongoing electrical signaling through the underlying endothelium
which is transmitted over long distances to influence upstream arteriolar diameter28. Together, the
present experiments support this idea and reveal that focal activation of the KATP channels in just
a single pericyte is sufficient to evoke rapid dilation of remote PAs at distances up to at least 421
µm (the furthest site we stimulated in our experiments). In stark contrast, we did not detect an
effect of direct stimulation of PAs with 10 μM pinacidil on diameter, although we note that
application of a 500-fold higher concentration onto PAs and 1st-3rd order capillaries caused
localized vasodilation in another study17. Our data using lower (but saturating) concentrations of
pinacidil suggest that functional KATP channels are absent, or present at too low of a density in the
arteriolar wall to generate sufficient hyperpolarization to elicit vasorelaxation under the conditions
used here, and this is consistent with previous findings in isolated and pressurized PAs which did
Pericyte KATP channels tune blood flow to local metabolism
14
not dilate to bath application of a KATP channel activator26. Moreover, pinacidil-stimulation of ECs
on a segment of 5th or higher order capillary lacking a pericyte cell body, or of pericytes with
genetically inactivated KATP channels, failed to dilate PAs. Collectively, these observations
strongly imply that pericytes represent the locus of KATP channel activity in the capillary bed, and
from this locus, hyperpolarization must then be transmitted to upstream SMCs to evoke dilation
at a distance. There are several possibilities as to how such long-range communication may be
achieved. The eponymous projections of thin-strand pericytes reach over long distances and
come into close contact with those of neighboring cells. However, these do not appear to closely
interdigitate and rather stay confined to their own territories8, and no evidence of direct pericyte-
pericyte transfer of charge or chemical agents has been reported to our knowledge, with the
exception of specialized interpericyte tunneling nanotube (IPNT) projections in the retina55. Thus,
it presently seems unlikely that capillary pericytes without IPNTs directly exchange electrical
signals. Instead, mounting evidence indicates that thin-strand pericytes directly interface with
capillary ECs via gap junctions10. Our prior work28 revealed that electrical signaling through the
brains capillary network to upstream arterioles is a major mechanism for blood flow control in the
brain. This mechanism relies on capillary EC KIR2.1 channels, which are activated by both external
K+ and membrane hyperpolarization and transmit electrical signals upstream at a velocity of
several millimeters per second23,28. Given that our data show that the KIR2 channel blocker Ba2+
eliminates pinacidil-evoked remote dilation of PAs, our observations in context with those of other
groups cumulatively suggest that the activation of KATP channels in pericytes generates
membrane hyperpolarization that is then injected via peg-socket junctions into the underlying ECs
to engage capillary EC electrical signaling and dilate upstream arterioles. We also recently
reported that capillary EC Ca2+ signals control blood flow through a nitric oxide-dependent
mechanism that relaxes contractile pericytes of the 1st to 4th order transitional segment of the
capillary bed18. Intriguingly, these signals are strongly influenced by ongoing electrical signaling
in the capillaries, with the hyperpolarization these provide likely increasing the driving force for
Ca2+ entry. Thus, it is possible that pericyte KATP-mediated electrical signals might also promote
capillary EC Ca2+ signaling, which could also be a contributory factor in the observed dilations
resulting from these.
Recasting Pericytes as Metabolic Sentinels
Pericytes play a range of roles in the brain, which include control of blood-brain barrier function12,
regulation of endothelial gene expression13, promotion of proper vascular development56,
provision of structural stability57, and regulation of blood flow15,58. Moreover, they appear to be
particularly vulnerable cells in the context of dementias and a range of other disorders impinging
on brain function (e.g. diabetes, hypertension and kidney dysfunction59), and contractile pericytes
have been noted to die in rigor which is thought to contribute to loss of brain blood flow control,
ultimately precipitating neuronal dysfunction and decline15. Our data evoke novel concepts
stemming from the sensitivity of thin-strand pericytes to subtle metabolic changes. Importantly, if
glucose drops below a critical threshold, a robust electrical response is generated through the
recruitment of pericyte KATP channels to increase local blood flow, thereby providing more glucose
to replenish local levels and protect ongoing neuronal function. This mechanism may be critical
Pericyte KATP channels tune blood flow to local metabolism
15
for the maintenance of brain health, and its disruption over long periods could contribute to the
mismatch between energy supply and demand that occurs in cognitive decline and dementia60–
62. Intriguingly, a recent VINE-seq atlas of human vascular cells suggest that the molecular players
that take center-stage in the electrical switch we have elucidated here (Kcnj8, Kcnj2, and Slc2a1)
are each profoundly downregulated in Alzheimer’s cerebrovasculature, which could potentially
disable protective responses to local glucose dips and imperil neuronal metabolism63. In support
of this idea, Kir2.1 function is known to be disrupted in the 5xFAD mouse model of AD64. Further
work is needed to address whether the pericyte energy switch is disabled in Alzheimer’s, and
ongoing experiments in our laboratory are now directly addressing these questions.
Our observation that pericytes are sensitive to glucose naturally evokes the question of whether
pericytes detect the levels of other energy substrates and metabolites. Pericytes exist in, and are
influenced by, a rich milieu of molecules and substrates, of which glucose is just one element.
Therefore, pericyte activity is likely to be regulated by a complex mix of factors which fluctuate in
concentration over widely varying timescales. One such factor, partnered with glucose to support
brain metabolism, might be oxygen. Oxidative phosphorylation relies on local oxygen tension,
which in turn is a direct function of local blood flow65. The oxidation of glucose provides vastly
more ATP than glycolysis alone and neuronal activity is primarily powered by oxidative
phosphorylation66. It is possible that the pericyte energy switch may also be activated by local
transient decreases in oxygen67, which might lead to an abrupt fall in intracellular ATP production,
influencing ATP:ADP ratio and engaging KATP channels. Interestingly, stalling (i.e. complete
cessation of RBC flux) behavior is relatively common in brain capillaries, with ~0.45% of capillaries
estimated to be stalled at any one time68. The function of this phenomenon is unclear, but it seems
likely that these events would lead to a localized decrease in oxygen tension due to the lack of
transiting RBCs loaded with oxygen. This may in turn activate the pericyte energy switch, leading
to signaling to increase blood flow to relieve the stall before it damages neurons. Still other
metabolites might be sensed by pericytes and evoke KATP-mediated hyperpolarizing responses.
As we previously noted19 pericytes express the A2a adenosine receptor, a Gs-coupled GPCR,
activation of which has recently been shown to lead to pericyte KATP channel activation through
protein kinase A25. As adenosine is released from neurons during their activity, this pathway may
also engage pericyte KATP channel activity to hyperpolarize the cell membrane and evoke
upstream arteriolar dilation as we have shown here. Pericytes might also possess mechanisms
to assess local carbon dioxide gradients69, which would reflect the degree of local metabolic
activity70, and may modulate blood flow in turn. It has also recently been demonstrated that
pericytes can sense lactate generated during glycolysis in ECs71, which could also serve as an
energy substrate that ultimately regulates pericyte KATP channel activity.
SUMMARY AND CONCLUSION
Despite their intimate association with capillaries, the precise contribution of thin-strand pericytes
to the control of blood flow in the brain is largely unknown. A rich complement of ion channels and
GPCRs equips pericytes to sense and respond to a wide range of stimuli19. KATP channels are the
most abundant ion channel expressed by pericytes19, and we demonstrate here that their
activation in response to decreased local metabolic substrate availability produces a robust
Pericyte KATP channels tune blood flow to local metabolism
16
increase in blood flow. Our data thus recast pericytes as metabolic sentinels that form a brain-
wide energy-sensing network, continually monitoring glucose concentrations and adjusting blood
flow to protect ongoing neuronal health and function. In conditions like sporadic Alzheimer’s
disease, brain glucose levels and metabolism are profoundly dysregulated72–75 and thus
determining the impact of this on pericyte energy sensing and accompanying blood flow control
may yield potential targets for improving clinical outcomes in neurological diseases with a
significant vascular and metabolic component.
.
Figure
6.
Illustrative
summary and model for
pericyte
KATP
channel-
mediated
coupling
of
electrical
activity
with
glucose availability. Under
conditions
of
abundant
glucose, pericyte ATP:ADP
ratio is high and keeps
pericyte
KATP
channels
closed. A decrease in GLUT-
1 mediated glucose import,
or a drop in local glucose
availability results in pericyte
KATP channel activation, likely
due
to
a
corresponding
decrease
in
cellular
ATP:ADP
ratio.
When
activated,
KATP
channels
robustly
hyperpolarize
pericyte membrane potential
and this electrical signal is
then fed into the underlying
capillary endothelium to be
rapidly transmitted upstream
via
a
KIR2.1
channel-
dependent mechanism. This
remotely dilates penetrating
arterioles
and
increases
blood
flow,
thereby
replenishing local glucose
levels and protecting ongoing
neuronal
metabolism and
function.
Pericyte KATP channels tune blood flow to local metabolism
17
METHODS
Animal husbandry. Adult (2–3 mo. old) male and female C57BL/6J mice, Cspg4-DsRed mice
(C57BL/6J background; Jackson Laboratories), Cspg4-Cre recombinase mice, and Cspg4-Cre-
KIR6.1AAA mice were group-housed on a 12-h light:dark cycle with environmental enrichment and
free access to food and water. Tamoxifen inducible Cspg4-Cre-KIR6.1AAA mice were generated by
crossing KIR6.1AAA mice expressing dominant-negative KIR6.1AAA with Cspg4-Cre recombinase
mice29,30. All animal procedures received prior approval from the University of Maryland
Institutional Animal Care and Use Committee.
KIR6.1AAA induction. 4-OHT, the active metabolite of tamoxifen, was dissolved in a corn
oil:ethanol solution (90:10% v/v) at a concentration of 2 mg/ml 76. Cspg4-Cre-KIR6.1AAA mice were
given either 4-OHT (10 mg/kg, intraperitoneal; KIR6.1AAA induction) or vehicle (corn oil:ethanol;
vehicle control), and control KIR6.1AAA mice were given 4-OHT (10mg/kg, intraperitoneal; Cre
control) once a day for 5 consecutive days. 4 weeks after the last injection, mice were imaged in
vivo as described below.
Chemicals. BAY-876 was purchased from Tocris Bioscience (USA). All other chemicals were
obtained from Sigma Aldrich (USA).
In vivo imaging. Cranial window preparation and in vivo imaging was performed as previously
described18,28. Briefly, mice were anesthetized with isoflurane (5% induction, 1.5-2%
maintenance). 150 µL of FITC-dextran (10mg/ml) or TRITC-dextran (40 mg/ml) dissolved in saline
was injected retro-orbitally. A midline incision was made on the scalp to expose the skull, and a
titanium head plate was affixed over the left hemisphere with a combination of dental cement and
superglue. On securing the headplate in a holding frame, a circular cranial window (~2 mm
diameter) was drilled in the skull over the somatosensory cortex. The skull piece was removed,
and the dura was carefully resected. The cranial surface was irrigated as necessary with saline.
Upon conclusion of surgery, isoflurane anesthesia was replaced with α-chloralose (50 mg/kg) and
urethane (750 mg/kg). Body temperature was maintained at 37°C throughout the experiment
using a rectal probe feedback-controlled electric heating pad (Harvard Apparatus). Oxygenated
and warmed (35-36 °C) aCSF (124 mM NaCl, 3 mM KCl, 2 mM CaCl2, 2 mM MgCl2, 1.25 mM
NaH2PO4, 26 mM NaHCO3, 4 mM glucose) was superfused over the exposed cortex for the
duration of the experiment at a rate of ~1 mL/min and continuously monitored at the window with
a temperature probe. Images were acquired through an Olympus 20x infinity-corrected Plan
Fluorite 1.0 NA water-immersion objective mounted on a Scientifica Hyperscope (Scientifica, UK)
coupled to a Coherent Chameleon Vision II Titanium-Sapphire pulsed fs laser (Coherent, USA).
FITC- and TRITC-dextran or DsRed were excited at 820 nm or 920 nm, respectively, and emitted
fluorescence was separated through 525/50 and 620/60 nm bandpass filters. Single-plane
imaging data to examine the time course of vessel diameter changes was collected at 30 Hz using
a resonant scanning mirror. 3D imaging data were typically gathered using standard galvo mirrors.
To measure RBC flux, we performed line scans at 1 kHz. Line scans were oriented along the
Pericyte KATP channels tune blood flow to local metabolism
18
lumen parallel to the flow of blood to maximize flux signal. For pressure-ejection of agents in aCSF
(vehicle) onto pericytes or endothelial cells, a pipette containing the agent of interest and FITC or
TRITC-dextran (to enable visualization) was maneuvered into the cortex and positioned adjacent
to the cell under study, after which the solution was ejected directly at 8–12 psi, for 30 ms. This
approach restricted agent delivery to the target cell and caused minimal displacement of the
surrounding tissue. For pharmacological and staining experiments, agents of interest were
applied to the cranial surface for a minimum of 20 min to allow penetration. All in vivo imaging
experiments were routinely ended with the application of aCSF containing 0 Ca2+ supplemented
with 5 mM EGTA and 200 µM diltiazem to elicit maximal relaxation of SMCs and contractile
pericytes to enable the measurement of maximum vessel diameters.
Microelectrode impalement of pericytes on isolated microvessels.
Membrane potential measurements were made by impaling pericytes on microvessels isolated
from Cspg4-DsRed mice using a papain-based Neural Tissue Dissociation kit (Miltenyi Biotec),
as described previously21,25. Cortical tissue from one hemisphere was carefully dissected and
minced into small pieces with microscalpels in an isolation solution containing 55 mM NaCl, 80
mM Na-glutamate, 5.6 mM KCl, 2 mM MgCl2, 10 mM HEPES and 4 mM glucose (pH 7.3). Minced
tissue was incubated with enzyme P from the kit for 18 min at 37°C, followed by addition of
enzyme A, homogenization by passing through a Pasteur pipette ~10 times and incubation for 15
min at 37°C. The homogenate was then passed through a 21 G needle 7 times and incubated for
12 min at 37°C. The cell suspension was filtered through a 62-μm nylon mesh and stored in ice-
cold isolation solution. Cells were transferred to a silicone elastomer (SYLGARD 182)-lined
perfusion chamber, and allowed to adhere for ~45 min. The chamber was perfused with bath
solution consisting of 137 mM NaCl, 3 mM KCl, 2 mM CaCl2, 1 mM MgCl2, 10 mM HEPES and 0-
4 mM glucose (pH 7.4). DsRed-positive pericytes on small capillary segments were identified
using brightfield microscopy, and impaled with a sharp microelectrode (pulled to ~100-200 MΩ)
filled with 0.5 M KCl. Only recordings fulfilling the following criteria were considered for analysis:
stable baseline prior to impalement, sharp negative deflection of membrane potential upon
impalement, immediate return to 0 mV upon withdrawing the electrode. Membrane potential was
recorded using an AxoClamp 900A digital amplifier and HS-2 headstage (Molecular Devices).
Signals were digitized and stored using Axon Digidata 1550B and pClamp 9 software (Molecular
Devices).
Immunohistochemistry
Brains were extracted from Cspg4-DsRed mice that underwent cardiac perfusion with 4%
paraformaldehyde. Tissues were stored in 4% paraformaldehyde overnight at 2–8ºC and
dehydrated in 30% sucrose in 1x phosphate buffered saline (PBS). Immunostaining and optical
clearing of brain samples were performed according to a modified CUBIC clearing method77,78.
Briefly, fixed brains were immunostained by first blocking non-specific binding with normal goat
serum (Vector Laboratories, USA). Blocked samples were incubated overnight at 2–8ºC with
rabbit anti-SLC2A1 polyclonal antibody (1:500 dilution, HPA031345, Atlas Antibodies, Stockholm,
Sweden) and developed with Alexa Fluor 488 goat anti-human IgG secondary antibody (1:1000).
Pericyte KATP channels tune blood flow to local metabolism
19
Samples were cleared by incubation in CUBIC R1 solution (see ref 77) at 37°C with shaking for
2-3 weeks, and then incubated in RIMS (refractive index matching solution; 88% w/v Histodenz
in 0.02 M PBS with 0.01% sodium azide) at 37°C until the samples were optically clear (~5 days)
with solution being replaced every 24 hours. Cleared tissue was mounted in RIMS and imaged
with a Nikon W1 spinning disk confocal microscope.
Data analysis and statistical testing
Diameter measurements were analyzed offline using ImageJ software. Vessel diameter was
calculated as the average of three measurements per vessel type made from Z-stacks of 3D
volume recordings using the full-width at half-maximum method. RBC flux data were binned at 1-
s intervals and analyzed using SparkAn software (A. Bonev, University of Vermont). For pressure-
ejection experiments, mean baseline diameter and flux were obtained by averaging the baseline
for each measurement before ejection of pinacidil or aCSF, and peak diameter and RBC flux
change was defined as the largest change from mean baseline. The distance from the site of
pressure ejection to the feed arteriole was estimated using the Simple Neurite Tracer plugin on
ImageJ software79. Statistical testing was performed using GraphPad Prism 7 software. Data are
expressed as means ± s.e.m., and a P-value ≤ 0.05 was considered significant. Stars denote
significant differences; ‘n.s.’ indicates comparisons that did not achieve statistical significance.
Statistical tests are noted in figure legends. All t-tests were two-sided. Statistical methods were
not used to pre-determine sample sizes, and ample size was estimated based on similar
experiments performed previously in our laboratory. Experiments were repeated to adequately
reduce confidence intervals and avoid errors in statistical testing. Data collection was not
performed blinded to the conditions of the experiments. Littermates were randomly assigned to
experimental groups; no further randomization was performed. No data were excluded.
ACKNOWLEDGEMENTS
The authors thank B. Huang and S. Edwards for animal husbandry and experimental support.
Support for this work was provided by the NIH National Institute on Aging and National Institute
of Neurological Disorders and Stroke (1R01AG066645, 5R01NS115401, and 1DP2NS121347-
01, to T.A.L), and the American Heart Association and the D.C. Women’s Board (Award 830093
to A.H,17SDG33670237 and 19IPLOI34660108 to T.A.L).
AUTHOR CONTRIBUTIONS
A.H. designed experiments, acquired and analyzed data, and edited the manuscript. C.R acquired
and analyzed pinacidil surface application data, D.G performed immunofluorescence and imaging
of GLUT1 staining, T.A.L directed the study, acquired and analyzed data, and edited the
manuscript. All authors reviewed the manuscript and approved its submission.
DECLARATION OF INTERESTS
The authors declare no financial or non-financial conflict of interest.
Pericyte KATP channels tune blood flow to local metabolism
20
REFERENCES
1.
Blinder P, Tsai PS, Kaufhold JP, Knutsen PM, Suhl H, Kleinfeld D. The cortical angiome:
an interconnected vascular network with noncolumnar patterns of blood flow. Nat
Neurosci. 2013;16:889–897. doi:10.1038/nn.3426
2.
Grubb S, Cai C, Hald BO, et al. Precapillary sphincters maintain perfusion in the cerebral
cortex. Nat Commun. 2020;11:395. doi:10.1038/s41467-020-14330-z
3.
Ratelade J, Klug NR, Lombardi D, et al. Reducing Hypermuscularization of the
Transitional Segment between Arterioles and Capillaries Protects Against Spontaneous
Intracerebral Hemorrhage. Circulation. 2020;141:2078–2094.
doi:10.1161/CIRCULATIONAHA.119.040963
4.
Grant RI, Hartmann DA, Underly RG, Berthiaume A-A, Bhat NR, Shih AY. Organizational
hierarchy and structural diversity of microvascular pericytes in adult mouse cortex. J
Cereb Blood Flow Metab. 2019;39(3):411–425. doi:10.1177/0271678X17732229
5.
Gonzales AL, Klug NR, Moshkforoush A, et al. Contractile pericytes determine the
direction of blood flow at capillary junctions. Proc Natl Acad Sci U S A. 2020.
doi:10.1073/pnas.1922755117
6.
Hartmann DA, Berthiaume AA, Grant RI, et al. Brain capillary pericytes exert a substantial
but slow influence on blood flow. Nat Neurosci. 2021;24(5):633-645. doi:10.1038/s41593-
020-00793-2
7.
Hartmann DA, Underly RG, Grant RI, Watson AN, Lindner V, Shih AY. Pericyte structure
and distribution in the cerebral cortex revealed by high-resolution imaging of transgenic
mice. Neurophotonics. 2015;2(4):041402. doi:10.1117/1.NPh.2.4.041402
8.
Berthiaume AA, Grant RI, McDowell KP, et al. Dynamic Remodeling of Pericytes In Vivo
Maintains Capillary Coverage in the Adult Mouse Brain. Cell Rep. 2018;22(1):8-16.
doi:10.1016/j.celrep.2017.12.016
9.
Armulik A, Abramsson A, Betsholtz C. Endothelial/pericyte interactions. Circ Res.
2005;97:512–523. doi:10.1161/01.RES.0000182903.16652.d7
10.
Kovacs-Oller T, Ivanova E, Bianchimano P, Sagdullaev BT. The pericyte connectome:
spatial precision of neurovascular coupling is driven by selective connectivity maps of
pericytes and endothelial cells and is disrupted in diabetes. Cell Discov. 2020;6:39.
doi:10.1038/s41421-020-0180-0
11.
Ornelas S, Berthiaume AA, Bonney SK, et al. Three-dimensional ultrastructure of the
brain pericyte-endothelial interface. J Cereb Blood Flow Metab. 2021:1-16.
doi:10.1177/0271678X211012836
12.
Armulik A, Genové G, Mäe M, et al. Pericytes regulate the blood-brain barrier. Nature.
2010;468(7323):557-561. doi:10.1038/nature09522
13.
Mäe MA, He L, Nordling S, et al. Single-Cell Analysis of Blood-Brain Barrier Response to
Pericyte Loss. Circ Res. 2021:E46-E62. doi:10.1161/CIRCRESAHA.120.317473
14.
Peppiatt CM, Howarth C, Mobbs P, Attwell D. Bidirectional control of CNS capillary
diameter by pericytes. Nature. 2006;443:700–704. doi:10.1038/nature05193
15.
Hall CN, Reynell C, Gesslein B, et al. Capillary pericytes regulate cerebral blood flow in
health and disease. Nature. 2014;508:55–60. doi:10.1038/nature13165
16.
Rungta RL, Chaigneau E, Osmanski B-F, Charpa S. Vascular Compartmentalization of
Functional Hyperemia from the Synapse to the Pia. Neuron. 2018;99(2):362-37.
doi:10.1016/j.neuron.2018.06.012
17.
Zambach SA, Cai C, Helms HCC, et al. Precapillary sphincters and pericytes at first-order
capillaries as key regulators for brain capillary perfusion. Proc Natl Acad Sci U S A.
Pericyte KATP channels tune blood flow to local metabolism
21
2021;118(26). doi:10.1073/pnas.2023749118
18.
Longden TA, Mughal A, Hennig GW, et al. Local IP(3) receptor-mediated Ca(2+) signals
compound to direct blood flow in brain capillaries. Sci Adv. 2021;7(30).
doi:10.1126/sciadv.abh0101
19.
Hariharan A, Weir N, Robertson C, He L, Betsholtz C, Longden TA. The Ion Channel and
GPCR Toolkit of Brain Capillary Pericytes (In Press). Front Cell Neurosci. 2020;14.
doi:10.3389/fncel.2020.601324
20.
He L, Vanlandewijck M, Mäe M, et al. Single cell RNAseq of mouse brain and lung
vascular and vessel-associated cell types. Sci Data. 2018;5:180160.
doi:10.1038/sdata.2018.160
21.
Vanlandewijck M, He L, Mäe MA, et al. A molecular atlas of cell types and zonation in the
brain vasculature. Nature. 2018;554(7693):475-480. doi:10.1038/nature25739
22.
Tarasov A, Dusonchet J, Ashcroft F. Metabolic regulation of the pancreatic beta-cell ATP-
sensitive K+ channel: a pas de deux. Diabetes. 2004;53(Suppl 3):S113-S122.
doi:10.2337/diabetes.53.suppl_3.S113
23.
Zhao G, Joca HC, Nelson MT, Lederer WJ. ATP- And voltage-dependent electro-
metabolic signaling regulates blood flow in heart. Proc Natl Acad Sci U S A.
2020;117(13):7461-7470. doi:10.1073/pnas.1922095117
24.
Davis MJ, Hill MA. Signaling mechanisms underlying the vascular myogenic response.
Physiol Rev. 1999;79(2):387-423. doi:10.1152/physrev.1999.79.2.387
25.
Sancho M, Klug NR, Mughal A, Heppner TJ, Hill-Eubanks D, Nelson MT. Electro-
Metabolic Sensing Through Capillary ATP-Sensitive K+ Channels and Adenosine to
Control Cerebral Blood Flow. bioRxiv. 2021:2021.03.12.435152.
doi:10.1101/2021.03.12.435152
26.
Dabertrand F, Nelson MT, Brayden JE. Acidosis dilates brain parenchymal arterioles by
conversion of calcium waves to sparks to activate BK channels. Circ Res.
2012;110(2):285-294. doi:10.1161/CIRCRESAHA.111.258145
27.
Díaz-Flores L, Gutiérrez R, Madrid JF, et al. Pericytes. Morphofunction, interactions and
pathology in a quiescent and activated mesenchymal cell niche. Histol Histopathol.
2009;24(7):909-969. doi:10.14670/HH-24.909
28.
Longden TA, Dabertrand F, Koide M, et al. Capillary K+-sensing initiates retrograde
hyperpolarization to locally increase cerebral blood flow. Nat Neurosci. 2017;20:717–726.
doi:10.1038/nn.4533
29.
Tong XY, Porter LM, Liu GX, et al. Consequences of cardiac myocyte-specific ablation of
KATP channels in transgenic mice expressing dominant negative Kir6 subunits. Am J
Physiol - Hear Circ Physiol. 2006;291(2). doi:10.1152/ajpheart.00051.2006
30.
Malester B, Tong X, Ghiu I, et al. Transgenic expression of a dominant negative K ATP
channel subunit in the mouse endothelium: effects on coronary flow and endothelin‐1
secretion . FASEB J. 2007;21(9):2162-2172. doi:10.1096/fj.06-7821com
31.
Damisah EC, Hill RA, Tong L, Murray KN, Grutzendler J. A fluoro-Nissl dye identifies
pericytes as distinct vascular mural cells during in vivo brain imaging. Nat Neurosci.
2017;20(7):1023-1032. doi:10.1038/nn.4564
32.
Akrouh A, Halcomb SE, Nichols CG, Sala-Rabanal M. Molecular biology of KATP
channels and implications for health and disease. IUBMB Life. 2009;61(10):971-978.
doi:10.1002/iub.246
33.
Brown PD, Davies SL, Speake T, Millar ID. Molecular mechanisms of cerebrospinal fluid
production. Neuroscience. 2004;129(4):955-968. doi:10.1016/j.neuroscience.2004.07.003
34.
Gruetter R, Novotny EJ, Boulware SD, et al. Direct measurement of brain glucose
Pericyte KATP channels tune blood flow to local metabolism
22
concentrations in humans by 13C NMR spectroscopy. Proc Natl Acad Sci U S A.
1992;89(3):1109-1112. doi:10.1073/pnas.89.3.1109
35.
Choi IY, Lee SP, Kim SG, Gruetter R. In vivo measurements of brain glucose transport
using the reversible michaelis-menten model and simultaneous measurements of
cerebral blood flow changes during hypoglycemia. J Cereb Blood Flow Metab.
2001;21(6):653-663. doi:10.1097/00004647-200106000-00003
36.
Ritter S. Monitoring and Maintenance of Brain Glucose Supply. Appet Food Intake.
2017:177-204. doi:10.1201/9781315120171-9
37.
De Vries MG, Arseneau LM, Lawson ME, Beverly JL. Extracellular Glucose in Rat
Ventromedial Hypothalamus during Acute and Recurrent Hypoglycemia. Diabetes.
2003;52(11):2767-2773. doi:10.2337/diabetes.52.11.2767
38.
Silver IA, Erecinska M. Extracellular glucose concentration in mammalian brain:
Continuous monitoring of changes during increased neuronal activity and upon limitation
in oxygen supply in normo-, hypo-, and hyperglycemic animals. J Neurosci.
1994;14(8):5068-5076. doi:10.1523/jneurosci.14-08-05068.1994
39.
McNay EC, Sherwin RS. From artificial cerebro-spinal fluid (aCSF) to artificial
extracellular fluid (aECF): Microdialysis perfusate composition effects on in vivo brain
ECF glucose measurements. J Neurosci Methods. 2004;132(1):35-43.
doi:10.1016/j.jneumeth.2003.08.014
40.
McNay EC, Gold PE. Extracellular Glucose Concentrations in the Rat Hippocampus
Measured byZero-Net-Flux. J Neurochem. 1999;72(2):785-790. doi:10.1046/j.1471-
4159.1999.720785.x
41.
Dunn-Meynell AA, Sanders NM, Compton D, et al. Relationship among brain and blood
glucose levels and spontaneous and glucoprivic feeding. J Neurosci. 2009;29(21):7015-
7022. doi:10.1523/JNEUROSCI.0334-09.2009
42.
Abi-Saab WM, Maggs DG, Jones T, et al. Striking differences in glucose and lactate
levels between brain extracellular fluid and plasma in conscious human subjects: Effects
of hyperglycemia and hypoglycemia. J Cereb Blood Flow Metab. 2002;22(3):271-279.
doi:10.1097/00004647-200203000-00004
43.
Longden TA, Hill-Eubanks DC, Nelson MT. Ion channel networks in the control of
cerebral blood flow. J Cereb Blood Flow Metab. 2016;36(3):492–512.
doi:10.1177/0271678X15616138
44.
Patching SG. Glucose Transporters at the Blood-Brain Barrier: Function, Regulation and
Gateways for Drug Delivery. Mol Neurobiol. 2017;54(2):1046-1077.
45.
Matsunami T, Suzuki T, Hisa Y, Takata K, Takamatsu T, Oyamada M. Gap junctions
mediate glucose transport between GLUT1-positive and -negative cells in the spiral
limbus of the rat cochlea. Cell Commun Adhes. 2006;13(1-2):93-102.
doi:10.1080/15419060600631805
46.
Rouach N, Koulakoff A, Abudara V, Willecke K, Giaume C. Astroglial metabolic networks
sustain hippocampal synaptic transmission. Science (80- ). 2008;322(5907):1551-1555.
doi:10.1126/science.1164022
47.
Longden TA, Nelson MT. Vascular inward rectifier K+ channels as external K+ sensors in
the control of cerebral blood flow. Microcirculation. 2015;22(3):183–196.
doi:10.1111/micc.12190
48.
Quignard J, Harley E, Duhault J, Vanhoutte P, Félétou M. K+ Channels in Cultured
Bovine Retinal Pericytes: Effects of β-Adrenergic Stimulation. J Cardiovasc Pharmacol.
2003;42(3):379-388.
49.
Ikematsu N, Dallas ML, Ross FA, et al. Phosphorylation of the voltage-gated potassium
Pericyte KATP channels tune blood flow to local metabolism
23
channel Kv2.1 by AMP-activated protein kinase regulates membrane excitability. Proc
Natl Acad Sci U S A. 2011;108(44):18132-18137. doi:10.1073/pnas.1106201108
50.
Fox PT, Raichle ME, Mintun MA, Dence C. Nonoxidative glucose consumption during
focal physiologic neural activity. Science. 1988;241:462-464.
51.
Lundgaard I, Li B, Xie L, et al. Direct neuronal glucose uptake heralds activity-dependent
increases in cerebral metabolism. Nat Commun. 2015;6. doi:10.1038/ncomms7807
52.
Hosford PS, Christie IN, Niranjan A, et al. A critical role for the ATP-sensitive potassium
channel subunit KIR6.1 in the control of cerebral blood flow. J Cereb Blood Flow Metab.
2019;39(10):2089–2095. doi:10.1177/0271678X18780602
53.
McManus R, Ioussoufovitch S, Froats E, St Lawrence K, Van Uum S, Diop M. Dynamic
response of cerebral blood flow to insulin-induced hypoglycemia. Sci Rep. 2020;10(1).
doi:10.1038/s41598-020-77626-6
54.
Fernández-Klett F, Offenhauser N, Dirnagl U, Priller J, Lindauer U. Pericytes in capillaries
are contractile in vivo, but arterioles mediate functional hyperemia in the mouse brain.
Proc Natl Acad Sci U S A. 2010;107(51):22290-22295. doi:10.1073/pnas.1011321108
55.
Alarcon-Martinez L, Villafranca-Baughman D, Quintero H, et al. Interpericyte tunnelling
nanotubes regulate neurovascular coupling. Nature. 2020;585(7823):91-95.
doi:10.1038/s41586-020-2589-x
56.
Payne LB, Hoque M, Houk C, Darden J, Chappell JC. Pericytes in Vascular
Development. Curr Tissue Microenviron Reports. 2020;1(3):143-154.
doi:10.1007/s43152-020-00014-9
57.
Benjamin LE, Hemo I, Keshet E. A plasticity window for blood vessel remodelling is
defined by pericyte coverage of the preformed endothelial network and is regulated by
PDGF-B and VEGF. Development. 1998;125(9):1591-1598. doi:10.1242/dev.125.9.1591
58.
Kisler K, Nelson AR, Rege S V, et al. Pericyte degeneration leads to neurovascular
uncoupling and limits oxygen supply to brain. Nat Neurosci. 2017;20:406–41.
doi:10.1038/nn.4489
59.
Hirunpattarasilp C, Attwell D, Freitas F. The role of pericytes in brain disorders: from the
periphery to the brain. J Neurochem. 2019;150(6):648-665. doi:10.1111/jnc.14725
60.
Kisler K, Nelson AR, Montagne A, Zlokovic B V. Cerebral blood flow regulation and
neurovascular dysfunction in Alzheimer disease. Nat Rev Neurosci. 2017;18(7):419-434.
doi:10.1038/nrn.2017.48
61.
Girouard H, Iadecola C. Neurovascular coupling in the normal brain and in hypertension,
stroke, and Alzheimer disease. J Appl Physiol. 2006;100(1):328-335.
doi:10.1152/japplphysiol.00966.2005
62.
De La Torre JC. Critically attained threshold of cerebral hypoperfusion: The CATCH
hypothesis of Alzheimer’s pathogenesis. Neurobiol Aging. 2000;21(2):331-342.
doi:10.1016/S0197-4580(00)00111-1
63.
Yang AC, Vest RT, Kern F, et al. A human brain vascular atlas reveals diverse mediators
of Alzheimer’s risk. Nature. 2022. doi:10.1038/s41586-021-04369-3
64.
Mughal A, Harraz OF, Gonzales AL, Hill-Eubanks D, Nelson MT. PIP2 Improves Cerebral
Blood Flow in a Mouse Model of Alzheimer’s Disease. Function. 2021;2(2).
doi:10.1093/function/zqab010
65.
Jespersen SN, Østergaard L. The roles of cerebral blood flow, capillary transit time
heterogeneity, and oxygen tension in brain oxygenation and metabolism. J Cereb Blood
Flow Metab. 2012;32(2):264–277. doi:10.1038/jcbfm.2011.153
66.
Hall CN, Klein-Flügge MC, Howarth C, Attwell D. Oxidative phosphorylation, not
glycolysis, powers presynaptic and postsynaptic mechanisms underlying brain
Pericyte KATP channels tune blood flow to local metabolism
24
information processing. J Neurosci. 2012;32(26):8940-8951.
doi:10.1523/JNEUROSCI.0026-12.2012
67.
Wei HS, Kang H, Rasheed IYD, et al. Erythrocytes Are Oxygen-Sensing Regulators of
the Cerebral Microcirculation. Neuron. 2016;91(4):851-862.
doi:10.1016/j.neuron.2016.07.016
68.
Erdener ŞE, Tang J, Sajjadi A, et al. Spatio-temporal dynamics of cerebral capillary
segments with stalling red blood cells. J Cereb Blood Flow Metab. 2019;39(5):886-900.
doi:10.1177/0271678X17743877
69.
Matsugi T, Chen Q, Anderson DR. Suppression of CO2-induced relaxation of bovine
retinal pericytes by angiotensin II. Investig Ophthalmol Vis Sci. 1997;38(3):652-657.
70.
Dienel GA. Brain glucose metabolism: Integration of energetics with function. Physiol
Rev. 2019;99(1):949-1045. doi:10.1152/physrev.00062.2017
71.
Lee H-W, Xu Y, Zhu X, et al. Endothelium-derived lactate is required for pericyte function
and blood–brain barrier maintenance. EMBO J. 2022;e109890.
doi:10.15252/embj.2021109890
72.
Mosconi L, Pupi A, De Leon MJ. Brain glucose hypometabolism and oxidative stress in
preclinical Alzheimer’s disease. Ann N Y Acad Sci. 2008;1147:180-195.
doi:10.1196/annals.1427.007
73.
Daulatzai MA. Cerebral hypoperfusion and glucose hypometabolism: Key
pathophysiological modulators promote neurodegeneration, cognitive impairment, and
Alzheimer’s disease. J Neurosci Res. 2017;95(4):943-972. doi:10.1002/jnr.23777
74.
An Y, Varma VR, Varma S, et al. Evidence for brain glucose dysregulation in Alzheimer’s
disease. Alzheimer’s Dement. 2018;14(3):318-329. doi:10.1016/j.jalz.2017.09.011
75.
Xu J, Begley P, Church SJ, et al. Elevation of brain glucose and polyol-pathway
intermediates with accompanying brain-copper deficiency in patients with Alzheimer’s
disease: metabolic basis for dementia. Sci Rep. 2016;6:27524. doi:10.1038/srep27524
76.
Cazzulino AS, Martinez R, Tomm NK, Denny CA. Improved specificity of hippocampal
memory trace labeling. Hippocampus. 2016;26(6):752-762. doi:10.1002/hipo.22556
77.
Muntifering M, Castranova D, Gibson GA, Meyer E, Kofron M, Watson AM. Clearing for
Deep Tissue Imaging. Curr Protoc Cytom. 2018;86(1). doi:10.1002/cpcy.38
78.
Susaki EA, Tainaka K, Perrin D, et al. Whole-brain imaging with single-cell resolution
using chemical cocktails and computational analysis. Cell. 2014;157(3):726-739.
doi:10.1016/j.cell.2014.03.042
79.
Longair MH, Baker DA, Armstrong JD. Simple neurite tracer: Open source software for
reconstruction, visualization and analysis of neuronal processes. Bioinformatics.
2011;27(17):2453-2454. doi:10.1093/bioinformatics/btr390
Pericyte KATP channels tune blood flow to local metabolism
25
SUPPLEMENTAL INFORMATION
Supplementary figure 1. Summary data of myogenic tone of PAs and capillaries in vivo. PA (8
vessels, 5 mice) vs. 1st order capillary (8 vessels, 5 mice): P = 0.9814, q38 = 0.41; PA vs. 2nd order
capillary (10 vessels, 5 mice): P = 0.8037, q38 = 0.8518; PA vs. 3rd order capillary (9 vessels, 5
mice): P = 0.4761, q38 = 1.336; PA vs. 4th order capillary (8 vessels, 5 mice): P = 0.1092, q38 =
2.185; one-way ANOVA with Dunnett's multiple comparison test.
Pericyte KATP channels tune blood flow to local metabolism
26
Supplementary figure 2. Vascular KATP channel activity is minimal at rest in vivo. (A) Effects of
glibenclamide (10 μM) on PA diameter. Left: PA diameter indicated by white line at baseline.
Right: The same PA after application of glibenclamide. (B) A PA and its downstream 1st-4th order
capillaries. Left: Baseline diameters of 1st-4th order capillaries indicated by respective colored
lines. Right: The same 1st-4th order capillaries after application of glibenclamide. (C) Summary
data of PA diameter before and after application of glibenclamide (n = 20 paired measurements,
6 mice, P = 0.7779, t19 = 0.2861, paired Student's t-test). (D) Summary data of 1st – 4th order
capillary diameter showing no change after glibenclamide application (n = 6 mice per group; 1st
order capillary: P = 0.4326, t196 = 1.512; 2nd order capillary: P = 0.2001, t196 = 1.936; 3rd order
capillary: P = 0.8198, t196 = 0.9398; 4th order capillary: P = 0.7720, t196 = 1.020; one-way ANOVA
with Sidak's multiple comparison test). (E) Summary RBC flux data from >5th order capillaries
demonstrating no change in blood flow after application of glibenclamide (n = 32 paired
measurements, 6 mice, P = 0.4487, t31 = 0.7674, paired Student's t-test).
Pericyte KATP channels tune blood flow to local metabolism
27
Supplementary figure 3. Direct stimulation of a pericyte with vehicle (aCSF) does not dilate PAs
or increase blood flow. (A) 1-s kymograph segments showing raw RBC flux of a >5th order
capillary at baseline, and after aCSF was ejected onto the overlying pericyte. (B) Representative
time course showing no change PA diameter after direct stimulation of a pericyte with aCSF. (C)
Summary data showing PA diameter before and after ejection of aCSF on a pericyte (n = 5 paired
measurements, 4 mice, P = 0.6317, t4 = 0.5181, paired Student's t-test). (D) Summary data of
RBC flux before and after aCSF-ejection on a pericyte (n = 8 paired measurements, 4 mice, P =
0.1108, t7 = 1.825, paired Student's t-test).
Pericyte KATP channels tune blood flow to local metabolism
28
Supplementary figure 4. Ejection of pinacidil on a PA does not affect its diameter (A) Focal
stimulation of a PA with 10 μM pinacidil. Left: PA diameter at baseline indicated by pink line, and
an ejection pipette containing 10 μM pinacidil positioned next to the PA. Middle: Ejection of
pinacidil directly onto the PA. Right: PA diameter 10 s after pinacidil ejection compared to control
diameter, indicated by pink line. (B) Summary data showing no change in PA diameter after direct
stimulation with pinacidil (n = 5 paired measurements, 4 mice, P = 0.5946, t4 = 0.5774, paired
Student's t-test).
Pericyte KATP channels tune blood flow to local metabolism
29
Supplementary figure 5. Example traces of Vm measurements with 2 mM bath glucose (A), 1
mM bath glucose (B), 750 μM bath glucose (C), 250 μM bath glucose (D), 750 μM bath glucose
with 10 μM glibenclamide (E), and 250 μM bath glucose in the presence of 10 μM glibenclamide
(F). (G) Summary data showing glibenclamide blocks hyperpolarizing effects of 750 μM and 250
μM glucose (750 μM glucose (20 cells, 4 mice) vs. 750 μM glucose + 10 μM glibenclamide (7
cells, 4 mice): **P = 0.0094, t97 = 3.343; 250 μM glucose (16 cells, 4 mice) vs. 250 μM glucose +
10 μM glibenclamide (9 cells, 4 mice): ****P < 0.0001, t97 = 5.009; One-way ANOVA with Sidak's
multiple comparison test).
Supplementary movie 1. Movie depicting imaging area for an experiment in which a deep
capillary pericyte was targeted by pressure-ejection of 10 μM pinacidil 170 μm downstream of the
imaging site focused on a PA with pre-capillary sphincter and 1st order capillary. Within seconds
of remote application of pinacidil, the PA, sphincter and 1st order capillary robustly dilate. Scale
bars on Z projections and single-plane recording are 50 and 5 μm, respectively.
| 2022 | Brain Capillary Pericytes are Metabolic Sentinels that Control Blood Flow through K Channel Activity | 10.1101/2022.03.14.484304 | [
"Hariharan Ashwini",
"Robertson Colin D.",
"Garcia Daniela C.G.",
"Longden Thomas A."
] | creative-commons |
Priestley, Baber, et al.
Page 1 of 23
Pan-cancer whole genome analyses of metastatic solid tumors
Peter Priestley1,2,*,#, Jonathan Baber1,2,*, Martijn P. Lolkema3,4, Neeltje Steeghs3,5, Ewart de Bruijn1,
Korneel Duyvesteyn1, Susan Haidari1,3, Arne van Hoeck6, Wendy Onstenk1,3,4, Paul Roepman1,
Charles Shale2, Mircea Voda1, Haiko J. Bloemendal7, Vivianne C.G. Tjan-Heijnen8, Carla M.L. van
Herpen9, Mariette Labots10, Petronella O. Witteveen11, Egbert F. Smit3,5, Stefan Sleijfer3,4, Emile E.
Voest3,5, Edwin Cuppen1,3,6,#
1 Hartwig Medical Foundation, Science Park 408, Amsterdam, The Netherlands
2 Hartwig Medical Foundation Australia, Sydney, Australia
3 Center for Personalized Cancer Treatment, The Netherlands
4 Erasmus MC Cancer Institute, Doctor Molewaterplein 40, Rotterdam, The Netherlands
5 Netherlands Cancer Institute/Antoni van Leeuwenhoekhuis, Plesmanlaan 121, Amsterdam, The Netherlands
6 Center for Molecular Medicine and Oncode Institute, University Medical Center Utrecht, Heidelberglaan 100,
Utrecht, The Netherlands
7 Meander Medisch Centrum, Maatweg 3, Amersfoort, The Netherlands
8 Maastricht University Medical Center, P. Debyelaan 25, Maastricht, The Netherlands
9 Radboud University Medical Center, Geert Grooteplein Zuid 10, Nijmegen, The Netherlands
10 VU Medical Center, De Boelelaan 1117, Amsterdam, The Netherlands
11 Cancer Center, University Medical Center Utrecht, Heidelberglaan 100, Utrecht, The Netherlands
* shared first author
# corresponding authors: p.priestley@hartwigmedicalfoundation.nl, e.cuppen@hartwigmedicalfoundation.nl
Abstract
Metastatic cancer is one of the major causes of death and is associated with poor
treatment efficiency. A better understanding of the characteristics of late stage cancer is
required to help tailor personalised treatment, reduce overtreatment and improve outcomes.
Here we describe the largest pan-cancer study of metastatic solid tumor genomes, including
2,520 whole genome-sequenced tumor-normal pairs, analyzed at a median depth of 106x and
38x respectively, and surveying over 70 million somatic variants. Metastatic lesions were
found to be very diverse, with mutation characteristics reflecting those of the primary tumor
types, although with high rates of whole genome duplication events (56%). Metastatic lesions
are relatively homogeneous with the vast majority (96%) of driver mutations being clonal and
up to 80% of tumor suppressor genes bi-allelically inactivated through different mutational
mechanisms. For 62% of all patients, genetic variants that may be associated with outcome of
approved or experimental therapies were detected. These actionable events were distributed
over the various mutation types (single and multiple nucleotide variants, insertions and
deletions, copy number alterations and structural variants) underlining the importance of
comprehensive genomic tumor profiling for cancer precision medicine for advanced cancer
treatment.
Introduction
Metastatic cancer is one of the leading causes of death globally and is a major burden for
society despite the availability of an increasing number of (targeted) drugs. Health care costs
associated with treatment of metastatic disease are increasing rapidly due to the high cost of novel
targeted treatments and immunotherapy, while many patients do not benefit from these approaches
with inevitable adverse effects for most patients. Metastatic cancer therefore poses a major challenge
for society to balance between individual and societal treatment responsibilities. Since cancer
Priestley, Baber, et al.
Page 2 of 23
genomes evolve over time, both in the highly heterogeneous primary tumor mass and as
disseminated metastatic cells1,2, a better understanding of metastatic cancer genomes is crucial to
further improve on tailoring treatment for late stage cancers.
In recent years, several large-scale whole genome sequencing (WGS) analysis efforts such
as TCGA and ICGC have yielded valuable insights in the diversity of the molecular processes driving
different types of adult3,4 and pediatric5,6 cancer and have fueled the promises of genome-driven
oncology care7. However, these analyses were primarily done on primary tumor material whereas
metastatic cancers, which cause the bulk of the disease burden and 90% of all cancer deaths, have
been less comprehensively studied at the whole genome level, with previous efforts focusing on
tumor-specific cohorts8–10 or at a targeted gene panel11 or exome level12.
Here we describe the first large-scale pan-cancer whole-genome landscape of metastatic
cancers based on the Hartwig Medical Foundation (HMF) cohort of 2,520 paired tumor and normal
genomes from 2,405 patients. The samples have been collected prospectively as fresh frozen
biopsies taken from a broad range of metastases (Extended Data Fig. 1) and blood controls from
patients with advanced cancer in a clinical study setup coordinated by the Center for Personalized
Cancer Treatment (CPCT) in 41 hospitals in the Netherlands (Supplementary Table 1). All samples
were paired with standardized clinical information (Supplementary Table 2). The sample distribution
over age and primary tumor types broadly reflects solid cancer incidence in the Western world,
including rare cancers (Fig. 1a-b).
The cohort has been analyzed with uniform and high depth paired-end (2 x 150 bp) whole
genome sequencing with a median depth of 106x for tumor samples and 38x for the blood control
(Extended Data Fig. 1). Sequencing data were analyzed for all types of somatic variants using an
optimized bioinformatic pipeline based on open source tools (Methods). We identified a total of
59,472,629 single nucleotide variants (SNVs), 839,126 multiple nucleotide variants (MNVs),
9,598,205 insertions and deletions (INDELs) and 653,452 structural variants (SVs) (Supplementary
Table 2). We found that the relative high sequencing depth is important for variant calling sensitivity
as downsampling of the tumor sample coverage to ~53x resulted in an average decrease in sensitivity
of 10% for SNV, 2% for INDEL, 15% for MNV, and 19% for SV (Extended Data Fig. 2).
Here we present a first characterization of this unique and comprehensive resource for a
better genomic understanding of advanced cancer.
Mutational landscape of metastatic cancer
We analysed the tumor mutational burden (TMB) of each class of variants per cancer type
based on tissue of origin (Fig. 1c-h, Supplementary Table 2). In line with previous studies on primary
cancers13, we found extensive variation in mutational load of up to 3 orders of magnitude both within
and across cancer types.
The median SNV counts per sample were highest in skin, predominantly consisting of
melanoma (44k) and lung (36k) tumors with ten-fold higher SNV counts than sarcomas (4.1k),
neuroendocrine tumors (NET) (3.5k) and mesotheliomas (3.4k). The variation for MNVs was even
greater with lung (median=815) and skin (median=764) tumors having five times the median MNV
counts of any other tumor type. This can be explained by the well-known mutational impact of UV
radiation (CC->TT MNV) and smoking (CC->AA MNV) mutational signatures, respectively (Fig. 1f).
Although only di-nucleotide substitutions are typically reported as MNVs, 10.7% of the MNVs involve
three nucleotides and 0.6% had four or more nucleotides affected.
INDEL counts were typically ten-fold lower than SNVs, with a lower relative rate for skin and
lung cancers (Fig. 1d, Extended Data Fig. 3). Genome-wide analysis of INDELs at microsatellite loci
identified 60 samples with microsatellite instability (MSI) (Supplementary Table 2), representing 2.4%
of all tumors. The highest rates of MSI were observed in central nervous system (CNS) (9.4%), uterus
(9.0%) and prostate (6.1%) tumors. For metastatic colorectal cancer lesions we found an MSI
frequency of only 4.0%, which is lower than reported for primary colorectal cancer, and in line with
better prognosis for patients with localized MSI colorectal cancer, which less often metastasizes14.
Priestley, Baber, et al.
Page 3 of 23
Remarkably, 67% of all INDELs in the entire cohort were found in the 60 MSI samples, and 85% of all
INDELs in the cohort were found in microsatellites or short tandem repeats. Only 0.33% of INDELs
(32k, ~1% of non-microsatellite INDELs) were found in coding sequences, of which the majority (88%)
had a predicted high impact by affecting the open reading frame of the gene.
The median rate of SVs across the cohort was 193 per tumor, with the highest median counts
observed in ovary (415) and esophageal (379) tumors, and the lowest in kidney tumors (71) and NET
(56) (Fig. 1h, Supplementary Table 2). Simple deletions were the most commonly observed SV
subtype (33% of all SVs) and were the most prevalent in every cancer type except stomach and
esophageal tumors which were highly enriched in translocations.
Copy number alteration landscape of metastatic cancer
Copy number alterations (CNAs) are important hallmarks of tumorigenesis15. Pan-cancer, the
most highly amplified regions in our metastatic cancer cohort contain the established oncogenes such
as EGFR, CCNE1, CCND1 and MDM2 (Fig. 2). Chromosomal arms 1q, 5p, 8q and 20q are also
highly enriched in moderate amplification across the cohort each affecting >20% of all samples. For
the amplifications of 5p and 8q this is likely related to the common amplification targets of TERT and
MYC, respectively. However, the targets of the amplifications on 1q, predominantly found in breast
cancers (>50% of samples) and amplifications on 20q, predominantly found in colorectal cancers
(>65% of samples), are less clear.
We identified some intriguing patterns of recurrent loss of heterozygosity (LOH) caused by
CNAs. Overall an average of 23% of the autosomal DNA per tumor has LOH. Unsurprisingly, TP53
has the highest LOH recurrence at 67% of samples. Many of the other LOH peaks are also explained
by well-known tumor suppressor genes (TSG). However, several clear LOH peaks are observed
which cannot easily be explained by known TSG selection, such as one on 8p (57% of samples). 8p
LOH has previously been linked to lipid metabolism and drug response16, although involvement of
individual genes has not been established. Alternatively, 8p LOH could potentially be the result of the
mechanism by which the amplification of 8q, the most commonly amplified part of the genome, is
established.
There are remarkable differences in LOH between cancer types (Fig. 2, Extended Data Fig.
4). For instance, we observed LOH events on the 3p arm in 90% of kidney samples17 and LOH of the
complete chromosome 10 in 72% of CNS tumors (predominantly glioblastoma multiforme18). Even in
the case of the TP53 region on chromosome 17, different tumor types display clearly different
patterns of LOH. Ovarian cancers exhibit LOH of the full chromosome 17 in 75% of samples whereas
in prostate cancer, which also has 70% LOH for TP53, this is nearly always caused by highly focal
deletions.
Unlike LOH events, homozygous deletions are nearly always restricted to small chromosomal
regions. Not a single example was found in which a complete autosomal arm was homozygously
deleted. Homozygous deletions of genes are surprisingly rare as well: we found only 4,915 autosomal
events (mean = 2.0 per tumor) where one or multiple consecutive genes are fully or partially
homozygously deleted. In 46% of these events a putative TSG was deleted. The scarcity of
passenger homozygous deletions underlines the fact that despite widespread copy number
alterations in metastatic tumors, the vast majority of genes or gross chromosomal organization likely
remain essential for tumor cell survival. Chromosome Y loss, which has been described anecdotally
for various tumor types19,20, is a special case and is deleted in 36% of all male tumor genomes but
varies strongly between tumor types from 5% to 68% for CNS and biliary tumors respectively
(Extended Data Fig. 5).
An extreme form of copy number change can be caused by whole genome duplication
(WGD). We found WGD events in 56% of all samples ranging between 17% in CNS to 80% in
esophageal tumors (Fig. 2d,e). This is much higher than reported previously for primary tumors (25%-
37%)21,22 and also higher than estimated from panel-based sequencing analyses of advanced tumors
(30%)23. Ploidy levels, in combination with accurate tumor purity information, are essential for correct
Priestley, Baber, et al.
Page 4 of 23
interpretation of the measured raw SNV and INDEL frequencies, e.g. to discriminate bi-allelic
inactivation of TSG from heterozygous events which are more likely to be passengers or to determine
(sub)clonality. Hence determining the WGD status of a tumor is highly relevant for diagnostic
applications. Furthermore, WGD has previously been found to correlate with a greater incidence of
cancer recurrence for ovarian cancer22 and has been associated with poor prognosis across cancer
types, independently of established clinical prognostic factors23.
Significantly mutated genes
To identify significantly mutated genes (SMGs) potentially specific for metastatic cancer, we
used the dNdScv approach24 with strict cutoffs (q<0.01) for the pan-cancer and tumor-type specific
datasets. In addition to reproducing previous results on cancer drivers, a few novel genes were
identified (Extended Data Fig. 6, Supplementary Table 3). In the pan-cancer analyses we found only a
single novel SMG, which was not either present in the curated COSMIC Cancer Gene Census or
found by Martincorena et al24. This gene, MLK4 (q = 2e-4), is a mixed lineage kinase that regulates
the JNK,P38 and ERK signaling pathways and has been reported to inhibit tumorigenesis in colorectal
cancer25. In addition, in our tumor type-specific analyses, which for several tumor types is limited by
the number of samples, we identified a novel metastatic breast cancer-specific SMG - ZFPM1 (also
known as Friend of GATA1 (FOG1), q = 8e-5), a zinc-finger transcription factor protein without clear
links with cancer. Nonetheless, we found six unique frameshift variants (all in a context of biallelic
inactivation) and three nonsense variants, which suggests a driver role for this gene in metastatic
breast cancer.
Our cohort also lends support to some prior SMG findings. In particular, eight significantly
mutated putative TSG in the HMF cohort were also found by Martincorena et al24 - GPS2 (pan-cancer,
q=1e-5 & breast, q=2e-3), SOX9 (colorectal & pan-cancer, q=0), TGIF1 (pan-cancer, q=3e-3 &
colorectal q=6e-3), ZFP36L1 (urinary tract q=3e-4, pan-cancer q=9e-4) and ZFP36L2 (colorectal &
pan-cancer, q=0), HLA-B (lymphoid, q=5e-5), MGA (pan-cancer, q=4e-03), KMT2B (skin, q=3e-3) and
RARG (urinary tract 8e-4). None of these genes are currently included in the COSMIC Cancer Gene
Census26. ZFP36L1 and ZFP36L2 are of particular interest as these genes are zinc-finger proteins
that normally play a repressive regulatory role in cell proliferation, presumably through a cyclin D
dependent and p53 independent pathway27. ZFP36L2 is also independently found as a significantly
deleted gene in our cohort, most prominently in colorectal and prostate cancers.
We also searched for genes that were significantly amplified or deleted (Supplementary Table
4). CDKN2A and PTEN were the most significantly deleted genes overall, but many of the top genes
involved common fragile sites (CFS) particularly FHIT and DMD, deleted in 5% and 4% of samples,
respectively. The role of CFS in tumorigenesis is unclear and aberrations affecting these genes are
frequently treated as passenger mutations reflecting localized genomic instability28. However, the
uneven distribution of the deletions across cancer types may indicate that some of these could be
genuine tumor-type specific cancer drivers. For example, we find deletions in DMD to be highly
enriched in esophageal tumors (deleted in 38% of samples, whilst SV burden in this tumortype is only
about 2-fold higher than average), GIST (Gastro-Intestinal Stromal Tumors; 24%) and pancreatic
neuroendocrine tumors (panNET; 41%), which is consistent with a recent study that indicated DMD as
a TSG in cancers with myogenic programs29. However, tissue type-specific gene expression and
differences in origins of replication may also contribute to the observed patterns28. We also identified
several significantly deleted genes not reported previously, including MLLT4 (13 samples) and
PARD3 (9 samples).
Unlike homozygous deletions, amplification peaks tend to be broad and often encompass
large number of genes, making identification of the amplification target challenging. However, SRY-
related high-mobility group box 4 gene (SOX4, 6p22.3) stands out as a significantly amplified single
gene peak (26 amplifications) and is highly enriched in urinary tract cancers (19% of samples highly
amplified) (Extended Data Fig. 4). SOX4 is known to be over-expressed in prostate, hepatocellular,
Priestley, Baber, et al.
Page 5 of 23
lung, bladder and medulloblastoma cancers with poor prognostic features and advanced disease
status and is a modulator of the PI3K/Akt signaling30.
Also notable was a broad amplification peak of 10 genes around ZMIZ1 at 10q22.3 (32
samples) which has not previously been reported. ZMIZ1 is a member of the Protein Inhibitor of
Activated STAT (PIAS)-like family of coregulators and is a direct and selective cofactor of Notch1 in T-
cell development and leukemia31. CDX2, previously identified as an amplified lineage-survival
oncogene in colorectal cancer32, is also highly amplified in our cohort with 20 out of 22 amplified
samples found in colorectal cancer, representing 5.4% of all colorectal samples.
Driver mutation catalog
We created a comprehensive catalog of all cancer driver mutations across all samples in our
cohort and all variant classes similar as described previously in primary tumors33 (N. Lopez, personal
communication). To do this, we combined our SMG discovery efforts with those from Martincorena et
al.24 and a panel of well known cancer genes (Cosmic Curated Genes)34, and added gene fusions,
TERT promoter mutations and germline predisposition variants found in our cohort. Accounting for the
proportion of SNV and INDELs estimated by dNdScv to be passengers, we found 13,423 somatic
driver events among the 20,125 identified mutations in the combined gene panel (Supplementary
table 5) together with 189 germline predisposition variants (Supplementary table 6). The somatic
drivers include 7,423 coding mutation, 615 non-coding point mutation drivers, 2,715 homozygous
deletions (25% of which are in common fragile sites), 2,393 focal amplifications and 277 fusion
events.
For the cohort as a whole, 55% of point mutations in the gene panel driver catalog were
predicted to be genuine driver events. To facilitate analysis of variants of unknown significance (VUS)
at a per patient level, we calculated a sample-specific likelihood for each point mutation to be a driver
taking into account the TMB of the sample as well as the biallelic inactivation status of the gene for
TSG and hotspot positions in oncogenes. Predictions of pathogenic variant overlap with known
biology, e.g. clustering of benign missense variants in the 3’ half of the APC gene (Extended Data Fig.
7b) fits with the absence of FAP-causing germline variants in this part of the gene35.
Overall, the catalog matches previous inventories of cancer drivers. TP53 (52% of samples),
CDKN2A (21%), APC (16%), PIK3CA (16%), KRAS (14%), PTEN (13%) and TERT (12%) were the
most common driver genes together making up 25% of all the driver mutations in the catalog (Fig. 3).
However, all of the ten most prevalent driver genes in our cohort were reported at a higher rate than
for primary cancers36, which may reflect the more advanced disease state. AR and ESR1 in particular
are more prevalent, with driver mutations in 44% of prostate and 18% of breast cancers, respectively.
Both genes are linked to resistance to hormonal therapy, a common treatment for these tumor types,
and have been previously reported as enriched in advanced metastatic cancer11 but are identified at
higher rates in this study.
Looking at a per patient level, the mean number of total driver events per patient was 5.7, with
the highest rate in urinary tract tumors (mean rate = 8.0) and the lowest in NET (mean rate = 2.8)
(Fig. 4). Esophageal and stomach tumors also had elevated driver counts, largely due to a much
higher rate of deletions in CFS genes (mean rate = 1.6 for stomach, 1.7 for esophageal) compared to
other cancer types (pan-cancer mean rate = 0.3). Fragile sites aside, the differential rates of drivers
between cancer types in each variant class do correlate with the relative mutational load, with the
exception of skin cancers, which have a lower than expected number of SNV drivers (Extended Data
Fig. 3f).
In 98.6% of all samples at least one somatic driver mutation or germline predisposition variant
was found. Of the 34 samples with no identified driver, 18 were NET of the small intestine
(representing 49% of all patients of this subtype). This likely indicates that small intestine NETs have
a distinct set of drivers that are not captured yet in any of the cancer gene resources used and are
also not prevalent enough in our relatively small NET cohort to be detected as significant.
Alternatively, NET tumors could be mainly driven by epigenetic mechanisms not detected by WGS37.
Priestley, Baber, et al.
Page 6 of 23
The number of amplified driver genes varied significantly between cancer types with highly
elevated rates per sample in breast cancer (mean = 2.1), esophageal, urinary tract and stomach (all
mean = 1.7) cancers and nearly no amplification drivers in kidney cancer (mean = 0.1) and none in
the mesothelioma cohort (Extended Data Fig. 8a). In tumor types with high rates of amplifications,
these amplifications are generally found across a broad spectrum of oncogenes (Extended Data Fig.
8b), suggesting there are mutagenic processes active in these tissues that favor amplifications, rather
than tissue-specific selection of individual driver genes. AR and EGFR are notable exceptions, with
highly selective amplifications in prostate, and in CNS and lung cancers, respectively, in line with
previous reports18,38,39. Intriguingly, we also found two-fold more amplification drivers in samples with
WGD (Extended Data Fig. 8c) despite amplifications being defined as relative to the average sample
ploidy.
The 189 germline variants identified in 29 cancer predisposition genes (present in 7.9% of the
cohort) consisted of 8 deletions and 181 point mutations (Fig. 3c, Supplementary Table 6). The top 5
affected genes were the well-known germline drivers CHEK2, BRCA2, MUTYH, BRCA1 and ATM,
and together contain nearly 80% of the observed predisposition variants (Fig 3c). The corresponding
wild type alleles were found to be lost in the tumor sample in more than half of the cases, either by
LOH or somatic point mutation, indicating a high penetrance for these variants, particularly in BRCA1
(89% of cases), APC (83%) and BRCA2 (79%).
The 277 fusions consisted of 168 in-frame coding fusions, 91 cis-activating fusions involving
repositioning of regulatory elements in 5’ genic regions, and 18 in-frame intragenic deletions where
one or more exons were deleted (Supplementary table 7). ERG (89 samples), BRAF(17 samples),
ERBB4 (16 samples), ALK(12 samples), NRG1(9 samples) and ETV4 (7 samples) were the most
commonly observed 3’ partners together making up more than half of the fusions. 77 of the 89 ERG
fusions were TMPRSS2-ERG affecting 38% of all prostate cancer samples in the cohort. 146 fusion
pairs were not previously recorded in CGI, OncoKb, COSMIC or CIViC34,40–42. A novel recurrent
KMT2A-BCOR fusion was observed in 2 samples (sarcoma and stomach cancer) and there were also
3 recurrent novel localized fusions resulting from adjacent gene pairs: YWHAE-CRK (2 samples),
FGFR2-ATE1 (2 samples) and BCR-GNAZ (2 samples).
Only promoter mutations in TERT were included in the study due to the current lack of robust
evidence for other recurrent oncogenic non-coding mutations43. A total of 257 variants were found at 5
known recurrent variant hotspots11 and included in the driver catalog.
Oncogene hotspots and novel activating variants
We found that the 70% of somatic driver mutations in oncogenes occur at or within 5
nucleotides of already known pathogenic mutational hotspots (Extended Data Fig. 7a). In the six most
prevalent oncogenes (KRAS, PIK3CA, BRAF, NRAS, CTNNB1 & ESR1) the rate was 96% (Fig. 5).
Furthermore, in many of the key oncogenes, we document several likely activating but non-canonical
variants near known mutational hotspots (Fig. 5). For example, we found activating MNVs in the well
known BRAF V600 hotspot (22 cases), but also novel non-hotspot MNVs in KRAS (8 cases) and
NRAS (4 cases) (Extended Data Fig 7b).
In-frame indels were even more striking, since despite being exceptionally rare overall (mean
= 1.7 per sample), we found an excess in known oncogenes including PIK3CA (19 cases), ERBB2
(10 cases) and BRAF(8 cases) frequently occurring at or near known hotspots44. Notably, many of the
in-frame indels are enriched in specific tumor types. For instance, all 18 KIT in-frame indels were
found in sarcomas, 6 out of 8 MUC6 in-frame indels in esophageal tumors, and 6 of 10 ERBB2 in-
frame indels in lung tumors. Finally, we identified 10 in-frame indels in FOXA1, which are highly
enriched in prostate cancer (7 of 10 cases) and clustered in two locations that were not previously
associated with pathogenic mutations45.
In CTNNB1 we identified an interesting novel recurrent in-frame deletion of the complete exon
3 in 12 samples, 9 of which are colorectal cancers. Surprisingly, these deletions were homozygous
Priestley, Baber, et al.
Page 7 of 23
but suspected to be activating as CTNNB1 normally acts as an oncogene in the WNT/beta-catenin
pathway and none of these nine colorectal samples had any APC driver mutations.
Biallelic tumor suppressor gene inactivation
Our results strongly support the Knudson two-hit hypothesis46 for tumor suppressor genes
with 80% of all TSG drivers explained by biallelic inactivation by genetic alterations (i.e. either by
homozygous deletion (32%), multiple somatic point mutations (7%), or a point mutation in
combination with LOH (41%)). This rate is the highest observed in any large-scale cancer WGS study.
For many key tumor suppressor genes the biallelic inactivation rate is almost 100% (more specifically:
TP53 (93%), CDKN2A (97%), RB1 (94%), PTEN (92%) and SMAD4 (96%); Fig. 3b), suggesting that
biallelic inactivation of these genes is a strict requirement for metastatic cancer.
Other prominent TSGs, however, have lower biallelic rates, including ARID1A (55%), KMT2C
(49%) and ATM (49%). It is unclear whether we systematically missed the second hit in these cases,
as this could potentially be mediated through non-mutational epigenetic mechanisms47, or if these
genes impact on tumorigenesis via a haploinsufficiency mechanism48.
Clonal and subclonal variants
To obtain insight into ongoing tumor evolution dynamics, we examined the clonality of all
variants. Surprisingly, only 6.5% of all SNV, MNV & INDELs across the cohort and just 3.7% of the
driver point mutations were found to be subclonal (Fig. 6). The low proportion of samples with
subclonal variants could be partially due to the detection limits of the sequencing approach
(sequencing depth, bioinformatic analysis settings), particularly for low purity samples. However, even
for samples with purities higher than 80% the total proportion of subclonal variants only reaches
10.2% (Fig. 6b). Furthermore, sensitized detection of variants at hotspot positions in cancer genes
showed that our analysis pipeline detected over 96% of variants with allele frequencies of > 3%.
Although the cohort contains some samples with (very) high fractions of subclonal variants, overall the
metastatic tumor samples are relatively homogeneous without the presence of multiple diverged
subclones. Low intratumor heterogeneity may be in part attributed to the fact that nearly all biopsies
were obtained by a core needle biopsy, which results in highly localized sampling, but is nevertheless
much lower compared to previous observations in primary cancers2.
In the 111 patients with independently collected repeat biopsies from the same patient
(Supplementary Table 8) we found 11% of all SNVs to be subclonal. Whilst 76% of clonal variants
were shared between biopsies, less than 30% of the subclonal variants were shared.
While we can not exclude the presence of larger amounts of lower frequency subclonal
variants, the low rate of high-frequency subclonal variants taken together with the observation that a
very high proportion of subclonal variants are private to a local metastasis, suggest a model where
individual metastatic lesions are dominated by a single clone at any one point in time and that more
limited tumor evolution and subclonal selection takes places after distant metastatic seeding. This
contrasts with observations in primary tumors, where larger degrees of subclonality and multiple
major subclones are more frequently observed2,49, but supports other recent studies which
demonstrate minimal driver gene heterogeneity in metastases8,50.
Co-occurrence of Drivers
We examined the pairwise co-occurrence of driver gene mutations per cancer type and found
ten combinations of genes that were significantly mutually exclusively mutated, and ten combinations
of genes which were significantly co-occurrently mutated (excluding pairs of genes on the same
chromosome which are frequently co-amplified or co-deleted) (Fig. 7). The 20 significant findings
include previously reported co-occurrence of mutated DAX|MEN1 in pancreatic NET (q=0.0007), and
CDH1|SPOP in prostate tumors (q=0.0005), as well as negative associations of mutated genes within
the same signal transduction pathway such as KRAS|BRAF (q=4e-4) and KRAS|NRAS (q=0.009) in
Priestley, Baber, et al.
Page 8 of 23
colorectal cancer, BRAF|NRAS in skin cancer (q=6e-12), CDKN2A|RB1 in lung cancer (q=8e-5) and
APC|CTNNB1 in colorectal cancer (q=8e-6). APC is also strongly negatively correlated with both
BRAF (q=1e-4) and RNF43 (q=2e-5) which together are characteristic of the serrated molecular
subtype of colorectal cancers51. We also found that SMAD2|SMAD3 are highly positively correlated in
colorectal cancer (q=0.02), mirroring a result reported previously in a large cohort of colorectal
cancers52.
In breast cancer, we found a number of significant novel relationships, including a positive
relationship for GATA3|VMP1(q=1e-4) and FOXA1|PIK3CA (q=2e-3), and negative relationships for
ESR1|TP53 (q=9e-4) and GATA3|TP53 (q=2e-3), which will need further validation and experimental
follow-up to understand underlying biology.
Actionability
We analyzed opportunities for biomarker-based treatment for all patients by mapping driver
events to three clinical annotation databases: CGI42, CIViC40 and OncoKB41. In 1,485 patients (62%)
at least one ‘actionable’ event was identified (Supplementary Table 9). Whilst these numbers are in
line with results from primary tumors33, longitudinal studies will be required to conclude if genomic
analyses for therapeutic guidance should be repeated when a patient experiences progressive
disease. Half of the patients with an actionable event (31% of total) contained a biomarker with a
predicted sensitivity to a drug at level A (approved anti-cancer drugs) and lacked any known
resistance biomarkers for the same drug (Fig. 8a). In 13% of patients the suggested therapy was a
registered indication, while in 18% of cases it was outside the labeled indication. In a further 31% of
patients a level B (experimental therapy) biomarker was identified. The actionable biomarkers
spanned all variant classes including 1,815 SNVs, 48 MNVs, 195 indels, 745 CNAs, 68 fusion genes
and 60 patients with microsatellite instability (Fig. 8b).
Tumor mutation burden is an important emerging biomarker for response to immune
checkpoint inhibitor therapy53 as it is a proxy for the amount of neo-antigens in the tumor cells. For
NSCLC it has been shown in at least 2 subgroup analyses of large phase III trials that both PFS and
OS are significantly improved with first line immunotherapy as compared to chemotherapy for patients
whose tumors have a TMB >10 mutations per Mb54,55. Although various clinical studies based on this
parameter are currently emerging, TMB was not yet included in the above actionability analysis.
However, when applying the same cut-off to all samples in our cohort, an overall 18% of patients
would qualify, varying from 0% for liver, mesothelioma and ovarian cancer patients to more than 50%
of lung and skin cancer patients (Extended Data Fig. 3b).
Discussion
Genomic testing of tumors faces numerous challenges in meeting clinical needs, including i)
the interpretation of variants of unknown significance (VUS), ii) the steadily expanding universe of
actionable genes, often with an increasingly small fraction of patients affected (e.g. NRG156 and
NTRK fusions57 in less than 2% of all patients), and iii) the development of advanced genome-derived
biomarkers such as tumor mutational load, DNA repair status and mutational signatures. Our results
demonstrate in several ways that WGS analyses of metastatic cancer can provide novel and relevant
insights and be instrumental in addressing some of these key challenges in cancer precision
medicine.
First, our systematic and large-scale pan-cancer analyses on metastatic cancer tissue
allowed for the identification of several novel (cancer type-specific) cancer drivers and mutation
hotspots. Second, the driver catalog analyses can be used to mitigate the problem of VUS
interpretation33 both by leveraging previously identified pathogenic mutations (accounting for more
than 2/3rds of oncogenic point-mutation drivers) and by careful analysis of the biallelic inactivation of
putative TSG which accounts for over 80% of TSG drivers in metastatic cancer. Third, we
demonstrate the importance of accounting for all types of variants, including large scale genomic
rearrangements (via fusions and copy number alteration events), which account for more than half of
Priestley, Baber, et al.
Page 9 of 23
all drivers, but also activating MNV and INDELs which we have shown are commonly found in many
key oncogenes. Fourth, we have shown that using WGS, even with very strict variant calling criteria,
we could find driver variants in more than 98% of all metastatic tumors, including putatively actionable
events in a clinical and experimental setting for up to 62% of patients.
Although we did not find metastatic tumor genomes to be fundamentally different to primary
tumors in terms of the mutational landscape or genes driving advanced tumorigenesis, we described
characteristics that could contribute to therapy responsiveness or resistance in individual patients. In
particular we showed that WGD is a more pervasive element of tumorigenesis than previously
understood affecting over half of all metastatic cancers. We also found metastatic lesions to be less
heterogeneous than reported in primary tumors, although the limited depth sequencing does not allow
for drawing conclusions regarding low-frequency subclonal variants.
It should be noted that differences between WGS cohorts should be interpreted with some
caution as inevitable differences between experimental and computational approaches may impact on
observations and can only be addressed in longitudinal studies including the different stages of
disease. Furthermore, the HMF cohort includes a mix of treatment-naive metastatic patients and
patients who have undergone (extensive) prior systemic treatments. While this may impact on specific
tumor characteristics, it also provides opportunities for studying treatment response and resistance as
this data is recorded in the studies.
Finally, we believe the resource described here is a valuable complementary resource to
comparable whole genome sequencing-based resources of primary tumors in advancing fundamental
and translational cancer research. Therefore, all non-privacy sensitive data is publicly available
through a local interface developed by ICGC58 (work in progress) and all other data is made freely
available for scientific research by a controlled access mechanism (see
www.hartwigmedicalfoundation.nl/en for details).
Acknowledgements
We thank the Hartwig Foundation and Barcode for Life for financial support of clinical studies and
WGS analyses. Development of the data portal was supported by a grant from KWF
Kankerbestrijding (HMF2017-8225, GENONCO). We are particularly grateful to all patients, nurses
and medical specialists for their essential contributions making this study possible. We would like to
specifically thank Hans van Snellenberg for operational management of the Hartwig Medical
Foundation. We would like to thank Stefan Willems, Wendy de Leng, Alexander Hoischen and
Winand Dinjens for support with pathology assessments and mutation validations and Jeroen de
Ridder, Wigard Kloosterman and Harmen van de Werken for critically reading the manuscript.
Priestley, Baber, et al.
Page 10 of 23
Figure Legends
Figure 1: Mutational load of metastatic cancer per tumor type
a) The number of samples of each tumor type cohort. Tumor types are ranked from lowest to highest
overall mutation burden (TMB)
b) Violin plot showing age distribution of each tumor type with 25th, 50th and 75th percentiles marked.
c)-d) cumulative distribution function plot (individual samples were ranked independently for each
panel) of mutational load for each tumor type for SNV and MNV (c) and INDEL and SV (d). The
median for each cohort is indicated with a vertical line.
e)-h) Mutational context or variant subtype per individual sample for each of (e) Single Nucleotide
Variant (SNV), (f) Multi Nucleotide Variant (MNV), (g) INsertion/DELetion (INDEL), (h) Structural
Variant (SV). Each column chart is ranked within tumor type by mutational load in that variant class.
MNVs are classified by the dinucleotide substitution with NN referring to any dinucleotide
combination. SVs are classified by type: INV = inversion, DEL = deletion, DUP = tandem duplication,
TRL = translocation, INS = insertion.
Figure 2: Copy number landscape of metastatic cancer
Proportion of samples with amplification and deletion events by genomic position per cohort - pan-
cancer (a), central nervous system (CNS) (b) and kidney (c). The inner ring shows the % of tumors
with homozygous deletion (orange), LOH and significant loss (copy number < 0.6x sample ploidy -
dark blue) and near copy neutral LOH (light blue). Outer ring shows % of tumors with high level
amplification (>3x sample ploidy - orange), moderate amplification (>2x sample ploidy - dark green)
and low level amplification (>1.4x amplification - light green). The scale on both rings is 0-100% and
inverted for the inner ring. The most frequently observed high-level gene amplifications (black text)
and homozygous deletions (red text) are shown.
d) Proportion of tumors with a whole genome duplication event (dark blue) grouped by tumor type.
e) Average sample ploidy distribution over the complete cohort. Samples with a WGD event (true) are
shown in darker blue.
Figure 3: Most prevalent driver genes in metastatic cancer
Most prevalent somatically mutated TSG (a) and oncogenes (b), and germline predisposition variants
(c) . From left to right, the heatmap shows the % of samples in each cancer type which are found to
have each gene mutated; absolute bar chart shows the pan-cancer % of samples with the given gene
mutated; relative bar chart shows the breakdown by type of alteration. For TSG, the % of samples
with a driver in which the gene is found biallelically inactivated, and for germline predisposition
variants the % of samples with loss of wild type in the tumor are also shown.
Figure 4: Drivers per sample by tumor type
a) Violin plot showing the distribution of the number of drivers per sample grouped by tumor type.
Black dots indicate the mean values for each tumor type.
b) Relative bar chart showing the breakdown per cancer type of the type of alteration.
Figure 5: Oncogenic Hotspots
Count of driver point mutations by variant type. Known pathogenic mutations curated from external
databases are categorized as hotspot mutations. Mutations within 5 bases of a known pathogenic
mutation are shown as near hotspot and all other mutations are shown as non-hotspot.
Figure 6: Subclonality
a) Count of samples per tumor purity bucket. b) Violin plot showing the percentage of point mutations
which are subclonal in each purity bucket. Black dots indicate the mean for each bucket. c)
Percentage of driver point mutations that are subclonal in each purity bucket.
Priestley, Baber, et al.
Page 11 of 23
Figure 7: Driver co-occurrence
a) Mutated driver gene pairs which are significantly positively (on the right) or negatively (on the left)
correlated in individual tumor types sorted by q-value. The color indicates the tumor type as depicted
below the chart.
Figure 8: Actionability
a) Percentage of samples in each cancer type with an actionable mutation based on data in CGI,
CIViC and OncoKB knowledgebases. Level ‘A’ represents presence of biomarkers with either an
approved therapy or guidelines and level B represents biomarkers having strong biological evidence
or clinical trials indicating they are actionable. On label indicates treatment registered by federal
authorities for that tumor type, while off-label indicates a registration for other tumor types.
b) Break down of the actionable variants by mutation type.
Extended Data Figures and Tables
Extended Data Figure 1: Hartwig sample workflow, biopsy locations and sequence coverage
a) Sample workflow from patient to high-quality WGS data. A total of 4,018 patients were enrolled in
the study between April 2016 and April 2018. For 9% of patients no blood and/or biopsy material was
obtained, mostly because conditions of patients prohibited further study participation. Up to 4 fresh-
frozen biopsies per patient were received, which were sequentially analyzed to identify a biopsy with
more than 30% tumor cellularity as determined by routine histology assessment. For 859 patients no
suitable biopsy was obtained and 2,796 patients were further processed for WGS. 44 and 29 samples
failed in either DNA isolation or library preparation and raw WGS data quality QC, respectively. For an
additional 385 samples the WGS data was of good quality, but the tumor purity determination based
on WGS data (PURPLE) was less than 20% making reliable and comprehensive somatic variant
calling and were therefore excluded. Eventually, 2,338 tumor-normal sample pairs with high-quality
WGS data were obtained, which were supplemented with 182 pairs from pre-April 2016, adding up to
2,520 tumor normal pairs that were included in this study.
b) Breakdown of cohort by biopsy location. Tumor biopsies were taken from a broad range of
locations. Primary tumor type is shown on the left and the biopsy location on the right.
c) Distribution of sample sequencing depth for tumor and blood reference.
Extended Data Figure 2: Impact of downsampling on variant calling
Comparison of variant calling of 10 randomly selected samples at normal depth and 50%
downsampled for purity (a), SNV counts (b), SV counts (c), Ploidy (d), MNV counts (e) and INDEL
counts (f). For the panels B, C, E and F, the black dots represent the % reduction per sample of
counts (right axis) and the dotted line represents the average % reduction across all tested samples.
Extended Data Figure 3: Mutational load, genome wide analyses and drivers
a) Proportion of samples by cancer type classified as microsatellite instable (MSISeq score > 4)
b) Proportion of samples with a high mutational burden (TMB > 10 SNV / Mb)
c)-e) Scatter plot of mutational load per sample for INDEL vs SNV (c), INDEL vs SV (d), and SV vs
SNV (e). MSI (MSISeq score > 4) and ‘high TMB’ (>10 SNV/ MB) thresholds are indicated.
f)-h) Mean mutational load vs driver rate for SNV (f), INDEL (g) and SV (h) grouped by cancer type.
MSI samples were excluded.
Extended Data Figure 4: Copy Number profile per cancer types
Circos plots showing the proportion of samples with amplification and deletion events by genomic
position per cancer type. The inner ring shows the % of tumors with homozygous deletion (red), LOH
Priestley, Baber, et al.
Page 12 of 23
and significant loss (copy number < 0.6x sample ploidy - dark blue) and near copy neutral LOH (light
blue). The outer ring shows the % of tumors with high level amplification (>3x sample ploidy - orange),
moderate amplification (>2x sample ploidy - dark green) and low level amplification (>1.4x
amplification - light green). Scales on both rings are 0-100% and inverted for the inner ring. The most
frequently observed high level gene amplifications (black text) and homozygous deletions (red text)
are labelled.
Extended Data Figure 5: Somatic Y chromosome Loss
Proportion of Male tumors with somatic loss of >50% of Y chromosome (dark blue) grouped by tumor
type.
Extended Data Figure 6: Significantly mutated genes
Tile chart showing genes found to be significantly mutated per cancer type cohort and pan-cancer
using dNdScv. Gene names marked in orange are also significant in Martincorena et al24, but not
found in COSMIC curated or census. Gene names marked in red are novel in this study.
Extended Data Figure 7: Coding mutation profiles by driver gene
Location and driver classification of all coding mutations (SNVs and indels) in oncogenes (a) and
tumor suppressor genes (TSG) (b) in the driver catalog. The lollipops on the chart show the location
(coding sequence coordinates) and count of mutations for all candidate drivers. The height of lollipop
represents the total count of each individual variant in the cohort (log scale). The height of the solid
line represents the sum of driver likelihoods for that variant, ie. the proportion that are expected to be
drivers. (Partially) dotted lines hence indicate variants for which driver role is uncertain. For TSG only,
variants are unshaded if all instances of that variant are monoallelic single hits with no LOH. The right
column chart shows the estimated number of drivers (calculated as the sum of driver likelihoods) and
passenger variants in each gene by cancer type.
Extended Data Figure 8: Amplifications
a) Mean rate of amplification drivers per cancer type. b) Breakdown of the number of amplification
drivers per gene by cancer type. c) Mean rate of drivers per variant type for samples with and without
WGD.
Supplementary Table 1: Overview of contributing organizations and local principal investigators.
Supplementary Table 2: Overview of cohort and sample characteristics
Supplementary Table 3: Pan-cancer and cancer type-specific dNdScv results
Supplementary Table 4: Recurring amplifications (a) and deletions (b) and associated target genes
Supplementary Table 5: Somatic driver catalog
Supplementary Table 6: Germline driver catalog
Supplementary Table 7: Gene Fusions
Supplementary Table 8: Overview of patients with multiple biopsies
Supplementary Table 9: Actionable mutations
Priestley, Baber, et al.
Page 13 of 23
References
1.
Klein, C. A. Selection and adaptation during metastatic cancer progression. Nature 501, 365–372
(2013).
2.
McGranahan, N. & Swanton, C. Clonal Heterogeneity and Tumor Evolution: Past, Present, and
the Future. Cell 168, 613–628 (2017).
3.
Cancer Genome Atlas Research, Network et al. The Cancer Genome Atlas Pan-Cancer analysis
project. Nat. Genet. 45, 1113–1120 (2013).
4.
International Cancer Genome, Consortium et al. International network of cancer genome
projects. Nature 464, 993–998 (2010).
5.
Grobner, S. N. et al. The landscape of genomic alterations across childhood cancers. Nature
555, 321-327 (2018).
6.
Ma, X. et al. Pan-cancer genome and transcriptome analyses of 1,699 paediatric leukaemias and
solid tumours. Nature 555, 371-376 (2018).
7.
Hyman, D. M., Taylor, B. S. & Baselga, J. Implementing Genome-Driven Oncology. Cell 168,
584–599 (2017).
8.
Yates, L. R. et al. Genomic Evolution of Breast Cancer Metastasis and Relapse. Cancer Cell 32,
169–184 e7 (2017).
9.
Naxerova, K. et al. Origins of lymphatic and distant metastases in human colorectal cancer.
Science 357, 55–60 (2017).
10. Gundem, G. et al. The evolutionary history of lethal metastatic prostate cancer. Nature 520, 353–
357 (2015).
11. Zehir, A. et al. Mutational landscape of metastatic cancer revealed from prospective clinical
sequencing of 10,000 patients. Nat. Med. 23, 703–713 (2017).
12. Robinson, D. R. et al. Integrative clinical genomics of metastatic cancer. Nature 548, 297–303
(2017).
13. Alexandrov, L. B. et al. Signatures of mutational processes in human cancer. Nature 500, 415–
421 (2013).
14. Gryfe, R. et al. Tumor microsatellite instability and clinical outcome in young patients with
colorectal cancer. N. Engl. J. Med. 342, 69–77 (2000).
15. Taylor, A. M. et al. Genomic and Functional Approaches to Understanding Cancer Aneuploidy.
Cancer Cell 33, 676–689.e3 (2018).
16. Cai, Y. et al. Loss of Chromosome 8p Governs Tumor Progression and Drug Response by
Altering Lipid Metabolism. Cancer Cell 29, 751–766 (2016).
17. Sato, Y. et al. Integrated molecular analysis of clear-cell renal cell carcinoma. Nat. Genet. 45,
860–867 (2013).
18. Brennan, C. W. et al. The somatic genomic landscape of glioblastoma. Cell 155, 462–477 (2013).
19. Hunter, S., Gramlich, T., Abbott, K. & Varma, V. Y chromosome loss in esophageal carcinoma:
an in situ hybridization study. Genes Chromosomes Cancer 8, 172–177 (1993).
20. Sauter, G. et al. Y chromosome loss detected by FISH in bladder cancer. Cancer Genet.
Cytogenet. 82, 163–169 (1995).
21. Zack, T. I. et al. Pan-cancer patterns of somatic copy number alteration. Nat. Genet. 45, 1134–
1140 (2013).
22. Carter, S. L. et al. Absolute quantification of somatic DNA alterations in human cancer. Nat.
Biotechnol. 30, 413–421 (2012).
23. Bielski, C. M. et al. Genome doubling shapes the evolution and prognosis of advanced cancers.
Nat. Genet. 50, 1189-1195 (2018).
24. Martincorena, I. et al. Universal Patterns of Selection in Cancer and Somatic Tissues. Cell 171,
1029–1041 e21 (2017).
25. Marusiak, A. A. et al. Recurrent MLK4 Loss-of-Function Mutations Suppress JNK Signaling to
Promote Colon Tumorigenesis. Cancer Res. 76, 724–735 (2016).
26. Forbes, S. A. et al. COSMIC: somatic cancer genetics at high-resolution. Nucleic Acids Res. 45,
D777–D783 (2017).
Priestley, Baber, et al.
Page 14 of 23
27. Suk, F.-M. et al. ZFP36L1 and ZFP36L2 inhibit cell proliferation in a cyclin D-dependent and p53-
independent manner. Sci. Rep. 8, 2742 (2018).
28. Glover, T. W., Wilson, T. E. & Arlt, M. F. Fragile sites in cancer: more than meets the eye. Nat.
Rev. Cancer 17, 489–501 (2017).
29. Wang, Y. et al. Dystrophin is a tumor suppressor in human cancers with myogenic programs.
Nat. Genet. 46, 601–606 (2014).
30. Mehta, G. A. et al. Amplification of SOX4 promotes PI3K/Akt signaling in human breast cancer.
Breast Cancer Res. Treat. 162, 439–450 (2017).
31. Pinnell, N. et al. The PIAS-like Coactivator Zmiz1 Is a Direct and Selective Cofactor of Notch1 in
T Cell Development and Leukemia. Immunity 43, 870–883 (2015).
32. Salari, K. et al. CDX2 is an amplified lineage-survival oncogene in colorectal cancer. Proc. Natl.
Acad. Sci. U. S. A. 109, E3196–205 (2012).
33. Sabarinathan, R. et al. The whole-genome panorama of cancer drivers. BioArchive (2017).
doi:10.1101/190330
34. Futreal, P. A. et al. A census of human cancer genes. Nat. Rev. Cancer 4, 177–183 (2004).
35. Friedl, W. et al. Can APC mutation analysis contribute to therapeutic decisions in familial
adenomatous polyposis? Experience from 680 FAP families. Gut 48, 515–521 (2001).
36. Bailey, M. H. et al. Comprehensive Characterization of Cancer Driver Genes and Mutations. Cell
173, 371–385.e18 (2018).
37. Cives, M., Simone, V., Rizzo, F. M. & Silvestris, F. NETs: organ-related epigenetic derangements
and potential clinical applications. Oncotarget 7, 57414–57429 (2016).
38. Viswanathan, S. R. et al. Structural Alterations Driving Castration-Resistant Prostate Cancer
Revealed by Linked-Read Genome Sequencing. Cell 174, 433-447 (2018).
39. Cancer Genome Atlas Research Network. Comprehensive genomic characterization of
squamous cell lung cancers. Nature 489, 519–525 (2012).
40. Griffith, M. et al. CIViC is a community knowledgebase for expert crowdsourcing the clinical
interpretation of variants in cancer. Nat. Genet. 49, 170–174 (2017).
41. Chakravarty, D. et al. OncoKB: A Precision Oncology Knowledge Base. JCO Precis Oncol 2017,
(2017).
42. Tamborero, D. et al. Cancer Genome Interpreter annotates the biological and clinical relevance
of tumor alterations. Genome Med. 10, 25 (2018).
43. Cuykendall, T. N., Rubin, M. A. & Khurana, E. Non-coding genetic variation in cancer. Current
Opinion in Systems Biology 1, 9–15 (2017).
44. Chang, M. T. et al. Accelerating Discovery of Functional Mutant Alleles in Cancer. Cancer Discov.
8, 174–183 (2018).
45. Yang, Y. A. & Yu, J. Current perspectives on FOXA1 regulation of androgen receptor signaling
and prostate cancer. Genes Dis 2, 144–151 (2015).
46. Knudson, A. G., Jr. Mutation and cancer: statistical study of retinoblastoma. Proc. Natl. Acad. Sci.
U. S. A. 68, 820–823 (1971).
47. Schlicker, A., Michaut, M., Rahman, R. & Wessels, L. F. A. OncoScape: Exploring the cancer
aberration landscape by genomic data fusion. Sci. Rep. 6, 28103 (2016).
48. Davoli, T. et al. Cumulative haploinsufficiency and triplosensitivity drive aneuploidy patterns and
shape the cancer genome. Cell 155, 948–962 (2013).
49. Andor, N. et al. Pan-cancer analysis of the extent and consequences of intratumor heterogeneity.
Nat. Med. 22, 105–113 (2016).
50. Reiter, J. G. et al. Minimal functional driver gene heterogeneity among untreated metastases.
Science 361, 1033–1037 (2018).
51. Bond, C. E. et al. RNF43 and ZNRF3 are commonly altered in serrated pathway colorectal
tumorigenesis. Oncotarget 7, 70589–70600 (2016).
52. Fleming, N. I. et al. SMAD2, SMAD3 and SMAD4 mutations in colorectal cancer. Cancer Res. 73,
725–735 (2013).
53. Goodman, A. M. et al. Tumor Mutational Burden as an Independent Predictor of Response to
Priestley, Baber, et al.
Page 15 of 23
Immunotherapy in Diverse Cancers. Mol. Cancer Ther. 16, 2598–2608 (2017).
54. Hellmann, M. D. et al. Nivolumab plus Ipilimumab in Lung Cancer with a High Tumor Mutational
Burden. N. Engl. J. Med. 378, 2093–2104 (2018).
55. Carbone, D. P. et al. First-Line Nivolumab in Stage IV or Recurrent Non-Small-Cell Lung Cancer.
N. Engl. J. Med. 376, 2415–2426 (2017).
56. Fernandez-Cuesta, L. & Thomas, R. K. Molecular Pathways: Targeting NRG1 Fusions in Lung
Cancer. Clin. Cancer Res. 21, 1989–1994 (2015).
57. Laetsch, T. W. et al. Larotrectinib for paediatric solid tumours harbouring NTRK gene fusions:
phase 1 results from a multicentre, open-label, phase 1/2 study. Lancet Oncol. 19, 705–714
(2018).
58. Zhang, J. et al. International Cancer Genome Consortium Data Portal--a one-stop shop for
cancer genomics data. Database 2011, bar026 (2011).
Priestley, Baber, et al.
Page 16 of 23
Figure 1
Priestley, Baber, et al.
Page 17 of 23
Figure 2
Priestley, Baber, et al.
Page 18 of 23
Figure 3
Priestley, Baber, et al.
Page 19 of 23
Figure 4
Priestley, Baber, et al.
Page 20 of 23
Figure 5
Priestley, Baber, et al.
Page 21 of 23
Figure 6
Priestley, Baber, et al.
Page 22 of 23
Figure 7
Priestley, Baber, et al.
Page 23 of 23
Figure 8
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 1 of 29
Detailed methods for
Pan-cancer whole genome analyses of metastatic solid tumors
Peter Priestley, Jonathan Baber, Martijn P. Lolkema, Neeltje Steeghs, Ewart de Bruijn, Korneel
Duyvesteyn, Susan Haidari, Arne van Hoeck, Wendy Onstenk, Paul Roepman, Charles Shale, Mircea
Voda, Haiko J. Bloemendal, Vivianne C.G. Tjan-Heijnen, Carla M.L. van Herpen, Mariette Labots,
Petronella O. Witteveen, Egbert F. Smit, Stefan Sleijfer, Emile E. Voest, Edwin Cuppen
Content
1. Sample collection
2
2. Sequencing workflow
2
3. Somatic point mutation calling
3
4. Validation of somatic point mutation calling
4
5. Somatic structural variant calling
6
6. Identification of gene fusions
6
7. Validation of gene fusions
7
8. Purity, ploidy and copy number calling
7
9. Validation of purity, ploidy and copy number output
11
10. Sample filtering based on copy number output
13
11. Impact of sequencing depth coverage on somatic variant calling sensitivity
13
12. Germline predisposition variant calling
13
13. Clonality and biallelic status of point mutations
15
14. WGD status determination
16
15. MSI status determination
17
16. Holistic gene panel for driver discovery
18
17. Significantly mutated driver genes discovery
18
18. Significantly amplified & deleted driver gene discovery
18
19. Fragile site annotation
20
20. Somatic driver catalog construction
21
21. Driver co-occurrence analysis
23
22. Actionability analysis
24
23. Data availability
27
24. References
28
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 2 of 29
1. Sample collection
Patients with advanced cancer not curable by local treatment options and being candidates for any type
of systemic treatment and any line of treatment were included as part of the CPCT-02 (NCT01855477)
and DRUP (NCT02925234) clinical studies, which were approved by the medical ethical committees
(METC) of the University Medical Center Utrecht and the Netherlands Cancer Institute, respectively. A
total of 41 academic, teaching and general hospitals across the Netherlands participated in these studies
and collected material and clinical data by standardized protocols1. Patients have given explicit consent
for whole genome sequencing and data sharing for cancer research purposes. Clinical data, including
primary tumor type, biopsy location, gender and birth year were collected in electronic case record forms
and stored in a central database.
Core needle biopsies were sampled from the metastatic lesion, or when considered not feasible or not
safe, from the primary tumor site when still in situ. One to four biopsies were collected (average of 2.1 per
patient) and frozen in liquid nitrogen directly after sampling and further processed at a central pathology
tissue facility. Frozen biopsies were mounted on a microtome in water droplets for optimal preservation of
all types of biomolecules (DNA, RNA and proteins) for subsequent and future omics-based analyses. A
single 6 micron section was collected for hematoxylin-eosin (HE) staining and estimation of tumor
cellularity by an experienced pathologist. Subsequently, 25 sections of 20 micron, containing an
estimated 25,000 to 500,000 cells, were collected in a tube for DNA isolation. In parallel, a tube of blood
was collected in CellSave (Menarini-Silicon Biosystems) tubes, which was shipped by room temperature
to the central sequencing facility at the Hartwig Medical Foundation. Left-over material (biopsy, DNA) after
sample processing was stored in biobanks associated with the studies at the University Medical Center
Utrecht and the Netherlands Cancer Institute.
2. Sequencing workflow
DNA was isolated from biopsy and blood on an automated setup (QiaSymphony) according to supplier's
protocols (Qiagen) using the DSP DNA Midi kit for blood and QIAsymphony DSP DNA Mini kit for tissue
and quantified (Qubit). Before starting DNA isolation from tissue, the biopsy was dissolved in 100
microliter Nuclease-free water by using the Qiagen TissueLyzer and split in two equal fractions for parallel
automated DNA and RNA isolation (QiaSymphony). Typically, DNA yield for the tissue biopsy ranged
between 50 and 5,000 ng. A total of 50 - 200 ng of DNA was used as input for TruSeq Nano LT library
preparation (Illumina), which was performed on an automated liquid handling platform (Beckman Coulter).
DNA was sheared using sonication (Covaris) to average fragment lengths of 450 nt. Barcoded libraries
were sequenced as pools (blood control 1 lane equivalent, tumor 3 lane equivalents) on HiSeqX (V2.5
reagents) generating 2 x 150 read pairs using standard settings (Illumina).
BCL output from the HiSeqX platform was converted using bcl2fastq tool (Illumina, versions 2.17 to 2.20
have been used) using default parameters. Reads were mapped to the reference genome GRCH37 using
BWA-mem v0.7.5a2, duplicates were marked for filtering and INDELs were realigned using GATK v3.4.46
IndelRealigner3. GATK HaplotypeCaller v3.4.464 was run to call germline variants in the reference
sample. For somatic SNV and INDEL variant calling, GATK BQSR5 is also applied to recalibrate base
qualities.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 3 of 29
3. Somatic point mutation calling
We called SNV & INDEL somatic variants using Strelka v1.0.146 with the following optimisations:
●
Preservation of known variants: From the raw Strelka output we marked all known pathogenic
variants from external databases such that these would be preserved from all subsequent
filtering. The list of pathogenic variants used was the union of:
○
Point mutations in CIViC7 with level C evidence or higher (download = 01-mar-2018)
○
Somatic variants from CGI8 (update: 17-jan-2018)
○
Oncogenic or likelyOncogenic variants from OncoKb9 (download = 01-mar-2018);
http://oncokb.org/api/v1/utils/allAnnotatedVariants.txt)
○
TERT promoter variants at genomic coordinates: 5:1295242, 5:1295228, 5:1295250
●
Modified quality score filtering
○
We split variants into high confidence (HC) and low confidence (LC) regions using the
NA12878 GIABv3.2.2 high confidence region definitions10, based on the observation that
we produce far higher rates of false positives variant calls in LC regions
○
Set quality score cutoffs for SNV & INDEL to 10 for HC regions and 20 for LC regions
(default = 15 for SNV, 30 for INDEL)
○
Added an additional quality filter to tighten filtering for low allelic frequency variants:
quality score * allele frequency > 1.3
●
Improved repeat sensitivity: Switched off the default Strelka repeat filter to improve indel calling
in microsatellites and short repeats.
●
Panel of normals (PON) to remove germline leakage: Filtered out any variants which were
found by GATK haplotypecaller in more than 5 samples in a germline PON consisting of 2000 of
our reference blood samples. PON available at (https://resources.hartwigmedicalfoundation.nl/)
●
PON to remove strelka-specific artefacts: Filtered any variant which was supported by 2 or
more reads in strelka in the reference sample in at least 4 patients in our cohort. PON available at
(https://resources.hartwigmedicalfoundation.nl/)
●
Removal of INDELS near a PON filtered INDEL - Regions of complex haplotype alterations are
often called as multiple long indels, which can make it more difficult to construct an effective
PON, and sometimes we find residual artefacts at these locations. Hence we also filter inserts or
deletes which are 3 bases or longer where there is a PON-filtered INDEL of 3 bases or longer
within 10 bases in the same sample.
●
MNV Correction - Variants occurring on consecutive positions, or 1 base apart were considered
potential multi nucleotide variants (MNVs). The BAM files were re-examined, and the variants
were merged into a single MNV if greater than 80% of the reads with a mapping quality score of
at least 10 and which are neither unmapped, duplicated, secondary, nor supplementary
containing any of the individual variants also contained the other variants of the potential MNV.
The attributes of the resulting MNV variant were determined by picking the minimum values from
the individual variants forming the MNV. MNVs were marked as PON filtered only if both
individual variants were PON filtered.
The settings and tools for this optimized HMF pipeline are available at https://github.com/hartwigmedical/.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 4 of 29
4. Validation of somatic point mutation calling
We performed three separate analyses to validate our somatic variant calling pipeline as follows:
4.1. Validation of somatic precision and sensitivity pipeline on a known benchmark
We tested the default Strelka and HMF optimized settings on a GIAB mix-in sample (ref = NA24385;
tumor = 70% NA24385 and 30% NA12878) to test sensitivity at a realistic purity and on a null tumor (ref =
NA12878, tumor = NA12878) to test precision. The results of this analysis are as follows:
Configuration
SNV sensitivity SNV false positive /
genome
INDEL sensitivity Indel false positive /
genome
Strelka default
93%
3500
24%
41
Optimized HMF pipeline
96%
109
77%
27
4.2. External independent validation of SNV and INDEL calling precision on real samples
We performed external validation of a set of single nucleotide variants (SNV) and short insertion/deletions
(indels) that have been detected by Whole Genome Sequencing (WGS) using the single molecule
Molecular Inversion Probe (smMIP) technology11. SNV and short indels variants were semi-randomly
selected from 30 patient samples. The first selection was to include every variant that was reported in a
panel of 114 ‘actionable’ cancer genes as used in the routine CPCT-02 study analysis. This way, a total
of 82 variants (67 SNVs, 15 indels) were selected in 45 genes. The second selection involved random
sampling adding up to a total of 256 coding and non-coding variants from the same 30 patient samples.
A custom smMIP panel was designed to cover the selected variants. For 45 variants (17.6%) no smMIP
design was possible, all of which were intergenic variants. For the other 211 variants probes could
successfully be designed. Analysis of the smMIP sequencing data indicated that for 17 of the 211 variants
(8.1%) the smMIP sequencing data was of insufficient quality (mostly due to repeat stretches), while the
WGS data seemed sufficiently reliable for accurate calling (confirmed by visual inspection of the read
data), including 3 coding variants (RB1, ERBB4 and BRCA2) and 14 intergenic regions. The retrospective
investigation of the WGS data indicated that for another three variants (1.4%) the smMIP as well as the
WGS data was of insufficient quality due to large homopolymer stretches.
In total 192 variants could be successfully sequenced and analyzed using the smMIP and could be used
for confirmation of the WGS findings. 189 SNVs and indel variants (98.4%) were confirmed by smMIP
sequencing, indicating a very high accuracy of WGS-derived variant calling results. All three variants that
could not be confirmed by smMIP were from intergenic regions, including 1 variant that showed a mixed
double-variant (chr3:75887550_G>T/C) and for which both technologies had difficulties in accurately
calling the genotype. For the remaining 2 variants (chr8:106533360_106533361insAC,
chr12:125662751_125662752insA), it remains unclear if these could not be detected by smMIP or were
falsely called by WGS, as they fall in repetitive genomic stretches.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 5 of 29
The 189 successfully confirmed variants showed a good linear correlation in variant allele frequency
between WGS and smMIP sequencing (average of duplicates) with an R2 of 0.733. This result indicated
that WGS, with its lower read depth (on average between 100-110x) than smMIP and without a read-
barcoding system, is accurate in quantitatively determining the variant frequency at frequencies above
5%. One variant (ch19:55276095C>T, indicated in red in the figure above) showed a large deviation in
variant frequency, which was likely due to the much lower than expected coverage of the variant, both in
the WGS (37 reads) as well as in the smMIP data (28 and 35 reads).
4.3. Validation of somatic variant calling sensitivity by reanalysis of known hotspots.
To validate somatic calling sensitivity and performance limitations of our pipeline on real samples, we built
a customised tool, SAGE (https://github.com/hartwigmedical/hmftools/tree/master/sage) to reanalyse all
10,211 known pathogenic hotspot variants in the coding region of the genome (sourced from CIVIC,
OncoKb and CGI as described above). These locations have a much higher prior likelihood of finding a
variant in cancer samples.
SAGE searches for each hotspot in the tumor BAM files directly and calls a variant if the sum of read
base qualities supporting the ALT > 100, effectively equating to 3 high quality reads of support. Our
standard somatic pipeline typically requires 6 or more reads support to call a variant. For the purposes of
this validation we excluded from SAGE a small number of variants in high repeat contexts (repeat count
>=8) and in regions with very high tumor copy number (tumor read depth > 300) as both these contexts
can cause low VAF artefacts which we want to avoid in a sensitivity test.
We evaluated on a randomly selected 1247 samples with the following results
Hotspot variants found in
standard somatic pipeline
Additional variants
found by SAGE
% variants missed by
somatic pipeline
1160
37
3.1%
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 6 of 29
Of the 37 additional variants found by SAGE but not in our standard somatic pipeline, 27 (2.3%) were
found to have been missed by Strelka due to low read count in the tumor (all with only 3 to 6 reads
supporting the ALT allele), 8 (0.7%) due to insufficient coverage in the reference sample, and 2 (0.2%) for
unknown reason.
Overall this analysis suggests that we capture more than 96% of all variants with 3 or more reads of
support in the tumor (equivalent to ~3% VAF).
5. Somatic structural variant calling
Structural Variants were called using Manta(v1.0.3)12 with default parameters. We then re-examined each
breakpoint, calculated variant allele frequencies for each break end and applied seven additional filters to
the Manta output to improve precision using an internally built tool called ‘Breakpoint-Inspector’ (BPI,
https://github.com/hartwigmedical/hmftools/tree/master/break-point-inspector) v1.5. Two main types of
filters are applied by BPI:
●
Evidence of variant in reference sample - Variants are filtered out if we can find any evidence
of paired read support, split read support or soft clipping concordance (5+ bases at exact
breakpoint) in the matching blood sample.
●
Inadequate support for variant in tumor sample - For all inversions and translocations and for
long deletions and tandem duplications (>1000 bases between breakpoints) we require at least 1
read with paired read support. For short deletions and duplications (<1000 bases between
breakpoints) we require at least 1 read with split read support. In both cases at least one of those
reads must be anchored with at least 30 bases at each breakpoint. We also require the minimum
read coverage across each breakpoint in the tumor to be > 10 depth.
Each breakend was annotated with it’s position in all transcripts from ‘KNOWN’ genes in Ensembl
v89.3713. Each gene was marked as disrupted if there was at least one structural variant that impacted on
the canonical transcript.
6. Identification of gene fusions
For each structural variant, every combination of annotated overlapping transcripts from each breakend
was tested to see if it could potentially form an intronic inframe fusion. A list of 411 curated known fusion
pairs was sourced by taking the union of known fusions from the following external databases:
●
Cosmic curated fusions14 (v83)
●
OncoKb9 (download = 01-mar-2018)
●
CGI8 (update: 17-jan-2018)
●
CIViC7 (download = 01-mar-2018)
We then also created a list of promiscuous fusion partners using the following rules
●
3’ promiscuous: Any gene which appears on the 3’ side in more than 3 of the curated fusion
pairs OR appears at least once on the 3’ side and is marked as promiscuous in either OncoKb,
CGI or CIVIC
●
5’ promiscuous: Any gene which appears on the 5’ side in more than 3 of the curated fusion
pairs OR appears at least once on the 5’ side and is marked as promiscuous in either OncoKb,
CGI or CIVIC
For each promiscuous partner we also curated a list of essential domains that must be preserved to form
a viable fusion partner.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 7 of 29
Finally, we report an intronic inframe fusion if the following conditions are met
●
Matches an exact fusion from the curated list OR is intergenic and matches 5’ promiscuous OR
matches 3’ promiscuous gene
●
Curated domains are preserved
●
Does not involve the 3’UTR region of either gene
●
For intragenic fusions, must start and end in coding regions of the gene
●
3’ partner is a protein coding gene and the transcript does not result in nonsense mediated decay
7. Validation of gene fusions
Whole transcriptome analysis (RNA-seq) of 60 samples with identified fusions was used to validate our
gene fusion calling pipeline.
RNA was isolated from the same biopsy material as used for DNA isolation using an automated setup
(QiaSymphony) using the QIAsymphony RNA kit (#931636, Qiagen) according to supplier's protocols.
RNA was quantified using Qubit RNA HS Assay Kit (Thermo Fisher). Typically, RNA yield for the tissue
biopsy ranged between 500 and 5,000 ng. 100 ng of total RNA was used as input for KAPA RNA
HyperPrep Kit with RiboErase (HMR) (#KR1351, Roche) and TruSeq DNA CD Indexes 96 Indexes (#PN
20015949, Illumina) performed on an automated liquid handling platform (Beckman Coulter). The
standard protocol used involved 240 sec 85 degrees Celcius fragmentation and 15 PCR cycles. Each
sample was subsequently sequenced in a multiplexed setup with 2x75 bp reads on a NextSeq 500/550
using the High Output Kit v2 (Illumina, #FC-404-2002), targeting 50M raw reads per sample. BCL output
from the NextSeq500 platform was converted using Illumina bcl2fastq tool (versions 2.17 to 2.20 have
been used) using default parameters.
STAR-Fusion15 was used with default settings to call fusion transcripts from the RNA. 38 out of 60 fusions
were readily identified independently in the RNA. Manual inspection of the expected chimeric junctions for
the remaining 22 fusions revealed RNA support for a further 6 fusions (4 of which were TMPRSS2-ERG),
although below the threshold to be called automatically in the RNA with the settings used. Overall, 73% of
the tested fusions were thus independently validated by the RNA analysis. The full results are
summarised below:
Total fusions tested
Transcript fusion found by
STAR-Fusion
Read support in
RNA but not called
by STAR-Fusion
No evidence of fusion
transcript in RNA
60
38 (63%)
6 (10%)
16 (27%)
8. Purity, ploidy and copy number calling
Accurate copy number calling is closely linked with correct sample purity determination. Currently, there is
not a clear consensus in the community for a preferred tool for this purpose. We tested several tools
(freeC, CANVAS and Sequenza) on the COLO829 benchmark, but none of them provided a correct fit16.
Therefore we developed PURPLE (PURity & PLoidy Estimator) as an alternative.
PURPLE combines B-allele frequency (BAF), read depth and structural variants to estimate the purity and
copy number profile of a tumor sample and follows a similar purity fitting methodology to several other
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 8 of 29
popular tools such as ASCAT, Sequenza and CANVAS, only with a different optimisation function to
determine the best fit.
The main advantages of PURPLE (v2.14) for the purposes of this study are:
●
extensive attention to removal of artefacts by filtering of inputs (see below sections 7.1, 7.2 and
7.3) and smoothing of output to avoid false positive copy number calling (section 7.5)
●
integrated SV and copy number calling allow single base accuracy of copy number calls and
accurately call each individual variant as heterozygous or homozygous as well as the detection of
partial loss of genes
There are five key steps in the PURPLE pipeline:
1. Calculate BAF in tumor at high confidence heterozygous germline loci
We determine the BAF of the tumor sample by finding heterozygous locations in the reference sample
from a panel of 796,447 common germline heterozygous SNP locations. To ensure that we only capture
heterozygous points, we filter the panel to only loci with allelic frequencies in the reference sample
between 40% and 60% and with depth between 50% and 150% of the reference sample genome wide
average. Typically, this yields 140k-200k heterozygous germline variants per patient. We then calculate
the allelic frequency of corresponding locations in the tumor.
2. Determine read depth ratios for tumor and reference genomes
The raw read counts per 1,000 base window for both normal and tumor samples, by counting the number
of alignment starts in the respective bam files with a mapping quality score of at least 10 that is neither
unmapped, duplicated, secondary, nor supplementary. Windows with a GC content less than 0.2 or
greater than 0.6 or with an average mappability below 0.85 are excluded from further analysis.
Next we apply a GC normalization to calculate the read ratios. We divide the read count of each window
by the median read count of all windows sharing the same GC content then normalise further to the ratio
of the median to mean read count of all windows.
Finally, the reference sample ratios have a further ‘diploid’ normalization applied to them to remove
megabase scale GC biases. This normalization assumes that the median ratio of each 10Mb window
(minimum 1Mb readable) should be diploid for autosomes and haploid for sex chromosomes in males in
the germline sample.
3. Segmentation
We segment the genome into regions of uniform copy number by combining segments generated from
the read ratios for both tumor and reference sample, from the BAF points with structural variant
breakpoints derived from Manta & BPI. Read ratios and BAF points are segmented independently using
the Bioconductor copynumber package17 which uses a piecewise constant fit (PCF) algorithm (with
custom settings gamma = 100, k =1). These segment breaks are then combined with the structural
variants breaks according to the following rules:
1. Every structural variant break starts a new segment, as does chromosome starts, ends and
centromeres. This is regardless of if they are distinguishable from existing segments or not.
2. Ratio and BAF segment breaks are only included if they are distinguishable from an existing
segment.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 9 of 29
3. To be distinguishable, a break must be at least one complete mappable read depth window away
from an existing segment.
Once the segments have been established we map our observations to them. In each segment we take
the median BAF of the tumor sample and the median read ratio of the tumor and reference samples. We
also record the number of BAF points within the segment as the BAFCount.
A reference sample copy number status is determined at this this stage based on the observed copy
number ratio in the reference sample, either ‘DIPLOID’ (0.8<= read depth ratio<=1.2),
‘HETEROZYGOUS_DELETION’ (0.1<=ratio<0.8), ‘HOMOZYGOUS_DELETION’
(ratio<0.1),’AMPLIFICATION’(1.2<ratio<=2.2)or ‘NOISE’ (ratio>2.2). The purity fitting and smoothing
steps below use only the DIPLOID germline segments.
4. Purity Fitting
Next we jointly fit tumor purity and sample ploidy (expressed as a normalisation factor) according to the
following principles:
1. The absolute copy number of each segment should be close to an integer ploidy
2. The BAF of each segment should be close to a % implied by integer major and minor allele
ploidies.
3. Higher ploidies have more degenerate fits but are less biologically plausible and should be
penalised
4. Segments are weighted by the count of BAF observations which is treated as a proxy for
confidence of BAF and read depth ratio inputs.
5. Segments with lower observed BAFs have more degenerate fits and are weighted less in the fit
For any given tumor purity and sample ploidy we calculate the score by first modelling the major and
minor allele ploidy of each segment and minimising the deviation between the observed and modelled
values according to the following formulas:
ModelDeviation = abs(ObservedRatio - ModelRatio) + abs(ObservedBaf - ModelBaf)
ModelBaf = (tumorPurity * (segmentMinorPloidy - 1) + 1) / (tumorPurity * (segmentPloidy - 2) + 2)
ModelRatio = sampleNormFactor + (segmentPloidy - 2) * tumorPurity * sampleNormFactor / 2d;
Once modelled, each segment is given a ploidy penalty:
PloidyPenalty = 1 +min(SingleEventDistance, WholeGenomeDoublingDistance);
WholeGenomeDoublingDistance = 1 + abs(segmentMajorAllele - 2) +abs(segmentMinorAllele - 2);
SingleEventDistance = abs(segmentMajorAllele - 1) + abs(segmentMinorAllele - 1);
Summing up over all the segments generates a score for each tumor purity / sample ploidy combination
from which we can select the minimum:
𝐹𝑖𝑡𝑒𝑑𝑃𝑢𝑟𝑖𝑡𝑦𝑆𝑐𝑜𝑟𝑒
=
1
𝑇𝑜𝑡𝑎𝑙𝐵𝑎𝑓𝐶𝑜𝑢𝑛𝑡 7
8
9 : ;
𝑃𝑙𝑜𝑖𝑑𝑦𝑃𝑒𝑛𝑎𝑙𝑡𝑦
9 × 𝑀𝑜𝑑𝑒𝑙𝐷𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛
9 × 𝐵𝑎𝑓𝐶𝑜𝑢𝑛𝑡
9
× 𝑂𝑏𝑠𝑒𝑟𝑣𝑒𝑑𝐵𝑎𝑓
9
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 10 of 29
If a sample has a fitted purity solution which is >98.5% diploid and a score within 10% of the best fitted
score, the sample is designated as highly diploid and a fit is determined by the highest vaf somatic ploidy
peak.
Given a fitted purity and sample ploidy we are then able to determine the purity adjusted copy number
and BAF of each segment in the tumor genome from the unadjusted read ratios and BAFs respectively.
5. Smoothing
Since the segmentation algorithm is highly sensitive, and there is a significant amount of noise in the read
depth in whole genome sequencing, many adjacent segments created above will have a similar copy
number and BAF profile and can be combined and averaged to form a larger, smoothed, region.
We apply a number of rules to merge adjacent regions to create a smooth copy number profile.
1. Never merge a segment break created from a structural variant break end.
2. Use the count of BAF points as a proxy for confidence or weight in the region. Note that some
segments may have a BAF count of 0.
3. Merge segments where the difference in BAF and copy number is within tolerances.
4. BAF tolerance is linear between 0.03 and 0.35 dependent on BAF count.
5. Copy number tolerance is linear between 0.3 and 0.7 dependent on BAF count. The tolerance
also increases linearly as purity of the tumor sample decreases below 20%.
6. Start from most confident segment and smooth outwards until we reach a segment outside of
tolerance. Move on to next highest unsmoothed section.
7. It is possible to merge in (multiple) segments that would otherwise be outside of tolerances if:
a. The total dubious region is sufficiently small (<30k bases or <50k bases if approaching
centromere); and
b. The dubious region does not end because of a structural variant; and
c. The dubious region ends at a centromere, telomere or a segment that is within
tolerances.
8. When the entire short arm of a chromosome is lacking copy number information (generally on
chromosome 13, 14, 15, 21 or 22), the copy number of the long arm is extended to the short arm.
9. Any remaining unknown segments are given the expected copy number of their associated
chromosome, i.e. 2 for autosomes and female allosomes, 1 for male allosomes.
Where clusters of SVs exist which are closer together than our read depth ratio window resolution of
1,000 bases, the segments in between will not have any copy number information associated with them.
To resolve this, we infer the ploidy from the surrounding copy number regions. The outermost segment of
any SV cluster will be associated with a structural variant with a ploidy that can be determined from the
adjacent copy number region and the VAF of the SV. We use this ploidy and the orientation of structural
variant to calculate the change in copy number across the SV and hence the copy number of the
outermost unknown segment. We repeat this process iteratively and infer the copy number of all regions
within a cluster.
Once region smoothing is complete, it is possible there will be regions of unknown BAF, if no BAF points
were present in a copy number region. We infer this BAF by assuming that they share their minor allele
ploidy with their neighbouring region. If there are multiple neighbouring regions with known BAF we use
the highest confident region (i.e. highest BAF count) to infer.
At this stage we have determined a copy number and minor allele ploidy for every base in the genome.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 11 of 29
9. Validation of purity, ploidy and copy number output
We performed three validations to evaluate the purity and ploidy estimates and copy number profile
obtained from PURPLE.
1. Validation of purity estimates through cell line in-silico dilutions
The purity estimates of PURPLE were validated using the tumor cell line COLO829. We created diluted
in-silico mixture models of the tumor and blood cell lines from COLO829 with simulated purities of 20%,
30%, 40%, 60%, 80% and 100%, and ran PURPLE on the simulated BAM files against the reference
sample.
The PURPLE estimates were found to match the simulation very closely as shown in the table below:
Simulated
Purity
PURPLE estimated purity
Difference
20%
20%
0%
30%
30%
0%
40%
40%
0%
50%
50%
0%
60%
60%
0%
80%
81%
1%
100%
100%
0%
2. Validation of absolute copy number predictions by FISH
We also validated the absolute copy number results for PURPLE by comparing the WGS analysis results
of the COLO-829 tumor vs normal cell line pair with DNA Fluorescence In Situ Hybridization (FISH)
results for the centromeric region of chromosome 9, 13, 16 and 18 (CEP9, CEP13, CEP16, CEP18) and
for the 2p23 ALK locus and the 9p24 JAK2 locus. In total, 100 COLO829 tumor cells were scored for
each of the six FISH probes. For both assays the local copy-number as well as the percentage of DNA
(PURPLE) or number of cells (FISH) is provided in the table below to indicate the intratumoral
heterogeneity. The FISH and sequencing based results showed a very high concordance for the
chromosomal copy numbers and the intratumoral heterogeneity (COLO-829 cell line heterogeneity has
been described previously18).
Genomic region
PURPLE ploidy and purity
FISH copy number
Centromere Chr 9
3.7-4.0 : 53-57%
2n : 33%
3n : 9%
4n : 58%
Centromere Chr 13
3.2 : 55%
2n : 41%
3n : 59%
Centromere Chr 16
2.0 : 100%
2n : 100%
Centromere Chr 18
2.8-2.9 : 67-71%
2n : 38%
3n : 62%
ALK (2p23)
3.1 : 67%
2n : 21%
3n : 79%
JAK2 (9p24)
2.0 : 100%
2n : 100%
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 12 of 29
3. Comparison of PURPLE purity and ploidy estimates on patient samples with ASCAT
To validate PURPLE on real patient data, we compared the purity and ploidy outputs from PURPLE to the
widely used copy number tool ASCAT19 for 65 randomly selected samples from our cohort. ASCAT was
run on GC corrected data using default parameters except for gamma which was set to 1 which is
recommended for massively parallel sequencing data.
The following charts show a comparison of ASCAT vs PURPLE purity and ploidy results with 55 of 65
samples (85%) in agreement to within 10% absolute purity and relative sample ploidy.
There are 2 types of differences observed in the remaining 10 samples:
●
Purity differences for highly diploid samples - this is unsurprising as PURPLE has additional
functionality which is not dependent on copy number alterations in the tumor for highly diploid
samples to fit the somatic ploidies whereas ASCAT does not.
●
Whole genome duplication (WGD) vs no whole genome duplication - In 5 of the samples ASCAT
calls a WGD event whereas PURPLE does not and in 2 samples the opposite occurs. This
reflects the tradeoff in the purity and ploidy determination between penalising higher ploidy
solutions which are more degenerate vs lower ploidy solutions with more subclonality. Manual
inspection of purity-corrected fitted minor allele ploidy plots reveals in all of the 5 cases where
ASCAT calls a WGD that whilst there is subclonality in each of these cases in the PURPLE
solution there is no subclonal peak at 0.5 copy number, nor is there a 0.5 somatic ploidy peak,
suggesting that the the WGD solution is less likely. Conversely, in the 2 cases where PURPLE
only calls a WGD, manual inspection shows that the ASCAT solution would be prefered in one
case and the PURPLE solution in the other.
In summary, overall concordance is very high between PURPLE and ASCAT. There appears to be little
systematic bias to either calling lower or higher ploidy solutions between methods, and where PURPLE
differs from ASCAT it more often than not appears to be the more plausible solution.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 13 of 29
10. Sample filtering based on copy number output
Following our copy number calling, samples were QC filtered from the analysis based on 4 criteria:
●
NO_TUMOR - If PURPLE fails to find any aneuploidy AND the number of somatic SNVs found is
less than 1,000 then the sample is marked as NO_TUMOR.
●
MIN_PURITY - We exclude samples with a fitted purity of <20%
●
FAIL_SEGMENT - We remove samples with more than 120 copy number segments unsupported
at either end by SV breakpoints. This step was added to remove samples with extreme GC bias,
with differences in depth of up to or in excess of 10x between high and low GC regions. GC
normalisation is unreliable when the corrections are so extreme so we filter.
●
FAIL_DELETED_GENES - We removed any samples with more than 280 deleted genes. This
QC step was added after observing that in a handful of samples with high MB scale positive GC
bias we sometimes systematically underestimate the copy number in high GC regions. This can
lead us to incorrectly infer homozygous loss of entire chromosomes, particularly on chromosome
19.
Where multiple biopsies exist for a single patient, we always choose the highest purity sample for our
analysis of mutational load and drivers.
11. Impact of sequencing depth coverage on somatic variant calling sensitivity
To assess the impact of our sequencing depth on variant calling sensitivity, we selected 10 samples at
random, downsampled the BAMs by 50%. We then reran the identical somatic variant calling pipeline.
Comparing the output to the original runs, we found near identical purities and ploidies for the down
sampled runs (Extended Data Fig. 2). We observed an average decrease in sensitivity of 10% for SNV,
15% for MNV, 19% for SV, and 2% for INDEL.
The relatively small drop in indel calling sensitivity upon downsampling is caused by hard-coded setting in
STRELKA. Strelka has a hard cutoff at 10% VAF for INDELs of less than 5 bases length (which is 99% of
INDELs in our dataset) for both 50x and 100x depth whereas for SNVs the cutoff is fixed at ~5 supporting
reads independent of read depth. This likely results in underestimation of subclonal INDELs in our dataset
but does not affect specificity.
12. Germline predisposition variant calling
We searched for germline variants in a broad list of 152 germline predisposition genes curated by Huang
et al20. For SNV and INDEL, using the germline variant calling outputs from the GATK HaplotypeCaller4,
we filtered for variants affecting the canonical transcript of these 152 genes which have the following
coding or splice effects:
●
All SNV Nonsense, INDEL Frameshift or SNV Splice Acceptor/Donor, excluding variants marked
in ClinVar21 as 'Benign/Likely_benign', 'Benign', 'Likely_benign’.
●
Missense and synonymous variants, only if marked in clinvar as ‘Pathogenic’ or ‘Likely
Pathogenic’, excluding pathogenic disease indications which are clearly unrelated to cancer.
Variants which were found with a median germline VAF across all samples of less than 0.2 or greater
than 0.8 were filtered as likely mapping artefacts. We further excluded frameshift variants which are found
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 14 of 29
to be exactly offset by other frameshift variants (thereby creating an in-frame protein product), which
actually involved more than 50% of samples in which such events occur.
This yielded 550 potential germline predisposition point mutations across the 2,405 samples in our cohort.
For each variant, we determined the genotype in the germline (HET or HOM) and also assessed in the
tumor sample whether there is a 2nd somatic hit, and whether the wild type or the variant itself has been
lost (see chapter 13: biallelic status evaluation methods). We also searched in the 152 genes for copy
number deletions that were heterozygous in the germline with subsequent homozygous loss in the tumor
and found an additional 16 of such germline copy number events, giving a total of 566 variants altogether.
We observed that for the variants in many of the 152 predisposition genes that a loss of wild type in the
tumor via LOH was lower than the average rate of LOH across the cohort and that fewer than 5% of
observed variants had a 2nd somatic hit in the same gene. Moreover, in many of these genes the ALT
variant was lost via LOH as frequently as the wild type, suggesting that a significant portion of the 566
variants may be passengers. For our downstream analysis and driver catalog, we therefore restricted our
analysis to a more conservative ‘High Confidence’ list including only the 25 cancer related genes in the
ACMG secondary findings reporting guidelines (v2.0)22, together with 4 curated genes (CDKN2A, CHEK2,
BAP1 & ATM), selected because these are the only additional genes from the larger list of 152 genes with
a significantly elevated proportion of called germline variants with loss of wild type in the tumor sample.
The following table summarises the statistics for the high confidence and low confidence genes:
Genes
Total germline
predisposition SNV &
INDEL
% with loss of wild type
OR somatic hit in
tumor
% with loss of germline
ALT variant in tumor
High Confidence: ACMG +
4 curated genes
211
53.1%
10.4%
Low Confidence: Rest of
152 panel
355
16.3%
13.1%
Outside the 29 high confidence genes, the germline variant itself is lost almost as frequently via LOH as
the remaining wild type in the tumor, whereas for the high confidence ACMG + curated genes, there is an
observed loss of wild type allele in over half of all variants.
For the additional 4 curated genes, the numbers are as follows:
Gene
Count germline
predisposition SNV &
INDEL
% with loss of wild
type
in tumor sample
% with loss of germline
variant in tumor
sample
ATM
17
52.9%
11.8%
BAP1
5
66.7%
0%
CHEK2
72
36.1%
13.9%
CDKN2A
3
66.7%
33.3%
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 15 of 29
Germline variants with loss of ALT variants in the tumor were also excluded from the final list used in our
analyses, leading to a final inclusion of 189 variants from the high confidence panel.
Supplementary Table 6 contains the full catalog of high and low confidence germline variants.
13. Clonality and biallelic status of point mutations
For each point mutation we determined the clonality and biallelic status by comparing the estimated
ploidy of the variant to the local copy number at the exact base of the variant. The ploidy of each variant
is calculated by adjusting the observed VAF by the purity and then multiplying by the local copy number
to work out the absolute number of chromatids that contain the variant.
We mark a mutation as biallelic (i.e. no wild type remaining) if Variant Ploidy > Local Copy Number - 0.5.
The 0.5 tolerance is used to allow for the binomial distribution of VAF measurements for each variant. For
example, if the local copy number is 2 than any somatic variant with measured ploidy > 1.5 is marked as
biallelic.
For each variant we also determine a probability that it is subclonal. This is achieved via a two-step
process
1. Fit the somatic ploidies for each sample into a set of clonal and subclonal peaks
We apply an iterative algorithm to find peaks in the ploidy distribution:
●
Determine the peak by finding the highest density of variants within +/- 0.1 of every 0.01 ploidy
bucket.
●
Sample the variants within a 0.05 ploidy range around the peak.
●
For each sampled variant, use a binomial distribution to estimate the likelihood that the variant
would appear in all other 0.05 ploidy buckets.
●
Sum the expected variants from the peak across all ploidy buckets and subtract from the
distribution.
●
Repeat the process with the next peak
This process yields a set of ploidy peaks, each with a ploidy and a total density (i.e. count of variants). To
avoid overfitting small amounts of noise in the distribution, we filter out any peaks that account for less
than 40% of the variants in the ploidy bucket at the peak itself. After this filtering we scale the fitted peaks
by a constant so that the sum of fitted peaks = the total variant count of the sample.
Finally we mark a peak as subclonal if the peak ploidy < 0.85.
2. Calculate the probability that each individual variant belongs to each peak
Once we have fitted the somatic ploidy peaks and determined their clonality, we can calculate the
subclonal likelihood for any individual variant as the proportion of subclonal variants at that same ploidy.
The following diagram illustrates this process for a typical sample. Figure A shows the histogram of
somatic ploidy for all SNV and INDEL in blue. Superimposed are four peaks in different colours fitted from
the sample as described above. The red filled peak is below the 0.85 threshold and is thus considered
subclonal. The black line shows the overall fitted ploidy distribution. Figure B shows the likelihood of a
variant being subclonal at any given ploidy.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 16 of 29
Subclonal counts in this paper are calculated as the total density of the subclonal peaks for each sample.
Subclonal driver counts are calculated as the sum across the driver catalog of subclonal probability *
driver likelihood (driver likelihood is explained in detail in chapter 20).
14. WGD status determination
We implement a simple heuristic that determines if Whole Genome Duplication has occurred:
Major allele Ploidy >1.5 on at least 50% of at least 11 autosomes
The principle behind this heuristic is that if sufficient independent chromosomes are predominantly
duplicated, the most parsimonious explanation is that the duplication occurred in a single genome-wide
event.
The number of duplicated autosomes per sample (ie. the number of autosomes which satisfy the above
rule) follows a bimodal distribution with 95% of samples have either <= 6 or > =15 autosomes duplicated.
Hence, the classification of a genome as WGD is not particularly sensitive to the choice of cut-off as is
evident the following chart:
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 17 of 29
15. MSI status determination
To determine the MSI status of all samples we used the method described by the MSISeq tool23. In brief,
we count the number of INDELS per million bases occuring in homopolymers of 5 or more bases or
dinucleotide, trinucleotide and tetranucleotide sequences of repeat count 4 or more. MSIseq scores
ranged from 0.004 up to 98.63, with a long tail towards lower MSI scores as shown in the following chart:
To be able to accurately set and validate the MSIseq cutoff for classification of MSI we compared the
WGS results with the standard, routinely used MSI assessment using a 5-marker PCR panel (BAT25,
BAT26, NR21, NR24 and MONO27 markers). For a batch of 48 pre-selected samples, the MSI PCR
assay was blindly performed by an independent ISO-accredited pathology laboratory. Both the binary MSI
and MSS classifications were determined, but also the number of positive markers.
A sample was considered as MSI if two or more of the five markers were score as positive (instable).
PCR-based analysis identified 16 MSI samples, all of which were also identified by MSIseq with scores
>4. MSIseq identified one sample that was missed by PCR-based analysis, although this sample showed
microsatellite instability for one out the five markers. The MSIseq scores thus highly correlate with the
number of positive MSI PCR markers and all, except one, samples with an elevated score are classified
as MSI by pathology. Based on this data we determined the best cutoff for MSIseq classification to be at
a score of 4.
Results of the PCR-based and WGS based MSI classification are summarized in the table below. The
sensitivity of WGS-based MSI classification on this set was 100% (95%CI 82.6 – 100%) with a specificity
of 97% (95%CI 88.2-96.9%). The calculated Cohen’s kappa score was 0.954 (95%CI 0-696-0.954),
indicative of a very high agreement.
PCR-MSS
PCR-MSI
Total
MSISeq -MSS
31
0
31
MSIseq - MSI
1
16
17
Total
32
16
48
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 18 of 29
16. Holistic gene panel for driver discovery
We used Ensembl13 release 89 as a basis for our gene definitions and have taken the union of Entrez
identifiable genes and protein coding genes as our base panel.
Certain genes have multiple definitions. NPIPA7 for example has two definitions which are
equally valid, ENSG00000214967 and ENSG00000183889. To solve this we select a single gene
definition based on the following steps:
1) Exclude non protein coding genes.
2) Favour genes that are present in both Havana and Ensembl.
3) Select gene with longest transcript.
This returns our final gene panel tally to 25,963 genes of which 20,083 genes are protein coding. For
each gene we chose the canonical transcript or the longest if no canonical transcript exists.
For CDKN2A, we included both the p16 and p14arf transcripts in the analysis given the known
importance of both transcripts to tumorigenesis24 and the fact that the two transcripts use alternate
reading frames in the same exon.
17. Significantly mutated driver genes discovery
Using all SNV and INDEL variants from the holistic gene panel, we ran dNdScv25 to find significantly
mutated genes (SMGs) and also to estimate the proportion of missense, nonsense, essential splice site
and INDEL variants which are drivers in each individual gene in the panel.
Pan cancer and at an individual cancer level we tested the normalised dNdS rates against a null
hypothesis that dNdS = 1 for each variant subtype. To identify SMGs in our cohort we used a strict
significance cutoff of q<0.01.
Two of the newly discovered SMG candidates were subsequently removed via manual curation as they
were deemed to be likely artefacts of our methods:
●
POM121L12 - found only to be significant due to an extreme covariate value in dNdScv
●
TRIM49B - found to have poor mappability on nearly all its variants and a known close paralog
18. Significantly amplified & deleted driver gene discovery
To search for significantly amplified and deleted genes we first calculated the minimum exonic copy
number per gene across our holistic gene panel. For amplifications, we searched for all the genes with
high level amplifications only (defined as minimum Exonic Copy number > 3 * sample ploidy). For
deletions, we searched for all the genes in each sample with either full or partial gene homozygous
deletions (defined as minimum exonic copy number < 0.5). The Y chromosome was excluded from the
deletion analysis since the Y chromosome is deleted altogether in 35% of all male cancer samples in our
cohort and hence is difficult to distinguish at the gene level.
We then searched separately for amplifications and deletions, on a per chromosome basis, for the most
significant focal peaks, using an iterative GISTIC-like peel off method26, specifically:
●
Find the highest scoring gene.
○
For deletions the score is simply the count of samples with homozygous deletions in the
gene.
○
For amplifications, we need to consider both the count and strength of the amplification
so we use:
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 19 of 29
■
score = sum(log2(copy number/ sample ploidy)).
●
Record gene as a peak, and mark all consecutive genes with a score within 15% and 25% of the
highest score for deletions and amplifications respectively as part of the candidate peak.
●
‘Peel’ off all samples which contributed to the peak across the entire chromosome.
●
Repeat the process.
A filter was applied where we removed deletions from a handful of noisy copy number regions in the
genome where we found more than 50% of the observed deletions were not supported on either
breakend by a structural variant.
Most of the deletion peaks resolve clearly to a single target gene reflecting the fact that homozygous
deletions are highly focal, but for amplifications this is not the case and the majority of our peaks have 10
or more candidates. We therefore annotated the peaks, to choose a single putative target gene using an
objective set of automated curation rules in order of precedence:
●
If more than 50% of the copy number events in the peeled samples include the telomere or
centromere than mark as <CHR>_<ARM>_<TELOMERE/CENTROMERE>
●
Else choose highest scoring candidate gene which matches a list of actionable amplifications
from OncoKB, CGI and CIViC clinical annotation DBs
●
Else choose highest scoring candidate gene found in our panel of significantly mutated genes.
●
Else choose highest scoring candidate gene found in cosmic census
●
Else choose highest scoring protein coding candidate gene
●
Else choose longest non-coding candidate gene
Finally, we filter the peaks to only highly significant deletions and amplifications using the following rules
●
Deletions => Keep any peak with > 5 homozygous deletions
●
Amplifications => Keep any peak with score > 29
These cut-offs were chosen using a binomial model which assumes the probability of any given gene
being observed to be randomly deleted or highly amplified is equal to the average number of genes
amplified or deleted in each event divided by the total number of genes considered. The cut-offs were
chosen to be the lowest score with a q-value below 0.25. Since amplifications are generally much broader
(averaged genes affected per event of 41.6 compared to just 5.4 for deletions) a much higher number of
genes is required to reach significance.
The calculation details for the cut-offs are presented in the table below.
Cohort data
Statistical Calculations
Count
of
events
Sum
Scores
Count
of
genes
affected
Avg
genes
affected
per event
Avg
score
/
event
Total
genes
tested
Probability
event
overlaps a
given gene
Score
cutoff
P value
of cutoff
Significant
findings
Q Value
Dels
4,915
4,915
26,676
5.4
1.0
25,965
0.00021
5
0.00068
117
0.15
Amps
3,925
6,959
163,393
41.6
1.8
25,965
0.00160
29
0.00030
33
0.23
This model is likely to be highly conservative as it assumes that all the events are passengers, whereas in
fact a high proportion contain driver genes.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 20 of 29
19. Fragile site annotation
Homozygous deletions were also annotated as common fragile site (CFS) based on their genomic
characteristics. This annotation is not definitive, but is useful as CFS are known to be regions of high
genomic instability. Hence despite being significantly deleted, their status as a genuine cancer driver
remains unclear.
There is no absolute agreement on which regions should be classified as CFS, but two well-known
features are a strong enrichment in long genes and a high rate of observed deletions of up to 1
megabase27. Hence for this analysis we classified a gene as a fragile site if it met all the following criteria:
●
Total length of gene > 500,000 bases
●
More than 30% of all SVs with breakpoints that disrupt the gene are deletions with length greater
than 20,000 bases and less than 1 megabase.
●
The gene is not found to be significantly mutated (by dNdScv) in our cohort or in Martincorena et
al.25.
Using these criteria we annotated the following list of 16 Genes as fragile:
Gene
Chr
Start
position
Length (bases)
Total Disruptive
SV Count
% of SV that are DELs
(>20kb & <1MB)
LRP1B
2
140,988,992
1,900,278
1,272
0.469
FHIT
3
59,735,036
1,502,097
2,128
0.596
LSAMP
3
115,521,235
2,194,860
1,306
0.364
NAALADL2
3
174,156,363
1,367,065
1,198
0.456
CCSER1
4
91,048,686
1,474,378
1,398
0.441
PDE4D
5
58,264,865
1,553,082
1,166
0.458
GMDS
6
1,624,041
621,885
399
0.441
PARK2
6
161,768,452
1,380,351
1,296
0.555
IMMP2L
7
110,303,110
899,463
1,028
0.444
PTPRD
9
8,314,246
2,298,477
1,264
0.309
PRKG1
10
52,750,945
1,307,165
781
0.318
GPHN
14
66,974,125
674,395
291
0.306
WWOX
16
78,133,310
1,113,254
1,319
0.541
MACROD2
20
13,976,015
2,057,827
3,039
0.605
DMD
X
31,115,794
2,241,764
789
0.328
DIAPH2
X
95,939,662
920,334
331
0.381
We also noted that 4 other significantly deleted genes (STS,HDHD1,LRRN3 and LINC00290), though not
fulfilling the length criteria above have a particularly high proportion of deletion SVs between 20kb and 1
megabase (over 60%) and hence were also marked as fragile:
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 21 of 29
Gene
Chr
Start
position
Length (bases)
Total Disruptive
SV Count
% of SV that are DELs
(>20kb & <1MB)
LINC00290
4
181,985,242
95,060
64
0.641
LRRN3
7
110,731,062
34,448
70
0.686
STS
X
7,137,497
135,354
168
0.649
HDHD1
X
6,966,961
99,270
126
0.659
Two of these genes (STS and HDHD1) fall in a previously identified CFS region (FRAXB) and a third,
LRNN3, falls in another knowns CFS region (FRAX7). The final one, LINC00290 is a long non-coding
RNA with an unknown status as cancer driver.
20. Somatic driver catalog construction
We created a catalog of each and every driver in our cohort across all variant types on a per patient
basis. This was done in a similar incremental manner to Sabarinathan et al28 (N. Lopez, personal
communication) whereby we first calculated the number of drivers in a broad panel of known and
significantly mutated genes across the full cohort, and then assigned the drivers for each gene to
individual patients by ranking and prioritising each of the observed variants. Key points of difference in
this study were both the prioritisation mechanism used and our choice to ascribe each mutation a
probability of being a driver rather than a binary cutoff based on absolute ranking.
The four detailed steps to create the catalog are described below:
1. Create a panel of driver genes for point mutations using significantly mutated genes and known
drivers
We created a gene panel using the union of
●
Martincorena significantly mutated genes25 (filtered to significance of q<0.01)
●
HMF significantly mutated genes (q<0.01) at global level or at cancer type level
●
Cosmic Curated Genes14 (v83)
2. Determine TSG or Oncogene status of each significantly mutated gene
We used a logistic regression model to classify the genes in our pane as either tumor suppressor gene
(TSG) or oncogene. We trained the model using unambiguous classifications from the Comic curated
genes, i.e. a gene was considered either a Oncogene or TSG but not both. We determined that the dNdS
missense and nonsense ratios (w_missense and w_nonsense) are both significant predictors of the
classification. The coefficients are given in the table below.
Estimate
Std. Error
z value
Pr(>|z|)
intercept
0.1830
0.3926
0.466
0.64106
w_missense
-0.6869
0.2643
-2.599
0.00936
w_nonsense
0.5237
0.1116
4.691
2.72e-06
We applied the model to all significantly mutated genes in Matincorena and HMF as well as any
ambiguous Cosmic curated genes.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 22 of 29
The following figure shows all genes that have classified using the logistic regression model. Figures A
and C show the likelihood of a gene being classified as a TSG under a single variate logistic model of
w_missense and w_nonsense respectively. Figure B shows the classification after the multivariate
regression using both predictors.
3. Add drivers from all variant classes to the catalog
Variants were added to the driver catalog which met any of the following criteria
●
All missense and inframe indels for panel oncogenes
●
All non synonymous and essential splice point mutations for tumor suppressor genes
●
All high level amplifications (min exonic copy number > 3 * sample ploidy) for both significantly
amplified target genes and panel oncogenes
●
All homozygous deletions for significantly deleted target genes and panel TSG (except for the Y
chromosome as described before)
●
All known or promiscuous inframe gene fusions as described above
●
Recurrent TERT promoter mutations
4. Calculate a per sample driver likelihood for each gene in the catalog
A driver likelihood estimate between 0 and 1 was calculated for each variant in the gene panel to ensure
that only excess mutations are used for determining the number of drivers in cancer cohort groups or at
the individual sample level. High level amplifications, Deletions, Fusions, and TERT promoter mutations
are all rare so were assumed to have a likelihood of 1 when found affecting a driver gene, but for coding
mutations we need to account for the large number of passenger point mutations that are present
throughout the genome and thus also in driver genes.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 23 of 29
For coding mutations we also marked coding mutations that are highly likely to be drivers and/or highly
unlikely to have occurred as passengers as driver likelihood of 1, specifically:
●
Known hotspot variants
●
Variants within 5 bases of a known pathogenic hotspot in oncogenes
●
Inframe indels in oncogenes with repeat count < 8 repeats. Longer repeat count contexts are
excluded as these are often mutated by chance in MSI samples
●
Biallelic variants in tumor suppressor genes
For the remaining variants (non-hotspot missense variants in oncogenes and non-biallelic variants in
TSG) these were only assigned a > 0 driver likelihood where there was a remaining excess of unallocated
drivers based on the calculated dNdS rates in that gene across the cohort after applying the above rules.
Any remaining point mutations were assigned a driver likelihood between 0 and 1 using a bayesian
statistic to calculate a sample specific likelihood of each gene based on the type of variant observed
(missense, nonsense, splice or INDEL) and taking into account the mutational load of the sample. The
principle behind the method is that the likelihood of a passenger variant occuring in a particular sample
should be approximately proportional to the tumor mutational burden and hence variants in samples with
lower mutational burden are more likely to be drivers.
The sample specific likelihood of a residual excess variant being a driver is estimated for each gene using
the following formula:
P(Driver|Variant) = P(Driver) / (P(Driver) + P(Variant|Non-Driver) * (1-P(Driver)))
where P(Driver) in a given gene is assumed to be equal across all samples in the cohort, ie:
P(Driver) = (residual unallocated drivers in gene) / # of samples in cohort
And P(Variant|Non-Driver), the probability of observing n or more passenger variants of a particular
variant type in a sample in a given gene, is assumed to vary according to tumor mutational burden, and is
modelled as a poisson process:
P(Variant|Non-Driver) = 1 - poisson(λ = TMB(Sample) / TMB(Cohort) * (# of passenger variants in
cohort),k=n-1)
All counts reported in the paper at a per cancer type or sample level refer to the sum of driver likelihoods
for that cancer type or sample.
21. Driver co-occurrence analysis
To examine the co-occurence of drivers, the driver-gene catalog was filtered to exclude fusions and any
driver with a driver likelihood of < 0.5. Separately for each cancer type, every pair of driver genes was
tested to see whether they co-occur more or less frequently than expected if they were independent using
Fisher’s Exact Test. The results were adjusted to a FDR using the number of gene-pair comparison being
tested in each cancer type cohort. Gene pairs with a positive correlation which were on the same
chromosome were excluded from the analysis as they are frequently co-amplified or deleted by chance.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 24 of 29
22. Actionability analysis
To determine clinical actionability of the variants observed in each sample, we mapped all variants to 3
external clinical annotation databases
●
OncoKB9 (download = 01-mar-2018)
●
CGI8 (update: 17-jan-2018)
●
CIViC7 (download = 01-mar-2018)
In order to be able to aggregate and compare this data, we have mapped each of the databases to a
common data model using the following rules:
1. Level of evidence mapping
The 3 databases we used in this study define different level for evidence items, depending on evidence
strength. In order to be able to aggregate and compare this data, we have mapped the CGI and OncoKB
evidence levels on the CIViC evidence levels defined at: https://civicdb.org/help/evidence/evidence-
levels.
HMF
CIViC
CGI
OncoKB
A
A
FDA guidelines,
NCCN guidelines, NCCN/CAP
guidelines, CPIC guidelines,
European Leukemia
Net guideline
1
2
R1
B
B
Clinical trials,
Late trials,
Late trials,Pre-clinical
3
R2
C
C
Early trials,
Case report
D
D
Pre-clinical
4,R3
In this study we considered only A and B level variants. This classification roughly corresponds to the
recently proposed ESMO Scale for Clinical Actionability of molecular Targets (ESCAT)29 as follows:
HMF A: ESCAT I-A+B (for on label) and I-C (for off-label)
HMF B: ESCAT II-A+B (for on label) and III-A (for off-label)
2. Response type Mapping
We also mapped response type to a common data model. First we filtered out evidence items from the
annotation databases that do not lead to clinical actionability (for example prognostic biomarkers). The
remaining evidence items were mapped as either responsive or resistant based on the following rules:
HMF
CIViC
CGI
OncoKB
Responsive
Sensitivity
Responsive
1
2
3
4
Resistant
Resistant or Non-Response
Resistant
R1
R2
R3
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 25 of 29
3. Mutation/Event type mapping
Each evidence item was mapped to HMF data as one of 4 event types according to the following criteria:
HMF Event type
Matching Criteria
Somatic Point Mutation
HGVS / genomic coordinates converted to chromosome, position, ref and alt and
mapped to exact variants in our database
Somatic Range Event
Matched to missense / inframe variants in Oncogenes and any non-synonymous
variant in TSG contained within a defined range, either exon level, transcript level or
specific coordinates. Where a transcript was not specified, the canonical transcript
was always used to map coordinates
Somatic CNA
‘Deletion’ mapped to homozygous deletions and ‘Amplification’ mapped to high level
amplification (>3x sample ploidy)
Fusion
Exact matching to an inframe fusion in our database. For OncoKB ‘loss-of-function’
fusions were excluded
A small number of items from CIViC level B evidence level were deemed either not specific enough or
insufficiently supportive of actionability for this study and were filtered:
●
Evidence items supporting TP53, KRAS & PTEN as actionable
●
Evidence items supporting actionability with ‘chemotherapy’ (ie. chemotherapy in general rather
than a specific treatment), ‘aspirin’ or ‘steroids’
Finally, a number of suspicious fusions from each of the databases were curated by either changing the
5’ and 3’ partners or filtered out altogether based on referring to the original evidence sources,
specifically:
HMF Curation
CIViC
CGI
OncoKB
Filtered Fusions
BRAF - CUL1
RET - TPCN1
5’ and 3’ partners
exchanged
ABL1 - BCR
PDGFRA - FIP1L1
PDGFB - COL1A1
ROS1 - CD74
EP300 - MLL
EP300 - MOZ
RET - CCDC6
Some of the more complex event types from the 3 databases have not been fully interpreted and have
been excluded from this analysis.
4. Cancer type mapping
Each evidence event mapped was also determined to be either on-label (ie. evidence supports treatment
in that specific cancer type) or off-label (evidence exists in another cancer type) for each specific sample.
To do this, we have annotated both the patient cancer types and the database cancer types with relevant
DOIDs, using the disease ontology database available at: http://disease-ontology.org.
Patient cancer types from the HMF database were annotated according to the following table:
HMF tumor type
DOID
Biliary
4607
Bone/Soft tissue
201;9253
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 26 of 29
Breast
1612
CNS
3620;3070
Colon/Rectum
9256;219
CUP
-
Esophagus
5041;4944
Head and neck
11934;8618
Kidney
263;8411
Liver
3571
Lung
1324
Mesothelioma
1790
NET
-
Other
-
Ovary
2394
Pancreas
1793
Prostate
10283
Skin
4159
Stomach
10534
Urinary tract
3996
Uterus
363
Database cancer types were mapped to a DOID by automatically querying the ontology on the disease
names. Some CIViC evidence items are already annotated with a DOID in the database, this was used if
present. We also manually annotated with DOIDs some of the database cancer types that failed the
automatic query:
cancerType
DOID
Ontology term
All Tumors
162
cancer
Any cancer type
162
cancer
B cell lymphoma
707
B-cell lymphoma
Billiary tract
4607
biliary tract cancer
Bladder
11054
urinary bladder cancer
Cervix
4362
cervical cancer
CNS Cancer
3620
central nervous system cancer
Dedifferentiated Liposarcoma
3382
liposarcoma
Endometrium
1380
endometrial cancer
Esophagogastric Cancer
5041
esophageal cancer
Gastrointestinal stromal
9253
gastrointestinal stromal tumor
Giant cell astrocytoma
3069
astrocytoma
Hairy-Cell leukemia
285
hairy cell leukemia
Head and neck
11934
head and neck cancer
Head and neck squamous
5520
head and neck squamous cell carcinoma
Hepatic carcinoma
686
liver carcinoma
Hepatocellular Mixed Fibrolamellar
Carcinoma
0080182
mixed fibrolamellar hepatocellular carcinoma
Inflammatory myofibroblastic
0050905
inflammatory myofibroblastic tumor
Lung
1324
lung cancer
Lung squamous cell
3907
lung squamous cell carcinoma
Melanoma
8923
Skin melanoma
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 27 of 29
Mesothelioma
1790
malignant mesothelioma
Neuroendocrine
169
neuroendocrine tumor
Non-small cell lung
3908
non-small cell lung carcinoma
Ovary
2394
ovarian cancer
Pancreas
1793
pancreatic cancer
Renal
263
kidney cancer
Salivary glands
8850
salivary gland cancer
Stomach
10534
stomach cancer
Thymic
3277
thymus cancer
Thyroid
1781
thyroid cancer
Well-Differentiated Liposarcoma
3382
liposarcoma
In case a matching DOID was found for the disease, we annotated the disease with a DOID set
consisting of: the disease DOID, all the children DOIDs and all the parent disease DOIDs.
A treatment is defined as on-label if any of the DOIDs of the patient cancer is present in the DOID set of
the disease.
5. MSI actionability
Samples classified as MSI in our driver catalog were also mapped as actionable at level A evidence
based on clinical annotation in the OncoKb database
6. Aggregation of evidence
For each actionable mutation in each sample, we aggregated all the mapped evidence that was available
supporting both on-label and off-label treatments at an A or B evidence level. Treatments that also had
evidence supporting resistance based on other biomarkers in the sample at the same or higher level were
excluded as non-actionable.
For each sample we reported the highest level of actionability, ranked first by evidence level and then by
on-label vs off-label.
23. Data availability
All data described in this study is freely available for academic use from the Hartwig Medical Foundation
through standardized procedures and request forms which can be found at
https://www.hartwigmedicalfoundation.nl/en. Briefly, a data request can be initiated by filling out the
standard form in which intended use of the requested data is motivated. First, an advice on scientific
feasibility and validity is obtained from experts in the field which is used as input by an independent Data
Access Board who also evaluates if the intended use of the data is compatible with the consent given by
the patients and if there would be any applicable legal or ethical constraints. Upon formal approval by the
Data Access Board, a standard license agreement which does not have any restrictions regarding
Intellectual Property resulting from the data analysis needs to be signed by an official organisation
representative before access to the data is granted. Raw data files will be made available through a
dedicated download portal with two-factor authentication.
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 28 of 29
24. References
1.
Bins, S. et al. Implementation of a Multicenter Biobanking Collaboration for Next-Generation
Sequencing-Based Biomarker Discovery Based on Fresh Frozen Pretreatment Tumor Tissue
Biopsies. Oncologist 22, 33–40 (2017).
2.
Li, H. & Durbin, R. Fast and accurate short read alignment with Burrows-Wheeler transform.
Bioinformatics 25, 1754–1760 (2009).
3.
McKenna, A. et al. The Genome Analysis Toolkit: a MapReduce framework for analyzing next-
generation DNA sequencing data. Genome Res. 20, 1297–1303 (2010).
4.
Poplin, R. et al. Scaling accurate genetic variant discovery to tens of thousands of samples. bioRxiv
201178 (2018). doi:10.1101/201178
5.
Van der Auwera, G. A. et al. From FastQ data to high confidence variant calls: the Genome Analysis
Toolkit best practices pipeline. Curr. Protoc. Bioinformatics 43, 11.10.1–33 (2013).
6.
Saunders, C. T. et al. Strelka: accurate somatic small-variant calling from sequenced tumor-normal
sample pairs. Bioinformatics 28, 1811–1817 (2012).
7.
Griffith, M. et al. CIViC is a community knowledgebase for expert crowdsourcing the clinical
interpretation of variants in cancer. Nat. Genet. 49, 170–174 (2017).
8.
Tamborero, D. et al. Cancer Genome Interpreter annotates the biological and clinical relevance of
tumor alterations. Genome Med. 10, 25 (2018).
9.
Chakravarty, D. et al. OncoKB: A Precision Oncology Knowledge Base. JCO Precis Oncol, July
2017, (2017) doi: 10.1200/PO.17.00011
10. Cleveland, M. H., Zook, J. M., Salit, M. & Vallone, P. M. Determining Performance Metrics for
Targeted Next-Generation Sequencing Panels Using Reference Materials. J. Mol. Diagn. 20, 583-
590 (2018).
11. Eijkelenboom, A. et al. Reliable Next-Generation Sequencing of Formalin-Fixed, Paraffin-Embedded
Tissue Using Single Molecule Tags. J. Mol. Diagn. 18, 851–863 (2016).
12. Chen, X. et al. Manta: rapid detection of structural variants and indels for germline and cancer
sequencing applications. Bioinformatics 32, 1220–1222 (2016).
13. Zerbino, D. R. et al. Ensembl 2018. Nucleic Acids Res. 46, D754–D761 (2018).
14. Forbes, S. A. et al. COSMIC: somatic cancer genetics at high-resolution. Nucleic Acids Res. 45,
D777–D783 (2017).
15. Haas, B. et al. STAR-Fusion: Fast and Accurate Fusion Transcript Detection from RNA-Seq. bioRxiv
120295 (2017). doi:10.1101/120295
16. Craig, D. W. et al. A somatic reference standard for cancer genome sequencing. Sci. Rep. 6, 24607
(2016).
17. Nilsen, G. et al. Copynumber: Efficient algorithms for single- and multi-track copy number
segmentation. BMC Genomics 13, 591 (2012).
18. Velazquez Villarreal, E. I., Kumar, V., Yin, Y., Carpten, J. D. & Craig, D. W. Abstract 437: Leveraging
new methods in single-cell copy number analysis and clonotype detection to uncover and
characterize hidden subclones within standard cell lines. Cancer Res. 78, 437–437 (2018).
19. Van Loo, P. et al. Allele-specific copy number analysis of tumors. Proc. Natl. Acad. Sci. U. S. A. 107,
16910–16915 (2010).
20. Huang, K.-L. et al. Pathogenic Germline Variants in 10,389 Adult Cancers. Cell 173, 355–370.e14
(2018).
21. Landrum, M. J. et al. ClinVar: public archive of interpretations of clinically relevant variants. Nucleic
Acids Res. 44, D862–8 (2016).
Pan-cancer whole genome analyses of metastatic solid tumors Priestley, Baber, et al.
Page 29 of 29
22. Kalia, S. S. et al. Recommendations for reporting of secondary findings in clinical exome and
genome sequencing, 2016 update (ACMG SF v2.0): a policy statement of the American College of
Medical Genetics and Genomics. Genet. Med. 19, 249–255 (2017).
23. Huang, M. N. et al. MSIseq: Software for Assessing Microsatellite Instability from Catalogs of
Somatic Mutations. Sci. Rep. 5, 13321 (2015).
24. Al-Kaabi, A., van Bockel, L. W., Pothen, A. J. & Willems, S. M. p16INK4A and p14ARF gene
promoter hypermethylation as prognostic biomarker in oral and oropharyngeal squamous cell
carcinoma: a review. Dis. Markers 2014, 260549 (2014).
25. Martincorena, I. et al. Universal Patterns of Selection in Cancer and Somatic Tissues. Cell 171,
1029–1041 e21 (2017).
26. Mermel, C. H. et al. GISTIC2.0 facilitates sensitive and confident localization of the targets of focal
somatic copy-number alteration in human cancers. Genome Biol. 12, R41 (2011).
27. Glover, T. W., Wilson, T. E. & Arlt, M. F. Fragile sites in cancer: more than meets the eye. Nat. Rev.
Cancer 17, 489–501 (2017).
28. Sabarinathan, R. et al. The whole-genome panorama of cancer drivers. BioArchive (2017).
doi:10.1101/190330
29. Mateo, J. et al. A framework to rank genomic alterations as targets for cancer precision medicine: the
ESMO Scale for Clinical Actionability of molecular Targets (ESCAT). Ann. Oncol. 29, 1895-1902
(2018).
| 2019 | Pan-cancer whole genome analyses of metastatic solid tumors | 10.1101/415133 | [
"Priestley Peter",
"Baber Jonathan",
"Lolkema Martijn P.",
"Steeghs Neeltje",
"Bruijn Ewart de",
"Duyvesteyn Korneel",
"Haidari Susan",
"Hoeck Arne van",
"Onstenk Wendy",
"Roepman Paul",
"Shale Charles",
"Voda Mircea",
"Bloemendal Haiko J.",
"Tjan-Heijnen Vivianne C.G.",
"van Herpen Carl... | null |
1
Molecular Survey for Selected Viral Pathogens in Wild Leopard Cats
1
(Prionailurus bengalensis) in Taiwan with an Emphasis on the Spatial and
2
Temporal Dynamics of Carnivore Protoparvovirus 1
3
4
Chen-Chih Chen,a,f#† Ai-Mei Chang,b† Wan-Jhen Chen,a Po-Jen Chang,c Yu-Ching
5
Lai,d Hsu-Hsun Leee
6
7
aInstitute of wildlife conservation, College of Veterinary Medicine, National Pingtung
8
University of Science and Technology, Pingtung, Taiwan
9
bGraduate Institute of Animal Vaccine Technology, College of Veterinary Medicine,
10
National Pingtung University of Science and Technology, Pingtung, Taiwan
11
cFormosan Wild Sound Conservation Science Center, Miaoli, Taiwan
12
dDepartment of Landscape Architecture and Environmental Design, Huafan University
13
eDepartment of Veterinary Medicine, College of Veterinary Medicine, National Pingtung
14
University of Science and Technology, Pingtung, Taiwan
15
fResearch Center for Animal Biologics, National Pingtung University of Science and
16
Technology, Pingtung, Taiwan
17
18
Running Head: viral pathogens in wild leopard cats
19
2
#Address correspondence to Chen-Chih Chen,
20
Email: ychih0502@gmail.com
21
†These authors contributed equally and listed as co-first authors
22
23
Abstract word count: 241
24
Main text word count: 3408
25
3
ABSTRACT The leopard cat (Prionailurus bengalensis) has been listed as an
26
endangered species under the Wildlife Conservation Act in Taiwan since 2009. In
27
this study, we targeted viral pathogens, included carnivore protoparvovirus 1
28
(CPPV-1), feline leukemia virus (FeLV), feline immunodeficiency virus (FIV),
29
coronavirus (CoV), and canine morbillivirus (CMV), using molecular screening. The
30
spatial and temporal dynamics of the target pathogens were evaluated. Through
31
sequencing and phylogenetic analysis, we aimed to clarify the phylogenetic
32
relationship of isolated viral pathogens between leopard cats and domestic
33
carnivores. Samples from 23 and 29 leopard cats that were live-trapped and found
34
dead, respectively, were collected from Miaoli County from 2015 to 2019 in
35
northwestern Taiwan. CPPV-1 and coronavirus were detected in leopard cats. The
36
prevalence (95% confidence interval) of CPPV-1, and CoV was 63.5% (50.4%–76.6%)
37
and 8.8% (0%–18.4%), respectively. The majority of sequences of each CPPV-1
38
strain amplified from Taiwanese leopard cats and domestic carnivores were
39
identical. All the amplified CoV sequences from leopard cats were identified as
40
feline coronavirus. The spatial and temporal aggregation of CPPV-1 infection in
41
leopard cats was not determined in the sampling area, which indicated a wide
42
distribution of CPPV-1 in the leopard cat habitat. We consider sympatric domestic
43
carnivores to be the probable primary reservoir for the pathogens identified. We
44
4
strongly recommend establishing efforts to manage CPPV-1 and FCoV in the
45
leopard cat habitat, with an emphasis on vaccination programs and population
46
control measures for free-roaming dogs and cats.
47
48
IMPORTANCE The leopard cat (Prionailurus bengalensis) is an endangered
49
species in Taiwan. The effects of infectious diseases on the wildlife population have
50
increasingly been recognized. In this study, we targeted highly pathogenic viral
51
pathogens in wild cat species, included carnivore protoparvovirus 1 (CPPV-1), feline
52
leukemia virus (FeLV), feline immunodeficiency virus (FIV), coronavirus (CoV),
53
and canine morbillivirus (CMV), using molecular screening. Furthermore, we
54
collected the epidemiological and phylogenetic data to understand the spatial and
55
temporal dynamics of the target pathogens in the wild leopard cat population and
56
identified the possible origin of target pathogens. Based on our study, we consider
57
sympatric domestic carnivores to be the probable primary reservoir for the
58
pathogens identified. Our study provides a deeper understanding related to the
59
distribution of target viral pathogens in the wild leopard cats. The information is
60
essential for leopard cat conservation and pathogen management.
61
62
KEYWORDS leopard cats, carnivore protoparvovirus 1, feline coronavirus, spatial
63
5
and temporal distribution, domestic carnivores
64
6
INTRODUCTION
65
he leopard cat (Prionailurus bengalensis) is an endangered felid species that
66
is distributed in East, Southeast, and South Asia (1). It was previously
67
commonly distributed in the lowland habitats throughout the island of Taiwan (2, 3).
68
However, the Wildlife Conservation Act of Taiwan listed the leopard cat as an
69
endangered species in 2009 after an island-wide decline in the population of this
70
species (4). Currently, the distribution of Taiwanese leopard cats is restricted to
71
small areas in 3 counties in Central Taiwan, namely Miaoli, Nantou, and Taichung
72
City. Studies in Miaoli County suggested that road traffic, habitat fragmentation and
73
degradation, illegal trapping, and poisoning are the principal threats to the
74
sustainability of the leopard cat population (5). However, the possible direct or
75
indirect effects of pathogens on the population of Taiwanese leopard cats have never
76
been evaluated. Moreover, information related to infectious agents distributed in the
77
wild Taiwanese leopard cat population has remained scarce. Our previous study
78
documented the distribution of carnivore protoparvovirus 1 in Taiwanese leopard
79
cats and its association with domestic carnivores (6). To our knowledge, this was the
80
only study on infectious agents in free-living leopard cats in Taiwan. The effects of
81
infectious diseases on the wildlife population have increasingly been recognized (7,
82
8). Conspicuous illness or the mass die-off of wild animals caused by specific agents
83
T
7
are easier to identify and are usually considered a threat to the abundance of wildlife
84
populations. Although unremarkable or sublethal diseases in wild animals are
85
difficult to identify, such diseases may reduce the fitness of wild animals through an
86
increased energy output or decreased food ingestion, arresting the growth of the
87
population substantially (7, 9).
88
Pathogen infection in wild felids has been documented worldwide with different
89
degrees of importance. Viral pathogens that have been identified in wild or captive
90
leopard cats include feline immunodeficiency virus (FIV) (10), carnivore
91
protoparvovirus 1 (CPPV-1) (6, 11, 12), feline herpesvirus type 1 (FHV-1) (11), and
92
feline calicivirus (FCV) (11). Furthermore, studies have recorded infection by
93
bacterial and parasitic agents including Anaplasma (13, 14), hemoplasma (13, 15),
94
Hepatozoon felis (16–18), and several helminths (19). Although the effects of the
95
recorded infectious agents on leopard cats remain unclear, identifying infectious
96
agents in the leopard cat population is essential for disease management and species
97
conservation.
98
Our previous study recorded carnivore protoparvovirus 1 (CPPV-1) infection in
99
free-living leopard cats, albeit with a limited sample size. In the present study, we
100
extended the target of viral pathogens for screening using a larger sample size. The
101
target viral pathogens were CPPV-1, feline leukemia virus (FeLV), FIV, coronavirus
102
8
(CoV), and canine morbillivirus (CMV).
103
Our objective was to identify the infection of selected viral pathogens based on
104
molecular screening. The spatial and temporal distribution of target pathogens was
105
described. Furthermore, through sequencing and phylogenetic analysis, we aimed to
106
clarify the phylogenetic relationship of isolated viral pathogens between leopard cats
107
and domestic carnivores.
108
109
MATERIALS AND METHODS
110
Study area. All the leopard cats samples were collected from Miaoli County in
111
northwestern Taiwan (Fig. 1). The sampling area has a hilly landscape with an
112
elevation of less than 320 m above sea level. The total area of Miaoli County is 1820
113
km2, consisting of 1245.3 km2 of forests (68.8%), 291.2 km2 of agricultural land
114
(16.1%), and 132.6 km2 of human construction (7.3%). A well-developed road
115
system, which includes a primary road (approximately 25 m wide), secondary roads
116
(approximately 10 m wide), and tertiary roads (approximately 5 m wide), and human
117
encroachment have fragmented the wildlife habitat in this rural area. The Taiwanese
118
leopard cat population was primarily distributed in the west half of Miaoli County
119
(20).
120
Although estimates of the population of stray or free-roaming dogs and cats were
121
9
not available, they were commonly observed and were sympatric with the leopard
122
cats in the study area (20).
123
124
Sample collection. The leopard cat samples were collected from January 2015
125
to April 2019. Free-living leopard cats were trapped for radio telemetry tracking or
126
relocation of leopard cats that invaded poultry farms. Permission for conducting this
127
study was issued by the Forest Bureau (Permit no.: COA, Forestry Bureau,
128
1061702029, 1081603388). Steel-mesh box traps (108-Rigid Trap, Tomahawk Live
129
Trap, LLC., Hazelhurst, Wisconsin, USA) baited with live quails were employed for
130
trapping the leopard cats. The trapped leopard cats were anesthetized by
131
veterinarians using a mixture of dexmedetomidine hydrochloride (100 µg/kg) and
132
tiletamine HCl/zolazepam HCl (2 mg/kg). The procedures for leopard cat trapping,
133
anesthesia administration, and sample collection were approved by the Institutional
134
Animal Care and Use Committee of National Pingtung University of Science and
135
Technology (Approval no.: NPUST-106-014, NPUST-107-041).
136
The carcasses of found-dead (FD) leopard cats, with the majority of deaths
137
caused by vehicle collision, were collected and submitted by the County
138
Government of Miaoli for additional necropsy and sample collection.
139
Tissues and swabs collected for PCR or reverse transcriptase (RT-PCR) screening
140
10
of selected pathogens are displayed in Table 1.
141
We recorded sex and age for each leopard cat. Age classification was based on
142
guidelines from Chen et al. (6). The criteria of age classification were deciduous
143
dentition for juveniles, permanent dentition but not full growth for subadults, full
144
growth of permanent dentition to mild abrasion of canine teeth for young adults, and
145
moderate to severe abrasion of canine teeth for old adults.
146
147
Nucleic acid extraction and (RT)PCR screening for selected viral pathogens.
148
Samples were homogenized prior to nucleic acid extraction. Total DNA was
149
extracted from the collected tissues and blood samples using the DNeasy blood and
150
tissue kit and total RNA was extracted using the RNeasy minikit and QIAamp RNA
151
blood minikit (Qiagen, Valencia, CA, USA). We performed rectal swabs using the
152
QIAamp DNA stool minikit as well as the QIAamp Viral RNA minikit (Qiagen,
153
Valencia, CA, USA) to extract DNA and RNA, respectively.
154
The manufacturer’s recommended procedures were followed for nucleic acid
155
extraction. Reverse transcription of total RNA to cDNA was performed with the
156
iScript cDNA synthesis kit (Bio-Rad, Hercules, CA) following the manufacturer’s
157
instructions.
158
We selected a consensus primer for each viral pathogen to avoid possible
159
11
genetic divergence of pathogens in wildlife, which cannot be amplified by a specific
160
primer designed for analyzing domestic animals (21). Samples and primers selected
161
for (RT)PCR screening are listed in Table 1. The limitation of detection of (RT)PCR
162
under designed conditions for amplifying the genes of targeted infectious agents
163
ranged from 1 to 1000 gene copies/µL (Table 2).
164
The PCR amplicons of collected samples were sequenced in an ABI377
165
sequencer using an ABI PRISM dye-terminator cycle sequencing ready reaction kit
166
with Amplitaq DNA polymerase (Perkin-Elmer, Applied Biosystems). To identify
167
sequences similar to those of the amplicons, a BLAST search was performed using
168
GeneBank with the nt/nr database available on the BLAST website (BLAST;
169
https://blast.ncbi.nlm.nih.gov/Blast.cgi).
170
171
Phylogenetic analysis. The nucleotide sequences of the infectious agents
172
amplified in this study and retrieved from NCBI Genbank
173
(https://www.ncbi.nlm.nih.gov/nucleotide/) accorded with CLUSTALW (28) in the
174
MEGA 7 software program (29). The maximum-likelihood method (30) was used to
175
model the phylogenetic relationship among sequences amplified from each
176
infectious agent. Prior to the construction of a maximum-likelihood tree, the most
177
suitable model was determined using MEGA 7 based on the lowest Bayesian
178
12
information criterion (BIC) score (31).
179
180
Data analysis. We first estimated the prevalence of each targeted infectious
181
agent and its 95% confidence interval (CI) (32). As leopard cats are endangered, our
182
sample size was limited; thus, we did not intend to exclude the possible distribution
183
of the targeted infectious agents in the population of leopard cats if all individual
184
samples screened negative.
185
The samples from live-trapped (LT) and FD leopard cats were pooled to
186
evaluate a possible spatial or temporal cluster of target pathogens using SaTScan
187
version 9 (33) with the Bernoulli model (34).
188
189
RESULTS
190
Leopard cat sample collection and distribution in Miaoli County. From 2015
191
to 2019, we collected samples from 52 leopard cats, of which 23 were LT and 29
192
were FD (Table 3; Table S1). No significant difference in sex was noted between LT
193
and FD individuals (Pearson’s chi-squared test; p = 0.157). However, there were
194
significantly more adults in the FD group than in the LT group (Fisher’s exact test; p
195
= 0.0026). Samples were collected from leopard cats across western Miaoli County
196
in a landscape of fragmented secondary forest habitat surrounded by farmland and
197
13
residential areas (Fig. 1), which corresponded to the current distribution of the
198
leopard cat population.
199
200
Prevalence and distribution of targeted viral pathogens. For the targeted viral
201
pathogens, only CPPV-1 and coronavirus were detected in the collected samples of
202
leopard cats. The prevalences (95% CI) of CPPV-1 , FeLV, FIV, CoV, and CMV
203
were 63.5% (50.4%–76.6%), 0% (0%–6%), 0% (0%–5.9%), 8.8% (0%–18.4%), and
204
0% (0%–6.3%), respectively (Table 4). The prevalence of CPPV-1 in FD cats was
205
significantly higher than that in LT cats (Fisher’s exact test, p = 0.002). Furthermore,
206
the prevalence was significantly higher in adults than in subadults (Fisher’s exact
207
test, p = 0.01). We did not determine any difference in prevalence between the type
208
of sample, sex, and age for CoV (Table 4).
209
The spatial distribution of CPPV-1-positive individuals was scattered in the
210
west of Miaoli County. Three positive CoV samples were distributed in northwest
211
Miaoli (Fig. 2). We did not determine any spatial and temporal aggregation of
212
CPPV-1 infection in the sampling area (SaTScan, Bernoulli model, p = 0.094).
213
Spatial and temporal analyses were not performed for CoV, CMV, FeLV, and FIV,
214
because very few or no positive samples were detected.
215
216
Viral strain identification and phylogenetic analysis. Viral strain identification
217
14
of CPPV-1 was based on the VP2 amino acid sequences obtained from the 29
218
CPPV-1-positive leopard cats. We determined that 11, 7, 6, and 5 leopard cats were
219
infected with CPV-2a, CPV-2b, CPV-2c, and feline panleukopenia virus (FPV),
220
respectively (Table S2). The occurrence of CPPV-1 strain was significantly different
221
from 2015 to 2018 (Fisher’s exact test, p = 0.006), with CPV-2b occurrence
222
decreasing and CPV-2c and FPV increasing (Fig. 3).
223
Partial VP2 sequences of all CPPV-1 strains amplified from 29 leopard cats, 27
224
dogs, and 9 cats in Miaoli County and accessed from Genbank were included for
225
phylogenetic analysis (Table S3). We adopted the Tamura-Nei model to construct a
226
CPPV-1 phylogenetic tree based on the lowest BIC scores. The phylogenetic tree
227
indicated that each CPPV-1 strain amplified from leopard cats and domestic
228
carnivores from Miaoli County was primarily located in the same subcluster (Fig. 4).
229
Furthermore, the majority of sequences of each CPPV-1 strain amplified from
230
Taiwanese leopard cats and domestic carnivores were identical, comprising
231
sequence types CPV-2a/1, CPV-2b/5, CPV-2b/8, CPV-2c/3, and FPV-4 (Fig 4, Table
232
S3). However, certain sequence types were detected in leopard cats but not in
233
domestic carnivores (Fig. 4, Table S3). Most of the nucleotide mutations of different
234
CPPV-1 variants amplified from leopard cats were synonyms, which did not change
235
the encoded amino acid (Fig 4; Table S2). Nonsynonymous mutations of sequence
236
15
types amplified from leopard cats were determined in CPV-2a/3 with P352L and
237
P356S substitution, CPV-2b/7 with S339N substitution, CPV-2c/5 with G437E
238
substitution, FPV/5 with A379V substitution, and FPV/6 with Q310L, A334T,
239
R377K, or R382K substitution.
240
Phylogenetic analysis of the 3 sequences amplified from the RNA-dependent
241
DNA polymerase (RdRP) gene of CoV from leopard cats was first performed using
242
the Tamura 3-parameter model with discrete Gamma distribution. The phylogenetic
243
tree indicated that all the amplified CoV sequences from leopard cats were located in
244
a cluster of viral species, Alphacoronavirus 1, and a feline coronavirus subcluster
245
(Fig. 5).
246
247
DISCUSSION
248
In this study, we screened the selected viral pathogens using (RT)PCR and
249
determined the distribution of CPPV-1 and CoV in free-living leopard cats.
250
Phylogenetic analysis revealed that the majority of identical genetic types of
251
CPPV-1 strains were circulated between leopard cats and domestic carnivores;
252
however, unique genetic types were identified in leopard cats. On the basis of the
253
sequences of the RdRp gene, all the amplified CoV strains were identified as strains
254
of feline coronavirus (FCoV) in species of Alphacoronavirus 1.
255
16
To our knowledge, CPPV-1 and FCoV infection in free-living leopard cats has
256
only been reported in Taiwan (6), although CPPV-1 infection has been previously
257
reported in captive leopard cats from Taiwan and Vietnam (11, 12). The worldwide
258
distribution of CPPV-1 has resulted in the infection of various wild carnivorous
259
species (22, 35–38). Mech et al. (38) determined that CPPV-1 contributed to a 40%
260
to 60% reduction in wolf pup survival and impeded the population growth rate.
261
Disease induced by CPPV-1 infection was commonly found in the juvenile or
262
subadult individuals of domestic carnivores. However, adult individuals with severe
263
clinical signs of CPPV-1 infection were recorded (39–41). Studies are increasingly
264
reporting severe CPPV-1 enteritis in adult dogs (40, 42). Furthermore, a higher risk
265
of developing chronic gastrointestinal disease had been determined in dogs after
266
CPPV-1 infection (43). However, we observed a higher prevalence of CPPV-1 in FD
267
and adults. A higher prevalence may represent a higher risk of infection or lower
268
mortality. Prevalence data alone are not sufficient to evaluate the effect of CPPV-1
269
on different sample types or age categories. Therefore information regarding the
270
physical effects, pathological changes, and mortality caused by CPPV-1 is required.
271
FCoV infection has been documented in various domestic and wild felids (44–
272
47). The infection can be asymptomatic or associated with a fatal systematic disease,
273
feline infectious peritonitis (FIP), and enteric disease (48, 49). Mochizuki et al. (50)
274
17
screened serum antibodies of 17 iriomote cats (Prionailurus bengalensis
275
iriomotensis), a subspecies of leopard cats, for coronavirus and found a prevalence
276
of 82%. This study indicated frequent exposure to and transmission of FCoV in
277
leopard cats. Although FCoV is commonly detected in wild felids worldwide, only a
278
few species, such as cheetahs (Acinoyx jubatus), have been reported to exhibit FIP
279
(44, 46, 48). In our study, 2 out of 3 positive samples were from FD cats and 1 was
280
from an LT cat. We did not determine any pathological changes or clinical signs
281
related to FCoV. Nevertheless, felids infected with FCoV that display no evidence of
282
disease are considered to be chronic carriers that may increase other felids’ risk of
283
contracting FIP (45, 49).
284
In this study, the effects of CPPV-1 and FCoV on individuals or the population
285
of these leopard cats were not evaluated. However, based on the documented effects
286
and cases of CPPV-1 and FCoV on wild felids, the effect of CPPV-1 and FCoV on
287
leopard cats should not be overlooked, and continuous surveillance will be required.
288
Moreover, the spatial aggregation of CPPV-1 infection in leopard cats was not
289
determined in the sampling area, which indicated a wide distribution of CPPV-1 in
290
the habitat of leopard cats. CPPV-1 is stable in the environment and infectiousness
291
can be maintained for several months (35). Free-roaming domestic carnivores are
292
commonly observed in the sampling area, which is an active area with
293
18
well-developed road systems (51, 52). Although the sample size was small, we
294
found a very high prevalence of CPPV-1 (90%; n = 10; data not shown) in
295
free-roaming dogs and cats in our sampling area. These conditions aggravate the
296
transmission and distribution of CPPV-1 in the leopard cat habitat. Future studies
297
should evaluate the influence of domestic carnivores on the transmission of CPPV-1
298
in the habitat.
299
In addition to pathogen surveillance, application of molecular analysis
300
techniques for pathogens has been suggested for investigating several aspects of
301
pathogenesis (53), including pathogen characterization and pathogen transmission
302
(53). We identified the infection of 4 strains of CPPV-1 and FCoV in leopard cats
303
based on the sequences of each positive amplification for selected pathogens.
304
Temporal dynamics revealed that the infection of CPV-2c and FPV was increased,
305
whereas CPV-2b infection was decreased. The distribution of CPV-2c in Taiwan was
306
first detected in dogs in 2015 (54). Since then, CPV-2c has gradually become the
307
predominant variant of CPPV-1 in dogs (54). We first detected CPV-2c in leopard
308
cats in 2017, which indicates an original transmission direction of CPPV-1 from
309
domestic carnivores to leopard cats. Background information and surveillance data
310
for FPV are scarce. Therefore, factors that increase FPV infection rates still need to
311
be assessed.
312
19
Our previous study found that the majority of sequences of CPPV-1 variants
313
were identical between domestic carnivores and the leopard cats based on the partial
314
VP2 gene sequences (6), which suggested frequent transmission of CPPV-1 between
315
domestic and wild carnivores. In this study, we collected 2 times of leopard cat
316
samples and we recorded several different sequence types for each CPPV-1 variant
317
circulating in the leopard cat population (Fig. 4). However, the majority of
318
amplifications from both domestic carnivores and leopard cats belonged to a specific
319
sequence type of each variant. These results support the assumption that CPPV-1 is
320
transmitted between domestic carnivores and leopard cats. Although we identified
321
nonsynonymous mutations of sequence types from leopard cats, the causes and
322
function of amino acid substitutions were undetermined. The amplified DNA
323
sequence of CPPV-1 VP2 encoded amino acid from 300 to 437 residues (Table S2),
324
located in the GH loop, an externally exposed loop in the antigenic region with the
325
greatest variability (55). The sequence types of each variant found only in leopard
326
cats does not indicate an adoption to the leopard cat, as only a few sequences from
327
domestic carnivores in the sampling area were reported.
328
Cross-species transmission of CPPV-1 between domestic and free-living
329
carnivores has been demonstrated or suspected in several countries (35, 36). Due to
330
the critically endangered situation of leopard cats in Taiwan, sustained CPPV-1
331
20
transmission in this low-density population is improbable (56). We considered the
332
domestic carnivores as the primary reservoirs based on the evidence that dogs and
333
cats exhibited the highest abundance among carnivores in the study area, a high
334
prevalence of CPPV-1, and the fact that CPV-2c occurrence in domestic dogs was
335
earlier than in leopard cats.
336
In this study, we did not detect any current infection of FIV, FeLV, and CDV in
337
leopard cats. The 95% CI of prevalence for FIV, FeLV, and CDV was 0% to 5.9%,
338
0% to 6%, and 0% to 6.3%, respectively. On the basis of the long-lasting disease and
339
proviral DNA in peripheral blood monocytic cells characteristics of the FIV and
340
FeLV in Felidae, the possibility of a false negative is low. Therefore we considered a
341
low occurrence of FIV and FeLV in Taiwanese leopard cats. Studies have been
342
conducted involving serological surveys of FIV and FeLV in leopard cats in Taiwan
343
and Vietnam and did not detect any positive cases (12, 57). However, Hayama et al.
344
(10) determined a prevalence of 3% (n = 86) and 13.6% (n = 280) of FIV infection
345
in Tsushima leopard cats and domestic cats, respectively, in Kami-Shima, Tsushima
346
Island, Japan. The domestic cat was considered as the reservoir of FIV in the
347
Tsushima case (10, 58). The prevalence of FIV and FeLV varies between different
348
felids and geographic regions (59).
349
The infection of CDV has been reported in both wild and domestic felids (59,
350
21
60). Ikeda et al. (61) reported a captive leopard cat in Taiwan having antibodies
351
against CDV. Furthermore, a serological survey for CDV found a prevalence of
352
77.8% in wild Taiwanese leopard cats (62). Exposure to CDV in Taiwanese leopard
353
cats is considered to be high. However, none of the leopard cats manifested clinical
354
signs of CDV. Although we targeted amplifying the nucleotide sequence of CDV
355
and identifying the strain from leopard cats, the low detection probability was
356
expected because of a short virus shedding period. Furthermore, a diseased
357
individual may reduce their activity and thus the probability that they would be
358
sampled.
359
Our study revealed CPPV-1 and FCoV infection in free-living leopard cats. The
360
sympatric domestic carnivores are considered the primary reservoir for the
361
pathogens identified in our study. Although the effects of CPPV-1 and FCoV on
362
individual leopard cats and populations of leopard cats were not evaluated in this
363
study, we strongly recommend the establishment of programs to manage CPPV-1
364
and FCoV in the leopard cat habitat with an emphasis on vaccination programs and
365
population control measures for free-roaming dogs and cats. Previous studies have
366
indicated that because of antigenic differences among CPPV-1 variants, new
367
vaccines that also provide protection against the CPV-2c variant may need to be
368
developed (40, 63).
369
22
370
ACKNOWLEDGMENTS
371
This study was supported by a grant from the Ministry of Science and
372
Technology (MOST)(108-2313-B-020-001) to C.-C. Chen. We thank the field crew
373
members, especially Dr. Esther van der Meer for her assistance in the sample
374
collection. This manuscript was edited by Wallace Academic Editing. We declare no
375
conflict of interest.
376
377
References
378
1.
Ross J, Brodie J, Cheyne S, Hearn A, Izawa M, Loken B, Lynam A,
379
McCarthy J, Mukherjee S, Phan C, Rasphone A, Wilting A. 2015.
380
Prionailurus bengalensis, on The IUCN red list of threatened species.
381
http://dx.doi.org/10.2305/IUCN.UK.2015-4.RLTS.T18146A50661611.en.
382
Retrieved 5 March.
383
2.
Chen JS. 1956. A synopsis of the vertebrates of Taiwan. Kaimin Press,
384
Taipei.
385
3.
Horikawa Y. 1931. A monograph of the mammals of Formosa. Zoological
386
Society of Formosa, Taiwan.
387
23
4.
Forest Bureau, Council of Agriculture, Executive Yuan. 2014. Schedule
388
of protected wildlife, on Forest Bureau, Council of Agriculture, Executive
389
Yuan. http://www.conservation.forest.gov.tw. Retrieved 20 August.
390
5.
Chen MT, Liang YJ, Kuo CC, Pei KJC. 2016. Home ranges, movements
391
and activity patterns of leopard cats (Prionailurus bengalensis) and threats to
392
them in Taiwan. Mammal Study 41:77–86.
393
6.
Chen CC, Chang AM, Wada T, Chen MT, Tu YS. 2019. Distribution of
394
Carnivore protoparvovirus 1 in free-living leopard cats (Prionailurus
395
bengalensis chinensis) and its association with domestic carnivores in
396
Taiwan. PLoS ONE 14:e0221990.
397
7.
Wobeser GA. 2006. Essentials of disease in wild animals. Blackwell
398
Publishing, Iowa, USA.
399
8.
Daszak P, Cunningham AA, Hyatt AD. 2000. Emerging infectious diseases
400
of wildlife-threats to biodiversity and human health. Science 287:443–449.
401
9.
Delahay RJ, Speakman JR, Moss R. 1995. The energetic consequences of
402
parasitism: effects of a developing infection of Trichostrongylus tenuis
403
(Nematoda) on red grouse (Lagopus lagopus scoticus) energy balance, body
404
weight and condition. Parasitology 110:473–482.
405
10.
Hayama S, Yamamoto H, Nakanishi S, Hiyama T, Murayama A, Mori H,
406
24
Sugitani A, Fujiwara SI. 2010. Risk analysis of feline immunodeficiency
407
virus infection in Tsushima leopard cats (Prionailurus bengalensis euptilurus)
408
and domestic cats using a geographic information system. J Vet Med Sci
409
72:1113–1118.
410
11.
Nakamura K, Ikeda Y, Miyazawa T, Nguyen NTP, Duong DD, Le KH,
411
Vo SD, Phan LV, Mikami T, Takahashi E. 1999. Comparison of prevalence
412
of feline herpesvirus type 1, calicivirus and parvovirus infections in domestic
413
and leopard cats in Vietnam. J Vet Med Sci 61:1313–1315.
414
12.
Ikeda Y, Miyazawa T, Nakamura K, Naito R, Inoshima Y, Tung KC, Lee
415
WM, Chen MC, Kuo TF, Lin JA. 1999. Serosurvey for selected virus
416
infections of wild carnivores in Taiwan and Vietnam. J Wildl Dis 35:578–
417
581.
418
13.
Hwang J, Oh DH, Lee H, Chun MS. 2015. Anaplasma sp. and hemoplasma
419
infection in leopard cats (Prionailurus bengalensis euptilurus) from Korea. J
420
Vet Sci 16:385–388.
421
14.
Tateno M, Nishio T, Sakuma M, Nakanishi N, Izawa M, Asari Y,
422
Okamura M, Maruyama S, Miyama TS, Setoguchi A. 2013. Molecular
423
epidemiologic survey of Bartonella, Ehrlichia, and Anaplasma infections in
424
Japanese Iriomote and Tsushima leopard cats. J Wildl Dis 49:646–652.
425
25
15.
Hirata M, Tateno M, Sakuma M, Nakanishi N, Izawa M, Asari Y,
426
Okamura M, Miyama TS, Setoguchi A, Endo Y. 2012. An epidemiological
427
survey of hemoplasma infection in Iriomote cats (Prionailurus bengalensis
428
iriomotensis). J Vet Med Sci 74:1531–1537.
429
16.
Kubo M, Miyoshi N, Yasuda N. 2006. Hepatozoonosis in two species of
430
Japanese wild cat. J Vet Med Sci 68:833–837.
431
17.
Sakuma M, Nishio T, Nakanishi N, Izawa M, Asari Y, Okamura M,
432
Miyama TS, Setoguchi A, Endo Y. 2011. A case of Iriomote cat
433
(Prionailurus bengalensis iriomotensis) with Hepatozoon felis parasitemia. J
434
Vet Med Sci 73:1381–1384.
435
18.
Tateno M, Nishio T, Matsuo T, Sakuma M, Nakanishi N, Izawa M, Asari
436
Y, Okamura M, Miyama TS, Setoguchi A. 2013. Epidemiological survey
437
of tick-borne protozoal infection in iriomote cats and tsushima leopard cats
438
in Japan. J Vet Med Sci 75:985–989.
439
19.
Yasuda N, Ezaki K, Akuzawa M, Izawa M, Doi T, Sakaguchi N, Tatara
440
M. 1994. Helminth survey of wildcats in Japan. J Vet Med Sci 56:1069–
441
1073.
442
20.
Pei JCK, Chen MT. 2008. Present status and conservation of small
443
carnivores at low elevation mountains in Shinchu County and Miaoli County
444
26
(3/3). Forestry Bureau CoA, Executive Yuan, Taipei, Taiwan.
445
21.
Anthony SJ, Epstein JH, Murray KA, Navarrete-Macias I,
446
Zambrana-Torrelio CM, Solovyov A, Ojeda-Flores R, Arrigo NC, Islam
447
A, Khan SA. 2013. A strategy to estimate unknown viral diversity in
448
mammals. MBio 4:e00598–00513.
449
22.
Steinel A, Munson L, Van Vuuren M, Truyen U. 2000. Genetic
450
characterization of feline parvovirus sequences from various carnivores. J
451
Gen Virol 81:345–350.
452
23.
Miyazawa T, Jarrett O. 1997. Feline leukaemia virus proviral DNA
453
detected by polymerase chain reaction in antigenaemic but non-viraemic
454
(‘discordant’) cats. Arch Virol 142:323–332.
455
24.
Adams HR, Vuuren Mv, Bosman AM, Kania S, Kennedy MA. 2011.
456
Detection and genetic analysis of feline immunodeficiency virus (FIVple) in
457
southern African lions (Panthera leo). S Afr J Wildl Res 41:173–180.
458
25.
Poon LLM, Chu DKW, Chan KH, Wong OK, Ellis TM, Leung YHC,
459
Lau SKP, Woo PCY, Suen KY, Yuen KY, Guan Y, Peiris JSM. 2005.
460
Identification of a Novel Coronavirus in Bats. J Virol 79:2001–2009.
461
26.
Vijgen L, Moës E, Keyaerts E, Li S, Van Ranst M. 2008. A
462
Pancoronavirus RT-PCR Assay for Detection of All Known Coronaviruses, p
463
27
3–12. In Cavanagh D (ed), SARS- and other coronaviruses: laboratory
464
protocols doi:10.1007/978-1-59745-181-9_1. Humana Press, Totowa, NJ.
465
27.
Tong S, Chern S-WW, Li Y, Pallansch MA, Anderson LJ. 2008. Sensitive
466
and broadly reactive reverse transcription-PCR assays to detect novel
467
paramyxoviruses. J Clin Microbiol 46:2652–2658.
468
28.
Thompson JD, Higgins DG, Gibson TJ. 1994. CLUSTAL W: improving
469
the sensitivity of progressive multiple sequence alignment through sequence
470
weighting, position-specific gap penalties and weight matrix choice. Nucleic
471
Acids Res 22:4673–4680.
472
29.
Tamura K, Stecher G, Peterson D, Filipski A, Kumar S. 2013. MEGA6:
473
molecular evolutionary genetics analysis version 6.0. Mol Biol Evol
474
30:2725–2729.
475
30.
Nei M, Kumar S. 2000. Molecular evolution and phylogenetics. Oxford
476
University Press, Oxford, USA.
477
31.
Tamura K, Peterson D, Peterson N, Stecher G, Nei M, Kumar S. 2011.
478
MEGA5: molecular evolutionary genetics analysis using maximum
479
likelihood, evolutionary distance, and maximum parsimony methods. Mol
480
Biol Evol 28:2731–2739.
481
32.
Hanley JA, Lippman-Hand A. 1983. If nothing goes wrong, is everything
482
28
all right? Interpreting zero numerators. J Am Med Assoc 249:1743–1745.
483
33.
Kulldorff M. 2014. SaTScan™ v9. 3: Software for the spatial and
484
space-time scan statistics. Information Management Services, Inc.
485
34.
Kulldorff M. 1997. A spatial scan statistic. Commun Stat-Theory Methods
486
26:1481–1496.
487
35.
Duarte MD, Henriques AM, Barros SC, Fagulha T, Mendonça P,
488
Carvalho P, Monteiro M, Fevereiro M, Basto MP, Rosalino LM, Barros
489
T, Bandeira V, Fonseca C, Cunha MV. 2013. Snapshot of viral infections
490
in wild carnivores reveals ubiquity of parvovirus and susceptibility of
491
Egyptian mongoose to feline panleukopenia virus. PLoS ONE 8:e59399.
492
36.
Allison AB, Kohler DJ, Fox KA, Brown JD, Gerhold RW,
493
Shearn-Bochsler VI, Dubovi EJ, Parrish CR, Holmes EC. 2013. Frequent
494
cross-species transmission of parvoviruses among diverse carnivore hosts. J
495
Virol 87:2342–2347.
496
37.
Filipov C, Desario C, Patouchas O, Eftimov P, Gruichev G, Manov V,
497
Filipov G, Buonavoglia C, Decaro N. 2016. A ten-year molecular survey on
498
parvoviruses infecting carnivores in Bulgaria. Transbound Emerg Dis
499
63:460–464.
500
38.
Mech LD, Goyal SM, Paul WJ, Newton WE. 2008. Demographic effects
501
29
of canine parvovirus on a free-ranging wolf population over 30 years. J Wildl
502
Dis 44:824–836.
503
39.
Cavalli A, Bozzo G, Decaro N, Tinelli A, Aliberti A, Buonavoglia D. 2001.
504
Characterization of a canine parvovirus strain isolated from an adult dog.
505
The new microbiologica 24:239–242.
506
40.
Decaro N, Desario C, Elia G, Martella V, Mari V, Lavazza A, Nardi M,
507
Buonavoglia C. 2008. Evidence for immunisation failure in vaccinated adult
508
dogs infected with canine parvovirus type 2c. Microbiologica 31:125–130.
509
41.
Mochizuki M, Horiuchi M, Hiragi H, San Gabriel MC, Yasuda N, Uno T.
510
1996. Isolation of canine parvovirus from a cat manifesting clinical signs of
511
feline panleukopenia. J Clin Microbiol 34:2101–2105.
512
42.
Decaro N, Cirone F, Desario C, Elia G, Lorusso E, Colaianni M,
513
Martella V, Buonavoglia C. 2009. Severe parvovirus in a 12-year-old dog
514
that had been repeatedly vaccinated. Vet Rec 164:593–595.
515
43.
Kilian E, Suchodolski JS, Hartmann K, Mueller RS, Wess G, Unterer S.
516
2018. Long-term effects of canine parvovirus infection in dogs. PLoS ONE
517
13:e0192198.
518
44.
Juan-Sallés C, Domingo M, Herráez P, Fernández A, Segalés J,
519
Fernández J. 1998. Feline infectious peritonitis in servals (Felis serval). Vet
520
30
Rec 143:535–535.
521
45.
Kennedy M, Citino S, McNabb AH, Moffatt AS, Gertz K, Kania S. 2002.
522
Detection of feline coronavirus in captive Felidae in the USA. J Vet Diagn
523
Investig 14:520–522.
524
46.
Kennedy M, Kania S, Stylianides E, Bertschinger H, Keet D, van Vuuren
525
M. 2003. Detection of feline coronavirus infection in southern African
526
nondomestic felids. J Wildl Dis 39:529–535.
527
47.
Van Rensburg IBJ, Silkstone MA. 1984. Concomitant feline infectious
528
peritonitis and toxoplasmosis in a cheetah (Acinonyx jubatus). J S Afr Vet
529
Assoc 55:205–207.
530
48.
Evermann JF, Benfield DA. 2001. Coronaviral infections, p 245–253. In
531
Barker EWAI (ed), Infectious diseases of wild mammals, 3rd ed. Iowa State
532
University Press, Ames, Iowa.
533
49.
O’Brien SJ, Troyer JL, Brown MA, Johnson WE, Antunes A, Roelke
534
ME, Pecon-Slattery J. 2012. Emerging viruses in the felidae: shifting
535
paradigms. Viruses 4:236–257.
536
50.
Mochizuki M, Akuzawa M, Nagatomo H. 1990. Serological survey of the
537
Iriomote cat (Felis iriomotensis) in Japan. J Wildl Dis 26:236–245.
538
51.
Matter HC, Daniels TJ. 2000. Dog ecology and population biology, p 17–
539
31
62. In Macpherson CNL, Meslin FX, Wandeler AI (ed), Dogs, zoonoses and
540
public health. CAB International, Wallingford.
541
52.
Sepúlveda M, Pelican K, Cross P, Eguren A, Singer R. 2015. Fine-scale
542
movements of rural free-ranging dogs in conservation areas in the temperate
543
rainforest of the coastal range of southern Chile. Mamm Biol 80:290–297.
544
53.
Blanchong JA, Robinson SJ, Samuel MD, Foster JT. 2016. Application of
545
genetics and genomics to wildlife epidemiology. J Wildl Manag 80:593–608.
546
54.
Chiang SY, Wu HY, Chiou MT, Chang MC, Lin CN. 2016. Identification
547
of a novel canine parvovirus type 2c in Taiwan. Virol J 13:160.
548
55.
Chapman MS, Rossmann MG. 1993. Structure, sequence, and function
549
correlations among parvoviruses. Virology 194:491–508.
550
56.
Laurenson K, Sillero-Zubiri C, Thompson H, Shiferaw F, Thirgood S,
551
Malcolm J. 1998. Disease as a threat to endangered species: Ethiopian
552
wolves, domestic dogs and canine pathogens. Anim Conserv 1:273–280.
553
57.
Miyazawa T, Ikeda Y, Maeda K, Horimoto T, Tohya Y, Mochizuki M, Vu
554
D, Vu GD, Cu DX, Ono K. 1998. Seroepidemiological survey of feline
555
retrovirus infections in domestic and leopard cats in northern Vietnam in
556
1997. J Vet Med Sci 60:1273–1275.
557
58.
Nishimura Y, Goto Y, Yoneda K, Endo Y, Mizuno T, Hamachi M,
558
32
Maruyama H, Kinoshita H, Koga S, Komori M. 1999. Interspecies
559
transmission of feline immunodeficiency virus from the domestic cat to the
560
Tsushima cat (Felis bengalensis euptilura) in the wild. J Virol 73:7916–
561
7921.
562
59.
Munson L, Terio KA, Ryser-Degiorgis MP, Lane EP, Courchamp F. 2010.
563
Wild felid diseases: conservation implications and management strategies, p
564
237-259. In Macdonald DW, Loveridge AJ (ed), Biology and conservation of
565
wild felids. Oxford University Press, New York.
566
60.
Deem SL, Spelman LH, Yates RA, Montali RJ. 2000. Canine distemper in
567
terrestrial carnivores: a review. J Zoo Wildl Med 31:441–451.
568
61.
Ikeda Y, Nakamura K, Miyazawa T, Chen MC, Kuo TF, Lin JA, Mikami
569
T, Kai C, Takahashi E. 2001. Seroprevalence of canine distemper virus in
570
cats. Clin Diagn Lab Immunol 8:641–644.
571
62.
Pei CJC. 2011. Study on the disease and conservation medicine of the
572
endangered leopard cat. Council of Agriculture, Executive Yuan, Council of
573
Agriculture EY, Taipei, Taiwan.
574
63.
Truyen U. 2006. Evolution of canine parvovirus—a need for new vaccines?
575
Vet Microbiol 117:9–13.
576
577
33
FIG 1 Sampling sites of leopard cats in Miaoli County. The map of Taiwan in the
578
box indicates the location of Miaoli County in Taiwan. Circles and triangles,
579
respectively, denote leopard cats that were live-trapped and found dead. Distribution
580
of land-use types, comprising agriculture, forest, wetland, and building area, are
581
denoted in the background.
582
583
FIG 2 Spatial distribution of CPPV-1(A) and CoV (B) in leopard cats. No
584
significant aggregation of positive samples was noted for either CPPV-1 or CoV.
585
586
FIG 3 Frequency of positive detection for CPPV-1 strains from 2015 to 2018. The
587
detection of CPV-2b decreased with an increase in CPV-2c and FPV detection.
588
589
FIG 4 Molecular phylogenetic relationship of the partial VP2 sequences of the
590
Carnivore protoparvovirus 1 amplified from leopard cats, domestic carnivores, and
591
sequences retrieved from GenBank. The bootstrap value is reported next to the node
592
with 1,000 replicates. Each strain and sequence type is labeled and followed by the
593
number of identical sequences within each group (e.g., CPV-2a/1 (19), indicating
594
that the sequence type 1 of the CPV-2a strain contains 19 identical sequences). The
595
host species and location of the isolates of each accession number was assessed
596
34
(Table S3).
597
598
FIG 5 Molecular phylogenetic relationship of the partial RNA-dependent RNA
599
polymerase gene of coronavirus amplified from leopard cats, and sequences
600
retrieved from GenBank. The bootstrap value is reported next to the node with 1,000
601
replicates. Three amplified sequences for leopard cats (Genbank accession number:
602
MN528739 – MN528741) were located in the feline coronavirus cluster.
603
35
TABLE 1 Samples collected from free-living leopard cats and PCR primers used for amplifying the target pathogens
604
Virus1
Sample for screening
Screening
method
Primer (Annealing temperature °C)
Primer
target
Amplified
gene
Reference
Live-trapped
Carcasses
CPPV-1
whole
blood,
rectal swab
spleen,
lymph node,
small intestine,
rectal swab
Nested PCR
First set (52°C):
M10: 5’-ACACATACATGGCAAACAAATAGA-3’
M11: 5’-ACTGGTGGTACATTATTTAATGCAG-3’
Second set (65°C):
M13: 5’-AATAGAGCATTGGGCTTACCACCATTTTT-3’
M14: 5’ATTCCTGTTTTACCTCCAATTGGATCTGTT-3’
CPPV-1
VP2 gene
(22)
FeLV
whole blood
spleen
Nested PCR
First set (50°C):
U3-F1: 5’- ACAGCAGAAGTTTCAAGGCC-3’
G-R1: 5’-GACCAGTGATCAAGGGTGAG-3’
Second set (52°C):
U3-F2: 5’-GCTCCCCAGTTGACCAGAGT-3’
G-R2: 5’-GCTTCGGTACCAAACCGAAA-3’
FeLV
Gag and
LTR gene
(23)
FIV
whole blood
spleen
Nested PCR
First set (52°C):
P1F: 5’-TGGCCWYTAWCWAATGAAAARATWGAAGC-3’
P2R: 5’-GTATTYTCTGCYTTTTTCTTYTGTCTA-3’
Second set (50°C):
P2F: 5’- TGAAAARATWGAAGCHTTAACAGAMATAG-3’
FIV
RNA-depe
ndent DNA
polymeras
e gene
(24)
36
P1R: 5’-GTAATTTRTCTTCHGGNGTYTCAAATCCCC-3’
CoV
whole blood,
rectal swab
spleen,
small
intestine,
lymph node,
rectal swab
RT-semi
nested PCR2
First set (54.7°C):
IN-6: 5’-GGTTGGGACTATCCTAAGTGTGA-3’
Cor-RV: 5’-TCRCAYTTDGGRTARTCCCA-3’
Second set (55°C):
IN-6: 5’-GGTTGGGACTATCCTAAGTGTGA-3’
IN-7: 5’- CCATCATCAGATAGAATCATCATA-3’
Coronaviri
dae
RNA-depe
ndent DNA
polymeras
e gene
(25, 26)
CDV
whole blood
Spleen, lung,
lymph node
RT-semi
nested PCR
First set (48°C):
RES-MOR-HEN-F1:
5’-TCITTYTTTAGRASITTYGGNCAYCC-3’
RES-MOR-HEN-R:
5’-CKCATTTTGTAIGTCATYTTNGCRAA-3’
Second set (55°C):
RES-MOR-HEN-F2:
GCYATATTYTGTGGRATAATHATHAAYGG
RES-MOR-HEN-R:
5’-CKCATTTTGTAIGTCATYTTNGCRAA-3’
Respirovir
us,
Morbillivir
us,
Henipaviru
s
RNA-depe
ndent RNA
polymeras
e gene
(27)
1CPPV-1: carnivore protoparvovirus 1; FeLV: feline leukemia virus; FIV: feline immunodeficiency virus; CoV: coronavirus; CDV: canine distemper virus. 2RT seminested
605
PCR: Reverse transcription seminested PCR
606
37
607
TABLE 2 Sensitivity of specific PCR assays for detecting CPPV-1, FeLV,
608
FIV, CoV, and CDV. The target genes were cloned into a plasmid vector
609
and the plasmid was diluted to 100 to 109 gene copies/µL for each detection
610
assay
611
Targeted
agent
Sensitivity (Gene copies/µl)
CPPV-1
10 gene copies/μl
FeLV
100 gene copies/μl
FIV
10 gene copies/μl
CoV
100 gene copies/μl
CDV
10 gene copies/μl
38
TABLE 3 Sex and age classification of leopard cats collected from live-trapped and
612
found-dead individuals
613
Type of animal
analyzed
Female (n = 19)
Male (n = 33)
Total
Adult
Subadult Juvenile
Adult Subadult Juvenile
Live-trapped
3
4
4
4
8
0
23
Road killed
6
2
0
16
4
1
29
Total
9
6
4
20
12
1
52
39
614
TABLE 4 Prevalence of targeted viral pathogens in the free-living leopard cat population according to sample type, sex, and age
615
Category
CPPV1 (n = 52)
CMV (n = 48)
Corona (n = 34)
FeLV (n = 50)
FIV (n = 51)
Positive
Prevalence (95%
CI)
Positive
Prevalence
(95% CI)
Positive
Prevalence
(95% CI)
Positive
Prevalence
(95% CI)
Positive
Prevalence
(95% CI)
Total
33
63.5% (50.4–76.5)
0
0% (0–6.3)
3
8.8% (0–18.4)
0
0% (0–6)
0
0% (0–5.9)
Type of sample
Live-trapped
9
39.1% (19.2–
59.1)
0
0% (0–
13.6)
1
7.1% (0–20.6)
0
0% (0–
13.6)
0
0% (0–13)
Found-dead
24
82.8% (69.0–
96.5)
0
0% (0–
11.5)
2
10.0% (0–
23.1)
0
0% (0–
10.7)
0
0% (0–
10.7)
40
Sex
Female
11
57.9% (35.7–
80.1)
0
0% (0–
16.7)
3
27.3% (0–
53.6)
0
0% (0–
16.7)
0
0% (0–
15.8)
Male
22
66.77% (50.6–
82.8)
0
0% (0–10)
1
4.3% (0–12.7)
0
0% (0–9.4)
0
0% (0–9.4)
Age
Adult
24
77.4% (62.7–
92.1)
0
0% (0–
10.7)
0
0% (0–15)
0
0% (0–10)
0
0% (0–10)
Subadult
6
37.5% (13.8–
61.2)
0
0% (0–20)
2
22.2% (0–
49.4)
0
0% (0–
18.8)
0
0% (0–
18.8)
41
Juvenile
3
60% (17–100)
0
0% (0–60)
1
20% (0–55.1)
0
0% (0–75)
0
0% (0–60)
616
617
004
00
Landuse type
Agriculture
WM Forest
5 km lll Wetland
Building
2015 2016 2017 2018
Year
“*Seguence retrieved from Genbank
MN528741/Leopard cat
‘} MNS28739/Leopard cat
MN528740/Leopard cat
=| DQ010921/Feline coronavirus/USA
Alphacoronaviru:
| KY06361 8/Canine coronavirus/China
— NC 002645/Human coronavirus 229E
| 2020 | Molecular Survey for Selected Viral Pathogens in Wild Leopard Cats () in Taiwan with an Emphasis on the Spatial and Temporal Dynamics of Carnivore Protoparvovirus 1 | 10.1101/2020.02.21.960492 | [
"Chen Chen-Chih",
"Chang Ai-Mei",
"Chen Wan-Jhen",
"Chang Po-Jen",
"Lai Yu-Ching",
"Lee Hsu-Hsun"
] | creative-commons |
Hyperoxia inhibits proliferation of retinal endothelial cells in Myc dependent manner
Charandeep Singh1, Andrew Benos1, Allison Grenell1,2, Sujata Rao1,3, Bela Anand-Apte1,3, Jonathan E.
Sears1,4
1 Ophthalmic Research, Cole Eye Institute, Cleveland Clinic, Cleveland, OH 44195, USA.
2 Department of Pharmacology, Case Western Reserve University School of Medicine, Cleveland, OH
44106, USA.
3 Department of Ophthalmology, Cleveland Clinic Lerner College of Medicine of Case Western Reserve
University, Cleveland, OH 44195, USA.
4 Cardiovascular and Metabolic Sciences, Cleveland Clinic, Cleveland, OH 44195, USA.
Abstract
Oxygen supplementation is necessary to prevent mortality of severely premature infants. However, the
supraphysiological concentration of oxygen utilized in these infants simultaneously creates retinovascular
growth attenuation and vasoobliteration that induces retinopathy of prematurity. Here, we report that
hyperoxia regulates the cell cycle and retinal endothelial cell proliferation in a previously unknown Myc
dependent manner which contributes to oxygen-induced retinopathy.
Introduction
Retinopathy of prematurity (ROP) is a leading cause of infant blindness world-wide, accounting for
184,700 new cases annually (Blencowe et al., 2013; Hoppe et al., 2016; Sears et al., 2008). Although
oxygen supplementation is necessary to prevent mortality in premature infants, oxygen supplementation
in severely low birthweight infants can be detrimental to the developing premature organs, such as the
retina, brain, and lung. Although ROP does not develop until corrected gestational age of 30-32 weeks, it
is retinovascular growth attenuation and vasoobliteration caused by higher than in utero oxygen
concentrations that creates increased avascular retinal tissue that causes pathological angiogenesis
followed by retinal detachment and blindness (Kim et al., 2018). One of the early clinical signs of ROP is
retinovascular growth suppression (Chen and Smith, 2007; Hartnett and Penn, 2012; Narayanan et al.,
2014). A similar phenotype can be recapitulated in the mouse and rat model of oxygen induced
retinopathy (OIR) (Barnett et al., 2010; Smith et al., 1994). This phenotype is often referred to as “oxygen
toxicity” as it bears the negative connotation reflecting ill effects of oxygen on the vascular development.
In vitro and in vivo studies have demonstrated that hyperoxia increases the formation of reactive oxygen
and nitrogen species (Auten and Davis, 2009; Zou et al., 2019). Furthermore, hyperoxia upregulates
neuronal apoptosis in the brain and the retina (Felderhoff-Mueser et al., 2004; Ikonomidou, 2009;
Taglialatela et al., 1998; Terraneo et al., 2017; Yiş et al., 2008). Although neurons are non-mitotic fully
differentiated cells, they do harbor cell cycle proteins and recent studies have demonstrated that
dysregulation in cell cycle protein levels in neurons can lead to apoptosis. However, mitotic cells of non-
neuronal origin can enter into a long G0 phase under unsuitable circumstances and can re-enter the cell
cycle when conditions become favorable (Foster et al., 2010; Linke et al., 1996). In mice, hyperoxia affects
the vasculature in early postnatal stages when the endothelial cells are still proliferating and migrating.
Smith et. al. (1993) demonstrated that once the vasculature is fully developed, these mice do not develop
the vaso-obliteration and neovascularization phenotype after exposer to 5 days of hyperoxia (Smith et al.,
1994). This implies that susceptibility to hyperoxia is not merely caused by oxidative damage but involves
more complex molecular pathways that are active in the early stages of retinal development. Like in retinal
tissue, postnatal oxygen rich environment inhibits proliferation of cardiomyocytes (Puente et al., 2014).
Mammalian cardiomyocytes have regenerative capacity at birth but lose this potential postnatally as the
oxygen rich environment prevents cell proliferation. In mice, after postnatal day 7, cardiomyocytes
become binucleated and permanently exit the cell cycle through DNA damage-induced cell cycle arrest
(Puente et al., 2014). These differences in response to hyperoxia amongst cell types demonstrate the
heterogeneity of cell cycle control and warrant closer examination of cell type-specific mechanisms.
Myc is a critical regulator of the cell cycle and cellular proliferation (Bretones et al., 2015b). Hypoxia-
induced increase in HIF1 levels result in decreased Myc RNA and protein expression (Okuyama et al., 2010;
Sun and Denko, 2014; Wise et al., 2011). Furthermore, Myc levels are inversely proportional to nutrient
availability and cell density. Myc is downregulated during starvation conditions, halting the cell cycle,
which leads to the loss of proliferation to protect the essential supplies for survival (Bretones et al.,
2015b). One of the mechanisms by which Myc regulates cellular proliferation is via upregulating
polyamine production (Bachmann and Geerts, 2018). The polyamine pathway is indispensable for normal
proliferation and growth (Li et al., 1999; Tabor and Tabor, 1984). Hypoxia increases glycolysis by
upregulating pyruvate dehydrogenase kinase-1 (PDK1) which phosphorylates pyruvate dehydrogenase
(PDH) and thereby inhibits entry of glycolytic carbon into the TCA cycle. This switch in metabolic flux
downregulates cell proliferation by inducing the expression of cyclin-dependent kinase inhibitor (CDKI).
Although phosphorylation of PDH is HIF dependent, upregulation or downregulation of Myc by HIF or vice-
versa is context dependent, and there have been no studies on effects of hyperoxia on Myc protein levels.
HIF can suppress cell proliferation by inhibiting the transcriptional activity of Myc (by destabilizing Myc’s
interaction with other transcriptional co-factors). Recent reports have shown that HIF1 displaces Myc
from MYC-associated protein X (MAX), resulting in destabilization of Myc (Eilers and Eisenman, 2008;
Grinberg et al., 2004). These findings appear to contradict the phenotype of OIR; if hyperoxia
downregulates HIF, one might assume that Myc would be induced by hyperoxia. In this investigation, we
analyzed the effect of hyperoxia on key cell cycle regulators. Our findings indicate a central effect of
hyperoxia on Myc protein levels, providing a molecular mechanism of how oxygen induces cell cycle arrest
in retinal endothelial cells.
Results
To study the effect of hyperoxia on retinal endothelial cell proliferation, we cultured primary human
retinal endothelial cells for 24 h under normoxic conditions, followed by hyperoxic or normoxic conditions
for 4-6 days. Cellular proliferation was significantly reduced under hyperoxic conditions (Fig. 1a), despite
the presence of mitogens such as VEGF, IGF and EGF (please refer to the materials and methods section
for the complete media recipe).
We next examined the expression of polyamine oxidation/breakdown genes, as polyamine levels are
critical regulators of cell proliferation in prokaryotes and eukaryotes (Igarashi and Kashiwagi, 2000).
Polyamines modulate translation by making complexes with RNA. Critical enzymes in the polyamine
pathway, such as ornithine decarboxylase (ODC), peak at G1/S and G2/M transition points, implying that
polyamine levels control these checkpoints (Yamashita et al., 2013). In addition, Nakayama and Nakayama
(1998) demonstrated that the cell cycle inhibitors p27Kip1 and p21Cip1/WAF1 were upregulated in
response to low polyamine concentration in the cells (Nakayama and Nakayama, 1998). This finding was
confirmed by Yamashita et al. (2013), who demonstrated that p27Kip1 translation was enhanced by
polyamine deficiency (Yamashita et al., 2013). In the retina, oxidation/breakdown of polyamines increases
in response to hyperoxia and induces neuronal death (Narayanan et al., 2014). Spermine oxidase (SMOX),
an enzyme that catabolizes early substrates of the growth-inducing polyamine pathway, is reported to be
increased in hyperoxic conditions (Narayanan et al., 2014). We investigated whether the expression levels
of polyamine oxidation genes are regulated at transcriptional levels in response to hyperoxia. Hyperoxia
indeed results in upregulation of SMOX in the endothelial cells as compared to normoxia (Fig. 1c). We also
measured expression of another gene responsible for polyamine oxidation, Peroxisomal N (1)-acetyl-
spermine/spermidine oxidase (PAOX) and found increased expression in response to hyperoxia (Fig. 1d).
This implies that the polyamine oxidation genes are transcriptionally controlled in hyperoxic conditions.
The SMOX inhibitor MDL 72527 has been shown to reduce retinal neuronal death in the OIR model
(Narayanan et al., 2014). We determined that SMOX inhibition could not rescue the cell proliferation
phenotype in endothelial cells cultured under hyperoxic conditions (Fig. 1b).
Given that the inhibition of enzymes that downregulate critical polyamines necessary for growth did not
rescue the growth of hyperoxic endothelial cells, we further examined upstream cell cycle regulators in
synchronized primary human retinal endothelial cells. Since Myc protein levels positively correlate with
cell proliferation in many different cell types, we measured Myc protein levels in normoxic vs. hyperoxic
endothelial cells. Myc protein levels were significantly reduced in hyperoxic endothelial cells compared to
normoxic conditions (Fig. 2a). We confirmed Myc levels with an additional antibody (Fig. S1). However,
we did not observe any changes in the Myc gene expression between normoxia and hyperoxia, implying
Myc levels may be controlled by a previously unknown post-translational modification of Myc protein (Fig.
S2). To further confirm the relationship between hyperoxia and cell cycle arrest, we next evaluated p53,
because this established regulator of the cell cycle is reported to regulate the phosphorylation of Rb (pRb)
(Kastenhuber and Lowe, 2017). p53 can either activate cell cycle arrest by inducing p21/Rb axis or
apoptosis by inducing BCL-2 pathway. Despite this duality of function, it is reported that only one of these
pathways is activated at a time; however it is not clear which cellular events determine which of these
pathways could be activated (Hafner et al., 2019; Kastenhuber and Lowe, 2017). We measured p53 and
p21 levels in normoxic and hyperoxic conditions. p53 (Fig. 2B,G) and p21 (Fig. 2D,I) levels were increased
in hyperoxic conditions confirming cell cycle arrest . Taken together, these findings establish that
hyperoxia causes cell cycle arrest in G1 phase, via p53 and Myc dependent pathways.
To further confirm that hyperoxia induces cell cycle arrest, we evaluated the phosphorylation of
Retinoblastoma protein (Rb). The first step in committing cells to cell division is transition from the G1 to
the S phase, which is dependent on phosphorylation of Retinoblastoma (Rb) protein (Bretones et al.,
2015b; Knudsen and Wang, 1997). Phosphorylated Rb leads to the increased concentration of E2F
elongation factor thereby signaling translation of the proteins required for S-phase (Bretones et al., 2015b;
Knudsen and Wang, 1997). There are 19 known phosphorylation sites on human Rb1 protein (source:
Uniport)(Consortium, 2018), of these the three most important sites involved in cell cycle regulation are
Ser 807, Ser 811, and Ser 795 (Rubin et al., 2005). Recent work by multiple teams have highlighted that
out of these three sites, Ser 807 and Ser 811 regulate c-Abl binding of Rb1 (Knudsen and Wang, 1997;
Rubin et al., 2005). Ser795 is involved in binding of Rb1 to E2F transcription factor (Knudsen and Wang,
1997; Rubin et al., 2005). We measured the levels of phosphorylated pRb Ser807/811 (Fig.2C,H) and pRb
Ser795 (Fig. 2E,J). Both the phosphorylated forms of Rb protein were decreased in response to hyperoxia,
indicating cell cycle arrest in G1 phase. We additionally looked at the levels of putrescine and found it to
be decreased in hyperoxic conditions (Fig. 2k). We confirmed these findings with an additional LC-MS/MS
method (Fig. S3 and S4). Spermidine was not statistically significantly changed and spermine quantity was
not high enough to be confidently measured.
Discussion
Our results clearly demonstrate that hyperoxia downregulates endothelial cell proliferation, without
inducing cell death, by decreasing expression of key cell cycle determinants. Although cell proliferation is
controlled by multiple mechanisms under physiological conditions, p53 and Myc are reported to be the
most important regulators of cell proliferation. Myc controls expression of positive regulators of cell cycle
and also induces growth by down regulating the expression of cell cycle inhibitors such as p21CIP1/WAF1
(for review see Bretones, Delgado and Leon (2015))(Bretones et al., 2015a). The most widely accepted
and recognized mechanism of p21 repression by Myc is through Miz-1. Miz-1, when in contact with Myc,
represses p21. Miz-1/Myc interaction also makes p21 insensitive to p53 signaling (Peukert et al., 1997;
Seoane et al., 2002). The significance of our observation of oxygen induced downregulation of Myc is that
it demonstrates that the central paradigm of the inverse relationship of HIF and Myc expression may not
hold true in hyperoxia. Downregulation of Myc in hyperoxia, when HIF1 levels are known to be decreased,
is unexpected as Myc in most cases works antagonistically to HIF (Okuyama et al., 2010; Sun and Denko,
2014; Wise et al., 2011). This warrants further studies on how and why hyperoxia downregulates Myc and
cell proliferation. Both Myc and p53 control these mechanisms in response to cellular stress like DNA
damage or nutrient deprivation (Puente et al., 2014; Stine et al., 2015). Biomass synthesis pathways like
serine/one-carbon and glutaminolysis involve Myc protein. These pathways were found to be altered by
hyperoxia in our previous studies (Singh et al., 2019; Singh et al., 2018; Singh et al., 2020).
A second important finding from our experiments is that standard, HIF-induced mitogens are unable to
override oxygen induced growth suppression, at least in retinal endothelial cells. VEGF and other mitogens
are known to activate endothelial cell proliferation. In our experiments, hyperoxia was able to block cell
proliferation of endothelial cells despite the presence of mitogens such as VEGF, IGF, and EGF – which
implies that hyperoxia inhibits cell proliferation by acting downstream of these targets. Both MAPK and
PI3K-Akt pathways control cell cycle progression and are downstream of VEGF and EGF/IGF. Our findings
suggest the relevance of these downstream pathways to OIR. Another independent possibility is that the
cells, in response to hyperoxia, have an aberrant VEFR2/R1 ratio rendering them less sensitive to
mitogens.
In conclusion, our investigation demonstrates that hyperoxia downregulates retinal endothelial cell
proliferation by downregulating Myc protein levels and upregulating p53 protein levels. The schema in
Fig. 3 provides a summary of our findings and a potential blueprint for examining how hyperoxia induces
retinal endothelial cell growth suppression.
Materials and methods
Cell proliferation assay
Primary human retinal endothelial cells were purchased from Cell Systems and used within 4-5 passages.
Cells were maintained in endothelial cell media from Cell Biologics (catalogue number H1168). Cells were
plated in black 96-well plates overnight and then incubated in normoxic (21% oxygen) or hyperoxic (75%
oxygen) incubator for 4-6 days. Cell proliferation was measured by using CyQuantTM NF cell proliferation
assay kit from Invitrogen following protocol provided with the kit. SMOX inhibitor, MDL 72527, was spiked
into the media at a final concentration of 100 µM.
Protein extraction form cultured cells
Cells were plated in 100 mm x 20 mm dishes (Corning) coated with the attachment factor (Cell systems
catalogue number 4Z0-201) and maintained in the media described above. Once the cells reached 70-80%
confluency, plates were either incubated in either normoxic (21% oxygen) or hyperoxic (75% oxygen)
incubators for the next 24 h. Both the incubators were set at 37 °C temperature and 5% CO2. After 24 h of
exposure to different levels of oxygen, proteins were extracted from these cells. To extract the proteins,
cells were briefly washed with normal saline, followed by addition of 300 µL of RIPA buffer containing
cOmpleteTM protease inhibitor and phosphatase inhibitor (both from Roche). Cells were scraped with cell
scrapers and transferred to 1.5 tubes. Cells were briefly sonicated and then spun down in a centrifuge at
15000 x g for 15 min at 4°C. The supernatant was transferred to fresh tubes and stored at -80°C until
further use.
SDS-PAGE and western blotting
Protein concentration in the cell lysates was measured using BCA protein assay reagent (PierceTM). Protein
sample 15-20 µg was mixed with tris-glycine SDS loading dye and 20 mM DTT. Samples were heated at
94°C for 3 min following centrifugation at 15000 x g at room temperature for 3 min. Supernatant 30 µL
was loaded into each well of 4-20 % or 12% Tris-glycine NovexTM WedgeWellTM precast gel (Invitrogen).
Equal quantities of protein samples were loaded in all the wells of each individual gel. Proteins were
separated at constant voltage of 150 V. Proteins were transferred from gel to 0.45-micron PVDF
membrane (Millipore) at 70 V for 2 h using wet-transfer in tris-glycine buffer. Following transfer,
membranes were dried for 1 h then quickly rinsed with methanol followed by rinsing with water.
Membranes were then washed with TBS and blocked with intercept TBS blocking buffer (LI-COR) for 1 h.
Following blocking, membranes were treated with primary antibodies diluted in intercept TBS blocking
buffer containing 0.2 % Tween 20 overnight at 4°C. Membranes were washed with TBST 3 times (5 minutes
per wash) then treated with secondary antibodies diluted in intercept TBS blocking buffer containing 0.2
% Tween 20 and 0.01% SDS (w/v) solution, for 1 h at room temperature in the dark. Following incubation
with secondary antibody, blots were washed 3 times with TBST and rinsed with TBS. Images were acquired
on Odyssey® CLx imaging system (LI-COR). Images were analyzed using Image Studio Lite version 5.2 (LI-
COR)
It has earlier been noted previously that the Myc antibodies binds to a non-specific band which co-elutes
with endogenous Myc (Tibbitts et al., 2012). We also observed the non-specific band which eluted very
closely with Myc. To circumvent this problem, we included an additional step of stripping and re-probing
the blot for Myc protein. This step was necessary to remove a second non-specific band seen in our Myc
blots. p21 western blot was stripped with 10 ml of RestoreTM PLUS western blot stripping buffer (Thermo
Fisher Scientific) for 20 min at room temperature followed by re-probing with Myc and β-actin antibody
for 1h at room temperature. After treatment with primary antibody, above described procedure was used
for secondary antibody treatment and imagining.
Following primary antibodies were used:
1) c-Myc (D84C12) Rabbit mAb catalog # 5605
2) Phospho-Rb (Ser 807/811) (D20B12) XP® Rabbit mAb catalog # 8516
3) Phospho-Rb (Ser 795) Rabbit antibody catalog # 9301
4) p21 Waf1/Cip1 (12D1) Rabbit mAb catalog # 2947
5) p53 Rabbit antibody catalog # 9282
6) ß-actin (8H10D10) Mouse mAb catalog # 3700
All the antibodies were purchased from Cell Signaling and were diluted as recommended by the vendor.
Following secondary antibodies were used:
IRDye® 800CW Donkey (polyclonal) anti-Rabbit IgG (H+L), catalog number 925-32213 from LI-COR.
IRDye® 680RD Donkey (polyclonal) anti-mouse IgG (H+L), catalog number 925-68072 from LI-COR. Both
the antibodies were used at 1:2000 dilution.
RNA extraction and quantitative RT-PCR
Cells were cultured in 6-well plates and maintained in endothelial cell media in normoxic incubator. At
around 70-80% confluence, cells were transferred to normoxic or hyperoxic incubator for 24 h, as
described above. Following which RNA was extracted using TRI reagent (Sigma-Aldrich) using protocol
provided with the reagent. The RNA was converted into cDNA using Verso cDNA synthesis kit (Thermo
Fisher Scientific). Two µL of this cDNA was mixed with 10 µL of 2x qPCR mix RadiantTM SYBR Green Lo-ROX
(Alkali Scientific), 1 µL of 10 µM forward (Fwd) primer, 1 µL of 10 µM reverse (Rev) primer and 6 µL of
nuclease free water. PCR settings were 50°C for 2 min, 95°C for 10 min, then 40 cycles at 95°C for 15 sec
and 60°C for 1 min. Following PCR completion, melting curve was recorded using these settings: 95°C for
15 sec, 60°C for 1 min and 95°C for 15 sec.
Sequences of the primers used for RT-PCR:
SMOX Fwd 5’ TCAAAGACAGCGCCCAT 3’; SMOX Rev 5’ CCGTGGGTGGTGGAATAGTA 3’
PAOX Fwd 5’ACTAGGGGGTCCTACAGCTA 3’; PAOX Rev 5 ‘CGTGGAGTAAAACGTGCGAT 3’
Metabolite extraction
Retinal endothelial cells were plated in 100 mm dishes coated with attachment factor (Cell Systems) at
density of 0.9 or 0.4 x 106 cells per plate and maintained in endothelial cell media (CellBiologics) in a 5%
CO2 incubator set at 37°C for 3 days. After 3 days of incubation, media was changed to high glucose DMEM
media (Cleveland Clinic Media lab) without FBS and cells were again incubated in normoxic incubator, to
synchronize the cells. After 6h, media was changed back to endothelial cell media (Cell Biologics) and
plates were either incubated in normoxic or hyperoxic (75% oxygen), 5% C02 incubator set at 37°C for 24
h. Following 24 h of incubation, metabolites were extracted. To extract metabolites, media was aspirated,
plates were washed with 10 ml of room temperature normal saline. To the washed cells 300 µL of 0.1%
formic acid (prepared in water) containing 1 µg of 13C5 ribitol per ml of solution. Next, 600 µL of -20°C cold
methanol was added to each plate. Cells were scraped with a cells scraper while keeping plates on ice and
cell lysates were transferred to tubes containing 450 µL of -20°C cold chloroform. Tubes were agitated on
a thermomixer at 4°C, 1400 rpm for 30 min. Tubes were then centrifuged for 5 min at 15000 x g at 4°C.
Six-hundred microliters of supernatant was transferred to fresh tubes and dried under vacuum in a -4°C
cold vacuum evaporator (Labconco). Samples were derivatized with the two step protocol as described
earlier (Singh et. al. 2020) and measured using GCMS method described earlier (Singh et. al. 2020).
Figure 1 Hyperoxia inhibits prolifeartion and increases expression of polyamine oxidation genes in
retinal endothelial cells a) Retinal endothelial cells cultured in hyperoxia demonstrated proliferation
defects even in the presence of growth factors (n=8 biological replicates per condition). b) Proliferation
defects were not rescued by spermidine oxidase inhibitor (SI) MDL72527 (n=6 biological replicates per
condition) c) Spermine oxidase (SMOX) t-test p-value <0.05 d) Peroxisomal N (1)-acetyl-
spermine/spermidine oxidase (PAOX) t-test p-value <0.05 (n=6 biological replicates per condition).
Figure 2 Cell cycle signaling proteins and polyamine (putrescine) affected by hyperoxia in endothelial
cells. Western blots of normoxic (N1-5; each number represent biological replicate) and hyperoxic (H1-5;
each number represent biological replicate) samples for a) Myc b) p53 c) pRb Ser 807/811 d) p21 e) pRb
Ser 795. Quantification of the western blot is provided in the histograms f) Myc, g) p53, h) pRb Ser 807/811
i) p21 j) pRb Ser 795 k) Putrescine. t-test p-values for all the quantifications were all less than 0.05.
a)
b)
c)
d)
e)
f)
g)
h)
i)
j)
k)
Figure 3 Cell cycle checkpoints and proposed routes affected by hyperoxia. Hyperoxia downregulates
Myc and upregulates P53 proteins thereby increasing p21 protein levels. P21 further downregulates pRb
levels, leading to cell cycle arrest in G1 phase. The cell cycle arrest can be due to anomalies in MAPK and
P13K/Akt pathways downstream of VEGF, EFG/IGF receptors. Alternatively, the cell cycle changes can be
a result of an aberrant VEGFR2/R1 ratio.
Acknowledgements
Grant Support: National Eye Institute (R01 EY024972 to JES; P30 EY025585 to Ophthalmic Research);
Research to Prevent Blindness Physician Scientist (RPB1801 to JES).
References
Auten, R.L., and Davis, J.M. (2009). Oxygen toxicity and reactive oxygen species: the devil is in the details.
Pediatr Res 66, 121-127.
Bachmann, A.S., and Geerts, D. (2018). Polyamine synthesis as a target of MYC oncogenes. The Journal of
biological chemistry 293, 18757-18769.
Barnett, J.M., Yanni, S.E., and Penn, J.S. (2010). The development of the rat model of retinopathy of
prematurity. Doc Ophthalmol 120, 3-12.
Blencowe, H., Lawn, J.E., Vazquez, T., Fielder, A., and Gilbert, C. (2013). Preterm-associated visual
impairment and estimates of retinopathy of prematurity at regional and global levels for 2010. Pediatric
research 74 Suppl 1, 35-49.
Bretones, G., Delgado, M.D., and León, J. (2015a). Myc and cell cycle control. Biochimica et Biophysica
Acta (BBA) - Gene Regulatory Mechanisms 1849, 506-516.
Bretones, G., Delgado, M.D., and León, J. (2015b). Myc and cell cycle control. Biochimica et biophysica
acta 1849, 506-516.
Chen, J., and Smith, L.E. (2007). Retinopathy of prematurity. Angiogenesis 10, 133-140.
Consortium, T.U. (2018). UniProt: a worldwide hub of protein knowledge. Nucleic Acids Research 47,
D506-D515.
Eilers, M., and Eisenman, R.N. (2008). Myc's broad reach. Genes & development 22, 2755-2766.
Felderhoff-Mueser, U., Bittigau, P., Sifringer, M., Jarosz, B., Korobowicz, E., Mahler, L., Piening, T.,
Moysich, A., Grune, T., Thor, F., et al. (2004). Oxygen causes cell death in the developing brain.
Neurobiology of disease 17, 273-282.
Foster, D.A., Yellen, P., Xu, L., and Saqcena, M. (2010). Regulation of G1 Cell Cycle Progression:
Distinguishing the Restriction Point from a Nutrient-Sensing Cell Growth Checkpoint(s). Genes Cancer 1,
1124-1131.
Grinberg, A.V., Hu, C.D., and Kerppola, T.K. (2004). Visualization of Myc/Max/Mad family dimers and the
competition for dimerization in living cells. Molecular and cellular biology 24, 4294-4308.
Hafner, A., Bulyk, M.L., Jambhekar, A., and Lahav, G. (2019). The multiple mechanisms that regulate p53
activity and cell fate. Nature reviews. Molecular cell biology 20, 199-210.
Hartnett, M.E., and Penn, J.S. (2012). Mechanisms and Management of Retinopathy of Prematurity. 367,
2515-2526.
Hoppe, G., Yoon, S., Gopalan, B., Savage, A.R., Brown, R., Case, K., Vasanji, A., Chan, E.R., Silver, R.B., and
Sears, J.E. (2016). Comparative systems pharmacology of HIF stabilization in the prevention of retinopathy
of prematurity. Proceedings of the National Academy of Sciences 113, E2516-E2525.
Igarashi, K., and Kashiwagi, K. (2000). Polyamines: Mysterious Modulators of Cellular Functions.
Biochemical and biophysical research communications 271, 559-564.
Ikonomidou, C. (2009). Triggers of apoptosis in the immature brain. Brain & development 31, 488-492.
Kastenhuber, E.R., and Lowe, S.W. (2017). Putting p53 in Context. Cell 170, 1062-1078.
Kim, S.J., Port, A.D., Swan, R., Campbell, J.P., Chan, R.V.P., and Chiang, M.F. (2018). Retinopathy of
prematurity: a review of risk factors and their clinical significance. Survey of ophthalmology 63, 618-637.
Knudsen, E.S., and Wang, J.Y. (1997). Dual mechanisms for the inhibition of E2F binding to RB by cyclin-
dependent kinase-mediated RB phosphorylation. Molecular and cellular biology 17, 5771-5783.
Li, L., Li, J., Rao, J.N., Li, M., Bass, B.L., and Wang, J.-Y. (1999). Inhibition of polyamine synthesis induces
p53 gene expression but not apoptosis. 276, C946-C954.
Linke, S.P., Clarkin, K.C., Di Leonardo, A., Tsou, A., and Wahl, G.M. (1996). A reversible, p53-dependent
G0/G1 cell cycle arrest induced by ribonucleotide depletion in the absence of detectable DNA damage.
Genes & development 10, 934-947.
Nakayama, K., and Nakayama, K. (1998). Cip/Kip cyclin-dependent kinase inhibitors: brakes of the cell
cycle engine during development. BioEssays : news and reviews in molecular, cellular and developmental
biology 20, 1020-1029.
Narayanan, S.P., Xu, Z., Putluri, N., Sreekumar, A., Lemtalsi, T., Caldwell, R.W., and Caldwell, R.B. (2014).
Arginase 2 deficiency reduces hyperoxia-mediated retinal neurodegeneration through the regulation of
polyamine metabolism. Cell Death Dis 5, e1075.
Okuyama, H., Endo, H., Akashika, T., Kato, K., and Inoue, M. (2010). Downregulation of c-MYC protein
levels contributes to cancer cell survival under dual deficiency of oxygen and glucose. Cancer research 70,
10213-10223.
Peukert, K., Staller, P., Schneider, A., Carmichael, G., Hänel, F., and Eilers, M. (1997). An alternative
pathway for gene regulation by Myc. The EMBO journal 16, 5672-5686.
Puente, B.N., Kimura, W., Muralidhar, S.A., Moon, J., Amatruda, J.F., Phelps, K.L., Grinsfelder, D.,
Rothermel, B.A., Chen, R., Garcia, J.A., et al. (2014). The oxygen-rich postnatal environment induces
cardiomyocyte cell-cycle arrest through DNA damage response. Cell 157, 565-579.
Rubin, S.M., Gall, A.-L., Zheng, N., and Pavletich, N.P. (2005). Structure of the Rb C-Terminal Domain Bound
to E2F1-DP1: A Mechanism for Phosphorylation-Induced E2F Release. Cell 123, 1093-1106.
Sears, J.E., Hoppe, G., Ebrahem, Q., and Anand-Apte, B. (2008). Prolyl hydroxylase inhibition during
hyperoxia prevents oxygen-induced retinopathy. Proceedings of the National Academy of Sciences 105,
19898-19903.
Seoane, J., Le, H.-V., and Massagué, J. (2002). Myc suppression of the p21Cip1 Cdk inhibitor influences the
outcome of the p53 response to DNA damage. Nature 419, 729-734.
Singh, C., Hoppe, G., Tran, V., McCollum, L., Bolok, Y., Song, W., Sharma, A., Brunengraber, H., and Sears,
J.E. (2019). Serine and 1-carbon metabolism are required for HIF-mediated protection against retinopathy
of prematurity. JCI insight 4.
Singh, C., Sharma, A., Hoppe, G., Song, W., Bolok, Y., and Sears, J.E. (2018). 3-Hydroxypyruvate Destabilizes
Hypoxia Inducible Factor and Induces Angiostasis. Investigative Ophthalmology & Visual Science 59, 3440-
3448.
Singh, C., Tran, V., McCollum, L., Bolok, Y., Allan, K., Yuan, A., Hoppe, G., Brunengraber, H., and Sears, J.E.
(2020). Hyperoxia induces glutamine-fuelled anaplerosis in retinal Müller cells. Nature Communications
11, 1277.
Smith, L.E., Wesolowski, E., McLellan, A., Kostyk, S.K., D'Amato, R., Sullivan, R., and D'Amore, P.A. (1994).
Oxygen-induced retinopathy in the mouse. Invest Ophthalmol Vis Sci 35, 101-111.
Stine, Z.E., Walton, Z.E., Altman, B.J., Hsieh, A.L., and Dang, C.V. (2015). MYC, Metabolism, and Cancer.
Cancer Discov 5, 1024-1039.
Sun, R.C., and Denko, N.C. (2014). Hypoxic regulation of glutamine metabolism through HIF1 and SIAH2
supports lipid synthesis that is necessary for tumor growth. Cell Metab 19, 285-292.
Tabor, C.W., and Tabor, H. (1984). Polyamines. Annual review of biochemistry 53, 749-790.
Taglialatela, G., Perez-Polo, J.R., and Rassin, D.K. (1998). Induction of apoptosis in the CNS during
development by the combination of hyperoxia and inhibition of glutathione synthesis. Free radical biology
& medicine 25, 936-942.
Terraneo, L., Paroni, R., Bianciardi, P., Giallongo, T., Carelli, S., Gorio, A., and Samaja, M. (2017). Brain
adaptation to hypoxia and hyperoxia in mice. Redox biology 11, 12-20.
Tibbitts, D.C., Escamilla-Powers, J.R., Zhang, X., and Sears, R.C. (2012). Studying c-Myc serine 62
phosphorylation in leukemia cells: concern over antibody cross-reactivity. Blood 119, 5334-5335.
Wise, D.R., Ward, P.S., Shay, J.E.S., Cross, J.R., Gruber, J.J., Sachdeva, U.M., Platt, J.M., DeMatteo, R.G.,
Simon, M.C., and Thompson, C.B. (2011). Hypoxia promotes isocitrate dehydrogenase-dependent
carboxylation of α-ketoglutarate to citrate to support cell growth and viability. Proceedings of the National
Academy of Sciences of the United States of America 108, 19611-19616.
Yamashita, T., Nishimura, K., Saiki, R., Okudaira, H., Tome, M., Higashi, K., Nakamura, M., Terui, Y.,
Fujiwara, K., Kashiwagi, K., et al. (2013). Role of polyamines at the G1/S boundary and G2/M phase of the
cell cycle. The international journal of biochemistry & cell biology 45, 1042-1050.
Yiş, U., Kurul, S.H., Kumral, A., Cilaker, S., Tuğyan, K., Genç, Ş., and Yılmaz, O. (2008). Hyperoxic exposure
leads to cell death in the developing brain. Brain and Development 30, 556-562.
Zou, D., Li, J., Fan, Q., Zheng, X., Deng, J., and Wang, S. (2019). Reactive oxygen and nitrogen species induce
cell apoptosis via a mitochondria-dependent pathway in hyperoxia lung injury. Journal of cellular
biochemistry 120, 4837-4850.
| 2020 | Hyperoxia inhibits proliferation of retinal endothelial cells in Myc dependent manner | 10.1101/2020.11.09.375220 | [
"Singh Charandeep",
"Benos Andrew",
"Grenell Allison",
"Rao Sujata",
"Anand-Apte Bela",
"Sears Jonathan E."
] | null |
Page 1 of 14
List of email addresses and ORCIDs for all authors:
1
Daniele Mercatelli, daniele.mercatelli2@unibo.it, ORCID 0000-0003-3228-0580
2
Luca Triboli, luca.triboli@studio.unibo.it, ORCID 0000-0002-1261-0637
3
Eleonora Fornasari, eleonora.fornasari@ordingbo.it, ORCID 0000-0002-7636-085X
4
Forest Ray, forest.ray@zoho.com, ORCID 0000-0002-8655-7066
5
Federico M. Giorgi, federico.giorgi@unibo.it, ORCID 0000-0002-7325-9908
6
7
coronapp: A Web Application to Annotate and Monitor
8
SARS-CoV-2 Mutations
9
Daniele Mercatelli1,#, Luca Triboli1,#, Eleonora Fornasari1, Forest Ray2, Federico M.
10
Giorgi1,*
11
1 Department of Pharmacy and Biotechnology, University of Bologna, Bologna,
12
40126, Italy
13
2 Department of Systems Biology, Columbia University Medical Center, New York
14
City, 10032, United States
15
# Equal contribution.
16
* Corresponding author.
17
E-mail: federico.giorgi@unibo.it (Giorgi FM)
18
19
Running title: Mercatelli D et al / coronapp – monitoring SARS-CoV-2 mutations
20
21
Word number: 3531
22
Figure number: 3
23
24
25
Page 2 of 14
Abstract
26
The avalanche of genomic data generated from the SARS-CoV-2 virus requires the
27
development of tools to detect and monitor its mutations across the World. Here, we
28
present a webtool, coronapp, dedicated to easily processing user-provided
29
SARS-CoV-2 genomic sequences, in order to detect and annotate protein-changing
30
mutations. This results in an up-to-date status of SARS-CoV-2 mutations, both
31
worldwide and in user-selected countries. The tool allows users to highlight and
32
prioritize the most frequent mutations in specific protein regions, and to monitor their
33
frequency in the population over time.
34
The tool is available at http://giorgilab.dyndns.org/coronapp/ and the full code is
35
freely shared at https://github.com/federicogiorgi/giorgilab/tree/master/coronapp
36
37
38
39
40
41
KEYWORDS: COVID-19; SARS-CoV-2; mutations; web application
42
43
44
Page 3 of 14
Introduction
45
SARS-CoV-2 is a novel pathogenic enveloped RNA beta-coronavirus causing a
46
severe illness in human hosts known as coronavirus disease-2019 (COVID-19). The
47
predominant COVID-19 illness is a viral pneumonia, often requiring hospitalization
48
and in some cases intensive care [1]. With almost 6 million laboratory-confirmed
49
positive cases worldwide as of 31 May 2020 and an estimated case fatality rate across
50
204 countries of 5.2%, COVID-19 has become a global health challenge in only a few
51
months [2]. SARS-CoV-2 infection depends on the recognition of host angiotensin
52
converting enzyme 2 (ACE2), exposed on the cell surface in human lung tissues [3,4].
53
SARS-CoV-2 spike glycoprotein binds ACE2, mediating membrane fusion and cell
54
entry [5]. Upon cell entry, the virus subverts host cell molecular processes, inducing
55
interferon responses and eventually apoptosis [6].
56
To date, much effort has been made to develop therapeutic strategies to limit
57
SARS-CoV-2 transmission and replication, but no treatment or vaccine has proven
58
effective against the virus, and repurposing of approved therapeutic agents has been
59
the main practical approach to manage the emergency so far [7]. As viruses mutate
60
during replication, the emergence of SARS-CoV-2 sub-strains and the challenge of a
61
probable antigenic drift require attention, especially for vaccine development [8].
62
Although sequence analyses of SARS-CoV-2 have shown that genomic variability
63
is very low [9], new SARS-CoV-2 mutation hotspots are emerging due to the high
64
number of infected individuals across countries and to viral replication rates [10].
65
Three major SARS-CoV-2 clades known as clade G, V, and S have emerged, showing
66
a different geographical prevalence [10]. The most frequent mutation detected so far
67
defines the G clade and causes an aminoacidic change, aspartate (D) or glycine (G), at
68
position 614 (D614G) of the viral Spike protein [11].
69
Continual genomic surveillance should be considered to monitor the possible
70
appearance of viral subtypes characterized by altered tropism, or causing more
71
aggressive symptoms. Constant and widespread monitoring of mutations is also a
72
Page 4 of 14
powerful means of informing drug development and global or local pandemic
73
management. The Global Initiative on Sharing All Influenza Data (GISAID) has
74
collected to date (31 May 2020) over 30,000 publicly accessible SARS-CoV-2
75
sequences. The GISAID effort has made it possible to compare genomes on a
76
geographical and temporal scale and an increasing number of laboratories have started
77
to sequence COVID-19 patient samples worldwide [13,14]. Several online tools have
78
been developed to monitor the evolution of the virus from a phylogenetic perspective,
79
such as Nextstrain [15], or to visualize epidemiological data such as number of cases
80
and deaths [16]. However, no tool currently exists to annotate user-provided
81
SARS-CoV-2 genomic sequences, which may derive from specific GISAID subsets
82
or from sequencing efforts of individual laboratories. Neither does any tool
83
specifically monitor the prevalence of specific SARS-CoV-2 mutations associated to
84
particular geographic regions or protein locations, nor their frequency in the
85
population over time.
86
To overcome these limitations, we have developed coronapp, a web application
87
with two purposes: real-time tracking of SARS-CoV-2 mutational status and
88
annotation of user-provided viral genomic sequences. Our tool enables users to easily
89
perform genomic comparisons and provides an instrument to monitor SARS-CoV-2
90
genomic variance, both worldwide and by uploading custom and locally produced
91
genomic sequences. The webtool is available at http://giorgilab.dyndns.org/coronapp/
92
and
the
full
source
code
is
shared
on
Github
93
https://github.com/federicogiorgi/giorgilab/tree/master/coronapp
94
95
Results
96
The
webtool
coronapp
is
available
at
the
website
97
http://giorgilab.dyndns.org/coronapp/ and it automatically provides the user with the
98
current status of SARS-CoV-2 mutations worldwide. The app also allows users to
99
Page 5 of 14
annotate user-provided sequences (Figure 1 A). There are multiple functionalities of
100
coronapp, described in the following paragraphs.
101
102
Current Status of SARS-CoV-2 mutational data
103
A worldwide analysis is shown, generated using data from GISAID. Specifically, we
104
processed all SARS-CoV-2 complete (>29,000 sequenced nucleotides) genomic
105
sequences, excluding low-quality sequences (>5% undefined nucleotide “N”) and
106
viruses extracted from non-human hosts.
107
The underlying database is updated weekly, and we provide the date of the last
108
version as a reference for studies based on the data provided. We indicate the number
109
of samples processed and the total number of mutational events detected (Figure 1 A).
110
We also show the number of distinct mutated loci. Currently, this number is slightly
111
below 11,000, meaning that less than half of the original Wuhan SARS-CoV-2
112
genome has been affected by mutations and/or sequencing errors (the full length of
113
the reference genome is 29,903 nucleotides, based on sequence id NC_045512.2).
114
115
Mutation frequency in SARS-CoV-2 proteins
116
We show the frequency of mutations along the length of every SARS-CoV-2 protein,
117
reporting in the X-axis the amino acid position and on the Y-axis its frequency, either
118
as number of observed samples carrying the mutation, the vase 10 logarithm of that
119
number, or the percentage over all sequenced samples. In the example in Figure 1 B,
120
we show the most frequent mutations affecting the viral Spike protein S,
121
distinguishing silent mutations and amino acid-changing mutations (including the
122
introduction of STOP codons). For Spike, the mutations appear to be evenly
123
distributed in frequency along the protein length, with the most frequent mutation
124
being the aforementioned D614G. Mouse-over functionality is provided to allow the
125
user to identify the selected mutation (N439K in Figure 1 B).
126
127
Page 6 of 14
The SARS-CoV-2 mutation table
128
The user can visualize or download the full table of mutations on which the webtool
129
operates (Figure 2 A). This table is frequently updated and allows the user to specify a
130
worldwide or a country-specific dataset. The table also provides a Search function to
131
look for specific variants or sample ids, and it can be viewed online or downloaded in
132
full as a Comma-Separated Values (CSV) file.
133
The table shows every mutation in a specific geographical area, reporting:
134
• the GISAID sample ID (useful for cross-reference with the GISAID database
135
and other analyses based on it, e.g. Nexstrain).
136
• The country where the sample was collected.
137
• The position of the mutation, on the reference genome (refpos) and on the
138
sample (qpos).
139
• The sequence at the mutation site, on the reference genome (refvar) and on the
140
sample (qvar).
141
• The length of the sample genome (qlength); the reference genome is 29,903
142
nucleotides long.
143
• The protein affected by the mutation or, if the mutation is extragenic, the
144
denomination of the untranslated region (UTR), e.g. 5’UTR or 3’UTR.
145
• The effect of the mutation on the amino acid sequence of the protein (variant).
146
This uses the canonical mutational standard, indicating the original amino
147
acid(s), the position on the protein, and the mutated amino acid(s). An asterisk
148
(*) indicates a STOP codon, while the letters indicate amino acids in IUPAC
149
code. E.g. a mutation P315L indicates a leucine mutation (L) on the amino
150
acid location 315, normally occupied by a proline (P). Nucleotide mutations
151
can be silent, i.e. not yielding any aminoacidic change, e.g. the mutation
152
F106F, where the codon of phenylalanine 106 is affected but without changing
153
the corresponding amino acid. As in the previous column, mutations affecting
154
UTR regions are simply reported as the location of the nucleotide affected.
155
Page 7 of 14
• The class of the mutation, of which there are currently 10 types:
156
o SNP: a change of one or more nucleotides, determining a change in
157
amino acid sequence.
158
o SNP_stop: a change of one or more nucleotides, yielding the generation
159
of one or more STOP codons.
160
o SNP_silent: a change of one or more nucleotides with no effect in
161
protein sequence.
162
o Insertion: the insertion of 3 (or multiples of 3) nucleotides, causing the
163
addition of 1 or more amino acids to the protein sequence.
164
o Insertion_stop: the insertion of 3 (or multiples of 3) nucleotides, causing
165
the generation of a novel STOP codon.
166
o Insertion_frameshift: the insertion of nucleotides not as multiples of 3,
167
causing a frameshift mutation.
168
o Deletion: the deletion of 3 (or multiples of 3) nucleotides, causing the
169
removal of 1 or more amino acids to the protein sequence.
170
o Deletion_stop: the removal of 3 (or multiples of 3) nucleotides, causing
171
the generation of a novel STOP codon.
172
o Deletion_frameshift: the deletion of nucleotides not as multiples of 3,
173
causing a frameshift mutation.
174
o Extragenic: a mutation affecting intergenic or UTR regions.
175
• The extended annotation of the protein region affected by the mutation (e.g.
176
“Spike” for “S” or “Predicted phosphoesterase, papain-like proteinase” for
177
NSP3, the Non-Structural Protein 3).
178
• The
full
name
of
the
variant
(varname),
in
the
format
179
proteinName:AApositionAA, to allow for unique denomination of viral
180
proteome variants.
181
182
Mutational overview
183
Page 8 of 14
The user is also provided with a general overview of the mutational status of the
184
selected country or the entire world (Figure 2 B). Six bar plots provide a summary and
185
highlights of the dataset, specifically:
186
• The most mutated samples, indicating which samples (in GISAID IDs) carry
187
the highest number of mutations
188
• The overall mutations per sample, indicating the distributions of mutations per
189
sample. It has been previously reported [10] that the current mode for
190
mutation number compared to the reference NC_045512.2 genome is 7.5.
191
• The most frequent events per class. Classes are the same as reported in the
192
mutation table and are described in the previous paragraph.
193
• The most frequent events per type. Individual mutation types are shown as
194
specific nucleotides events, e.g. cytosine to thymidine transitions (C>T),
195
guanosine to thymidine transversion (G>T) or even multinucleotide mutations
196
(e.g. GGG>AAC, observed in the Nucleocapsid protein). As reported before,
197
nucleotide transitions seem to be the most abundant SARS-CoV-2 type of
198
mutational event worldwide [11].
199
• The most frequent events, either in nucleotide coordinates or in aminoacidic
200
coordinates. Currently, the most frequent events are four mutations affecting
201
SARS-CoV-2 genomes belonging to clade G, which is the most sequenced
202
worldwide and predominant in Europe. These mutations are A23403G
203
(associated to the already mentioned D614G mutation in the Spike protein),
204
C3037T, C14408T and C241T.
205
206
Analysis of mutations over time
207
The coronapp webtool allows users to monitor the abundance and frequency of any
208
SARS-CoV-2 mutation in any country specified (Figure 3). Both plots in this section
209
report continuous dates on the X-axis, starting on the day of the first collected
210
SARS-CoV-2 genome available on GISAID: December 24, 2019.
211
Page 9 of 14
The “abundance” plot reports on the Y-axis the number of samples carrying a
212
selected mutation in a particular day, in the specified country or worldwide. Since the
213
date reported is the collection date (not the submission date to the GISAID database),
214
there is usually a drop towards the right part of the plot, as there are fewer sequences
215
collected approaching the day of the analysis. The “frequency” plot on the other hand
216
normalizes the abundance of mutations by the total number of sequences generated on
217
each day. The plot currently shows a sharp increase in clade G-associated mutations
218
(e.g. S:D614G), as these mutations are most frequent in countries where sequencing is
219
more pervasive (e.g. United Kingdom).
220
221
Annotation of user-provided SARS-CoV-2 genomic sequence.
222
coronapp provides the user with the optional possibility of uploading one or more
223
SARS-CoV-2 genomic sequences, which can be complete or partial. The format of
224
the sequences is standard FASTA, and an example input FASTA containing 12
225
sequences is provided (Figure 1 A). The analysis is almost instantaneous and shows
226
an overall breakdown of the most mutated samples and most frequent mutations in the
227
dataset. Moreover, a full table of all detected mutations is provided: this can be
228
visualized and searched on the web browser or downloaded as a standard CSV file.
229
Finally, a mutation frequency plot is provided, allowing the user to visualize mutation
230
frequency in selected proteins.
231
The user can easily return to the worldwide status of the app by refreshing or
232
reopening the page.
233
234
Discussion
235
Our webtool coronapp provides a fast, simple tool to annotate user-provided
236
SARS-CoV-2 genomes and visualize all mutations currently present in viral
237
sequences collected worldwide. The results provided by this instrument can have
238
several applications. The main purpose of coronapp is to help medical laboratories at
239
Page 10 of 14
the front lines of COVID-19 fight with the opportunity to quickly define the
240
mutational status of their sequences, even without dedicated bioinformaticians.
241
Additionally, it enables scientists to perform mutational co-variance analyses and
242
to identify present and future significant functional interactions between viral
243
mutations, as previously attempted for the influenza virus and the human
244
immunodeficiency virus (HIV) [17]. Another application is the identification of the
245
most frequent mutations in specific protein regions: for example, our tool can quickly
246
identify that the most frequent mutation in the Spike protein, D614G, lies outside the
247
known interaction domain with the human protein ACE2, which spans roughly
248
between Spike amino acids 330 and 530 [18].
249
A recently published structural model simulating the effect of the D614G mutation
250
on the 3D structure of the spike protein has suggested that this mutation may result in
251
a viral particle which binds ACE2 receptors less efficiently, due to the masking of the
252
host receptor binding site on viral spikes [12]. The same researchers have reported a
253
possible correlation of the D614G form with increased case fatality rates,
254
hypothesizing that this mutation may lead to a viral form which is better suited to
255
escape immunologic surveillance by eliciting a lower immunologic response [12].
256
The coronapp analysis highlighted in Figure 1 B shows that a mutation located within
257
the Spike/ACE2 interaction domain is the change of Asparagine (N) to a Lysine (K)
258
in position 439 of the Spike sequence; this mutation could affect the protein folding or
259
its affinity with ACE2, as Asparagine is less charged than the basic amino acid
260
Lysine.
261
One of coronapp’s key strengths is to help prioritize scientific efforts on specific
262
aminoacidic variations that could affect the efficacy of anti-viral strategies or the
263
development of a vaccine by tracking the most frequent mutations in the population.
264
A further novelty of coronapp is that it provides a mean to assess the growth or
265
decline of specific mutations over time, in order to identify possible viral adaptation
266
mechanisms.
267
Page 11 of 14
We provide not only the webtool, but also all the underlying code for the
268
annotation and visualization steps on a public Github repository, in order to help other
269
computational scientists in the ongoing battle against COVID-19. Furthermore, the
270
coronapp structure and concept could be expanded to other current and future
271
pathogens as well (e.g. the seasonal influenza or HIV), in order to monitor the
272
mutational status across proteins, countries and time.
273
274
Materials and methods
275
The webtool coronapp has been developed using the programming language R and is
276
based on a Shiny server (current version 1.4.0.2) running on R version 3.6.1. The app
277
is based on two distinct files, server.R and ui.R, managing the server functionalities
278
and the browser visualization processes, respectively. The results visualization utilizes
279
both basic R functions and Shiny functionalities; for tooltip functionality, coronapp
280
uses the R package googleVis v0.6.4, which provides an interface between R and the
281
Google visualization API [19].
282
The core of the annotation of the user-provided sequences rests in the NUCMER
283
(Nucleotide Mummer) alignment tool, version 3.1 [20]. Nucmer output is processed
284
by UNIX and R scripts provided in Github within the server.R file.
285
286
287
Page 12 of 14
Authors’ contributions
288
DM drafted the manuscript and performed the mutational analysis and literature
289
search. LT developed the user interface code and drafted the methodological parts of
290
the manuscript. EF worked on graphical interface of the webtool. FR wrote the
291
manuscript and performed literature search. FMG designed the study, developed the
292
server code, finalized the manuscript and provided financial support. All authors
293
tested the webtool and provided original contributions to its development. All authors
294
read and approve the final manuscript.
295
296
Competing interests
297
The authors have declared no competing interests.
298
299
Acknowledgements
300
We thank the Italian Ministry of University and Research for their support, under the
301
Montalcini Grant 2016.
302
303
References
304
[1] Guan W-J, Ni Z-Y, Hu Y, Liang W-H, Ou C-Q, He J-X, et al. Clinical
305
Characteristics of Coronavirus Disease 2019 in China. N Engl J Med
306
2020;382:1708–20. https://doi.org/10.1056/NEJMoa2002032.
307
[2] Phua J, Weng L, Ling L, Egi M, Lim C-M, Divatia JV, et al. Intensive care
308
management of coronavirus disease 2019 (COVID-19): challenges and
309
recommendations. Lancet Respir Med 2020.
310
https://doi.org/10.1016/S2213-2600(20)30161-2.
311
[3] Zhang H, Penninger JM, Li Y, Zhong N, Slutsky AS. Angiotensin-converting
312
enzyme 2 (ACE2) as a SARS-CoV-2 receptor: molecular mechanisms and
313
potential therapeutic target. Intensive Care Med 2020;46:586–90.
314
https://doi.org/10.1007/s00134-020-05985-9.
315
[4] Guzzi PH, Mercatelli D, Ceraolo C, Giorgi FM. Master Regulator Analysis of the
316
SARS-CoV-2/Human Interactome. J Clin Med 2020;9:982.
317
https://doi.org/10.3390/jcm9040982.
318
[5] Ou X, Liu Y, Lei X, Li P, Mi D, Ren L, et al. Characterization of spike
319
glycoprotein of SARS-CoV-2 on virus entry and its immune cross-reactivity with
320
Page 13 of 14
SARS-CoV. Nat Commun 2020;11:1620.
321
https://doi.org/10.1038/s41467-020-15562-9.
322
[6] Blanco-Melo D, Nilsson-Payant BE, Liu W-C, Uhl S, Hoagland D, Møller R, et
323
al. Imbalanced Host Response to SARS-CoV-2 Drives Development of
324
COVID-19. Cell 2020;181:1036-1045.e9.
325
https://doi.org/10.1016/j.cell.2020.04.026.
326
[7] Tu Y-F, Chien C-S, Yarmishyn AA, Lin Y-Y, Luo Y-H, Lin Y-T, et al. A Review
327
of SARS-CoV-2 and the Ongoing Clinical Trials. Int J Mol Sci 2020;21.
328
https://doi.org/10.3390/ijms21072657.
329
[8] Koyama T, Weeraratne D, Snowdon JL, Parida L. Emergence of Drift Variants
330
That May Affect COVID-19 Vaccine Development and Antibody Treatment.
331
Pathog Basel Switz 2020;9. https://doi.org/10.3390/pathogens9050324.
332
[9] Ceraolo C, Giorgi FM. Genomic variance of the 2019�nCoV coronavirus. J Med
333
Virol 2020;92:522–8. https://doi.org/10.1002/jmv.25700.
334
[10] Mercatelli D, Giorgi FM. Geographic and Genomic Distribution of SARS-CoV-2
335
Mutations. Preprints; 2020. https://doi.org/10.20944/preprints202004.0529.v1.
336
[11] Chiara M, Horner DS, Gissi C, Pesole G. Comparative genomics suggests limited
337
variability and similar evolutionary patterns between major clades of
338
SARS-CoV-2. BioRxiv; 2020. https://doi.org/10.1101/2020.03.30.016790.
339
[12] Becerra-Flores M, Cardozo T. SARS-CoV-2 viral spike G614 mutation exhibits
340
higher case fatality rate. Int J Clin Pract 2020. https://doi.org/10.1111/ijcp.13525.
341
[13] Gudbjartsson DF, Helgason A, Jonsson H, Magnusson OT, Melsted P, Norddahl
342
GL, et al. Spread of SARS-CoV-2 in the Icelandic Population. N Engl J Med
343
2020. https://doi.org/10.1056/NEJMoa2006100.
344
[14] Fauver JR, Petrone ME, Hodcroft EB, Shioda K, Ehrlich HY, Watts AG, et al.
345
Coast-to-Coast Spread of SARS-CoV-2 during the Early Epidemic in the United
346
States. Cell 2020;181:990-996.e5. https://doi.org/10.1016/j.cell.2020.04.021.
347
[15] Hadfield J, Megill C, Bell SM, Huddleston J, Potter B, Callender C, et al.
348
Nextstrain: real-time tracking of pathogen evolution. Bioinformatics
349
2018;34:4121–3. https://doi.org/10.1093/bioinformatics/bty407.
350
[16] Max Roser EO-O Hannah Ritchie, Hasell J. Coronavirus Pandemic (COVID-19).
351
Our World Data 2020.
352
[17] Sruthi CK, Prakash MK. Statistical characteristics of amino acid covariance as
353
possible descriptors of viral genomic complexity. Sci Rep 2019;9:18410.
354
https://doi.org/10.1038/s41598-019-54720-y.
355
[18] Lan J, Ge J, Yu J, Shan S, Zhou H, Fan S, et al. Structure of the SARS-CoV-2
356
spike receptor-binding domain bound to the ACE2 receptor. Nature
357
2020;581:215–20. https://doi.org/10.1038/s41586-020-2180-5.
358
[19] Gesmann M, de Castillo D. Using the Google visualisation API with R. R J
359
2011;3:40–44.
360
Page 14 of 14
[20] Delcher AL, Salzberg SL, Phillippy AM. Using MUMmer to Identify Similar
361
Regions in Large Sequence Sets. Curr Protoc Bioinforma 2003;00:10.3.1-10.3.18.
362
https://doi.org/10.1002/0471250953.bi1003s00.
363
364
Figure legends
365
Figure 1 Overview of coronapp
366
A. Screenshot of the entry page of coronapp showing the basic tool description, the
367
interface to upload user-provided sequences and the overall summary of the mutations
368
detected worldwide. B. Common interface showing mutation frequency in
369
SARS-CoV-2 proteins, with occurrence of the mutation on the Y-axis and protein
370
coordinate on the Y-axis. Red dots indicate amino acid (aa)-changing mutations, and
371
blue dots indicate silent mutations. Tooltip functionality is also provided to identify
372
and quantify each mutation on mouse-over.
373
374
Figure 2 Mutation table and overview in coronapp
375
A. Result table of coronapp, available both for worldwide-precomputed and
376
user-input analyses. A “download full table” button is provided to allow the user to
377
perform larger-scale analyses autonomously. B. Barplots showing the most mutated
378
samples, overall sample mutations and most frequent mutation events, classes and
379
types. This analysis is also available both for worldwide-precomputed and user-input
380
analyses.
381
382
Figure 3 Analysis of mutations over time
383
The final output of coronapp, showing the abundance of each user-specified mutation
384
in any user-specified country (or worldwide). The left graph indicates the absolute
385
amount of samples where the indicated mutation is detected. The right graph shows
386
the same data normalized by total number of samples, as the percentage of samples
387
sequenced in a specific day and carrying the mutation.
388
389
Mutation frequency for protein S (Spike) in World
Maac
5 BB siler
e
4
N439K
aa: 439
3 occurrence: 2.111 e
status: aa chan:
. ge ce ° e
2 ° a
septa Sey im
600 900 1,200
updated May 30, 2020
Number of samples: 29668
Number of distinct mutated loci: 10458
Total number of mutational events: 203292
Select Select Protein:
@ Logio
ee s .
World » CO Percentage
Mutation frequency for protein S (Spike) in World
mews §=§=CUrrent status Of SARS-CoV-2 mutational data
: ; Number of samples: 29668
Provide your own (multi)FASTA file Number of distinct mutated loci: 10458
Total number of mutational events: 203292
Select Select Protein:
ect Brovein Logio
Country:
s -
Word + Percentage
Mutation frequency for protein S (Spike) in World
Mutation frequency for protein S (Spike) in World
lB aa ch
5 BB silent
e
4
N439K
aa: 439
3 occurrence: 2.111 e
status: aa change
Occurrence of event (Log10)
0 bume ONO) 6 O1MEO © OF CIOL) 6 AOD @ U6 (CHT (A1ES (O18 0 @ OG) @)@@lNtees)
0 300 600 900 1,200
Current Status of SARS-CoV-2 mutational data
Mutation frequency for protein S (Spike) in World
Most frequent events per class
he
Overall mutations per sample
Most mutated samples
= 10K
00
50K
40K
f ox
3 20K
doys gNS
uonafep
quebenxe
walls dNS
NS
nr of mutations
x “« =~ *¥
so 6 F¥ AW
20
15
10
5
°
‘SUO;JE;NW jo JU
t2terr ISI Ida
Sizeer ISI 1d3
steszy ISI Id3
Most frequent events (protein)
i
Most frequent events (nucleotide)
Mh
Most frequent events per type
10K
8K
6K
4K
2K
00
10K
8K
6K
4K
2K
00
50K
SERS S
393 =
‘sojdwes yo su
NOSSI:ZASN
SS8Sd'Z7dSN
ALSz:2¢48O
AQPPA!GZL SN
ALETASN
YE0ZOU'N
49014:€dSN
WeYLns
TWhed:d2tdSN
Ori9a's
iw
oo
“<O1lV
wo
OSL
vo
ovy<999
Lo
ow
4<o
Show 10 ~ entries
i
sample country refpos, refvar qvar qpos qlength protein variant varclass annotation varname
EPLISL_415706 Switzerland 4 A T 4 29903 S'uTR 4 extragenic S'UTR:4
EPLISL_415706 Switzerland 241 c T 241 29903 s'uTR 241 extragenic S'UTR:241
EPLISL_415706 Switzerland 3037 c T 3037 29903 NSPS FL06F ‘SNP_silent Fredicted pretphocstarsee, pape: NSP3:F 106F
like proteinase
EPLISL_415706 Switzerland 14408 C T 14408 29903 NSPI2 = PS1AL SNP RNA-dependent RNAPOmeras®, cosh ps 141
post-ribosomal frameshift
RNA-dependent :
EPLISL_415706 Switzerland 15324 c Tt 15324 29903 NSPI2b NGION ‘SNP_silent a eee NSP12b:N619N
post-ribosomal frameshift
EPLISL_415706 Switzerland 23403 A 6 23403 29903 s 06146 ‘SNP ‘Spike S:06146
EPL_ISL_416497 France 4 A Tt 4 29862 s'uTR 4 extragenic SUTRA
EPLISL_416497 France 241 c T 241 29862 sume 241 extragenic S'UTR:241
EPLISL_416497 France 2416 c T 2416 29862 NSP2 YS37Y ‘SNP_silent —_ Non-Structural protein 2 NSP2:YS3TY
EPLISL_416497 France 3037 c Tt 3037 29862 NSP Fi06F ‘SNP_silent Predicted phoey erase, papeln- NSP3:F 106
like proteinase
‘Showing 1 to 10 of 192,208 entries Previous | 1 2 3 4 s 19221 Next
Mutational overview for United Kingdom
Most mutated samples Overall mutations per sample Most frequent events per class
20
« 50K
§ 15 ox 40K
5 10 g 30K
3 20K
Bs § 4k :
= 3 10K Oo
° = x 00 ———
25 e588
00 — * 2 3 fy
onwooon? 2 3 a &
nr of mutations i
Most frequent events per type Most frequent events (nucleotide) Most frequent events (protein)
50K 10K 10K
3 40K 8K 8K
5 30K 6K 6K
3 20K 3 4K 3 4k
= 10k Bo = 2k = 2k
0 Meo 0 ales 00 seGGe:
j
°o
oe
°
C14805T
G26144
G11083T
55
S:D614G
NSP12b:P314L
‘446Y
SUTR:241
NSP3:F106F
N:RG203KR
NSP6:L37F
NSP 12b:Y
ORF3a:G251V
NSP2:P585S
NSP2:I559V
‘Showing 1 to 10 of 192,208 entries
$:D614G abundance in World
800
600 ~|
3
=
pa}oejep suopeynw jo su
T
g
8 8 8 g g °
Peouenbes sajdwes jo %
$:D614G abundance in World
$:D614G frequency in World
S:D614G frequency in World
'D614G abundance in World
s 3 3 g &
Peouenbes sajdwes jo %
!
2
ofa Bate
cn
.y
4¢
T T T
g g g g
Pa}sjep SUOHE;NW 4O JU
O1-S0-0:
%0-S0-0
82-P0-0:
22-v0-0
€2-€0-0:
L1-€0°0
11-€0-0)
0-0-0
82-200:
22-20-0:
91-20-0
01-20-0:
%0-20-0
62-10-0:
€2-10°0
2110-0:
bb-b0-0)
S0-10-0:
O€-Z1-6
ve-21-6
01-S0-0:
¥0-S0-0
22-¥0-0
| 2020 | : A Web Application to Annotate and Monitor SARS-CoV-2 Mutations | 10.1101/2020.05.31.124966 | [
"Mercatelli Daniele",
"Triboli Luca",
"Fornasari Eleonora",
"Ray Forest",
"Giorgi Federico M."
] | creative-commons |
Dopamine enhances model-free credit assignment through boosting of
retrospective model-based inference
Lorenz Deserno1,2,3*, Rani Moran1,2*, Jochen Michely1,2, Ying Lee1,2, Peter Dayan1,4,5,
Raymond J. Dolan1,2
1 Max Planck UCL Centre for Computational Psychiatry and Ageing Research, London, United Kingdom;
2 Wellcome Centre for Human Neuroimaging, Institute of Neurology, University College London, London,
United Kingdom;
3 Department of Child and Adolescent Psychiatry, Psychotherapy and Psychosomatics, University of
Würzburg, Würzburg, Germany;
4 Max Planck Institute for Biological Cybernetics, Max Planck Ring 8, 72076 Tübingen, Germany;
5 University of Tübingen, 72074 Tübingen, Germany
* denotes equal contribution to authorship
Contact Information:
Prof. Dr. Lorenz Deserno
Dr. Rani Moran
Margarete-Höppel-Platz 1
10-12 Russel Square
97080 Würzburg, Germany
London WC1B5EH, United Kingdom
deserno_l@ukw.de
rani.moran@gmail.com
2
Abstract
Dopamine is implicated in signalling model-free (MF) reward prediction errors and various
aspects of model-based (MB) credit assignment and choice. Recently, we showed that
cooperative interactions between MB and MF systems include guidance of MF credit
assignment by MB inference. Here, we used a double-blind, placebo-controlled, within-
subjects design to test the hypothesis that enhancing dopamine levels, using levodopa, boosts
the guidance of MF credit assignment by MB inference. We found that levodopa enhanced
retrospective guidance of MF credit assignment by MB inference, without impacting on MF and
MB influences per se. This drug effect positively correlated with working memory, but only in
a context where reward needed to be recalled for MF credit assignment. The dopaminergic
enhancement in MB-MF interactions correlated negatively with a dopamine-dependent change
in MB credit assignment, possibly reflecting a potential trade-off between these two
components of behavioural control. Thus, our findings demonstrate that dopamine boosts MB
inference during guidance of MF learning, supported in part by working memory, but trading-
off with a dopaminergic enhancement of MB credit assignment. The findings highlight a novel
role for a DA influence on MB-MF interactions.
Introduction
1
Dual system theories of reinforcement learning (RL) propose behaviour is controlled by
2
competitive and cooperative interactions between a prospective, model-based (MB), planning
3
system and a retrospective, model-free (MF), value-caching system (Daw and Dayan, 2014;
4
Dolan and Dayan, 2013). MF value-caching is driven by reward prediction error (RPE)
5
signalling via phasic dopamine (DA, Montague et al., 1996; Schultz et al., 1997; Steinberg et
6
al., 2013), a finding mirrored in human neuroimaging studies (D’Ardenne et al., 2008;
7
O’Doherty et al., 2004). While DA RPEs are assumed to train MF values (a process we refer
8
to as MF credit assignment or MFCA), there is evidence that DA neuromodulation also impacts
9
MB learning (MB credit assignment or MBCA) and control (Doll et al., 2012; Langdon et al.,
10
2018). For example, the activity of DA neurons reflects MB values (Sadacca et al., 2016), DA
11
RPEs reflect hidden-state inference (Starkweather et al., 2017), and optogenetic activation
12
and silencing of DA neurons impact the efficacy of MB learning (Sharpe et al., 2017). Human
13
studies also show that higher DA levels are linked to enhanced MB influences (Deserno et al.,
14
2015; Doll et al., 2016; Sharp et al., 2016; Wunderlich et al., 2012), which was confirmed in a
15
non-human animal study (Groman et al., 2019), potentially mediated by a modulation in the
16
efficiency of working memory or motivation.
17
RL theory has proposed cooperative interactions between MB and MF systems,
18
including the idea that a MB controller instructs a MF system about the structure of the
19
environment (Daw and Dayan, 2014; Mattar and Daw, 2018; Sutton, 1991). For instance,
20
inferences made in a MB manner can disambiguate different possible states of the world in
21
cases in which the MF system is otherwise unable to learn properly because it does not know
22
the state. We recently provided empirical evidence for this sort of MB-MF cooperation, showing
23
that retrospective MB inference guides MFCA via provision of knowledge regarding the
24
environment’s transition structure (Moran et al., 2019). Given DA’s contribution to both MF and
25
MB systems, we set out to examine whether this aspect of MB-MF cooperation is subject to
26
DA influence.
27
4
To address this question, we used a dual-outcome bandit task (Moran et al., 2019) in
1
a double-blind, placebo-controlled, within-subjects pharmacological study, employing
2
levodopa to boost the brain’s overall DA levels. This task allows a separate measurement of
3
MB and MF systems, and specifically probes guidance of MF learning based on MB knowledge
4
of the environmental transition structure. Our hypothesis was that enhancing DA would
5
strengthen the guidance of MFCA by MB inference. Importantly, at the time MB inference is
6
possible, some rewards are no longer perceptually available to participants. Thus, we expected
7
that DA-induced boosting of the MB guidance of MFCA would depend on working memory
8
capacity exclusively for perceptually absent rewards. Finally, in light of previous reports that
9
levodopa enhanced MB influences (Sharp et al., 2016; Wunderlich et al., 2012), we examined
10
whether this is also true for our dual-outcome task, expecting that inter-individual differences
11
in the effect of boosting DA on MB influences and on MB guidance of MFCA would be related.
12
Foreshadowing our results, we found that boosting DA levels via levodopa enhanced
13
guidance of a MFCA by MB inference, an effect moderated by inter-individual differences in
14
working memory but only when reward needed to be recalled. While boosting DA did not alter
15
the overall influence of a MB system on choice per se, the drug effects on guidance of MFCA
16
by MB inference and on MB choice were negatively correlated.
17
18
Results
19
Study design and task logic. We conducted a placebo-controlled, double-blind,
20
within-subjects pharmacological study using levodopa to enhance presynaptic DA levels, as
21
in previous studies (Chowdhury et al., 2013; Wunderlich et al., 2012). Participants were tested
22
twice, once under the influence of 150mg levodopa, and once on placebo, where drug order
23
was counterbalanced across individuals (n=62, Figure 1A; cf. Methods). On each lab visit,
24
participants performed a task first introduced previously by Moran et al. (2019). The task was
25
framed as a treasure hunt game called the “Magic Castle”. Initially, participants were trained
26
extensively on a transition structure between states, under a cover narrative of four vehicles
27
5
and four destinations. Subjects learned that each vehicle (state) travelled to two different
1
sequential destinations in a random order (Figure 1B). The mapping of vehicles and
2
destinations remained stationary throughout a session, but the two test sessions featured
3
different vehicles and destinations. At each destination, participants could potentially earn a
4
reward with a probability that drifted across trials according to four independent random walks
5
(Figure 1C).
6
7
Figure 1. A) Illustration of within-subjects design. On each of two testing days, approximately
8
7 days apart, participants started with either a medical screening and brief physical exam (day
9
1) or a working memory test (day 2). Subsequently they drank an orange squash containing
10
either levodopa (D) or placebo (P). B) Task structure of the Magic Castle Game. Following a
11
choice of vehicle, participants “travelled” to two associated destinations. Each vehicle shared
12
a destination with another vehicle. At each destination, participants could win a reward (10
13
pence) with a probability that drifted slowly as Gaussian random walks, illustrated in C). D)
14
Depiction of trial types and sequences. (1) On standard trials (2/3 of the trials), participants
15
made a choice out of two options in trial-n (max. choice 2s). The choice was then highlighted
16
(.25s) and participants subsequently visited each destination (.5s displayed alone). Reward, if
17
obtained, was overlaid to each of the destinations for 1s. (2) On uncertainty trials, participants
18
made a choice between two pairs of vehicles. Subsequently, the ghost nominates, unbeknown
19
to the participant, one vehicle out of the chosen pair. Firstly, the participant is presented the
20
destination shared by the chosen pair of vehicles and this destination is therefore non-
21
informative about the ghost’s nominee. Secondly, the destination unique to the ghost-
22
nominated vehicle is then shown. This second destination is informative because it enables
23
inference of the ghost’s nominee with perfect certainty based on a MB inference that relies on
24
task transition structure. Trial timing was identical for standard and uncertainty trials.
25
26
6
The task included two trial types (Figure 1D): (1) standard trials (2/3 of the trials) and
1
(2) uncertainty trials (1/3 of the trials). On standard trials, participants were offered two vehicles
2
and upon choosing one, they visited both its associated destinations where they could earn
3
rewards. On uncertainty trials, participants likewise chose a pair of vehicles (from two offered
4
vehicle-pairs). Next, an unseen ghost randomly nominated a choice of one of the vehicles in
5
the chosen pair, and a visit to its two destinations followed. Critically, participants were not
6
privy to which vehicle was nominated by the ghost. However, they could resolve this
7
uncertainty after seeing both visited destinations based on their knowledge of task transition
8
structure. We refer to this as retrospective MB inference. Such inference can only occur after
9
exposure to the second destination, as only then can subjects know which of the two vehicles
10
the ghost had originally selected.
11
We first present ‘model-agnostic’ analyses focusing on how events on trial n affect
12
choices on trial n+1. This allows identification of MF and MB choice signatures, the guidance
13
of MFCA by retrospective MB inference, and, crucially, whether these signatures varied as a
14
function of drug treatment (levodopa vs. placebo). These analyses are supported by validating
15
simulations using computational models as provided in a later section.
16
Logic of model-free and model-based contributions to choices. A MF system
17
updates values based on earned rewards only for a chosen vehicle (illustrated in Figure 2A).
18
A MB system does not maintain and update values for the vehicles directly. Instead, the MB
19
system updates the values of destinations and calculates prospectively on-demand values for
20
each offered vehicle (see computational modelling). This enables the MB system to generalize
21
value across vehicles which share a common destination (illustrated in Figure 2B).
22
No evidence of dopaminergic modulation for MF choice repetition. Consider a pair
23
of standard trials n and n+1 for which the vehicle chosen on the former is also offered on the
24
latter, against another vehicle (Figure 2A). The two vehicles offered on trial n+1 reach a
25
common destination, but the vehicle previously chosen on trial n also visits a unique
26
destination. In a logistic mixed effects model, we regressed a choice repetition of this vehicle
27
on whether the common and/or unique destinations were rewarded on trial n (reward/non-
28
7
reward) and on drug status (levodopa/placebo). Replicating a previous finding (Moran et al.,
1
2019), we found a main effect for common reward (b=0.67, t(7251)=9.14, p<.001). This effect
2
constitutes MF choice repetition, as the MB system appraises that the common destination
3
favours both trial n+1 vehicles (see Figure S1 for validating simulations). As expected on both
4
MB and MF grounds, there was a main effect for unique reward (b=1.54, t(7251)=17.40,
5
p<.001). There was no drug x common-reward interaction (b=0.07, t(7251)=.67, p=.500),
6
providing no evidence for a drug-induced change in MF choice repetition on standard trials
7
(Figure, 2B). None of the remaining (main or interaction) effects were significant (Table S1).
8
9
Figure 2. A) Illustration of MF choice repetition. We consider only standard trials n+1 that offer
10
for choice the standard trial n chosen vehicle (e.g. green antique car) alongside another vehicle
11
(e.g. yellow racing car), sharing a common destination. Following choice of a vehicle in trial n
12
(framed in red), participants visited two destinations of which one can be labelled on trial n+1
13
as common to both offered vehicles (C, e.g. forest, which was also rewarded in the example)
14
and the other labelled as unique (U, e.g. city highway, unrewarded in this example) to the trial
15
n chosen vehicle. The trial n common-destination reward effect on the probability to repeat the
16
previously chosen vehicle constitutes a MF choice repetition. B) The empirical reward effect
17
at the common destination (i.e., the difference between rewarded and unrewarded on trial n,
18
see Figure S3 for a more detailed plot) on repetition probability in trial n+1 is plotted for placebo
19
and levodopa (L-DOPA) conditions. There was a positive common-reward main effect and this
20
reward effect did not differ significantly between placebo and levodopa conditions. C)
21
Illustration of the MB contribution. We considered only standard trials n+1 that excluded from
22
the choice set the standard trial n chosen vehicle (e.g. green antique car). One of the vehicles
23
offered on trial n+1 shared one destination in common with the trial-n chosen vehicle (e.g.,
24
yellow racing car and we term its choice a generalization). A reward (on trial n) effect for the
25
common destination on the probability to generalize on trial n+1 constitutes a signature of MB
26
choice generalization. D) The empirical reward effect at the common destination (i.e., the
27
difference between rewarded and unrewarded, see Figure S3 for a more detailed plot) on
28
generalization probability is plotted for placebo and levodopa conditions. E) In the regression
29
analysis described in the text, we also include the current (subject- and trial-specific) state of
30
8
the drifting reward probabilities (at the common destination) because we previously found this
1
was necessary to control for temporal auto correlations in rewards (Moran et al., 2019). For
2
completeness, we plot beta regression weights of reward versus no reward at the common
3
destination (indicated as MB) and for the common reward probability (RewProbC) each for
4
placebo and levodopa conditions. No significant interaction with drug session was observed.
5
Error bars correspond to SEM reflecting variability between participants.
6
7
No evidence of dopaminergic modulation for MB choice generalization. Consider a
8
standard trial-n+1, which excludes the vehicle chosen on trial n from the choice set. This trial-
9
n chosen vehicle shares a destination with one of the trial-n+1 offered vehicles, allowing an
10
analysis of MB choice generalization. Using a logistic mixed effects model, wherein we
11
regressed choice generalization on trial-n rewards at the common destination, on the current
12
reward probability of the common destination and on drug session, replicated our previous
13
finding (Moran et al., 2019) of a positive main effect for the common-reward (b=0.40,
14
t(7177)=6.22, p<.001). This positive common trial-n reward-effect on choice constitutes a MB
15
choice generalization (even after controlling for the drifting reward probability at the common
16
destination, see Figure S1 for validating simulations). The common-reward x drug interaction
17
was not significant (b=0.05, t(7177)=0.39, p=.695), providing no evidence for a drug-induced
18
change in MB choice (Figure 2D & E). Except for the main effect of the drifting reward
19
probability at the common destination, no other effects were significant (Table S1).
20
In summary, we replicate previous findings (Moran et al., 2019) of mutual MF and MB
21
contributions to choices. There was no evidence, however, that these contributions were
22
modulated by levodopa.
23
24
Retrospective MB inference guides MFCA. We next addressed our main question:
25
Does levodopa administration boost a MB guidance of MFCA through a retrospective MB
26
inference? In an uncertainty trial, participants choose one out of the two pairs of vehicles
27
(Figure 1D). Next, a ghost randomly nominates a vehicle from the chosen pair (Figure 3).
28
Participants then observe a destination common to both of the vehicles of the chosen pair,
29
followed by a destination unique to the ghost-nominated vehicle. As participants are
30
uninformed about the ghost nominee, they have a 50-50% belief initially and observing the first
31
9
destination is non-informative with respect to the ghost’s nominee (as it is shared between
1
vehicles). Critically, following observation of the second destination, a MB system can infer the
2
ghost-nominated vehicle with absolute certainty based upon knowledge of the task transition
3
structure. Thus, the second destination is retrospectively informative with respect to inference
4
of the ghost’s nominee. Subsequently, the inferred vehicle information can be shared with a
5
MF system to direct MFCA towards the ghost-nominated vehicle. We predicted guidance of
6
MFCA occurs for both vehicles in the chosen pair, but to a different extent. Specifically,
7
guidance of MFCA for the ghost-nominated, as compared to the ghost-rejected, vehicle would
8
support an hypothesis that retrospective MB inference preferentially guides MFCA (Moran et
9
al., 2019). See Figure S2 for validating model simulations. Our novel hypothesis here is that
10
this effect will be strengthened under levodopa as compared to placebo, which we examine,
11
firstly via the informative and, secondly, via the non-informative destination.
12
Dopamine enhances preferential guidance of MFCA for the informative
13
destination. MFCA for the ghost-nominated vehicle is tested in a “repeat” standard trial n+1
14
that follows an uncertainty trial n, as depicted in Figure 3 A1. MFCA of the ghost-rejected
15
vehicle is examined in a “switch” standard trial n+1 following an uncertainty trial n, as depicted
16
in Figure 3 A2. For a detailed analysis of repeat and switch trials, see Supplementary
17
Information (SI) and Figure S4. The key metric of interest for our drug analysis is the contrast
18
between MFCA for ghost-nominated versus ghost-rejected vehicles, based on the reward
19
effects at the informative destination in repeat and switch trials (repeat or ghost-nominated /
20
switch or ghost-rejected), separately for each nomination trial type (repeat/switch) x drug
21
condition (levodopa /placebo) (Figure 3B). In a mixed effects model (Table S2), we found no
22
main effect either of nomination (b=.043 t(239)=1.60, p=.110) or of drug (b=.01, t(239)=.40,
23
p=.690). Crucially, we found a significant nomination x drug interaction (b=.11, t(239)=2.56,
24
p=.011). A simple effects analysis revealed a preferential MFCA of the ghost-nominated over
25
the ghost-rejected vehicle was significant under levodopa (b=.09, F(243,1)=9.07, p=.003) but
26
not under placebo (b=-.02, F(243,1)=.53, p=.472). This supports our hypothesis that levodopa
27
preferentially enhanced MFCA for the ghost-nominated, compared to ghost-rejected, vehicle
28
10
under the guidance of retrospective MB inference. The nomination x drug interaction was not
1
affected by session order (see Table S2).
2
3
Figure 3. In an uncertainty trial n, participants choose a pair of vehicles. The ghost nominates
4
one vehicle out of this pair (e.g., green antique car). Participants have a chance belief about
5
the ghost-nominated vehicle. The firstly presented destination holds no information about the
6
ghost-nominated vehicle, the non-informative (“N) destination. The destination presented
7
second enables retrospective MB inference about the ghost’s nomination and is therefore
8
informative (“I”). A1. Illustration of the repeat condition. The ghost-nominated vehicle (e.g.,
9
green antique car) is offered for choice in standard trial n+1 alongside a vehicle from the non-
10
chosen pair (e.g., blue building crane). A higher probability to repeat the ghost-nominated
11
vehicle in standard trial n+1 after a reward as compared to no reward at the informative
12
destination constitutes MFCA for the ghost’s nomination (GN). A2. Illustration of the switch
13
condition. The ghost-rejected vehicle (e.g., the yellow racing car) is offered for choice in
14
standard trial n+1 alongside a vehicle from the non-chosen pair (e.g. brown farming tractor). A
15
higher probability to choose the ghost-rejected vehicle in standard trial n+1 after a reward as
16
compared to no reward at the informative destination constitutes MFCA for the ghost’s
17
rejection (GR). Both ghost-based assignments depend on retrospective MB inference. B.
18
Preferential effect of retrospective MB inference on MFCA (effects of GN>GR) based on the
19
informative destination is enhanced under levodopa (L-Dopa) as compared to placebo. This is
20
indicated by a significant trial type (GN/GR) x drug (placebo/ levodopa) interaction. Under
21
levodopa, MFCA for GN is significantly higher than of GR, which is not the case under placebo
22
(see Figure S4 for a more detailed plot). C. Illustration of the clash condition. The previously
23
chosen pair is offered for choice in standard trial n+1. A higher probability to repeat the ghost-
24
nominated vehicle in standard trial n+1 following reward (relative to non-reward) at the non-
25
informative destination constitutes a signature of preferential MFCA for GN over GR. D. Choice
26
repetition in clash trial is plotted as a function of reward and drug-group (see Figure S5 for a
27
more detailed plot). While there was a main effect for drug, there was no interaction of non-
28
informative reward x drug, providing no evidence that drug modulated MFCA based on the
29
non-informative outcome. R+: reward; R-: non-reward. Error bars correspond to SEM reflecting
30
variability between participants.
31
32
11
Dopaminergic modulation of preferential MFCA for the non-informative
1
destination. A second means to examine MB influences over MFCA is to consider the non-
2
informative destination. In a standard “clash” trial-n+1 following an uncertainty trial-n, the
3
ghost-nominated vehicle is offered for choice alongside the ghost-rejected vehicle as depicted
4
in Figure 3C. We previously showed that a positive effect of reward at the non-informative
5
destination on choice repetition (i.e., a choice of the previously ghost-nominated vehicle)
6
implicates a preferential guidance of MFCA towards the ghost-nominated vehicle guided by
7
retrospective MB inference (Moran et al., 2019). In contrast, a MB system has knowledge that
8
a non-informative destination is common to both standard trial n+1 vehicles. Note, this effect
9
of reward at the non-informative destination can only occur when uncertainty about the ghost’s
10
nomination was resolved retrospectively, once the informative destination was encountered.
11
In a logistic mixed effects model, we regressed choice repetition on trial-n rewards at
12
informative and non-informative destinations as well as on drug session. A marginally
13
significant main effect for the reward at the non-informative destination provides some support
14
for preferential MFCA of the ghost-nominated vehicle (b=0.13, t(4861)=1.96, p=.051).
15
Additionally, we found a main effect for reward at the informative destination (b=1.01,
16
t(4861)=9.95, p<.001), as predicted by both the enhanced MFCA for the ghost-nominated
17
vehicle and by an MB contribution. The interaction effect between drug and non-informative
18
reward, however, was not significant (b=0.05, t(4861)=.39, p=.696, Figure 3D), nor were any
19
other interactions in the model (Table S2). This analysis yielded no evidence that levodopa
20
enhanced preferential guidance of MFCA based on reward at a non-informative destination.
21
Unexpectedly, we found a positive main effect of drug (b=0.15, t(4861)=2.31, p=.021, Figure
22
3D), indicating that participants’ tendency to repeat choices of the ghost-nominated vehicle
23
was generally enhanced under levodopa, but this finding that was only seen in this specific
24
subset of trials and could not be corroborated based on computational modelling. We further
25
dissect effects at the non-informative destination, in particular with respect to inter-individual
26
differences in working memory, using computational modelling.
27
12
Computational Modelling. One limitation of the analyses reported above is that they
1
isolate the effects of the immediately preceding trial on a current choice. However, values and
2
actions of RL agents are influenced by an entire task history and, to take account of such
3
extended effects, we formulated a computational model that specified the likelihood of choices
4
(Moran et al., 2019, also see Moran et al., in press, 2021). In brief, at choice, MF values (𝑄!")
5
of the two presented vehicles feed into a decision module. During learning, the MF system
6
updates 𝑄!" of the chosen vehicle based on earned rewards alone. By contrast, the MB
7
system prospectively calculates on-demand 𝑄!#-values for each offered vehicle based on an
8
arithmetic sum of the values of its two destinations:
9
(𝐸𝑞. 1) 𝑄!#(vehicle) = 𝑄!#(corresponding destination 1) + 𝑄!#(corresponding destination 2)
10
During learning, the MB system updates the values of the two visited destinations. We
11
refer to these updates as MB credit assignment (MBCA). Unlike MFCA, which does not
12
generalize credit from one vehicle to another, MBCA generalizes across the two vehicles which
13
share a common destination. Thus, when a reward is collected in the forest destination,
14
𝑄!#(forest) increases. As the forest is a shared destination, both vehicles that lead to this
15
destination benefit during ensuing calculations of the on-demand 𝑄!#-values. Critically, our
16
model included five free “MFCA parameters” of focal interest, quantifying the extent of MFCA
17
on standard trials (one parameter), on uncertainty trials (four parameters) for each of the
18
objects in the chosen pair (nominated/rejected), and for each destination (informative/non-
19
informative). We verified that the inclusion of these parameters was warranted using
20
systematic model comparisons. A description of the sub-models and the model selection
21
procedure is reported in the methods section and in Figure S6. We fitted our full model to each
22
participant’s data in drug and placebo sessions based on Maximum Likelihood Estimation (see
23
methods).
24
Absence of dopaminergic modulation for MBCA and MFCA on standard trials. In
25
line with our model-agnostic analyses of standard trials, we found positive contributions of
26
MFCA (parameter c$%&'(&)(
*+
; Fig. 4A) for both levodopa (M= 0.381, t(61)= 6.84, p<.001) and
27
13
placebo (M= 0.326, t(61)= 5.76 p<.001), with no difference between drug conditions (t(61)= -
1
0.78, p=.442). Likewise, MBCA (parameter 𝑐*, ; Fig. 4B) contributed positively for both
2
levodopa (M= 0.255, t(61)= 7.88, p<.001) and placebo (M= 0.29, t(61)= 8.88, p<.001), with no
3
significant difference between drugs (t(61)= 0.88, p=.3838). Thus, while both MBCA and MFCA
4
contribute to choice, there was no evidence for a drug-related modulation. Forgetting and
5
perseveration parameters of the model did not differ as a function of drug (see SI).
6
Levodopa enhances guidance of preferential MFCA by retrospective MB
7
inference on uncertainty trials. To test our key hypothesis, that guidance of preferential
8
MFCA by retrospective MB inference on uncertainty trials is enhanced by levodopa, we
9
focused on the four computational parameters that pertaining to MFCA on uncertainty trials
10
(c'-.,0'1-
*+
, c)23,0'1-
*+
, c'-.,'-'0'1-
*+
, c423,'-'0'1-
*+
, Figure 4B,C). In a mixed effects model, we regressed
11
these MFCA parameters on their underlying features: nomination (nominated / rejected),
12
informativeness (informative / non-informative) and drug session (levodopa / placebo).
13
Crucially, we found a positive nomination x drug interaction (b=0.10, t(480)=2.43, p=.015). A
14
simple effects analysis revealed preferential MFCA (the effect of nomination) to be significant
15
under levodopa (b=.13, F(488,1)=9.71, p=.002), and stronger than in the placebo condition
16
(b=0.08, F(488,1)=4.83, p=.029), indicating that preferential MFCA was stronger under
17
levodopa as compared to placebo. Importantly, this interaction was not qualified by a triple
18
interaction (b=.02, t(480)=0.32, p=.738), providing no evidence that the extent of preferential
19
MFCA differed for informative and non-informative outcomes. No other effect pertaining to drug
20
reached significance (Table S3).
21
To examine in more fine-grained detail whether a MFCA is indeed preferential, we
22
calculated, for each participant, in each session (drug/placebo), and for each level of
23
informativeness (informative/non-informative), the extent to which MFCA was preferential for
24
the ghost-nominated as opposed to the ghost-rejected vehicle (as quantified by c'-.,0'1-
*+
−
25
c)23,0'1-
*+
, c'-.,'-'0'1-
*+
− c423,'-'0'1-
*+
; Figure 4D). Using a mixed effects model, we regressed
26
preferential MFCA (PMFCA), based on MB guidance on informativeness and drug session.
27
14
We found a positive main effect for drug (b=0.10, t(240)= 2.41, p=.017), but neither the main
1
effect of informativeness (b=-0.03, t(240)=-0.57, p=.568) nor the informativeness x drug
2
interaction (b=.02, t(240)=0.33, p=.739) were significant. Using simple effects, MFCA preferred
3
the ghost-nominated vehicle in the levodopa condition (b= 0.15, F(1,244)= 15.45, p<.001),
4
while the same effect was only marginally significant in the placebo condition (b= 0.05,
5
F(1,244)= 2.86, one-sided p=.046). Thus, our computational modelling analysis indicates that
6
preferential MFCA is boosted by levodopa as compared to placebo across informative and
7
non-informative destinations.
8
9
Figure 4. Analyses based on estimated credit assignment (CA) parameters from
10
computational modelling. A) Model-free and model-based credit assignment parameters
11
(MFCA; MBCA) did not differ significantly for placebo and levodopa conditions. B) MFCA
12
parameters based on the informative outcome for the ghost-nominated and the ghost-rejected
13
destinations as a function of drug condition. D) Same as C but for the non-informative
14
destination. E) The extent to which MFCA prefers the nominated over the rejected vehicle for
15
each destination and drug condition. We name this preferential MFCA (PMFCA).
16
17
Drug effect correlates positively with working memory only for reward at the non-
18
informative destination. We hypothesized that working memory (WM) would moderate the
19
boosting effect of levodopa, but only based on reward at the non-informative destination. When
20
the informative destination is delivered on uncertainty trials, a MB system can infer the hidden
21
choice and guide PMFCA. PMFCA based on reward at the non-informative destination can
22
prefer the ghost-nominated vehicle only if it is at least partially postponed until uncertainty has
23
been resolved by retrospective MB inference, in other words after delivery of the informative
24
destination. At this time, reward received at the non-informative destination is no longer
25
perceptually available and needs to be recalled (as illustrated in Figure 5A). Subjects’ WM
26
15
capacity, as ascertained with the digit span test, showed a positive across-participants
1
Spearman correlation with the drug effect (levodopa vs placebo) on PMFCA in the non-
2
informative (r= .278, p=.029, Figure 5B), but not for the informative destination (r= -.057,
3
p=.659, Figure 5B). The difference between these correlations was significant (p=. 044,
4
permutation test; see methods). There was no significant correlation of WM capacity with drug-
5
induced change in MBCA or with MBCA at levodopa or placebo (see SI).
6
Inter-individual differences in drug effects. Previous studies, using a task that
7
cannot dissociate cooperative and competitive interactions between MB and MF systems,
8
reported that boosting DA levels leads to enhanced MB choices (Sharp et al., 2016;
9
Wunderlich et al., 2012), an effect we did not observe at a group level on our measure of
10
MBCA. To explore the possibility that drug effects in different task conditions (guidance of
11
MFCA vs. MBCA) are related, we analyzed inter-individual differences in the effects of boosting
12
DA levels on guidance of MFCA and on MBCA. Because WM capacity correlated positively
13
with drug effects at the non-informative destination as reported above, we included WM in the
14
analysis of inter-individual differences in drug effects. Thus, we regressed DA-dependent
15
differences (levodopa vs placebo) in PMFCA against informativeness, DA-dependent
16
differences in MBCA and WM capacity. This model revealed an informativeness x MBCA x
17
WM interaction (b=0.16, t(116)=2.16, p=.032). To unpack the interaction, we ran the model
18
separately at high and low WM capacity based on a median split. In individuals with high WM
19
capacity, this revealed a negative main effect of MBCA (b=-0.13, t(48)=-2.45, p=.018, see
20
Figure 5C) which was not qualified by an interaction between informativeness x MBCA (b=-
21
0.07, t(48)=-0.86, p=.40). This means that, for high WM individuals, the drug-effects on PMFCA
22
and MBCA are negatively related for informative and non-informative destinations. In contrast,
23
in individuals with low WM capacity, there was a significant negative informativeness x MBCA
24
interaction (b=-0.23, t(68)=-2.43, p=.018; Figure 5D). A simple effects analysis revealed that
25
the drug-effect on MBCA had a significant negative relation on the drug effect on PMFCA for
26
the informative destination (b=-.18, F(1,68)=6.13, p=.015; Figure 5D) but not for the non-
27
informative destination (b=.05, F(1,68)=0.42, p=.517; Figure 5D). Using model-agnostic
28
16
metrics of DA-dependent change in guidance of MFCA and in MB choice, the negative
1
correlation was also significant (see Figure S7). These inter-individual differences may reflect
2
a trade-off between PMFCA and MBCA under boosted DA levels.
3
4
Figure 5. Inter-individual differences. A) Illustration of MFCA based on rewards at informative
5
and non-informative destination. The latter is likely to depend more on memory recall because
6
the reward is no longer perceptually available when MFCA can take place (after state
7
uncertainty was resolved). B) Scatter plots of the drug effect (levodopa minus placebo) on
8
preferential MFCA (∆ PMFCA) based on the informative destination reward and for the non-
9
informative destination reward against working memory (WM). C) Scatter plot of the drug effect
10
(levodopa minus placebo) on preferential MFCA (∆ PMFCA) based on the informative
11
destination reward (info, red) and for the non-informative destination reward (non-info, blue)
12
against drug-induced change in MBCA (∆ MBCA) at high working memory (WM) capacity. D)
13
Same scatter plot as in C) but at low working memory (WM) capacity. In panels B, C and D
14
regression lines are dashed. r refers to the Spearman correlation coefficient in panel B and
15
Pearson correlation coefficient in C and D.
16
17
18
19
17
Discussion
1
We show that enhancing dopamine boosted the guidance of model-free credit
2
assignment by retrospective model-based inference. This pharmacological effect was
3
associated with higher working memory capacity just for rewards that were no longer
4
perceptually available and had to be recalled for credit assignment to be correct. Whereas both
5
MF and MB influences were unaffected by the drug manipulation at the group level, analysis
6
of inter-individual differences in drug effects showed that enhanced guidance of MFCA by
7
retrospective MB inference was negatively correlated with drug-related change in MBCA. The
8
findings provide, to our knowledge, the first human evidence that DA directly influences
9
cooperative interactions between MB and MF systems, highlighting a novel role for DA in how
10
MB information guides MFCA.
11
The effect of levodopa on prefrontal DA levels can lead to the enhancement of general aspects
12
of cognition, for example WM (Cools and D’Esposito, 2011), probably depending on DA
13
synthesis capacity in an inverted U-curved manner. The latter is likely to be important for
14
supporting the computationally sophisticated operation of a MB system (Otto et al., 2013). One
15
might therefore expect a primary drug effect on prefrontal DA to result in boosted MB
16
influences (Sharpe et al., 2017; Wunderlich et al., 2012) – but we found no such influence.
17
Equally, a long-standing proposal that phasic DA relates to a MF learning signal might predict
18
that the main effect of the drug would be to speed or bias MF learning (Pessiglione et al.,
19
2006). We observed no such effect, nor has it been seen in two previous studies (Sharp et al.,
20
2016; Wunderlich et al., 2012). Instead, we found levodopa had a more specific influence,
21
impacting the preferential MB guidance of MFCA in a situation where individuals needed to
22
rely on retrospective MB inference to resolve state uncertainty. Thus, MB instruction about
23
what (unobserved or inferred) state the MF system might learn about, was boosted under
24
levodopa. In other words, DA boosts an exploitation of a model of task structure so as to
25
facilitate retrospective learning about the past. These findings indicate an enhanced integration
26
of MB information in DA signalling (Sadacca et al., 2016). Our results thus may provide a fine-
27
grained view of the various processes involved – with the specificities of our task allowing us
28
18
to separate out a rather particular component of WM, and an important, but restricted influence
1
of MB information on MFCA.
2
First, preferential MFCA based on reward at the uninformative destination can only take
3
place after seeing the informative destination and inferring the ghost’s choice. Thus, the
4
uninformative destination’s reward has to be maintained in WM to support preferential MFCA.
5
In other words, an ability to maintain information in working memory is a prerequisite for a DA-
6
dependent boosting of preferential MFCA based on the uninformative destination. In line with
7
this, we found a DA-boosting of MB guidance of MFCA depended on WM for the non-
8
informative destination alone. This underlines the importance of accounting for inter-individual
9
differences in supportive cognitive processes particularly when it comes to providing a detailed
10
understanding of DA drug effects of interest (Cools, 2019; Kroemer et al., 2019).
11
Second, given that the information about the uninformative destination is stored in WM,
12
what might be the neural mechanisms associated with its use in MB guidance of MFCA. Animal
13
and human work points to a crucial role for orbitofrontal cortex in representing the model of a
14
task model, including unobserved and inferred states, and in guiding behaviour accordingly
15
(Howard et al., 2020; Jones et al., 2012; Schuck et al., 2016). This orbitofrontal function has
16
also been related to the degree of sequential offline replay in the hippocampus (Schuck and
17
Niv, 2019). Theoretical treatments of hippocampal offline neural replay proposes it informs
18
credit assignment based on RPE (Mattar and Daw, 2018), a suggestion gaining support in
19
recent empirical evidence in humans (Eldar et al., 2020; Liu et al., 2019, 2020). In our task,
20
offline replay seems especially necessary to support preferential MFCA based on the first,
21
uninformative, destination, because at this stage participants are still uncertain about the
22
ghost’s choice. Under this account, we would predict enhanced offline replay (during rest
23
between trials) of the non-informative destination (including its reward) and the inferred ghost’s
24
choice under the influence of L-Dopa. Whether this enhanced replay occurs indirectly, via the
25
interaction with WM, or is also a direct consequence of the L-Dopa is a pressing question for
26
future work.
27
19
Previous studies, using a task not designed to test cooperative interactions between
1
MB and MF systems (Daw et al., 2011), indicated a positive relationship between boosted DA
2
and MB contributions to choice (Deserno et al., 2015; Doll et al., 2012; Sharp et al., 2016;
3
Wunderlich et al., 2012). While MB choice contributions were not elevated at the group level
4
by the drug in our data, we found a negative correlation between drug-related change on these
5
contributions and on MB guidance of MFCA, in keeping with a trade-off between DA influences
6
on these two components of behavioural control. In arbitrating between MB choice and
7
retrospective MB inference to guide MFCA, participants need to weigh their respective
8
cognitive costs vs. instrumental value. In independent recent work, a balance of costs and
9
benefits was recently shown to be modulated by DA (Westbrook et al., 2020). Future studies
10
will be needed to detail how the relative costs of planning vs. retrospective state-inference are
11
influenced by DA, which can also inform DA contributions to trade-offs pertaining to strategy
12
selection.
13
A limitation in our study is that guidance of informative MFCA by MB inference was
14
significant in the levodopa condition alone but not in the placebo condition in model-agnostic
15
measures (which are based on a subset of trials and consider only very recent influences on
16
choice). However, computational modelling, informed by the entire trial-by-trial history of one’s
17
experiences is arguably more sensitive, and this consideration enabled us to capture a
18
preferential guidance of MFCA by MB inference also in the placebo condition.
19
In sum, our study provides first evidence that DA enhances cooperative interactions
20
between MB and MF systems. The finding provides a unified perspective on previous
21
research in humans and animals, suggesting a closely integrated architecture of how MF and
22
MB systems interact under the guidance of DA-mediated so as to improve learning. DA-
23
mediated cooperation between MB and MF control is a potentially exciting target for
24
disentangling the precise role played by MB control in the development of impulsive and
25
compulsive psychiatric symptoms.
26
27
20
Methods
1
Procedures. A total of 64 participants (32 females) completed a bandit at each of the
2
two sessions with drug or placebo in counterbalanced order in a double-blinded design. One
3
participant failed to reach required performance during training (see below) and task data could
4
not be collected. Out of remaining 63 participants, one participant experienced side effects
5
during task performance and was therefore excluded. Results reported above are based on a
6
sample of n=62. All participants attended on two sessions approximately 1 week apart.
7
Participants were screened to have no psychiatric or somatic condition, no regular intake of
8
medication before invitation and received a short on-site medical screening at the beginning
9
of their day 1 visit. At the beginning of the day 2 visit, they performed a working memory test,
10
the digit span, which was thus only collected once.
11
Drug protocol. The order of drug and placebo was counterbalanced. The protocol
12
contained two decision-making tasks, which started at least 60min after ingestion of either
13
levodopa (150 mg of levodopa + 37.5 mg of benserazide dispersed in orange squash) or
14
placebo (orange squash alone with ascorbic acid). Benserazide reduces peripheral
15
metabolism of levodopa, thus, leads to higher levels of DA in the brain and minimizes side
16
effects such as nausea and vomiting. To achieve comparable drug absorption across
17
individuals, subjects were instructed not to eat for up to 2h before commencing the study.
18
Repeated physiological measurements (blood pressure and heart rate) and subjective mood
19
rating scales were recorded under placebo and levodopa. A doctor prepared the orange
20
squash such that data collection was double-blinded.
21
Task Description. Participants were introduced to a minor variant of a task developed
22
by Moran et al. (2019) using pictures of vehicles and destinations rather than objects and
23
coloured rooms, and lasting slightly less time. The was presented as a treasure hunt called
24
the ‘Magic Castle”. Before playing the main task, all participants were instructed that they can
25
choose out of four vehicles from the Magic Castle’s garage that each vehicle could take them
26
to two destinations (see Figure 1B). The mapping between vehicles and destination was
27
randomly created for each participant and each session (sessions also had different sets of
28
21
stimuli) but remained fixed for one session. They were then extensively trained on the specific
1
vehicle-destination mapping. In this training, participants first saw a vehicle and had to press
2
the space bar in self-paced time to subsequently visits the two associated destinations in
3
random. The initial training run contained 12 repetitions per vehicle-destination mapping (48
4
trials). This training was followed by two types of each 8 quiz trials which asked to match one
5
destination out of two to a vehicle or to match a vehicle out of two to a destination (time limit
6
of 3sec). Each quiz trial had to be answered correctly and in time otherwise another training
7
session was started with only 4 repetitions per vehicle-destination mapping (16 trials) followed
8
again by the quiz. This procedure was repeated until participants passed all quiz. Participants
9
were then introduced to the general structure of standard trials of bandit task (18 practice
10
trials). This was followed by instructions introducing the ghost trials, which were complemented
11
by another 16 practice trials including standard and ghost trials. Before starting the main
12
experiment, participants performed a shorter refresher training of the vehicle-destination
13
mapping with 4 repetitions per vehicle-destination mapping followed by the same quiz trials to
14
passed as described above. In case of not passing at this stage, the refresher training was
15
repeated with 2 repetitions per vehicle-destination mapping until the quiz was passed.
16
During the subsequent main task, participants should try to maximize their earnings. In
17
each trial, they could probabilistically find a treasure (reward) at each of the two destinations
18
(worth 1 penny). Reward probabilities varied over time independently for each of the four
19
destinations according to Gaussian random walks with boundaries at p=0 and p=1 and a
20
standard deviation of .025 per trial (Figure 1C). Random walks were generated anew per
21
participant and session. A total of 360 trials split in 5 blocks of each 72 trials were played with
22
short enforced breaks between blocks. Two of three trials were ‘standard trials’, in which a
23
random pair of objects was offered for choice sharing one common outcome (choice time <=
24
2s). After making a choice, they visited each destination subsequently in random order. Each
25
destination was presented for 1s and overlaid with treasure or not (indicating a reward or not).
26
The lag between the logged choice and the first destination as well as between first and second
27
destinations was 500ms. Every third trial was an “uncertainty trial” in which two disjoint pairs
28
22
of vehicles were offered for choice. Crucially, each of the presented pairs of vehicles shared
1
one common outcome. Participants were told before the main task that after their choice of a
2
pair of vehicles, the ghost of the Magic Castle would randomly pick one vehicle out of the
3
chosen pair. Because this ghost was transparent, participants could not see the ghost’s choice.
4
However, participants visited the two destinations subsequently and collected treasure reward
5
(or not). Essentially, when the ghost nominated a vehicle, the common destination was
6
presented first and the destination unique to this vehicle was presented second. At this time of
7
presentation of the unique destination, participants could retrospectively infer the choice made
8
by the ghost. Trial timing was identical for standard and ghost trials. The 120 standard trials
9
following a previous trial n-1 standard trial included 30 presentations of each of the four eligible
10
pairs of vehicles in a random order. The 120 uncertainty trials included 60 presentations of the
11
two eligible pairings in a random order. The standard trials following uncertainty trials were
12
defined according to the observed transition based on the (ghost’s) choice in the preceding
13
(uncertainty) trial. These 120 trials contained 40 presentations of each of the “repeat”, “switch”
14
or “clash” trial types in a random order. A repeat trial presented the ghost-nominated object
15
alongside its vertical counterpart, a switch trial presented the ghost-rejected object alongside
16
its vertical counterpart and a clash trial presented the previously selected pair.
17
Model-agnostic analysis. Model agnostic analyses were performed with logistic
18
mixed effects models using MATLAB’s “fitglme” function with participants serving as random
19
effects with a free covariance matrix. All models included the variable ORDER as regressor
20
(coded as +.5 for the first and -.5 for the second session) to control for unspecific effects and
21
participants (PART) served as random effects. Details of are reported in Table S1.
22
The analysis of MF and MB contributions is restricted to standard trials followed by a
23
standard trial. For MF contributions, we consider only a trial-n+1, which offers the trial-n chosen
24
object for choice (against another object). Regressors C (common destination) and U (unique
25
destination) indicated whether rewards were received at trial n (coded as +.5 for reward and -
26
.5 for no reward) and were included to predict the variable REPEAT indicating whether the
27
previously chosen vehicle was repeated or not. The variable DRUG was included as regressor
28
23
indicating within-subject Levodopa or placebo session (coded as +.5 for levopdopa and -.5 for
1
placebo). The model, in Wilkinson notation, can be found on Table S1. For MB contributions,
2
we specifically examined trials in which the trial-n chosen vehicle was excluded on trial n+1.
3
The regressors C, PART and DRUG were coded as for the analysis of the MF contribution.
4
One additional regressor P was included, which coded the reward probability of the common
5
destination and was centralized by subtracting .5. These regressors were included to predict
6
the variable GENERALIZE indicated whether the choice on trial n+1 was generalized
7
(choosing the vehicle not shown in trial n+1 that shares a destination with the trial-n chosen
8
vehicle). The model, in Wilkinson notation, can be found on Table S1.
9
The analysis of how retrospective MB inference preferentially guides MFCA focused
10
on standard trials following uncertainty trials. The key analysis reported above focuses on MF
11
choice repetition for the ghost-nominated in contrast to the ghost-rejected vehicle. This was
12
achieved by extracting empirical choice proportions from “repeat trials” and from “switch trials”.
13
More specifically, we computed the proportion of repeating or switching after a reward minus
14
no reward at the informative destination averaged across rewards at the non-informative
15
destination (reflecting the main effect of the informative destination, “I”) for each trial type.
16
These two metrics were subjected to a mixed-effects models as dependent variable and with
17
TYPE (nominated / rejected coded as +.5 and -.5) and, as before, DRUG and PART as
18
predictors. The model, in Wilkinson notation, can be found on Table S2. A detailed analysis
19
using separate mixed effects models for repeat and switch conditions is reported in the SI.
20
Another model-agnostic analysis examined learning for the ghost-nominated and -
21
rejected vehicles based on the uncertainty trial n non-informative destination and therefore
22
focused on n+1 “clash” trials, which offer for choice the same pair of objects as chosen on the
23
previous uncertainty trial (the ghost-nominated and ghost-rejected objects). Choice repetition
24
was defined as choice of the ghost-nominated vehicle from uncertainty trial n indicated by the
25
variable REPEAT. Regressors PART, N, I and DRUG are coded as previously. The model, in
26
Wilkinson notation, can be found on Table S2.
27
24
Computational Models. We formulated a hybrid RL model to account for the series of
1
choices for each participant. In the model, choices are contributed by both the MB and MF
2
systems. The MF system caches a 𝑄!"-value for each vehicle, subsequently retrieved when
3
the vehicle is offered for choice. During learning on standard trials, following reward-feedback,
4
rewards from the two visited destinations are used to update the 𝑄!"-value for the chosen
5
vehicle as follows:
6
(𝐸𝑞. 2) 𝑄!"(chosen vehicle) ← (1 − 𝑓*+) ∗ 𝑄!"(chosen vehicle) + 𝑐56789749
*+
∗ (𝑟: + 𝑟;)
7
where 𝑐56789749
*+
is a free MFCA parameter on standard trials and the r’s are the rewards for
8
each of the two obtained outcomes (coded as 1 for reward or -1 for non-reward) and
9
𝑓*+(between 0-1) is a free parameter corresponding to forgetting in the MF system.
10
During learning on uncertainty trials, the MF values of the ghost nominated and ghost
11
rejected options were updated according to:
12
(𝐸𝑞. 3) 𝑄!"(nominated vehicle)
13
← (1 − 𝑓*+) ∗ 𝑄!"(nominated vehicle) + 𝑐'-.,0'1-
*+
∗ 𝑟0'1- + 𝑐'-.,'-'0'1-
*+
14
∗ 𝑟'-'0'1-
15
(𝐸𝑞. 4) 𝑄!"(rejected vehicle)
16
← (1 − 𝑓*+) ∗ 𝑄!"(rejected vehicle) + 𝑐)23,0'1-
*+
∗ 𝑟0'1- + 𝑐)23,'-'0'1-
*+
∗ 𝑟'-'0'1-
17
Where the c’s are free MFCA parameters on uncertainty trials for each destination
18
(informative/non-informative) and vehicle type (ghost nominated/rejected) in the chosen pair.
19
The r’s are rewards (once more, coded as 1 or -1) for the informative and non-informative
20
outcomes.
21
The MF values of the remaining vehicles (3 on standard trials; 2 on uncertainty trials)
22
were subject to forgetting:
23
(𝐸𝑞. 5) 𝑄!"(non chosen vehicles) ← (1 − 𝑓*+) ∗ 𝑄!"(non chosen vehicles)
24
25
Unlike MF, the MB system maintains 𝑄!#-values for the four different destinations.
1
During choices the 𝑄!#- value for each offered vehicle is calculated based on the transition
2
structure (i.e., the two destinations associated with a vehicle):
3
(𝐸𝑞. 6) 𝑄!#(vehicle) = 𝑄!#(detstination 1) + 𝑄!#(detstination 2)
4
Following a choice (on both standard and uncertainty trials), the MB system updates the 𝑄!#-
5
values of each of the two observed destination based on its own reward:
6
(𝐸𝑞 8) 𝑄!#(destination) ← (1 − 𝑓*,) ∗ 𝑄!#(destination) + 𝑐*, ∗ 𝑟
7
Where 𝑓*, (bet. 0-1) is a free parameter corresponding to forgetting in the MB system, 𝑐*, is
8
a free MBCA parameter and r corresponds to the reward (1 or -1) obtained at the destination.
9
Our model additionally included progressive perseveration for vehicles. After each
10
standard trial the perseveration values of each of the 4 vehicles updated according to
11
(𝐸𝑞. 9) 𝑃𝐸𝑅𝑆(vehicle) ← (1 − 𝑓<) ∗ 𝑃𝐸𝑅𝑆(vehicle) + pr$%&'(&)( ∗ 1=2>0?@2A?>-$2'
12
Where 1=2>0?@2A?>-$2' is the chosen vehicle indicator, pr$%&'(&)( is a free perseveration
13
parameter for standard trials, and 𝑓<(bet. 0-1) is a free perseveration forgetting parameter.
14
Similarly after each uncertainty trials perseverations values were updated according to:
15
(𝐸𝑞. 10) 𝑃𝐸𝑅𝑆(vehicle) ← (1 − 𝑓<) ∗ 𝑃𝐸𝑅𝑆(vehicle) + prB'?2)%&0'%C ∗ 1=2>0?@2A'-.
16
where 1=2>0?@2A'-. is the ghost-nominated vehicle indicator, and prB'?2)%&0'%C is a free
17
perseveration parameter for uncertainty trials.
18
During a standard trial choice a net Q value was calculated for each offered vehicle:
19
(𝐸𝑞. 11) 𝑄'2%(vehicle) = 𝑄!#(vehicle) + 𝑄!"(vehicle) + 𝑃𝐸𝑅𝑆(vehicle)
20
Similarly, during an uncertainty-trial choice the 𝑄'2% value of each offered vehicle-pair was
21
calculated as a sum of the MB, MF and PERS values of that pair. MF, MB, and PERS values
22
for a vehicle-pair in turn were each calculated as the corresponding average value of the two
23
vehicles in that pair. For example:
24
26
(𝐸𝑞. 12) 𝑄!"(vehicle pair) ← 𝑄!"(vehicle 1) + 𝑄!"(vehicle 2)
2
1
The 𝑄'2% values for the 2 vehicles offered for choice on standard trials are then injected
2
into a softmax choice rule such that the probability to choose an option is:
3
(𝐸𝑞. 13) 𝑃𝑟𝑜𝑏(vehicle) =
𝑒D!"#(=2>0?@2)
𝑒[D!"#(=2>0?@2)HD!"#(-%>2) =2>0?@2)]
4
Similarly, on uncertainty trials the probability to choice a vehicle pair was based on softmaxing
5
the net Q-values of the two offered pairs. 𝑄!" and 𝑃𝐸𝑅𝑆 person-values and 𝑄!# vegetables-
6
values where initialized to 0 at the beginning of the experiment.
7
Model Comparison and Fitting. Our full hybrid agents, which allowed for contributions
8
from both an MB and an MF system, served as a super-model in a family of six nested sub-
9
models of interest: 1) a pure MB model, which was obtained by setting the contribution of the
10
MF to 0 (i.e. c$%&'(&)(
*+
= c'-.,0'1-
*+
= c'-.,'-'0'1-
*+
= c)23,0'1-
*+
= c)23,'-'0'1-
*+
= 0), 2) a pure MF-
11
action model, which was obtained by setting the contribution of the MB system to choices to 0
12
(i.e. 𝑐*, = 0; Note that in this model, MB inference was still allowed to guide MF inference),
13
3) a ‘no informativeness effect on MFCA’ sub-model obtained by constraining equality between
14
the MFCA for the informative and non-informative destination (i.e., c'-.,0'1-
*+
= c'-.,'-'0'1-
*+
,
15
c)23,0'1-
*+
= c)23,'-'0'1-
*+
), 4) a ‘no MB guided MFCA’ sub-model obtained by constraining equality
16
between the MFCA parameters, for both the informative and non-informative destination, for
17
the ghost-nominated and rejected objects (c'-.,0'1-
*+
= c)23,0'1-
*+
, c'-.,'-'0'1-
*+
= c423,'-'0'1-
*+
), 5) a
18
‘no MB guidance of MFCA for the informative outcome’ obtained by constraining equality
19
between the MFCA parameters for the ghost-nominated and ghost-rejected objects for the
20
informative outcome (c'-.,0'1-
*+
= c)23,0'1-
*+
) , and 6) a ‘no MB guidance of MFCA for the non-
21
informative outcome’ which was similar to 5 but for the non-informative outcome (c'-.,'-'0'1-
*+
=
22
c)23,'-'0'1-
*+
).
23
We conducted a bootstrapped generalized likelihood ratio test, BGLRT (Moran and
24
Goshen-Gottstein, 2015), for the super-model vs. each of the sub-models separately. In a
25
27
nutshell, this method is based on the classical-statistics hypothesis testing approach and
1
specifically on the generalized-likelihood ratio test (GLRT). However, whereas GLRT assumes
2
asymptotic Chi-squared null distribution for the log-likelihood improvement of a super model
3
over a sub-model, in BGLRT these distributions are derived empirically based on a parametric
4
bootstrap method. In each of our model comparison the sub model serves as the H0 null
5
hypothesis whereas the full model as the alternative H1 hypothesis. For each participant and
6
drug condition, we created 1001 synthetic experimental sessions by simulating the sub-agent
7
with the ML parameters on novel trial sequences which were generated as in the actual data.
8
We next fitted both the super-agent and the sub-agent to each synthetic dataset and calculated
9
the improvement in twice the logarithm of the likelihood for the full model. For each participant
10
and drug condition, these 1001 likelihood-improvement values served as a null distribution to
11
reject the sub-model. The p-value for each participant in each drug condition was calculated
12
based on the proportion of synthetic dataset for which the twice logarithm of the likelihood-
13
improvement was at least as large as the empirical improvement. Additionally, we performed
14
the model comparison at the group level. We repeated the following 10,000 times. For each
15
participant and drug condition we chose randomly, and uniformly, one of his/her 1,000
16
synthetic twice log-likelihood super-model improvements and we summed across participant
17
and drug conditions. These 10,000 obtained values constitute the distribution of group super-
18
model likelihood improvement under the null hypothesis that a sub-model imposes. We then
19
calculated the p-value for rejecting the sub-agent at the group level as the proportion of
20
synthetic datasets for which the super-agent twice logarithm of the likelihood improvement was
21
larger or equal to the empirical improvement in super-model, summed across participants.
22
Results, as display in Figure S6 in detail, fully supported the use of our full model including all
23
effects of interest regarding MFCA in uncertainty trials.
24
We next fit our choice models to the data of each individual, separately for each drug
25
condition (levodopa/placebo) maximizing the likelihood (ML) of their choices (we optimized
26
likelihood using MATLAB’s ‘fmincon’, with 200 random starting points per participant * drug
27
28
condition; Table S4 for distribution best-fitting parameters). See Table S2 for the distribution
1
of full model’s fitted parameters.
2
Model simulations. To generate model predictions with respect to choices, we
3
simulated for each participant and each drug condition, 25 synthetic experimental sessions
4
(novel trial sequences were generated as in the actual experiment), based on ML parameters
5
obtained from the corresponding model fits. We then analysed these data in the same way as
6
the original empirical data (but with datasets that were 25 times larger, as compared to the
7
empirical data, per participant). Results are reported in Figures S1 and S2 of the SI. We also
8
tested recoverability of model parameters (see Figure S7).
9
Analysis of model parameters. All models included the variable ORDER as regressor
10
(coded as +.5 for the first and -.5 for the second session) to control for unspecific effects and
11
participants (PART) served as random effects. Details of are reported in Table S2.
12
For each participant in each drug condition, we obtained, based on the full model, four
13
MFCA parameter estimates corresponding to destination (informative/non-informative) and
14
vehicle (nominated/rejected) types. We conducted a mixed effects model (again implemented
15
with MATLAB’s function “fitglme”) with TYPE (nominated/rejected coded as +.5 and -.5), INFO
16
(informative/non-informative coded as +.5 and -.5) and DRUG (drug/placebo coded as +.5 and
17
-.5) as regressors. The model, in Wilkinson notation, can be found on Table S3.
18
After finding significant drug by NOM * DRUG interaction, we followed this up in detail:
19
we calculated for each participant in each drug condition and for each destination type the
20
“preferential MFCA” (denoted PMFCA) effect as the difference between the corresponding
21
nominated and rejected MFCA parameters. We next ran a mixed effects model for PMFCA.
22
Our regressors where the destination type (denoted INFO; coded as before), and DRUG
23
(coded as before). The model, in Wilkinson notation, can be found on Table S3.
24
Correlations between preferential MFCA and WM. For each participant and
25
destination type (Informative/non-informative), we contrasted the “preferential MFCA”
26
estimates (as defined in the previous section) for levodopa minus placebo to obtain a drug-
27
29
induced PMFCA effect. For each destination, we calculated across-participants Spearman
1
correlations between these drug induced effects and WM. We compared the two correlations
2
(for informative and non-informative destinations) using a permutation test. First, we z-scored
3
the PMFCA separately for each destination type. Next we repeated the following steps (1-3),
4
10,000 times: 1) For each participant we randomly reshuffled (independent of other
5
participants) the outcome type labels “informative” and “non-informative”, 2) We calculated the
6
“synthetic” Spearman correlations between drug induced PMFCA effects and WM for each
7
outcome type subject to the relabelling scheme and, 3) We subtracted the two correlations
8
(non-informative minus informative). These 10,000 correlation-differences constituted a null
9
distribution for testing the null hypothesis that the two correlations are equal. Finally, we
10
calculated the p-value for testing the hypothesis of a stronger correlation for the non-
11
informative destination as the percentage (of the 10,000) synthetic correlation-differences that
12
were at least as large (in absolute value) as the empirical correlation-difference.
13
Relationship between drug effects. We used the same score for drug-dependent
14
change in PMFCA (levodopa minus placebo) and regressed it against informativeness, drug-
15
dependent change in MBCA and working memory capacity in a mixed effects model.
16
17
18
Acknowledgements. RJD is supported by a Wellcome Trust Investigator Award
(098362/Z/12/Z) under which the above study was carried out. This work was carried out whilst
R.J.D. was in receipt of a Lundbeck Visiting Professorship (R290-2018-2804) to the Danish
Research Centre for Magnetic Resonance. RM is supported by the Max Panck Society and
LD was at the time when the study was performed. The UCL-Max Planck Centre for
Computational Psychiatry and Ageing is funded by a joint initiative between UCL and the Max
Planck Society. RJD and LD are supported by a grant from the German Research Foundation
(DFG TRR 265, project A02) and YL was at the time when the study was performed. For the
purpose of Open Access, the authors have applied a CC BY public copyright license to any
Author Accepted Manuscript version arising from this submission.
31
References
1
Chowdhury, R., Guitart-Masip, M., Lambert, C., Dayan, P., Huys, Q., Duzel, E., and Dolan,
2
R.J. (2013). Dopamine restores reward prediction errors in old age. Nat Neurosci 16, 648–
3
653.
4
Cools, R. (2019). Chemistry of the Adaptive Mind: Lessons from Dopamine. Neuron 104,
5
113–131.
6
Cools, R., and D’Esposito, M. (2011). Inverted-U-shaped dopamine actions on human
7
working memory and cognitive control. Biol Psychiatry 69, e113-25.
8
D’Ardenne, K., McClure, S.M., Nystrom, L.E., and Cohen, J.D. (2008). BOLD responses
9
reflecting dopaminergic signals in the human ventral tegmental area. Science 319, 1264–
10
1267.
11
Daw, N.D., and Dayan, P. (2014). The algorithmic anatomy of model-based evaluation.
12
Philos Trans R Soc Lond B Biol Sci 369.
13
Daw, N.D., Gershman, S.J., Seymour, B., Dayan, P., and Dolan, R.J. (2011). Model-based
14
influences on humans’ choices and striatal prediction errors. Neuron 69, 1204–1215.
15
Deserno, L., Huys, Q., Boehme, R., Buchert, R., Heinze, H.J., Grace, A.A., Dolan, R.J.,
16
Heinz, A., and Schlagenhauf, F. (2015). Ventral striatal presynaptic dopamine reflects
17
behavioral and neural signatures of model-based control during sequential decision-making.
18
Proc Natl Acad Sci U S A.
19
Dolan, R.J., and Dayan, P. (2013). Goals and habits in the brain. Neuron 80, 312–325.
20
Doll, B.B., Simon, D.A., and Daw, N.D. (2012). The ubiquity of model-based reinforcement
21
learning. Curr Opin Neurobiol 22, 1075–1081.
22
Doll, B.B., Bath, K.G., Daw, N.D., and Frank, M.J. (2016). Variability in Dopamine Genes
23
Dissociates Model-Based and Model-Free Reinforcement Learning. J Neurosci 36, 1211–
24
1222.
25
Eldar, E., Lièvre, G., Dayan, P., and Dolan, R.J. (2020). The roles of online and offline replay
26
in planning. ELife 9, e56911.
27
Groman, S.M., Massi, B., Mathias, S.R., Curry, D.W., Lee, D., and Taylor, J.R. (2019).
28
Neurochemical and Behavioral Dissections of Decision-Making in a Rodent Multistage Task.
29
J Neurosci 39, 295–306.
30
Howard, J.D., Reynolds, R., Smith, D.E., Voss, J.L., Schoenbaum, G., and Kahnt, T. (2020).
31
Targeted Stimulation of Human Orbitofrontal Networks Disrupts Outcome-Guided Behavior.
32
Current Biology 30, 490-498.e4.
33
Jones, J.L., Esber, G.R., McDannald, M.A., Gruber, A.J., Hernandez, A., Mirenzi, A., and
34
Schoenbaum, G. (2012). Orbitofrontal Cortex Supports Behavior and Learning Using Inferred
35
But Not Cached Values. Science 338, 953–956.
36
Kroemer, N.B., Lee, Y., Pooseh, S., Eppinger, B., Goschke, T., and Smolka, M.N. (2019). L-
37
DOPA reduces model-free control of behavior by attenuating the transfer of value to action.
38
NeuroImage 186, 113–125.
39
32
Langdon, A.J., Sharpe, M.J., Schoenbaum, G., and Niv, Y. (2018). Model-based predictions
1
for dopamine. Curr Opin Neurobiol 49, 1–7.
2
Liu, Y., Dolan, R.J., Kurth-Nelson, Z., and Behrens, T.E.J. (2019). Human Replay
3
Spontaneously Reorganizes Experience. Cell 178, 640-652.e14.
4
Liu, Y., Mattar, M.G., Behrens, T.E.J., Daw, N.D., and Dolan, R.J. (2020). Experience replay
5
supports non-local learning (Neuroscience).
6
Mattar, M.G., and Daw, N.D. (2018). Prioritized memory access explains planning and
7
hippocampal replay. Nat Neurosci 21, 1609–1617.
8
Montague, P.R., Dayan, P., and Sejnowski, T.J. (1996). A framework for mesencephalic
9
dopamine systems based on predictive Hebbian learning. J Neurosci 16, 1936–1947.
10
Moran, R., and Goshen-Gottstein, Y. (2015). Old processes, new perspectives: Familiarity is
11
correlated with (not independent of) recollection and is more (not equally) variable for targets
12
than for lures. Cognitive Psychology 79, 40–67.
13
Moran, R., Dayan, P., and Dolan, R.J. (in press). Human subjects exploit a cognitive map for
14
credit assignment. Proc Natl Acad Sci U S A.
15
Moran, R., Keramati, M., Dayan, P., and Dolan, R.J. (2019). Retrospective model-based
16
inference guides model-free credit assignment. Nat Commun 10, 750.
17
Moran, R., Keramati, M., and Dolan, R.J. (2021). Model based planners reflect on their
18
model-free propensities. PLoS Comput Biol 17, e1008552.
19
O’Doherty, J., Dayan, P., Schultz, J., Deichmann, R., Friston, K., and Dolan, R.J. (2004).
20
Dissociable roles of ventral and dorsal striatum in instrumental conditioning. Science 304,
21
452–454.
22
Otto, A.R., Gershman, S.J., Markman, A.B., and Daw, N.D. (2013). The curse of planning:
23
dissecting multiple reinforcement-learning systems by taxing the central executive. Psychol
24
Sci 24, 751–761.
25
Pessiglione, M., Seymour, B., Flandin, G., Dolan, R.J., and Frith, C.D. (2006). Dopamine-
26
dependent prediction errors underpin reward-seeking behaviour in humans. Nature 442,
27
1042–1045.
28
Sadacca, B.F., Jones, J.L., and Schoenbaum, G. (2016). Midbrain dopamine neurons
29
compute inferred and cached value prediction errors in a common framework. ELife 5.
30
Schuck, N.W., and Niv, Y. (2019). Sequential replay of nonspatial task states in the human
31
hippocampus. Science 364, eaaw5181.
32
Schuck, N.W., Cai, M.B., Wilson, R.C., and Niv, Y. (2016). Human Orbitofrontal Cortex
33
Represents a Cognitive Map of State Space. Neuron 91, 1402–1412.
34
Schultz, W., Dayan, P., and Montague, P.R. (1997). A neural substrate of prediction and
35
reward. Science 275, 1593–1599.
36
Sharp, M.E., Foerde, K., Daw, N.D., and Shohamy, D. (2016). Dopamine selectively
37
remediates “model-based” reward learning: a computational approach. Brain 139, 355–364.
38
33
Sharpe, M.J., Chang, C.Y., Liu, M.A., Batchelor, H.M., Mueller, L.E., Jones, J.L., Niv, Y., and
1
Schoenbaum, G. (2017). Dopamine transients are sufficient and necessary for acquisition of
2
model-based associations. Nat Neurosci 20, 735–742.
3
Starkweather, C.K., Babayan, B.M., Uchida, N., and Gershman, S.J. (2017). Dopamine
4
reward prediction errors reflect hidden-state inference across time. Nat Neurosci 20, 581–
5
589.
6
Steinberg, E.E., Keiflin, R., Boivin, J.R., Witten, I.B., Deisseroth, K., and Janak, P.H. (2013).
7
A causal link between prediction errors, dopamine neurons and learning. Nat Neurosci 16,
8
966–973.
9
Sutton, R.S. (1991). Dyna, an integrated architecture for learning, planning, and reacting.
10
SIGART Bull. 2, 160–163.
11
Westbrook, A., van den Bosch, R., Määttä, J.I., Hofmans, L., Papadopetraki, D., Cools, R.,
12
and Frank, M.J. (2020). Dopamine promotes cognitive effort by biasing the benefits versus
13
costs of cognitive work. Science 367, 1362–1366.
14
Wunderlich, K., Smittenaar, P., and Dolan, R.J. (2012). Dopamine enhances model-based
15
over model-free choice behavior. Neuron 75, 418–424.
16
17
18
34
Supplementary Information
1
2
Repeat and switch standard trials following uncertainty trials. We showed
3
previously on “repeat” trials (Moran et al., 2019), a positive effect of an informative destination
4
reward (on trial n) on choice-repetition implicates MFCA to the ghost-nominated object (while
5
the MB system knows that the value of the informative destination favours both vehicles on
6
trial n+1). We also ran a separate analysis that examined MFCA for the ghost-nominated
7
alone. In trial n+1 “repeat” trials, the ghost-nominated vehicle from trial-n is offered for choice
8
alongside a vehicle from the trial-n non-chosen pair that shares the inference-allowing
9
destination with the ghost-nominated object. Choice repetition was defined as choice of the
10
ghost-nominated vehicle from uncertainty trial n as indicated by the variable REPEAT.
11
Regressor PART is coded as previously. Regressors N (non-informative destination) and I
12
(informative destination) indicate whether a reward was received at the destinations or not in
13
trial n (coded as +.5/-.5). The model is REPEAT ~ N*I + (N*I | PART). This showed a main
14
effect for the informative (I) destination (b=0.60, t(4885)=7.56, p=4e-14), supporting MFCA to
15
the ghost-nominated object. Additionally, we found a main effect for the non-informative (N)
16
destination (b=1.23, t(4885)=10.83, p=9e-42) as predicted by both MF and by MB
17
contributions, and a significant interaction between the Informative and Non-informative
18
destinations (b=0.31, t(4885)=2.09, p=.04). See Figure S4, A & B.
19
We showed previously on switch trials that a positive main effect of the informative
20
outcome reward on choice-switching implicates MFCA for the ghost-rejected vehicle (because
21
the MB system knows the informative destination is unrelated to both vehicles on trial n+1). A
22
second separate analysis examined MFCA for the ghost-rejected vehicle In uncertainty
23
trialn+1 “switch” trials, the ghost-rejected vehicle from trial-n is offered for choice alongside a
24
vehicle from the trial-n non-chosen pair that shares a destination with the ghost-rejected object.
25
Choice switching was defined as choice of the ghost-rejected vehicle from uncertainty trial n
26
as indicated by the variable SWITCH. Regressors PART, N and I are coded as previously. The
27
model is SWITCH ~ N*I + (N*I | PART). This showed a main effect for the reward at informative
28
35
destination (b=0.38, t(4866)=5.60, p=2e-8), supporting MFCA to the ghost-rejected vehicle.
1
While, this challenges any notion of perfect MB guidance of MFCA, it is consistent with the
2
possibility that some participants, at least some of the time, do not rely on MB-inference
3
because when MB inference does not occur, or when it fails to guide MF credit-assignment,
4
the MF system has no basis to assign credit unequally to both vehicles in the selected pair.
5
Additionally, we found a main effect for the non-informative destination reward (b=0.98,
6
t(4866)=13.18, p=5e-39), as predicted by an MF credit-assignment to the ghost-rejected
7
vehicle account but also by MB contributions. We found no significant interaction between
8
rewards at the informative and non-informative destinations (Table S1). In Figure S2, we plot
9
empirical choice proportions from both repeat and switch conditions (reflecting the effects
10
reported above) in the manner as in the original paper by Moran et al. (2019) but separately
11
for drug and placebo conditions. See Figure S4 C & D.
12
13
Absence of drug effects on perseveration and forgetting parameters of the
14
computational model. No difference between drug conditions was observed for
15
perseveration parameter on standard trials (t(61)= 0.48, p=.63), perseveration parameter on
16
uncertainty trials (t(61)= 0.51, p=.61), MF forgetting parameter (t(61)= 1.37, p=.17), MB
17
forgetting parameter (t(61)= -0.33, p=.74), perseveration forgetting parameter (t(61)=0.30,
18
p=.77).
19
20
Correlation between WM and MBCA. Working memory moderated the boosting drug effect
21
on guidance of MFCA based on retrospective MB inference but only based on non-informative
22
reward (see main text). No moderating effect of working memory on a drug-dependent
23
difference in MBCA was observed (r=-.07, p=.59). Working memory correlated positively with
24
MBCA separately at placebo and at drug but this was non-significant (placebo: r=.21, p=.08;
25
drug: r=.15, p=.23).
26
Table S1. Mixed-effects models on model-agnostic choice data from standard trials
Name
Estimate
SE
tStat
DF
pValue
LowerCI
UpperCI
MF choice (standard trials)
REPEAT ~ 1+ C*U*DRUG*ORDER + (C+U+DRUG+ORDER | PART)
(Intercept)
0.34
0.06
5.54
7251
.000
0.22
0.46
C (common)
0.67
0.07
9.14
7251
.000
0.53
0.81
U (unique)
1.54
0.09
17.40
7251
.000
1.36
1.71
DRUG
0.03
0.07
0.46
7251
.643
-0.11
0.18
ORDER
0.07
0.07
0.91
7251
.365
-0.08
0.21
C*U
0.19
0.11
1.72
7251
.085
-0.03
0.40
C*DRUG
0.07
0.11
0.67
7251
.500
-0.14
0.29
U*DRUG
0.06
0.11
0.56
7251
.577
-0.15
0.27
C*ORDER
0.12
0.11
1.09
7251
.276
-0.09
0.33
U*ORDER
-0.11
0.11
-0.99
7251
.321
-0.32
0.11
DRUG*ORDER
-0.25
0.24
-1.02
7251
.309
-0.73
0.23
C*U*DRUG
0.14
0.22
0.64
7251
.524
-0.29
0.57
C*U*ORDER
0.13
0.22
0.59
7251
.554
-0.30
0.56
C*DRUG*ORDER
-0.02
0.29
-0.06
7251
.952
-0.59
0.56
U*DRUG*ORDER
-0.18
0.35
-0.51
7251
.609
-0.87
0.51
C*U*DRUG*ORDER
-0.22
0.44
-0.50
7251
.618
-1.07
0.64
MB choice (standard trials)
GENERALIZE ~ C*P*DRUG*ORDER + (C+P+DRUG+ORDER | PART)
(Intercept)
0.30
0.04
6.96
7177
.000
0.22
0.38
C (common)
0.40
0.06
6.22
7177
.000
0.27
0.52
P (common reward probability)
1.33
0.21
6.39
7177
.000
0.92
1.74
DRUG
-0.13
0.08
-1.65
7177
.099
-0.29
0.03
ORDER
-0.13
0.08
-1.57
7177
.116
-0.29
0.03
C*P
-0.23
0.23
-1.01
7177
.311
-0.67
0.21
C*DRUG
0.05
0.12
0.39
7177
.695
-0.19
0.28
P*DRUG
-0.34
0.23
-1.48
7177
.140
-0.79
0.11
C*ORDER
-0.06
0.12
-0.52
7177
.606
-0.30
0.17
P*ORDER
0.16
0.23
0.70
7177
.482
-0.29
0.61
DRUG*ORDER
-0.24
0.17
-1.41
7177
.158
-0.58
0.09
C*P*DRUG
-0.08
0.45
-0.18
7177
.856
-0.97
0.80
C*P*ORDER
0.57
0.45
1.26
7177
.207
-0.31
1.45
C*DRUG*ORDER
-0.38
0.25
-1.48
7177
.140
-0.87
0.12
P*DRUG*ORDER
0.46
0.83
0.55
7177
.583
-1.18
2.09
C*P*DRUG*ORDER
1.40
0.91
1.54
7177
.123
-0.38
3.17
37
Table S2. Mixed-effects models on model-agnostic choice data from uncertainty trials
Name
Estimate
SE
tStat
DF
pValue
LowerCI
UpperCI
Preferential MFCA for the informative destination (Ghost-nominated, “repeat trials” > ghost-rejected, “switch trials”)
MFCA ~ NOM*DRUG*ORDER + (NOM*DRUG+ORDER | PART)
(Intercept)
0.10
0.01
8.54
239
.000
0.07
0.12
NOM (Nomination)
0.03
0.02
1.60
239
.110
-0.01
0.08
DRUG
0.01
0.02
0.40
239
.690
-0.04
0.05
ORDER
0.00
0.02
0.15
239
.878
-0.04
0.05
NOM*DRUG
0.11
0.04
2.56
239
.011
0.03
0.20
NOM*ORDER
0.02
0.04
0.45
239
.650
-0.07
0.11
DRUG*ORDER
-0.03
0.04
-0.74
239
.463
-0.12
0.06
NOM*DRUG*ORDER
-0.01
0.09
-0.06
239
.951
-0.18
0.16
MFCA for non-informative destination (Ghost-nominated > ghost-rejected, “clash trials”)
REPEAT ~ N*I*DRUG*ORDER + (N*I*DRUG+ORDER | PART)
(Intercept)
0.05
0.04
1.27
4861
.203
-0.03
0.12
N (non-informative)
0.13
0.07
1.96
4861
.051
0.00
0.26
I (informative)
1.01
0.10
9.95
4861
.000
0.81
1.21
DRUG
0.15
0.07
2.31
4861
.021
0.02
0.29
ORDER
0.03
0.07
0.41
4861
.684
-0.10
0.16
N*U
0.08
0.14
0.57
4861
.568
-0.19
0.35
N*DRUG
0.05
0.13
0.39
4861
.696
-0.21
0.31
I*DRUG
0.03
0.14
0.24
4861
.810
-0.25
0.32
N*ORDER
-0.05
0.13
-0.34
4861
.733
-0.30
0.21
I*ORDER
0.06
0.14
0.43
4861
.664
-0.22
0.35
DRUG*ORDER
-0.20
0.15
-1.37
4861
.171
-0.49
0.09
N*I*DRUG
0.07
0.29
0.26
4861
.798
-0.49
0.64
N*I*ORDER
0.25
0.29
0.86
4861
.388
-0.32
0.81
N*DRUG*ORDER
-0.47
0.26
-1.80
4861
.072
-0.99
0.04
I*DRUG*ORDER
-0.12
0.41
-0.31
4861
.759
-0.92
0.67
N*I*DRUG*ORDER
0.86
0.55
1.56
4861
.118
-0.22
1.94
38
Table S3. Mixed-effects models on parameters of the computational model.
Table S4. Distribution of parameters from the full computational model.
Cond.
%
MFCA
standard
MFCA
info-
nom
MFCA
info-
rej
MFCA
non-
info-
nom
MFCA
non-
info-rej
MBCA
persev
eration
-
standar
d
persev
eration
-
nomina
ted
forget_
MF
forget_
MB
forget_
Pers
Placebo
25
0.053
-0.056
-0.026
-0.070
-0.074
0.059
-0.197
-0.093
0.002
0.038
0.010
50
0.147
0.168
0.149
0.048
0.030
0.273
0.042
0.071
0.058
0.148
0.123
75
0.364
0.479
0.391
0.333
0.204
0.454
0.383
0.353
0.519
0.521
0.428
L-DOPA
25
0.060
-0.025
-0.073
-0.011
-0.098
0.026
-0.086
-0.047
0.019
0.022
0.008
50
0.272
0.165
0.130
0.178
0.070
0.278
0.098
0.084
0.190
0.127
0.089
75
0.574
0.517
0.383
0.390
0.291
0.367
0.346
0.374
0.598
0.508
0.492
Name
Estimate
SE
tStat
DF
pValue LowerCI
UpperCI
MFCA for ghost-nominated vs. ghost-rejected and informative vs non-informative
MFCA ~ NOM*INFO*DRUG* + (NOM*INFO*DRUG | PART)
(Intercept)
0.18
0.02
7.60
480
.000
0.14
0.23
NOM (nomination)
0.10
0.03
3.72
480
.000
0.05
0.15
INFO (informativeness)
0.08
0.04
2.19
480
.029
0.01
0.15
DRUG
0.05
0.05
1.00
480
.316
-0.05
0.15
ORDER
0.04
0.05
0.78
480
.434
-0.06
0.14
NOM*INFO
-0.03
0.05
-0.57
480
.567
-0.12
0.06
NOM*DRUG
0.10
0.04
2.43
480
.015
0.02
0.18
INFO*DRUG
-0.08
0.07
-1.16
480
.247
-0.22
0.06
NOM*ORDER
0.02
0.04
0.37
480
.715
-0.07
0.10
INFO*ORDER
0.10
0.07
1.42
480
.157
-0.04
0.23
DRUG:ORDER
-0.09
0.10
-0.98
480
.328
-0.28
0.10
NOM*INFO*DRUG
0.02
0.07
0.33
480
.738
-0.12
0.17
NOM*INFO*ORDER
-0.01
0.07
-0.08
480
.934
-0.15
0.14
NOM*DRUG*ORDER
-0.06
0.11
-0.60
480
.551
-0.27
0.15
INFO*DRUG*ORDER
0.16
0.14
1.10
480
.272
-0.12
0.44
NOM*INFO*DRUG*ORDER
0.10
0.19
0.55
480
.585
-0.26
0.47
Preferential MFCA for informative vs. non-informative
PMFCA ~ INFO*DRUG*ORDER + (INFO*DRUG+ORDER | PART)
(Intercept)
0.10
0.03
3.72
240
.000
0.05
0.15
INFO (informativeness)
-0.03
0.05
-0.57
240
.568
-0.12
0.07
DRUG
0.10
0.04
2.41
240
.017
0.02
0.18
ORDER
0.02
0.04
0.36
240
.717
-0.07
0.10
INFO*DRUG
0.02
0.07
0.33
240
.739
-0.12
0.17
INFO*ORDER
-0.01
0.07
-0.08
240
.934
-0.15
0.14
DRUG*ORDER
-0.06
0.11
-0.60
240
.551
-0.27
0.15
INFO*DRUG*ORDER
0.10
0.19
0.55
240
.585
-0.27
0.47
39
Figure S1. Simulations for standard trials based on the full model and sub-models. NR=no reward,
R=reward. Rew=reward at the common destination, RewProBC=Reward Probability at the common
destination.
40
Figure S2. Simulations for uncertainty trials based on the full model and sub-models. GS=Ghost-
selected, GR=Ghost-rejected.
41
Figure S3. Empirical probabilities of model-agnostic MF (A & B) and MB (C & D) choice contribution
under placebo and levodopa (L-DOPA). U-Non=no reward at unique destination, U-Rew= reward at
unique destination, C-Non=no reward at common destination, C-Rew= reward at common destination.
42
Figure S4. Retrospective MB inference using the informative destination based on repeat and switch
signatures after uncertainty trials. I-Non=no reward at informative destination, I-Rew= reward at
informative destination, N-Non=no reward at non-informative destination, N-Rew= reward at non-
informative destination.
43
Figure S5. Retrospective MB inference using the non-informative destination based on choice repetition
in “clash” trials n+1 following an uncertainty trial-n. I-Non=no reward at informative destination, I-Rew=
reward at informative destination, N-Non=no reward at non-informative destination, N-Rew= reward at
non-informative destination.
44
Figure S6. Model-comparison results. A) Results of the bootstrap-GLRT model-comparison for the pure
MB sub-model. The blue bars show the histogram of the group twice log-likelihood improvement (model
vs. sub-model) for synthetic data simulated using the sub-model (10000 simulations). The blue line
displays the smoothed null distribution (using Matlab’s “ksdensity”). The red line shows the empirical
group twice log-likelihood improvement. p-value reflect the proportion of 10000 simulations that yielded
an improvement in likelihood that was at least as large as the empirical improvement. B-E) Same as
(A), but for the pure MF choice, the no informativeness effects on MFCA, the no MB-guidance for MFCA,
the no MB-guidance for the informative destination and the no-MB guidance for the non-informative
destination sub models.
45
Figure S7. Parameter recoverability. For each of the 2*62 full model parameter-combinations
1000 synthetic (simulated) datasets were created by simulating the full model on experimental
sessions as in the true experiment. Then the full model was fit to each of these generated
datasets. For each MFCA parameter (info/non-info x nom/rej), we plot the recovered against
the generating parameters (and impose black diagonals where "recovered = generating").
46
Figure S8. When using a model-agnostic measure of MB choice (probability to generalize after
reward minus no-reward) and of preferential MFCA at the informative destination (repeat or
ghost-nominate minus switch or ghost-rejected), dopamine dependent differences (levodopa
minus placebo) in those measures were correlated negatively (r=-.29, p=.021) mirroring the
finding as reported on parameters from the computational model in the main text.
-1
-0.5
0
0.5
1
MB: L-Dopa > Placebo
-1.5
-1
-0.5
0
0.5
1
1.5
2
Info GN>GR: L-Dopa > Placebo
r=-0.29, p=0.02
| 2021 | Dopamine enhances model-free credit assignment through boosting of retrospective model-based inference | 10.1101/2021.01.15.426639 | [
"Deserno Lorenz",
"Moran Rani",
"Michely Jochen",
"Lee Ying",
"Dayan Peter",
"Dolan Raymond J."
] | creative-commons |
1
TITLE: Developing an empirical model for spillover and emergence: Orsay virus host range in
Caenorhabditis
AUTHOR LIST:
Clara L. Shaw1 Department of Biology, The Pennsylvania State University, University Park, PA 16802
David A. Kennedy2 Department of Biology, The Pennsylvania State University, University Park, PA 16802
Corresponding Author: David A. Kennedy
1cls6630@psu.edu
2dak30@psu.edu
2
ABSTRACT
A lack of tractable experimental systems in which to test hypotheses about the ecological and
evolutionary drivers of disease spillover and emergence has limited our understanding of these
processes. Here we introduce a promising system: Caenorhabditis hosts and Orsay virus, a positive-
sense single-stranded RNA virus that naturally infects C. elegans. We assayed the susceptibility of
species across the Caenorhabditis tree and found 21 of 84 wild strains belonging to 14 of 44 species to
be susceptible to Orsay virus. Confirming patterns documented in other systems, we detected effects of
host phylogeny on susceptibility. We then tested whether susceptible strains were capable of
transmitting Orsay virus by transplanting exposed hosts and determining whether they transmitted
infection to conspecifics during serial passage. We found no evidence of transmission in 10 strains (virus
undetectable after passaging), evidence of low-level transmission in 5 strains (virus lost between
passage 1 and 5), and evidence of sustained transmission in 6 strains (including all 3 experimental C.
elegans strains). Transmission was associated with host phylogeny and with viral amplification in
exposed populations. Variation in Orsay virus susceptibility and transmission among Caenorhabditis
species suggests that the system could be powerful for studying spillover and emergence.
KEYWORDS: host range, spillover, emergence, Caenorhabditis, Orsay virus
INTRODUCTION
Disease spillover and emergence can have catastrophic consequences for the health of humans and
other species. For example, SARS-CoV-2 spilled over into human populations [1] and became pandemic,
killing more than 5 million people when this study was published [2]. Moreover, the frequency of
3
spillover events and the rate of new disease emergence has been increasing in the recent past [3],
endowing urgency to the task of understanding drivers of spillover and the progression of emergence.
Studies in wild systems with ongoing spillover have provided substantial insights into the spillover and
emergence process [4–6], but experimental manipulation to test hypotheses in these systems can be
impractical due to ethical and logistical concerns. Moreover, disease emergence is so rare that it
typically can only be studied retrospectively. Therefore, it remains a challenge to understand what
factors facilitate emergence and how evolution proceeds in emerging pathogens.
Spillover requires that pathogens have the opportunity and the ability to exploit a new host;
emergence requires that this opportunity and ability persist through time [5,7]. Opportunity could arise
if hosts share habitats or resources. Ability may arise through mutations or pre-exist due to pathogen
plasticity or host similarity. Studies of natural spillover and emergence events have identified
characteristics of pathogens, hosts, and their interactions that generally support the above. For
example, pathogens that successfully spill over are likely to be RNA viruses with large host ranges [8,9].
Likewise, hosts with close phylogenetic relationships are more likely to share pathogens than more
distantly related hosts [9–14]. In addition, geographic overlap between hosts is associated with sharing
pathogens [12], meaning that changes in host population distributions that bring new species into
contact could potentially promote spillover and emergence events [9,15–17].
Ecological factors (e.g. host densities, distributions, diversity, condition, and behavior) can
promote or hinder spillover by modulating host exposure risk or host susceptibility [5,7]. Likewise, it is
believed that ecological factors can promote or hinder emergence through the modulation of onward
transmission in spillover hosts, which determines whether pathogens meet dead ends in novel hosts,
transmit in stuttering chains, or adapt and persist [18–20]. Conclusively demonstrating the influence of
ecological factors, however, requires experimental manipulation, and it has so far been difficult to
perform such studies.
4
Experimental model systems have been essential for testing hypotheses about infectious
disease biology [21–23]. Indeed, major discoveries in immunity, pathogenesis, and pathogen ecology
and evolution come from model systems such as Mus musculus [24], Drosophila melanogaster [25],
Daphnia species [21], Arabadopsis thaliana [26], and Caenorhabditis elegans [27]. These systems have
important traits that make them amenable to experimentation: they are inexpensive, have fast
generation times, and have simplified genetics since they are usually hermaphroditic, asexual, or inbred.
In addition, experimental tools and knowledge have accumulated in these systems, lowering the barriers
to novel findings. However, few model systems exist to study the ecology and evolution of disease
spillover and emergence, and the systems that do exist lack key features known to drive disease
dynamics (e.g. host behavior or transmission ecology). A perfect model system would have large host
population sizes, naturally transmitting, fast-evolving pathogens (e.g. viruses), and multiple potential
host species with variable susceptibility and transmission.
Caenorhabditis nematode species are appealing model host candidates. Indeed, C. elegans and
various bacterial and microsporidian parasites are staples of evolutionary disease ecology [22,28].
Specifically, the trivial manipulation and sampling of laboratory host populations means that population-
level processes like disease transmission and evolution can be observed, and the tractable replication of
large populations makes possible the observation of rare events like spillover and emergence. However,
until recently, there were no known viruses of any nematodes including C. elegans. That changed with
the recent discovery of Orsay virus [29].
Orsay virus, a natural gut pathogen of C. elegans, is a bipartite, positive-sense, single-stranded
RNA (+ssRNA) virus that transmits readily in laboratory C. elegans populations through the fecal-oral
route [29]. This virus is an appealing model pathogen candidate since +ssRNA viruses have high
mutation rates [30] and typically evolve quickly [31]. Moreover, since Orsay virus transmits between
hosts in the lab, this system allows transmission itself to evolve, a critical component of emergence [19]
5
that cannot be readily studied in other animal laboratory systems of disease emergence. To develop
Caenorhabditis hosts and Orsay virus as a system for studying spillover and emergence, it is necessary to
know the extent to which the virus can infect and transmit in non-elegans Caenorhabditis species. So
far, such exploration been limited to one other species, C. briggsae, which was determined to be
refractory to infection [29]. Notably, an ancestral virus likely crossed at least one host species boundary
in the past since C. briggsae has been found to be susceptible to three related viruses [29,32,33].
To explore the suitability of the Caenorhabditis-Orsay virus system for studies of disease
spillover and emergence, we first test a suite of Caenorhabditis species for susceptibility to Orsay virus,
and then we test the extent to which susceptible host species can transmit the virus. For both traits
(susceptibility and transmission ability), we test for effects of host phylogeny. Though host ranges of
pathogens have been studied by infection assays (e.g. [34–37]) or by sampling infected hosts from
natural systems (e.g. [11,38]), these studies do not typically distinguish between dead-end infections,
stuttering chains of transmission, and sustained transmission. Therefore, to our knowledge, our study is
the first to empirically link phylogeny with disease transmission dynamics in novel species following
spillover.
METHODS
Susceptibility Assays
We assayed susceptibility of Caenorhabditis species to Orsay virus by measuring virus RNA in
previously virus-exposed host populations using quantitative PCR (qPCR). We obtained 84 wild isolate
strains belonging to 44 Caenorhabditis species (1-3 strains per species) from the Caenorhabditis Genetics
Center (CGC) and from Marie-Anne Félix. We tested each strain for Orsay virus susceptibility using 8
experimental blocks (Table 1, Table S1). Species identities were confirmed by sequencing the small
6
ribosomal subunit internal transcribed spacer ITS2 and/or by mating tests. For each Caenorhabditis
strain, we initiated three replicate populations with five adult animals. For sexual species, we used five
mated females, and for hermaphroditic species, we used five hermaphrodites. All populations were
maintained on nematode growth medium (NGM) in 60 mm diameter plates with a lawn of bacterial
food (lawns were seeded with 200 µL E. coli strain OP50 in Luria-Bertani (LB) broth and allowed to grow
at room temperature for approximately 24 hours [39]). We exposed populations to virus by pipetting 3
µL of Orsay virus filtrate, prepared as described in [29], onto the center of the bacterial lawn. We
determined the concentration of the filtrate to be 428.1 (95% CI: 173.4-972.3) x the median tissue
culture infectious dose (TCID50) per µL (Supplement A) [40]. We maintained populations at 20°C until
freshly starved (i.e. plates no longer had visible bacterial lawns). Depending on the strain, this took
anywhere from 3 to 28 days (Table S1). While this meant that strains may have experienced variable
numbers of generations, this method ensured that all the exposure virus was consumed. We collected
nematodes from freshly starved plates by washing plates with 1,800 µL water and transferring
suspended animals to 1.7 mL microcentrifuge tubes. We centrifuged tubes at 1000 x g for 1 minute to
pellet nematodes. We removed the supernatant down to 100 µL (including the pellet of nematodes) and
‘washed’ external virus from nematodes by adding 900 µL of water and removing it 5 times, centrifuging
at 1000 x g for 1 minute between each wash. After the five washes, we lysed the nematodes by
transferring the nematode pellet along with 500 µL water to 2 mL round-bottom snap cap tubes, adding
approximately 100 µL of 0.5 mm silica beads, and shaking in a TissueLyser II (Qiagen) for 2 minutes at a
frequency of 30 shakes per second. We then removed debris with two centrifugation steps of 17,000 x g
for 5 minutes, each time keeping the supernatant and discarding the pellet. Samples were stored at -80
°C.
We used qPCR to measure viral RNA in these samples. Primers and probe were: Forward: GTG
GCT GTG CAT GAG TGA ATT T, Reverse: CGA TTT GCA GTG GCT TGC T, Probe: 6-FAM-ACT TGC TCA GTG
7
GTC C-MGB. We performed 10 µL reactions composed of 1.12X qScript XLT One-Step RT-qPCR ToughMix
(Quantabio), 200 nM each of forward and reverse primers and probe, and 2 µL of sample. Reaction
conditions were: 50 °C (10 min), 95 °C (1 min), followed by 40 cycles of 95 °C (3 sec), 60 °C (30 sec).
Assays were run on a 7500 Fast Real-Time qPCR System (Thermo Fisher Scientific, Applied Biosystems).
Cycle threshold (Ct) values were determined using the auto-baseline and auto-threshold functions of the
7500 Fast Real-Time software (Thermo Fisher Scientific, Applied Biosystems).
Each experimental block also contained five sets of controls and benchmarks (Table 2): a
negative control where virus was never added (control 1), two positive controls where strains with
known susceptibilities were exposed (control 2, strain N2: mean(Ct)=15.7, sd(Ct)=2.0; control 3, strain
JU1580: mean(Ct)=12.7, sd(Ct)=2.2), a benchmark to determine a Ct threshold for infection (benchmark
4: mean(Ct)=38.4, sd(Ct)=2.6), and a benchmark that gives a conservative Ct threshold for viral
replication (benchmark 5: mean(Ct)=22.0, sd(Ct)=0.6). Species were considered susceptible if at least
one replicate population amplified virus to levels higher than our infection threshold (one standard
deviation more virus than the maximum value of benchmark 4 across all blocks which translates to
Ct<29.5).
Transmission Assays
We conducted transmission assays for all strains where at least one replicate population was
determined to be infected in our susceptibility assay. First, three replicate populations were initiated as
above and exposed to 3 µL of virus filtrate. At the same time, we initiated three replicate positive
control populations of C. elegans laboratory strain N2 exposed to 3 µL virus filtrate and three replicate
negative control populations of N2s exposed to 3 µL of water. When populations were recently starved,
20 adult nematodes (mated females for sexual species or hermaphrodites for hermaphroditic species)
8
were chosen at random and passaged to virus-free plates with fresh food (E. coli strain OP50 lawns
prepared as above). Remaining animals were washed from the starved plates, virus was extracted, and
viral RNA quantified via qPCR as above (Table S2). We passaged each replicate line 5 times, or until there
was no detectable viral RNA by qPCR. Controls were passaged 5 times regardless of virus detection.
We assigned each passage line a transmission score of 0, 1, 2, or 3 based on detection of viral
RNA through the passages. A value of 0 was assigned when viral RNA was not detected in the exposure
population; a value of 1 was assigned when viral RNA was detected in the exposure population but not
in the first passage population; a value of 2 was assigned when viral RNA was detected in the first
passage population but became undetectable on or before the fifth passage population; and a value of 3
was assigned when viral RNA was still detectable in the fifth passage population.
Statistical Analysis
To test for phylogenetic effects, we fit Bayesian phylogenetic mixed effects models to the
susceptibility and transmission data using the ‘MCMCglmm’ package [35,41,42] in R [43]. For these
models, we used the most recent published phylogeny of Caenorhabditis [44]. We rooted the phylogeny
with Diploscapter pachys as the outgroup and constrained it to be ultrameric (i.e. tips are all equidistant
from the root) using the ‘chronopl’ function in the ‘ape’ package [45] with a smoothing parameter of 1.
Since our susceptibility data are binomial, we fit them using logistic regression with a logit link. In
practice this was achieved by setting family to ‘multinomial2’. Our transmission data are continuous,
and we fit them using linear regression by setting family to ‘gaussian’. Data from controls and
benchmarks were excluded from the analysis. For both the susceptibility and transmission data, all
models included a random effect of species and all transmission models also included a random effect of
strain. These random effects were included to prevent pseudo-replication. Other factors were included
9
or excluded as described below. For the susceptibility data, our most complicated model included
effects of phylogenetic distance to the native host C. elegans (calculated by the ‘cophenetic.phylo’
function in ‘ape’ [45]) and phylogenetic distance between pairwise sets of species (calculated by the
‘inverseA’ function in ‘MCMCglmm’ [41,46]). Note that ‘inverseA’ calculates the inverse relatedness
matrix (i.e. the inverse of the matrix that contains the time from the root to the common ancestor of
each species pair), but we refer to this metric as “phylogenetic distance between pairwise sets of
species” for simplicity. For the transmission data, our most complicated model included these effects
and an additional effect of viral amplification in the primary exposure population measured as Ct.
Phylogenetic distance from C. elegans and viral amplification in the primary exposure population were
treated as fixed effects, and phylogenetic distance between pairwise sets of species was treated as a
random effect. We generated a suite of nested models that included all possible combinations of
including or excluding these effects (Table 3, Table 4).
We used the MCMCglmm default priors for fixed effects (normal distribution with mean = 0 and
variance = 108) and parameter expanded priors for random effects that result in scaled multivariate F
distributions with V=1, nu=1, alpha.mu=0, alpha.V=1000 [47]. Residuals were assigned inverse Wishart
priors with V=1 n=0.002 [41]. We ran models for 100,000,000 iterations with a burn in of 300,000 and
thinning interval of 10,000. We visualized traces to affirm convergence of MCMC chains.
We used the deviance information criterion (DIC) to describe the relative support of models and
to understand the importance of parameters [48]. We calculated DIC weights for each model, each
parameter, and the phylogenetic parameters combined [49]. The DIC weight of a model, calculated as
������/�
∑ ������/�
�
where � is the set of all models, gives the relative support for each model. Similarly, the DIC
weight of a parameter, calculated as
∑ ������/�
�
∑ ������/�
�
where � refers to the set of models that includes a given
parameter and � is the set of all models, is the posterior probability that a given factor is included in the
10
‘true’ model assuming the ‘true’ model has been designated. Thus, parameters with DIC weights greater
than 0.5 are more likely than not to be included in the ‘true’ model.
11
Table 1. Strains assayed for susceptibility to Orsay virus with the number of replicates processed in each
block. When strains were assayed in multiple blocks, replicate numbers are given in the respective order
of the blocks. Strains were acquired from the Caenorhabditis Genetics Center (University of Minnesota)
and from Marie-Anne Felix (IBENS).
Strain
Species
Block
Number of
Replicates
Strain
Species
Block
Number of
Replicates
JU1199
C. afra
2
3
JU2613
C. portoensis
7
3
JU1198
C. afra
4
3
JU2745
C. quiockensis
2
3
JU1593
C. afra
7
3
MY28
C. remanei
2
3
NIC1040
C. astrocarya
3
1
PB206
C. remanei
6
3
QG704
C. becei
2
3
JU1082
C. remanei
6
3
SB280
C. brenneri
1
3
JU1201
C. sinica
1
3
SB129
C. brenneri
6
3
JU4053
C. sinica
4
3
LKC28
C. brenneri
6
3
JU1202
C. sinica
6
3
JU1038
C. briggsae
1,2,31
3,3,3
JU2203
C. sp. 8
5
2
EG4181
C. briggsae
6
3
QG555
C. sp. 24
3
3
ED3083
C. briggsae
6
3
JU2867
C. sp. 24
5,7
1,3
JU1426
C. castelli
3,7
3,3
JU2837
C. sp. 24
6
3
JU1333
C. doughertyi
1
3
ZF1092
C. sp. 25
3
3
JU1328
C. doughertyi
4
3
QX2263
C. sp. 27
1,3
2,3
JU1331
C. doughertyi
5
3
DF5152
C. sp. 30
3
3
DF5112
C. drosophilae
3
3
NIC1070
C. sp. 43
2
3
GXW1
C. elegans
6
3
JU4050
C. sp. 62
5
3
JU1401
C. elegans
6
3
JU4045
C. sp. 62
7
3
ED3042
C. elegans
6
3
JU4056
C. sp. 63
6
3
NIC113
C. guadaloupensis
1
3
JU4061
C. sp. 64
6
3
EG5716
C. imperialis
3
3
JU4087
C. sp. 65
4
3
JU1905
C. imperialis
7
3
JU4093
C. sp. 65
5
3
NKZ352
C. inopinata
3
3
JU4092
C. sp. 65
5
3
QG122
C. kamaaina
2
3
JU4094
C. sp. 66
4
3
VX80
C. latens
1
3
JU4096
C. sp. 66
4
3
JU3325
C. latens
4
3
JU4088
C. sp. 66
4
3
JU724
C. latens
5,7
1,3
SB454
C. sulstoni
2
3
JU1857
C. macrosperma
2
3
JU2774
C. tribulationis
1
3
JU1865
C. macrosperma
5
3
JU2776
C. tribulationis
5
3
JU1853
C. macrosperma
7
3
JU2775
C. tribulationis
5
3
JU28843
C. monodelphis
8
3
JU1373
C. tropicalis
1
3
JU16673
C. monodelphis
8
3
JU1428
C. tropicalis
2
3
JU1325
C. nigoni
1,2,3
2, 1, 3
JU2469
C. uteleia
2
3
JU2617
C. nigoni
4
3
JU2458
C. uteleia
4
3
EG5268
C. nigoni
6
3
JU1968
C. virilis
3
3
JU1825
C. nouraguensis
1
3
JU2758
C. virilis
5
3
JU1833
C. nouraguensis
5
3
NIC564
C. waitukubuli
1
3
JU1854
C. nouraguensis
6
3
JU1873
C. wallacei
1
3
QG702
C. panamensis
2
3
EG6142
C. yunquensis
3
3
JU2770
C. parvicauda
7
3
JU2156
C. zanzibari
1
3
EG4788
C. portoensis
1
3
JU3236
C. zanzibari
6
3
JU3126
C. portoensis
5
3
JU2161
C. zanzibari
7
3
1JU1038 was included in the first three blocks as a type of negative control since a previous study found
that C. briggsae was not susceptible. We discontinued this practice given the number of strains we
needed to test.
2Strain NKZ35 was maintained at 23°C according to Caenorhabditis Genetics Center recommendation.
3Populations were initiated with 12 juvenile animals due to challenges rearing animals with standard
methods.
12
Table 2. Description of controls and benchmarks included in triplicate in each of the 8 blocks of the
susceptibility assays.
Control/benchmark
Description
Type
1
Laboratory C. elegans strain N2
exposed to 3 µL water
Negative control
2
Laboratory C. elegans strain N2
exposed to 3 µL Orsay virus filtrate
Positive control
3
Highly susceptible C. elegans strain
JU1580 exposed to 3 µL of Orsay virus
filtrate
Positive control
4
3 µL Orsay virus filtrate pipetted on
the center of bacterial lawn with no
nematodes
Thresholda
5
3 µL Orsay virus filtrate added directly
to 497 µL water, yielding the final
extraction volume for experimental
populations.
Thresholdb
aThe purpose of this benchmark was to quantify exposure virus remaining in samples after 5 rounds of
washing.
bThe purpose of this benchmark was to quantify the maximum amount of virus that could be present
without replication (i.e. total amount of virus added to each plate).
Table 3. Models compared for analysis of susceptibility patterns. All models included an intercept. The
random effect of species is retained in all models to avoid pseudo-replication.
Model
ΔDIC
DIC weight
Suscep. ~ fixed = phylo. dist., random = pairwise phylo. dist. + species
0
0.486
Suscep. ~ fixed = phylo. dist., random = species
1.121
0.277
Suscep. ~ fixed = random = pairwise phylo. dist. + species
2.189
0.163
Suscep. ~ fixed = random = species
3.761
0.074
‘phylo. dist’ indicates the effect of phylogenetic distance from C. elegans whereas ‘pairwise phylo. dist.’
indicates the effect of phylogenetic distance between species pairs.
13
Table 4. Models compared for analysis of transmission scores. All models included an intercept. Random
effects of species and strain are retained in all models to avoid pseudo-replication.
Model
ΔDIC
DIC weight
Trans. ~ fixed = Ct + phylo. dist., random = pairwise phylo. dist. + species + strain
0
0.269
Trans. ~ fixed = Ct , random = pairwise phylo. dist. + species + strain
0.533
0.206
Trans. ~ fixed = Ct + phylo. dist., random = species + strain
0.585
0.201
Trans. ~ fixed = Ct , random = species + strain
0.790
0.181
Trans. ~ fixed = phylo. dist., random = pairwise phylo. dist. + species + strain
3.942
0.038
Trans. ~ fixed = random = species + strain
4.086
0.035
Trans. ~ fixed = random = pairwise phylo. dist. + species + strain
4.091
0.035
Trans. ~ fixed = phylo. dist., random = species + strain
4.112
0.034
‘Ct’ indicates viral amplification on primary exposure plates. ‘phylo.dist’ indicates the effect of
phylogenetic distance from C. elegans whereas ‘pairwise phylo. dist.’ indicates the effect of phylogenetic
distance between species pairs.
RESULTS
Susceptibility Assays
In our assays of host susceptibility to Orsay virus, we identified 21 susceptible Caenorhabditis strains of
the 84 experimental strains tested. These included three (non-control) strains of C. elegans (note that
one of these strains JU1401 had been previously documented to be susceptible [50]) and 18 strains
belonging to 13 other species. The strains were distributed broadly across the Caenorhabditis
phylogenetic tree and in species that do not currently have a well determined phylogenetic placement
(Figure 1). In total, we found that Orsay virus is capable of infecting hosts from at least 14 of 44
Caenorhabditis species.
14
Our statistical analysis uncovered the importance of host phylogeny in explaining differences in
susceptibility. Our best model included both phylogenetic effects tested: phylogenetic distance from C.
elegans and phylogenetic distance between pairwise sets of species (Table 3). The model lacking these
phylogenetic effects had a ΔDIC of 3.761 demonstrating support for the importance of phylogenetic
effects [51,52]. We also computed DIC weights of parameters to show the relative importance of each
on model fit. Distance from C. elegans had a weight of 0.763 and pairwise phylogenetic distance
between sets of species had a weight of 0.648. Since both weights are greater than 0.5, each
phylogenetic effect is more likely than not to be included in the ‘true’ model. Moreover, models that
included at least one of these phylogenetic effects had a weight of 0.926, demonstrating very strong
support for phylogenetic effects on susceptibility.
15
Figure 1. Species across the Caenorhabditis phylogeny are susceptible to Orsay virus (i.e. Ct values
smaller than the infection determination cut off (dashed line, see methods). Note that smaller Ct values
imply more virus). The asterisk on the left side of the y-axis shows the Ct value from ‘benchmark 5’ for
the sample with the most detectable virus (Table 2). The phylogeny (bottom left) is pruned from [44].
Many species currently have uncertain phylogenetic placement (right). Species for which a clade is
hypothesized are color-coded accordingly. These hypotheses were obtained from [53]. However, clades
are unknown for C. sp. 62, C. sp. 63, C. sp. 64, C. sp. 65, C. sp. 66. Shapes indicate different strains within
a species, colors differentiate clades, but are otherwise only varied to aid visualization. Open gold circles
and diamonds indicate Ct values for positive controls (‘control 2’ and ‘control 3’ plates respectively;
Table 2).
Transmission Assays
The primary exposure populations (passage 0) in our transmission assay were treated nearly identically
to populations in our susceptibility assay. As an internal control, we thus note high concordance
between Ct measures in both assays (correlation coefficient = 0.85). Most replicates of C. elegans strains
as well as positive control replicates (C. elegans strain N2) maintained high levels of virus through five
16
passages (Figure 2). However, virus was lost in 1 out of 3 control replicates in both blocks; in retrospect,
this is unremarkable since the N2 strain used for controls is known to be more resistant to Orsay virus
than many other C. elegans strains [29]. Non-elegans strains did not transmit the virus as well in most
cases. Virus was undetectable in the first passage population in all replicates of C. doughertyi, C.
wallacei, C. latens strain JU3325, C. waitukubuli, C. sp. 25, C. castelli, C. sp. 24, C. sp. 63, and C. sp. 66
strains JU4088 and JU4096. Virus was also undetectable in the first passage population in one or two
replicates of C. tropicalis, C. latens strain 724, C. macrosperma, C. sulstoni, C. sp. 65 strain JU4087, and C.
sp. 66 strain JU4094. Virus was maintained for 1-4 passages in at least one replicate of strains of C.
tropicalis, C. latens strain VX80, C. macrosperma, C. sulstoni, C. sp. 65 strains JU4093 and JU4087, and C.
sp. 66 strain JU4094. Virus was detectable through the 5th passage in four non-elegans replicates
belonging to three strains of different species: 1 replicate of C. sulstoni strain SB454, 1 replicate of C.
latens strain JU724, and 2 replicates of C. sp. 65 strain JU4093 (Figure 2).
As with the susceptibility data, we again identified factors associated with differences in
transmission through model analysis. Our best model again included phylogenetic effects of distance
from C. elegans and phylogenetic distance between pairwise sets of species. This model additionally
included an effect of viral amplification (Ct) in primary exposure populations (Table 4), which was
correlated with phylogenetic distance from C. elegans (correlation coefficient = 0.461). DIC weights were
as follows: amplification (Ct) in primary exposure populations = 0.858, phylogenetic distance from C.
elegans = 0.542, pairwise phylogenetic distance between sets of species = 0.548. Models including at
least one of the phylogenetic effects had a weight of 0.784. These weights indicate strong support for an
effect of viral amplification in primary exposure populations and at least some support for each
phylogenetic effect in explaining transmission ability.
17
Figure 2. Orsay virus persisted to different extents when susceptible hosts were sequentially passaged to
virus-free plates. “Passage 0” denotes the primary exposure population. This experiment was carried out
in two blocks indicated by shape (circle=block 1, triangle=block 2). N2 controls were present in both
blocks, shown in black. Colors match color-coded phylogeny in Figure 1. Shades represent different
strains within a species: C. elegans GXW1 (dark green), ED3042 (medium green), JU1401 (light green); C.
doughertyi JU1331; C. tropicalis JU1428; C. wallacei JU1873; C. latens JU724 (dark green; one of the
three replicate lines was removed from analysis due to bacterial contamination), VX80 (medium green),
JU3325 (light green); C. macrosperma JU1857; C. sulstoni SB454; C. waitukubuli NIC564; C. sp. 25
ZF1092, C. castelli JU1426; C. sp. 24 JU2837; C. sp. 63 JU4056; C. sp. 65 JU4093 (dark gray), JU4087
(medium gray); C. sp. 66 JU4094 (dark gray), JU4088 (medium gray), JU4096 (light gray).
DISCUSSION
In our study examining the host range of Orsay virus, we determined that at least 13
Caenorhabditis species in addition to C. elegans are susceptible and that hosts varied in their ability to
18
transmit the virus. Specifically, we found 21 susceptible Caenorhabditis strains (including 3 out of 3 C.
elegans strains) out of 84 tested strains belonging to 44 species. When susceptible strains were assayed
for transmission ability, 10 strains were dead-end hosts in all replicates, and 6 strains (3 C. elegans
strains, 1 C. sulstoni strain, 1 C. latens strain, and 1 C. sp. 65 strain) showed virus persistence for at least
five passages in at least one replicate. The remaining 5 susceptible strains showed stuttering chains of
transmission in at least one replicate. Both susceptibility and transmission ability were associated with
two phylogenetic effects: distance from C. elegans and phylogenetic distance between pairwise sets of
species. Transmission ability was also positively associated with viral amplification in primary exposure
populations. Overall, we argue that this study primes the Caenorhabditis-Orsay virus system to be
valuable for experimental studies on the ecology and evolution of pathogen spillover and emergence.
Replicating findings from several other experimental studies of host range [34–36], we found
evidence of phylogenetic effects on susceptibility. Host species more closely related to the native host C.
elegans were more likely to be susceptible to infection, and closely related hosts had more similar
susceptibilities regardless of their relationship to the native host. These patterns may arise because
closely related hosts likely have similar receptors, pathogen defenses, and within-host environments
[10]. We expect that the importance of phylogenetic effects would only become more readily detectable
if our unplaced Caenorhabditis species were placed on the phylogeny, since their lack of placement cost
us statistical power. Importantly, we recovered an effect of phylogenetic distance from C. elegans even
though few species are closely related to C. elegans (Figure 1). We hypothesize that the statistical
support for this phylogenetic effect would become stronger if this work were repeated with related
viruses of C. briggsae, which is a member of a clade with more closely related species.
We also found detectable effects of phylogeny on transmission ability. Although patterns
consistent with a phylogenetic effect on transmission have been identified [10,35,54], to the best of our
knowledge, this study is the first to empirically document such a pattern. In comparison to susceptibility,
19
however, the association between phylogeny and transmission ability had weaker statistical support.
This reduction in statistical support may have resulted from the small number of hosts tested, since we
were only able to assay transmission in susceptible strains. Moreover, the susceptible species were less
well distributed across the phylogenetic tree than random (i.e. the mean distance from C. elegans for
strains in this assay was 0.149 and ranged from 0 to 0.419, while the mean distance from C. elegans
across all strains in the susceptibility assay was 0.220 and ranged from 0 to 0.794). In addition, the
moderate correlation between phylogenetic distance from C. elegans and our other focal fixed effect,
viral amplification in primary exposure populations, may have made a phylogenetic distance effect more
difficult to detect.
The strongest predictor of transmission ability in our study was viral amplification in primary
exposure populations. We can imagine at least three reasons why amplification in primary exposure
populations may matter for transmission. First, high levels of viral amplification may be indicative of
some level of “pre-adaptation”, the ability to infect and transmit among novel hosts before additional
evolutionary changes [55]. Indeed, the correlation between viral amplification in primary exposure
populations with phylogenetic distance from C. elegans is consistent with this idea. Second, if hosts can
shed the virus, high levels of viral amplification may expose conspecifics to higher doses, which could
increase infection prevalence. If this was the case in our experiment, animals passaged from primary
exposure populations with more viral amplification may have been more likely to have been infected.
Third, larger virus populations may harbor more genetic variation, increasing opportunities for adaptive
evolution that could maintain persistence of the virus in the spillover host. Indeed, evolutionary rescue
theory has shown that larger populations are more likely to persist in comparison to smaller ones [56].
Here we have documented spillover and transmission of Orsay virus in Caenorhabditis hosts. It is
important to note, however, that the patterns we see with our susceptibility and transmission assays
may not fully predict spillover and emergence patterns among Caenorhabditis hosts in the wild.
20
Exposure risk is a key determinant of spillover and emergence [7], but in our experiments, we exposed
all hosts equally. Orsay virus exposure risk for Caenorhabditis species in nature is unknown since we
know little about the distributions of Caenorhabditis species and their viruses [57,58]. The two host
species that have been most extensively studied in the wild, C. elegans and C. briggsae, do have
overlapping distributions [59], but appear to be refractory to each other's viruses [29]. However, the fact
that three viruses related to Orsay virus have been found in C. briggsae [29,32,33] suggests that at least
one host jump has occurred in the past, since the viruses appear to be much more closely related [33]
than C. briggsae and C. elegans [60].
C. elegans has long been used as a model system to study infectious disease [22]. We argue that
the Caenorhabditis-Orsay virus system will be useful for studying virus spillover and emergence since the
system has many attractive features, including large populations, short experimental timelines,
replicable experimental manipulations, natural transmission, and related hosts with variable viral
competence. In particular, this system can be used to understand how ecological attributes of host
populations (e.g. density, diversity, immunity, heterogeneity) facilitate or impede emergence and how
evolution proceeds as a virus adapts to a new host species (e.g. phenotypic changes, genetic changes,
predictability, repeatability).
The Caenorhabditis-Orsay virus system joins a small set of empirical systems suitable for
studying spillover and emergence. Prior studies using other systems have yielded useful insights into
these processes. For example, bacteria-phage systems have been used to show that the probability of
virus emergence is highest when host populations contain intermediate combinations of native and
novel hosts [61], that pathogen variation in reservoir hosts drives emergence in novel hosts [62], and
that mutations that allow phages to infect novel hosts also constrain further host range expansion [63].
Plant-virus systems have been used to document the effects of host species on the fitness distribution of
viral mutations [64], to determine the importance of dose, selection, and viral replication for adaptation
21
to resistant hosts [65], and to characterize how spillover can impact competition among host species
[66,67]. Drosophila-virus systems have been used to show that viruses evolve in similar ways when
passaged through closely related hosts [42] and to show that spillover dynamics can depend on
temperature [68].
The Caenorhabditis-Orsay virus model can be uniquely useful for studying how ecology impacts
spillover and emergence in animal systems since population characteristics like density, genetic
variation, and immunity can be readily manipulated and virus transmission occurs without intervention
by a researcher. Caenorhabditis hosts have complex animal physiology, immune systems, and behavior,
meaning that this system can be useful for revealing the importance of variation in these traits. In this
study, we identified multiple susceptible spillover hosts that have variation in transmission ability. In the
future, these hosts can be used not only to probe how ecology impacts spillover and emergence, but
also to better understand how and why spillover and emergence patterns may differ across hosts.
ACKNOWLEDGEMENTS
We thank Marie-Anne Félix and Aurélien Richaud for sending Caenorhabditis strains and for advising on
their propagation and on molecular species identification. We are also grateful to Marie-Anne Félix for
her comments on an earlier version of this manuscript. We thank Anton Aluquin for help with viral
extractions. We thank Beth Tuschhoff and Charles Geyer for helpful discussion about analysis and
Andrew Wood for providing his expertise with Roar, the Penn State supercomputing cluster. We thank
Lewis Stevens for technical guidance on working with phylogenetic data. We thank Amrita Bhattacharya,
Heverton Dutra, Beth McGraw, and Andrew Read for lively discussion of spillover science and pattern
interpretation. This work was partially supported by National Science Foundation grant DEB-1754692.
22
The funders had no role in study design, data collection and analysis, decision to publish, or preparation
of this article.
REFERENCES
1.
Zhou P et al. 2020 A pneumonia outbreak associated with a new coronavirus of probable bat origin.
Nature 579, 270–273.
2.
In press. Coronavirus resource center. COVID-19 Dashboard by the Center for Systems Science and
Engineering (CSSE) at Johns Hopkins University (JHU). See https://coronavirus.jhu.edu/map.html
(accessed on 13 October 2021).
3.
Jones KE, Patel NG, Levy MA, Storeygard A, Balk D, Gittleman JL, Daszak P. 2008 Global trends in
emerging infectious diseases. Nature 451, 990–993.
4.
Chua KB et al. 2000 Nipah virus: a recently emergent deadly paramyxovirus. Science 288, 1432–
1435.
5.
Plowright RK et al. 2015 Ecological dynamics of emerging bat virus spillover. Proc. Biol. Sci. 282,
20142124.
6.
Métras R et al. 2020 Estimation of Rift Valley fever virus spillover to humans during the Mayotte
2018-2019 epidemic. Proc. Natl. Acad. Sci. U. S. A. 117, 24567–24574.
7.
Plowright RK, Parrish CR, McCallum H, Hudson PJ, Ko AI, Graham AL, Lloyd-Smith JO. 2017 Pathways
to zoonotic spillover. Nat. Rev. Microbiol. 15, 502–510.
8.
Kreuder Johnson C et al. 2015 Spillover and pandemic properties of zoonotic viruses with high host
plasticity. Sci. Rep. 5, 14830.
9.
Olival KJ, Hosseini PR, Zambrana-Torrelio C, Ross N, Bogich TL, Daszak P. 2017 Host and viral traits
predict zoonotic spillover from mammals. Nature 546, 646–650.
10. Streicker DG, Turmelle AS, Vonhof MJ, Kuzmin IV, McCracken GF, Rupprecht CE. 2010 Host
phylogeny constrains cross-species emergence and establishment of rabies virus in bats. Science
329, 676–679.
11. Barrow LN, McNew SM, Mitchell N, Galen SC, Lutz HL, Skeen H, Valqui T, Weckstein JD, Witt CC.
2019 Deeply conserved susceptibility in a multi-host, multi-parasite system. Ecol. Lett. 22, 987–998.
12. Davies TJ, Pedersen AB. 2008 Phylogeny and geography predict pathogen community similarity in
wild primates and humans. Proc. Biol. Sci. 275, 1695–1701.
13. Charleston MA, Robertson DL. 2002 Preferential host switching by primate lentiviruses can account
for phylogenetic similarity with the primate phylogeny. Syst. Biol. 51, 528–535.
23
14. Ssebuliba E, Davies TJ. 2021 Assessing the phylogenetic host breadth of millet pathogens and its
implication for disease spillover. Ecological Solutions and Evidence 2. (doi:10.1002/2688-
8319.12040)
15. Alexander KA, Carlson CJ, Lewis BL, Getz WM, Marathe MV, Eubank SG, Sanderson CE, Blackburn JK.
2018 The Ecology of Pathogen Spillover and Disease Emergence at the Human-Wildlife-
Environment Interface. In The Connections Between Ecology and Infectious Disease (ed CJ Hurst),
pp. 267–298. Springer International Publishing.
16. Jones RAC. 2020 Disease pandemics and major epidemics arising from new encounters between
indigenous viruses and introduced crops. Viruses 12. (doi:10.3390/v12121388)
17. Faust CL, McCallum HI, Bloomfield LSP, Gottdenker NL, Gillespie TR, Torney CJ, Dobson AP,
Plowright RK. 2018 Pathogen spillover during land conversion. Ecol. Lett. 21, 471–483.
18. Lloyd-Smith JO, George D, Pepin KM, Pitzer VE, Pulliam JRC, Dobson AP, Hudson PJ, Grenfell BT.
2009 Epidemic dynamics at the human-animal interface. Science 326, 1362–1367.
19. Wasik BR, de Wit E, Munster V, Lloyd-Smith JO, Martinez-Sobrido L, Parrish CR. 2019 Onward
transmission of viruses: how do viruses emerge to cause epidemics after spillover? Philos. Trans. R.
Soc. Lond. B Biol. Sci. 374, 20190017.
20. Dennehy JJ. 2017 Evolutionary ecology of virus emergence. Ann. N. Y. Acad. Sci. 1389, 124–146.
21. McLean KD, Duffy MA. 2020 Ecological Context Influences Evolution in Host-Parasite Interactions:
Insights from the Daphnia-Parasite Model System. In Evolution in Action: Past, Present and Future:
A Festschrift in Honor of Erik D. Goodman (eds W Banzhaf, BHC Cheng, K Deb, KE Holekamp, RE
Lenski, C Ofria, RT Pennock, WF Punch, DJ Whittaker), pp. 289–307. Springer International
Publishing.
22. Gibson AK, Morran LT. 2017 A model for evolutionary ecology of disease: The case for
Caenorhabditis nematodes and their natural parasites. J. Nematol. 49, 357–372.
23. O’Callaghan D, Vergunst A. 2010 Non-mammalian animal models to study infectious disease:
worms or fly fishing? Curr. Opin. Microbiol. 13, 79–85.
24. Minkah NK, Schafer C, Kappe SHI. 2018 Humanized mouse models for the study of human malaria
parasite biology, pathogenesis, and immunity. Front. Immunol. 9, 807.
25. Harnish JM, Link N, Yamamoto S. 2021 Drosophila as a model for infectious diseases. Int. J. Mol. Sci.
22. (doi:10.3390/ijms22052724)
26. Pagán I, Fraile A, Fernandez-Fueyo E, Montes N, Alonso-Blanco C, García-Arenal F. 2010 Arabidopsis
thaliana as a model for the study of plant-virus co-evolution. Philos. Trans. R. Soc. Lond. B Biol. Sci.
365, 1983–1995.
27. Marsh EK, May RC. 2012 Caenorhabditis elegans, a model organism for investigating immunity.
Appl. Environ. Microbiol. 78, 2075–2081.
24
28. Gammon DB. 2017 Caenorhabditis elegans as an emerging model for virus-host interactions. J.
Virol. 91. (doi:10.1128/JVI.00509-17)
29. Félix M-A et al. 2011 Natural and experimental infection of Caenorhabditis nematodes by novel
viruses related to nodaviruses. PLoS Biol. 9, e1000586.
30. Sanjuán, Rafael, Nebot, Miguel R., Chirico, Nicola, Mansky, Louis M., Belshaw, Robert. 2010 Viral
mutation rates. J. Virol. 84, 9733–9748.
31. Holmes EC. 2010 The comparative genomics of viral emergence. Proc. Natl. Acad. Sci. U. S. A. 107
Suppl 1, 1742–1746.
32. Frézal L, Jung H, Tahan S, Wang D, Félix M-A. 2019 Noda-like RNA viruses infecting Caenorhabditis
nematodes: sympatry, diversity, and reassortment. J. Virol. 93. (doi:10.1128/JVI.01170-19)
33. Franz CJ, Zhao G, Félix M-A, Wang D. 2012 Complete genome sequence of Le Blanc virus, a third
Caenorhabditis nematode-infecting virus. J. Virol. 86, 11940.
34. Perlman SJ, Jaenike J. 2003 Infection success in novel hosts: an experimental and phylogenetic
study of Drosophila-parasitic nematodes. Evolution 57, 544–557.
35. Longdon B, Hadfield JD, Webster CL, Obbard DJ, Jiggins FM. 2011 Host phylogeny determines viral
persistence and replication in novel hosts. PLoS Pathog. 7, e1002260.
36. Hoverman JT, Gray MJ, Haislip NA, Miller DL. 2011 Phylogeny, life history, and ecology contribute to
differences in amphibian susceptibility to ranaviruses. Ecohealth 8, 301–319.
37. Gilbert GS, Webb CO. 2007 Phylogenetic signal in plant pathogen-host range. Proc. Natl. Acad. Sci.
U. S. A. 104, 4979–4983.
38. Levitt AL, Singh R, Cox-Foster DL, Rajotte E, Hoover K, Ostiguy N, Holmes EC. 2013 Cross-species
transmission of honey bee viruses in associated arthropods. Virus Res. 176, 232–240.
39. Stiernagle T. 2006 Maintenance of C. elegans. WormBook , 1–11.
40. Chen K, Franz CJ, Jiang H, Jiang Y, Wang D. 2017 An evolutionarily conserved transcriptional
response to viral infection in Caenorhabditis nematodes. BMC Genomics 18, 303.
41. Hadfield J. 2010 MCMC methods for multi-response generalized linear mixed models: The
MCMCglmm R package. J. Stat. Softw. 33, 1–22.
42. Longdon B, Day JP, Alves JM, Smith SCL, Houslay TM, McGonigle JE, Tagliaferri L, Jiggins FM. 2018
Host shifts result in parallel genetic changes when viruses evolve in closely related species. PLoS
Pathog. 14, e1006951.
43. R Core Team. 2020 R: A Language and Environment for Statistical Computing.
44. Stevens L et al. 2020 The Genome of Caenorhabditis bovis. Curr. Biol. 30, 1023-1031.e4.
45. Paradis E, Schliep K. 2019 ape 5.0: an environment for modern phylogenetics and evolutionary
analyses in R. Bioinformatics 35, 526–528.
25
46. Hadfield JD, Nakagawa S. 2010 General quantitative genetic methods for comparative biology:
phylogenies, taxonomies and multi-trait models for continuous and categorical characters. J. Evol.
Biol. 23, 494–508.
47. Gelman A. 2006 Prior distributions for variance parameters in hierarchical models(Comment on
Article by Browne and Draper). Bayesian Analysis 1, 515–533.
48. Spiegelhalter DJ, Best NG, Carlin BP, van der Linde A. 2014 The deviance information criterion: 12
years on. J. R. Stat. Soc. Series B Stat. Methodol. 76, 485–493.
49. Burnham KP, Anderson DR. 2003 Model Selection and Multimodel Inference: A Practical
Information-Theoretic Approach. Springer Science & Business Media.
50. Ashe A, Bélicard T, Le Pen J, Sarkies P, Frézal L, Lehrbach NJ, Félix M-A, Miska EA. 2013 A deletion
polymorphism in the Caenorhabditis elegans RIG-I homolog disables viral RNA dicing and antiviral
immunity. Elife 2, e00994.
51. Mazerolle MJ. 2006 Improving data analysis in herpetology: using Akaike’s Information Criterion
(AIC) to assess the strength of biological hypotheses. Amphib-reptil. 27, 169–180.
52. Bolker BM, Brooks ME, Clark CJ, Geange SW, Poulsen JR, Stevens MHH, White J-SS. 2009
Generalized linear mixed models: a practical guide for ecology and evolution. Trends Ecol. Evol. 24,
127–135.
53. Kiontke K. 2018 RhabditinaDB. See https://wormtails.bio.nyu.edu/ (accessed on 2 July 2021).
54. Guth S, Visher E, Boots M, Brook CE. 2019 Host phylogenetic distance drives trends in virus
virulence and transmissibility across the animal-human interface. Philos. Trans. R. Soc. Lond. B Biol.
Sci. 374, 20190296.
55. Mollentze N, Biek R, Streicker DG. 2014 The role of viral evolution in rabies host shifts and
emergence. Curr. Opin. Virol. 8, 68–72.
56. Gonzalez A, Ronce O, Ferriere R, Hochberg ME. 2013 Evolutionary rescue: an emerging focus at the
intersection between ecology and evolution. Philos. Trans. R. Soc. Lond. B Biol. Sci. 368, 20120404.
57. Félix M-A, Jovelin R, Ferrari C, Han S, Cho YR, Andersen EC, Cutter AD, Braendle C. 2013 Species
richness, distribution and genetic diversity of Caenorhabditis nematodes in a remote tropical
rainforest. BMC Evol. Biol. 13, 10.
58. Cutter AD. 2015 Caenorhabditis evolution in the wild. Bioessays 37, 983–995.
59. Félix M-A, Duveau F. 2012 Population dynamics and habitat sharing of natural populations of
Caenorhabditis elegans and C. briggsae. BMC Biol. 10, 59.
60. Cutter AD, Félix M-A, Blaxter M. 2016 Myths and misconceptions about C. elegans ecology and
evolution. Worm Breed. Gaz. 20.
26
61. Chabas H, Lion S, Nicot A, Meaden S, van Houte S, Moineau S, Wahl LM, Westra ER, Gandon S. 2018
Evolutionary emergence of infectious diseases in heterogeneous host populations. PLoS Biol. 16,
e2006738.
62. Dennehy JJ, Friedenberg NA, McBride RC, Holt RD, Turner PE. 2010 Experimental evidence that
source genetic variation drives pathogen emergence. Proc. Biol. Sci. 277, 3113–3121.
63. Zhao L, Seth-Pasricha M, Stemate D, Crespo-Bellido A, Gagnon J, Draghi J, Duffy S. 2019 Existing
host range mutations constrain further emergence of RNA viruses. J. Virol. 93.
(doi:10.1128/JVI.01385-18)
64. Lalić J, Cuevas JM, Elena SF. 2011 Effect of host species on the distribution of mutational fitness
effects for an RNA virus. PLoS Genet. 7, e1002378.
65. Rousseau E et al. 2018 Impact of genetic drift, selection and accumulation level on virus adaptation
to its host plants. Mol. Plant Pathol. 19, 2575–2589.
66. Mordecai EA. 2013 Despite spillover, a shared pathogen promotes native plant persistence in a
cheatgrass-invaded grassland. Ecology 94, 2744–2753.
67. Beckstead J, Meyer SE, Connolly BM, Huck MB, Street LE. 2010 Cheatgrass facilitates spillover of a
seed bank pathogen onto native grass species. J. Ecol. 98, 168–177.
68. Roberts KE, Hadfield JD, Sharma MD, Longdon B. 2018 Changes in temperature alter the potential
outcomes of virus host shifts. PLoS Pathog. 14, e1007185.
| 2021 | Developing an empirical model for spillover and emergence: Orsay virus host range in | 10.1101/2021.12.10.472097 | [
"Shaw Clara L.",
"Kennedy David A."
] | creative-commons |
Blood and site of disease inflammatory profiles differ in HIV-1-infected pericardial
1
tuberculosis patients
2
3
Hygon Mutavhatsindia,i*, Elsa Du Bruyna, Sheena Ruzivea, Patrick Howletta, Alan Sherb,
4
Katrin D. Mayer-Barberc, Daniel L. Barberd, Mpiko Ntsekhea,e,f, Robert J. Wilkinsona,e,g,h and
5
Catherine Rioua,i
6
7
a Wellcome Centre for Infectious Disease Research in Africa, Institute of Infectious Disease
8
and Molecular Medicine, University of Cape Town, Observatory, 7925, South Africa.
9
b Immunobiology Section, Laboratory of Parasitic Diseases, National Institute of Allergy and
10
Infectious Diseases, National Institutes of Health, Bethesda, MD, USA.
11
c Inflammation and Innate Immunity Unit, Laboratory of Clinical Immunology and
12
Microbiology, National Institute of Allergy and Infectious Diseases, National Institutes of
13
Health, Bethesda, MD, USA.
14
d T Lymphocyte Biology Section, Laboratory of Parasitic Diseases, National Institute of
15
Allergy and Infectious Diseases, National Institutes of Health, Bethesda, MD, USA.
16
e Department of Medicine, University of Cape Town, Observatory, 7925, South Africa.
17
f Division of Cardiology, Department of Medicine, University of Cape Town, Observatory,
18
7925, South Africa.
19
g Imperial College London, SW7 2AZ, UK.
20
h The Francis Crick Institute, 1 Midland Rd, London NW1 1AT, UK.
21
i Division of Medical Virology, Department of Pathology, University of Cape Town,
22
Observatory, 7925, South Africa.
23
24
*Corresponding Author: Hygon Mutavhatsindi, CIDRI-Africa, IDM, University of Cape
25
Town, 1 Anzio Road, Observatory, 7925, Cape Town, South Africa. Email:
26
Hygon.mutavhatsindi@uct.ac.za / h.mutavhatsindi@gmail.com
27
28
Abstract
29
Objectives. To better understand the pathogenesis of pericardial tuberculosis (PCTB), we
30
sought to characterize the systemic inflammatory profile in HIV-1-infected participants with
31
latent TB infection (LTBI), pulmonary TB (PTB) and PCTB.
32
Methods. Using Luminex, we measured 39 analytes in pericardial fluid (PCF) and paired
33
plasma from 18 PCTB participants, and plasma from 16 LTBI and 20 PTB. Follow-up
34
plasma samples were also obtained from PTB and PCTB participants. HLA-DR expression
35
on Mtb-specific CD4 T cells was measured in baseline samples using flow cytometry.
36
Results. Assessment of the overall systemic inflammatory profile by principal component
37
analysis showed that the inflammatory profile of active TB participants was distinct from the
38
LTBI group, while PTB patients could not be distinguished from those with PCTB. In the
39
LTBI group, 12 analytes showed a positive association with plasma HIV-1 viral load, and
40
most of these associations were lost in the diseased groups. When comparing the
41
inflammatory profile between PCF and paired blood, we found that the concentrations of
42
most analytes (24/39) were elevated at site of disease. However, the inflammatory profile in
43
PCF partially mirrored inflammatory events in the blood. After TB treatment completion, the
44
overall plasma inflammatory profile reverted to those observed in the LTBI group. Lastly,
45
HLA-DR expression showed the best performance for TB diagnosis compared to previously
46
described biosignatures built from soluble markers.
47
Conclusion. Our results describe the inflammatory profile associated with PTB and PCTB
48
and emphasize the potential role of HLA-DR as a promising biomarker for TB diagnosis.
49
50
Key words: Pericardial tuberculosis, Inflammatory profile, site of disease, diagnosis,
51
treatment response
52
1. Introduction
53
Tuberculosis (TB) is the leading cause of death amongst human immunodeficiency virus
54
(HIV-1)-infected individuals [1]. Moreover, 15 to 20% of all TB cases in developing
55
countries are accounted for by extrapulmonary TB (EPTB) [2,3] which disproportionately
56
affects immunocompromised patients [4,5]. Pericardial TB (PCTB), a severe form of EPTB,
57
is the most common cause of pericarditis in TB endemic countries in Africa and Asia [6–8].
58
PCTB related morbidity is significant, with mortality (which generally occurs early in the
59
onset of the disease), as high as 26% and increasing to approximately 40% in cohorts of
60
predominantly HIV-infected persons [9,10].
61
HIV impairs both innate and adaptive immune responses, with the most obvious immune
62
defect being a progressive reduction in absolute CD4+ T cell numbers and systemic hyper
63
activation [11]. HIV-1 has also been shown to alter the balance of Mtb-specific T helper
64
subsets, through the reduction of Th17 cells and T regulatory (Treg) cells [12–14], suggesting
65
that HIV shifts Mtb-specific responses toward a more pathogenic/inflammatory profile [12].
66
Pulmonary TB-induced systemic inflammation has been studied extensively showing high
67
concentrations of acute phase proteins and pro-inflammatory cytokines including C-reactive
68
protein (CRP), serum amyloid P component (SAP), interferon gamma (IFN-γ), interferon
69
gamma-induced protein 10 (IP-10), chemokine (C-C motif) ligand 1 (CCL1) and tumor
70
necrosis factor alpha (TNF-α) in serum/plasma of active TB participants in comparison to
71
other respiratory diseases, LTBI or healthy controls [15–18]. Furthermore, in patients with
72
pulmonary TB admitted to intensive care units, serum levels of inflammatory factors such as
73
interleukin (IL)-1, IL-6, IL-10, IL-12, and IL-4 are upregulated compared to healthy controls
74
[19]. Based on these results several host inflammatory marker signatures have been proposed
75
as biomarkers for TB diagnosis and the monitoring of treatment response, with superior
76
performance compared to smear microscopy [15,16,20,21].
77
However, the influence of HIV-1 co-infection on the immune response to Mtb in the context
78
of pulmonary and extrapulmonary TB remains poorly understood. Moreover, studies
79
assessing immune responses at site of disease are scarce [22–24]. These studies reported
80
higher levels of cytokines/chemokines at the site of disease in comparison to paired
81
peripheral blood with exception of a few analytes, such as interferon gamma (IFN-γ), IL-1β
82
and IL-8 which were reported to be significantly higher in peripheral blood instead [22–24].
83
Thus, in the current study, we measured 39 soluble markers in blood and at site of disease
84
(pericardial fluid) to 1) compare the systemic cytokine environment between pulmonary and
85
pericardial TB (PCTB) patients coinfected with HIV-1, 2) define the relationship between
86
HIV viral load and the inflammatory profiles, 3) define whether peripheral inflammation
87
signatures mirrors those at site of infection, 4) assess the impact of TB treatment on systemic
88
inflammation and 5) evaluate the performance of previously described blood-based
89
biomarkers to discriminate latent from active TB.
90
2. Materials and methods
91
2.1. Study population
92
Participants included in this study (n = 54) were recruited from the Ubuntu Clinic, Site B,
93
Khayelitsha or the Groote Schuur Hospital Cardiology Unit (Cape Town, South Africa)
94
between June 2017 and April 2019. Participants were divided in three groups according to
95
their TB status: i) Pericardial tuberculosis (PCTB, n=18), ii) Pulmonary tuberculosis (PTB,
96
n=20) and iii) Latent tuberculosis infection (LTBI, n=16).
97
The PCTB group (n = 18) included patients with either definite (Mtb culture positive in
98
pericardial fluid (PCF), n = 9) or probable PCTB (n = 9). Probable PCTB was defined based
99
on evidence of pericarditis with microbiologic confirmation of Mtb-infection elsewhere in the
100
body and/or an exudative, lymphocyte predominant pericardial effusion with elevated
101
adenosine deaminase (≥35 U/L), according to Mayosi et al [25]. Only three PCTB patients
102
were HIV negative. Paired PCF and Blood were collected at the same time for PCTB
103
patients. Patients from the PTB group (n = 20) were all HIV positive, tested sputum Xpert
104
MTB/RIF (Xpert, Cepheid, Sunnyvale, CA) positive and had clinical symptoms and/or
105
radiographic evidence of tuberculosis. All were infected by drug sensitive isolates of Mtb and
106
had received no more than one dose of anti-tubercular treatment (ATT) at the time of baseline
107
blood sampling. The LTBI group (n = 16) were all asymptomatic, had a positive IFN-γ
108
release assay (IGRA, QuantiFERON-TB Gold In-Tube, Qiagen, Hilden, Germany), tested
109
sputum Xpert MTB/RIF negative and exhibited no clinical evidence of active TB. All LTBI
110
participants were HIV positive. Clinical characteristics of the study participants are shown in
111
Table 1. Sputum and PCF Mtb culture, CD4 count, and HIV VL were performed by the
112
South African National Health Laboratory Services. Active TB patients (PTB or PCTB) were
113
followed up over the duration of their ATT and additional blood draws were performed at
114
week 6 for PCTB, week 8 for PTB and week 24 for both diseased groups. All participants
115
were adults (age ≥ 18 years) and provided written informed consent. The study was approved
116
by the University of Cape Town Human Research Ethics Committee (050/2015 and
117
271/2019).
118
119
2.2. Pericardial fluid, blood collection and whole blood assay
120
Pericardial fluid was obtained at the time of pericardiocentesis, placed in sterile Falcon tubes
121
and transported to the laboratory at 4°C. Blood was collected in sodium heparin tubes and
122
processed within 3 hours of collection. The whole blood or whole PCF assay were adapted
123
from the protocol described by Hanekom et al [26]. Briefly, 0.5 mL of whole blood or 1 mL
124
of whole PCF were stimulated with a pool of 300 Mtb-derived peptides (Mtb300, 2 μg mL-1)
125
[27] at 37°C for 5 hours in the presence of the co-stimulatory antibodies, anti-CD28 and anti-
126
CD49d (1 μg mL-1 each; BD Biosciences, San Jose, CA, USA) and Brefeldin-A (10 μg mL-1;
127
Sigma-Aldrich, St Louis, MO, USA). Unstimulated cells were incubated with co-stimulatory
128
antibodies and Brefeldin-A only. Red blood cells were then lysed in a 150 mM NH4Cl, 10
129
mM KHCO3 and 1 mM Na4EDTA solution. Cells were stained with a Live/Dead near-
130
infrared dye (Invitrogen, Carlsbad, CA, USA) and then fixed using a transcription factor
131
fixation buffer (eBioscience, San Diego, CA, USA), cryopreserved in freezing media (50%
132
fetal calf serum, 40% RPMI and 10% dimethyl sulfoxide) and stored in liquid nitrogen until
133
use.
134
135
2.3. Cell staining and flow cytometry
136
Cryopreserved cells were thawed, washed and permeabilized with a transcription factor
137
perm/wash buffer (eBioscience). Cells were then stained at room temperature for 45 min with
138
the following antibodies: CD3 BV650 (OKT3; BioLegend, San Diego, CA, USA), CD4
139
BV785 (OKT4; BioLegend), CD8 BV510 (RPA-T8; BioLegend), HLA-DR BV605 (L243;
140
BioLegend), IFN-γ BV711 (4S.B3; BioLegend), TNF-α PE-Cy7 (Mab11; BioLegend
141
eBioscience) and IL-2 PE/Dazzle (MQ1-17H12; BioLegend). Samples were acquired on a
142
BD LSR-II and analysed using FlowJo (v10.8.1, FlowJo LCC, Ashland, OR, USA). A
143
positive cytokine response was defined as at least twice the background of unstimulated cells.
144
To define the phenotype of Mtb300-specific CD4 T cells, a cut-off of 30 events was used.
145
146
2.4. Luminex® Multiplex Immunoassay
147
Using Luminex® technology, we measured the levels of 39 analytes using antibodies
148
supplied by Merck Millipore (Billerica, Massachusetts, USA) and R&D Systems
149
(Minneapolis, MN, USA). The analytes measured included: Granzyme B (GrB), interleukin 2
150
(IL-2), interleukin 8 (IL-8), interleukin 12p40 (IL-12p40), macrophage colony-stimulating
151
factor (M-CSF), tumor necrosis factor alpha (TNF-α), transforming growth factor beta (TGF-
152
β), complement component 3 (C3), complement component 4 (C4), C-reactive protein (CRP),
153
serum amyloid P (SAP), interleukin 22 (IL-22), Galectin-3 (Gal-3), intercellular adhesion
154
molecule 1 (ICAM-1), neural cell adhesion molecule 1 (NCAM-1), granulocyte colony-
155
stimulating factor (G-CSF), interferon gamma (IFN-γ), interleukin 6 (IL-6), interleukin 10
156
(IL-10), interleukin 27 (IL-27) and vascular endothelial growth factor (VEGF), monokine
157
induced by gamma (MIG), monocyte chemoattractant protein 2 (MCP-2), granulocyte
158
chemoattractant protein 2 (GCP-2), chemokine (C-X-C motif) ligand 11 (CXCL11),
159
macrophage inflammatory protein 1 beta (MIP-1β), chemokine (C-C motif) ligand 1 (CCL1)
160
and interferon gamma-induced protein 10 (IP-10), cluster of differentiation 163 (CD163),
161
interleukin 6 receptor alpha (IL-6Rα), cluster of differentiation 30 (CD30), interleukin 2
162
receptor alpha (IL-2Rα), apolipoprotein A-I (ApoA-I), apolipoprotein C-III (Apo-CIII),
163
oncostatin M (OSM), interleukin 33 receptor (IL-33R), osteopontin (OPN), platelet derived
164
growth factor BB (PDGF-BB) and thrombomodulin (TM). All samples were evaluated
165
undiluted or diluted according to the manufacturer’s recommendations. Samples were
166
randomized to assay plates with the experimenter blinded to sample data. All assays were
167
performed and read at UCT on the Bio-Plex platform (Bio-Rad), with the Bio-Plex Manager
168
Software (v6·1) used for bead acquisition and analysis.
169
170
2.5. Statistical Analyses
171
Statistical tests were performed in Prism (v9.1.3, GraphPad Software Inc, San Diego, CA,
172
USA). Non-parametric tests were used for all comparisons. The Kruskal-Wallis test with
173
Dunn’s multiple comparison test was used for multiple comparisons, the Spearman rank test
174
for correlation and the Mann-Whitney and Wilcoxon matched pairs test for unmatched and
175
paired samples, respectively. When the measured analyte was below the limit of detection in
176
more than 20% of the samples (i.e., M-CSF and IL-10), the analyte was not included in the
177
correlation with plasma HIV VL and HLA-DR expression on Mtb-specific CD4 T cells.
178
Unsupervised hierarchical clustering analysis (HCA, Ward method), principal component
179
analyses (PCA) were carried out in JMP (v16.0.0; SAS Institute, Cary, NC, USA). For HCA
180
and PCA, the min-max normalization method (i.e., feature scaling, analyte value - min / max
181
- min) was used to scale data in the 0 to 1 range. The predictive abilities of combinations of
182
analytes were investigated by general discriminant analysis (GDA) in JMP. The diagnostic
183
ability of HLA-DR expression on Mtb-specific CD4 T cells were assessed by receiver
184
operator characteristics (ROC) curve analysis. Optimal cut off values and associated
185
sensitivity and specificity were determined based on the Youden’s Index [28]. Analyte
186
network analysis was performed using Gephi (v0.9.2, University of Technology of
187
Compiègne, Compiègne, France). The Bonferroni method [29] was used to adjust for
188
multiple comparisons. A p-value of <0.05 was considered statistically significant.
189
3. Results
190
3.1 Study population
191
The clinical characteristics of participants are presented in Table 1. Participants (n = 54)
192
were classified into three groups according to their TB status: PCTB (n = 18), PTB (n = 20)
193
and LTBI (n = 16). Median age was comparable between the three groups. All participants
194
were HIV-infected except for three PCTB patients. LTBI participants had a lower plasma
195
HIV-1 viral load (VL) and higher absolute CD4 count compared to the PCTB and PTB
196
groups (median Log10 VL: 3.28 vs 4.68 and 4.79 copies mL-1, respectively and median CD4:
197
409 vs 141 and 176 cells mm-3, respectively, Table 1).
198
199
3.2 Comparison of the systemic inflammatory profile between LTBI, PTB and
200
PCTB.
201
Plasma levels of 39 analytes, including cytokines, chemokines, apolipoproteins, chemokine,
202
protein receptors, and fibrosis-related analytes, were measured in all participants (the
203
complete list of measured analytes is presented in the material and methods section).
204
Assessing the overall systemic inflammatory profile using unsupervised hierarchical
205
clustering (Fig. 1a) and principal component analysis (Fig. 1b) we showed an evident
206
separation between LTBI and active TB participants (PCTB and PTB), driven by elevated
207
levels of most of the measured inflammatory markers. However, there was no noticeable
208
separation between the PCTB and PTB groups, suggesting comparable systemic
209
inflammation in these groups. Individual analysis of measured analytes showed that 15
210
markers were significantly higher in both PTB and PCTB compared to the LTBI group,
211
including innate-related inflammation markers (such as IL-6, TNF-⍺, and IL-8), acute phase
212
protein (CRP) and chemokines (CCL1, MIG, IP-10 and CXCL11). VEGF also showed a
213
similar profile, with the p-value between LTBI and PTB being borderline significant (p =
214
0.0503) (Supplementary fig. 1 and Supplementary table 1). IL-6Rα and G-CSF were the
215
only markers that were observed to be differentially expressed between PTB and PCTB
216
(Supplementary fig. 1 and Supplementary table 1), highlighting similarities between the
217
different clinical forms of TB. Only one marker, OPN showed increased expression levels
218
only in the PCTB group compared to LTBI (p = 0.0063) while no significant difference was
219
observed for the PTB group (p = 0.374) (Supplementary fig. 1 and Supplementary table
220
1). Elevated OPN levels have been associated with severe tuberculosis [30]. Next, we defined
221
the interplay between markers, using network analysis (Fruchterman-Reingold algorithm,
222
Fig. 1c). In LTBI participants, TNF-α and MIP-1β were the most central nodes, showing the
223
most connections (positive associations) with other analytes. In active TB patients (both PTB
224
and PCTB), the network structure was substantially altered; and while MIP-1β remained a
225
predominant node, TGF-β emerged as a new influential node, with multiple negative
226
associations with analytes such as IL-12p40, ApoA-I or G-CSF (Fig. 1c). Overall, these
227
results illustrate that active TB disease significantly increases systemic inflammation and
228
PCTB and PTB participants share similar inflammatory signatures.
229
230
3.3 Relationship between inflammatory profile and HIV viral load
231
To examine the interplay between HIV viral load (VL) and cytokine profile, we defined the
232
associations between cytokine concentrations and HIV VL in plasma. Of the 39 measured
233
analytes, 12 markers positively associated with HIV VL in the LTBI group (Fig. 2a). Several
234
of those have been previously reported as HIV-associated systemic inflammation markers,
235
including IL-2Rα [31], CXCL11 [32], IL-6 [33], IFN-γ [34], IP-10 [35], TNF-α [35], and
236
CD30 [36]. In both the PTB and PCTB groups, most of these correlations were disrupted
237
with six analytes correlating with HIV VL in the PTB group and only one in the PCTB group
238
(Fig. 2a). The only cytokine which maintained significant correlation with HIV VL in all
239
groups was IL-12p40, albeit the correlation strength was weaker in the diseased groups (r =
240
0.83, p = 0.0002 vs r = 0.49, p = 0.028 in the PTB group and r = 0.63, p = 0.012 in the PCTB
241
group) (Fig. 2b). IP-10 concentration only showed a significantly positive correlation with
242
HIV VL in the LTBI group (r = 0.82, p = 0.0002), and was largely disrupted in both the PTB
243
and PCTB groups (r = 0.29, p = 0.26 and r = 0.25, p = 0.37, respectively) (Fig. 2b). No
244
negative associations were observed in the LTBI and PTB groups, however, TGF-β showed a
245
strong negative association with HIV VL in the PCTB group (r = -0.65, p = 0.0133) (Fig. 2a).
246
These findings suggest that active TB disease disrupts HIV-associated systemic
247
inflammation.
248
249
3.4 Profile of soluble markers in plasma compared to pericardial fluid
250
To better understand compartmentalization, we compared the profiles of expression of the 39
251
measured analytes in plasma and PCF from PCTB participants, using hierarchical clustering
252
analysis and PCA (Fig. 3a and b). There was a clear separation between sample types, where
253
PC1 accounted for 42% and PC2 11.2% of the variance (Fig. 3b). Furthermore, visualizing
254
sample clustering using a constellation plot, we observed that cluster 2 (comprised of PCF
255
samples) was divided into 2 distinct sub-clusters, where cluster 2b was enriched in
256
participants who were PCF culture positive (5/7, 72%) compared to patients included in
257
cluster 2a (4/12, 33%) (Fig. 3c). However, looking at individual analytes, we did not find
258
significant difference between PCF culture negative and PCF culture positive samples (data
259
not shown).
260
Univariate analysis of analytes showed that the concentrations of 25 out of the 39 measured
261
analytes were significantly higher in PCF in comparison to paired plasma samples, only 9/39
262
were significantly higher in plasma compared to PCF, and 5/39 showed no significant
263
difference in expression between the two sample types after correction of the p-values for
264
multiple testing (Supplementary fig. 2 and Supplementary table 2).
265
To better understand the relationship between peripheral and site of disease inflammation,
266
pairwise comparisons (plasma vs PCF) were assessed. Significant positive correlations were
267
observed for 18 out of the 39 analytes (with r and p ranging from 0.98 - 0.47 and <0.0001 -
268
0.048, respectively), the highest Spearman’s rank r values for significant positive correlations
269
were observed for ICAM-1, SAP, and ApoA-I (Fig. 3d). A summarized representation of the
270
associations between plasma and PCF for each analyte is shown in fig. 3d and individual
271
correlation plots of all the significant associations are presented in supplementary fig. 3. We
272
then defined the interplay between markers in PCF, using network analysis (Fruchterman-
273
Reingold algorithm, Fig. 3e). OSM, MCP-2 and ApoA-I were the most central nodes, with
274
OSM and MCP-2 showing positive associations with other analytes. While ApoA-I showing
275
mostly negative associations with analytes such as TGF-β, IP-10 and Apo-CIII (Fig. 3e).
276
Overall, these results show that inflammatory response at site of disease was greater than in
277
blood. However, inflammatory profile in PCF partially mirrored inflammatory events in
278
blood.
279
280
3.5 Associations between systemic inflammation and the activation of Mtb-
281
specific CD4+ T cells in blood and at site of disease.
282
HLA-DR expression on peripheral Mtb-specific CD4+ T cells has been shown to
283
discriminate latent from active TB infection [37–39]. To better understand the relationship
284
between inflammation and T cell activation, we measured the expression of HLA-DR on
285
Mtb-specific CD4+ T cells in blood from LTBI, PTB, PCTB and PCF from PCTB
286
participants. As expected, HLA-DR expression on peripheral Mtb-specific CD4+ T cells was
287
significantly higher in the aTB groups (PTB and PCTB) compared to LTBI (medians:
288
62.30% and 70.85% vs 17.20%, respectively, p >0.0001). Moreover, HLA-DR expression on
289
Mtb-specific CD4+ T cells in PCF was significantly higher compared to blood in the PCTB
290
group (medians: 78.30% vs 69.90%, respectively, p= 0.0341) (Fig. 4a and b). We then
291
assessed the association of HLA-DR expression on Mtb-specific CD4 T cells and the
292
concentrations of each measured analyte at the site of disease (PCF) and in blood from PCTB
293
participants as well as blood from PTB participants (Fig. 4c). At disease site, we observed
294
positive associations between HLA-DR expression on Mtb-specific CD4 T cells and 10
295
analytes, including CCL1, G-CSF, OSM, IL-8, IL-2 and IL-2Rα (with r value > than 0.6).
296
Negative associations were observed with C4 (r = -0.71, p = 0.002) and IL-6Rα (r = -0.54, p
297
= 0.017) (Fig. 4d). None of these associations were observed in peripheral blood (Fig. 4c). In
298
PTB participants, HLA-DR expression on peripheral Mtb-specific CD4+ T cells associated
299
with only 2 analytes, namely IP-10 (r = 0.57, p = 0.0102) and IL-6Rα (r = -0.54, p = 0.0174)
300
(Fig. 4c). These data suggest a coordinated and compartmentalized immune response at the
301
disease site.
302
303
3.6 Impact of TB treatment on the inflammatory profile in plasma
304
Monitoring of TB treatment response is challenging mainly due to the lack of specific and
305
sensitive blood-based tools. In the current study, we examined the effect of TB treatment on
306
the expression of inflammation markers. First, we compared the overall systemic
307
inflammatory profile in participants with LTBI and in aTB patients (PTB and PCTB) 24
308
weeks after TB treatment initiation using unsupervised hierarchical clustering (Fig. 5a) and
309
principal component analysis (Fig. 5b). No specific clustering was observed between the
310
groups, showing a global normalization of the inflammation signature at treatment
311
completion. Furthermore, we performed univariate analysis comparing the level of
312
expression of each analyte at baseline (before TB treatment initiation), week 6 or 8 and week
313
24 post treatment initiation (Supplementary fig. 4 and Supplementary table 3). Of the 39
314
measured analytes, 13 showed significant reduction between baseline, week 6/8 and/or week
315
24 in both the PTB and PCTB groups (Supplementary fig. 4a and Supplementary table 3).
316
An additional eight analytes showed reduction between the three time points in the PTB
317
group only (Supplementary fig. 4b and Supplementary table 3).
318
Representative plots of analytes including, CXCL11, MIG, IL-6 and CRP depict the
319
significant reduction of expression of analytes with TB treatment from baseline, week 6/8 to
320
end of treatment (week 24) in both PTB and PCTB groups (Fig. 5c). These data suggest that
321
the overall inflammatory profile normalized upon TB treatment completion in both PTB and
322
PCTB.
323
324
3.7 Comparison of HLA-DR expression and biosignatures derived from soluble
325
analytes in discriminating LTBI from active TB
326
Previous studies have shown the potential of blood-based markers to distinguish LTBI from
327
aTB, including biosignatures derived from soluble markers and HLA-DR expression on
328
MTB-specific T cells [15,16,20,21,37,38]. Although this study was not designed as a
329
diagnostic study, we explored this aspect, wherein we assessed the ability of HLA-DR
330
expression to distinguish LTBI from PTB, PCTB or any aTB (PTB + PCTB) and compared it
331
with previously described biosignatures that included analytes measured in this study. We
332
generated receiver operating characteristic (ROC) curves from data obtained in Mtb-specific
333
CD4 T cells. Consistent with previous reports, HLA-DR expression on Mtb-specific CD4 T
334
cells showed a great capability to distinguish LTBI from PTB (p<0.0001, area-under-the-
335
curve (AUC) = 0.97, 95% CI: 0.92 – 1.00, sensitivity: 97.75%, specificity: 100%, at an
336
optimal cut-off of 48.5%) (Supplementary fig. 5a and b). Moreover, HLA-DR expression
337
also discriminated LTBI from PCTB (p<0.0001, AUC = 0.94, 95% CI: 0.82 – 1.00,
338
sensitivity: 93.75%, specificity: 100%, at an optimal cut-off of 46.9%) and LTBI from any
339
aTB (p<0.0001, AUC = 0.96, 95% CI: 0.90 – 1.00, sensitivity: 94.29%, specificity: 100%, at
340
an optimal cut-off of 46.9%) (Supplementary fig. 5a and b).
341
We assessed the performance of previously described soluble biosignatures our data set to
342
and compared soluble biosignature performance to HLA-DR expression. We identified six
343
different published biosignatures which include analytes measured in this study: [IL-12p40 +
344
IL-10] [21], [IFN-γ + IL-10 + IL-12p40] [21], [TNF-α + IL-12p40] [21], [CCL1 + CRP] [15],
345
[CCL1 + TNF-α] [16], and [IL-6Rα + IL-2Rα] [20].
346
These biosignatures discriminated LTBI from PTB with AUCs ranging from 0.72-0.9 and
347
corresponding sensitivity and specificity ranging from 55% - 85% and 75% - 100%,
348
respectively. They also discriminated LTBI from PCTB with AUCs ranging from 0.64 - 1.00
349
and corresponding sensitivity and specificity ranging from 61.11% - 83.33% and 62.5% -
350
93.75%, respectively, while they discriminated LTBI from any aTB (PTB + PCTB) with
351
AUCs ranging from 0.69 - 0.98 and corresponding sensitivity and specificity ranging from
352
52.63% - 76.32% and 62.50% - 100%, respectively (Supplementary table 4). Detailed
353
performances of these signatures in comparison to HLA-DR expression are shown in
354
supplementary table A.4.
355
None of these biosignatures out-performed HLA-DR expression in discriminating LTBI from
356
the diseased groups (Supplementary table 4). These findings suggest that HLA-DR is a
357
better biomarker than soluble markers for discriminating between the different TB groups.
358
4. Discussion
359
EPTB represents a small but significant proportion of all TB cases globally, particularly in
360
HIV-infected patients and is frequently difficult to diagnose. However, immune and
361
inflammatory responses at the site of disease remains understudied. In this study, we
362
compared the TB-associated inflammatory response in HIV-infected participants between
363
latent, pulmonary, and pericardial TB infection. We also compared the inflammatory
364
signature in blood and at site of disease (i.e., PCF) in PCTB patients. Moreover, we measured
365
HLA-DR expression on Mtb-specific CD4 T cells from whole blood and compared its
366
diagnostic potential to previously described biosignatures derived from different
367
combinations of soluble markers.
368
We show that PTB in HIV-infected patients is characterized by increased systemic
369
inflammation compared to LTBI persons. This is in accordance with previous reports
370
showing elevated inflammatory markers (such as CRP, IP-10, IFN-γ, CCL1, and VEGF) in
371
unstimulated plasma or serum in aTB compared to LTBI or other respiratory diseases
372
regardless of HIV status [15,16,18]. In HIV negative individuals, distinct inflammatory
373
profiles in PTB versus extra pulmonary TB have been reported, which were speculated to be
374
the consequence of differences between disseminated versus more localized infection [40].
375
However, here, we observed a similar inflammatory profile in HIV-infected PTB individuals
376
and HIV-infected PCTB individuals. These differences may be explained by the different
377
analytes measured in the Vinhaes et al [40] study and the current study, with only seven
378
analytes overlapping between the two studies (namely, IL-2, IL-6, IL-8, IL-10, IL-27, TNF-α,
379
and IFN-γ). Moreover, the Vinhaes et al [40] study included patients with different types of
380
EPTB (including Pleural TB, TB lymphadenitis and Miliary TB) while our study focused
381
exclusively on PCTB patients.
382
To improve our understanding of immunological mechanisms at the disease site, we
383
compared inflammatory profile at disease site and in plasma. A study by Matthews et al [22],
384
assessing the inflammatory response at the disease site, showed compartmentalization of
385
inflammatory proteins (including IL-6, IL-8 and IFN-γ) in PCF compared to blood. Our
386
results are in accordance with this study, showing that inflammation was greater at the site of
387
disease compared to the periphery and further demonstrate that there was a partial mirroring
388
of the innate-associated inflammatory response (such as CCL1, IL-12p40, TGF-β and IL-8)
389
between blood and disease site. Interestingly, Th1 cytokines levels (IFN-γ and IL-2) in PCF
390
did not correlate with plasma levels. We previously reported that there was no correlation
391
between the frequency of Mtb-specific CD4 T cells in blood and PCF [41] and recent data
392
from murine model suggests that the rate of migration of T cell to the disease site is mostly
393
regulated by the pattern of chemokine receptors they expressed [42].
394
TB diagnosis is challenging due to the lack of rapid, accurate, blood-based diagnostic tests.
395
HLA-DR expression on Mtb-specific CD4 T cells has been shown to be a robust marker in
396
discriminating latent TB from aTB [37–39] and EPTB [43]. In this study, we observed HLA-
397
DR to be significantly highly expressed in blood of aTB compared to LTBI, it was also
398
highly expressed, at the site of disease (PCF) in PCTB participants compared to blood of the
399
same participants. Our findings are in agreement with previously published studies [37–
400
39,43] and further suggest that the extent of activation of infiltrating CD4 T cells associate
401
with the inflammatory profile at the disease site.
402
Several biosignatures consisting of host soluble inflammatory markers have been described
403
as promising tools for TB diagnosis [15,16,20,21]. Here, we used our cohort as a validation
404
cohort to compare their performance in discriminating LTBI from aTB, and several
405
previously identified biosignatures continued to show promise in our cohort. However, none
406
of these biosignatures showed better performance compared to the measure of HLA-DR
407
expression on Mtb-specific CD4 T cells, which met the WHO target product profile (TPP)
408
recommendations for a point of care non-sputum-based triage test [44]. These data further
409
emphasize the role of HLA-DR as a promising biomarker for TB diagnosis.
410
Sputum culture conversion at two months post treatment initiation remains the most widely
411
used tool for the evaluation of TB treatment response [45,46]. However, in individuals with
412
PCTB who are sputum smear or culture negative for Mtb, monitoring of treatment response is
413
solely assessed clinically as there are no validated blood biomarkers to assist in this regard.
414
Changes in blood biomarker levels during antitubercular treatment in either PTB or EPTB
415
cases has been previously reported in a number of prospective studies [18,47–58], showing
416
the normalization of several inflammatory markers (such as CRP, IP-10, CCL1, IFN-γ and
417
TNF-α) after successful TB treatment. Our findings are in accordance with these results and
418
add to the current knowledge, showing that the concentrations of several of the biomarkers
419
tested (21 out of 39 and 13 out of 39) decreased at treatment completion to levels observed in
420
LTBI participants in both the PTB and PCTB groups, respectively. The discrepancy in the
421
normalization of inflammatory profile after treatment between PTB and PCTB could be
422
related to disease severity, where disseminated disease has been shown to present with
423
elevated systemic bacterial burden and higher mortality [59] and limited drug penetration at
424
the site of disease. Thus, our study confirms that measuring blood biomarkers may have
425
utility to monitor treatment response in both pulmonary and extra-pulmonary TB.
426
Our study has several limitations. First, most of the participants were HIV infected, we were
427
thus unable to define the impact of HIV infection on TB-induced inflammatory profiles.
428
Second, we did not have long-term follow-up clinical data to identify potential TB relapse, so
429
long-term outcome could not be related to inflammatory profiles. Third, the current study was
430
not designed to identify novel diagnostic markers, thus we confined our analysis to
431
previously described blood-based biomarkers. However, further assessments of HLA-DR
432
expression on Mtb-specific CD4 T cells are required in well-designed diagnostic studies.
433
Finally, further experiments including patients with non-tuberculous pericardial effusion will
434
be necessary to define whether the observed inflammatory signatures in plasma and at site of
435
disease are TB specific. Regardless of the limitations, our results show that in a largely HIV-
436
infected cohort with advanced immunosuppression, PCTB and PTB share similar
437
inflammatory signature and aTB disrupts the relationship between HIV VL and soluble
438
analytes. These results also reveal that profiles of markers at the site of disease are distinct
439
from peripheral blood though some markers strongly correlate. Furthermore, upon
440
completion of TB treatment, levels of soluble analytes normalized and lastly, we showed that
441
in HIV-infected patients, assessing the expression of HLA-DR on Mtb-specific CD4 T cells
442
had a better potential to discriminate PCTB and PTB from LTBI compared to biosignatures
443
derived from soluble markers.
444
Acknowledgments
445
The authors thank the study participants, the clinical staff at the Khayelitsha Site B
446
Community Health Centre in Cape Town and the laboratory staff at the Wellcome Centre for
447
Infectious Disease Research in Africa at the University of Cape Town.
448
449
Funding
450
This work was supported by the European and Developing Countries Clinical Trials
451
Partnership EDCTP2 programme; the European Union (EU)’s Horizon 2020 programme
452
(Training and Mobility Action TMA2017SF-1951-TB-SPEC to CR), the NIH (R21AI148027
453
to CR) and the South African Medical Research Council (MRC-UFSP-1-IMPI-2 to MN).
454
RJW is supported by the Francis Crick Institute, which receives funds from Cancer Research
455
UK(FC00110218), Wellcome (FC00110218) and the UK Medical Research Council
456
(FC00110218). RJW is also supported by Wellcome (203135), and NIH (U01/115940;
457
U01/152103). HM is supported by National Research Foundation of South Africa, (Grant
458
number: 129614), CIDRI-Africa Fellowship and in part by the Fogarty International Center
459
of the National Institutes of Health (D43TW010559). DLB, KDMB, and AS are supported by
460
the National Institute of Allergy and Infectious Diseases, National Institutes of Health,
461
Division of Intramural Research.
462
Competing interests
463
The authors declare that they have no competing interests associated with this publication.
464
REFERANCES
465
[1] Global HIV & AIDS statistics — Fact sheet, (n.d.).
466
https://www.unaids.org/en/resources/fact-sheet (accessed January 31, 2022).
467
[2] S.K. Sharma, A. Mohan, Extrapulmonary tuberculosis., Indian J. Med. Res. 120 (2004)
468
316–353.
469
[3] A.A. Cagatay, Y. Caliskan, S. Aksoz, L. Gulec, S. Kucukoglu, Y. Cagatay, H. Berk, H.
470
Ozsut, H. Eraksoy, S. Calangu, Extrapulmonary tuberculosis in immunocompetent
471
adults, Scand. J. Infect. Dis. 36 (2004) 799–806.
472
https://doi.org/10.1080/00365540410025339.
473
[4] R.W. Shafer, D.S. Kim, J.P. Weiss, J.M. Quale, Extrapulmonary tuberculosis in patients
474
with human immunodeficiency virus infection, Medicine (Baltimore). 70 (1991) 384–
475
397. https://doi.org/10.1097/00005792-199111000-00004.
476
[5] H.L. Rieder, D.E. Snider, G.M. Cauthen, Extrapulmonary tuberculosis in the United
477
States, Am. Rev. Respir. Dis. 141 (1990) 347–351.
478
https://doi.org/10.1164/ajrccm/141.2.347.
479
[6] J.J. Noubiap, V.N. Agbor, A.L. Ndoadoumgue, J.R. Nkeck, A. Kamguia, U.F. Nyaga, M.
480
Ntsekhe, Epidemiology of pericardial diseases in Africa: a systematic scoping review,
481
Heart. 105 (2019) 180–188. https://doi.org/10.1136/heartjnl-2018-313922.
482
[7] P. Howlett, E. Du Bruyn, H. Morrison, I.C. Godsent, K.A. Wilkinson, M. Ntsekhe, R.J.
483
Wilkinson, The immunopathogenesis of tuberculous pericarditis, Microbes Infect. 22
484
(2020) 172–181. https://doi.org/10.1016/j.micinf.2020.02.001.
485
[8] H. Reuter, L.J. Burgess, A.F. Doubell, Epidemiology of pericardial effusions at a large
486
academic hospital in South Africa, Epidemiol. Infect. 133 (2005) 393–399.
487
https://doi.org/10.1017/S0950268804003577.
488
[9] B.M. Mayosi, M. Ntsekhe, J. Bosch, S. Pandie, H. Jung, F. Gumedze, J. Pogue, L.
489
Thabane, M. Smieja, V. Francis, L. Joldersma, K.M. Thomas, B. Thomas, A.A.
490
Awotedu, N.P. Magula, D.P. Naidoo, A. Damasceno, A. Chitsa Banda, B. Brown, P.
491
Manga, B. Kirenga, C. Mondo, P. Mntla, J.M. Tsitsi, F. Peters, M.R. Essop, J.B.W.
492
Russell, J. Hakim, J. Matenga, A.F. Barasa, M.U. Sani, T. Olunuga, O. Ogah, V. Ansa,
493
A. Aje, S. Danbauchi, D. Ojji, S. Yusuf, Prednisolone and Mycobacterium indicus
494
pranii in Tuberculous Pericarditis, N. Engl. J. Med. 371 (2014) 1121–1130.
495
https://doi.org/10.1056/NEJMoa1407380.
496
[10] B.M. Mayosi, C.S. Wiysonge, M. Ntsekhe, F. Gumedze, J.A. Volmink, G. Maartens, A.
497
Aje, B.M. Thomas, K.M. Thomas, A.A. Awotedu, B. Thembela, P. Mntla, F. Maritz,
498
K.N. Blackett, D.C. Nkouonlack, V.C. Burch, K. Rebe, A. Parrish, K. Sliwa, B.Z. Vezi,
499
N. Alam, B.G. Brown, T. Gould, T. Visser, N.P. Magula, P.J. Commerford, Mortality in
500
patients treated for tuberculous pericarditis in sub-Saharan Africa : original article, S.
501
Afr. Med. J. 98 (2008) 36–40. https://doi.org/10.10520/EJC69118.
502
[11] C. Geldmacher, A. Zumla, M. Hoelscher, Interaction between HIV and Mycobacterium
503
tuberculosis: HIV-1-induced CD4 T-cell depletion and the development of active
504
tuberculosis, Curr. Opin. HIV Aids. 7 (2012) 268–275.
505
https://doi.org/10.1097/coh.0b013e3283524e32.
506
[12] C. Riou, N. Strickland, A.P. Soares, B. Corleis, D.S. Kwon, E.J. Wherry, R.J.
507
Wilkinson, W.A. Burgers, HIV Skews the Lineage-Defining Transcriptional Profile of
508
Mycobacterium tuberculosis–Specific CD4+ T Cells, J. Immunol. 196 (2016) 3006–
509
3018. https://doi.org/10.4049/jimmunol.1502094.
510
[13] J.M. Brenchley, M. Paiardini, K.S. Knox, A.I. Asher, B. Cervasi, T.E. Asher, P.
511
Scheinberg, D.A. Price, C.A. Hage, L.M. Kholi, A. Khoruts, I. Frank, J. Else, T.
512
Schacker, G. Silvestri, D.C. Douek, Differential Th17 CD4 T-cell depletion in
513
pathogenic and nonpathogenic lentiviral infections, Blood. 112 (2008) 2826–2835.
514
https://doi.org/10.1182/blood-2008-05-159301.
515
[14] S. Clark, E. Page, T. Ford, R. Metcalf, A. Pozniak, M. Nelson, D.C. Henderson, D.
516
Asboe, F. Gotch, B.G. Gazzard, P. Kelleher, Reduced T(H)1/T(H)17 CD4 T-cell
517
numbers are associated with impaired purified protein derivative-specific cytokine
518
responses in patients with HIV-1 infection, J. Allergy Clin. Immunol. 128 (2011) 838-
519
846.e5. https://doi.org/10.1016/j.jaci.2011.05.025.
520
[15] H. Mutavhatsindi, G.D. Van Der Spuy, S. Malherbe, J.S. Sutherland, A. Geluk, H.
521
Mayanja Kizza, A.C. Crampin, D. Kassa, R. Howe, A. Mihret, J.A. Sheehama, E.
522
Nepolo, G. Günther, H.M. Dockrell, P.L. Lam Corstjens, K. Stanley, G. Walzl, N.N.
523
Chegou, Validation and optimisation of host immunological bio-signatures for a point-
524
of-care test for TB disease, Front. Immunol. 12 (2021).
525
https://doi.org/10.3389/fimmu.2021.607827.
526
[16] B.H. Chendi, H. Tveiten, C.I. Snyders, K. Tonby, S. Jenum, S.D. Nielsen, M. Hove-
527
Skovsgaard, G. Walzl, N.N. Chegou, A.M. Dyrhol-Riise, CCL1 and IL-2Ra
528
differentiate Tuberculosis disease from latent infection Irrespective of HIV infection in
529
low TB burden countries, J. Infect. 83 (2021) 433–443.
530
https://doi.org/10.1016/j.jinf.2021.07.036.
531
[17] N.N. Chegou, J.S. Sutherland, S. Malherbe, A.C. Crampin, P.L.A.M. Corstjens, A.
532
Geluk, H. Mayanja-Kizza, A.G. Loxton, G. van der Spuy, K. Stanley, L.A. Kotzé, M.
533
van der Vyver, I. Rosenkrands, M. Kidd, P.D. van Helden, H.M. Dockrell, T.H.M.
534
Ottenhoff, S.H.E. Kaufmann, G. Walzl, Diagnostic performance of a seven-marker
535
serum protein biosignature for the diagnosis of active TB disease in African primary
536
healthcare clinic attendees with signs and symptoms suggestive of TB, Thorax. 71
537
(2016) 785–794. https://doi.org/10.1136/thoraxjnl-2015-207999.
538
[18] R. Jacobs, S. Malherbe, A.G. Loxton, K. Stanley, G. van der Spuy, G. Walzl, N.N.
539
Chegou, Identification of novel host biomarkers in plasma as candidates for the
540
immunodiagnosis of tuberculosis disease and monitoring of tuberculosis treatment
541
response, Oncotarget. 7 (2016). https://doi.org/10.18632/oncotarget.11420.
542
[19] Q.-Y. Liu, F. Han, L.-P. Pan, H.-Y. Jia, Q. Li, Z.-D. Zhang, Inflammation responses in
543
patients with pulmonary tuberculosis in an intensive care unit, Exp. Ther. Med. 15
544
(2018) 2719–2726. https://doi.org/10.3892/etm.2018.5775.
545
[20] O.A. Eribo, M.S. Leqheka, S.T. Malherbe, S. McAnda, K. Stanley, G.D. van der Spuy,
546
G. Walzl, N.N. Chegou, Host urine immunological biomarkers as potential candidates
547
for the diagnosis of tuberculosis, Int. J. Infect. Dis. 99 (2020) 473–481.
548
https://doi.org/10.1016/j.ijid.2020.08.019.
549
[21] J.S. Sutherland, B.C. de Jong, D.J. Jeffries, I.M. Adetifa, M.O.C. Ota, Production of
550
TNF-α, IL-12(p40) and IL-17 Can Discriminate between Active TB Disease and Latent
551
Infection in a West African Cohort, PLOS ONE. 5 (2010) e12365.
552
https://doi.org/10.1371/journal.pone.0012365.
553
[22] K. Matthews, A. Deffur, M. Ntsekhe, F. Syed, J.B.W. Russell, K. Tibazarwa, J. Wolske,
554
J. Brink, B.M. Mayosi, R.J. Wilkinson, K.A. Wilkinson, A Compartmentalized
555
Profibrotic Immune Response Characterizes Pericardial Tuberculosis, Irrespective of
556
HIV-1 Infection, Am. J. Respir. Crit. Care Med. 192 (2015) 1518–1521.
557
https://doi.org/10.1164/rccm.201504-0683LE.
558
[23] H. Reuter, L.J. Burgess, M.E. Carstens, A.F. Doubell, Characterization of the
559
immunological features of tuberculous pericardial effusions in HIV positive and HIV
560
negative patients in contrast with non-tuberculous effusions, Tuberculosis. 86 (2006)
561
125–133. https://doi.org/10.1016/j.tube.2005.08.018.
562
[24] Q. Yang, Y. Cai, W. Zhao, F. Wu, M. Zhang, K. Luo, Y. Zhang, H. Liu, B. Zhou, H.
563
Kornfeld, X. Chen, IP-10 and MIG Are Compartmentalized at the Site of Disease
564
during Pleural and Meningeal Tuberculosis and Are Decreased after Antituberculosis
565
Treatment, Clin. Vaccine Immunol. 21 (2014) 1635–1644.
566
https://doi.org/10.1128/CVI.00499-14.
567
[25] B.M. Mayosi, L.J. Burgess, A.F. Doubell, Tuberculous Pericarditis, Circulation. 112
568
(2005) 3608–3616. https://doi.org/10.1161/CIRCULATIONAHA.105.543066.
569
[26] W.A. Hanekom, J. Hughes, M. Mavinkurve, M. Mendillo, M. Watkins, H. Gamieldien,
570
S.J. Gelderbloem, M. Sidibana, N. Mansoor, V. Davids, R.A. Murray, A. Hawkridge,
571
P.A.J. Haslett, S. Ress, G.D. Hussey, G. Kaplan, Novel application of a whole blood
572
intracellular cytokine detection assay to quantitate specific T-cell frequency in field
573
studies, J. Immunol. Methods. 291 (2004) 185–195.
574
https://doi.org/10.1016/j.jim.2004.06.010.
575
[27] C.S.L. Arlehamn, D.M. McKinney, C. Carpenter, S. Paul, V. Rozot, E. Makgotlho, Y.
576
Gregg, M. van Rooyen, J.D. Ernst, M. Hatherill, W.A. Hanekom, B. Peters, T.J. Scriba,
577
A. Sette, A Quantitative Analysis of Complexity of Human Pathogen-Specific CD4 T
578
Cell Responses in Healthy M. tuberculosis Infected South Africans, PLOS Pathog. 12
579
(2016) e1005760. https://doi.org/10.1371/journal.ppat.1005760.
580
[28] R. Fluss, D. Faraggi, B. Reiser, Estimation of the Youden Index and its Associated
581
Cutoff Point, Biom. J. 47 (2005) 458–472. https://doi.org/10.1002/bimj.200410135.
582
[29] J.M. Bland, D.G. Altman, Multiple significance tests: the Bonferroni method, BMJ. 310
583
(1995) 170. https://doi.org/10.1136/bmj.310.6973.170.
584
[30] D. Wang, X. Tong, L. Wang, S. Zhang, J. Huang, L. Zhang, H. Fan, The association
585
between osteopontin and tuberculosis: A systematic review and meta-analysis, PLOS
586
ONE. 15 (2020) e0242702. https://doi.org/10.1371/journal.pone.0242702.
587
[31] D.M. Muema, N.A. Akilimali, O.C. Ndumnego, S.S. Rasehlo, R. Durgiah, D.B.A.
588
Ojwach, N. Ismail, M. Dong, A. Moodley, K.L. Dong, Z.M. Ndhlovu, J.M. Mabuka,
589
B.D. Walker, J.K. Mann, T. Ndung’u, Association between the cytokine storm, immune
590
cell dynamics, and viral replicative capacity in hyperacute HIV infection, BMC Med. 18
591
(2020) 81. https://doi.org/10.1186/s12916-020-01529-6.
592
[32] J.E. Teigler, L. Leyre, N. Chomont, B. Slike, N. Jian, M.A. Eller, N. Phanuphak, E.
593
Kroon, S. Pinyakorn, L.A. Eller, M.L. Robb, J. Ananworanich, N.L. Michael, H.
594
Streeck, S.J. Krebs, Distinct biomarker signatures in HIV acute infection associate with
595
viral dynamics and reservoir size, JCI Insight. 3 (n.d.) e98420.
596
https://doi.org/10.1172/jci.insight.98420.
597
[33] Á.H. Borges, J.L. O’Connor, A.N. Phillips, F.F. Rönsholt, S. Pett, M.J. Vjecha, M.A.
598
French, J.D. Lundgren, Factors Associated With Plasma IL-6 Levels During HIV
599
Infection, J. Infect. Dis. 212 (2015) 585–595. https://doi.org/10.1093/infdis/jiv123.
600
[34] L. ROBERTS, J.-A.S. PASSMORE, C. WILLIAMSON, F. LITTLE, L.M. BEBELL, K.
601
MLISANA, W.A. BURGERS, F. VAN LOGGERENBERG, G. WALZL, J.F. DJOBA
602
SIAWAYA, Q. ABDOOL KARIM, S.S. ABDOOL KARIM, Plasma cytokine levels
603
during acute HIV-1 infection predict HIV disease progression, AIDS Lond. Engl. 24
604
(2010) 819–831. https://doi.org/10.1097/QAD.0b013e3283367836.
605
[35] R. Bunjun, A.P. Soares, N. Thawer, T.L. Müller, A. Kiravu, Z. Ginbot, B. Corleis, B.D.
606
Murugan, D.S. Kwon, F. von Groote-Bidlingmaier, C. Riou, R.J. Wilkinson, G. Walzl,
607
W.A. Burgers, Dysregulation of the Immune Environment in the Airways During HIV
608
Infection, Front. Immunol. 12 (2021).
609
https://www.frontiersin.org/article/10.3389/fimmu.2021.707355 (accessed April 12,
610
2022).
611
[36] G.P. Rizzardi, W. Barcellini, G. Tambussi, F. Lillo, M. Malnati, L. Perrin, A. Lazzarin,
612
Plasma levels of soluble CD30, tumour necrosis factor (TNF)-alpha and TNF receptors
613
during primary HIV-1 infection: correlation with HIV-1 RNA and the clinical outcome,
614
AIDS Lond. Engl. 10 (1996) F45-50. https://doi.org/10.1097/00002030-199611000-
615
00001.
616
[37] T. Adekambi, C.C. Ibegbu, S. Cagle, A.S. Kalokhe, Y.F. Wang, Y. Hu, C.L. Day, S.M.
617
Ray, J. Rengarajan, Biomarkers on patient T cells diagnose active tuberculosis and
618
monitor treatment response, J. Clin. Invest. 125 (2015) 1827–1838.
619
https://doi.org/10.1172/JCI77990.
620
[38] C. Riou, N. Berkowitz, R. Goliath, W.A. Burgers, R.J. Wilkinson, Analysis of the
621
Phenotype of Mycobacterium tuberculosis-Specific CD4+ T Cells to Discriminate
622
Latent from Active Tuberculosis in HIV-Uninfected and HIV-Infected Individuals,
623
Front. Immunol. 8 (2017).
624
https://www.frontiersin.org/article/10.3389/fimmu.2017.00968 (accessed April 13,
625
2022).
626
[39] C. Riou, E. Du Bruyn, S. Ruzive, R.T. Goliath, C.S. Lindestam Arlehamn, A. Sette, A.
627
Sher, D.L. Barber, R.J. Wilkinson, Disease extent and anti‐tubercular treatment
628
response correlates with Mycobacterium tuberculosis‐specific CD4 T‐cell phenotype
629
regardless of HIV‐1 status, Clin. Transl. Immunol. 9 (2020) e1176.
630
https://doi.org/10.1002/cti2.1176.
631
[40] C.L. Vinhaes, D. Oliveira-de-Souza, P.S. Silveira-Mattos, B. Nogueira, R. Shi, W. Wei,
632
X. Yuan, G. Zhang, Y. Cai, C.E. Barry, L.E. Via, K.F. Fukutani, B.B. Andrade, K.D.
633
Mayer-Barber, Changes in Inflammatory Protein and Lipid Mediator Profiles Persist
634
After Antitubercular Treatment of Pulmonary and Extrapulmonary Tuberculosis: A
635
Prospective Cohort Study, Cytokine. 123 (2019) 154759.
636
https://doi.org/10.1016/j.cyto.2019.154759.
637
[41] E. Du Bruyn, S. Ruzive, P. Howlett, A.J. Jacobs, C.S. Lindestam Arlehamn, A. Sette, A.
638
Sher, K.D. Mayer-Barber, D.L. Barber, B. Mayosi, M. Ntsekhe, R.J. Wilkinson, C.
639
Riou, Profile of Mycobacterium tuberculosis-specific CD4 T cells at the site of disease
640
and blood in pericardial tuberculosis, BioRxiv. (2022) 2022.05.12.491749.
641
https://doi.org/10.1101/2022.05.12.491749.
642
[42] S.G. Hoft, M.A. Sallin, K.D. Kauffman, S. Sakai, V.V. Ganusov, D.L. Barber, The Rate
643
of CD4 T Cell Entry into the Lungs during Mycobacterium tuberculosis Infection Is
644
Determined by Partial and Opposing Effects of Multiple Chemokine Receptors, Infect.
645
Immun. 87 (2019) e00841-18. https://doi.org/10.1128/IAI.00841-18.
646
[43] P.S. Silveira-Mattos, B. Barreto-Duarte, B. Vasconcelos, K.F. Fukutani, C.L. Vinhaes,
647
D. Oliveira-De-Souza, C.C. Ibegbu, M.C. Figueiredo, T.R. Sterling, J. Rengarajan, B.B.
648
Andrade, Differential Expression of Activation Markers by Mycobacterium
649
tuberculosis-specific CD4+ T Cell Distinguishes Extrapulmonary From Pulmonary
650
Tuberculosis and Latent Infection, Clin. Infect. Dis. Off. Publ. Infect. Dis. Soc. Am. 71
651
(2020) 1905–1911. https://doi.org/10.1093/cid/ciz1070.
652
[44] World Health Organization, High priority target product profiles for new tuberculosis
653
diagnostics: report of a consensus meeting, 28-29 April 2014, Geneva, Switzerland,
654
World Health Organization, 2014. https://apps.who.int/iris/handle/10665/135617
655
(accessed June 3, 2022).
656
[45] D.A. Mitchison, Assessment of New Sterilizing Drugs for Treating Pulmonary
657
Tuberculosis by Culture at 2 Months, Am. Rev. Respir. Dis. 147 (1993) 1062–1063.
658
https://doi.org/10.1164/ajrccm/147.4.1062.
659
[46] R.S. Wallis, M. Pai, D. Menzies, T.M. Doherty, G. Walzl, M.D. Perkins, A. Zumla,
660
Biomarkers and diagnostics for tuberculosis: progress, needs, and translation into
661
practice, The Lancet. 375 (2010) 1920–1937. https://doi.org/10.1016/S0140-
662
6736(10)60359-5.
663
[47] C. Riou, B.P. Peixoto, L. Roberts, K. Ronacher, G. Walzl, C. Manca, R. Rustomjee, T.
664
Mthiyane, D. Fallows, C.M. Gray, G. Kaplan, Effect of Standard Tuberculosis
665
Treatment on Plasma Cytokine Levels in Patients with Active Pulmonary Tuberculosis,
666
PLOS ONE. 7 (2012) e36886. https://doi.org/10.1371/journal.pone.0036886.
667
[48] J.Y. Hong, H.J. Lee, S.Y. Kim, K.S. Chung, E.Y. Kim, J.Y. Jung, M.S. Park, Y.S. Kim,
668
S.K. Kim, J. Chang, S.-N. Cho, Y.A. Kang, Efficacy of IP-10 as a biomarker for
669
monitoring tuberculosis treatment, J. Infect. 68 (2014) 252–258.
670
https://doi.org/10.1016/j.jinf.2013.09.033.
671
[49] C. Martins, A.C. de C. Gama, D. Valcarenghi, A.P. de B. Batschauer, C. Martins, A.C.
672
de C. Gama, D. Valcarenghi, A.P. de B. Batschauer, Markers of acute-phase response in
673
the treatment of pulmonary tuberculosis, J. Bras. Patol. E Med. Lab. 50 (2014) 428–433.
674
https://doi.org/10.5935/1676-2444.20140052.
675
[50] B.H. Chendi, C.I. Snyders, K. Tonby, S. Jenum, M. Kidd, G. Walzl, N.N. Chegou, A.M.
676
Dyrhol-Riise, A Plasma 5-Marker Host Biosignature Identifies Tuberculosis in High
677
and Low Endemic Countries, Front. Immunol. 12 (2021).
678
https://www.frontiersin.org/article/10.3389/fimmu.2021.608846 (accessed June 3,
679
2022).
680
[51] L. Liang, R. Shi, X. Liu, X. Yuan, S. Zheng, G. Zhang, W. Wang, J. Wang, K. England,
681
L.E. Via, Y. Cai, L.C. Goldfeder, L.E. Dodd, C.E. Barry, R.Y. Chen, Interferon-gamma
682
response to the treatment of active pulmonary and extra-pulmonary tuberculosis, Int. J.
683
Tuberc. Lung Dis. 21 (2017) 1145–1149. https://doi.org/10.5588/ijtld.16.0880.
684
[52] N. Rockwood, D.L. Costa, E.P. Amaral, E. Du Bruyn, A. Kubler, L. Gil-Santana, K.F.
685
Fukutani, C.A. Scanga, J.L. Flynn, S.H. Jackson, K.A. Wilkinson, W.R. Bishai, A. Sher,
686
R.J. Wilkinson, B.B. Andrade, Mycobacterium tuberculosis induction of heme
687
oxygenase-1 expression is dependent on oxidative stress and reflects treatment
688
outcomes, Front. Immunol. 8 (2017). https://doi.org/10.3389/fimmu.2017.00542.
689
[53] R.S. Wallis, P. Kim, S. Cole, D. Hanna, B.B. Andrade, M. Maeurer, M. Schito, A.
690
Zumla, Tuberculosis biomarkers discovery: developments, needs, and challenges,
691
Lancet Infect. Dis. 13 (2013) 362–372. https://doi.org/10.1016/S1473-3099(13)70034-3.
692
[54] X. Bai, H. Li, Y. Yang, J. Zhang, Y. Liang, X. Wu, Cytokine and soluble adhesion
693
molecule profiles and biomarkers for treatment monitoring in Re-treated smear-positive
694
patients with pulmonary tuberculosis, Cytokine. 108 (2018) 9–16.
695
https://doi.org/10.1016/j.cyto.2018.03.009.
696
[55] P. Miranda, L. Gil-Santana, M.G. Oliveira, E.D.D. Mesquita, E. Silva, A. Rauwerdink,
697
F. Cobelens, M.M. Oliveira, B.B. Andrade, A. Kritski, Sustained elevated levels of C-
698
reactive protein and ferritin in pulmonary tuberculosis patients remaining culture
699
positive upon treatment initiation, PLOS ONE. 12 (2017) e0175278.
700
https://doi.org/10.1371/journal.pone.0175278.
701
[56] G.B. Sigal, M.R. Segal, A. Mathew, L. Jarlsberg, M. Wang, S. Barbero, N. Small, K.
702
Haynesworth, J.L. Davis, M. Weiner, W.C. Whitworth, J. Jacobs, J. Schorey, D.M.
703
Lewinsohn, P. Nahid, Biomarkers of Tuberculosis Severity and Treatment Effect: A
704
Directed Screen of 70 Host Markers in a Randomized Clinical Trial, EBioMedicine. 25
705
(2017) 112–121. https://doi.org/10.1016/j.ebiom.2017.10.018.
706
[57] A. Singanayagam, K. Manalan, D.W. Connell, J.D. Chalmers, S. Sridhar, A.I. Ritchie,
707
A. Lalvani, M. Wickremasinghe, O.M. Kon, Evaluation of serum inflammatory
708
biomarkers as predictors of treatment outcome in pulmonary tuberculosis, Int. J. Tuberc.
709
Lung Dis. 20 (2016) 1653–1660. https://doi.org/10.5588/ijtld.16.0159.
710
[58] J. Nouhin, P. Pean, Y. Madec, M.F. Chevalier, C. Didier, L. Borand, F.-X. Blanc, D.
711
Scott-Algara, D. Laureillard, L. Weiss, Interleukin-1 receptor antagonist, a biomarker of
712
response to anti-TB treatment in HIV/TB co-infected patients, J. Infect. 74 (2017) 456–
713
465. https://doi.org/10.1016/j.jinf.2017.01.016.
714
[59] C. Schutz, D. Barr, B.B. Andrade, M. Shey, A. Ward, S. Janssen, R. Burton, K.A.
715
Wilkinson, B. Sossen, K.F. Fukutani, M. Nicol, G. Maartens, R.J. Wilkinson, G.
716
Meintjes, Clinical, microbiologic, and immunologic determinants of mortality in
717
hospitalized patients with HIV-associated tuberculosis: A prospective cohort study,
718
PLoS Med. 16 (2019) e1002840. https://doi.org/10.1371/journal.pmed.1002840.
719
720
Table 1. Clinical characteristics of study participants.
721
722
PCTB
PTB
LTBI
N
18
20
16
Age (years) †
36 [29 – 44]
39 [32 – 45]
37 [32 – 41]
Gender (F/M)
8/10
8/12
16/0
HIV status (Neg/Pos)
3/15
0/20
0/16
CD4 count (cells/mm3) † 141 [61 – 195.3]
176 [107 – 246]
409 [264 – 524]
Log10 VL (mRNA
copies/mL) †
4.68 [2.903 – 5.278]
4.79 [4.23 – 5.11]
3.28 [1.44 – 4.18]
Mtb Culture positive (n,
%)
9/16 (56.2%) in PCF‡
19 (95%) in sputum
0 (0%) in sputum
723
LTBI = Latent TB infection, PCTB = Pericardial TB, PTB = Pulmonary TB, F = Female, M
724
= Male, VL = HIV viral load, NA = not applicable
725
†Median and interquartile range.
726
‡Mtb culture data were not available for two PCTB patients.
727
728
Figure legends:
729
Figure 1. Analyte profiles in the different TB groups at baseline. (a) A non-supervised
730
two-way hierarchical cluster analysis (HCA, Ward method) was employed to evaluate the TB
731
groups using the 39 measured analytes. TB status (PCTB in red, PTB in blue and LTBI in
732
green) of each patient is indicated at the top of the dendrogram. Data are depicted as a
733
heatmap colored from minimum to maximum normalized values for each marker. (b)
734
Principal component analysis (PCA) on correlations based on the 39 analytes was used to
735
explain the variance of the data distribution in the cohort. Each dot represents a participant.
736
The two axes represent principal components 1 (PC1) and 2 (PC2). Their contribution to the
737
total data variance is shown as a percentage. (c) Analyte network analysis (Fruchterman-
738
Reingold algorithm) in plasma of LTBI, PTB and PCTB participants. Size of nodes indicate
739
the number of connections. Size of edges indicate the spearman r value (only r > 0.6 were
740
included). Blue lines: positive correlation. Red lines: negative correlation.
741
742
Figure 2. Univariate associations between HIV VL and analyte concentrations in the
743
different TB groups. (a) Spearman’s rank values of the univariate correlation between each
744
analyte and the HIV VL in LTBI participants, PTB participants, and PCTB participants
745
plasma samples. Red bars indicate positive correlations, Black bars indicate negative
746
correlations, and grey bars indicate non-significant correlations. (b) Depicts the examples of
747
IL-12p40 (maintained relationship between the TB groups) and IP-10 (disrupted relationship
748
between the TB groups). The line indicates linear regression for statistically significant
749
correlations.
750
Figure 3. Analyte profiles in peripheral blood (Plasma) and site of disease (Pericardial
751
fluid) in PCTB participants. (a) A non-supervised two-way hierarchical cluster analysis
752
(HCA, Ward method) was employed to evaluate the two sites using the 39 analytes. The
753
sample type and Mtb culture results (PCF in purple, Plasma in red; Mtb culture negative in
754
white and positive in black) of each patient is indicated at the top of the dendrogram. Data are
755
depicted as a heatmap colored from minimum to maximum normalized values detected for
756
each marker. (b) Principal component analysis (PCA) on correlations based on the 39
757
analytes was used to explain the variance of the data distribution in the subgroup. Each dot
758
represents a participant. The two axes represent principal components 1 (PC1) and 2 (PC2).
759
Their contribution to the total data variance is shown as a percentage. (c) Constellation Plot-
760
cluster analysis based on all measured analytes. Each dot represents a participant and is color-
761
coded according to sample type. Each cluster obtained for the HCA is identified by a number.
762
(d) Pairwise correlation of the 39 analytes. Red bars indicate a positive correlation, Black
763
bars indicate a negative correlation, and grey bars indicate a non-significant correlation. (e)
764
Analyte network analysis in PCF of PCTB participants. Size of nodes indicate the number of
765
connections. Size of edges indicate the spearman r (only r > 0.6 were included). Blue lines:
766
positive correlation. Red lines: negative correlation.
767
768
Figure 4. Univariate associations between HLA-DR and analyte concentrations in the
769
different TB groups. (a) Representative flow cytometry plots of the expression of HLA-DR.
770
(b) Expression of HLA-DR on Mtb-specific CD4 T cells in response to Mtb300. (c)
771
Spearman’s rank values of the univariate correlation between each analyte and between Mtb-
772
specific CD4 T cell activation (HLA-DR) level at the site of disease (PCF) in PCTB
773
participants, in blood of PCTB and PTB participants, respectively. Red bars indicate a
774
positive correlation, Black bars indicate a negative correlation, and the grey bars indicate
775
non-significant correlation. (d) Representative graphs showing the positive (CCL1 and G-
776
CSF) and negative (C4) correlation to HLA-DR frequency at the site of disease (PCF).
777
Statistical comparisons were performed using a Kruskal-Wallis test, adjusted for multiple
778
comparisons (Dunn’s test) for blood LTBI vs PTB vs PCTB, Wilcoxon test for blood PCTB
779
vs PCF PCTB and the Mann-Whitney test to compare blood LTBI and PCF PCTB.
780
Figure 5. Analyte profiles in the different TB groups before, during and post TB
781
treatment. (a) A non-supervised two-way hierarchical cluster analysis (HCA, Ward method)
782
was employed to grade the TB groups using the 39 analytes. TB status (PCTB in red, PTB in
783
blue and LTBI in green) of each patient is indicated at the top of the dendrogram. Data are
784
depicted as a heatmap colored from minimum to maximum normalized values detected for
785
each marker. (b) Principal component analysis (PCA) on correlations based on the 39
786
analytes was used to explain the variance of the data distribution in the cohort. Each dot
787
represents a participant. The two axes represent principal components 1 (PC1) and 2 (PC2).
788
Their contribution to the total data variance is shown as a percentage. (c) Representative
789
graphs showing the change of concentrations of CXCL11, MIG, IL-6 and CRP with
790
treatment and no statistical difference between week 24 post-treatment initiation and LTBI in
791
both PTB and PCTB groups, respectively. Statistical comparisons were performed using a
792
Friedman test, adjusted for multiple comparisons (Dunn’s test) for BL v W6/W8, BL v W24
793
and W6/W8 v W24 and the Mann-Whitney test to compare LTBI with W24, p-values were
794
adjusted using the Bonferroni method.
795
796
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
Group
Apo-AI
Apo-CIII
CC3
CC4
CRP
SAP
TM
IL-22
Galectin-3
PDGF-BB
MCP-2
ICAM-1
NCAM-1
IL-33R
IP-10
IL-27
OPN
CCL1
CD30
IL-2R!
MIG
G-CSF
IFN-훾
IL-6
IL-10
IL-6R!
OSM
VEGF
MIP-1β
CD163
GCP-2
CXCL11
Granzyme B
IL-2
IL-8
IL-12 (p40)
M-CSF
TNF-!
TGF-β
Fig. 1
(b)
(a)
Groups (plasma)
PCTB
PTB
LTBI
0
1
PC2 (12.2%)
PC1 (26.6%)
Groups
(c)
ApoA-I
Apo-CIII
OPN
IL-33R
IL-6
TM
G-CSF
M-CSF
IL-10
IL-6Rα
CD163
IL-22
OSM
GCP-2
CD30
MIP-1β
IL-12p40
TNF-α
IL-2
IFN-γ
GrB
C3
MCP-2
C4
SAP
PDGF-BB
TGF-β
ICAM-1
VEGF
NCAM-1
CRP
Gal-3
IP-10
MIG
CCL1
IL-2Rα
IL-27
CXCL11
IL-8
Normalized analyte values
MCP-2
PCTB
PTB
LTBI
797
1
2
3
4
5
6
7
0
200
400
600
1
2
3
4
5
6
7
0
200
400
600
1
2
3
4
5
6
7
0
200
400
600
1
2
3
4
5
6
7
0
200
400
600
1
2
3
4
5
6
7
0
200
400
600
1
2
3
4
5
6
7
0
200
400
600
(a)
(b)
p = 0.0002
r = 0.83
p = 0.012
r = 0.63
p = 0.028
r = 0.49
Spearman’s rank (r) values
PCTB
PTB
LTBI
IL-12p40 (pg/ml)
Fig. 2
PCTB
PTB
LTBI
HIV viral load (log10 RNA copies/ml)
IP-10 (pg/ml)
p = 0.0002
r = 0.82
p = 0.37
r = 0.25
p = 0.26
r = 0.26
-0.5
0.0
0.5
1.0
SAP
OPN
CRP
TGF-β
NCAM-1
VEGF
Gal-3
PDGF-BB
C4
IL-33R
MCP-2
TM
C3
IL-22
Apo-CIII
ICAM-1
CCL1
IL-8
GCP-2
IL-2
IL-6Rα
MIG
G-CSF
OSM
ApoA-I
IL-27
CD163
CXCL11
IL-6
IFN-γ
GrB
IP-10
IL-2Rα
IL-12p40
TNF-α
MIP-1β
CD30
Spearman’s rank values
HIV VL vs LTBI
-0.5
0.0
0.5
1.0
SAP
OPN
CRP
TGF-β
NCAM-1
VEGF
Gal-3
PDGF-BB
C4
IL-33R
MCP-2
TM
C3
IL-22
Apo-CIII
ICAM-1
CCL1
IL-8
GCP-2
IL-2
IL-6Rα
MIG
G-CSF
OSM
ApoA-I
IL-27
CD163
CXCL11
IL-6
IFN-γ
GrB
IP-10
IL-2Rα
IL-12p40
TNF-α
MIP-1β
CD30
Spearman’s rank values
HIV VL vs PCTB
-0.5
0.0
0.5
1.0
SAP
OPN
CRP
TGF-β
NCAM-1
VEGF
Gal-3
PDGF-BB
C4
IL-33R
MCP-2
TM
C3
IL-22
Apo-CIII
ICAM-1
CCL1
IL-8
GCP-2
IL-2
IL-6Rα
MIG
G-CSF
OSM
ApoA-I
IL-27
CD163
CXCL11
IL-6
IFN-γ
GrB
IP-10
IL-2Rα
IL-12p40
TNF-α
MIP-1β
CD30
Spearman’s rank values
HIV VL vs PTB
<0.0001
<0.0001
0.0001
0.0002
0.0002
0.0002
0.0012
0.0027
0.0056
0.0031
0.011
0.024
0.0053
0.0072
0.0285
0.0058
0.011
0.045
0.012
0.013
798
ICAM-1
SAP
Apo-AI
CCL1
IL-22
CRP
NCAM-1
IL-2Rα
MIG
IL-12 (p40)
CC4
TGF-β
CD163
IL-6Rα
IL-8
TM
Apo-CIII
IP-10
OPN
IL-27
M-CSF
IL-10
TNF-α
VEGF
GCP-2
IL-6
CXCL11
IL-2
CD30
IL-33R
Galectin-3
OSM
MIP-1 beta
MCP-2
G-CSF
IFN-gamma
CC3
PDGF-BB
Granzyme B
0.0
0.5
1.0
(a)
(d)
(e)
Mtb culture
Sample Type
-50
0
50
Y
-50
0
50
X
1a
1b
2b
2a
n=12
n=7
33%
50%
17%
72%
14%
14%
Negative
Positive
No data
Plasma
PCF
(b)
Cult-
Cult+
No data
Fig. 3
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
PCF
Plasma
Sample Type
Mtb culture
Apo-AI
Apo-CIII
CC3
CC4
CRP
SAP
Thrombomodulin
IL-22
Galectin-3
PDGF-BB
MCP-2
ICAM-1
NCAM-1
IL-33R
IP-10
IL-27
OPN
I-309
CD30
IL-2R alpha
MIG
G-CSF
IFN-gamma
IL-6
IL-10
IL-6R alpha
OSM
VEGF
MIP-1 beta
CD163
GCP-2
I-TAC
Granzyme B
IL-2
IL-8
IL-12
M-CSF
TNF-alpha
TGF-beta
PCF vs Plasma
<0.0001
<0.0001
<0.0001
0.0001
0.0005
0.0005
0.0016
0.004
0.006
0.0081
0.0083
0.017
0.018
0.018
0.034
0.044
0.045
0.048
Spearman’s rank (r) values
-5
0
5
PC2 (11.2%)
-5
0
5
PC1 (42%)
PC1 (42%)
PC2 (11.2%)
(c)
1a
1b
2a
2b
1
2
ApoA-I
Apo-CIII
TM
GCP-2
Gal-3
CCL1
G-CSF
MIG
MCP-2
OSM
IL-8
IL-2
TNF-α
MIP-1β
IL-12p40
CD30
IL-2Rα
IP-10
IL-6
VEGF
IL-10
CD163
ICAM-1
IL-27
IFN-γ
CXCL11
GrB
OPN
M-CSF
TGF-β
C3
PDGF-BB
IL-6Rα
C4
SAP
CRP
IL-22
NCAM-1
IL-33R
0
1
Normalized analyte values
PCF
-0.5
0.0
0.5
1.0
GrB
PDGF-BB
C3
IFN-γ
G-CSF
MCP-2
MIP-1β
OSM
IL-33R
Gal-3
CD30
IL-2
CXCL11
IL-6
VEGF
GCP-2
TNF-α
IL-10
M-CSF
IL-27
OPN
IP-10
Apo-CIII
TM
IL-8
CD163
IL-6Rα
TGF-β
C4
IL-12p40
MIG
IL-2Rα
NCAM-1
CRP
IL-22
CCL1
ApoA-I
SAP
ICAM-1
Spearman’s rank values
Plasma vs PCF
799
0
20 40
0
100000
200000
300000
400000
40
60
80
100
0
20 40
0
500
1000
1500
2000
2500
40
60
80
100
0
20 40 40
60
80
100
0
1000
2000
3000
4000
0.002
0.017
(a)
Fig. 4
HLADR (% all CK) / LTBI
HLADR (% all CK) / PTB
HLADR (% all CK) / PCTB
HLADR (% all CK) / PCF
0
20
40
60
80
100
HLA-DR
IFN-g
% HLA-DR in Mtb-specific CD4 cells
73%
80%
8.1%
57%
LTBI
PTB
PCTB
PCTB
Blood
PCF
>0.0001
>0.0001
>0.0001
(b)
(c)
Spearman’s rank (r) values
PTB / Blood
PCTB / Blood
PCTB / PCF
CCL1
G-CSF
C4
pg/mL
ng/mL
p=0.002, r=0.71
p=0.0022, r=0.70
(d)
p=0.002, r=-0.71
% HLA-DR on Mtb-sp CD4 T cells
% HLA-DR on Mtb-sp CD4 T cells
% HLA-DR on Mtb-sp CD4 T cells
mg/L
0.0102
0.0174
0.0341
LTBI
PTB
PCTB / Blood
PCTB / PCF
-0.5
0.0
0.5
C4
IL-6Rα
IL-33R
CRP
GCP-2
C3
Apo-CIII
SAP
ApoA-I
IL-6
IL-22
NCAM-1
IL-27
VEGF
ICAM-1
TM
CD30
CXCL11
OPN
IP-10
CD163
Gal-3
PDGF-BB
TNF-α
TGF-β
GrB
MCP-2
IFN-γ
MIP-1β
IL-12p40
MIG
IL-2Rα
IL-2
IL-8
OSM
G-CSF
CCL1
Spearman’s rank values
HLA-DR (% MTB-spec CD4 cells) vs PCTB
-0.5
0.0
0.5
C4
IL-6Rα
IL-33R
CRP
GCP-2
C3
Apo-CIII
SAP
ApoA-I
IL-6
IL-22
NCAM-1
IL-27
VEGF
ICAM-1
TM
CD30
CXCL11
OPN
IP-10
CD163
Gal-3
PDGF-BB
TNF-α
TGF-β
GrB
MCP-2
IFN-γ
MIP-1β
IL-12p40
MIG
IL-2Rα
IL-2
IL-8
OSM
G-CSF
CCL1
Spearman’s rank values
HLA-DR (% MTB-spec CD4 cells) vs PTB
-0.5
0.0
0.5
C4
IL-6Rα
IL-33R
CRP
GCP-2
C3
Apo-CIII
SAP
ApoA-I
IL-6
IL-22
NCAM-1
IL-27
VEGF
ICAM-1
TM
CD30
CXCL11
OPN
IP-10
CD163
Gal-3
PDGF-BB
TNF-α
TGF-β
GrB
MCP-2
IFN-γ
MIP-1β
IL-12p40
MIG
IL-2Rα
IL-2
IL-8
OSM
G-CSF
CCL1
Spearman’s rank values
HLA-DR (% MTB-spec CD4 cells) vs PCF
0.002
0.0022
0.0034
0.0069
0.0074
0.0077
0.0173
0.0252
0.031
0.032
0.043
800
BL
W8
W24
LTBI
0
1
2
3
4
BL
W6
W24
LTBI
0
500
1000
BL
W8
W24
LTBI
0
500
1000
1500
(a)
(c)
(b)
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
LTBI
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
PCTB
Group
Apo-AI
Apo-CIII
CC3
CC4
CRP
SAP
TM
IL-22
Galectin-3
PDGF-BB
MCP-2
ICAM-1
NCAM-1
IL-33R
IP-10
IL-27
OPN
CCL1
CD30
IL-2R!
MIG
G-CSF
IFN-훾
IL-6
IL-10
IL-6R!
OSM
VEGF
MIP-1β
CD163
GCP-2
CXCL11
Granzyme B
IL-2
IL-8
IL-12 (p40)
M-CSF
TNF-!
TGF-β
Fig. 5
Groups (plasma)
PCTB (W24)
PTB (W24)
LTBI (BL)
CRP (ml/L)
CXCL11 (pg/mL)
IL-6 (pg/mL)
MIG (ng/mL)
<0.0001
0.0027
BL
W6
W24
LTBI
0
500
1000
1500
0.0052
0.011
BL
W8
W24
LTBI
0
500
1000
1500
2000
<0.0001
0.0047
0.0024
0.022
BL
W8
W24
LTBI
0
10
20
30
40
400
600
800
<0.0001
0.006
BL
W6
W24
LTBI
0
20
40
60
80
0.011
0.005
<0.0001
0.0047
0.043
BL
W6
W24
LTBI
0
1
2
3
4
0.0003
0.016
PCTB
PTB
-5
0
5
PC2 (14 %)
-5
0
5
PC1 (22 %)
PC2 (14%)
PC1 (22%)
ApoA-I
TM
GCP-2
IL-10
M-CSF
IL-22
OSM
ICAM-1
IL-27
IP-10
CCL1
IL-2Rα
MIG
CD30
MIP-1β
IL-12p40
TNF-α
IFN-γ
GrB
Apo-CIII
C4
SAP
Gal-3
C3
MCP-2
CRP
IL-6
PDGF-BB
TGF-β
VEGF
CXCL11
IL-8
IL-2
NCAM-1
OPN
IL-33R
G-CSF
IL-6Rα
CD163
0
1
Normalized analyte values
| 2022 | Blood and site of disease inflammatory profiles differ in HIV-1-infected pericardial tuberculosis patients | 10.1101/2022.10.21.513232 | [
"Mutavhatsindi Hygon",
"Du Bruyn Elsa",
"Ruzive Sheena",
"Howlett Patrick",
"Sher Alan",
"Mayer-Barber Katrin D.",
"Barber Daniel L.",
"Ntsekhe Mpiko",
"Wilkinson Robert J.",
"Riou Catherine"
] | creative-commons |
Linguistic Analysis of the bioRxiv Preprint
Landscape
This manuscript (permalink) was automatically generated from greenelab/annorxiver_manuscript@2034e45 on May 12,
2021.
Authors
David N. Nicholson
0000-0003-0002-5761 ·
danich1 ·
dnicholson329
Department of Systems Pharmacology and Translational Therapeutics, Perelman School of Medicine University of
Pennsylvania, Philadelphia PA, USA · Funded by The Gordon and Betty Moore Foundation (GBMF4552); The National
Institutes of Health (T32 HG000046)
Vincent Rubinetti
·
vincerubinetti ·
vincerubinetti
Department of Systems Pharmacology and Translational Therapeutics, Perelman School of Medicine University of
Pennsylvania, Philadelphia PA, USA; Center for Health AI, University of Colorado School of Medicine, Aurora, CO, USA ·
Funded by The Gordon and Betty Moore Foundation (GBMF4552); The National Institutes of Health (R01 HG010067)
Dongbo Hu
·
dongbohu
Department of Systems Pharmacology and Translational Therapeutics, Perelman School of Medicine University of
Pennsylvania, Philadelphia PA, USA · Funded by The Gordon and Betty Moore Foundation (GBMF4552); The National
Institutes of Health (R01 HG010067)
Marvin Thielk
0000-0002-0751-3664 ·
MarvinT ·
TheNeuralCoder
Elsevier, Philadelphia PA, USA
Lawrence E. Hunter
0000-0003-1455-3370 ·
LEHunter ·
ProfLHunter
Center for Computational Pharmacology, University of Colorado School of Medicine, Aurora CO, USA · Funded by The
Gordon and Betty Moore Foundation (GBMF4552)
Casey S. Greene
0000-0001-8713-9213 ·
cgreene ·
greenescientist
Department of Systems Pharmacology and Translational Therapeutics, Perelman School of Medicine University of
Pennsylvania, Philadelphia PA, USA; Department of Biochemistry and Molecular Genetics, University of Colorado
School of Medicine, Aurora CO, USA; Center for Health AI, University of Colorado School of Medicine, Aurora, CO, USA ·
Funded by The Gordon and Betty Moore Foundation (GBMF4552); The National Institutes of Health (R01 HG010067)
Abstract
Preprints allow researchers to make their �ndings available to the scienti�c community before they
have undergone peer review. Studies on preprints within bioRxiv have been largely focused on article
metadata and how often these preprints are downloaded, cited, published, and discussed online. A
missing element that has yet to be examined is the language contained within the bioRxiv preprint
repository. We sought to compare and contrast linguistic features within bioRxiv preprints to
published biomedical text as a whole as this is an excellent opportunity to examine how peer review
changes these documents. The most prevalent features that changed appear to be associated with
typesetting and mentions of supplementary sections or additional �les. In addition to text
comparison, we created document embeddings derived from a preprint-trained word2vec model. We
found that these embeddings are able to parse out di�erent scienti�c approaches and concepts, link
unannotated preprint-peer reviewed article pairs, and identify journals that publish linguistically
similar papers to a given preprint. We also used these embeddings to examine factors associated with
the time elapsed between the posting of a �rst preprint and the appearance of a peer reviewed
publication. We found that preprints with more versions posted and more textual changes took
longer to publish. Lastly, we constructed a web application (https://greenelab.github.io/preprint-
similarity-search/) that allows users to identify which journals and articles that are most linguistically
similar to a bioRxiv or medRxiv preprint as well as observe where the preprint would be positioned
within a published article landscape.
Introduction
The dissemination of research �ndings is key to science. Initially, much of this communication
happened orally [1]. During the 17th century, the predominant form of communication shifted to
personal letters that were shared from one scientist to another [1]. Scienti�c journals didn’t become a
predominant mode of communication until the 19th and 20th centuries, when the �rst journal was
created [1,2,3]. Although scienti�c journals became the primary method of communication, they
added high maintenance costs and long publication times to scienti�c discourse [2,3]. Some
scientists’ solutions to these issues has been to communicate through preprints, which are scholarly
works that have yet to undergo peer review process [4,5].
Preprints are commonly hosted on online repositories, where users have open and easy access to
these works. Notable repositories include arXiv [6], bioRxiv [7] and medRxiv [8]; however, there are
over 60 di�erent repositories available [9]. The burgeoning uptake of preprints in life sciences has
been examined through research focused on metadata from the bioRxiv repository. For example, life
science preprints are being posted at an increasing rate [10]. Furthermore, these preprints are being
rapidly shared on social media, routinely downloaded, and cited [11]. Some preprint categories are
shared on social media by both scientists and non-scientists [12]. About two-thirds to three-quarters
of preprints are eventually published [13,14] and life science articles that have a corresponding
preprint version are cited and discussed more often than articles without them [15,16,17]. Preprints
take an average of 160 days to be published in the peer-reviewed literature [18], and those with
multiple versions take longer to publish[18].
The rapid uptake of preprints in the life sciences also poses challenges. Preprint repositories receive a
growing number of submissions [19]. Linking preprints with their published counterparts is vital to
maintaining scholarly discourse consistency but is challenging to perform manually [16,20,21]. Errors
and omissions in linkage result in missing links and consequently erroneous metadata. Furthermore,
repositories based on standard publishing tools are not designed to show how textual content of
preprints is altered due to the peer review process [19]. Certain scientists have expressed concern
that competitors could scoop them by making results available before publication [19,22]. Preprint
repositories by de�nition do not perform in-depth peer review, which can result in posted preprints
containing inconsistent results or conclusions [17,20,23,24]; however, an analysis of preprints posted
at the beginning of 2020 revealed that most underwent minor changes as they were published [25].
Despite a growing emphasis on using preprints to examine the publishing process within life sciences,
how these �ndings relate to the text of all documents in bioRxiv has yet to be examined.
Textual analysis uses linguistic, statistical, and machine learning techniques to analyze and extract
information from text [26]. For instance, scientists analyzed linguistic similarities and di�erences of
biomedical corpora [27,28]. Scientists have provided the community with a number of tools that aide
future text mining systems [29,30,31] as well as advice on how to train and test future text processing
systems [32,33,34]. Here, we use textual analysis to examine the bioRxiv repository, placing a
particular emphasis on understanding the extent to which full-text research can address hypotheses
derived from the study of metadata alone.
To understand how preprints relate to the traditional publishing ecosystem, we examine the linguistic
similarities and di�erences between preprints and peer-reviewed text and observe how linguistic
features change during the peer review and publishing process. We hypothesize that preprints and
biomedical text are pretty similar, especially when controlling for the di�erential uptake of preprints
across �elds. Furthermore, we hypothesize that document embeddings [35,36] provide a versatile
way to disentangle linguistic features along with serving as a suitable medium for improving preprint
repository functionality. We test this hypothesis by producing a linguistic landscape of bioRxiv
preprints, detecting preprints that change substantially during publication, and identify journals that
publish manuscripts that are linguistically similar to a target preprint. We encapsulate our �ndings
through a web app that projects a user-selected preprint onto this landscape and suggests journals
and articles that are linguistically similar. Our work reveals how linguistically similar and dissimilar
preprints are to peer-reviewed text, quanti�es linguistic changes that occur during the peer review
process, and highlights the feasibility of document embeddings with respect to preprint repository
functionality and peer review’s e�ect on publication time.
Materials and Methods
Corpora Examined
Text analytics is generally comparative in nature, so we selected three relevant text corpora for
analysis: the BioRxiv corpus, which is the target of the investigation, the PubMedCentral Open Access
corpus, which represents the peer-reviewed biomedical literature, the New York Times Annotated
Corpus, which is used a representative of general English text.
BioRxiv Corpus
BioRxiv [7] is a repository for life sciences preprints. We downloaded an XML snapshot of this
repository on February 3rd, 2020, from bioRxiv’s Amazon S3 bucket [37]. This snapshot contained the
full text and image content of 98,023 preprints. Preprints on bioRxiv are versioned, and in our
snapshot, 26,905 out of 98,023 contained more than one version. When preprints had multiple
versions, we used the latest one unless otherwise noted. Authors submitting preprints to bioRxiv can
select one of twenty-nine di�erent categories and tag the type of article: a new result, con�rmatory
�nding, or contradictory �nding. A few preprints in this snapshot were later withdrawn from bioRxiv;
when withdrawn, their content is replaced with the reason for withdrawal. As there were very few
withdrawn preprints, we did not treat these as a special case.
PubMed Central Open Access Corpus
PubMed Central (PMC) is a digital archive for the United States National Institute of Health’s Library of
Medicine (NIH/NLM) that contains full text biomedical and life science articles [38]. Paper availability
within PMC is mainly dependent on the journal’s participation level [39]. PMC articles can be closed
access ones from research funded by the NIH appearing after an embargo period or be published
under Gold Open Access [40] publishing schemes. Individual journals have the option to fully
participate in submitting articles to PMC, selectively participate sending only a few papers to PMC,
only submit papers according to NIH’s public access policy [41], or not participate at all. As of
September 2019, PMC had 5,725,819 articles available [42]. Out of these 5 million articles, about 3
million were open access (PMCOA) and available for text processing systems [30,43]. PMC also
contains a resource that holds author manuscripts that have already passed the peer review process
[44]. Since these manuscripts have already been peer-reviewed, we excluded them from our analysis
as the scope of our work is focused on examining the beginning and end of a preprint’s life cycle. We
downloaded a snapshot of the PMCOA corpus on January 31st, 2020. This snapshot contained many
types of articles: literature reviews, book reviews, editorials, case reports, research articles, and more.
We used only research articles, which aligns with the intended role of bioRxiv, and we refer to these
articles as the PMCOA corpus.
The New York Times Annotated Corpus
The New York Times Annotated Corpus (NYTAC) is [45] is a collection of newspaper articles from the
New York Times dating from January 1st, 1987, to June 19th, 2007. This collection contains over 1.8
million articles where 1.5 million of those articles have undergone manual entity tagged by library
scientists [45]. We downloaded this collection on August 3rd, 2020, from the Linguistic Data
Consortium (see Software and Data Availability section) and used the entire collection as a negative
control for our corpora comparison analysis.
Mapping bioRxiv preprints to their published counterparts
We used CrossRef [46] to identify bioRxiv preprints linked to a corresponding published article. We
accessed CrossRef on July 7th, 2020, and were able to link 23,271 preprints to their published
counterparts successfully. Out of those 23,271 preprint-published pairs, only 17,952 pairs had a
published version present within the PMCOA corpus. For our analyses that involved published links,
we only focused on this subset of preprints-published pairs.
Comparing Corpora
We compared the bioRxiv, PMCOA, and NYTAC corpora to assess the similarities and di�erences
between them. We used the NYTAC corpus as a negative control to assess the similarity between two
life sciences repositories when compared with non-life sciences text. All corpora contain both words
and non-word entities (e.g., numbers or symbols like
), which we refer to together as tokens to avoid
confusion. We calculated the following characteristic metrics for each corpus: the number of
documents, the number of sentences, the total number of tokens, the number of stopwords, the
average length of a document, the average length of a sentence, the number of negations, the
number of coordinating conjunctions, the number of pronouns and the number of past tense verbs.
Spacy is a lightweight and easy-to-use python package designed to preprocess and �lter text [47]. We
used spaCy’s “en_core_web_sm” model [47] (version 2.2.3) to preprocess all corpora and �lter out 326
spaCy-provided stopwords.
Following that cleaning process, we calculated the frequency of every token across all corpora.
Because many tokens were unique to one set or the other and observed at low frequency, we focused
on the union of the top 0.05% (~100) most frequently occurring tokens within each corpus. We
generated a contingency table for each token in this union and calculated the odds ratio along with
±
the 95% con�dence interval [48]. We measured corpora similarity by calculating the Kullback–Leibler
(KL) divergence across all corpora along with token enrichment analysis. This metric measures the
extent to which two distributions di�er. A low value of KL divergence implicates that two distributions
are similar and vice versa for high values. The optimal number of tokens used to calculate the KL
divergence is unknown, so we calculated this metric using a range of the 100 most frequently
occurring tokens between two corpora to the 5000 most frequently occurring tokens.
Constructing a Document Representation for Life Sciences Text
We sought to build a language model to quantify linguistic similarities of biomedical preprint and
articles. Word2vec is a suite of neural networks designed to model linguistic features of words based
on their appearance in the text. These models are trained to either predict a word based on its
sentence context, called a continuous bag of words (CBOW) model, or predict the context based on a
given word, called a skipgram model [35]. Through these prediction tasks, both networks learn latent
linguistic features that can be used for downstream tasks, such as identifying similar words. We used
gensim [49] (version 3.8.1) to train a CBOW [35] model over all the main text within each preprint in
the bioRxiv corpus. Determining the best number of dimensions for word embeddings can be a non-
trivial task; however, it has been shown that optimal performance is between 100-1000 dimensions
[50]. We chose to train the CBOW model using 300 hidden nodes, a batch size of 10000 words, and
for 20 epochs. We set a �xed random seed and used gensim’s default settings for all other
hyperparameters. Once trained, every token present within the CBOW model is associated with a
dense vector representing latent features captured by the network. We used these word vectors to
generate a document representation for every article within the bioRxiv and PMCOA corpora. For
each document, we used spaCy to lemmatize each token and then took the average of every
lemmatized token present within the CBOW model and the individual document [36]. Any token
present within the document but not in the CBOW model is ignored during this calculation process.
Visualizing and Characterizing Preprint Representations
We sought to visualize the landscape of preprints and determine the extent to which their
representation as document vectors corresponded to author-supplied document labels. We used
principal component analysis (PCA) [51] to project bioRxiv document vectors into a low-dimensional
space. We trained this model using scikit-learn’s [52] implementation of a randomized solver [53] with
a random seed of 100, an output of 50 principal components (PCs), and default settings for all other
hyperparameters. After training the model, every preprint within the bioRxiv corpus is assigned a
score for each generated PC. We sought to uncover concepts captured the generated PCs and used
the cosine similarity metric to examine these concepts. This metric takes two vectors as input and
outputs a score between -1 (most dissimilar) and 1 (most similar). We used this metric to score the
similarity between all generated PCs and every token within our CBOW model for our use case. We
report the top 100 positive and negative scoring tokens as word clouds. The size of each word
corresponds to the magnitude of similarity, and color represents positive (orange) or negative (blue)
association.
Discovering Unannotated Preprint-Publication Relationships
The bioRxiv maintainers have automated procedures to link preprints to peer-reviewed versions, and
many journals require authors to update preprints with a link to the published version. However, this
automation is primarily based on the exact matching of speci�c preprint attributes. If authors change
the title between a preprint and published version (e.g., [54] and [55]), then this change will prevent
bioRxiv from automatically establishing a link. Furthermore, if the authors do not report the
publication to bioRxiv, the preprint and its corresponding published version are treated as distinct
entities despite representing the same underlying research. We hypothesize that close proximity in
the document embedding space could match preprints with their corresponding published version. If
this �nding holds, we could use this embedding space to �ll in links missed by existing automated
processes. We used the subset of paper-preprint pairs annotated in CrossRef as described above to
calculate the distribution of available preprint to published distances. This distribution was calculated
by taking the Euclidean distance between the preprint’s embedding coordinates and the coordinates
of its corresponding published version. We also calculated a background distribution, which consisted
of the distance between each preprint with an annotated publication and a randomly selected article
from the same journal. We compared both distributions to determine if there was a di�erence
between both groups as a signi�cant di�erence would indicate that this embedding method can parse
preprint-published pairs apart. Following the comparison of the two distributions, we calculated
distances between preprints without a published version link with PMCOA articles that weren’t
matched with a corresponding preprint. We �ltered any potential links with distances greater than the
minimum value of the background distribution as we considered these pairs to be true negatives.
Lastly, we binned the remaining pairs based on percentiles from the annotated pairs distribution at
the [0,25th percentile), [25th percentile, 50th percentile), [50th percentile, 75th percentile), and [75th
percentile, minimum background distance). We randomly sampled 50 articles from each bin and
shu�ed these four sets to produce a list of 200 potential preprint-published pairs with a randomized
order. We supplied these pairs to two co-authors to manually determine if each link between a
preprint and a putative matched version was correct or incorrect. After the curation process, we
encountered eight disagreements between the reviewers. We supplied these pairs to a third scientist,
who carefully reviewed each case and made a �nal determination. Using this curated set, we
evaluated the extent to which distance in the embedding space revealed valid but unannotated links
between preprints and their published versions.
Measuring Time Duration for Preprint Publication Process
Preprints that are published can take varying amounts of time to be published. We sought to measure
the time required for preprints to be published in the peer-reviewed literature and compared this
time measurement across author-selected preprint categories as well as individual preprints. First, we
queried bioRxiv’s application programming interface (API) to obtain the date a preprint was posted
onto bioRxiv as well as the date a preprint was accepted for publication. We measured time elapsed
as the di�erence between the date at which a preprint was �rst posted on bioRxiv and its publication
date. Along with calculating the amount of time elapsed, we also recorded the number of di�erent
preprint versions posted onto bioRxiv.
Using this captured data, we used the Kaplan-Meier estimator [56] via the KaplanMeierFitter function
from the lifelines [57] (version 0.25.6) python package to calculate the half-life of preprints across all
preprint categories within bioRxiv. We considered survival events as preprints that have yet to be
published. There were a limited number of cases in which authors appeared to post preprints after
the publication date, which results in preprints receiving a negative time di�erence, as previously
reported [58]. We removed these preprints for this analysis as they were incompatible with the rules
of the bioRxiv repository.
Following our half-life calculation, we measured the textual di�erence between preprints and their
corresponding published version by calculating the Euclidean distance for their respective embedding
representation. This metric can be di�cult to understand within the context of textual di�erences, so
we sought to contextualize the meaning of a distance unit. We accomplish this by �rst randomly
sampled with replacement a pair of preprints from the Bioinformatics topic area as this was well
represented within bioRxiv and contains a diverse set of research articles. Next, we calculated the
distance between two preprints 1000 times and reported the mean. We repeated the above
procedure using every preprint within bioRxiv as a whole. These two means serve as normalized
benchmarks to compare against as distance units are only meaningful when compared to other
distances within the same space. Following our contextualization approach, we performed linear
regression to model the relationship between preprint version count with a preprint’s time to
publication. We also performed linear regression to measure the relationship between document
embedding distance and a preprint’s time to publication. For this analysis, we retained preprints with
negative time within our linear regression model, and we observed that these preprints had minimal
impact on results. We visualize our version count regression model as a violin plot and our document
embeddings regression model as a square bin plot.
Building Classi�ers to Detect Linguistically Similar Journal Venues and
Published Articles
Preprints are more likely to be published in journals that contained similar content to work in
question. We assessed this claim by building classi�ers based on document and journal
representations. First, we removed all journals that had fewer than 100 papers in the PMC corpus. We
held our preprint-published subset (see above section ‘Mapping bioRxiv preprints to their published
counterparts’) and treated it as a gold standard test set. We used the remainder of the PMCOA corpus
for training and initial evaluation for our models.
Speci�c journals publish articles in a focused topic area, while others publish articles that cover many
topics. Likewise, some journals have a publication rate of at most hundreds of papers per year, while
others publish at a rate of at least ten thousand papers per year. Accounting for these characteristics,
we designed two approaches - one centered on manuscripts and another centered on journals.
We identi�ed manuscripts that were most similar to the preprint query for the manuscript-based
approach and evaluated where these documents were published. We embedded each query article
into the space de�ned by the word2vec model (see above section ‘Constructing a Document
Representation for Life Sciences Text’). We selected manuscripts close to the query via Euclidean
distance in the embedding space. Once identi�ed, we return the journal in which these articles were
published. We also return the articles that led to each journal being reported as this approach allows
for journals that frequently publish papers to engulf our results.
We constructed a journal-based approach to accompany the manuscript-based process to account for
the overrepresentation of these high publishing frequency journals. We identi�ed the most similar
journals for this approach by constructing a journal representation in the same embedding space. We
computed this representation by taking the average embedding of all published papers within a given
journal. We then projected a query article into the same space and returned journals close to the
query.
Both models were constructed using the scikit-learn k-Nearest Neighbors implementation [59] with
the number of neighbors set to 10 as this is an appropriate number for our use case. We consider a
prediction to be a true positive if the correct journal appears within our reported list of neighbors and
evaluate our performance using 10-fold cross-validation on the training set along with test set
evaluation.
Web Application for Discovering Similar Preprints and Journals
We developed a web application that places any bioRxiv or medRxiv preprint into the overall
document landscape and identi�es similar papers and journals. The application downloads a pdf
version of any preprint hosted on the bioRxiv or medRxiv server uses PyMuPDF [60] to extract text
from the downloaded pdf and feeds the extracted text into our CBOW model to construct a document
embedding representation. We pass this representation onto our journal and manuscript search to
identify journals based on the ten closest neighbors of individual papers and journal centroids. We
implemented this search using the scikit-learn implementation of k-d trees. To run it more cost-
e�ectively in a cloud computing environment with limited available memory, we sharded the k-d trees
into four trees.
The app provides a visualization of the article’s position within our training data to illustrate the local
publication landscape, We used SAUCIE [61], an autoencoder designed to cluster single-cell RNA-seq
data, to build a two-dimensional embedding space that could be applied to newly generated preprints
without retraining, a limitation of other approaches that we explored for visualizing entities expected
to lie on a nonlinear manifold. We trained this model on document embeddings of PMC articles that
did not contain a matching preprint version. We used the following parameters to train the model: a
hidden size of 2, a learning rate of 0.001, lambda_b of 0, lambda_c of 0.001, and lambda_d of 0.001 for
5000 iterations. When a user requests a new document, we can then project that document onto our
generated two-dimensional space; thereby, allowing the user to see where their preprint falls along
the landscape. We illustrate our recommendations as a shortlist and provide access to our network
visualization at our website (see Software and Data Availability).
Analysis of the Preprints in Motion Collection
Our manuscript describes the large-scale analysis of bioRxiv. Concurrent with our work, another set of
authors performed a detailed curation and analysis of a subset of bioRxiv [25] that was focused on
preprints posted during the initial stages of the COVID-19 pandemic. The curated analysis was
designed to examine preprints at a time of increased readership [62] and includes certain preprints
posted from January 1st, 2020 to April 30th, 2020 [25]. We sought to contextualize this subset, which
we term “Preprints in Motion” after the title of the preprint [25], within our global picture of the
bioRxiv preprint landscape. We extracted all preprints from the set reported in Preprints in Motion
[25] and retained any entries in the bioRxiv repository. We manually downloaded the XML version of
these preprints and mapped them to their published counterparts as described above. We used
Pubmed Central’s DOI converter [63] to map the published article DOIs with their respective PubMed
Central IDs. We retained articles that were included in the PMCOA corpus and performed a token
analysis as described to compare these preprints with their published versions. As above, we
generated document embeddings for every obtained preprint and published article. We projected
these preprint embeddings onto our publication landscape to visually observe the dispersion of this
subset. Finally, we performed a time analysis that paralleled our approach for the full set of preprint-
publication pairs to examine relationships between linguistic changes and the time to publication.
Results
Comparing bioRxiv to other corpora
bioRxiv Metadata Statistics
The preprint landscape is rapidly changing, and the number of bioRxiv preprints in our data download
(71,118) was nearly double that of a recent study that reported on a snapshot with 37,648 preprints
[13]. Because the rate of change is rapid, we �rst analyzed category data and compared our results
with previous �ndings. As in previous reports [13], neuroscience remains the most common category
of preprints, followed by bioinformatics (Supplemental Figure S1). Microbiology, which was �fth in the
most recent report [13], has now surpassed evolutionary biology and genomics to move into third.
When authors upload their preprints, they select from three result category types: new results,
con�rmatory results, or contradictory results. We found that nearly all preprints (97.5%) were
categorized as new results, consistent with reports on a smaller set [64]. The results taken together
suggest that while bioRxiv has experienced dramatic growth, how it is being used appears to have
remained consistent in recent years.
Global analysis reveals similarities and di�erences between bioRxiv and
PMC
Table 1: Summary statistics for the bioRxiv, PMC, and NYTAC corpora.
Metric
bioRxiv
PMC
NYTAC
document count
71,118
1,977,647
1,855,658
sentence count
22,195,739
480,489,811
72,171,037
token count
420,969,930
8,597,101,167
1,218,673,384
stopword count
158,429,441
3,153,077,263
559,391,073
avg. document length
312.10
242.96
38.89
avg. sentence length
22.71
21.46
19.89
negatives
1,148,382
24,928,801
7,272,401
coordinating conjunctions
14,295,736
307,082,313
38,730,053
coordinating conjunctions%
3.40%
3.57%
3.18%
pronouns
4,604,432
74,994,125
46,712,553
pronouns%
1.09%
0.87%
3.83%
passives
15,012,441
342,407,363
19,472,053
passive%
3.57%
3.98%
1.60%
A
B
C
D
E
Figure 1: A. The Kullback–Leibler divergence measures the extent to which the distributions, not speci�c tokens, di�er
from each other. The token distribution of bioRxiv and PMC corpora is more similar than these biomedical corpora are
to the NYTAC one. B. The signi�cant di�erences in token frequencies for the corpora appear to be driven by the �elds
with the highest uptake of bioRxiv, as terms from neuroscience and genomics are relatively more abundant in bioRxiv.
We plotted the 95% con�dence interval for each reported token. C. Of the tokens that di�er between bioRxiv and PMC,
the most abundant in bioRxiv are “et” and “al” while the most abundant in PMC is “study.” D. The signi�cant di�erences
in token frequencies for preprints and their corresponding published version often appear to be associated with
typesetting and supplementary or additional materials. We plotted the 95% con�dence interval for each reported token.
E. The tokens with the largest absolute di�erences in abundance appear to be stylistic.
Documents within bioRxiv were slightly longer than those within PMCOA, but both were much longer
than those from the control (NYTAC) (Table 1). The average sentence length, the fraction of pronouns,
and the use of the passive voice were all more similar between bioRxiv and PMC than they were to
NYTAC(Table 1). The Kullback–Leibler (KL) divergence of term frequency distributions between bioRxiv
and PMCOA were low, especially among the top few hundred tokens (Figure 1A). As more tokens were
incorporated, the KL divergence started to increase but remained much lower than the biomedical
corpora compared against NYTAC. These �ndings support our notion that bioRxiv is linguistically
similar to the PMCOA repository.
Terms like “neurons”, “genome”, and “genetic”, which are common in genomics and neuroscience,
were more common in bioRxiv than PMCOA while others associated with clinical research, such as
“clinical” “patients” and “treatment” were more common in PMCOA (Figure 1B and 1C). When
controlling for the di�erences in the body of documents to identify textual changes associated with
the publication process, we found that tokens such as “et” “al” were enriched for bioRxiv while “
”, “–”
were enriched for PMCOA (Figure 1D and 1E). Furthermore, we found that speci�c changes appeared
to be related to journal styles: “�gure” was more common in bioRxiv while “�g” was relatively more
common in PMCOA. Other changes appeared to be associated with an increasing reference to content
external to the manuscript itself: the tokens “supplementary”, “additional” and “�le” were all more
common in PMCOA than bioRxiv, suggesting that journals are not simply replacing one token with
another but that there are more mentions of such content after peer review.
These results taken together suggest that the structure of the text within preprints on bioRxiv are
similar to published articles within PMCOA. The di�erences in uptake across �elds are supported by
di�erences in authors’ categorization of their articles and by the text within the articles themselves. At
the level of individual manuscripts, the terms that change the most appear to be associated with
typesetting, journal style, and an increasing reliance on additional materials after peer review.
Document embeddings derived from bioRxiv reveal �elds and
sub�elds
±
A
PC 1
B
PC 2
C
D
E
Figure 2: A. Principal components (PC) analysis of bioRxiv word2vec embeddings groups documents based on author-
selected categories. We visualized documents from key categories on a scatterplot for the �rst two PCs. The �rst PC
separated cell biology from informatics-related �elds, and the second PC separated bioinformatics from neuroscience
�elds. B. A word cloud visualization of PC1. Each word cloud depicts the cosine similarity score between tokens and the
�rst PC. Tokens in orange were most similar to the PC’s positive direction, while tokens in blue were most similar to the
PC’s negative direction. The size of each token indicates the magnitude of the similarity. C. A word cloud visualization of
PC2, which separated bioinformatics from neuroscience. Similar to the �rst PC, tokens in orange were most similar to
the PC’s positive direction, while tokens in blue were most similar to the PC’s negative direction. The size of each token
indicates the magnitude of the similarity. D. Examining PC1 values for each article by category created a continuum
from informatics-related �elds on the top through cell biology on the bottom. Speci�c article categories (neuroscience,
genetics) were spread throughout PC1 values. E. Examining PC2 values for each article by category revealed �elds like
genomics, bioinformatics, and genetics on the top and neuroscience and behavior on the bottom.
Document embeddings provide a means to categorize the language of documents in a way that takes
into account the similarities between terms [36,65,66]. We found that the �rst two PCs separated
articles from di�erent author-selected categories (Figure 2A). Certain neuroscience papers appeared
to be more associated with the cellular biology direction of PC1, while others seemed to be more
associated with the informatics-related direction Figure 2A). This suggests that the concepts captured
by PCs were not exclusively related to their �eld.
Visualizing token-PC similarity revealed tokens associated with certain research approaches (Figures
2B and 2C). Token association of PC1 shows the separation of cell biology and informatics-related
�elds through tokens: “empirical”, “estimates” and “statistics” depicted in orange and “cultured” and
“overexpressing” shown in blue (Figure 2B). Association for PC2 shows the separation of
bioinformatics and neuroscience via tokens: “genomic”, “genome” and “genomes” depicted in orange
and “evoked”, “stimulus” and “stimulation” shown in blue (Figure 2C).
Examining the value for PC1 across all author-selected categories revealed an ordering of �elds from
cell biology to informatics-related disciplines (Figure 2D). These results suggest that a primary driver
of the variability within the language used in bioRxiv could be the divide between informatics and cell
biology approaches. A similar analysis for PC2 suggested that neuroscience and bioinformatics
present a similar language continuum (Figure 2E). This result supports the notion that bioRxiv
contains an in�ux of neuroscience and bioinformatics-related research results. For both of the top
two PCs, the submitter-selected category of systems biology preprints was near the middle of the
distribution and had a relatively large interquartile range when compared with other categories
(Figure 2D and 2E), suggesting that systems biology is a broader sub�eld containing both informatics
and cellular biology approaches.
Examining the top �ve and bottom �ve preprints within the systems biology �eld reinforces PC1’s
dichotomous theme (Table 2). Preprints with the highest values [67,68,69,70,71] included software
packages, machine learning analyses, and other computational biology manuscripts, while preprints
with the lowest values [72,73,74,75,76] were focused on cellular signaling and protein activity. We
provide the rest of our 50 generated PCs in our online repository (see Software and Data Availability).
Table 2: PC1 divided the author-selected category of systems biology preprints along an axis from computational to
molecular approaches.
Title [citation]
PC1
License
Figure Thumbnail
Conditional Robust Calibration (CRC): a new
computational Bayesian methodology for
model parameters estimation and
identi�ability analysis [67]
4.522818390064091
None
FPtool a software tool to obtain in silico
genotype-phenotype signatures and
�ngerprints based on massive model
simulations [77]
4.348956760251298
CC-BY
GpABC: a Julia package for approximate
Bayesian computation with Gaussian process
emulation [70]
4.259104249060651
CC-BY-NC-ND
Notions of similarity for computational biology
models [69]
4.079855550647664
CC-BY-NC-ND
Title [citation]
PC1
License
Figure Thumbnail
SBpipe: a collection of pipelines for
automating repetitive simulation and analysis
tasks [71]
4.022240241143516
CC-BY-NC-ND
Bromodomain inhibition reveals FGF15/19 as a
target of epigenetic regulation and metabolic
control [78]
-3.4783803547922414
None
Inhibition of Bruton’s tyrosine kinase reduces
NF-kB and NLRP3 in�ammasome activity
preventing insulin resistance and
microvascular disease [75]
-3.6926161167521476
None
Spatiotemporal proteomics uncovers
cathepsin-dependent host cell death during
bacterial infection [72]
-3.728443135960558
CC-BY-ND
NADPH consumption by L-cystine reduction
creates a metabolic vulnerability upon glucose
deprivation [74]
-3.7363965062637288
None
AKT but not MYC promotes reactive oxygen
species-mediated cell death in oxidative
culture [76]
-3.8769231933681176
None
Document embedding similarities reveal unannotated preprint-
publication pairs
A
B
C
Figure 3: A. Preprints are closer in document embedding space to their corresponding peer-reviewed publication than
they are to random papers published in the same journal. B. Potential preprint-publication pairs that are unannotated
but within the 50th percentile of all preprint-publication pairs in the document embedding space are likely to represent
true preprint-publication pairs. We depict the fraction of true positives over the total number of pairs in each bin.
Accuracy is derived from the curation of a randomized list of 200 potential pairs (50 per quantile) performed in duplicate
with a third rater used in the case of disagreement. C. Most preprints are eventually published. We show the publication
rate of preprints since bioRxiv �rst started. The x-axis represents months since bioRxiv started, and the y-axis
represents the proportion of preprints published given the month they were posted. The light blue line represents the
publication rate previously estimated by Abdill et al. [13]. The dark blue line represents the updated publication rate
using only CrossRef-derived annotations, while the dark green line includes annotations derived from our embedding
space approach. The horizontal lines represent the overall proportion of preprints published as of the time of the
annotation snapshot.
Distances between preprints and their corresponding published versions were nearly always lower
than preprints paired with a random article published in the same journal (Figure 3A). This suggests
that embedding distances can identify documents with similar textual content. Approximately 98% of
our 200 pairs with an embedding distance in the 0-25th and 25th-50th percentile bins were scored as
true matches (Figure 3B). These two bins contained 1,542 preprint-article pairs, suggesting that many
preprints may have been published but not previously connected with their published versions. There
is a particular enrichment for preprints published but unlinked within the 2017-2018 interval (Figure
3C). We expected a higher proportion of such preprints before 2019 (many of which may not have
been published yet); however, observing relatively few missed annotations before 2017 was against
our expectations. There are several possible explanations for this increasing fraction of missed
annotations. As the number of preprints posted on bioRxiv grows, it may be harder for bioRxiv to
establish a link between preprints and their published counterparts simply due to the scale of the
challenge. It is possible that the set of authors participating in the preprint ecosystem is changing and
that new participants may be less likely to report missed publications to bioRxiv. Finally, as familiarity
with preprinting grows, it is possible that authors are posting preprints earlier in the process and that
metadata �elds that bioRxiv uses to establish a link may be less stable.
Preprints with more versions or more text changes took longer to
publish
A
B
C
Figure 4: A. Author-selected categories were associated with modest di�erences in respect to publication half-life.
Author-selected preprint categories are shown on the y-axis, while the x-axis shows the median time-to-publish for each
category. Error bars represent 95% con�dence intervals for each median measurement. B. Preprints with more versions
were associated with a longer time to publish. The x-axis shows the number of versions of a preprint posted on bioRxiv.
The y-axis indicates the number of days that elapsed between the �rst version of a preprint posted on bioRxiv and the
date at which the peer-reviewed publication appeared. The density of observations is depicted in the violin plot with an
embedded boxplot. C. Preprints with more substantial text changes took longer to be published. The x-axis shows the
Euclidean distance between document representations of the �rst version of a preprint and its peer-reviewed form. The
y-axis shows the number of days elapsed between the �rst version of a preprint posted on bioRxiv and when a preprint
is published. The color bar on the right represents the density of each hexbin in this plot, where more dense regions are
shown in a brighter color.
The process of peer review includes several steps, which take variable amounts of time [79], and we
sought to measure if there is a di�erence in publication time between author-selected categories of
preprints (Figure 4A). Of the most abundant preprint categories microbiology was the fastest to
publish (140 days, (137, 145 days) [95% CI]) and genomics was the slowest (190 days, (185, 195 days)
[95% CI]) (Figure 4A). We did observe category-speci�c di�erences; however, these di�erences were
generally modest, suggesting that the peer review process did not di�er dramatically between
preprint categories. One exception was the Scienti�c Communication and Education category, which
took substantially longer to be peer-reviewed and published (373 days, (373, 398 days) [95% CI]). This
hints that there may be di�erences in the publication or peer review process or culture that apply to
preprints in this category.
Examining peer review’s e�ect on individual preprints, we found a positive correlation between
preprints with multiple versions and the time elapsed until publication (Figure 4B). Each new version
adds additional 51 days before a preprint is published. This time duration seems broadly compatible
with the amount of time it would take to receive reviews and revise a manuscript, suggesting that
many authors may be updating their preprints in response to peer reviews or other external
feedback. The embedding space allows us to compare preprint and published documents to
determine if the level of change that documents undergo relates to the time it takes them to be
published. Distances in this space are arbitrary and must be compared to reference distances. We
found that the average distance of two randomly selected papers from the bioinformatics category
was 4.470, while the average distance of two randomly selected papers from bioRxiv was 5.343.
Preprints with large embedding space distances from their corresponding peer-reviewed publication
took longer to publish (Figure 4C): each additional unit of distance corresponded to roughly forty-
three additional days.
Overall, our �ndings support a model where preprints are reviewed multiple times or require more
extensive revisions take longer to publish.
Preprints with similar document embeddings share publication
venues
We developed an online application that returns a listing of published papers and journals closest to a
query preprint in document embedding space. This application uses two k-nearest neighbor classi�ers
that achieved better performance than our baseline model (Supplemental Figure S2) to identify these
entities. Users supply our app with digital object identi�ers (DOIs) from bioRxiv or medRxiv, and the
corresponding preprint is downloaded from the repository. Next, the preprint’s PDF is converted to
text, and this text is used to construct a document embedding representation. This representation is
supplied to our classi�ers to generate a listing of the ten papers and journals with the most similar
representations in the embedding space (Figures 5A, 5B and 5C). Furthermore, the user-requested
preprint’s location in this embedding space is then displayed on our interactive map, and users can
select regions to identify the terms most associated with those regions (Figures 5D and 5E). Users can
also explore the terms associated with the top 50 PCs derived from the document embeddings, and
those PCs vary across the document landscape.
Figure 5: The preprint-similarity-search app work�ow allows users to examine where an individual preprint falls in the
overall document landscape. A. Starting with the home screen, users can paste in a bioRxiv or medRxiv DOI, which
sends a request to bioRxiv or medRxiv. Next, the app preprocesses the requested preprint and returns a listing of (B)
the top ten most similar papers and (C) the ten closest journals. D. The app also displays the location of the query
preprint in PMC. E. Users can select a square within the landscape to examine statistics associated with the square,
including the top journals by article count in that square and the odds ratio of tokens.
Contextualizing the Preprints in Motion Collection
A
B
C
D
E
Figure 6: The Preprints in Motion Collection results are similar to all preprint results, except that their time to
publication was independent of the number of preprint versions and amount of linguistic change. A. Tokens that
di�ered included those associated with typesetting and those related to the nomenclature of the virus that causes
COVID-19. Error bars show 95% con�dence intervals for each token. B. Of the tokens that di�er between Preprints in
Motion and their published counterparts, the most abundant were associated with the nomenclature of the virus. C.
The Preprints in Motion fall across the landscape of PMCOA with respect to linguistic properties. This square bin plot
depicts the binning of all published papers within the PMCOA corpus. High-density regions are depicted in yellow, while
low-density regions are in dark blue. Red dots represent the Preprints in Motion Collection. D. The Preprints in Motion
were published faster than other bioRxiv preprints, and the number of versions was not associated with an increase in
time to publication. The x-axis shows the number of versions of a preprint posted on bioRxiv. The y-axis indicates the
number of days that elapsed between the �rst version of a preprint posted on bioRxiv and the date at which the peer-
reviewed publication appeared. The density of observations is depicted in the violin plot with an embedded boxplot. The
red dots and red regression line represent Preprints in Motion. D. The Preprints in Motion were published faster than
other bioRxiv preprints, and no dependence between the amount of linguistic change and time to publish was
observed. The x-axis shows the Euclidean distance between document representations of the �rst version of a preprint
and its peer-reviewed form. The y-axis shows the number of days elapsed between the �rst version of a preprint posted
on bioRxiv and when a preprint is published. The color bar on the right represents the density of each hexbin in this
plot, where more dense regions are shown in a brighter color. The red dots and red regression line represent Preprints
in Motion.
The Preprints in Motion collection included a set of preprints posted during the �rst four months of
2020. We examined the extent to which preprints in this set were representative of the patterns that
we identi�ed from our analysis on all of bioRxiv. As with all of bioRxiv, typesetting tokens changed
between preprints and their paired publications. Our token-level analysis identi�ed certain patterns
consistent with our �ndings across bioRxiv (Figure 6A and 6B). However, in this set, we also observe
changes likely associated with the fast-moving nature of COVID-19 research: the token “2019-ncov”
became less frequently represented while “sars” and “cov-2” became more represented, likely due to a
shift in nomenclature from “2019-nCoV” to “SARS-CoV-2”. The Preprints in Motion were not strongly
colocalized in the linguistic landscape, suggesting that the collection covers a diverse set of research
approaches (Figure 6C). Preprints in this collection were published faster than the broader set of
bioRxiv preprints (Figure 6D and 6E). The relationship between time to publication and the number of
versions (Figure 6D) and the relationship between time to publication and the amount of linguistic
change (Figure 6E) were both lost in the Preprints in Motion set. Our �ndings suggest that Preprints in
Motion changed during publication in ways aligned with changes in the full preprint set but that peer
review was accelerated in ways that broke the time dependences observed with the full bioRxiv set.
Discussion and Conclusions
BioRxiv is a constantly growing repository that contains life science preprints. The majority of research
involving bioRxiv focuses on the metadata of preprints; however, the language contained within these
preprints has not previously been systematically examined. Throughout this work, we sought to
analyze the language within these preprints and understand how it changes in response to peer
review. Our global corpora analysis found that writing within bioRxiv is consistent with the biomedical
literature in the PMCOA repository, suggesting that bioRxiv is linguistically similar to PMCOA. Token-
level analyses between bioRxiv and PMCOA suggested that research �elds drive signi�cant
di�erences; e.g., more patient-related research is prevalent in PMCOA than bioRxiv. This observation
is expected as preprints focused on medicine are supported by the complementary medRxiv
repository [8]. Token-level analyses for preprints and their corresponding published version suggest
that peer review may focus on data availability and incorporating extra sections for published papers;
however, future studies are needed to ascertain individual token level changes as preprints venture
through the publication process.
Document embeddings are a versatile way to examine language contained within preprints,
understanding peer review’s e�ect on preprints, and provide extra functionality for preprint
repositories. Examining linguistic variance within document embeddings of life science preprints
revealed that the largest source of variability was informatics. This observation bisects the majority of
life science research categories that have integrated preprints within their publication work�ow.
Preprints are typically linked with their published articles via bioRxiv manually establishing links or
authors self-reporting that their preprint has been published; however, gaps can occur as preprints
change their appearance through multiple versions or authors do not notify bioRxiv. Our work
suggests that document embeddings can help �ll in missing links within bioRxiv. Furthermore, our
analysis reveals that the publication rate for preprints is higher than previously estimated, even
though our analysis can only account for published open access papers. Our results raise the lower
bound of the total preprint publication fraction; however, the true fraction is necessarily higher.
Future work, especially that which aims to assess the fraction of preprints that are eventually
published, should account for the possibility of missed annotations.
Preprints take a variable amount of time to become published, and we examined factors that
in�uence a preprint’s time to publication. Our half-life analysis on preprint categories revealed that
preprints in most bioRxiv categories take similar amounts of time to be published. An apparent
exception is the scienti�c communication and education category, which contained preprints that
took much longer to publish. Regarding individual preprints, each new version adds several weeks to
a preprints time to publication, which is roughly aligned with authors making changes after a round of
peer review; furthermore, preprints that undergo substantial changes take longer to publish. Overall,
these results illustrate that bioRxiv is a practical resource for obtaining insight into the peer-review
process.
Lastly, we found that document embeddings were associated with the eventual journal at which the
work was published. We trained two machine learning models to identify which journals publish
linguistically similar papers towards a query preprint. Our models achieved a considerably higher fold
change over the baseline model, so we constructed a web application that makes our models
available to the public and returns a list of the papers and journals that are linguistically similar to a
bioRxiv or medRxiv preprint.
Software and Data Availability
An online version of this manuscript is available under a Creative Commons Attribution License at
https://greenelab.github.io/annorxiver_manuscript/. Source for the research portions of this project is
dual licensed under the BSD 3-Clause and Creative Commons Public Domain Dedication Licenses at
https://github.com/greenelab/annorxiver. The preprint similarity search website can be found at
https://greenelab.github.io/preprint-similarity-search/, and code for the website is available under a
BSD-2-Clause Plus Patent License at https://github.com/greenelab/preprint-similarity-search. Full text
access for the bioRxiv repository is available at https://www.biorxiv.org/tdm. Access to PubMed
Central’s Open Access subset is available on NCBI’s FTP server at
https://www.ncbi.nlm.nih.gov/pmc/tools/ftp/. Access to the New York Times Annotated Corpus
(NYTAC) is available upon request with the Linguistic Data Consortium at
https://catalog.ldc.upenn.edu/LDC2008T19.
Acknowledgments
The authors would like to thank Ariel Hippen Anderson for evaluating potential missing preprint to
published version links. We also would like to thank Richard Sever and the bioRxiv team for their
assistance with access to and support with questions about preprint full text downloaded from
bioRxiv.
Funding
This work was supported by grants from the Gordon Betty Moore Foundation (GBMF4552) and the
National Institutes of Health’s National Human Genome Research Institute (NHGRI) under awards T32
HG00046 and R01 HG010067.
Competing Interests
Marvin Thielk receives a salary from Elsevier Inc. where he contributes NLP expertise to health
content operations. Elsevier did not restrict the results or interpretations that could be published in
this manuscript. The opinions expressed here do not re�ect the o�cial policy or positions of Elsevier
Inc.
References
1. Scienti�c communication pathways: an overview and introduction to a symposium
David F. Zaye, W. V. Metanomski
Journal of Chemical Information and Computer Sciences (2002-05-01) https://doi.org/bwsxhg
DOI: 10.1021/ci00050a001
2. The trouble with medical journals
Richard Smith
Journal of the Royal Society of Medicine (2006)
3. The Transition from Paper to Electronic Journals
Hak Joon Kim
The Serials Librarian (2001-11-19) https://doi.org/d7rnh2
DOI: 10.1300/j123v41n01_04
4. Preprints: What Role Do These Have in Communicating Scienti�c Results?
Susan A. Elmore
Toxicologic Pathology (2018-04-08) https://doi.org/ghdd7c
DOI: 10.1177/0192623318767322 · PMID: 29628000 · PMCID: PMC5999550
5. The prehistory of biology preprints: A forgotten experiment from the 1960s
Matthew Cobb
PLOS Biology (2017-11-16) https://doi.org/c6wv
DOI: 10.1371/journal.pbio.2003995 · PMID: 29145518 · PMCID: PMC5690419
6. arXiv.org: the Los Alamos National Laboratory e‐print server
Gerry McKiernan
International Journal on Grey Literature (2000-09) https://doi.org/fg8pw7
DOI: 10.1108/14666180010345564
7. bioRxiv: the preprint server for biology
Richard Sever, Ted Roeder, Samantha Hindle, Linda Sussman, Kevin-John Black, Janet Argentine,
Wayne Manos, John R. Inglis
Cold Spring Harbor Laboratory (2019-11-06) https://doi.org/ggc46z
DOI: 10.1101/833400
8. medRxiv.org - the preprint server for Health Sciences https://www.medrxiv.org/
9. The Second Wave of Preprint Servers: How Can Publishers Keep A�oat?
By
The Scholarly Kitchen (2019-10-16) https://scholarlykitchen.sspnet.org/2019/10/16/the-second-
wave-of-preprint-servers-how-can-publishers-keep-a�oat/
10. Rxivist.org: Sorting biology preprints using social media and readership metrics
Richard J. Abdill, Ran Blekhman
PLOS Biology (2019-05-21) https://doi.org/dm27
DOI: 10.1371/journal.pbio.3000269 · PMID: 31112533 · PMCID: PMC6546241
11. How the Scienti�c Community Reacts to Newly Submitted Preprints: Article Downloads,
Twitter Mentions, and Citations
Xin Shuai, Alberto Pepe, Johan Bollen
PLoS ONE (2012-11-01) https://doi.org/f4cw6r
DOI: 10.1371/journal.pone.0047523 · PMID: 23133597 · PMCID: PMC3486871
12. Quantifying and contextualizing the impact of bioRxiv preprints through automated social
media audience segmentation
Jedidiah Carlson, Kelley Harris
Cold Spring Harbor Laboratory (2020-03-10) https://doi.org/ghdd66
DOI: 10.1101/2020.03.06.981589
13. Tracking the popularity and outcomes of all bioRxiv preprints
Richard J Abdill, Ran Blekhman
eLife (2019-04-24) https://doi.org/gf2str
DOI: 10.7554/elife.45133 · PMID: 31017570 · PMCID: PMC6510536
14. An analysis of published journals for papers posted on bioR X iv
Hiroyuki Tsunoda, Yuan Sun, Masaki Nishizawa, Xiaomin Liu, Kou Amano
Proceedings of the Association for Information Science and Technology (2019-10-18)
https://doi.org/ggz7f9
DOI: 10.1002/pra2.175
15. The relationship between bioRxiv preprints, citations and altmetrics
Nicholas Fraser, Fakhri Momeni, Philipp Mayr, Isabella Peters
Quantitative Science Studies (2020-04-01) https://doi.org/gg2cz3
DOI: 10.1162/qss_a_00043
16. Releasing a preprint is associated with more attention and citations for the peer-reviewed
article
Darwin Y Fu, Jacob J Hughey
eLife (2019-12-06) https://doi.org/ghd3mv
DOI: 10.7554/elife.52646 · PMID: 31808742 · PMCID: PMC6914335
17. Preprints and Scholarly Communication: An Exploratory Qualitative Study of Adoption,
Practices, Drivers and Barriers
Andrea Chiarelli, Rob Johnson, Stephen Pin�eld, Emma Richens
F1000Research (2019-11-25) https://doi.org/ghp38z
DOI: 10.12688/f1000research.19619.2 · PMID: 32055396 · PMCID: PMC6961415
18. The Need for Speed: How Quickly Do Preprints Become Published Articles?
Rachel Herbert, Kate Gasson, Alex Ponsford
SSRN Electronic Journal (2019) https://doi.org/ghd3mt
DOI: 10.2139/ssrn.3455146
19. Technical and social issues in�uencing the adoption of preprints in the life sciences
Naomi C. Penfold, Jessica K. Polka
PLOS Genetics (2020-04-20) https://doi.org/dtt2
DOI: 10.1371/journal.pgen.1008565 · PMID: 32310942 · PMCID: PMC7170218
20. Biologists urged to hug a preprint
Ewen Callaway, Kendall Powell
Nature (2016-02-16) https://doi.org/ghdd62
DOI: 10.1038/530265a · PMID: 26887471
21. Day-to-day discovery of preprint–publication links
Guillaume Cabanac, Theodora Oikonomidi, Isabelle Boutron
Scientometrics (2021-04-18) https://doi.org/gjr9k4
DOI: 10.1007/s11192-021-03900-7 · PMCID: PMC8053368
22. On the value of preprints: An early career researcher perspective
Sarvenaz Sarabipour, Humberto J. Debat, Edward Emmott, Steven J. Burgess, Benjamin
Schwessinger, Zach Hensel
PLOS Biology (2019-02-21) https://doi.org/gfw8hd
DOI: 10.1371/journal.pbio.3000151 · PMID: 30789895 · PMCID: PMC6400415
23. Prepublication Communication of Research Results
Michael J. Adams, Reid N. Harris, Evan H. C. Grant, Matthew J. Gray, M. Camille Hopkins, Samuel A.
Iverson, Robert Likens, Mark Mandica, Deanna H. Olson, Alex Shepack, Hardin Waddle
EcoHealth (2018-08-07) https://doi.org/ghn66s
DOI: 10.1007/s10393-018-1352-3 · PMID: 30088185 · PMCID: PMC6245104
24. Peer Review and bioRxiv
Leslie M. Loew
Biophysical Journal (2016-08) https://doi.org/ghdd6x
DOI: 10.1016/j.bpj.2016.06.035 · PMID: 27508451 · PMCID: PMC4982934
25. Preprints in motion: tracking changes between posting and journal publication
Jessica K Polka, Gautam Dey, Máté Pálfy, Federico Nanni, Liam Brierley, Nicholas Fraser, Jonathon
Alexis Coates
Cold Spring Harbor Laboratory (2021-04-04) https://doi.org/gh5mhm
DOI: 10.1101/2021.02.20.432090
26. Textual Analysis in Accounting and Finance: A Survey
TIM LOUGHRAN, BILL MCDONALD
Journal of Accounting Research (2016-09) https://doi.org/gc3hf7
DOI: 10.1111/1475-679x.12123
27. The textual characteristics of traditional and Open Access scienti�c journals are similar
Karin Verspoor, K Bretonnel Cohen, Lawrence Hunter
BMC Bioinformatics (2009-06-15) https://doi.org/b973tn
DOI: 10.1186/1471-2105-10-183 · PMID: 19527520 · PMCID: PMC2714574
28. Current �ndings from research on structured abstracts
James Hartley
Journal of the Medical Library Association : JMLA (2004-07)
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC442180/
PMID: 15243644 · PMCID: PMC442180
29. A survey on annotation tools for the biomedical literature
M. Neves, U. Leser
Brie�ngs in Bioinformatics (2012-12-18) https://doi.org/f5vzsj
DOI: 10.1093/bib/bbs084 · PMID: 23255168
30. PubTator central: automated concept annotation for biomedical full text articles
Chih-Hsuan Wei, Alexis Allot, Robert Leaman, Zhiyong Lu
Nucleic Acids Research (2019-07-02) https://doi.org/ggzfsc
DOI: 10.1093/nar/gkz389 · PMID: 31114887 · PMCID: PMC6602571
31. Coreference annotation and resolution in the Colorado Richly Annotated Full Text (CRAFT)
corpus of biomedical journal articles
K. Bretonnel Cohen, Arrick Lanfranchi, Miji Joo-young Choi, Michael Bada, William A. Baumgartner,
Natalya Panteleyeva, Karin Verspoor, Martha Palmer, Lawrence E. Hunter
BMC Bioinformatics (2017-08-17) https://doi.org/ghmbw2
DOI: 10.1186/s12859-017-1775-9 · PMID: 28818042 · PMCID: PMC5561560
32. The structural and content aspects of abstracts versus bodies of full text journal articles
are di�erent
K Bretonnel Cohen, Helen L Johnson, Karin Verspoor, Christophe Roeder, Lawrence E Hunter
BMC Bioinformatics (2010-09-29) https://doi.org/b9f6rn
DOI: 10.1186/1471-2105-11-492 · PMID: 20920264 · PMCID: PMC3098079
33. A corpus of full-text journal articles is a robust evaluation tool for revealing di�erences in
performance of biomedical natural language processing tools
Karin Verspoor, Kevin Bretonnel Cohen, Arrick Lanfranchi, Colin Warner, Helen L Johnson,
Christophe Roeder, Jinho D Choi, Christopher Funk, Yuriy Malenkiy, Miriam Eckert, … Lawrence E
Hunter
BMC Bioinformatics (2012-08-17) https://doi.org/gb8t7v
DOI: 10.1186/1471-2105-13-207 · PMID: 22901054 · PMCID: PMC3483229
34. From POS tagging to dependency parsing for biomedical event extraction
Dat Quoc Nguyen, Karin Verspoor
BMC Bioinformatics (2019-02-12) https://doi.org/ggsrkw
DOI: 10.1186/s12859-019-2604-0 · PMID: 30755172 · PMCID: PMC6373122
35. E�cient Estimation of Word Representations in Vector Space
Tomas Mikolov, Kai Chen, Greg Corrado, Je�rey Dean
arXiv (2013-09-10) https://arxiv.org/abs/1301.3781
36. Distributed Representations of Sentences and Documents
Quoc V. Le, Tomas Mikolov
arXiv (2014-05-26) https://arxiv.org/abs/1405.4053
37. https://www.biorxiv.org/tdm
38. PubMed Central: The GenBank of the published literature
R. J. Roberts
Proceedings of the National Academy of Sciences (2001-01-16) https://doi.org/bbn9k8
DOI: 10.1073/pnas.98.2.381 · PMID: 11209037 · PMCID: PMC33354
39. How Papers Get Into PMC https://www.ncbi.nlm.nih.gov/pmc/about/submission-methods/
40. Gold open access: the best of both worlds
M. A. G. van der Heyden, T. A. B. van Veen
Netherlands Heart Journal (2017-12-01) https://doi.org/ggzfr9
DOI: 10.1007/s12471-017-1064-2 · PMID: 29196877 · PMCID: PMC5758455
41. 8.2.2 NIH Public Access Policy
https://grants.nih.gov/grants/policy/nihgps/html5/section_8/8.2.2_nih_public_access_policy.htm
42. PMC Overview https://www.ncbi.nlm.nih.gov/pmc/about/intro/
43. PMC text mining subset in BioC: about three million full-text articles and growing
Donald C Comeau, Chih-Hsuan Wei, Rezarta Islamaj Doğan, Zhiyong Lu
Bioinformatics (2019-09-15) https://doi.org/ggzfsb
DOI: 10.1093/bioinformatics/btz070 · PMID: 30715220 · PMCID: PMC6748740
44. Author Manuscripts in PMC https://www.ncbi.nlm.nih.gov/pmc/about/authorms/
45. The new york times annotated corpus
Evan Sandhaus
Linguistic Data Consortium, Philadelphia (2008)
46. CrossRef Text and Data Mining Services
Rachael Lammey
Insights the UKSG journal (2015-07-07) https://doi.org/gg4hp9
DOI: 10.1629/uksg.233
47. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural
networks and incremental parsing
Matthew Honnibal, Ines Montani
(2017)
48. Odds Ratio
Steven Tenny, Mary R. Ho�man
StatPearls (2021) http://www.ncbi.nlm.nih.gov/books/NBK431098/
49. Software Framework for Topic Modelling with Large Corpora
Radim Řehůřek, Petr Sojka
Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks (2010-05-22)
50. On the Dimensionality of Word Embedding
Zi Yin, Yuanyuan Shen
arXiv (2018-12-12) https://arxiv.org/abs/1812.04224
51. Probabilistic Principal Component Analysis
Michael E. Tipping, Christopher M. Bishop
Journal of the Royal Statistical Society: Series B (Statistical Methodology) (1999-08)
https://doi.org/b3hjwt
DOI: 10.1111/1467-9868.00196
52. Scikit-learn: Machine learning in Python
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P.
Prettenhofer, R. Weiss, V. Dubourg, … E. Duchesnay
Journal of Machine Learning Research (2011)
53. Finding structure with randomness: Probabilistic algorithms for constructing approximate
matrix decompositions
Nathan Halko, Per-Gunnar Martinsson, Joel A. Tropp
arXiv (2014-04-29) https://arxiv.org/abs/0909.4061
54. The Drosophila Cortactin Binding Protein 2 homolog, Nausicaa, regulates lamellipodial actin
dynamics in a Cortactin-dependent manner
Meghan E. O’Connell, Divya Sridharan, Tristan Driscoll, Ipsita Krishnamurthy, Wick G. Perry, Derek
A. Applewhite
Cold Spring Harbor Laboratory (2018-07-24) https://doi.org/gg4hp7
DOI: 10.1101/376665
55. The Drosophila protein, Nausicaa, regulates lamellipodial actin dynamics in a Cortactin-
dependent manner
Meghan E. O’Connell, Divya Sridharan, Tristan Driscoll, Ipsita Krishnamurthy, Wick G. Perry, Derek
A. Applewhite
Biology Open (2019-06-15) https://doi.org/gg4hp8
DOI: 10.1242/bio.038232 · PMID: 31164339 · PMCID: PMC6602326
56. Understanding survival analysis: Kaplan-Meier estimate
Jugal Kishore, ManishKumar Goel, Pardeep Khanna
International Journal of Ayurveda Research (2010) https://doi.org/fdft75
DOI: 10.4103/0974-7788.76794 · PMID: 21455458 · PMCID: PMC3059453
57. CamDavidsonPilon/lifelines: v0.25.6
Cameron Davidson-Pilon, Jonas Kalderstam, Noah Jacobson, Sean-Reed, Ben Kuhn, Paul Zivich,
Mike Williamson, AbdealiJK, Deepyaman Datta, Andrew Fiore-Gartland, … Jlim13
Zenodo (2020-10-26) https://doi.org/ghh2d3
DOI: 10.5281/zenodo.4136578
58. Medium – Where good ideas �nd you.
Medium
https://medium.com
59. Scikit-learn: Machine Learning in Python
Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier
Grisel, Mathieu Blondel, Andreas Müller, Joel Nothman, Gilles Louppe, … Édouard Duchesnay
arXiv (2018-06-06) https://arxiv.org/abs/1201.0490
60. Introduction — PyMuPDF 1.18.13 documentation
https://pymupdf.readthedocs.io/en/latest/intro.html
61. Assessing the Heterogeneity of Cardiac Non-myocytes and the E�ect of Cell Culture with
Integrative Single Cell Analysis
Brian S. Iskra, Logan Davis, Henry E. Miller, Yu-Chiao Chiu, Alexander R. Bishop, Yidong Chen,
Gregory J. Aune
Cold Spring Harbor Laboratory (2020-03-05) https://doi.org/gg9353
DOI: 10.1101/2020.03.04.975177
62. Preprinting the COVID-19 pandemic
Nicholas Fraser, Liam Brierley, Gautam Dey, Jessica K Polka, Máté Pálfy, Federico Nanni, Jonathon
Alexis Coates
Cold Spring Harbor Laboratory (2021-02-05) https://doi.org/dxdb
DOI: 10.1101/2020.05.22.111294
63. PMCID - PMID - Manuscript ID - DOI Converter https://www.ncbi.nlm.nih.gov/pmc/pmctopmid/
64. Altmetric Scores, Citations, and Publication of Studies Posted as Preprints
Stylianos Serghiou, John P. A. Ioannidis
JAMA (2018-01-23) https://doi.org/gftc69
DOI: 10.1001/jama.2017.21168 · PMID: 29362788 · PMCID: PMC5833561
65. E�cient Vector Representation for Documents through Corruption
Minmin Chen
arXiv (2017-07-11) https://arxiv.org/abs/1707.02377
66. Document Network Projection in Pretrained Word Embedding Space
Antoine Gourru, Adrien Guille, Julien Velcin, Julien Jacques
arXiv (2020-01-17) https://arxiv.org/abs/2001.05727
67. Conditional Robust Calibration (CRC): a new computational Bayesian methodology for
model parameters estimation and identi�ability analysis
Fortunato Bianconi, Chiara Antonini, Lorenzo Tomassoni, Paolo Valigi
Cold Spring Harbor Laboratory (2017-10-02) https://doi.org/gg9393
DOI: 10.1101/197400
68. Machine learning of stochastic gene network phenotypes
Kyemyung Park, Thorsten Prüstel, Yong Lu, John S. Tsang
Cold Spring Harbor Laboratory (2019-10-31) https://doi.org/gg94bm
DOI: 10.1101/825943
69. Notions of similarity for computational biology models
Ron Henkel, Robert Hoehndorf, Tim Kacprowski, Christian Knüpfer, Wolfram Liebermeister,
Dagmar Waltemath
Cold Spring Harbor Laboratory (2016-03-21) https://doi.org/gg939z
DOI: 10.1101/044818
70. GpABC: a Julia package for approximate Bayesian computation with Gaussian process
emulation
Evgeny Tankhilevich, Jonathan Ish-Horowicz, Tara Hameed, Elisabeth Roesch, Istvan Kleijn, Michael
PH Stumpf, Fei He
Cold Spring Harbor Laboratory (2019-09-18) https://doi.org/gg94bj
DOI: 10.1101/769299
71. SBpipe: a collection of pipelines for automating repetitive simulation and analysis tasks
Piero Dalle Pezze, Nicolas Le Novère
Cold Spring Harbor Laboratory (2017-02-09) https://doi.org/gg9392
DOI: 10.1101/107250
72. Spatiotemporal proteomics uncovers cathepsin-dependent host cell death during bacterial
infection
Joel Selkrig, Nan Li, Jacob Bobonis, Annika Hausmann, Anna Sueki, Haruna Imamura, Bachir El
Debs, Gianluca Sigismondo, Bogdan I. Florea, Herman S. Overkleeft, … Athanasios Typas
Cold Spring Harbor Laboratory (2018-11-07) https://doi.org/gg94bc
DOI: 10.1101/455048
73. Systems analysis by mass cytometry identi�es susceptibility of latent HIV-infected T cells to
targeting of p38 and mTOR pathways
Linda E. Fong, Victor L. Bass, Serena Spudich, Kathryn Miller-Jensen
Cold Spring Harbor Laboratory (2018-07-19) https://doi.org/gg9398
DOI: 10.1101/371922
74. NADPH consumption by L-cystine reduction creates a metabolic vulnerability upon glucose
deprivation
James H. Joly, Alireza Delfarah, Philip S. Phung, Sydney Parrish, Nicholas A. Graham
Cold Spring Harbor Laboratory (2019-08-13) https://doi.org/gg94bf
DOI: 10.1101/733162
75. Inhibition of Bruton’s tyrosine kinase reduces NF-kB and NLRP3 in�ammasome activity
preventing insulin resistance and microvascular disease
Gareth S. D. Purvis, Massimo Collino, Haidee M. A. Tavio, Fausto Chiazza, Caroline E. O’Riodan,
Lynda Zeboudj, Nick Guisot, Peter Bunyard, David R. Greaves, Christoph Thiemermann
Cold Spring Harbor Laboratory (2019-08-28) https://doi.org/gg94bg
DOI: 10.1101/745943
76. AKT but not MYC promotes reactive oxygen species-mediated cell death in oxidative culture
Dongqing Zheng, Jonathan H. Sussman, Matthew P. Jeon, Sydney T. Parrish, Alireza Delfarah,
Nicholas A. Graham
Cold Spring Harbor Laboratory (2019-09-01) https://doi.org/gg94bh
DOI: 10.1101/754572
77. FPtool a software tool to obtain in silico genotype-phenotype signatures and �ngerprints
based on massive model simulations
Guido Santos, Julio Vera
Cold Spring Harbor Laboratory (2018-02-18) https://doi.org/gjr9m9
DOI: 10.1101/266775
78. Bromodomain inhibition reveals FGF15/19 as a target of epigenetic regulation and
metabolic control
Chisayo Kozuka, Vicencia Sales, Soravis Osataphan, Yixing Yuchi, Jeremy Chimene-Weiss,
Christopher Mulla, Elvira Isganaitis, Jessica Desmond, Suzuka Sanechika, Joji Kusuyama, … Mary-
Elizabeth Patti
Cold Spring Harbor Laboratory (2019-12-12) https://doi.org/gjr9m8
DOI: 10.1101/2019.12.11.872887
79. Peer review and the publication process
Parveen Azam Ali, Roger Watson
Nursing Open (2016-03-16) https://doi.org/c4g8
DOI: 10.1002/nop2.51 · PMID: 27708830 · PMCID: PMC5050543
Supplemental Figures
Figure S1: Neuroscience and bioinformatics are the two most common author-selected topics for bioRxiv preprints.
Figure S2: Both classi�ers outperform the randomized baseline when predicting a paper’s journal endpoint. This
bargraph shows each model’s accuracy in respect to predicting the training and test set.
| 2021 | Linguistic Analysis of the bioRxiv Preprint Landscape | 10.1101/2021.03.04.433874 | [
"Nicholson David N.",
"Rubinetti Vincent",
"Hu Dongbo",
"Thielk Marvin",
"Hunter Lawrence E.",
"Greene Casey S."
] | creative-commons |
1
7 Tesla MRI of the ex vivo human brain at 100 micron resolution
1
2
Brian L. Edlow1,2, Azma Mareyam2, Andreas Horn3, Jonathan R. Polimeni2, M. Dylan
3
Tisdall4, Jean Augustinack2, Jason P. Stockmann2, Bram R. Diamond2, Allison
4
Stevens2, Lee S. Tirrell2, Rebecca D. Folkerth5, Lawrence L. Wald2, Bruce Fischl2,* &
5
Andre van der Kouwe2,*
6
* co-senior authors
7
8
1. Center for Neurotechnology and Neurorecovery, Massachusetts General Hospital,
9
Department of Neurology, Boston, MA 02114, USA
10
2. Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General
11
Hospital, Department of Radiology, Charlestown, MA 02129, USA
12
3. Movement Disorders & Neuromodulation Section, Department for Neurology,
13
Charité – University Medicine Berlin, Germany
14
4. Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia,
15
PA 19104, USA
16
5. City of New York Office of the Chief Medical Examiner, and New York University
17
School of Medicine, New York, NY, USA
18
19
corresponding author: Brian Edlow (bedlow@mgh.harvard.edu)
20
2
Abstract
21
We present an ultra-high resolution MRI dataset of an ex vivo human brain
22
specimen. The brain specimen was donated by a 58-year-old woman who
23
had no history of neurological disease and died of non-neurological causes.
24
After fixation in 10% formalin, the specimen was imaged on a 7 Tesla MRI
25
scanner at 100 µm isotropic resolution using a custom-built 31-channel
26
receive array coil. Single-echo multi-flip Fast Low-Angle SHot (FLASH) data
27
were acquired over 100 hours of scan time (25 hours per flip angle), allowing
28
derivation of a T1 parameter map and synthesized FLASH volumes. This
29
dataset
provides
an
unprecedented
view
of
the
three-dimensional
30
neuroanatomy of the human brain. To optimize the utility of this resource, we
31
warped the dataset into standard stereotactic space. We now distribute the
32
dataset in both native space and stereotactic space to the academic
33
community via multiple platforms. We envision that this dataset will have a
34
broad range of investigational, educational, and clinical applications that will
35
advance understanding of human brain anatomy in health and disease.
36
37
Design Type(s)
Single measure design
Measurement Type(s)
Nuclear magnetic resonance assay
Technology Type(s)
7 Tesla MRI scanner
Factor Type(s)
Sample Characteristic(s)
Homo sapiens • brain
38
3
Background & Summary
39
Postmortem ex vivo MRI provides significant advantages over in vivo MRI for
40
visualizing the microstructural neuroanatomy of the human brain. Whereas in
41
vivo MRI acquisitions are constrained by time (i.e. ~hours) and affected by
42
motion, ex vivo MRI can be performed without time constraints (i.e. ~days)
43
and without cardiorespiratory or head motion. The resultant advantages for
44
characterizing neuroanatomy at microscale are particularly important for
45
identifying cortical layers and subcortical nuclei1-5, which are difficult to
46
visualize even in the highest-resolution in vivo MRI datasets6,7. Ex vivo MRI
47
also provides advantages over histological methods that are associated with
48
distortion and tearing of human brain tissue during fixation, embedding, and
49
slide-mounting.
50
As the field of ex vivo MRI has developed over the past two decades,
51
several laboratories have focused on imaging blocks of tissue from human
52
brain specimens using small-bore scanners2,8 and specialized receive coils9-11.
53
This approach allows for spatial resolutions of up to 35–75 microns for
54
analyses of specific neuroanatomic regions9,11-13. However, ultra-high
55
resolution imaging of whole human brain specimens at high magnetic field
56
strengths has been far more challenging, due to the need for multi-channel
57
receive coils and large-bore clinical scanners that can accommodate a whole-
58
brain specimen. Whole-brain imaging is required to observe neuroanatomic
59
relationships across distant brain regions, as well as to provide a complete
60
view of human neuroanatomy in standard stereotactic space.
61
Here, we report the results of a multidisciplinary effort to image a whole
62
human brain specimen ex vivo at an unprecedented spatial resolution of 100
63
4
µm isotropic. Central to this effort was the construction of an integrated
64
system consisting of a custom-built 31-channel receive array coil and volume
65
transmit coil, which was designed to accommodate and tightly enclose an ex
66
vivo human brain14. The scans were performed on a 7 Tesla whole-body
67
human MRI scanner using four single-echo spoiled gradient-recalled echo
68
(SPGR/GRE) or Fast Low-Angle SHot (FLASH) sequences. We used varying
69
flip-angles (FA15°, FA20°, FA25°, FA30°) to generate multiple synthesized
70
volumes, each of which provides a different tissue contrast. The scans,
71
performed over ~100 hours (~25 hours per FA), generated an ~8 TB dataset
72
(~2 TB per flip angle) that required custom-built computational tools for offline
73
MRI reconstruction and creation of the synthesized volumes. Offline MRI
74
reconstruction considerably reduces the data amount. We release the
75
resulting FA25° acquisition, as well as the synthesized FLASH25 volume here,
76
both in native space and coregistered to standard stereotactic space, for use
77
by the academic community. We envision a broad range of investigational,
78
educational, and clinical applications for this dataset that have the potential to
79
advance understanding of human brain anatomy in health and disease.
80
81
Methods
82
Specimen acquisition and processing
83
A 58-year-old woman with a history of lymphoma and stem cell
84
transplantation, but no history of neurological or psychiatric disease, died in a
85
medical intensive care unit. She was initially admitted to the hospital for
86
fevers, chills, and fatigue, and then was transferred to the intensive care unit
87
for hypoxic respiratory failure requiring mechanical ventilation. Her hospital
88
course was also notable for a deep venous thromboses and a pulmonary
89
5
embolism. The cause of her death on hospital day 15 was determined to be
90
hypoxic respiratory failure due to viral pneumonia. At the time of her death,
91
her surrogate decision-maker provided written informed consent for a clinical
92
autopsy and for donation of her brain for research, as part of a protocol
93
approved by our Institutional Review Board.
94
At autopsy, her fresh brain weighed 1,210 grams (normal range = 1,200
95
to 1,500 grams). The brain was fixed in 10% formalin 14 hours after death.
96
Gross examination revealed a normal brain (Fig. 1), without evidence of mass
97
lesions or cerebrovascular disease. To ensure adequate fixation and prevent
98
specimen flattening (which can prevent specimens from fitting into custom ex
99
vivo MRI coils), we followed a series of standard specimen processing
100
procedures, as previously described15.
101
102
Specimen preparation for scanning
103
After remaining in fixative for 35 months, the brain specimen was transferred
104
to Fomblin Y LVAC 06/6 (perfluoropolyether, Solvay Specialty Polymers USA,
105
LLC, West Deptford, NJ), which is invisible to MR and reduces magnetic
106
susceptibility artifacts. The specimen, immersed in Fomblin, was then
107
secured inside a custom-built, air-tight brain holder made of rugged
108
urethane16. The brain holder contains degassing ports for removal of air
109
bubbles, which further reduces magnetic susceptibility artifacts.
110
111
Construction of a receive array coil and transmit volume coil for ex vivo
112
imaging of the whole human brain
113
We built a receive coil apparatus consisting of a 31-channel surface coil loop
114
array with two halves. The apparatus was fabricated using a 3D printer of
115
6
slightly larger dimensions than the brain holder, which slides inside the single-
116
channel birdcage volume transmit coil (Fig. 2). The brain holder is an oblate
117
spheroid (16 × 19 cm) that conforms to the shape of a whole brain (cerebral
118
hemispheres + cerebellum + brainstem)16 (Fig. 2d). It is made of two separate
119
halves that can be sealed together with a silicone gasket after packing the
120
brain inside. This holder must also withstand the degassing process when
121
under vacuum pressure. Degassing is performed in three steps: 1) introducing
122
vacuum suction into the container with the brain inside, which allows the
123
bubbles to expand under decreased pressure and exit tissue cavities; 2)
124
opening the valve to fill the holder with fomblin and then sealing off the fill
125
valve; and 3) continuation of vacuum suction with low-amplitude vibration of
126
the holder for 2-6 hours. The vibration facilitates the removal of bubbles from
127
tissue cavities. All three steps are performed inside a fume hood.
128
The coil former (Fig. 2c) consists of two halves and encloses the brain
129
holder. The receive array coil consists of 31 detectors (Fig. 2a), with 15
130
elements on the top half (diameter = 5.5 cm) and 16 on the bottom half
131
(diameter = 8.5 cm). Coil elements were constructed using 16 AWG wire
132
loops 17, each with four or five evenly spaced capacitors (Supplementary Fig.
133
1). All elements were tuned to 297.2 MHz and matched to a loaded
134
impedance of 75 Ω to minimize preamplifier noise. Preamplifier decoupling
135
was achieved with a cable length of 6 cm. Preamplifiers were placed directly
136
on the coil elements, yielding a substantial reduction in cable losses compared
137
to a previous 30-channel ex vivo brain array18. The active detuning circuit was
138
formed across the match capacitor using an inductor and PIN diode.
139
Tuning, matching, and decoupling of neighboring elements was
140
optimized on the bench with a brain sample immersed in periodate-lysine-
141
7
paraformaldehyde (PLP) solution. Because coil loading varies with the fixative
142
used, the coil must be tuned and matched on the bench using a brain sample
143
with the correct fixative. (For example, testing can be performed with a brain
144
sample immersed in PLP or formalin, but not the regular loading phantom
145
comprised of water and salt). Loops tuned/matched on PLP showed
146
unloaded-to-loaded quality factor ratio (Q-ratio) of QUL/QL = 210/20 = 10.5,
147
corresponding to an equivalent noise resistance of 11 ohms for the loaded coil
148
(Q = wL/R). By contrast, formalin is a less lossy fixative, giving a coil Q-ratio
149
of QUL/QL = 210/60 = 3.5, corresponding to an equivalent noise resistance of 4
150
ohms.
151
A shielded detunable volume coil (Fig. 2) was built for excitation, with
152
the following parameters and features: band-pass birdcage, diameter 26.7 cm,
153
and an extended length of 32 cm to accommodate brain samples of larger
154
dimensions. For the detuning circuit we used diodes in every leg of the
155
birdcage. These diodes are powered with the high-power chokes, which can
156
withstand high voltage and short duration inversion pulses.
157
In summary, this coil system incorporates an improved mechanical
158
design, preamps mounted at the coil detectors, and an extended transmit coil
159
design capable of producing high-power pulses.
160
161
7 Tesla MRI data acquisition
162
The brain specimen was scanned on a whole-body human 7 Tesla (7T)
163
Siemens Magnetom MRI scanner (Siemens Healthineers, Erlangen,
164
Germany) with the custom-built coil described above. We utilized a GRE
165
sequence19 at 100 µm isotropic spatial resolution with the following acquisition
166
parameters: TR = 40 msec, TE = 14.2 msec, bandwidth = 90 Hz/px, FA = 15˚,
167
8
20˚, 25˚, 30˚. Total scan time for each FA was 25:01:52 [hh:mm:ss], and each
168
FA acquisition generated 1.98 TB of raw k-space data. To improve the signal-
169
to-noise ratio (SNR) and optimise T1 modelling, we collected FLASH scans at
170
four FAs: 15°, 20°, 25°, 30° (Fig. 3). Accounting for localizers, quality
171
assurance (QA) scans, and adjustment scans, the total scan time was 100
172
hours and 8 minutes, and we collected nearly 7.92 TB of raw k-space data.
173
174
MRI data reconstruction
175
The size of the k-space data exceeded the storage capacity of the RAID
176
provided by the scanner image reconstruction computer. The image
177
reconstruction also required more RAM than what was available. We
178
therefore implemented software on the scanner to stream the data directly
179
via TCP/IP to a server on an external computer added to the scanner network,
180
which saved the data as they were received. Because of additional limitations
181
related to the total size of the raw data for any single scan, as dictated by the
182
imager RAID size, we also divided each acquisition into segments. The
183
server on the external computer stored the data as they were acquired,
184
creating date stamps for every k-space segment.
185
After the scan was completed, the streamed k-space data were
186
transferred to a computational server where we ran custom software to stitch
187
together the segments, reconstruct the images for each channel (via a 3D FFT
188
on each volume per channel20), and combine the images derived from the 31
189
channels via the root-sum-of-squares of the signal magnitudes at each voxel.
190
These signal magnitudes were channel-wise decorrelated using a covariance
191
matrix of the channels’ thermal noise. The output from coil combination was
192
the final acquired image (Data Citation 1; Videos 1, 2 and 3).
193
9
194
MRI data processing
195
The acquired data underwent a series of processing steps, culminating in the
196
creation of a T1 parameter map and synthesized FLASH volumes (Fig. 3 and
197
Fig. 4; Videos 4, 5, and 6; Data Citations 1 and 2). The volumes were
198
estimated directly from the four FLASH acquisitions using the DESPOT1
199
algorithm19,21 with the program ‘mri_ms_fitparms’ distributed in FreeSurfer
200
(http://surfer.nmr.mgh.harvard.edu) to quantify tissue properties independent
201
of scanner and sequence types. This algorithm fits the tissue parameters (i.e.
202
T1) of the signal equation for the FLASH scan at each voxel using multiple
203
input volumes. The volumes at the originally acquired TRs and flip angles
204
were then regenerated from the parameter maps by evaluating the FLASH
205
signal equation. In principle, a volume with any TR and flip angle combination
206
could be synthesized. These synthesized volumes are created from all the
207
acquired data, and therefore have better SNR than the individually acquired
208
input volumes. We choose to release the 25 degree synthetic volume as it
209
has maximal SNR and the best apparent contrast for cortical and subcortical
210
structures9.
211
Of note, ex vivo MRI of the fixed human brain yields a different contrast
212
than in vivo MRI, mainly from a shortened T1, but also from a decrease in T2
*,
213
both of which are related to formalin fixation22. The predominant source of
214
signal contrast in ex vivo MRI is likely myelin23 and/or iron24. Specifically,
215
myelin appears to be a source of T1 contrast, while cortical iron appears to be
216
a source of T2
* contrast25.
217
218
Coregistration of the dataset to standard stereotactic space
219
10
The dataset was spatially normalized into the MNI ICBM 2009b NLIN ASYM
220
template26 (Supplementary Fig. 2a). This template constitutes the newest
221
version of the “MNI space” and is considered a high-resolution version of MNI
222
space because it is available at 0.5 mm isotropic resolution. To combine
223
structural information present on T1 and T2 versions of the template, we
224
created a joint template using PCA, as previously described27. The four
225
synthesized FLASH volumes (FA15, FA20, FA25, and FA30) were
226
downsampled to isotropic voxel-sizes of 0.5 mm for spatial normalization and
227
initially registered into template space in a multispectral approach using
228
Advanced Normalization Tools (ANTs; http://stnava.github.io/ANTs/; 28). This
229
multispectral approach simultaneously accounts for intensity data in all four
230
volumes. The initial normalization was performed in four stages (rigid body,
231
affine, whole brain SyN and subcortically focused SyN) as defined in the
232
“effective: low variance + subcortical refine” preset implemented in Lead-DBS
233
software (www.lead-dbs.org; 29).
234
To refine the warp, we introduced fiducial regions of interest (ROI)
235
iteratively using a tool developed for this task (available within Lead-DBS).
236
Specifically, we manually drew line and point fiducial markers in both native
237
and template spaces (Supplementary Fig. 2b). In addition, we manually
238
segmented four structures in native space (subthalamic nucleus, internal and
239
external pallidum and red nucleus). The three types of fiducials (line ROI,
240
spherical ROI and manual segmentations of key structures) were then added
241
as “spectra” in subsequent registration refinements (Supplementary Fig. 2c).
242
Thus, the final registration consisted of a large number of pairings between
243
native and template space (the first four being the actual anatomical volumes,
244
the subsequent ones being manual segmentations and paired helper
245
11
fiducials). To achieve maximal registration precision, the warp was refined in
246
over 30 iterations with extensive manual expert interaction, each refinement
247
continuing directly from the last saved state. We used linear interpolation to
248
create the normalized files in the data release (Data Citations 1 and 3).
249
250
Code availability
251
Neuroimaging data were processed using standard processing pipelines
252
(http://surfer.nmr.mgh.harvard.edu/,
https://github.com/freesurfer/freesurfer).
253
All code used for registration of volumes into standard stereotactic space are
254
available
within
the
open-source
Lead-DBS
software
255
(https://github.com/leaddbs/leaddbs). Because registration involved multiple
256
manual user interface steps, no ready-made code is provided, but the process
257
can be readily reproduced with the provided data and software.
258
259
Data Records
260
The native space FA25˚ acquisition and synthesized FLASH25 volume are
261
available for download at https://datadryad.org (Data Citation 1). Additional
262
synthesized volumes are available upon request to the corresponding author.
263
Axial, coronal, and sagittal videos of the native space FA25˚ acquisition
264
(Videos 1, 2, and 3) and synthesized FLASH25 volume (Videos 4, 5, and 6)
265
are also available at the Dryad data repository (Data Citation 1). The
266
synthesized FLASH25 volume is available for interactive, online viewing at
267
https://histopath.nmr.mgh.harvard.edu (Data Citation 2). The normalized
268
FLASH25 volume in standard stereotactic space is available at the Dryad data
269
repository (Data Citation 1) and is hosted on www.lead-dbs.org (preinstalled
270
as part of the LEAD-DBS software package; Data Citation 3).
271
12
272
Technical Validation
273
Coil signal-to-noise ratio (SNR) measurements
274
The receive coil has a QUL/QL ratio that ranged from 6 in the top half elements
275
to 8 in the bottom half elements due to larger coil diameter. The S12 coupling
276
between neighbouring elements, measured with all other coils active detuned,
277
ranged from −10.9 to −24 dB. All individual elemnts had S11 < −20 dB and
278
active detuning of > 30 dB. We evaluated the performance of the transmit coil
279
by examining the B1
+ profile14, which shows the efficiency throughout the
280
entire spatial distribution of the brain specimen. The efficiency was greatest in
281
the center of the specimen and fell off gradually towards the edges, as
282
expected for a whole brain specimen at 7T.
283
We compared the SNR of the 31-channel ex vivo array to that of a
284
standard 31-channel 7T head coil and a 64-channel 3T head coil. SNR maps
285
were computed following the method of Kellman & McVeigh30. We calibrated
286
the voltage required for 180˚ pulse using a B1
+ map (estimated with the AFI
287
method)31 with an ROI of 3-cm diameter at the center of the brain. We
288
estimated array noise covariance from thermal noise data acquired without RF
289
excitation. The SNR gain with the 31-channel ex vivo array was 1.6-fold
290
versus the 31-channel 7T standard coil and 3.3-fold versus the 64-channel 3T
291
head array (Fig. 5). The noise coupling between channels was 11% for the
292
31-channel ex vivo array, a 2-fold improvement relative to our previous
293
array18.
294
295
Coregistration accuracy
296
13
We assessed the neuroanatomic accuracy of the final registration results (i.e.
297
the fit between structures on the normalized FLASH volumes versus the high-
298
resolution MNI template) by visual inspection using a tool specifically designed
299
for this task (implemented in Lead-DBS). An example of this visual inspection
300
assessment for the subthalamic nucleus and globus pallidus interna is
301
provided in Supplementary Fig. 3. The final maps are stored in NIfTI and mgz
302
files in isotropic 150 μm resolution (Data Citation 1). The normalized
303
FLASH25 volume is additionally distributed pre-installed within Lead-DBS
304
software and can be selected for visualization in the 3D viewer (Data Citation
305
3). Fig. 6 shows an example in synopsis with deep brain stimulation electrode
306
reconstructions in a hypothetical patient being treated for Parkinson’s
307
Disease.
308
14
Acknowledgements
309
We thank Michelle Siciliano and Terrence Ott for assistance in obtaining and
310
processing the brain specimen. We thank Simon Sigalovsky for assistance
311
with coil construction, and Gunjan Madan for assistance with coil testing and
312
evaluation. We thank L. Daniel Bridgers for constructing the brain container and
313
coil array housing. We thank Andrew Hoopes for assistance with creation of
314
visual media. This work was supported by the NIH National Institute for
315
Neurological Disorders and Stroke (K23-NS094538, R01-NS052585, R21-
316
NS072652, R01-NS070963, R01-NS083534, U01-NS086625), the National
317
Institute for Biomedical Imaging and Bioengineering (P41-EB015896, R01-
318
EB006758,
R21-EB018907,
R01-EB019956,
R01-EB023281,
R00-
319
EB021349), the National Institute on Aging (R01-AG057672, R01-AG022381,
320
R01-AG008122, R01-AG016495, R01-AG008122, U01-AG006781, R21-
321
AG046657, P41-RR014075, P50-AG005136), the National Center for
322
Alternative Medicine (RC1-AT005728), the Eunice Kennedy Shriver National
323
Institute of Child Health and Human Development (K01-HD074651, R01-
324
HD071664, R00-HD074649), and the Centers for Disease Control and
325
Prevention (R49-CE001171). This research also utilized resources provided
326
by the National Center for Research Resources (U24-RR021382), Additional
327
support was provided by the NIH Blueprint for Neuroscience Research (U01-
328
MH093765), as part of the multi-institutional Human Connectome Project.
329
This research also utilized resources provided by National Institutes of Health
330
shared instrumentation grants S10-RR023401, S10-RR019307, and S10-
331
RR023043. Additional support for this project comes from the James S.
332
McDonnell Foundation, Rappaport Foundation, the Tiny Blue Dot Foundation
333
15
as well as the German Research Foundation (Emmy Noether Grant
334
410169619).
335
336
16
Author contributions
337
B.L.E. designed the study, analyzed the data, and prepared the manuscript.
338
A.M. built the coil, acquired and analyzed the data, and contributed to the
339
manuscript.
340
A.H. created the warp from native space to standard stereotactic space,
341
performed the coregistration for Lead-DBS implementation, and contributed to
342
the manuscript.
343
J.R.P. designed the study, acquired and analyzed the data, and contributed to
344
the manuscript.
345
M.D.T. acquired and analyzed the data, and contributed to the manuscript.
346
J.A. designed the study, acquired and analyzed the data, and contributed to
347
the manuscript.
348
J.P.S. advised on the building and testing of the coil, and contributed to the
349
manuscript.
350
B.R.D. analyzed the data and contributed to the manuscript.
351
A.S. acquired and analyzed the data, and contributed to the manuscript.
352
L.S.T. processed and analyzed the data, and contributed to the manuscript.
353
R.D.F. performed the pathological assessment and contributed to the
354
manuscript.
355
L.L.W. supervised the building of the coil and contributed to the manuscript.
356
B.F. supervised and designed the study, analyzed the data, and contributed to
357
the manuscript.
358
A.v.d.K. supervised and designed the study, acquired and analyzed the data,
359
and contributed to the manuscript.
360
17
Additional Information
361
Competing interests
362
None of the authors has a conflicting financial interest. Dr. Fischl and Mr.
363
Tirrell have financial interest in CorticoMetrics, a company whose medical
364
pursuits focus on brain imaging and measurement technologies. Their
365
interests were reviewed and are managed by Massachusetts General Hospital
366
and Partners HealthCare in accordance with their conflict of interest policies.
367
18
Figures
368
369
Figure 1. Human brain specimen. The human brain specimen that
370
underwent ex vivo MRI is shown from inferior (a), superior (b), right lateral (c)
371
and left lateral (d) perspectives. Gross pathological examination of the brain
372
was normal.
373
374
Figure 2. Receive array coil and transmit volume coil for ex vivo imaging
375
of the whole human brain. (a) The 31-channel receive array has 15
376
elements on the top half (with a diameter of 5.5 cm) and 16 on the bottom half
377
(with a diameter of 8.5 cm), each made of 16 AWG wire loops with four or five
378
evenly spaced capacitors. All elements are tuned to 297.2 MHz. (c) The coil
379
former has slightly larger dimensions than the brain holder, which slides inside
380
a volume coil (b). (d) A custom air-tight brain holder was designed to conform
381
to the shape of a whole human brain. The brain holder is an oblate spheroid
382
container (16 x 19 cm) with degassing ports that are used to apply a vacuum
383
suction, thereby reducing air bubbles in the specimen and surrounding fomblin
384
solution.
385
386
Figure 3. Comparison of FA25˚ acquisition and synthesized FLASH25
387
volume. Representative images from the FA25˚ acquisition (left column) and
388
the synthesized FLASH25 volume (right column) are displayed in the sagittal
389
(top row), coronal (middle row) and axial (bottom row) planes. These images
390
provide a qualitative comparison of the respective signal-to-noise properties of
391
19
the FA25˚ acquisition (~25 hours) and the synthesized FLASH25 volume
392
(~100 hours). All images are shown in radiologic convention.
393
394
Figure 4. Delineation of subcortical neuroanatomy. Representative axial
395
sections from the synthesized FLASH25 volume are shown at the level of the
396
rostral pons and caudal midbrain (a-c, see inset in panel c). Zoomed views of
397
the brainstem, medial temporal lobe, and anterior cerebellum (within the white
398
rectangles in a-c) are shown in the bottom row (d-f). The anatomic detail that
399
can be visualized in this ex vivo 100 μm resolution MRI dataset is beyond that
400
which can be seen in typical in vivo MRI datasets. All images are shown in
401
radiologic convention. Neuroanatomic abbreviations: Amg = amygdala; Cb =
402
cerebellum; CP = cerebral peduncle; MB = mammillary body; P = pons; SCP =
403
superior cerebellar peduncle; VTA = ventral tegmental area; xSCP =
404
decussation of the superior cerebellar peduncle; Th = thalamus.
405
406
Figure 5. Signal-to-noise ratio (SNR) analysis of coil performance.
407
Representative SNR maps are shown in the sagittal (top row), coronal (middle
408
row) and axial (bottom row) planes for a test brain sample immersed in
409
periodate-lysine-paraformaldehyde. The maps demonstrate an SNR gain of
410
1.6-fold for the 31-channel 7 Tesla (7T) ex vivo coil (left column) compared to
411
the 31-channel 7T standard coil (middle row), and a gain of 3.3-fold compared
412
to the 64-channel 3T head coil (right column). The noise coupling between
413
channels was 11% for the 31-channel ex vivo coil array, a 2-fold improvement
414
relative to our previous array18.
415
416
20
Figure 6. Normalization of the ex vivo MRI dataset into standard
417
stereotactic space and integration into the Lead-DBS software platform.
418
(a) Exemplary use-case of the normalized FLASH25 volume in deep brain
419
stimulation (DBS). DBS electrodes are visualized for a hypothetical patient
420
using Lead-DBS software (https://www.lead-dbs.org)29. An axial image from
421
the normalized scan, at the level of the rostral midbrain, is shown as a
422
backdrop, with 3D-structures defined by the DISTAL atlas32 (right subthalamic
423
and left red nucleus hidden for optimal visualization of the underlying
424
anatomy). Panels (b) and (c) show zoomed views of key DBS target regions:
425
the left globus pallidus interna (GPi in b) and the subthalamic nucleus (STN in
426
c). The images in (b) and (c) are shown in radiologic convention.
427
21
References
428
1
Augustinack, J. C., van der Kouwe, A. J. & Fischl, B. Medial temporal
429
cortices in ex vivo magnetic resonance imaging. J Comp Neurol 521,
430
4177-4188, doi:10.1002/cne.23432 (2013).
431
2
Edlow, B. L. et al. Neuroanatomic connectivity of the human ascending
432
arousal system critical to consciousness and its disorders. J
433
Neuropathol
Exp
Neurol
71,
531-546,
434
doi:10.1097/NEN.0b013e3182588293 (2012).
435
3
McNab, J. A. et al. The Human Connectome Project and beyond: Initial
436
applications of 300mT/m gradients. Neuroimage 80, 234-245,
437
doi:10.1016/j.neuroimage.2013.05.074 (2013).
438
4
Roebroeck, A., Miller, K. L. & Aggarwal, M. Ex vivo diffusion MRI of the
439
human brain: Technical challenges and recent advances. NMR Biomed
440
32, e3941, doi:10.1002/nbm.3941 (2019).
441
5
Aggarwal, M. et al. Feasibility of creating a high-resolution 3D diffusion
442
tensor imaging based atlas of the human brainstem: a case study at
443
11.7
T.
Neuroimage
74,
117-127,
444
doi:10.1016/j.neuroimage.2013.01.061 (2013).
445
6
Lusebrink, F., Sciarra, A., Mattern, H., Yakupov, R. & Speck, O. T1-
446
weighted in vivo human whole brain MRI dataset with an ultrahigh
447
isotropic
resolution
of
250
mum.
Sci
Data
4,
170032,
448
doi:10.1038/sdata.2017.32 (2017).
449
7
Horn, A. et al. Teaching NeuroImages: In vivo visualization of Edinger
450
comb
and
Wilson
pencils.
Neurology
92,
e1663-e1664,
451
doi:10.1212/WNL.0000000000007252 (2019).
452
22
8
Takahashi, E., Song, J. W., Folkerth, R. D., Grant, P. E. &
453
Schmahmann, J. D. Detection of postmortem human cerebellar cortex
454
and white matter pathways using high angular resolution diffusion
455
tractography:
a
feasibility
study.
Neuroimage
68,
105-111,
456
doi:10.1016/j.neuroimage.2012.11.042 (2013).
457
9
Augustinack, J. C. et al. Detection of entorhinal layer II using 7Tesla
458
[corrected] magnetic resonance imaging. Ann Neurol 57, 489-494,
459
doi:10.1002/ana.20426 (2005).
460
10
Gangolli, M. et al. Quantitative validation of a nonlinear histology-MRI
461
coregistration method using generalized Q-sampling imaging in
462
complex human cortical white matter. Neuroimage 153, 152-167,
463
doi:10.1016/j.neuroimage.2017.03.059 (2017).
464
11
van Veluw, S. J. et al. Microbleed and microinfarct detection in amyloid
465
angiopathy: a high-resolution MRI-histopathology study. Brain 139,
466
3151-3162, doi:10.1093/brain/aww229 (2016).
467
12
Sengupta, S. et al. High resolution anatomical and quantitative MRI of
468
the entire human occipital lobe ex vivo at 9.4T. Neuroimage 168, 162-
469
171, doi:10.1016/j.neuroimage.2017.03.039 (2018).
470
13
van der Kouwe, A. et al. High Resolution Structural and Diffusion MRI
471
of Ex Vivo Human Motor Cortex. Organization for Human Brain
472
Mapping (2011).
473
14
Mareyam, A. et al. Array coil and sample preparation and support
474
system for whole brain ex vivo imaging at 100 μm. ISMRM. #3130
475
(2015).
476
15
Edlow, B. L. et al. Multimodal Characterization of the Late Effects of
477
Traumatic Brain Injury: A Methodological Overview of the Late Effects
478
23
of Traumatic Brain Injury Project. J Neurotrauma 35, 1604-1619,
479
doi:10.1089/neu.2017.5457 (2018).
480
16
Bridgers, L. D. Design and Manufacture of an Ultra-High Field Ex Vivo
481
Coil Assembly, Massachusetts Institute of Technology, (2012).
482
17
Keil, B. et al. Size-optimized 32-channel brain arrays for 3 T pediatric
483
imaging. Magn Reson Med 66, 1777-1787, doi:10.1002/mrm.22961
484
(2011).
485
18
Mareyam, A., Polimeni, J. R., Alagappan, V., Fischl, B. & Wald, L. L. A
486
30 channel receive-only 7T array for ex vivo brain hemisphere imaging.
487
ISMRM. #106 (2009).
488
19
Fischl, B. et al. Sequence-independent segmentation of magnetic
489
resonance
images.
Neuroimage
23
Suppl
1,
S69-84,
490
doi:10.1016/j.neuroimage.2004.07.016 (2004).
491
20
Frigo, M. & Johnson, S. G. The Design and Implementation of FFTW3.
492
Proceedings of the IEEE 93, 216-231 (2005).
493
21
Deoni, S. C., Peters, T. M. & Rutt, B. K. High-resolution T1 and T2
494
mapping of the brain in a clinically acceptable time with DESPOT1 and
495
DESPOT2. Magn Reson Med 53, 237-241, doi:10.1002/mrm.20314
496
(2005).
497
22
Tovi, M. & Ericsson, A. Measurements of T1 and T2 over time in
498
formalin-fixed human whole-brain specimens. Acta Radiol 33, 400-404
499
(1992).
500
23
Eickhoff, S. et al. High-resolution MRI reflects myeloarchitecture and
501
cytoarchitecture of human cerebral cortex. Hum Brain Mapp 24, 206-
502
215, doi:10.1002/hbm.20082 (2005).
503
24
24
Fukunaga, M. et al. Layer-specific variation of iron content in cerebral
504
cortex as a source of MRI contrast. Proc Natl Acad Sci U S A 107,
505
3834-3839, doi:10.1073/pnas.0911177107 (2010).
506
25
Stuber, C. et al. Myelin and iron concentration in the human brain: a
507
quantitative study of MRI contrast. Neuroimage 93 Pt 1, 95-106,
508
doi:10.1016/j.neuroimage.2014.02.026 (2014).
509
26
Fonov, V. S., Evans, A. C., McKinstry, R. C., Almli, C. R. & Collins, D.
510
Unbiased nonlinear average age-appropriate brain templates from birth
511
to adulthood. Neuroimage 47, S102 (2009).
512
27
Horn,
A.
PCA
MNI
2009b
NLIN
template.
(2017).
513
<doi:10.6084/m9.figshare.4644472.v2>.
514
28
Avants, B. B., Epstein, C. L., Grossman, M. & Gee, J. C. Symmetric
515
diffeomorphic image registration with cross-correlation: evaluating
516
automated labeling of elderly and neurodegenerative brain. Medical
517
image analysis 12, 26-41, doi:10.1016/j.media.2007.06.004 (2008).
518
29
Horn, A. et al. Lead-DBS v2: Towards a comprehensive pipeline for
519
deep
brain
stimulation
imaging.
Neuroimage
184,
293-316,
520
doi:10.1016/j.neuroimage.2018.08.068 (2019).
521
30
Kellman, P. & McVeigh, E. R. Image reconstruction in SNR units: a
522
general method for SNR measurement. Magn Reson Med 54, 1439-
523
1447, doi:10.1002/mrm.20713 (2005).
524
31
Yarnykh, V. L. Actual flip-angle imaging in the pulsed steady state: a
525
method for rapid three-dimensional mapping of the transmitted
526
radiofrequency
field.
Magn
Reson
Med
57,
192-200,
527
doi:10.1002/mrm.21120 (2007).
528
25
32
Ewert, S. et al. Toward defining deep brain stimulation targets in MNI
529
space: A subcortical atlas based on multimodal MRI, histology and
530
structural
connectivity.
Neuroimage
170,
271-282,
531
doi:10.1016/j.neuroimage.2017.05.015 (2018).
532
533
26
Data Citations
534
1. Edlow, B.L., Mareyam, A., Horn, A., Polimeni, J.R., Tisdall M.D.,
535
Augustinack, J., Stockmann, J.P., Diamond B.R., Stevens, A., Tirrell, L.,
536
Folkerth, R.D., Wald, L.L., Fischl, B. & Kouwe, A.v.d. Dryad Digital
537
Repository. https://datadryad.org. doi:10.5061/dryad.119f80q (2019).
538
539
2. Edlow, B.L., Mareyam, A., Horn, A., Polimeni, J.R., Tisdall M.D.,
540
Augustinack, J., Stockmann, J.P., Diamond B.R., Stevens, A., Tirrell, L.,
541
Folkerth, R.D., Wald, L.L., Fischl, B. & Kouwe, A.v.d. Biolucida
542
https://histopath.nmr.mgh.harvard.edu (2019).
543
544
3. Edlow, B.L., Mareyam, A., Horn, A., Polimeni, J.R., Tisdall M.D.,
545
Augustinack, J., Stockmann, J.P., Diamond B.R., Stevens, A., Tirrell, L.,
546
Folkerth, R.D., Wald, L.L., Fischl, B. & Kouwe, A.v.d. Lead-DBS
547
https://www.lead-dbs.org (2019).
548
549
27
Videos
550
551
Video 1. Axial images from the FA25˚ acquisition. These images were
552
acquired in ~25 hours of scan time. The images are shown in radiologic
553
convention.
554
555
Video 2. Coronal images from the FA25˚ acquisition. These images were
556
acquired in ~25 hours of scan time. The images are shown in radiologic
557
convention.
558
559
Video 3. Sagittal images from the FA25˚ acquisition. These images were
560
acquired in ~25 hours of scan time.
561
562
Video 4. Axial images from the synthesized FLASH25 volume. These
563
images were acquired in ~100 hours of scan time. The images are shown in
564
radiologic convention.
565
566
Video 5. Coronal images from the synthesized FLASH25 volume. These
567
images were acquired in ~100 hours of scan time. The images are shown in
568
radiologic convention.
569
570
Video 6. Sagittal images from the synthesized FLASH25 volume. These
571
images were acquired in ~100 hours of scan time.
572
rt"
A
err dl
ie |
aN
yi oa
q a ;
a
PL LS
—
@ a
Synthesized FLASH25
(~100 hours)
Ee gy ee
eu. F5
we WE ON
fF AZAETMAT
ex vivo coil standard coil head coil
~™s
—
ra
120
100
80
Un
60
CAUDATE
EXTERNAL
Y ERNAL
|
PUTAMEN
ade “Se
sh . NUCLEUS ; WP
rs BU a0 NaN COMB
-. el Ua x SYSTEM
~\_ CAUDATE
7.
| 2019 | 7 Tesla MRI of the human brain at 100 micron resolution | 10.1101/649822 | [
"Edlow Brian L.",
"Mareyam Azma",
"Horn Andreas",
"Polimeni Jonathan R.",
"Tisdall M. Dylan",
"Augustinack Jean",
"Stockmann Jason P.",
"Diamond Bram R.",
"Stevens Allison",
"Tirrell Lee S.",
"Folkerth Rebecca D.",
"Wald Lawrence L.",
"Fischl Bruce",
"van der Kouwe Andre"
] | creative-commons |
1
Transcription Co-Factor LBH Is Necessary for Maintenance of Stereocilia Bundles and
Survival of Cochlear Hair Cells
Huizhan Liu1#, Kimberlee P. Giffen1#, Grati M’Hamed2, Seth W. Morrill1, Yi Li1,3
Xuezhong Liu2, Karoline J. Briegel4*, David Z. He1*
1Department of Biomedical Sciences, Creighton University School of Medicine, Omaha, Nebraska
68178
2Department of Otorhinolaryngology-Head and Neck Surgery, University of Miami Miller School
of Medicine, Miami, Florida 33136
3Department of Otorhinolaryngology-Head and Neck Surgery, Beijing Tongren Hospital, Capital
Medical University, Beijing, China
4Department of Surgery, University of Miami Miller School of Medicine, Miami, Florida 33136
# these authors contribute equally
* Correspondence:
Karoline Briegel: KBriegel@med.miami.edu
David He: hed@creighton.edu
Conflict of interest statement: The authors declare no conflict of interest.
Acknowledgments
This work has been supported by the NIH grants R01DC016807 from the NIDCD to DH, R01
GM113256 from the NIGMS to KJB and R01DC005575 and R01DC012115 from the NIDCD to
XL. YL is supported by National Science Foundation of China (#81600798 and #81770996). We
acknowledge the use of the University of Nebraska DNA Sequencing Core Facility for performing
RNA-seq. The University of Nebraska DNA Sequencing Core receives partial support from the
NCRR (RR018788).
2
Abstract
Hearing loss affects ~10% of adults worldwide and is irreversible. Most sensorineural hearing
loss is caused by progressive loss of mechanosensitive hair cells (HCs) in the cochlea of the inner
ear. The molecular mechanisms underlying HC maintenance and loss are largely unknown. Our
previous cell-specific transcriptome analysis showed that Limb-Bud-and-Heart (LBH), a
transcription co-factor implicated in development, is abundantly expressed in outer hair cells
(OHCs). We used Lbh-null mice to identify its role. Surprisingly, Lbh deletion did not affect
differentiation and early development of HCs, as nascent HCs in Lbh knockout mice had normal
looking stereocilia bundles. Whole-cell recording showed that the stereocilia bundle was
mechanosensitive and OHCs exhibited the characteristic electromotility. However, Lbh-null mice
displayed progressive hearing loss, with stereocilia bundle degeneration and OHC loss as early
as postnatal day 12. Cell-specific RNA-seq and bioinformatic analyses identified Spp1, Six2,
Gps2, Ercc6, Snx6 as well as Plscr1, Rarb, Per2, Gmnn and Map3k5 among the top five
transcription factors up- or down-regulated in Lbh-null OHCs. Furthermore, this analysis showed
significant gene enrichment of biological processes related to transcriptional regulation, cell cycle,
DNA damage/repair and autophagy. In addition, Wnt and Notch pathway-related genes were
found to be dysregulated in Lbh-deficient OHCs. We speculate that LBH may promote
maintenance of HCs and stereocilia bundles by regulating Notch and Wnt signaling activity. Our
study implicates, for the first time, loss of LBH function in progressive hearing loss, and
demonstrates a critical requirement of LBH in promoting HC survival.
Keywords: LBH, hair cell, hearing loss, stereocilia, RNA-seq
3
Introduction
466 million people worldwide are estimated to be living with hearing loss. Most sensorineural
hearing loss is caused by progressive degeneration of hair cells (HCs) in the cochlea of the inner
ear. These cells are specialized mechanoreceptors which transduce mechanical forces
transmitted by sound to electrical activities (Hudspeth, 2014; Fettiplace, 2017). HCs in adult
mammals are terminally differentiated and unable to regenerate once they are lost due to aging
or exposure to noise and ototoxic drugs. Although HCs have been well characterized
morphologically and biophysically, the key molecules that control their differentiation,
homeostasis and aging remain to be identified.
Inner and outer HCs (IHCs and OHCs) are the two types of HCs, with distinct morphology and
function in the mammalian cochlea (Dallos, 1992). IHCs are the true sensory receptor cells and
transmit information to the brain while the OHCs are a mammalian innovation with a unique
capability of changing its length in response to changes in receptor potential (Brownell et al.,
1985; Zheng et al., 2000). OHC motility is believed to confer the mammalian cochlea with high
sensitivity and exquisite frequency selectivity (Liberman et al., 2002; Dallos et al., 2007). We
recently compared cell type-specific transcriptomes of IHC and OHC populations collected from
adult mouse cochleae to identify genes commonly and differentially expressed in these cells (Liu
et al., 2014; Li et al., 2018). Our analysis showed that Limb-bud-and-heart (Lbh), a transcription
co-factor implicated in development (Briegel & Joyner 2001; Briegel et al., 2005; Ai et al. 2008;
Al-Ali et al., 2010; Lindley et al, 2015), is expressed in both IHCs and OHCs (Liu et al., 2014; Li
et al., 2018). Lbh is also expressed in vestibular HCs (Scheffer et al., 2015) and upregulated
during transdifferentiation from supporting cells to HCs (Ebeid et al., 2017; Yamashita et al.,
2018). We, therefore, asked whether LBH is necessary for HC differentiation, development, and
maintenance. Because Lbh expression in OHCs is 4.7 Log2 fold greater than in IHCs, we also
questioned whether LBH plays a role in regulating cell specialization underlying OHC morphology
and function.
Lbh conditional knockout mice have been generated by LoxP and Cre recombination (Lindley
and Briegel, 2013). The role of LBH in HCs was examined by comparing changes in morphology,
function and gene expression between HCs from Lbh-null and wildtype mice. Results showed
that HC differentiation, formation of the mechanotransduction apparatus and OHC specialization
were unaffected by loss of LBH. However, stereocilia bundles and HCs, especially OHCs, showed
signs of degeneration as early as P12. Moreover, adult Lbh-null mice displayed progressive loss
of hearing and otoacoustic emissions, suggesting that LBH is critical for maintenance of
stereocilia bundles and survival of HCs. Cell-specific transcriptome and bioinformatics analyses
showed a significant enrichment of genes associated with transcription, cell cycle, DNA
damage/repair, and autophagy in the Lbh-null OHCs. Wnt and Notch pathway-related genes,
known for their important roles in regulating HC differentiation and regeneration in vertebrate HCs
(Raft and Groves, 2015), were found to be dysregulated. We speculate that dysregulated
Notch/Wnt activity following LBH ablation may lead to degeneration of stereocilia bundles and
OHCs. Our study implicates, for the first time, loss of transcription co-factor LBH function in
progressive hearing loss, and demonstrates a critical requirement of LBH in promoting cochlear
HC survival.
Results
1.Expression of Lbh/LBH in inner ear HCs
Lbh gene expression in HCs and supporting cells in the adult murine organ of Corti was
examined using our published cell type-specific RNA-seq data sets (Liu et al., 2018). Lbh was
expressed in all four cell types, IHCs, OHCs, pillar cells and Deiters’ cells, however, Lbh transcript
levels were highest in OHCs (Fig. 1A, left panel). We also examined expression of Lbh during
4
development using RNA-seq data by Scheffer et al (Scheffer et al., 2015). Lbh was expressed
with comparable levels in both cochlear and vestibular HCs at embryonic day 16 (E16) and
upregulated at postnatal day 7 (P7) (Fig. 1A; right panel). In contrast, low level Lbh expression in
non-sensory supporting cells did not change. We next used LBH-specific antibodies to examine
LBH protein expression in inner ears from neonatal and adult C57BL/6 mice. Fig. 1B shows a
micrograph obtained from a cryosection of a P3 cochlea. LBH was expressed in both OHCs and
IHCs with no obvious expression in supporting cells in the organ of Corti at this neonatal stage
(Fig. 1B). LBH positivity was also detected in some cells in the greater epithelial ridge. In P12
cochlea, LBH was still expressed in both IHCs and OHCs, as revealed by confocal microscopy,
however, expression was strongest in OHCs (Fig. 1C). Of note, in OHCs LBH was predominately
cytoplasmic although weaker expression was also seen in the nuclei of these cells, and in IHCs
(Fig. 1C). This expression pattern was LBH-specific, as in the age-matched Lbh-null mice, no
LBH protein was detected in IHCs and OHCs (Fig. 1D). In adult cochlea, strong LBH expression
in OHCs persisted, while LBH expression in IHCs remained weak (Fig. 1E). This pattern of
expression is consistent with the predominant expression of Lbh mRNA in adult OHCs (Liu et al.,
2014, 2018). In contrast, LBH was not expressed in vestibular HCs, as no LBH-specific
immunopositivity was detected in utricular HCs of P12 wildtype mice (Figs. 1F,G).
2. Auditory function of Lbh-mutant mice
To determine if LBH expression in cochlear HCs is required for hearing, we examined auditory
function in Lbh-mutant mice by measuring auditory brainstem response (ABR). Fig. 2A shows
the ABR thresholds of homozygous (Lbh∆2/∆2), heterozygous (Lbh+/∆2) and wildtype (Lbh+/+) mice
at 1 month of age. As shown, the threshold of Lbh∆2/∆2 null mice is elevated by ~10 dB at lower
frequencies to ~40 dB in higher frequencies relative to their wildtype littermates. Heterozygous
Lbh+/∆2 mice also showed 10 to 25 dB hearing loss at higher frequencies when compared with the
wildtype controls, suggesting that even minor decreases in Lbh gene dosage impairs hearing.
We next measured distortion product otoacoustic emission (DPOAE) thresholds at 8, 16 and 32
kHz in these mice. DPOAEs are generated by motor activity of OHCs (Liberman et al., 2002;
Dallos et al., 2007) and reflect OHC function/condition. Consistent with our ABR measurements,
DPOAE thresholds (Fig. 2B) were also elevated at higher frequencies in Lbh∆2/∆2, and Lbh+/∆2
mice. We further measured the cochlear microphonic (CM) response to an 8 kHz tone burst in
Lbh∆2/∆2 and Lbh+/+ mice. A significant reduction of the CM magnitude (Fig. 2C) in response to the
same level of sound stimulation was observed in Lbh∆2/∆2 mice (n = 6, p = 4.29E-06). Since weak
expression of Lbh was also detected in intermediate cells of the stria vascularis during
development (20), we measured endocochlear potential (EP) from one-month-old Lbh∆2/∆2 and
Lbh+/+ mice to determine whether stria development and function are affected by deletion of Lbh.
This is necessary since stria function (i.e., the EP) can influence HC survival (Liu et al., 2016). An
increase in the EP magnitude was observed in Lbh∆2/∆2 mice (Fig. 2D). The fact that no EP
reduction was observed suggests that loss of LBH does not affect stria function. Finally, ABR and
DPOAE measurements at 3 months of age showed that hearing was further decreased in both
Lbh∆2/∆2 and Lbh+/∆2 mice (Figs. 2E,F), indicating LBH deficiency causes progressive hearing loss.
3. Morphological changes of HCs in Lbh∆2/∆2 mice
We next asked if there was progressive HC loss in Lbh-deficient mice. To this end, we
examined HC frequency at the base and apex of the cochleae at four different ages in Lbh∆2/∆2
and Lbh+/+ mice (n=3 each). Fig. 3A shows representative confocal images at P12 and 1 month.
The total number of IHCs and OHCs at the two cochlear locations were also counted. Fig. 3B
shows the percentage of surviving HCs at P3, P12, 1 and 3 months. No HC loss was apparent
at either location in P3 Lbh∆2/∆2 or Lbh+/+ cochleae. At P12, Lbh∆2/∆2 cochlea exhibited sporadic
HC loss in the basal turn region, whereby OHC loss was more severe than IHC loss (Figs. 3A,
B). At 1 month, OHC loss also occurred at the apical turn, and nearly 50% of OHCs were lost in
5
the basal turn region of these mice (Figs. 3A, B). IHC loss at the basal turn region remained mild.
Finally, more OHCs were lost in both apical and basal turns at 3 months, with only ~10% of OHCs
remaining in the basal turn region of Lbh∆2/∆2 KO cochleae (Fig. 3B). Interestingly, the majority of
IHCs in the basal turns survived and apical IHCs were unaffected, despite substantial OHCs loss
at 3 months. We also examined HC survival in the vestibular end organs at 3 months, but did not
find any noticeable HC loss in the utricle and crista ampullaris of Lbh∆2/∆2 KO mice (Fig. 3C).
Scanning electron microscopy (SEM) was used to examine stereocilia bundle morphology in
Lbh∆2/∆2 mice to determine whether LBH is necessary for morphogenesis and maintenance of
stereocilia and for differentiation of IHCs and OHCs. Fig. 4A shows an electron micrograph of
stereocilia bundles in a P5 Lbh∆2/∆2 mouse cochlea. The characteristic one row of IHC and three
rows of OHC stereocilia bundles were well organized and properly oriented. At higher
magnification (Figs. 4B,C), the stereocilia were arranged in a normal staircase fashion, with OHCs
(Fig. 4B) and IHCs (Fig. 4C) having distinct morphologies. Thus, no signs of abnormality or
degeneration of the stereocilia bundles were visible at P5. In one-month-old Lbh∆2/∆2 cochleae,
stereocilia bundles in the apical turn appeared largely normal, although sporadic OHC stereocilia
bundle loss was observed (asterisk in Fig. 4D). However, degeneration and loss of OHC
stereocilia bundles in the basal turn were more pronounced (Fig. 4E). Some of the remaining
bundles showed signs of degeneration such as absorption (marked by arrows in Fig. 4F),
corruption and recession of the stereocilia on the edge of the bundle (Fig. 4G). While the majority
of IHC stereocilia bundles in the basal turn were present (Fig. 4E), some signs of IHC
degeneration (such as fusion of the stereocilia in Fig. 4H) were also observed. The fact that the
stereocilia bundles of IHCs and OHCs looked normal in P5 Lbh∆2/∆2 mice suggests that LBH is not
essential for morphogenesis of the stereocilia bundles. However, degeneration and loss of
stereocilia bundles in adult Lbh∆2/∆2 HCs suggest that the maintenance of the bundles and survival
of HCs, especially OHCs, depend on LBH.
4. Mechanotransduction (MET) and electromotility of OHCs in Lbh∆2/∆2 mice
We questioned if LBH plays a role in development of MET apparatus as LBH expression
appeared to be limited to HCs. Voltage-clamp technique was used to measure MET current of
OHC stereocilia bundles in response to bundle deflection in Lbh∆2/∆2 mice. A coil preparation from
the mid-cochlear region was used for recording (Jia and He, 2005). The bundle was deflected
with the fluid jet technique (Kros et al., 1992, Jia et al., 2009) and the deflection-evoked MET
current was recorded (Fig. 5A). Two examples of the maximal MET current from OHCs of Lbh∆2/∆2
and Lbh+/+ mice at P12 are presented in Fig. 5A. We compared maximal MET currents from 9 and
8 OHCs from four Lbh+/+ and four Lbh∆2/∆2 mice, respectively. The magnitude of the current was
614 ± 90 pA (mean ± SD) for Lbh+/+ and 449 ± 57 pA for Lbh∆2/∆2 OHCs. Despite significant
reduction (p=0.00048), the presence of MET current suggests that the mechanotransduction
apparatus is functional in Lbh∆2/∆2 OHCs.
Prestin-based somatic motility is a unique property of OHCs (Zheng et al., 2000). As LBH is
predominantly expressed in OHCs, we asked if LBH regulates prestin expression. OHC
electromotility occurs after birth (He et al., 1994; He, 1997); thus, we measured nonlinear
capacitance (NLC), an electric signature of electromotility (Ashmore, 1989; Santos-Sacchi, 1991;
He et al., 2010), from Lbh∆2/∆2 OHCs at P12 when OHC degeneration was observed to be mild.
Fig. 5B shows NLC measured from 9 OHCs from the mid-cochlear region in Lbh∆2/∆2 and Lbh+/+
mice. A two-state Boltzmann function relating nonlinear charge movement to voltage (Ashmore,
1989; Santos-Sacchi, 1991) was used to compute four parameters, the maximum charge
transferred through the membrane’s electric field (Qmax), the slope factor of the voltage
dependence (α), the voltage at peak capacitance (Vpkcm), and the linear membrane capacitance
(Clin). No statistically significant differences in any of these parameters were found between
Lbh∆2/∆2 and Lbh+/+ OHCs (Fig. 5B). Thus, OHC motility is not affected by loss of LBH.
6
5. Changes in OHC gene expression after deletion of Lbh
To identify molecular mechanism underlying the observed hearing and HC loss in Lbh-
deficient mice, we performed OHC-specific RNA-seq transcriptome analyses. OHCs were
isolated from P12 mice Lbh∆2/∆2 and Lbh+/+ mice (Fig. 6A), as at this stage HC degeneration in Lbh
null mice had just begun (Fig. 3A,B). The raw data of transcriptomes of P12 OHCs from Lbh∆2/∆2
and Lbh+/+ mice are available from the National Center for Biotechnology Information BioProject’s
metadata (PRJNA552016). For similarity comparison, Fig. 6B shows a Euclidean distance
heatmap of 10,000 genes with a cutoff Z-score calculated as the absolute values from the mean.
Comparison of the gene expression profiles between Lbh∆2/∆2 and Lbh+/+ OHCs identified 2,779
differentially upregulated and 2,065 differentially downregulated genes (defined as those whose
expression levels were ≥ 1.0 log2 fold change in expression between the two cell types with
statistical significance (FDR p value ≤ 0.01)) in Lbh∆2/∆2 OHCs (Fig. 6C). Among those genes,
biological processes related to gene expression, protein metabolic process, and organelle
organization were significantly enriched in Lbh∆2/∆2 OHCs, as assessed by ShinyGO analysis
(Figs. 6D,E). In contrast, Lbh+/+ OHCs showed greater enrichment in genes associated with
cytoskeletal and actin filament organization, membrane bound cell projection organization,
anatomical structure morphogenesis, RNA splicing, and axon ensheathment (Fig. 6E).
Additionally, gene set enrichment analysis (GSEA) was performed using the Broad Institute
software. Enriched pathways in Lbh∆2/∆2 compared to Lbh+/+ OHCs included Wnt and Notch
signaling pathways, as well as cell cycle regulation, regulation of nucleic acid-templated
transcription, DNA damage/repair and autophagy (Fig. 7). As Wnt and Notch play important roles
in HC differentiation and regeneration, and LBH is a Wnt target gene known to regulate cell
differentiation states in other cell types )(Briegel et al., 2005; Conen et al., 2009; Rieger et al.,
2010; Lindley et al, 2015; Li et al., 2015), the expression of genes related to Wnt and Notch
signaling was examined in more detail. While Notch1 (although not among the top 20), Wnt4,
Fzd4, Ctnnb1 (β-catenin) and Fuz were all significantly upregulated, key target genes of Wnt (e.g.
Axin2, Lgr5, Lrp6) and Notch (i.e. Hey1, Hey2), that mirror signaling activity, were downregulated
in Lbh∆2/∆2 OHCs (Fig. 7A,B). For cell cycle control, 192 genes were upregulated while 107 were
downregulated (Fig. 7C). Analysis of transcription factors shows 430 and 281 transcription factors
up- or down-regulated in the Lbh∆2/∆2 OHCs, respectively (Fig. 7D). As LBH is implicated in DNA
damage/repair in some cells (Deng et al., 2010; Matusda et al., 2017), the enrichment in these
genes was also analyzed (Figs. 7E,F). 90 and 55 genes associated with DNA damage/repair are
up- and down-regulated in Lbh∆2/∆2 OHCs, respectively. Interestingly, autophagy-related genes
were also found to be enriched, whereby 81 genes were upregulated and 46 downregulated in
Lbh∆2/∆2 OHCs.
RT-qPCR was used to validate selected differentially expressed genes identified by the RNA-
seq analysis, using RNA from P12 Lbh∆2/∆2 and Lbh+/+ OHCs. Seventeen genes involved in key
biological processes related to HC maintenance/degeneration were chosen for comparison. As
shown in Fig. 7G, the trend of differential expression of these genes is highly consistent between
the two analyses, confirming LBH-dependent gene expression changes in the global RNA-seq
analysis.
Discussion
LBH, a transcriptional regulator highly conserved in evolution from zebrafish to human, is
implicated in heart (Briegel and Joyner, 2001; Briegel et al., 2005; Ai et al. 2008), bone (Conen et
al. 2009), and mammary gland (Lindley et al., 2015) development. A zebrafish LBH homologue,
lbh-like, is necessary for photoreceptor differentiation (Li et al., 2015). Here, we identified an
unanticipated novel role of LBH in the maintenance of the adult auditory sensory epithelium.
Unlike in heart, bone, mammary gland, and eye development, where LBH or lbh-like proteins
7
control progenitor/stem cell fate, self-renewal, and/or differentiation, LBH does not appear to be
critical for cochlear HC differentiation, specification, and stereocilia morphogenesis.
Morphologically distinct IHCs and OHCs were present at birth, with no HC loss in the cochleae of
Lbh-null mice at early postnatal stages. Stereocilia bundles of Lbh-null HCs also appeared normal
and were functional, as mechanical stimulus was able to evoke MET currents. Furthermore, LBH
is not necessary for expression of prestin, a specialization of OHCs, despite the fact that Lbh is
preferentially expressed in adult OHCs. We, therefore, conclude that LBH is not necessary for
stereocilia morphogenesis and HC differentiation, specification and development.
Surprisingly, however, we found that LBH is critical for stereocilia bundle maintenance and
survival of HCs in adult mice. When Lbh was deleted, stereocilia and HCs began to degenerate
as early as P12. The degeneration was progressive from OHCs to IHCs and from base to apex
of the cochlea, similar to the pattern seen during age-related hearing loss. Our findings, to the
best of our knowledge, are the first demonstration that loss of LBH causes degeneration of
cochlear HCs, leading to progressive hearing loss. Furthermore, these results provide evidence
that LBH is required for adult tissue maintenance. While LBH has been previously implicated in
tissue maintenance and regeneration of the postnatal mammary gland by promoting the self-
renewal and maintenance of the basal mammary epithelial stem cell pool (Lindley et al., 2015);
HCs are terminally differentiated, postmitotic cells that have lost the ability to proliferate and
regenerate. In this regard, it is worth noting that loss of LBH is also associated with Alzheimer’s,
a neurodegenerative disease affecting postmitotic neurons (Yamaguchi-Kabata et al., 2018).
Thus, LBH appears to be required for tissue maintenance in both regenerative and non-
regenerative adult tissues.
OHC-specific RNA-seq and bioinformatic analyses examined the potential molecular
mechanisms underlying HC degeneration after Lbh deletion. Our analyses showed that a greater
number of genes were upregulated in Lbh-null OHCs compared to wildtype littermate OHCs.
Importantly, genes/pathways associated with transcriptional regulation, cell cycle, DNA
repair/maintenance, autophagy, as well as Wnt and Notch signaling were significantly enriched in
Lbh-null OHCs. Notch and Wnt signaling are known to be critical for differentiation and
specification of HCs and supporting cells during inner ear morphogenesis and development (Raft
and Groves, 2015). Notch and Wnt signaling are also important for transdifferentiation of
supporting cells to HCs during regeneration (Waqas et al., 2016; Jansson et al., 2015; Zak et al.,
2015). Blocking Notch signaling leads to generation of supernumerary HCs in vivo and in vitro
(Lanford et al., 1999; Li et al., 2018). Notch and Wnt signaling are normally downregulated
(Kiernan, 2013), whereas Lbh is upregulated during HC maturation (Scheffer et al., 2015).
Interestingly, our RNA-seq and pathway analyses showed that, although many WNT and Notch
pathway genes were upregulated in Lbh-null OHCs, key target genes of WNT (Axin2, Lgr5, Lrp6)
and Notch (Hey1, Hey2, Jag1), mirroring endogenous signaling activity, were downregulated. This
suggests that normal adult OHCs retain low levels of Notch and WNT signaling for their
maintenance. It also suggests that LBH is required for maintaining low level Notch and WNT
activity in OHCs, and that dysregulated Notch/Wnt activity following LBH ablation, as measured
by altered WNT/Notch target gene expression in Lbh null OHCs, may lead to OHC degeneration.
Thus, LBH may promote maintenance of HCs and stereocilia bundles by regulating Notch and
Wnt signaling activity. Alternatively, OHC degeneration caused by LBH loss may be due to
increased genotoxic and cell stress, as cell cycle, DNA repair/maintenance, and autophagy genes
were also deregulated. Indeed, a recent study showed that LBH is involved in cell cycle regulation,
and LBH-deficiency induced S-phase arrest and increased DNA damage in articular cartilage
(Matusda et al., 2017). Cell-based transcriptional reporter assays further indicate LBH may
repress the transcriptional activation of p53 (Deng et al., 2010), a key regulator of DNA damage
control and apoptosis.
8
Our finding that Wnt pathway genes were dysregulated in Lbh-null OHCs was unexpected.
LBH is a direct WNT/ß-catenin target gene induced by canonical Wnt signaling in epithelial
development and cancer (Rieger et a., 2010). It is required downstream of WNT to promote
mammary epithelial cell proliferation, while blocking differentiation (Rieger et a., 2010; Ashad-
Bishop et al., 2019). Interestingly, LBH knockout in a WNT-driven breast cancer mouse model,
MMTV-Wnt1, reduced cell proliferation and hyperplasia, but also increased cell death (Ashad-
Bishop et al., 2019). This supports our notion that LBH is required for epithelial cell maintenance.
Conversely, knockdown of zebrafish lbh-like increased cell proliferation and Notch target gene
(Hes5) expression (Li et al., 2015); while we find essential Notch target effectors (Hey1, Hey2) to
be downregulated in Lbh-null OHCs. Regardless of whether LBH acts upstream or downstream
of Wnt and Notch signaling, all these studies suggest that dysregulation of Notch and Wnt
signaling, and alterations in LBH levels can perturb the balance between proliferation,
differentiation, and maintenance, with different outcomes in different epithelial tissues (Rieger et
al., 2010; Li et al., 2015; Ashad-Bishop et al., 2019).
LBH function as a transcription cofactor has been shown by multiple studies ((Briegel and
Joyner, 2001; Briegel et al., 2005; Ai et al 2008; Deng et al 2010). Interestingly, while LBH is
predominantly localized to the nucleus in most cells (Briegel and Joyner, 2001; Lindley et al.,
2015; Liu et al., 2015; Ashad-Bishop et al., 2019), LBH expression was predominantly cytoplasmic
in HCs, although weak nuclear LBH positivity was also observed. In fibroblast-like COS-7 cells,
co-localization analysis shows that LBH proteins are localized to both the nucleus and the
cytoplasm (Ai et al., 2008). In postmitotic neurons, LBH is also found to be more cytoplasmic than
nuclear (unpublished observation). Some transcriptional co-factors, such as TAZ/YAP, are
detected in the cytoplasm and can translocate into the nucleus upon mechano-stimulation (Low
et al., 2014). The STAT (signal transducer and activator of transcription) transcription factors are
constantly shuttling between nucleus and cytoplasm irrespective of cytokine stimulation (Meyer
and Vinkemeier, 2004). It is therefore plausible that cytoplasmic LBH in OHCs may translocate
to the nucleus upon sensory input. It is also possible that in the cytoplasm LBH may interact with
different proteins and have a different function than in the nucleus.
Collectively, this is the first study showing that transcription co-factor LBH can influence
stereocilia bundle maintenance and survival of cochlear HCs, especially OHCs. Although the
underlying mechanisms remain to be further investigated, based on our analyses, we entertain
the possibility that dysregulation of Notch and Wnt signaling caused by LBH loss is detrimental to
maintenance of stereocilia bundles and survival of adult cochlear HCs. Importantly, our work
points to LBH as a novel causative factor and putative molecular target in progressive hearing
loss. It also identifies LBH as paramount for adult tissue maintenance, which could be exploited
therapeutically to slow aging of HCs.
Materials and Methods
1. Lbh knockout mice: Lbh-mutant mice aged between P0 and 3 months were used for
experiments. Mice with a conditional null allele of Lbh were generated by flanking exon 2
with loxP sites (Lbhflox) (17) (Lindley and Briegel, 2013). LbhloxP mice were then crossed with a
Rosa26-Cre line, resulting in ubiquitous deletion of exon 2 and abolishment of LBH protein
expression, which was confirmed by western plot and negative anti-LBH antibody staining
(Lindley and Briegel, 2013). Ubiquitous Lbh-null mice are viable and fertile. Care and use of the
animals in this study were approved by the Institutional Animal Care and Use Committees of
Creighton University and the University of Miami.
2. ABR and DPOAE measurements: ABRs were recorded in response to tone bursts from 4 to 50
kHz using standard procedures described previously (Zhang et al., 2013). Response signals were
amplified (100,000x), filtered, averaged and acquired by TDT RZ6 (Tucker-Davis Technologies,
9
Alachua, FL). Threshold is defined visually as the lowest sound pressure level (in decibel) at which
any wave (wave I to wave IV) is detected and reproducible above the noise level.
The DPOAE at the frequency of 2f1 -f2 was recorded in response to f1 and f2, with f2/f1 = 1.2
and the f2 level 10 dB lower than the f1 level. The sound pressure obtained from the microphone
in the ear-canal was amplified and Fast-Fourier transforms were computed from averaged
waveforms of ear-canal sound pressure. The DPOAE threshold is defined as the f1 sound
pressure level (measured in decibels) required to produce a response above the noise level at
the frequency of 2f1-f2.
3. Recording of CM and EP: Procedures for recording CM and EP were described before (Zhang
et al., 2014; Liu et al., 2016). A silver electrode was placed on the ridge near the round window
for recording CM. An 8 kHz tone burst was delivered through a calibrated TDT MF1 multi-field
magnetic speaker. The biological signals were amplified using an Axopatch 200B amplifier
(Molecular Devices, Sunnyvale, CA) and acquired by software pClamp 9.2 (Molecular Devices)
running on an IBM-compatible computer. The sampling frequency was 50 kHz.
For recording the EP, a basal turn location was chosen. A hole was made using a fine drill. A
glass capillary pipette electrode (5 MΩ) was mounted on a hydraulic micromanipulator and
advanced until a stable positive potential was observed. The signals were filtered and amplified
under current-clamp mode using an Axopatch 200B amplifier and acquired by software pClamp
9.2. The sampling frequency was 10 kHz.
4. Immunocytochemistry and HC count: The cochlea and vestibule from the Lbh∆2/∆2 and Lbh+/+
were fixed for 24 hours with 4% paraformaldehyde (PFA). The basilar member, including the
organ of Corti, the utricle and ampulla were dissected out. Antibodies against MYO7A (Proteus,
product # 25-6790) or LBH (Sigma, Lot# HPA034669) and secondary antibody (Life
Technologies, Lot# 1579044) were used. Tissues were mounted on glass microscopy slides and
imaged using a Leica Confocal Microscope (Leica TCS SP8 MP). HC counts from two areas
(approximately 1.4 and 4.5 mm from the hook, each 800 µm in length) were obtained for HC count
from confocal images off-line (Zhang et al., 2013).
5. SEM: The cochleae from Lbh-mutant mice were fixed for 24 hours with 2.5% glutaraldehyde in
0.1 M sodium cacodylate buffer (pH 7.4) containing 2 mM CaCl2, washed in buffer. After the
cochlear wall was removed, the cochleae were then post-fixed for 1 hour with 1% OsO4 in 0.1 M
sodium cacodylate buffer and washed. The cochleae were dehydrated via an ethanol series,
critical point dried from CO2 and sputter-coated with gold. The morphology of the HCs was
examined in a FEI Quanta 200 scanning electron microscope and photographed.
6. Whole-cell voltage-clamp techniques for recording MET current and NLC: Details for recording
MET currents from auditory sensory epithelium are provided elsewhere (Kros et al., 1992; Jia and
He, 2005). A segment of auditory sensory epithelium was prepared from the mid-cochlear and
bathed in extracellular solution containing (in mM) 120 NaCl, 20 TEA-Cl, 2 CoCl2, 2 MgCl2, 10
HEPES, and 5 glucose at pH 7.4. The patch electrodes were back-filled with internal solution,
which contains (in mM) CsCl 140; CaCl2: 0.1; MgCl2 3.5; MgATP: 2.5; EGTA-KOH 5; HEPES-
KOH 10. The solution was adjusted to pH 7.4 and osmolarity adjusted to 300 mOsm with glucose.
The pipettes had initial bath resistances of ~3-5 MΩ. After the whole-cell configuration was
established and series resistance was ~70% compensated, the cell was held under voltage-clamp
mode to record MET currents in response to bundle deflection by a fluid jet positioned ~10–15
µm away from the bundle. Sinusoidal bursts (100 Hz) with different magnitudes were used to drive
the fluid jet as described previously (Kros et al., 1992; Jia and He, 2005). Holding potential was
normally set near -70 mV. The currents (filtered at 2 kHz) were amplified using an Axopatch 200B
amplifier and acquired using pClamp 9.2. Data were analyzed using Clampfit in the pClamp
software package and Igor Pro (WaveMetrics, Inc).
For recording NLC, the cells were bathed in extracellular solution containing (in mM) 120
10
NaCl, 20 TEA-Cl, 2 CoCl2, 2 MgCl2, 10 HEPES, and 5 glucose at pH 7.4. The internal solution
contains (in mM): 140 CsCl, 2 MgCl2, 10 EGTA, and 10 HEPES at pH 7.4. The two-sine voltage
stimulus protocol (10 mV peak at both 390.6 and 781.2 Hz) with subsequent fast Fourier
transform-based admittance analysis (jClamp, version 15.1) was used to measure membrane
capacitance using jClamp software (Scisoft). Fits to the capacitance data were made in IgorPro
(Wavemetrics). The maximum charge transferred through the membrane’s electric field (Qmax),
the slope factor of the voltage dependence (α), the voltage at peak capacitance (Vpkcm), and the
linear membrane capacitance (Clin) were calculated.
7. Cell isolation, RNA preparation, and RNA-sequencing: Lbh∆2/∆2 and Lbh+/+ mice at P12 were
used for gene expression analysis. Details for cell isolation and collection are provided elsewhere
(Liu et al., 2014). Approximately 1,000 OHCs were collected from 7-8 mice for one biological
repeat per genotype. Three biological replicates were prepared for each genotype.
Total RNA, including small RNAs (> ~18 nucleotides), were extracted and purified using the
Qiagen RNeasy mini plus Kit (Qiagen, Germantown, MD). To eliminate DNA contamination in the
collected RNA, on-column DNase digestion was performed. The quality and quantity of RNA were
examined using an Agilent 2100 BioAnalyzer (Agilent, Santa Clara, CA).
Genome-wide transcriptome libraries were prepared from three biological replicates
separately for Lbh∆2/∆2 and Lbh+/+ OHCs. The SMART-Seq V4 Ultra Low Input RNA kit (Clontech
Laboratories, Inc., Mountain View, CA) and the Nextera Library preparation kit (Illumina, Inc., San
Diego, CA) were used. An Agilent 2100 Bioanalyzer and a Quibit fluorometer (Invitrogen, Thermo
Fisher Scientific) were used to assess library size and concentration prior to sequencing.
Transcriptome libraries were sequenced using the HiSeq 2500 Sequencing System (Illumina).
Four samples per lane were sequenced, generating approximately 60 million, 100 bp single-end
reads per sample. The files from the multiplexed RNA-seq samples were demulitplexed and fastq
files were obtained.
The CLC Genomics Workbench software (CLC bio, Waltham, MA, USA) was used to
individually map the reads to the exonic, intronic, and intergenic sections of the mouse genome
(mm10, build name GRCm38). Gene expression values were normalized as reads per kilobase
of transcript per million mapped reads (RPKM). Log fold changes and FDR p-values were
calculated, and the dataset was exported for further analysis. The raw data are available from the
National Center for Biotechnology Information BioProject’s metadata (PRJNA552016) and the
normalized RPKM values are accessible as an Excel file. Transcriptomes and differentially
expressed genes as well as significantly enriched genes associated with Wnt and Notch signaling,
transcription, cell cycle, DNA damage/repair, and autophagy in the Lbh-null OHCs are also
provided as Supplemental File 1.
8. Real-time quantitative PCR for validation: OHCs were collected as described above from
fourteen additional Lbh∆2/∆2 mice (aged P12) and 15 age-matched Lbh+/+ mice for RT-qPCR. Total
RNA was isolated using the Qiagen miRNeasy kit and quantified by nanodrop. cDNA libraries
were prepared from isolated RNA with the iSCRIPT master mix (BioRad). Oligonucleotide primers
were acquired from Integrated DNA Technologies (Coralville, Iowa). The sequences of
oligonucleotide primers are shown as follows:
Gene
NCBI Ref Seq
Forward (5' > 3')
Reverse (5' > 3')
Actb
NM_007393.5
GTACTCTGTGTGGATCGGTGG
ACGCAGCTCAGTAACAGTCC
Arrb1
NM_177231.2
AAGGGACACGAGTGTTCAAGA
GATCCACCAGGACCACACCA
Bcl2
NM_009741.5
GAGTTCGGTGGGGTCATGTG
AGTTCCACAAAGGCATCCCAG
Chchd4
NM_133928.2
CGGGAACAACCATGTCCTACT
GGCAGTATCAACCCGTGCTC
Cinp
NM_026048.4
CCATCTTGGACGGCTTGACTA
ACGTGTGAAATAGAGGGGGC
Ercc6
NM_001081221.2
TGAGCAGGTCTTATTTTGCCG
AAAGAGGTCAGGGTGGTTGC
Fuz
NM_027376.3
CTGAAGAAAGAATTGAGGGCCAG
CCTCTGCAAACCCTGAAAGG
Itgb1bp1
NM_008403.5
ACACTTGTTCCACTGCGGC
CCACAGACTTGCTCTTTGTACTG
11
Mrpl28
NM_024227.3
CACTCGGGAGCTTTACAGTGA
GCTTCAGGTCCATGCCAAAC
Mrps12
NM_001360250.1
CCGCTAGGTTGGTGAGGTG
AAAACAGAAAGTCCCCTCGCA
Nprl2
NM_018879.2
CTGTCCTACGTCACCAAGCA
CTGGATCAGCTTCCTTTCATCA
Ruvbl2
NM_011304.3
CACACCATTCACAGCCATCG
CTCTGTCTCCTCCTTGATCCG
Scfd1
NM_029825.3
CGTCCGAGGTTGATTTGGAG
TAGTGTTTCCGTAGCTGGCA
Six2
NM_011380.2
CGCAAGTCAGCAACTGGTTC
GAACTGCCTAGCACCGACTT
Slain1
NM_001361639.1
TCAGCCCTTATAGCAATGGCA
ACTGTCGATGGATGACTGCG
Spp1
NM_001204201.1
ATCCTTGCTTGGGTTTGCAG
TGGTCGTAGTTAGTCCCTCAGA
Uqcrfs1
NM_025710.2
TTCTGGATGTGAAGCGACCC
CAGAGAAGTCGGGCACCTTG
Zfp365
NM_178679.2
GAAGCCCAGATGCCTAAGCC
GACTCAGCCGGTTCGTGAAT
RT-qPCR reactions were prepared as 10 l reactions including Lbh∆2/∆2 or Lbh+/+ OHC cDNA,
PowerUp SYBR green master mix (Thermo Fisher), gene-specific forward and reverse primers
and run in triplicate on a BioRad CFX96 Touch real-time PCR machine. Primer specificity was
confirmed by melt curve analysis. Quantified expression (Ct) of each gene (gene of interest or
GOI) was normalized to the Ct value of a house-keeping gene (Actb) (∆Ct = Ct(GOI) ― CtAVG Actb).
Then differential expression of the gene between Lbh∆2/∆2 and Lbh+/+ OHCs was calculated as
∆∆Ct (∆∆Ct = ∆Ct Lbh∆2/∆2 - ∆Ct Lbh+/+). The relationships between the RNA-seq derived-log2 fold
change values and Cq values from RT-qPCR between Lbh∆2/∆2 and Lbh+/+ OHCs were
compared to confirm trends in expression.
9. Bioinformatic Analyses: The expressed genes were examined for enrichment using GSEA v.
3.0 (Broad Institute) (Mootha et al., 2003; Subramanian et al., 2005), iDEP 0.85 and ShinyGO
(Ge-lab.org) (Ge et al., 2018). Enriched biological processes and molecular functions, classified
according to gene ontology (GO) terms, as well as signaling pathways in the Lbh∆2/∆2 and Lbh+/+
OHCs were examined (FDR cutoff < 0.05). With the RPKM expression value arbitrarily set at ≥
0.10 (FDR p-value ≤ 0.05) (Li et al., 2018; Liu et al., 2018), expression values from Lbh∆2/∆2 and
Lbh+/+ OHCs were input into iDEP for analyses and log transformed. All transcripts detected in
Lbh∆2/∆2 and Lbh+/+ OHCs is provided in Supplementary Data 1. For reference and verification,
additional
resources
such
as
the
Ensembl
database,
AmiGO
(http://amigo.geneontology.org/amigo),
gEAR
(www.umgear.org)
and
SHIELD
(https://shield.hms.harvard.edu/index.html) were also used. No custom code was used in the
analysis.
10. Statistical analysis: Means and standard deviations (SD) were calculated based on
measurements from three different types of mice. For each parameter, student’s t-test was used
to determine statistical significance between two different conditions and genotypes. Probability
(P) value ≤ 0.01 was regarded as significant. For transcriptome analysis, means and SD were
calculated for three biological repeats from Lbh∆2/∆2 and Lbh+/+ OHCs, with four technical replicates
each. ANOVA False Discovery Rate-corrected p-values were used to compare average
expression (RPKM) values for each transcript and FDR p < 0.05 was considered statistically
significant.
Author Contributions
HL performed electrophysiological and morphological experiments, KG performed transcriptome
analysis and contributed to manuscript writing, GMH, SM, YL, XL performed some morphological
and electrophysiological experiments, KB generated Lbh-mutant mice, contributed to
experimental design, data analysis, and manuscript writing. DH designed research, performed
SEM experiments and wrote the manuscript.
12
Reference
1. Ai J, Wang Y, Tan K, Deng Y, Luo N, Yuan W, Wang Z, Li Y, Wang Y, Mo X, Zhu C, Yin Z,
Liu M, Wu X. A human homolog of mouse Lbh gene, hLBH, expresses in heart and activates
SRE and AP-1 mediated MAPK signaling pathway. Mol Biol Rep. 35(2):179-87 (2008).
2. Al-Ali H, Rieger ME, Seldeen KL, Harris TK, Farooq A, Briegel KJ. Biophysical
characterization reveals structural disorder in the developmental transcriptional regulator
LBH. Biochem Biophys Res Commun. 391(1):1104-9 (2010).
3. Ashad-Bishop K, Garikapati K, Lindley LE, Jorda M, Briegel KJ. Loss of Limb-Bud-and-Heart
(LBH) attenuates mammary hyperplasia and tumor development in MMTV-Wnt1 transgenic
mice. Biochem Biophys Res Commun. 508(2):536-542 (2019).
4. Ashmore JF. Motor coupling in mammalian outer hair cells. In: Wilson, J.P., Kemp, D.T. (Eds.),
Cochlear Mechanisms. Plenum, New York, pp. 107–114 (1989).
5. Belyantseva IA, Adler HJ, Curi R, Frolenkov GI, Kachar B. Expression and localization of
prestin and the sugar transporter GLUT-5 during development of electromotility in cochlear
outer hair cells. J Neurosci. 20(24):RC116 (2000).
6. Briegel KJ, Joyner AL. Identification and characterization of Lbh, a novel conserved nuclear
protein expressed during early limb and heart development. Dev Biol. 233(2):291-304 (2001).
7. Briegel KJ, Baldwin HS, Epstein JA, Joyner AL. Congenital heart disease reminiscent of partial
trisomy 2p syndrome in mice transgenic for the transcription factor Lbh. Development.
132(14):3305-16 (2005).
8. Brownell WE, Bader CR, Bertrand D, de Ribaupierre Y. Evoked mechanical responses of
isolated cochlear outer hair cells. Science. 227(4683):194-6 (1985).
9. Conen KL, Nishimori S, Provot S, Kronenberg HM. The transcriptional cofactor Lbh regulates
angiogenesis and endochondral bone formation during fetal bone development. Dev Biol.
333(2): 348–358 (2009).
10. Dallos P. The active cochlea. J Neurosci. 12(12):4575-85 (1992).
11. Dallos P, Wu X, Cheatham MA, Gao J, Zheng J, Anderson CT, Jia S, Wang X, Cheng WH,
Sengupta S, He DZ, Zuo J. Prestin-based outer hair cell motility is necessary for mammalian
cochlear amplification. Neuron. 58(3):333-9 (2008).
12. Deng Y, Li Y, Fan X, Yuan W, Xie H, Mo X, Yan Y, Zhou J, Wang Y, Ye X, Wan Y, Wu X.
Synergistic efficacy of LBH and alphaB-crystallin through inhibiting transcriptional activities of
p53 and p21. BMB Rep. 43(6):432-7 (2010).
13. Ebeid M, Sripal P, Pecka J, Beisel KW, Kwan K, Soukup GA. Transcriptome-wide comparison
of the impact of Atoh1 and miR-183 family on pluripotent stem cells and multipotent otic
progenitor cells. PLoS One. 12(7):e0180855 (2017).
14. Fettiplace R. Hair cell transduction, tuning, and synaptic transmission in the mammalian
cochlea. Compr Physiol. 7(4):1197-1227 (2017).
15. Ge SX, Son EW, Yao R. iDEP: an integrated web application for differential expression and
pathway analysis of RNA-Seq data. BMC Bioinformatics. 19(1):534 (2018).
16. He DZ, Evans BN, Dallos P. First appearance and development of electromotility in neonatal
gerbil outer hair cells. Hear Res. 78(1):77-90 (1994).
17. He DZ. Relationship between the development of outer hair cell electromotility and efferent
innervation: a study in cultured organ of corti of neonatal gerbils. J Neurosci. 17(10):3634-43
(1997).
18. He DZ, Jia S, Sato T, Zuo J, Andrade LR, Riordan GP, Kachar B Changes in plasma
membrane structure and electromotile properties in prestin deficient outer hair cells.
13
Cytoskeleton 67(1):43-55 (2010).
19. Hudspeth AJ. Integrating the active process of hair cells with cochlear function. Nat Rev
Neurosci. 15(9):600-14 (2014).
20. Jansson L, Kim GS, Cheng AG. Making sense of Wnt signaling-linking hair cell regeneration
to development. Front Cell Neurosci. 9:66 (2015).
21. Jia S, He DZ. Motility-associated hair-bundle motion in mammalian outer hair cells. Nat
Neurosci. 8(8):1028-34 (2005).
22. Jia S, Yang S, Guo W, He DZ. Fate of mammalian cochlear hair cells and stereocilia after
loss of the stereocilia. J Neurosci. 29(48):15277-85 (2009).
23. Kiernan AE. Notch signaling during cell fate determination in the inner ear. Semin Cell Dev
Biol. 24(5):470-9 (2013).
24. Kros CJ, Rüsch A, Richardson GP. Mechano-electrical transducer currents in hair cells of the
cultured neonatal mouse cochlea. Proc Biol Sci. 249(1325):185-93 (1992).
25. Lanford PJ, Lan Y, Jiang R, Lindsell C, Weinmaster G, Gridley T, Kelley MW. Notch signaling
pathway mediates hair cell development in mammalian cochlea. Nat. Genet. 21, 289–292
(1999).
26. Li WH, Zhou L, Li Z, Wang Y, Shi JT, Yang YJ, Gui JF. Zebrafish Lbh-like is required for Otx2-
mediated photoreceptor differentiation. Int J Biol Sci. 11(6):688-700 (2015).
27. Li Y, Liu H, Giffen KP, Chen L, Beisel KW, He DZZ. Transcriptomes of cochlear inner and
outer hair cells from adult mice. Sci Data. 5:180199 (2018).
28. Li Y, Jia S, Liu H, Tateya T, Guo W, Yang S, Beisel KW, He DZZ. Characterization of hair cell-
like cells converted from supporting cells after Notch inhibition in cultures of the organ of Corti
from neonatal gerbils. Front Cell Neurosci. 12:73 (2018).
29. Liberman MC, Gao J, He DZ, Wu X, Jia S, Zuo J. Prestin is required for electromotility of the
outer hair cell and for the cochlear amplifier. Nature. 419(6904):300-4 (2002).
30. Lindley LE, Briegel KJ. Generation of mice with a conditional Lbh null allele. Genesis.
51(7):491-7 (2013).
31. Lindley LE, Curtis KM, Sanchez-Mejias A, Rieger ME, Robbins DJ, Briegel KJ. The WNT-
controlled transcriptional regulator LBH is required for mammary stem cell expansion and
maintenance of the basal lineage. Development. 142(5):893-904 (2015).
32. Liu H, Pecka JL, Zhang Q, Soukup GA, Beisel KW, He DZ. Characterization of transcriptomes
of cochlear inner and outer hair cells. J Neurosci. 34(33):11085-95 (2014).
33. Liu Q, Guan X, Lv J, Li X, Wang Y, Li L. Limb-bud and Heart (LBH) functions as a tumor
suppressor of nasopharyngeal carcinoma by inducing G1/S cell cycle arrest. Sci Rep. 5:7626
(2015).
34. Liu H, Li Y, Chen L, Zhang Q, Pan N, Nichols DH, Zhang WJ, Fritzsch B, He DZ. Organ of
Corti and stria vascularis: Is there an interdependence for survival? PLoS One.
11(12):e0168953 (2016).
35. Liu H, Chen L, Giffen KP, Stringham ST, Li Y, Judge PD, Beisel KW, He DZZ. Cell-specific
transcriptome analysis shows that adult pillar and Deiters' cells express genes encoding
machinery for specializations of cochlear hair cells. Front Mol Neurosci. 11:356 (2018).
36. Low BC, Pan CQ, Shivashankar GV, Bershadsky A, Sudol M, Sheetz M. YAP/TAZ as
mechanosensors and mechanotransducers in regulating organ size and tumor growth. FEBS
Lett. 588(16):2663-70 (2014).
37. Matsuda S, Hammaker D, Topolewski K, Briegel KJ, Boyle DL, Dowdy S, Wang W, Firestein
GS. Regulation of the cell cycle and inflammatory arthritis by the transcription cofactor LBH
14
gene. J Immunol. 199(7):2316-2322 (2017).
38. Meyer T, Vinkemeier U. Nucleocytoplasmic shuttling of STAT transcription factors. Eur J
Biochem. 271(23-24):4606-12 (2004).
39. Mootha VK, Lindgren CM, Eriksson KF, Subramanian A, Sihag S, Lehar J, Puigserver P,
Carlsson E, Ridderstråle M, Laurila E, Houstis N, Daly MJ, Patterson N, Mesirov JP, Golub
TR, Tamayo P, Spiegelman B, Lander ES, Hirschhorn JN, Altshuler D, Groop LC. PGC-
1alpha-responsive
genes
involved
in
oxidative
phosphorylation
are
coordinately
downregulated in human diabetes. Nat Genet. 34(3):267-73 (2003).
40. Oliver D, Fakler B. Expression density and functional characteristics of the outer hair cell
motor protein are regulated during postnatal development in rat. J Physiol. 519 Pt 3:791-800
(1999).
41. Raft S, Groves AK. Segregating neural and mechanosensory fates in the developing ear:
patterning, signaling, and transcriptional control. Cell Tissue Res. 359(1):315-32 (2015).
42. Rieger ME, Sims AH, Coats ER, Clarke RB, Briegel KJ. The embryonic transcription cofactor
LBH is a direct target of the Wnt signaling pathway in epithelial development and in aggressive
basal subtype breast cancers. Mol Cell Biol. 30(17):4267-79 (2010).
43. Santos-Sacchi J. Reversible inhibition of voltage-dependent outer hair cell motility and
capacitance. J Neurosci. 11(10):3096-110 (1991).
44. Scheffer DI, Shen J, Corey DP, Chen ZY. Gene expression by mouse inner ear hair cells
during development. J Neurosci. 35(16):6366-80 (2015).
45. Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A,
Pomeroy SL, Golub TR, Lander ES, Mesirov JP. Gene set enrichment analysis: a knowledge-
based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci USA.
102(43):15545-50 (2005).
46. Waqas M, Zhang S, He Z, Tang M, Chai R. Role of Wnt and Notch signaling in regulating hair
cell regeneration in the cochlea. Front Med. 10(3):237-49 (2016).
47. Yamaguchi-Kabata Y, Morihara T, Ohara T, Ninomiya T, Takahashi A, Akatsu H, Hashizume
Y, Hayashi N, Shigemizu D, Boroevich KA, Ikeda M, Kubo M, Takeda M, Tsunoda T.
Integrated analysis of human genetic association study and mouse transcriptome suggests
LBH and SHF genes as novel susceptible genes for amyloid-β accumulation in Alzheimer's
disease. Hum Genet. 137(6-7):521-533 (2018).
48. Yamashita T, Zheng F, Finkelstein D, Kellard Z, Carter R, Rosencrance CD, Sugino K, Easton
J, Gawad C, Zuo J. High-resolution transcriptional dissection of in vivo Atoh1-mediated hair
cell conversion in mature cochleae identifies Isl1 as a co-reprogramming factor. PLoS Genet.
14(7):e1007552 (2018).
49. Żak M, Klis SF, Grolman W. The Wnt and Notch signalling pathways in the developing
cochlea: Formation of hair cells and induction of regenerative potential. Int J Dev Neurosci.
47(Pt B):247-58 (2015).
50. Zhang Q, Liu H, McGee J, Walsh EJ, Soukup GA, He DZ. Identifying microRNAs involved in
degeneration of the organ of corti during age-related hearing loss. PLoS One. 8(4):e62786
(2013).
51. Zhang Q, Liu H, Soukup GA, He DZ. Identifying microRNAs involved in aging of the lateral
wall of the cochlear duct. PLoS One. 9(11): e112857 (2014).
52. Zheng J, Shen W, He DZ, Long KB, Madison LD, Dallos P. Prestin is the motor protein of
cochlear outer hair cells. Nature. 405(6783):149-55 (2000).
15
Figure 1: Expression of LBH in cochlear and vestibular hair cells. A: Cell type-specific expression
of Lbh mRNA in Deiters’ cells (D), pillar cells (P), IHCs (I) and OHCs (O), as well as in vestibular
hair cells (VHCs), and non-sensory cells (VSCs) during development. B: Fluorescent microscopy
picture of antibody staining of LBH protein in a cryosection of the cochlea from a P3 wildtype
mouse. Stria vascularis (SV), IHCs, OHCs, and greater epithelium ridge (GER) are marked. Bar:
10 µm. C: LBH expression (red) in the organ of Corti from a P12 wildtype mouse using optical
sectioning with confocal microscopy. D: Lack of LBH protein expression in hair cells in a P12
Lbh∆2/∆2 null mouse. Bars: 5 µm in C and D. E: Fluorescence microscopy picture of antibody
staining of LBH in P30 cochlear hair cells from wildtype mouse. F and G: Cryosection of the
utricle from a P12 wildtype mouse. The nuclei of vestibular hair cells (VHCs) and supporting cells
(VSCs) are marked white arrows. The area delineated by the white frame in F is displayed at
higher magnification in G. Bars: 5 µm in F and G.
16
Figure 2: Auditory function of Lbh-mutant mice. A: ABR thresholds of the three genotypes of
mice (color-coded) at 1 month-of-age. Eight mice for each genotype from three different litters
were used. B: DPOAE thresholds at 1 month. C: Representative CM responses together with
CAP measured in Lbh∆2/∆2 null (red) and Lbh+/+ wildtype (black) mice. 8 kHz tone bursts (80 dB
SPL) were used to evoke response. Peak-to-peak magnitude (mean ± SD, n= 6 per genotype) of
the CM is presented in the right panel. D: Representative EP measured from Lbh∆2/∆2 (red) and
Lbh+/+ (black) mice at 1 month. EP magnitude (mean ± SD, n = 6 per genotype) is also presented.
E: ABR thresholds at 3 months. F: DPOAE thresholds at 3 months.
17
Figure 3: Hair cell status in the cochlear and vestibular sensory epithelia. A: Confocal
micrographs of IHCs and OHCs labelled by anti-MYO7A-antibody. Images were obtained from an
apical and a basal region in the cochleae of Lbh∆2/∆2 null mice at P12 and 1 month. Bar: 10 µm
for all panels in A. B: IHC and OHC count (mean ± SD) from apical (A) and basal (B) regions of
four Lbh∆2/∆2 (pink color) and four Lbh+/+ (black lines) mice at P3, P12, 1 month and 3 months. C:
Utricle and crista ampulla of Lbh+/+ and Lbh∆2/∆2 mice at 3 months (top panel). Bar: 50 µm. Higher
magnification images of areas within white frames are presented in the bottom panels. Bar: 10
µm.
18
Figure 4: SEM micrographs of stereocilia bundles of cochlear hair cells in Lbh∆2/∆2 null mice. A:
Micrograph of stereocilia bundles from the low-apical region of a cochlea at P5. Bar: 10 µm. B
and C: Higher magnification images of the stereocilia bundle of an OHC (B) and an IHC (C) from
the basal turn of the same cochlea shown in panel A. Bar: 1 µm for B and C. D and E: Micrographs
of stereocilia bundles from an apical turn region (D) and basal turn region (E) from an 1 month-
old Lbh∆2/∆2 mouse. Asterisk marks a missing OHC while black arrows mark fusion of stereocilia.
Bar: 10 µm. A magnified image of the area within the white frame is highlighted in panel F. F, G
and H: Representative images of degenerating stereocilia bundles of OHCs (F,G), and an IHC
(H) from mid-basal turn region of an 1-month-old Lbh∆2/∆2 mouse. Black arrows in panel F indicate
near complete absorption of stereocilia bundles. Bars: 2 µm (F), 1.5 µm (G), and 2.5 µm (H).
19
Figure 5: OHC function examined using whole-cell voltage-clamp technique. A: Recording of
MET current in vitro and representative MET current recorded from OHCs in lower apical turn of
Lbh∆2/∆2 (red) and Lbh+/+ (black) mice at P12. Means and SDs are plotted in the right panel.
Asterisk marks statistical significance (p<0.05). Bar: 5 µm. B: NLC measured from 9 and 8 OHCs
in the lower apical turn of Lbh∆2/∆2 (red) and Lbh+/+ (black) mice, respectively, at P12. Curve fitting
using a two-states Boltzmann function yielded four parameters: Qmax, slope (ɑ), Vpkcm, and Clin.
The means and SDs from the two types of OHCs are plotted in the right panels. Student’s t-tests
yielded p = 0.29, 0.33, 0.47 and 0.42, respectively, for the four parameters.
20
Figure 6: RNA-seq transcriptome analysis of Lbh∆2/∆2 and Lbh+/+ OHCs. A: Workflow of the
experimental design for RNA-seq analysis of OHCs isolated from Lbh∆2/∆2 and Lbh+/+ mice. B:
Euclidean distance heatmap of 10,000 genes (Z-score cutoff = 4), depicting average linkage
21
between genes expressed in Lbh∆2/∆2 and Lbh+/+ OHCs. C: Upregulated and downregulated genes
in Lbh∆2/∆2 compared to Lbh+/+ OHCs. The top 20 genes up- or down-regulated are shown on
either side of the plot. D: ShinyGO biological processes enriched in upregulated genes in Lbh∆2/∆2
compared to Lbh+/+ OHCs. E: ShinyGO biological processes enriched in downregulated genes in
Lbh∆2/∆2 compared to Lbh+/+ OHCs.
22
Figure 7: Gene set enrichment analysis (GSEA) of Lbh∆2/∆2 and Lbh+/+ OHCs transcriptomes.
Enriched pathways (FDR < 0.25) Lbh∆2/∆2 null OHCs include regulation of Wnt signaling (A), Notch
signaling (B), cell cycle (C), nucleic acid-templated transcription (D), DNA damage/repair (E), and
autophagy (F). The total numbers of up- (red) and down- (green) regulated genes within each
23
pathway are indicated, and the top 20 genes in each category are listed on either side of the
graph, with greatest to least fold change in downward direction (arrow). G: Validation of
differentially expressed cell survival genes using RT-qPCR. Log2 fold changes (Lbh∆2/∆2 vs.
Lbh+/+) from RNA-seq and ∆∆Ct values (normalized to Gapdh) from RT-qPCR for each gene are
shown.
Other supplementary materials for this manuscript include the following:
Datasets (in Excel format): RNA-seq dataset of transcriptomes of Lbh-null and wildtype outer hair
cells. Differentially expressed genes as well as significantly enriched genes associated with Wnt
and Notch signaling, transcription, cell cycle, DNA damage/repair, and autophagy in the Lbh-null
OHCs are also included.
| 2020 | Transcription Co-Factor LBH Is Necessary for Maintenance of Stereocilia Bundles and Survival of Cochlear Hair Cells | 10.1101/2020.05.13.093377 | [
"Liu Huizhan",
"Giffen Kimberlee P.",
"M’Hamed Grati",
"Morrill Seth W.",
"Li Yi",
"Liu Xuezhong",
"Briegel Karoline J.",
"He David Z."
] | null |
Highlights
Toolkit for Oscillatory Real-time Tracking and Estimation (TORTE)
Mark J Schatza,Ethan B Blackwood,Sumedh S Nagrale,Alik S Widge
• TORTE provides a toolkit to investigate closed loop oscillation-informed experiments.
• The toolkit is versatile and open-source promoting replicability across scientists.
• The analytic signal algorithm within TORTE preforms equally to existing algorithms.
Toolkit for Oscillatory Real-time Tracking and Estimation (TORTE)
Mark J Schatzaa, Ethan B Blackwooda, Sumedh S Nagralea and Alik S Widgea
aUniversity of Minnesota, Minneapolis, MN
A R T I C L E I N F O
Keywords:
Closed-Loop
Oscillations
Toolkit
Analytic Signal
Translational
Open-Source
A B S T R A C T
Background
Closing the loop between brain activity and behavior is one of the most active areas of development
in neuroscience. There is particular interest in developing closed-loop control of neural oscillations.
Many studies report correlations between oscillations and functional processes. Oscillation-informed
closed-loop experiments might determine whether these relationships are causal and would provide
important mechanistic insights which may lead to new therapeutic tools. These closed-loop perturba-
tions require accurate estimates of oscillatory phase and amplitude, which are challenging to compute
in real time.
New Method
We developed an easy to implement, fast and accurate Toolkit for Oscillatory Real-time Tracking
and Estimation (TORTE). TORTE operates with the open-source Open Ephys GUI (OEGUI) system,
making it immediately compatible with a wide range of acquisition systems and experimental prepa-
rations.
Results
TORTE efficiently extracts oscillatory phase and amplitude from a target signal and includes a
variety of options to trigger closed-loop perturbations. Implementing these tools into existing exper-
iments is easy and adds minimal latency to existing protocols.
Comparison with Existing Methods
Most labs use in-house lab-specific approaches, limiting replication and extension of their experi-
ments by other groups. Accuracy of the extracted analytic signal and accuracy of oscillation-informed
perturbations with TORTE match presented results by these groups. However, TORTE provides ac-
cess to these tools in a flexible, easy to use toolkit without requiring proprietary software.
Conclusion
We hope that the availability of a high-quality, open-source, and broadly applicable toolkit will
increase the number of labs able to perform oscillatory closed-loop experiments, and will improve the
replicability of protocols and data across labs.
1. Introduction
1.1. Importance of Oscillations
Oscillations in continuous neural data are implicated in a
wide range of functional processes, including decision mak-
ing, learning and memory, sensory coordination, and emo-
tion regulation. Dominant theories argue that cross-regional
oscillatory synchrony (phase-phase and/or phase-amplitude
coupling) enables and may be necessary for inter-regional
communication (Engel et al., 2001; Buzsáki et al., 1994).
That causal model has not yet been proven, as most prior
work only shows correlations between oscillations, synchrony,
and behavior. There is still a strong possibility that oscilla-
tions have no causal role, but are solely epiphenomena of
spike-level processes (Schneider et al., 2020; Wilson et al.,
2018; Tort et al., 2018). On the other hand, some early re-
sults suggest that oscillation-informed perturbations can al-
ter brain circuit function in ways that are not possible with
oscillation-blind approaches. Phase-locked stimulation can
induce plasticity (Zanos et al., 2018; Zrenner et al., 2018), as
can stimulation optimized to interact with a dominant cross-
regional oscillation (Lo et al., 2020). Stimulation locked to
the amplitude of a tremor-related oscillation can be more ef-
schat107@umn.edu (M.J. Schatza); ethanbblackwood@gmail.com (E.B.
Blackwood); nagra007@umn.edu (S.S. Nagrale); awidge@umn.edu (A.S.
Widge)
ORCID(s):
ficient in suppressing that tremor (Rosin et al., 2011; Bronte-
Stewart et al., 2009), as can stimulation delivered at spe-
cific phases of a tremor cycle (Cagnan et al., 2019). Sim-
ilar phase-aware approaches may be useful in manipulating
circuits relevant to psychiatric illness (Herman and Widge,
2019; Widge and Miller, 2019; Kanta et al., 2019; Knudsen
and Wallis, 2020).
1.2. Difficulty in Creating Closed Loop
Experiments
Early closed-loop results are promising but highlight a
major challenge in broadly testing causal claims about oscil-
latory synchrony – the need for accurate real-time estimates
of an oscillation’s state. To demonstrate that cross-region or
within-region phase-phase, or phase-amplitude, phenomena
are causally linked to a functional process, neuroscientists
and clinicians need tools to perturb those phenomena and/or
to lock stimuli to specific oscillatory events. The standard
algorithm to extract oscillatory features, the Hilbert trans-
form (Cohen, 2014), is not well suited to a real-time situa-
tion. A Hilbert transform requires a large window of data
around the time of interest, or else edge effects appear in
its output. Individual labs have addressed this problem by
developing special-purpose hardware and/or alternate algo-
rithms, each with limitations. Some approaches have esti-
mation inaccuracies that are too large to provide consistent
results (Siegle and Wilson, 2014). Others are accurate but
MJ Schatza et al.: Preprint submitted to Elsevier
Page 1 of 14
TORTE
require special hardware that is not easily maintained with-
out dedicated engineering staff. They may be prohibited by
high cost and may be difficult to implement into existing ex-
perimental protocols (Kanta et al., 2019; Rodriguez Rivero
and Ditterich, 2021; Escobar Sanabria et al., 2020; Zrenner
et al., 2018; Shirinpour et al., 2020). Many solutions are
built atop proprietary, closed-source software such as MAT-
LAB and its toolboxes (Hassan et al., 2020; Zelmann et al.,
2020). These factors greatly limit reproducibility. Further,
many existing systems are only capable of identifying peak
or trough phases of an ongoing oscillation (Siegle and Wil-
son, 2014; Rodriguez Rivero and Ditterich, 2021). They
cannot support other paradigms such as detecting interme-
diate phases (Zanos et al., 2018), estimating phase response
curves (Ermentrout et al., 2012; Holt et al., 2014) or oscilla-
tory amplitude. There is a need for a toolkit that provides ac-
curate oscillatory calculations in a pure software solution (to
maximize flexibility) and that can readily be implemented in
many labs and experimental settings.
1.3. Introducing TORTE
Here we provide a Toolkit for Oscillatory Real-time Track-
ing and Estimation (TORTE) that enables closed-loop oscil-
latory experiments. This toolkit implements a real-time al-
gorithm to extract the analytic signal of the continuous neu-
ral data and is built within the Open Ephys GUI (OEGUI)
system (Siegle et al., 2017). We developed TORTE with
the intention of providing an easy to use, flexible and accu-
rate system for scientists across a broad range of disciplines.
OEGUI is interoperable with a variety of recording setups
commonly used for rodent and non-human primate experi-
ments, supports next-generation high-density silicon probes
such as Neuropixels, and has been integrated with both inva-
sive and non-invasive human recordings (Schatza and Black-
wood, 2020; Black et al., 2017). A single software process-
ing chain usable across many different preparations could ac-
celerate scientific progress, just as open-source neural anal-
ysis toolkits have improved both speed and reproducibility
(Oostenveld et al., 2011; Boki et al., 2010).
TORTE enables closed-loop experiments where pertur-
bations are locked to arbitrary values of either the oscilla-
tory phase or amplitude of continuous neural data. Fig. 1A
presents an example of locking an event to the 180° phase
of the slow frequency component of a local field potential
(LFP) recording. This event in slow wave oscillations dur-
ing sleep was used to trigger an auditory stimulus which
led to enhanced memory consolidation (Ngo et al., 2013).
When paired with transcranial magnetic stimulation (TMS),
this same event in the alpha band facilitated long term po-
tentiation in humans (Zrenner et al., 2018). In Fig. 1B, we
depict a perturbation being presented during a time when the
oscillation is at a high amplitude. Using the amplitude of
a motor cortical oscillation to trigger stimulation has been
used to create a brain computer interface to restore motor
function in nonhuman primates (Fetz, 2015), and in humans
using the cortico-thalamic circuit to suppress tremor (Opri
et al., 2020; Bronte-Stewart et al., 2009). These are exam-
Figure 1: A) Events triggered at trough (180°) of low frequency
oscillation. B) Event triggered by high power activity of a low
frequency oscillation.
ples of two event targets and a few output stimuli. The versa-
tility of TORTE allows users to target events for any phase or
amplitude, at any frequency band in which true oscillatory
activity occurs. This event can then be used to trigger any
relevant perturbation, which may include presentation of a
task stimulus, delivery of a reward/outcome, or direct brain
electrical/optical/magnetic stimulation.
MJ Schatza et al.: Preprint submitted to Elsevier
Page 2 of 14
TORTE
Figure 2: A) Overview of the software (Open Ephys System). B) Closed-loop hardware of the system includes an acquisition
system and output logic (Closed-Loop Hardware). C) The experiment includes the subject and the experimental stimuli presented
to them (Experiment).
2. Materials and Methods
2.1. TORTE Overview
TORTE provides closed-loop tools to lock perturbations
to neural oscillatory events. The toolkit, developed within
OEGUI, can be run on any standard lab grade computer, in-
cluding laptops for portable data acquisition, and can com-
municate with a variety of neural acquisition systems. The
toolkit and OEGUI are freely available and modifiable. The
GitHub repository includes extensive documentation on con-
figuration and typical use cases (Schatza, 2021b).
Fig. 2 provides a system overview of the toolkit, high-
lighting the three main components. The “Open Ephys Sys-
tem” extracts oscillatory features from the neural data (Fig.
2A). The “Closed-Loop Hardware” layer handles communi-
cation between the experiment and OEGUI (Fig. 2B). The
“Experiment” includes the subject and the presented/delivered
perturbations (Fig. 2C). Starting with closed-loop hardware,
an acquisition system records continuous neural data. This
data is brought into OEGUI by a data interface plugin. Plug-
ins currently available include: Open Ephys acquisition box,
Alpha Omega intraoperative monitoring systems, Neuralynx
systems, Neuropixels, EEG via a custom interface board (Black
et al., 2017), and EEG using the LSL-inlet plugin (Schatza,
2021a). Additional systems are occasionally being added.
Collectively, these systems cover common platforms for hu-
man, non-human primate, and rodent recordings. The neural
data is passed on to the Real-Time Analytic Signal plugin.
This plugin outputs either the phase or amplitude of the sig-
nal. The Analytic Signal Crossing Detector plugin continu-
ously monitors the output from the Real-Time Analytic Sig-
nal plugin and triggers events when a threshold is crossed.
The threshold is set to either a specific phase of interest or
an amplitude value. Additional logic can denoise these sig-
nals if needed, e.g., by requiring the amplitude to cross and
remain on one side of a threshold for M of N samples. A
rudimentary artifact suppression algorithm is included that
limits the jump size between samples. We have not, at the
present time, implemented more advanced approaches such
as stimulation artifact template subtraction. These would
likely best be achieved by separate plugins, to maximize use
of OEGUI’s modular design. TORTE is best used for sit-
uations where the presented perturbation does not induce
artifact (e.g., optical or sensory stimulation with electrical
recordings), or in situations where a long recovery time can
MJ Schatza et al.: Preprint submitted to Elsevier
Page 3 of 14
TORTE
be given between pulses for amplifiers to settle. The Event
Broadcaster plugin then uses the common, very low latency
interprocess communication framework ZeroMQ (0MQ, 2021)
to output the Crossing Detector event to the closed-loop hard-
ware using a publisher/subscriber mode of communication.
Any of the 26 programming languages that ZeroMQ sup-
ports can be used to create output logic to receive this event.
Output logic code is not provided within this toolkit, as it is
heavily dependent on the specific perturbation to be deliv-
ered. However, example code that receives ZeroMQ events
in Python can be found in the provided OEGUI Python tools
repository (Siegle, 2017) and LabVIEW code can be found
on our GitHub (Blackwood, 2021). We specifically include
an example of how to generate a 5 V rising-edge square pulse,
the most common signal used to trigger both brain stimula-
tion and task-related hardware, but the output logic can be
designed to create stimuli of any kind. Further, ZeroMQ
supports communication within a single computer (e.g., for
controlling physiology and direct brain stimulation in closed-
loop) or network communication (e.g., for synchronizing mul-
tiple experimental machines via standard Internet protocols).
This decoupling of detection from output allows easier in-
corporation of this toolkit into existing experimental proto-
cols. For versatility and computational efficiency OEGUI
and TORTE are C++ based. To access the toolkit the code
can either be compiled from source for advanced users, or
simply installed using binaries provided across all major plat-
forms.
To improve accuracy in phase-locked closed-loop exper-
iments and improve data analysis, it is recommended to pro-
vide feedback of perturbation timing back into the system.
There are a few methods to implement this feedback. It is
recommended to provide either a digital input into the sys-
tem or send a software TTL pulse to OEGUI with ZeroMQ.
If neither of these are possible, an alternative technique would
be to send perturbation event markers through an analog in-
put (e.g., by consuming one recording channel). Two types
of pulses are typically sent, a perturbation pulse (Fig. 2 red
arrow) and a sham pulse (Fig. 2 orange arrow). The per-
turbation pulse is tightly time-locked to the simultaneously
acquired neural data and can be used in later analysis to ex-
tract data at the time of perturbation. The sham pulse can be
used to improve accuracy in phase-locked closed-loop ex-
periments in real time and/or used to verify the oscillations’
status during the event trigger timing without perturbation
related artifacts. The sham pulse is sent in place of present-
ing a perturbation, but with the same timing of a perturbation
pulse. If used to improve accuracy in phase-locked closed-
loop in real time, the Analytic Signal Plugin and Analytic
Signal Detector are set up to listen for these events. As de-
scribed further below, this enables a self-adjusting algorithm
that compensates for experimental hardware latency and bias
in phase estimates.
2.2. Analytic Signal Calculation
TORTE transforms continuous neural data into its ana-
lytic signal in real-time utilizing a Hilbert transformer. The
Analytic Signal Plugin GUI allows a user to adjust the al-
gorithm for their experiment. Fig. 3 shows the flowchart for
the Hilbert transformer (Fig. 3A), how the customizable val-
ues on the plugin’s user interface affects the algorithm (Fig.
3B), and the corresponding output (Fig. 3C).
Fig. 3A provides a flow chart of the algorithm. It starts
with raw continuous neural data, e.g., a single EEG or LFP
channel. This may be a derived channel, e.g., a bipolar-
referenced pair as in (Zelmann et al., 2020) or local Lapla-
cian as in (Zrenner et al., 2018). This data is causally filtered
through a 2nd order forward Butterworth bandpass filter to
extract a frequency of interest. For the next steps of the algo-
rithm, the data is downsampled to 500 Hz. An autoregres-
sive (AR) model then predicts enough samples ahead in time
to compensate for the Hilbert transformer’s group delay, as
described below, similar in principle to (Blackwood et al.,
2018). The AR order and model refresh rate for comput-
ing the AR model coefficients can be configured by the user
to achieve the optimal efficiency/accuracy tradeoff for their
system. With default parameters, the model coefficients are
computed using the last 1 second of data and are updated ev-
ery 50 ms for a 20th order AR model. Decreasing the update
frequency and the order of the model can greatly improve
computational efficiency. It is expected that no artifacts are
present within the data used for the AR prediction during
event detection. With default parameters this would be 20
samples at 500 Hz or 40 ms of data. To ensure no artifacts are
present, the user should enable a timeout/lockout period be-
tween successive detections and perturbations. This is an in-
cluded feature of the TORTE plugins. A Hilbert transformer
is then applied to the predicted and observed values, return-
ing the imaginary component at quadrature with the data.
The Hilbert transformer is a finite impulse response (FIR) fil-
ter that has a phase response with a constant group delay (off-
set) equal to half the filter order. The AR model predicts the
bandpassed signal sufficiently far into the future to compen-
sate for this delay. TORTE currently includes five Hilbert
transformers that provide well-behaved amplitude responses
that are close to flat in the band of interest and are reason-
ably flat and suppressed outside the band of interest. Filters
are provided for oscillatory bands of alpha/theta (4-18 Hz),
beta (10-40 Hz), low gamma (30-55 Hz), mid gamma (40-
90 Hz), and high gamma (60-120 Hz). Users can add new
filters that provide a better response for their band of inter-
est into the plugin. This could be used to extend TORTE
for detection of, e.g., sharp wave ripples near 200 Hz. In-
structions are provided in the TORTE GitHub repository to
implement a new filter, which is trivial to add to the plu-
gin once the filter has been created. Parks-McClellan op-
timal FIR filter design, e.g., MATLAB’s firpm or Python’s
scipy.signal.remez function, is generally adequate. The al-
gorithm downsamples to 500 Hz so filters up to 250 Hz can
easily be created. The downsampling frequency was cho-
sen as it provides good tradeoffs for efficiency, accuracy and
range of frequencies. Motivated users could adjust this sam-
pling rate for highly specific needs, but it is not an easy to
adjust feature of this toolkit.
MJ Schatza et al.: Preprint submitted to Elsevier
Page 4 of 14
TORTE
Figure 3: A) Flow chart overviewing the algorithm transforming neural data into oscillatory activity. B) GUI for the analytic
signal plugin. C) GUI elements showing output of analytic signal plugin.
The band selection region (green) of the phase calculator
can be used for updating the frequency band of interest for
the initial bandpass. The filter selection region (purple) is
used to choose which of the Hilbert transformers best fits
the frequency band of interest. The AR configuration region
(yellow) is used to adjust the efficiency/accuracy tradeoff for
the AR model. The output selection region (blue) allows the
user to select either phase or amplitude for the output. The
output visualization region (red) shows real time accuracy if
using phase-locked closed-loop.
TORTE was compared to the standard peak/trough algo-
rithm built into OEGUI (Siegle et al., 2017). The standard
algorithm can compute phase, but notably cannot compute
amplitude, of a target signal. Further, it can only detect 0,
90, 180, or 270° phase events. To find the target, the stan-
dard algorithm bandpasses the data down to the frequency of
interest and then detects zero crossings or slope inversions.
Further optimization of this type of algorithm is possible,
but the standard algorithm tested in this publication is a di-
rect representation of the next best option in the current Open
Ephys environment.
2.3. Learning Algorithm
TORTE is interoperable with a wide range of systems
and output hardware. Each laboratory setup will have unique
sources of latency between the triggering oscillatory event
and the delivery of the matched perturbation. To improve
phase-locked closed-loop experimentation and reduce the ef-
fects of such delay, TORTE includes a learning algorithm.
Fig. 4A shows an example intending to lock a perturbation
to a phase of 180º, but with an expected communication la-
tency that will produce an approximately 20° offset at the
MJ Schatza et al.: Preprint submitted to Elsevier
Page 5 of 14
TORTE
Figure 4: A) Flowchart of learning algorithm. Right side shows the crossing detector plugin being updated. Left side shows the
logical owchart for the learning algorithm. B) Crossing detector GUI showing variables con gurable for the learning algorithm.
center frequency of the desired band. Thus, initially events
are commanded to trigger when the phase crosses 160° to
account for the offset. The learning algorithm will itera-
tively improve this offset throughout the experiment, to opti-
mize perturbation delivery at the target phase. For common
cases of electrical/optical stimulation that cause recording
artifacts, learning requires sham pulses. For perturbations
that do not cause artifacts (e.g., oscillation-locked delivery
of sensory stimuli), the perturbation pulses may be used in-
stead. Fig. 4B then shows learning, as implemented in the
Analytic Signal Crossing Detector GUI. In this window, the
event channel, target phase, learning rate and other parame-
ters are configured. The event channel is set to track what-
ever source is receiving sham/perturbation pulse events. Af-
ter the event is received, TORTE waits for additional neural
data. It then performs an acausal calculation of phase using a
bidirectional filter and full (not approximated) Hilbert trans-
form. This acausal phase will be more accurate than the real
time estimate. Using the acausal phase calculation, TORTE
compares the phase at which the sham pulse arrived to the
target phase. The circular difference between the phases is
multiplied by the current learning rate to adjust the thresh-
old value. The learning rate decays over time as configured
by the user, and the process usually can converge in a few
minutes.
2.4. Real-time Feedback of Coherence and
Spectrogram
A likely use case for TORTE is closed-loop control of os-
cillations, and an experimenter may wish to verify that this
control is effective as the experiment progresses. TORTE
thus includes a Coherence and Spectrogram Viewer, which
displays either the coherence between multiple channels or
the spectrogram of individual channels. Both of these val-
ues can be determined from a time frequency representation
(TFR) of the data. See Fig. 5A for a flowchart describing
the TFR decomposition. The decomposition starts by stor-
ing data into a buffer of a size configured by the user. The
default buffer size is 8 seconds, which provides a reasonable
balance between estimation accuracy (number of oscillatory
cycles contained in a buffer) and frequency of updates. Any
buffer size above 4 seconds will provide reasonable calcula-
tions for both coherence and spectrogram at most frequency
bands. Once the buffer is filled, a 2 second Hann window is
used to perform TF decomposition using a sliding window
Fourier transform. With the TFR calculated, the power and
covariance (cross spectral) matrices for the buffer can be cal-
culated. Because TFR windows are calculated in real time,
the user can choose to either weight these matrices linearly
over the entire experiment or can use exponential weighting
to emphasize recent changes in activity. Channels, frequen-
MJ Schatza et al.: Preprint submitted to Elsevier
Page 6 of 14
TORTE
Figure 5: A) Flowchart describing calculation of real-time feedback using a TFR decomposition to generate either a spectrogram
or coherence plot. B) Example screenshot of the UI for the coherence spectrogram visualization. The snapshot shown here is to
demonstrate what a user may expect to see when using the plugin.
cies of interest, and weighting factors are all configurable in
this plugin. With the TFR computed, the plugins can show
either the coherence between two groups of channels, or the
spectrogram of the channels selected.
2.5. Experimental Validation
Four data sets were used to assess the algorithms de-
scribed above. These datasets will be called Rodent, EEG,
Simulated and Human. The Rodent dataset consists of 5
minute LFP recordings collected from the infralimbic cor-
tex (IL) and basolateral amygdala (BLA) in freely behaving
Long Evans rats (Lo et al., 2020). LFP were acquired con-
tinuously at 30 kHz (OpenEphys, Cambridge, MA, USA).
An adaptor connected the recording head stage (RHD 2132,
Intan Technologies LLC, Los Angeles, CA, USA) to two
Millmax male-male connectors (8 channels each, Part num-
ber: ED90267-ND, Digi-Key Electronics, Thief River Falls,
MN, USA). Sham and stimulation pulses were triggered at
180° for a 4-8 Hz oscillation. Pulses were commanded by
a Python application which received events from an Open
Ephys Event Broadcaster plugin. Sham pulses were used
to assess accuracy post hoc. This dataset was procured in
compliance with relevant laws and institutional guidelines;
all animal procedures were reviewed and approved by the
University of Minnesota IACUC. The EEG dataset consisted
of two single channel 30 minute recordings located over the
left prefrontal cortex (Brain Vision, Morrisville, NC, USA
). This dataset was procured in compliance with relevant
laws and institutional guidelines; all EEG recordings were
reviewed and approved by the University of Minnesota IRB.
MJ Schatza et al.: Preprint submitted to Elsevier
Page 7 of 14
TORTE
The Human dataset was collected from two individuals with
refractory epilepsy (1 male) undergoing invasive monitoring
as part of their clinical care at the Northwestern Memorial
Hospital Comprehensive Epilepsy Center. sEEG depth elec-
trodes (~1 mm diameter, ~2 mm contact length; AD-Tech
Medical Instruments Co., Oak Creek, WI) were implanted
according to clinical need prior to participation. Record-
ings were acquired using a Neuralynx ATLAS system with a
scalp electrode reference and ground (Neuralynx, Boseman,
MT). FIR digital bandpass filters were applied from 0.1 to
5000 Hz at the time of recording. Data were recorded at
a resolution of 0.15 µV (5000 µV input range) and a sam-
pling rate of 20 kHz, and were subsequently downsampled
to 1 kHz for analysis. Data were streamed from the ATLAS
to a separate computer running OEGUI via fiber optic ca-
ble. Sham and stimulation pulses were triggered at 180° for
a 4-8 Hz oscillation through the OEGUI machine’s parallel
port, using a custom OE plugin. Sham pulses were used to
assess accuracy post hoc. The Simulated dataset was created
in MATLAB by creating a sin wave with the frequency of in-
terest with an amplitude that varies slightly over time. Pink
noise was added on top of the sin wave with an amplitude
equal to the peak amplitude from the frequency of interest.
An example of a single Rodent and EEG dataset are avail-
able in the GitHub repository with Human data accessible
by request.
For saline testing one channel of data within the infral-
imbic cortex of the Rodent dataset from each recording was
converted into an MP3 file and played through an auxiliary
port of the test computer with Audacity®. The exposed male
end of the auxiliary cable was placed in a saline bath to emu-
late a brain. Recordings from the saline were then taken us-
ing the same headstage and electrodes as the in-vivo setup,
but now measuring the replayed LFP data. In this testing
setup, we did not deliver electrical stimuli or other perturba-
tions back into the saline, but sent a 1 V rising-edge square
pulse to the Open Ephys Acquisition system to track the tim-
ing of the detection events. To further characterize TORTE’s
performance across a wide range of frequencies and condi-
tions, and to collect more data points per condition, we fur-
ther implemented a software simulation of this saline test.
For the simulation, Rodent, EEG and Simulated datasets were
loaded into MATLAB and processed by a MATLAB imple-
mentation of TORTE’s real-time buffered processing. Sys-
tem latencies between components were simulated by gen-
erating random numbers from a gamma distribution whose
peak matched the median latency observed in the saline tests.
We verified that the simulated saline test produced identical
results to its physical counterpart. However, because data
buffers could be "acquired" faster than real time in the simu-
lation, we could complete each test thousands of times more
quickly and use more datasets.
The MATLAB simulation was used to compare the TORTE
and standard algorithm parameters by setting them to trig-
ger events targeted at 180° and at 300° for oscillations rang-
ing from 5 Hz to 55 Hz using the Simulated dataset. We
chose 180° because it has been a target in practical closed-
loop experiments (Zrenner et al., 2018) and 300° because it
is not easily detectable by peak-trough or zero-crossing de-
tectors. Phase-locked closed-loop experiments typically tar-
get lower frequency oscillations because they are commonly
implicated in functional processes (Watrous et al., 2015) and
their estimation is less affected by inherent system latencies.
Using the MATLAB replication software, a phase-locked
protocol was run targeting oscillations between 5 Hz and
55 Hz, with a step size of 1 Hz, for the two phase targets.
The difference between the ground truth phase at the sham
pulse time and the target phase was calculated. Ground truth
phase was calculated using the standard offline approach of
a forward-backward bandpass filter over the frequency band
of interest, followed by a Hilbert transform. The TORTE
Hilbert transformer method utilized its learning algorithm to
iteratively improve the threshold for event triggers. The stan-
dard algorithm was set in trough (180°) mode for both tar-
gets; this highlighted certain features of that algorithm more
clearly than setting it to 270° for the 300° target.
The MATLAB implementation was also used to simulate
an in-vivo experiment in both the Rodent and EEG datasets
to assess if any differences were present across recording
techniques/species in both algorithms. The standard algo-
rithm’s best parameters, targeting either 0, 90, 180 or 270°,
were used on both algorithms and tests were completed tar-
geting the theta band (4-8 Hz). Once again the learning al-
gorithm was implemented for TORTE. To assess the impact
on peak frequency within the oscillation of interest the spec-
trogram of the data across 1 minute windows was calculated
using the pwelch function in MATLAB. The corresponding
output was smoothed using a 5 minute gaussian window.
As a demonstration of how the overall system architec-
ture and performance can vary depending on the specific
acquisition hardware, and to further demonstrate TORTE’s
viability for human closed-loop experiments, we performed
a further test using an ATLAS human-grade electrophysio-
logic rig (Neuralynx, Bozeman, MT, USA). The Rodent data
was played into the analog input of the ATLAS system using
a USB DAQ and LabVIEW. The ATLAS system then broad-
cast the data as UDP packets over an Ethernet cable. A freely
available acquisition plugin (Schatza and Blackwood, 2020)
reassembled these packets as a datastream within OEGUI.
On phase detection event triggers, a 1 V rising-edge square
pulse was sent to the ATLAS system. ATLAS rebroadcasted
the received square pulses alongside the data within the UDP
packets.
All results were compared to a standard offline analysis
procedure in MATLAB using Fieldtrip (Oostenveld et al.,
2011). The MATLAB library CircStat (Berens, 2009) was
used to determine circular mean and circular standard devi-
ation (SD).
To assess the real-time feedback visualizations of power
and coherence, the Rodent dataset was replayed in OEGUI
using the file reader plugin with the IL and BLA channels
split into separate groups for coherence measurements. The
output for each trial was recorded by TORTE. The builtin
functions included in Fieldtrip, ft_freqanalysis and
MJ Schatza et al.: Preprint submitted to Elsevier
Page 8 of 14
TORTE
Table 1
TORTE and Standard algorithm results from the Rodent dataset in saline and Simulated
dataset in MATLAB.
Saline
MATLAB
MATLAB
MATLAB
MATLAB
4-8 Hz 180°
5 Hz 180°
55 Hz 180°
5 Hz 300°
55 Hz 300°
Mean TORTE
0.42°
-0.86°
-2.4°
-0.39°
-2.26°
SD TORTE
63.5°
16.42°
40.53°
11.9°
40.98°
Mean Standard
-49.00°
-50.71°
-178.03°
69.38°
-57.94°
SD Standard
67.81°
3.07°
26.55°
3.07°
26.55°
Table 2
TORTE results across recording types.
TORTE Rodent (In-Vivo)
TORTE Human
TORTE EEG (MATLAB)
Mean
0.42°
2.4°
0.50°
SD
63.5°
60.5°
80.56°
ft_connectivityanalysis, were used to generate coherence and
spectrogram outputs for the data at the same timepoints as
the TORTE output from the Rodent data.
3. Results
3.1. Results Overview
In this section the efficacy of TORTE in a standard lab
setup is shown. First we show that the system accurately
estimates phase and amplitude. We then describe sources
of latency within the system and how the online learning
algorithm reduces latency effects. Finally, the coherence
and power calculated by the casual monitoring algorithm are
compared to an acausal MATLAB implementation to show
adequate accuracy.
3.2. Phase Accuracy
Using the software replication experiment with the Sim-
ulated dataset, results for event phase accuracy at two phase
targets for TORTE and the standard algorithm are shown in
Fig. 6A-B and compiled in Table 1. TORTE triggers events
(n=200000) within 2.4° of the target phase and with less than
41° SD at all frequencies and both phase targets. Target-
ing the trough, the standard algorithm triggers events with
a mean error of 50° and a SD of 3° from the target phase at
lower frequencies, with worsening performance as frequen-
cies increase. While targeting 300° at the lowest frequency,
the standard algorithm starts 70° from target with a SD of 3°.
As the frequencies increase, the inaccuracy of the standard
algorithm caused by the latency of the system makes the al-
gorithm “accidentally” hit the target phase at around 40 Hz.
The user could use this to their advantage, but would greatly
limit target phase and frequency combinations.
The saline bath phase-locked closed-loop experiments
were set to target 180° phase within the 4-8 Hz band within
the selected IL data channel (n=115). As seen in Fig. 6C
a similar offset appears between the mean of the standard
algorithm event phases and the target phase. In-vivo exper-
iments with the same target are shown in Fig. 6D for both
Rodent (n=2) and Human (n=2) closed loop experiments.
Compiled results for the in-vivo experiments are found in
Table 2 and are comparable to saline and software tests.
Further MATLAB software implementation experiments
tested the standard algorithm’s best case targets of 0, 90,
180 and 270° for an oscillation of 4-8 Hz. These tests were
performed on the Rodent (Fig. 7A) and EEG (Fig. 7C)
datasets. The highest power frequency within the Rodent
dataset was found for each minute of the recordings (Fig.
7B). The peak oscillation frequency varies during the ex-
periment, but phase accuracy remains high. These three ex-
periments all verify the Hilbert transformer algorithm as be-
ing more versatile and more accurate than the standard algo-
rithm.
The TORTE results presented in Table 2 are also com-
parable to other approaches to high-accuracy phase predic-
tion. (Zanos et al., 2018) reported an equivalent mean er-
ror to TORTE, but with a 30° higher standard deviation of
phase accuracy. A similar AR prediction algorithm reported
a mean error of 1° and a SD of 53° from their target phase
(Zrenner et al., 2018). An alternative technique called Edu-
cated Temporal Prediction (ETP) has been proposed that es-
timates the phase oscillations and makes an educated guess
at phase timings in the future. This method again yielded
similar results, with a mean phase error of 0.37° and a SD of
67.35° (Shirinpour et al., 2020). Compiled results are shown
in Table 3, emphasizing that the differences between each al-
gorithm’s phase-locking performance are numerically quite
small.
3.3. Amplitude Accuracy
Using the software replication setup with both the Ro-
dent and EEG datasets, the real-time amplitude output from
the TORTE Hilbert Transformer algorithm was recorded.
The resulting output was compared sample by sample with
an amplitude ground truth calculation using the standard of-
fline Hilbert transform approach. Fig. 7D shows the differ-
ence in amplitude between both outputs as the percentage
of the three sigma envelope of all amplitude values. Ampli-
MJ Schatza et al.: Preprint submitted to Elsevier
Page 9 of 14
TORTE
Table 3
Comparing reported results of state-of-the-art real-time phase estimates.
TORTE
Zanos et al. (2018)
Zrenner et al. (2018)
Shirinpour et al. (2020)
Mean
0.42°
41°
1°
0.37°
SD
63.5°
66°
53°
67.35°
Figure 6:
A) Comparison of circular distance from target phase in MATLAB across frequencies for TORTE and the standard
algorithm targeting 180° and B) 300°. Standard error of the mean is shown as a shaded region, but is too small to be seen in the
gure. C) Circular distance from 180° phase target in the saline bath setup. D) Circular distance from target phase in Rodent
and Human in-vivo experiments.
tude differences are minimal across the experiment for both
datasets.
3.4. Latency
A wide variety of acquisition systems can provide data
to the OEGUI, each with unique latency and jitter, ranging
from µs to ms. As an example, we show how the latency of
the Open Ephys acquisition board is driven by its USB-based
communication, and compare this to the Ethernet-based Neu-
ralynx ATLAS. Fig. 8A shows the latency between an event
occurring in the neural data and a 1 V rising-edge square
pulse being sent back to the preparation in response. The
Ethernet-based system has a lower mean latency and much
narrower spread. Fig. 8B shows the processing time of the
TORTE algorithm for a buffer with 18.3ms of neural data.
Using these two data sets, we can determine what percent-
age of the real-time closed-loop latency is attributable to the
TORTE algorithms. TORTE’s internal calculations com-
prise about 0.9% of the latency in the Open Ephys acqui-
sition system and 4% of the latency in the Neuralynx AT-
LAS, demonstrating that the majority of latency lies in inter-
system communication.
3.5. Learning Algorithm
TORTE uses a learning algorithm to improve the accu-
racy of its phase targeting in the presence of system latency,
estimation errors and phase bias. Fig. 8C shows phase-
locking performance over time in our saline preparation, again
targeting 180° in the theta (4-8 Hz) band within the selected
IL data channel. Over the first 200 events, accuracy im-
proves by >10°, then stays consistent for the remainder of
the experiment. As the learning rate approaches zero during
the experiment, the standard deviation decreases.
3.6. Real-time Feedback
TORTE’s real-time visualizations closely approximate
acausal calculations of the same signals. The mean differ-
ence of the coherence output is 0.0065 with a SD of 0.0258
MJ Schatza et al.: Preprint submitted to Elsevier
Page 10 of 14
TORTE
Figure 7:
A) Comparison of circular distance from target phase in the theta band (4-8 Hz) using the MATLAB simulation
for TORTE and the standard algorithm, targeting 0, 90, 180, and 270° in the Rodent dataset. B) Peak frequency within the
theta band over time. C) Comparison of circular distance from target phase in the theta band using the MATLAB simulation
for TORTE and the standard algorithm, targeting 0, 90, 180, and 270° in the EEG dataset. E) Dierence in real-time estimated
amplitude to ground truth.
as shown in Fig. 8D. The scale of coherence values are 0 to
1 so these differences can be considered negligible. The dif-
ference in the spectrogram causal and acausal measurements
were calculated as a percentage of the envelope of the data.
The mean difference of the spectrogram is 0.0012% and a SD
of 0.0064% as shown in Fig. 8E. These represent small per-
centages of the full scale, demonstrating the validity of these
visualizers for real-time experimental performance tracking.
4. Discussion
4.1. TORTE
We have presented TORTE, a toolkit to enable scientists
to easily implement closed-loop experiments based upon os-
cillatory activity within continuous neural data. This fills a
resource gap by providing an open-source toolkit that can
readily be adapted into most existing experimental systems.
Further, TORTE is a sub-component of the larger open-source
framework of OEGUI. OEGUI takes a modular approach,
where plugins can be separately created and compiled with-
out dependence on the main package maintainer. Thus, TORTE
leverages other labs’ work to create plugins that stream in
data from several commonly used neural acquisition systems.
Although TORTE is a complete toolkit, the plugin architec-
ture also allows TORTE to be extended upon by other plug-
ins that provide additional functionality within OEGUI. An
example would be using a behavior to gate the presentation
of oscillation informed perturbations. A video tracking plu-
gin could pause the event output of TORTE during periods
of non-desired behavior such as grooming, and only allow
oscillation-informed perturbations during non-grooming pe-
riods. On top of the code being freely available with am-
ple documentation, our lab provides support for users imple-
menting TORTE and the Open Ephys team provides support
for using OEGUI. Potential applications include locking an
event to the 180° phase of a slow frequency component of
a LFP recording to modulate synchrony, sending perturba-
tions in response to high gamma power as a proxy for local
spiking rate, or stimulating at periods of high power in the
beta band to target tremor-related activity.
4.2. Limitations
The toolkit presented is easy to use and flexible, but does
have limitations. For processing efficiency, both OEGUI and
TORTE are developed in C++, which is extremely compu-
tationally efficient but not a commonly-used programming
MJ Schatza et al.: Preprint submitted to Elsevier
Page 11 of 14
TORTE
Figure 8: A) The latency between event trigger and perturbation delivery in two widely used acquisition systems.
B) The
latency between receiving a buer of data into the algorithm and return of the corresponding analytic signal. C) Phase accuracy
improvements over the experiment from the learning algorithm over 114 thirteen minute trials consisting of 200 perturbation
pulses. Standard error of the mean is shown as a shaded region. D-E) Histogram displaying the dierence between acausal
MATLAB and causal C++ implementation of the (D) coherence and (E) spectrogram of a single channel across 40 frequencies.
language among life scientists. Where possible, we have
made TORTE components easily configurable without a di-
rect code rewrite, but extracting the analytic signal from fre-
quency bands that fall outside of the provided configurations
requires knowledge of designing digital filters. TORTE uti-
lizes a Hilbert transformer algorithm which works well for
many use cases, however other algorithms may better suit
some users’ needs. Only a very rudimentary artifact sup-
pression technique is implemented as described, which is
sufficient for intermittent locking to low-frequency oscilla-
tions, but may not cover all use cases. Artifacts typically
show up as a phase reset, jump to zero phase, or a momentary
increase in amplitude. More advanced artifact suppression
techniques would need to be assessed on a case by case basis
according to recording technique wherein each will include
unique artifacts such as blinks, stimulation, etc. Finally, al-
though the toolkit is free and open-source, the user still needs
to “assemble” the experiment themselves including the out-
put logic for perturbation presentation. The TORTE team
can assist in this process, but may not be able to provide the
same level of support that a private company may provide
for its products. Several neural acquisition systems are sup-
MJ Schatza et al.: Preprint submitted to Elsevier
Page 12 of 14
TORTE
ported, but there are many that OEGUI cannot receive data
from yet. As shown, the toolkit is compatible and initial
testing has been completed with human invasive and non-
invasive systems. It is suitable for basic science experiments,
but TORTE is not currently suitable for clinical research, as
it has not undergone FDA-compatible design controls.
4.3. Algorithm Comparison
As shown in Results, the TORTE Hilbert transformer
algorithm provides improved phase accuracy compared to
the standard algorithm included in OEGUI. It also provides
real time amplitude information which the standard algo-
rithm does not. The Hilbert transformer algorithm works
well for many use cases, however other algorithms may bet-
ter suit some users’ needs. For instance, novel state-space
approaches (SSPE) have been proposed as a more principled
and reliable way to track oscillations (Wodeyar et al., 2021).
TORTE is extendable to new algorithms (we have imple-
mented an initial version of that state-space approach), but
each new approach would need to be converted into C++,
which is not trivial. The Hilbert transformer algorithm is
preferred when the data is narrow band and well character-
ized or the user wants to implement the simplest solution.
The novel SSPE algorithm should improve performance when
frequencies of interest are variable or multiple peaks exist
within a frequency band. The standard peak/trough detector
included in OEGUI may be sufficient in the case of extremely
stable, near-sinusoidal oscillations, e.g., occipital eyes-closed
alpha.
4.4. Future Directions
TORTE is continuously being improved and extended
upon, driven by experiments in our own lab as well as our
collaborators. The current real-time analytic signal algo-
rithm is fast and reasonably accurate, but oscillatory signal
processing is rapidly advancing. For instance, latent vari-
able approaches may estimate and predict oscillatory signals
in ways that our current filtering approaches cannot (Yang
et al., 2021).
We expect to implement these innovations
into TORTE as they become available. Similarly, OEGUI it-
self is rapidly evolving as the Open Ephys platform spreads.
A second-generation hardware system will dramatically im-
prove latency by removing USB communication, but will
also re-factor OEGUI into the Bonsai architecture. TORTE
will be made compatible with these future evolutions, as
we intend to adopt them in our own experiments. Future
versions may also extend visualization of cross-region os-
cillatory synchrony to include phase-amplitude coupling or
spike-field locking based on real-time spike sorting.
4.5. Conclusion
TORTE provides a platform for rapidly and reproducibly
creating oscillation-informed closed-loop experiments. Such
experiments are already being implemented, in a prelimi-
nary fashion, in areas such as motor rehabilitation, epilepsy,
and movement disorders. They are theorized to be appli-
cable to understanding and developing treatments for more
complex domains such as mental disorders (Cho et al., 2015).
The availability of a common and flexible toolkit should make
these paradigms easier to apply for testing a wide variety of
brain functions, accelerating progress in both basic neuro-
science and clinical translation.
5. Acknowledgments
We would like to thank Dr. Joel Voss, Dr. James Kragel
and Sarah Lurie for assistance with generating the Human
dataset. We thank Dr. Mo Chen, Dr. Saydra Wilson and
Dr. Sarah Olsen for providing the EEG dataset. We thank
Dr. Meng-Chen Lo and Rebecca Younk for assistance with
generating the Rodent dataset. This work was supported by
the Brain & Behavior Research Foundation, Picower Family
Foundation, Kent and Liz Dauten Bipolar Disorders Seed
Fund at Harvard University, the MnDRIVE Brain Condi-
tions and Medical Discovery Team - Addictions initiatives
at the University of Minnesota, and the National Institutes
of Health (R21MH109722, R21MH113103, R01EB026938,
and R01MH119384).
5.1. Author Contributions
M.J.S., E.B.B, and A.S.W. designed the toolkit. M.J.S,
E.B.B., and S.S.N. implemented, programmed and preformed
analyses on the toolkit. M.J.S. and A.S.W. wrote the paper
with input from the other authors. All authors gave final ap-
proval to the paper.
5.2. Conflict of Interest
A.S.W. and E.B.B. are named inventors on granted and
pending patents related to oscillation-locked stimulation.
References
0MQ, 2021. Network data. https://zeromq.org/.
Berens, P., 2009. Circstat: A matlab toolbox for circular statistics. Journal
of Statistical Software, Articles 31, 1–21. URL: https://www.jstatsoft.
org/v031/i10, doi:10.18637/jss.v031.i10.
Black, C., Voigts, J., Agrawal, U., Ladow, M., Santoyo, J., Moore, C.,
Jones, S., 2017. Open ephys electroencephalography (open ephys EEG):
a modular, low-cost, open-source solution to human neural recording.
Journal of Neural Engineering 14, 035002. URL: https://doi.org/10.
1088/1741-2552/aa651f, doi:10.1088/1741-2552/aa651f.
Blackwood, E., 2021. closed-loop-stim. doi:10.5281/zenodo.5539699.
Blackwood, E., Lo, M., Widge, A., 2018. Continuous phase estimation
for phase-locked neural stimulation using an autoregressive model for
signal prediction. 2018 40th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC) , 4736–
4739doi:https://doi.org/10.1109/EMBC.2018.8513232.
Boki, H., Andrews, P., Kulkarn, J., Mehta, S., Mitra, P., 2010. Chronux a
platform for analyzing neural signals. Journal of neuroscience methods
doi:https://doi.org/10.1016/j.jneumeth.2010.06.020.
Bronte-Stewart, H., Barberini, C., Koop, M.M., Hill, B.C., Henderson,
J.M., Wingeier, B., 2009.
The STN beta-band profile in parkinson’s
disease is stationary and shows prolonged attenuation after deep brain
stimulation.
Experimental Beurology doi:https://doi.org/10.1016/j.
expneurol.2008.09.008.
Buzsáki, G., Bragin, A., Chrobak, J.J., Nádasdy, Z., Sik, A., Hsu, M., Yli-
nen, A., 1994. Oscillatory and intermittent synchrony in the hippocam-
pus: Relevance to memory trace formation, in: Buzsáki, G., Llinás, R.,
Singer, W., Berthoz, A., Christen, Y. (Eds.), Temporal Coding in the
Brain, Springer Berlin Heidelberg, Berlin, Heidelberg. pp. 145–172.
MJ Schatza et al.: Preprint submitted to Elsevier
Page 13 of 14
TORTE
Cagnan, H., Mallet, N., Moll, C.K.E., Gulberti, A., Holt, A.B., Westphal,
M., Gerloff, C., Engel, A.K., Hamel, W., Magill, P.J., Brown, P., Sharott,
A., 2019. Temporal evolution of beta bursts in the parkinsonian cortical
and basal ganglia network. Proceedings of the National Academy of
Sciences 116, 16095–16104. URL: https://www.pnas.org/content/116/
32/16095, doi:10.1073/pnas.1819975116.
Cho, K.K., Hoch, R., Lee, A.T., Patel, T., Rubenstein, J.L., Sohal, V.S.,
2015. Gamma rhythms link prefrontal interneuron dysfunction with cog-
nitive inflexibility in dlx5/6 mice. Neuron 85, 1332–1343. doi:https:
//doi.org/10.1016/j.neuron.2015.02.019.
Cohen, M., 2014. Analyzing Neural Time Series Data. The MIT Press.
Engel, A., Fries, P., Singer, W., 2001. Dynamic predictions: Oscillations
and synchrony in top–down processing. Nature Reviews Neuroscience
doi:https://doi.org/10.1038/35094565.
Ermentrout, G.B., Beverlin, B., Netoff, T., 2012. Phase Response Curves
to Measure Ion Channel Effects on Neurons. Springer New York, New
York, NY. chapter 1. pp. 207–236. doi:10.1007/978-1-4614-0739-3_9.
Escobar Sanabria, D., Johnson, L.A., Yu, Y., Busby, Z., Nebeck, S., Zhang,
J., Harel, N., Johnson, M.D., Molnar, G.F., Vitek, J.L., 2020. Real-
time suppression and amplification of frequency-specific neural activity
using stimulation evoked oscillations. Brain Stimulation 13, 1732–1742.
doi:https://doi.org/10.1016/j.brs.2020.09.017.
Fetz, E., 2015. Restoring motor function with bidirectional neural inter-
faces. Progress in Brain Research doi:https://doi.org/10.1016/bs.pbr.
2015.01.001.
Hassan, U., Pillen, S., Zrenner, C., Bergmann, T., 2020. P170 best toolbox:
Brain electrophysiological recording & stimulation toolbox.
Clinical
Neurophysiology 131, e109–e110. doi:10.1016/j.clinph.2019.12.281.
Herman, A.B., Widge, A.S., 2019. Dynamic network targeting for closed-
loop deep brain stimulation. Neuropsychopharmacology doi:https://
doi.org/10.1038/s41386-018-0210-x.
Holt, A., Wilson, D., Shinn, M., Moehlis, J., TI, N., 2014. Phasic burst
stimulation: A closed-loop approach to tuning deep brain stimulation
parameters for parkinson’s disease. Nature PLoS Computational Biology
Neuroscience doi:doi:10.1371/journal.pcbi.1005011.
Kanta, V., Pare, D., Headley, D., 2019. Closed-loop control of gamma
oscillations in the amygdala demonstrates their role in spatial memory
consolidation.
Nature Communications doi:https://doi.org/10.1038/
s41467-019-11938-8.
Knudsen, E.B., Wallis, J.D., 2020. Closed-loop theta stimulation in the
orbitofrontal cortex prevents reward-based learning. Nature Communi-
cations doi:https://doi.org/10.1016/j.neuron.2020.02.003.
Lo, M.C., Younk, R., Widge, A.S., 2020. Paired electrical pulse trains for
controlling connectivity in emotion-related brain circuitry. IEEE Trans-
actions on Neural Systems and Rehabilitation Engineering 28, 2721–
2730. doi:10.1109/TNSRE.2020.3030714.
Ngo, H., Martinetz, T., Born, J., Mölle, M., 2013. Auditory closed-loop
stimulation of the sleep slow oscillation enhances memory.
Neuron
doi:https://doi.org/10.1016/j.neuron.2013.03.006.
Oostenveld, R., Fries, P., Maris, E., Schoffelen, J., 2011. Fieldtrip: Open
source software for advanced analysis of MEG, EEG, and invasive elec-
trophysiological data.
Computational Intelligence and Neuroscience
doi:https://doi.org/10.1155/2011/156869.
Opri, E., Cernera, S., Molina, R., Eisinger, R.S., Cagle, J.N., Almeida, L.,
Denison, T., Okun, M.S., Foote, K.D., Gunduz, A., 2020. Chronic em-
bedded cortico-thalamic closed-loop deep brain stimulation for the treat-
ment of essential tremor. Science Translational Medicine 12. doi:10.
1126/scitranslmed.aay7680.
Rodriguez Rivero, C., Ditterich, J., 2021. A user-friendly algorithm for
adaptive closed-loop phase-locked stimulation. Journal of Neuroscience
Methods 347, 108965. doi:https://doi.org/10.1016/j.jneumeth.2020.
108965.
Rosin, B., Slovik, M., Mitelman, R., Rivlin-Etzion, M., Haber, S.N., Israel,
Z., Vaadia, E., Bergman, H., 2011. Closed-loop deep brain stimulation is
superior in ameliorating parkinsonism. Neuron 72, 370–384. doi:https:
//doi.org/10.1016/j.neuron.2011.08.023.
Schatza, M., 2021a. Lsl-inlet. doi:10.5281/zenodo.5539563.
Schatza, M., 2021b. Torte. doi:10.5281/zenodo.5539831.
Schatza, M., Blackwood, E., 2020. neuralynx-plugin. doi:10.5281/zenodo.
4900097.
Schneider, M., Dann, B., Swathi, S., Scherberger, H., Vinck, M., 2020. A
general theory of coherence between brain areas. bioRxiv doi:10.1101/
2020.06.17.156190.
Shirinpour, S., Alekseichuk, I., Mantell, K., Opitz, A., 2020. Experimen-
tal evaluation of methods for real-time EEG phase-specific transcra-
nial magnetic stimulation.
Journal of Neural Engineering doi:https:
//doi.org/10.1088/1741-2552/ab9dba.
Siegle, J., 2017. open-ephys-python-tools. https://github.com/open-ephys/
open-ephys-python-tools.
Siegle, J.H., López, A.C., Patel, Y.A., Abramov, K., Ohayon, S., Voigts,
J., 2017.
Open ephys: an open-source, plugin-based platform for
multichannel electrophysiology.
Journal of Neural Engineering 14,
045003. URL: https://doi.org/10.1088/1741-2552/aa5eea, doi:10.1088/
1741-2552/aa5eea.
Siegle, J.H., Wilson, M.A., 2014. Enhancement of encoding and retrieval
functions through theta phase-specific manipulation of hippocampus.
eLife doi:https://doi.org/10.7554/eLife.03061.
Tort, A., Brankačk, J., Draguhn, A., 2018.
Respiration-entrained brain
rhythms are global but often overlooked.
Trends in Neurosciences
doi:https://doi.org/10.1016/j.tins.2018.01.0070.
Watrous, A., Fell, J., Ekstrom, A., Axmacher, N., 2015. More than spikes:
common oscillatory mechanisms for content specific neural representa-
tions during perception and memory. Current Opinion in Neurobiology
doi:https://doi.org/10.1016/j.conb.2014.07.024.
Widge, A.S., Miller, E.K., 2019. Targeting cognition and networks through
neural oscillations: Next-generation clinical brain stimulation. JAMA
psychiatry doi:https://doi.org/10.1001/jamapsychiatry.2019.0740.
Wilson, C., Higgs, M., Simmons, D., Morales, J., 2018.
Oscillations
and spike entrainment. F1000Research doi:https://doi.org/10.12688/
f1000research.16451.1.
Wodeyar, A., Schatza, M., Widge, A.S., Eden, U.T., Kramer, M.A., 2021.
A state space modeling approach to real-time phase estimation. bioRxiv
doi:10.1101/2021.03.25.437024.
Yang, Y., Qiao, S., Sani, O., Sedillo, J., Ferrentino, B., Pesaran, B.,
Shanechi, M., 2021.
Modelling and prediction of the dynamic re-
sponses of large-scale brain networks during direct electrical stimu-
lation.
Nature Biomedical Engineering doi:https://doi.org/10.1038/
s41551-020-00666-w.
Zanos, S., Rembado, I., Chen, D., Fetz, E.E., 2018. Phase-locked stim-
ulation during cortical beta oscillations produces bidirectional synap-
tic plasticity in awake monkeys. Current Biology 28, 2515–2526.e4.
doi:https://doi.org/10.1016/j.cub.2018.07.009.
Zelmann, R., Paulk, A.C., Basu, I., Sarma, A., Yousefi, A., Crocker,
B., Eskandar, E., Williams, Z., Cosgrove, G.R., Weisholtz, D.S.,
Dougherty, D.D., Truccolo, W., Widge, A.S., Cash, S.S., 2020.
Closes:
A platform for closed-loop intracranial stimulation in hu-
mans. NeuroImage 223, 117314. URL: https://www.sciencedirect.com/
science/article/pii/S1053811920308004,
doi:https://doi.org/10.1016/
j.neuroimage.2020.117314.
Zrenner, C., Desideri, D., Belardinelli, P., Ziemann, U., 2018. Real-time
EEG-defined excitability states determine efficacy of TMS-induced plas-
ticity in human motor cortex. Brain Stimulation doi:https://doi.org/10.
1016/j.brs.2017.11.016.
MJ Schatza et al.: Preprint submitted to Elsevier
Page 14 of 14
| 2021 | Toolkit for Oscillatory Real-time Tracking and Estimation (TORTE) | 10.1101/2021.06.21.449019 | [
"Schatza Mark J",
"Blackwood Ethan B",
"Nagrale Sumedh S",
"Widge Alik S"
] | creative-commons |
Susceptibility rhythm to bacterial
endotoxin in myeloid clock-knockout mice
Veronika Lang1, Sebastian Ferencik1, Bharath Ananthasubramaniam1,2, Achim Kramer1, Bert
Maier1,*
1 Laboratory of Chronobiology, Charit´e Universit¨atsmedizin Berlin, Germany, 2 Institute for Theoretical Biology,
Humboldt-Universit¨at zu Berlin, Germany, * corresponding author: e-mail bert.maier@charite.de
Abstract
Local circadian clocks are active in most cells of our body. However, their impact on circadian
physiology is still under debate. Mortality by endotoxic (LPS) shock is highly time-of-day de-
pendent and local circadian immune function such as the cytokine burst after LPS challenge
has been accounted for the large differences in survival. Here, we investigate the roles of light
and myeloid clocks on mortality by endotoxic shock. Strikingly, mice in constant darkness
(DD) show a three-fold increased susceptibility to LPS as compared to mice in light-dark
conditions. Mortality by endotoxic shock as a function of circadian time is independent of
light-dark cycles as well as myeloid CLOCK or BMAL1 as demonstrated in conditional knock-
out mice. Unexpectedly, despite the lack of a myeloid clock these mice still show rhythmic
patterns of pro- and anti-inflammatory cytokines such as TNFα, MCP-1, IL-18 and IL-10 in
peripheral blood as well as time-of-day and site dependent traffic of myeloid cells. We spec-
ulate that systemic time-cues are sufficient to orchestrate innate immune response to LPS
by driving immune functions such as cell trafficking and cytokine expression.
Introduction
Timing of immune-functions is crucial for initiating, establishing, maintaining and resolving
immune-responses. The temporal organization of the immune system also applies for daily
recurring tasks and may even help to anticipate times of environmental challenges. In hu-
mans, many parameters and functions of the immune system display diurnal patterns [43],
which impact on disease severity and symptoms [13]. While the concepts of chronobiology
are increasingly acknowledged in life-science and medicine [5], a deep comprehension of
how time-of-day modulates our physiology in health and disease is still lacking.
The fundamental system behind the time-of-day dependent regulation of an organism,
its behavior, physiology and disease is called the circadian clock. In mammals, this clock is
organized in a hierarchical manner: a central pacemaker in the brain synchronized to environ-
mental light-dark cycles via the eyes and peripheral clocks receiving and integrating central
as well as peripheral (e.g. metabolic) time information. Both central and peripheral clocks
are essentially identical in their molecular makeup: Core transcription factors form a nega-
tive feedback loop consisting of the activators, CLOCK and BMAL1, and the repressors, PERs
and CRYs. Additional feedback loops and regulatory factors amplify, stabilize and fine-tune
the cell intrinsic molecular oscillator to achieve an about 24-hour (circadian) periodicity of
cell- and tissue-specific clock output functions [3].
Circadian patterns of various immune-functions have been reported in mice [43] and
other species [24,30,48] including cytokine response to bacterial endotoxin and pathogens,
white blood cell traffic [7, 38, 44] and natural killer cell activity [2]. Cell-intrinsic clocks
have been described for many leukocyte subsets of lymphoid as well as myeloid origin,
including monocytes/macrophages [28]. Furthermore, immune-cell intrinsic clocks have
been connected to cell-type specific output function such as the TNFα response to LPS
in macrophages [28].
Sepsis is a severe life threatening condition with more than 31 million incidences per year
worldwide [19]. Mouse models of sepsis show a strong time-of-day dependency in mortal-
1/15
ity rate when challenged at different times of the day [12,17,18,23,26]. Most studies agree
about the times of highest (around Zeitgeber time [ZT]8 - i.e. 8 hours past lights on) or low-
est (around ZT20) mortality across different animal models and investigators [17,18,23,26].
The fatal cascade in the pathomechanism of endotoxic shock is initiated by critical doses
of LPS recognized by CD14 bearing monocytes/macrophages leading to a burst of pro-
inflammatory cytokines. A viscous cycle of leukocyte recruitment, activation and tissue fac-
tor (III) expression triggers disseminated intravascular coagulation and blood pressure de-
compensation and final multi-organ dysfunction [37].
Here, we investigate the impact of light-dark cycles and local myeloid clocks on time-of-
day dependent survival rates in endotoxic shock using conditional clock-knockout mouse
models. We show that peripheral blood cytokine levels as well as mortality triggered by
bacterial endotoxin depend on time-of-day despite a functional clock knockout in myeloid
cells. Our work thus challenges current models of local regulation of immune responses.
Results
Time-of-day dependent survival in endotoxic shock
Diurnal patterns of LPS-induced mortality (endotoxic shock) have been reported numerous
times in different laboratory mouse strains [23, 36, 44]. To address the question, whether
time-of-day dependent susceptibility to LPS is under control of the circadian system rather
than being directly or indirectly driven by light, we challenged mice kept either under light-
dark conditions (LD 12:12) or in constant darkness (DD) at four different times during the
cycles. As expected from previous reports, survival of mice housed in LD was dependent on
the time of LPS injection, being highest during the light phase and lowest at night (Fig. 1A).
Surprisingly, mice challenged one day after transfer in constant darkness using the same
dose showed a more than 60% increase in overall mortality compared to mice kept in LD
(Fig. 1A). Furthermore, time-of-day dependent differences in mortality were much less pro-
nounced under these conditions.
To discriminate, whether constant darkness alters overall susceptibility to LPS leading to
a ceiling effect rather than eliminating time-dependent effects, we systematically reduced
LPS dosage in DD conditions. By challenging the mice at four times across the day we deter-
mined the half-lethal dose of LPS in constant darkness (Fig. 1B). As suspected, mice in DD
showed a 3-fold increased susceptibility to LPS-induced mortality. For circadian rhythm
analysis, we extended the number of time points at approximately half-lethal doses both in
LD and DD groups, respectively. This revealed diurnal/circadian patterns in mortality rate in
LD (p-value=0.06) as well as in DD conditions (p-value=0.001) (Fig. 1C) demonstrating that
the circadian system controls susceptibility to LPS. In addition, this susceptibility is overall
increased in constant darkness.
Circadian cytokine response upon LPS challenge in mice
We and others have recently found that cells in the immune system harbor self-sustained
circadian oscillators, which shape immune functions in a circadian manner [25, 28]. Pro-
inflammatory responses in murine ex vivo macrophage culture [4,28] are controlled by cell-
intrinsic clocks and are most prominent during the day and lowest during the activity phase.
Furthermore, pro-inflammatory cytokines such as TNFα, IL-1α/β, IL-6, IL-18 as well as MCP-1
have been linked to the pathomechanism of endotoxic shock [1,22,31,35,49].
Thus, we hypothesized that the cytokine response in LPS-challenged mice has a time-
of-day dependent profile, which might govern the time-of-day dependent mortality rates
in the endotoxic shock model. Indeed, plasma of mice collected two hours after adminis-
tratioin of half-lethal doses of LPS (either in LD (30mg/kg) or DD (13mg/kg) conditions at
various times during the day), exhibited an up to two-fold time-of-day difference in absolute
cytokine concentrations (Fig. 1D, and E for amplitude and phase informatin as determined by
sine fit). In animals kept in LD, TNFα, showed highest levels around ZT8, IL-18 levels peaked
at ZT18 and IL-12 as well as the anti-inflammatory cytokine IL-10 had their peak-time around
ZT14. Interestingly, cytokine profiles from DD mice differed substantially from LD profiles:
the peak-times of IL-12 was phase-advanced by 6 hours, CXCL5 completely reversed its
phase, whereas IL-18 remained expressed predominantly in the night. Taken together, cy-
tokine profiles of LPS-challenged mice parallel endotoxic shock-induced mortality patterns,
2/15
IL6
ZT8 CT8
0
20
40
60
80
100
***
MCP1
ZT8 CT8
0
10
20
30
***
plasma concentration (ng/ml)
IFN-
ZT8 CT8
0
5
10
15
20
25
**
plasma concentration (pg/ml)
LD
DD
0
20
40
60
80
100
mortality (%)
***
A
D
B
0
20
40
60
80
100
7.5
15 20
30
5
50
10
25
40
LPS (mg/kg)
mortality (%)
0
4
8
12
16
20
24
0
20
40
60
80
100
ZT/CT LPS injection (h)
mortality (%)
0
4
8
12
16
20
24
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0
4
8
12
16
20
24
0
20
40
60
80
100
ZT LPS injection (h)
mortality (%)
C
0
4
8
12
16
20
24
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0
4
8
12
16
20
24
0
20
40
60
80
100
CT LPS injection (h)
E
0
2
4
6
8
10
12
14
16
18
20
22
rel. amplitude
2.2
1.9
1.6
1.3
1
ZT/CT peak level (h)
IL-10
IL-18
IL-12
TNFα
MCP-1
IL-1α
IL-6
CXCL5
F
IL10
ZT8 CT8
0.0
0.3
0.6
0.9
1.2
*
IFNγ
MCP-1
IL-6
IL-10
IL-1
0
4
8
12
16
20
24
15
25
35
45
ZT/CT LPS injection (h)
IL-1 (ng/ml)
IL-6
0
4
8
12
16
20
24
60
80
100
120
ZT/CT LPS injection (h)
IL-6 (ng/ml)
IL-10
0
4
8
12
16
20
24
0
1
2
3
4
ZT/CT LPS injection (h)
IL-10 (ng/ml)
CXCL5
0
4
8
12
16
20
24
40
60
80
100
120
140
ZT/CT LPS injection (h)
CXCL5 (pg/ml)
IL-12
0
4
8
12
16
20
24
0.4
0.8
1.2
1.6
ZT/CT LPS injection (h)
plasma conc. (ng/ml)
IL-18
0
4
8
12
16
20
24
0.6
0.8
1.0
1.2
1.4
1.6
ZT/CT LPS injection (h)
IL-18 (ng/ml)
MCP-1
0
4
8
12
16
20
24
10
15
20
ZT/CT LPS injection (h)
MCP-1 (ng/ml)
TNF
0
4
8
12
16
20
24
1.5
2.0
2.5
3.0
3.5
ZT/CT LPS injection (h)
TNF (ng/ml)
Figure 1. Time-of-day dependent mortality in LPS treated mice is controlled by the circadian system and light conditions. A) LPS (30mg/kg,
i.p.) induced mortality in C57Bl/6 mice (n=10 per time point) kept either in LD 12:12 (yellow) or in DD (grey). Left graph: single time points;
right graph: mean mortality of LD or DD light condition. Error bars represent 95% confidence intervals (n=40 per group, *** p<0.001). B)
LPS dose-mortality curves of mice challenged at 4 time points in LD versus DD (overall n=40 mice per dose). Gray lines were calculated
by fitting an allosteric model to each group. C) Mice (n=10-14 per time point) were challenged with half-lethal doses of LPS (30 mg/kg, i.p.,
for mice kept in LD (left panel) or 13mg/kg, i.p., for mice kept in DD (right panel). Mortality was assessed 60 hours after LPS injection. To
perform statistical analyses, mortality rates were transformed to probability of death in order to compute sine fit using logistic regression and
F-test (LD, p=0.06; DD, p=0.001; gray shaded areas indicate 95% confidence intervals. D) and E) Time-of-day dependent cytokine profiles
in peripheral blood of C57Bl/6 mice (n=10 per time point) challenged with half-lethal doses of LPS (30mg/kg, i.p., for mice kept in LD (yellow)
or 13mg/kg, i.p., for mice kept in DD (gray)) and sacrificed 2 hours later. E) Relative amplitudes and phases of cytokines shown in D). Light-
colored circles represent non-significant circadian rhythms (p-value>0.05 ) as determined by non-linear least square fit and consecutive
F-test (see also Methods section). F) Cytokine levels in peripheral blood from mice (n=10 per time point) challenged with LPS (13mg/kg, i.p.)
at either ZT8 (LD) or CT8 (DD) conditions (T-test, *** p<0.001, ** p<0.01, *p<0.05).
although most circadian cytokines show variable phase relations between free-running and
entrained conditions (Fig. 1E).
Next, we investigated, whether the increased overall mortality in DD was correlated with
an increased pro-inflammatory cytokine response. We thus injected mice at either ZT8
(mice kept in LD) or CT8 (mice kept in DD) with the same dose of LPS (13mg/kg) and took
blood samples two hours later. IFNγ, MCP-1, IL-6 and the anti-inflammatory cytokine IL-
10 showed significantly altered levels between mice kept in LD or DD (Fig. 1F), suggest-
ing that in DD conditions the sensitivity to endotoxin is increased leading to enhanced pro-
inflammatory cytokine secretion and subsequently increased mortality.
Dispensable role of myeloid clocks in circadian endotoxin reactivity
Local clocks are thought to play important roles in mediating circadian modulation of cell-
and tissue-specific functions [47]. In fact, depletion of local immune clocks has been shown
to disrupt circadian patterns of tissue function [9,14,20,21,39]. To test whether local clocks
in cells of the innate immune system are responsible for circadian time dependency in the
response to bacterial endotoxin, we challenged myeloid lineage Bmal1-knockout mice (LysM-
Cre+/+ x Bmal1flox/flox, hereafter called myBmal-KO) with half-lethal doses of LPS at various
times across the circadian cycle. These mice lack physiological levels of Bmal1 mRNA and
protein in cells of myeloid origin and have been characterized elsewhere [21]. To our surprise,
mortality of these mice was still dependent on circadian time of LPS administration (Fig. 2A)
indicating that a functional circadian clock in myeloid cells is not required for time-of-day
dependent LPS-sensitivity. However, the overall susceptibility to LPS decreased two-fold
compared to wild-type mice (Fig. 2B) suggesting that BMAL1 levels in myeloid cells directly
3/15
WT
LysM-Cre+/+
myBmal-KO
0
20
40
60
80
100
mortality (%)
***
**
0
4
8
12
16
20
24
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0
4
8
12
16
20
24
0
20
40
60
80
100
CT LPS injection (h)
mortality (%)
A
B
C
0
20
40
60
80
100
7.5
15 20
30
5
50
10
25
40
LPS (mg/kg)
mortality (%)
Figure 2. myBmal-KO mice show time-of-day dependent and increased susceptibility to LPS. A)
Circadian mortality in myBmal-KO mice. Mice (n=10-14 per group) were challenged with half-lethal
doses of LPS (30mg/kg, i.p.) at indicated time points. Mortality was assessed 60 hours after LPS injec-
tion. Statistics were performed as in Fig. 1C, (p=0.0009; gray shaded area indicates 95% confidence
interval). B) LPS dose-mortality curves of mice challenged at 4 time points in constant dark conditions.
About 3-fold decrease of susceptibility to LPS in myBmal-KO mice (blue circles) as compared to wild-
type mice (gray circles – replotted from Fig. 1B) kept in DD. Gray lines were calculated by fitting an
allosteric model to each group. C) Reduced mean mortality in myBmal-KO mice (n=84) compared to
control strains LysM-Cre+/+ or C57Bl/6 (wild-type, n=40). Error bars represent 95% confidence intervals
(n=40 per group, *** p<0.001, ** p<0.01).
or indirectly modulate susceptibility towards LPS. The latter effect could only in part be at-
tributed to genetic background (note decreased susceptibility to LPS in LysM-Cre+/+ control
mice (Fig. 2C)) together arguing for a tonic rather than temporal role of myeloid BMAL1 in
regulating LPS sensitivity.
Despite its essential role for circadian clock function, BMAL1 has been linked with a num-
ber of other non-rhythmic processes such as adipogenesis [45], sleep regulation [16] and
cartilage homeostasis [15]. Thus, we asked whether the decreased susceptibility to LPS ob-
served in myBmal-KO mice was due to non-temporal functions of BMAL1 rather than to
the disruption of local myeloid clocks. CLOCK, like its heterodimeric binding partner BMAL1,
has also been shown to be an indispensable factor for peripheral clock function [11]. If non-
temporal outputs of myeloid clocks rather than gene specific functions of myeloid BMAL1
controls the susceptibility to LPS, a depletion of CLOCK should copy the myBmal-KO phe-
notype. We therefore generated conditional, myeloid lineage specific Clock-KO mice (here-
after called myClock-KO). As in myBmal1-KO mice, the expression of Cre recombinase is
driven by a myeloid specific promoter (LysM) consequently leading to excision of LoxP
flanked exon 5 and 6 of the clock gene [10]. As expected, the expression of Clock mRNA
and protein was substantially reduced in peritoneal cavity cells but was normal in liver (Fig.
3A-B).
To investigate, whether clock gene rhythms were truly abolished in myClock-KO mice, we
harvested peritoneal macrophages from myClock-KO or control mice (LysM-Cre) in regular
4-hour intervals over the course of 24 hours. Rhythmicity of Bmal1, Cry1, Cry2, Dbp, Npas2
and Nr1d1 mRNA levels was essentially eliminated, while a low amplitude rhythmicity was
detected for Per1 and Per2 mRNA (Fig. 3C and D). Moreover, circadian oscillations were dis-
rupted in peritoneal cavity cells (mainly macrophages and B-cells) from myClock-KO mice,
but not in tissue explants from SCN or lung (Fig. 3E).
Given the disruption of rhythmicity in myClock-KO myeloid cells, we asked whether
4/15
rel. amplitude
100
60
20
2
1
10
40
4
6
WT
LysM-Cre+/+
myClock-KO
0
20
40
60
80
100
mortality (%)
***
ns
0
4
8
12
16
20
24
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0
4
8
12
16
20
24
0
20
40
60
80
100
CT LPS injection (h)
mortality (%)
A
B
C
D
E
F
G
0
2
4
6
8
10
12
14
16
18
20
22
peak phase CT (h)
peritoneal
cavity cells
liver
myClock-KO
Clockfl/fl
myClock-KO
Clockfl/fl
B-ACTIN
CLOCK
B-ACTIN
CLOCK
Clock
Nr1d1
Cry1
Bmal1
Per1
Npas2
LysM-Cre+/+
myClock-KO
0
40
80
120
rel. Clock mRNA expression
Cry1
0
2
4
6
8
CT (h)
relative expression
0
4
8
12 16 20
0
4
Npas2
0
2
4
6
8
CT (h)
relative expression
0
4
8
12 16 20
0
4
Nr1d1
0
25
50
75
CT (h)
relative expression
0
4
8
12 16 20
0
4
Per1
0
2
4
6
CT (h)
relative expression
0
4
8
12 16 20
0
4
Bmal1
0
1
2
3
4
5
CT (h)
relative expression
0
4
8
12 16 20
0
4
Clock
0.0
0.5
1.0
1.5
2.0
CT (h)
relative expression
0
4
8
12 16 20
0
4
SCN
24
48
72
96
120 144 168
0.5
1.0
1.5
···········································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································
···································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································
time (h)
relative bioluminescence
lung
24
48
72
96
120 144 168
0.5
1.0
1.5
··································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································
·········································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································································
time (h)
relative bioluminescence
24
48
72
96 120
0.5
1.0
1.5
168 192 216 240 264 288
··················································································································
·····························································································································································································································································································································································································································································································
·································
··············································
····························································································································································································································································
·····
····················
······································································
·····················
······
··················
···········
····
·······
·························
·································································
···································
···
··························································
········
·································
·····································
······
·
·································································
···
······································
··················
············
·················
····
··················································································································
·····························································································································································································································································································································································································································································································
·································
··············································
····························································································································································································································································
·····
····················
······································································
·····················
······
··················
···········
····
·······
·························
·································································
···································
···
··························································
········
·································
·····································
······
·
·································································
···
······································
··················
············
·················
····
················································································································································································································································································································································································································································································································································································································································································································································································································································
·····················································································································································································································································
·
·
································································
··························································································································································································································································································································
······································································
····························
···········································
·············································································
················································
····················································
······················································································
·························································································································································································
·········
·····
·
···
··
························································································································································································································································································································
················
·······················
············
················································································································································································································································································································································································································································································································································································································································································································································································································································
·····················································································································································································································································
·
·
································································
··························································································································································································································································································································
······································································
····························
···········································
·············································································
················································
····················································
······················································································
·························································································································································································
·········
·····
·
···
··
························································································································································································································································································································
················
·······················
············
time (h)
relative bioluminescence
peritoneal
macrophages
Figure 3. Conditional myClock-KO mice show circadian pattern in mortality by endotoxic shock. A) and B) Reduced levels of Clock mRNA
and protein in myeloid lineage cells of myClock-KO mice. A) Mean values of normalized mRNA expression values of time-series shown
in C) (n=21 mice per condition, p<0.0001, t-test). B) Protein levels by immunoblot in peritoneal cavity cells and liver of myClock-KO and
Clockfl/fl control mice (n=3). C) Relative mRNA levels of selected clock genes in peritoneal macrophages from LysM-cre+/+ (brown circles) or
myClock-KO (red circles) mice at indicated circadian times. Phase and amplitude information are depicted in D) as analyzed by Chronolyse.
Non-significant circadian expression (p>0.05) are depicted in light red (myClock-KO) or light brown (LysM-cre control). E) Representative
bioluminescence recordings of peritoneal macrophages, SCN or lung tissue from myClock-KO or wild-type mice crossed with PER2:Luc
reporter mice (color coding as before). Black arrow indicates time of re-synchronization by dexamethason treatment (detrended data). F)
Circadian pattern in endotoxic shock mortality despite deficiency of CLOCK in myeloid lineage cells. Mice (n=10-14 per time point) were
challenged with half-lethal doses of LPS (30mg/kg, i.p.) at indicated time points. Mortality was assessed 60 hours after LPS injection. Statistic
were performed as in Fig 1C (p=0.005, gray shaded area indicates 95% confidence interval). G) Reduced mean mortality (at 30mg/kg LPS) in
mice deficient of myeloid CLOCK (n=84) compared to control strains LysM-Cre+/+ (n=40) or C57Bl/6 (wild-type, n=40). First two bars where
re-plotted from Fig. 2C. Error bars represent 95% confidence intervals (n=40 per group, ns p>0.05, p*** p<0.001.
CLOCK in myeloid cells is required for time-of-day dependent mortality in endotoxic shock.
To this end, we challenged myClock-KO mice at various times during the circadian cycle with
half-lethal doses of LPS. Again, mortality in these mice was significantly time-of-day depen-
dent (Fig. 3F). As in myBmal1-KO mice, myClock-KO mice showed strongly reduced suscep-
tibility to LPS compared to wild-type mice (Fig. 3G). Taken together, our data unequivocally
show that myeloid clockworks are dispensable for the time-of-day dependency in endotoxic
shock. In addition, decreased overall susceptibility suggest a non-temporal, sensitizing role
of myeloid CLOCK/BMAL1 in the regulation of endotoxic shock.
Circadian cytokine response in myClock-KO mice
Our initial hypothesis was built on the assumption that a time-of-day dependent cytokine
response determines the outcome in endotoxic shock. Previous results from us and oth-
ers [21,25,28] suggested that local myeloid clocks govern the timing of the pro-inflammatory
cytokine response. However, circadian mortality profiles in LPS-challenged myeloid clock-
knockout mice (Fig. 2A and 3F) led us to question this model: The circadian cytokine re-
sponse in plasma is either independent of a myeloid clock or the circadian mortality by en-
dotoxic shock does not require a circadian cytokine response.
To test these mutually not exclusive possibilities, we challenged myClock-KO mice with
half-lethal doses of LPS in regular 4-hour intervals over the course of one day. Unexpect-
edly, cytokine levels in plasma, collected two hours after LPS administration still exhibited
circadian patterns for TNFα, IL-18, IL-10 (Fig. 4A) (p-values=0.046, 0.001 and 0.009, re-
spectively), very similar to those observed in wild-type animals (Fig. 4B). Other cytokines
remained below statistical significance threshold for circadian rhythmicity tests (IL-1α) or
5/15
B
0
2
4
6
8
10
12
14
16
18
20
22
CT peak level (h)
IL-10
IL-18
IL-12
IL-6
TNFα
CXCL5
MCP-1
IL-1α
rel. amplitude
2.2
1.9
1.6
1.3
1
C
A
TNF
0
4
8
12
16
20
24
1.0
1.5
2.0
2.5
CT LPS injection (h)
TNF (ng/ml)
CXCL5
0
4
8
12
16
20
24
80
100
120
140
CT LPS injection (h)
CXCL5 (pg/ml)
IL-1
0
4
8
12
16
20
24
20
25
30
35
40
CT LPS injection (h)
IL-1 (ng/ml)
IL-6
0
4
8
12
16
20
24
100
200
300
400
500
CT LPS injection (h)
IL-6 (ng/ml)
IL-10
0
4
8
12
16
20
24
0.5
1.0
1.5
2.0
2.5
CT LPS injection (h)
IL-10 (ng/ml)
IL-12
0
4
8
12
16
20
24
0.3
0.4
0.5
0.6
0.7
0.8
CT LPS injection (h)
IL-12 (ng/ml)
IL-18
0
4
8
12
16
20
24
0.6
0.9
1.2
1.5
1.8
CT LPS injection (h)
IL-18 (ng/ml)
MCP-1
0
4
8
12
16
20
24
5
10
15
20
CT LPS injection (h)
MCP-1 (ng/ml)
IL-6
0
100 200 300 400 500
0
20
40
60
80
100
r2 = 0.01
plasma concentration (ng/ml)
mortality (%)
IL-10
500
1200 1900 2600 3300
0
20
40
60
80
100
r2 = 0.39
plasma concentration (pg/ml)
mortality (%)
IL-18
700
950
1200 1450 1700
20
40
60
80
100
r2 = 0.56
plasma concentration (pg/ml)
mortality (%)
MCP-1
8000
12000
16000
20000
0
20
40
60
80
100
r2 = 0.38
plasma concentration (pg/ml)
mortality (%)
TNF
1000
2000
3000
0
20
40
60
80
100
r2 = 0.26
plasma concentration (pg/ml)
mortality (%)
CCL7
700
800
900
1000 1100
20
40
60
80
100
r2 = 0.57
plasma concentration (pg/ml)
mortality (%)
Figure 4. Circadian time dependent cytokine levels in plasma of myClock-KO mice. A) Plasma cytokine levels in myClock-KO mice kept in
constant darkness, two hours after injection of 30mg/kg LPS. Data represent mean values ± SEM (n=14 per time point). B) Polar plot showing
amplitude and phase distribution of pro- and anti-inflammatory cytokines from A), red circles and wild-type DD (Fig. 1E). Light circles indicate
non-significant (p-values>0.05, non-linear least square fit statistics by ChronoLyse) circadian abundance. C) Overall correlation of cytokine
levels with mortality independent of time-of-day of LPS injection and mouse model. Colors indicate data source (wild-type, LD - yellow;
wild-type, DD - gray; myClock-KO, DD - red; linear regression - gray line; statistics: spearman correlation).
displayed large trends within this period of time (IL-6). These data suggest that a LPS-
induced circadian cytokine response does not depend on a functional circadian clock in cells
of myeloid origin. Thus, circadian cytokine expression might still be responsible for time-of-
day dependent mortality in endotoxic shock.
To identify those cytokines, whose levels best explain mortality upon LPS challenge, we
correlated cytokine levels and mortality rate for all conditions - independent of time-of-day
or mouse strain - in a linear correlation analysis. Levels of CCL7, MCP-1 and TNFα showed
strong positive correlation with mortality (p-values=0.0003, 0.0065, 0.0319, respectively).
Others, such as IL-10 and IL-18 correlated negatively (p-values=0.0054 and 0.0004, re-
spectively) (Fig. 4C). Interestingly, while TNFα and IL-10 are well known factors in the path-
omechanism of endotoxic shock, CCL7, MCP-1 and protective effects of IL-18 have not been
reported in this context.
Together, our data suggest that local circadian clocks in myeloid lineage cells are dis-
pensable for time-of-day dependent plasma cytokine levels upon LPS challenge. However,
where does time-of-day dependency in endotoxic shock originate instead?
Persistent circadian traffic in myeloid clock-knockout mice
Circadian patterns in immune cell trafficking and distribution have been recently reported
and linked to disease models and immune functions [14, 28, 39]. Similarly, homing and re-
lease/egress of hematopoetic stem cells (HSPCs), granulocytes, and lymphocytes to bone
marrow and lymph nodes, respectively, have been shown to vary in a time-of-day depen-
dent manner requiring the integrity of an immune-cell intrinsic circadian clock [7,14,38,44].
Thus, we asked, whether circadian traffic of myeloid cells can be associated with mortality
rhythms in our endotoxic shock model.
To test this, we measured the number of immune cells in various immunological com-
partments at two distinct circadian time points representing peak (CT8) and trough (CT20)
of circadian mortality rate upon LPS challenge. Cells from wild-type, genotype control and
myeloid clock-knockout mice (n=5 per time point) were collected from a broad spectrum
6/15
A
B
D
2 weeks entrainment
12:12 hours LD
0
4
8
12
16
20
24
4
8
12
16
20
24
0
4
8
12
16
20
24
day 1 in DD
tissue sampling at day 2
in DD
C
E
WT
LysM-cre
myBmal-KO
myClock-KO
FACS
Tissue sampling
Mouse strains
Blood
Spleen
Bone-marrow
Lymph-nodes
Macrophages
Neutrophils
bone marrow
spleen
lymph nodes
0
2×107
4×107
6×107
***
***
***
total # cells
in compartment
WT
LysM-cre
myBmal1-KO
myClock-KO
0%
5%
10%
15%
***
*
ns
ns
frequency of F4/80+
cells in blood
WT
LysM-cre
myBmal1-KO
myClock-KO
0
1×106
2×106
3×106
4×106
5×106
***
***
***
**
# of Ly6G+ cells
in bone marrow
WT
LysM-cre
myBmal1-KO
myClock-KO
0
1×106
2×106
3×106
4×106
# of F4/80+ cells
in spleen
***
***
**
**
Figure 5. Circadian traffic of myeloid cells despite depletion of myeloid CLOCK or BMAL1. A) Exper-
imental scheme to investigate time-of-day dependent immune cell traffic in various compartments
and genetic mouse models. B) Total cell counts of femoral bone marrow (blue)), spleen (brown) or in-
guinal lymph nodes (green) at CT8 (light colors) or CT20 (dark colors). C-E) Cell number or frequency
from wild-type and various conditional circadian clock mice at two different circadian time points (wild-
type - gray, LysM-cre - brown, myBmal-KO - blue, myClock-KO - red, CT8 - light, CT20 - dark). C) Total
number of F4/80+ macrophages in spleen. D) Relative number of F4/80+ macrophages in blood. As-
terisks indicate level of significance as determined by t-test . E) Total number of Ly6G+ neutrophils in
bone marrow (significance levels: ns p>0.05, * p<0.05, ** p<0.01, *** p<0.001).
of immune system compartments (blood, peritoneal cavity, bone marrow, spleen, inguinal
lymph nodes, thymus)(Fig. 5A). As expected from previous studies [14,28,44] the number of
cells were significantly time-of-day dependent in bone marrow, spleen and inguinal lymph
nodes of wild-type mice (Fig. 5B).
If circadian traffic of myeloid cells and therefore lymphoid organ composition would
be a main factor in regulating sensitivity to bacterial endotoxin, similar patterns in myeloid
clock-knockout mice should be observed. However, results obtained from myBmal-KO and
myClock-KO mice did not support this hypothesis. While in spleen, F4/80+ macrophages
of both myeloid clock knockout strains were still found at higher numbers at CT8 compared
to CT20 (Fig. 5C), significant time-of-day differences of macrophage numbers in wild-type
blood diminished in myeloid clock knockouts (Fig. 5D). Different patterns of time-of-day
dependency were also observed in bone marrow (Fig. 5E), together suggesting a rather com-
plex than mono-causal relation between myeloid clocks, circadian traffic and mortality risk
in endotoxic shock.
7/15
Discussion
In this study we tested the hypothesis that local myeloid circadian clocks regulate the devas-
tating immune response in endotoxic shock. A number of findings by various labs including
our own pointed to such a possibility. First, monocytes/macrophages have been identified
as important cellular entities relaying the endotoxin (LPS) triggered signal to the immune
system and other organ systems by means of massive secretion of pro-inflammatory cy-
tokines [40, 41]. Second, a high amplitude circadian clock in monocytes/macrophages has
been shown to control circadian cytokine output invitro and exvivo [4,9,28]. Third, a high am-
plitude time-of-day dependent mortality has been demonstrated in endotoxic shock [23].
Fourth, a number of experiments done at two circadian time points, including a cecal liga-
tion and puncture model [12], demonstrating loss of time-of-day dependency of mortality
in myeloid Bmal1 knockout mice [9] further supported this hypothesis.
On the other hand, some studies suggested other mechanisms, e.g. Marpegan and col-
leagues reported that mice challenged with LPS at two different time points in constant
darkness did not exhibit differences in mortality rates [36], arguing for a light-driven pro-
cess in regulating time-of-day dependency in endotoxic shock. However, we suspected that
this result may be caused by a ceiling effect induced by constant darkness, since mortality
rates at both time points were close to 100 percent. Indeed, when we compared suscepti-
bility of mice challenged with a similar dose of LPS in light-dark versus constant darkness,
we observed a marked increase of overall mortality. Similarly, depletion/mutation of Per2
rendered mice insensitive to experimental time dependency in endotoxic shock [32]. Sur-
prisingly, however, we found that depletion of either CLOCK or BMAL1 in myeloid lineage
derived cells both did not abolish time-of-day dependency in mortality to endotoxic shock,
which led us to reject our initial hypothesis.
Our data also exclude light as a stimulusdriving mortality: First, data fromHalberg, Marpe-
gan, Scheiermann as well as our own lab [23,28,36,44] suggest that mice are more suscep-
tible to endotoxic shock during the light phase compared to dark phase - whereas switching
light schedules from light-dark to constant darkness led to an increase in overall mortal-
ity. Second, irrespective of the genetic clock-gene depletion tested, our data unequivocally
demonstrate circadian mortality rhythms upon LPS challenge even under DD conditions.
Strikingly, peripheral blood cytokine levels showed - though altered - circadian time depen-
dency in the myClock-KO strain.
A burstof cytokines, followinga lethal dose ofLPS is generallythought to bean indispens-
able factor causing multi-organ dysfunction and leading to death. However, the contribution
of single cytokines has been difficult to tease apart due to complex nature of interconnected
feedback systems. By adding time-of-day as an independent variable in a number of differ-
ent mouse models we were able to correlate individual cytokines’ contribution to mortality
in the context of the complex response to LPS. While the pro-inflammatory cytokine TNFα
was confirmed to act detrimentally, IL-18 surprisingly turned out to likely be protective. This
interpretation is not only supported by the negative correlation of IL-18 levels in blood and
mortality risk, but also by an anti-phasic oscillation of IL-18 levels paralleling those of the
known protective cytokine IL-10.
While circadian regulation of trafficking lymphocytes by cell intrinsic clockworks have
been demonstrated to impact the pathophysiology of an autoimmune disease model such
as EAE [14], our data on distribution patterns of immune cells at two circadian time points
draw a more complex picture. In spleen, absolute numbers of macrophages, but not neu-
trophils and monocytes, are independent of their local clocks. In contrast, patterns of
macrophages and neutrophils in peripheral blood and bone marrow (respectively) change
upon local myeloid clock depletion. Thus it appears that myeloid cell traffic is regulated at
multiple levels including cell-intrinsic, endothelial and site specific factors. However, our
data do not support the hypothesis of myeloid cell distribution being a main factor for time-
of-day dependent mortality risk in endotoxic shock. Hence, what remains as the source of
these rhythms?
Following the patho-physiology of endotoxic shock on its path from cause to effect, it is
important to note that the bio-availability of bacterial endotoxin injected intraperitoneally
depends on multiple factors (i.e. pharmaco-kinetics), many of which themselves might un-
derlie circadian regulation. As a consequence, same doses of LPS administered i.p. at differ-
ent times-of-day might result in highly diverging concentrations at the site of action.
8/15
Along this path, rhythmic feeding behavior driven by the central pacemaker was sug-
gested to alter immune-responses directly or indirectly [29, 33]. Also, nutrition related fac-
tors could drive immune cells to respond differently to stimuli independent of local clocks
[27,39].
Finally, the ability of cells and organs to resist all sorts of noxa might as well be regulated
in a time-of-day dependent manner. In this case, rather than same doses of toxin leading to
time dependent cytokine responses (noxa) resulting in diverging rates of multi organ failure,
same amount of noxa would cause time dependent rates of multi organ dysfunction and
death. Indeed, work from Hrushesky [26] showed that mice challenged with same doses
of TNFα at various times throughout the day exhibited time-dependent survival paralleling
phenomena in endotoxic shock. However, our data showing fluctuating cytokine levels in
myClock-KO mice challenged with LPS in constant darkness (Fig. 4A) question the mecha-
nism of organ vulnerability as the only source of circadian regulation.
While our work suggests that local myeloid clocks do not account for time-of-day depen-
dent mortality in endotoxic shock it unequivocally argues for a strong enhancing effect of
myeloid CLOCK and BMAL1 on overall susceptibility, which adds on the effect of light con-
ditions. However, it is important to note that our data do stay in conflict with findings from
other labs which rather reported attenuating effects of BMAL1 on inflammation [9, 12, 39]
but align well with a report on clock mutant mice [4]. Differences in genetic background of
mouse strains, animal facility dependent microbiomes [42] or animal care procedures might
account for this but remain unsatisfying explanations.
One of the most striking findings of our work is the large increase in susceptibility to endo-
toxic shock when mice were housed under DD as compared to LD conditions. This increase
was observed in wild-type as well as in myBmal-KO mice, which implies that myeloid BMAL1
is not required for this effect. Interestingly, Carlson and Chiu reported similar effects in a ce-
cal ligation and puncture model in rats upon transfer to LL (constant light) or DD conditions,
where they found decreased survival as compared to rats remaining in LD conditions [6]. It is
tempting to speculate that rhythmic light conditions, rather than light itself promote survival
in endotoxemia. In either case, it will be important to further investigate these phenomena
not only in respect to animal housing conditions, which need to be tightly light controlled in
immunological, physiological and behavioral experiments but also for their apparent impli-
cations on health-care in intensive care units.
Materials and Methods
Animals
All procedures were authorized by and performed in accordance with the guidelines and regulations of
the German animal protection law (Deutsches Tierschutzgesetz). Mice were housed in macrolon type
II cages supplied with nesting material, food and water ad libitum at a 12h:12h light/dark (LD) cycle. For
endotoxic shock and running wheel experiments mice were individually housed. For all other exper-
iment mice were group-housed. Manipulations during the dark phase of the cycle were performed
under infrared light. Male C57Bl/6 mice (Jackson Laboratories strain) mice were purchased from our
animal facility (Charit´e FEM, Berlin, Germany) at 8-10 weeks of age. Homozygous, male LysMCre/Cre
(LysM-Cre) [8], myBmal-KO and myClock-KO were bred and raised in our animal facility (FEM, Berlin
Germany) and used at 8-12 weeks. Female LysMCre/Cre Per2:Luc, either wild-type or homozygous for
Clock-flox, were bred and raised in our animal facility (FEM, Berlin Germany) and used at 14 weeks.
Generation of myeloid clock knockout mice
Bmal1flox/flox (Bmal-flox) [46] or Clockflox/flox (Clock-flox) [10] were bred with LysM-Cre to target Bmal1
or Clock for deletion in the myeloid lineage. Offspring were genotyped to confirm the presence of the
loxP sites within Bmal1 or Clock and to determine presence of the Cre recombinase. Upon successful
recombination the loxP flanked exon 8 of Bmal1 or in case of Clock the floxed exon 5 and 6 were deleted.
LysMCre/Cre x Clockflox/flox were crossed onto a Per2:Luc [50] background for further characterization in
bioluminescence reporter assays. All mice have been genotyped before experiments.
Endotoxic shock experiments
8-12 weeks old mice were entrained to 12h:12h light-dark cycles for 2 weeks. Dosing of LPS injection
was adjusted for individual body weight prior injection. Injection volume did not exceed 10µl/g body
weight. For LD experiments, mice were injected i.p. on day 14. For DD experiments mice were trans-
9/15
ferred to DD on day 14 and were injected i.p. on second day in DD. Animals were kept in the respective
lighting conditions until the termination of experiment 60h post LPS injection. For all endotoxic shock
experiments, human endpoints were applied to determine survival (for definition of human endpoints
see respective section).
Intraperitoneal LPS injection
E. coli LPS (055.B5, Sigma Aldrich) stock solution (10mg/ml) was diluted to appropriate concentration
in sterile PBS and thoroughly vortexed before use. LPS injection was performed i.p. using Legato 100
Syringe Pump (KD Scientific). The following settings were applied: Mode: infuse only; syringe: BD,
plastic, 5ml; rate: 2ml/min. A cannula (26G x 3/8”) was attached to Legato 100 by microbore extension
line (60cm, MedEx, Smiths Medical) for the LPS injections.
Time-of-day dependent mortality experiments
Endotoxic shock mortality experiments were comprised of two parts for each mouse strain and con-
dition tested: First, lethal dose 50 (LD50) of LPS was determined by injecting groups of mice (n=10)
at four 6-hour spaced time points. The LD50 describes the concentration at which approx. 50% of all
animals injected (averaged over all injection time points) survive. This ensures most dynamic range
for the detection of potential circadian rhythms. Second, experimentally determined LD50 of LPS was
used to investigate, whether mortality by endotoxic shock was dependent on time-of-day. To this end,
mice (n=14 per group) were injected i.p. at six 4-hour spaced time points to increase statistical power
for circadian rhythm analysis (see statistical data analysis section).
Definition of human endpoints
A scoring system was developed in order to detect irreversibly moribund mice before the occurrence of
death by endotoxic shock. It is based on previous reports by [32,45] and required further refinements
according to our experience. Mice in the endotoxic shock experiments were monitored and scored
every 2-4 hours for up to 60 hours post LPS injection. In addition, surface body temperature was
measured at the sternum every 12h and body weight was measured every 24 h. Mice with a score of
0-2 were monitored every 4h. As of a score of 3, the monitoring frequency was increased to every
2h. In addition, softened, moisturized food was provided in each cage as of a score of 3. Weight loss
exceeding 20%, 3 consecutive scores of 4, or one score of 5 served as human endpoints. A mouse was
defined as a non-survivor when human endpoints applied and was subsquently sacrificed by cervical
dislocation. Mice which did not display any signs of endotoxic shock such as weight or temperature
loss and no increasing severeness in score were excluded from the analysis.
Peripheral blood cytokine concentrations in endotoxic shock
To determine the blood cytokine levels in the endotoxic shock model, mice were injected with cor-
responding LD50 of LPS at six 4-hour spaced time points (n=14). 2h post LPS injection mice were
terminally bled by cardiac puncture using a 23G x 1” cannula (Henke-Sass Wolf). Syringes (1ml, Braun)
were coated with heparin (Ratiopharm) to avoid blood coagulation. After isolation blood was kept at
4◦C for subsequent plasma preparation. To this end, blood was centrifugated for 15min, 370g at 4◦C.
Plasma was isolated, aliquoted and frozen at -80◦C for further analysis.
Isolation of peritoneal macrophages (PM)
Mice were sacrificed by cervical dislocation. Peritoneal cavity cells (PEC) were isolated by peritoneal
lavage with ice cold PBS. Lavage fluid visibly containing red blood cells was dismissed. For RNA mea-
surements or bioluminescence recordings peritoneal macrophages were further purified by MACS
sorting (Miltenyi) according to manufacturer’s protocol using mouse/human CD11b Microbeads and
LS columns. Eluates containg positively sorted macrophages were analyzed for sorting efficiency by
FACS. All steps were performed at 4◦C.
Isolation of bone marrow cells
Mice were sacrificed by cervical dislocation. One tibia and femur were excised per mouse. Femur and
tibia were flushed with supplemented RPMI 1640 and erythrocytes were lysed using GEYS solution
(2min at 4◦C). After erythrocyte lysis, cells were suspended in supplemented RPMI 1640 medium and
filtered through a 30µM filter (Miltenyi). Cell numbers were determined using a Neubauer chamber.
All steps were performed at 4◦C. All centrifugation steps were performed at 4◦C, 300g, 7min.
Isolation of spleen cells
Mice were sacrificed by cervical dislocation. Spleen was removed and a single cell single cell suspension
was obtained using gentleMACS, Miltenyi (program: m spleen 01) with C-tubes (Miltenyi) in PBS. Next
single cell solution was filtered using 100µM cell strainers (Thermo Fischer). GEYS solution was used
10/15
for erythrocyte lysis for 2min at 4◦C. After erythrocyte lysis cells were suspended in supplemented
RPMI 1640 and filtered through a 30µM filter (Miltenyi). Cell number was determined using a Neubauer
chamber. All steps were performed at 4◦C. All centrifugation steps were performed at 4◦C, 300g,
7min.
Bioluminescence recordings
PER::LUC protein bioluminescence recordings were used to characterize circadian clock function in
peritoneal macrophages, SCN and lung tissue. Mice were sacrificed by cervical dislocation. Brains and
lungs were isolated and transferred to chilled Hank’s buffered saline solution, pH 7.2. (HBSS). For tissue
culture, 300µm coronal sections of the brain and 500µm sections of the lung were obtained using a
tissue chopper. The lung and SCN slices were cultured individually on a Millicell membrane (Millipore)
in a Petri dish in supplemented DMEM containing 1µM luciferin (Promega). PECs were isolated by peri-
toneal lavage and immediately CD11b-MACS sorted. The CD11b-sorted peritoneal macrophages were
cultured in Petri dishes in supplemented DMEM containing 1µM luciferin (Promega). For biolumines-
cence recording, tissues/primary cell cultures were placed in light-tight boxes (Technische Werkstaet-
ten Charite, Berlin, Germany) equipped with photo-multiplier tubes (Hamamatsu, Japan) at standard
cell culture conditions. Bioluminescence was recorded in 5min bins. On fifth day in culture PMs were
treated with 1µM dexamethason for 1h, followed by a medium change to supplemented DMEM con-
taining 1µM luciferin (Promega). Data were further processed and analyzed using Chronostar 3.0 [34].
Generation of whole cell protein lysates
Liver and peritoneal cavity cells were homogenized in ice cold RIPA buffer containing 1x protease-
inhibitor-cocktail (Sigma Aldrich) and incubated on ice for 30min. Homogenized cells were then cen-
trifuged at maximum speed for 30min at 4◦C to pellet the insoluble cell debris. The supernatant frac-
tion was then removed and used for cellular protein analysis or frozen at -80◦C. Protein concentrations
were determined using standard BCA assay.
Immunoblotting
Samples were denaturated for SDS-PAGE in NuPAGE SDS Sample Buffer (4x) (Invitrogen) containing
0.8% 2-β-mercaptoethanol (Sigma) and boiled for 5-10min at 95◦C. SDS-PAGE using 4-12% Bis-
Tris gels (Thermo Scientific) at 200V for 60min in NuPAGE MES SDS Running Buffer. Proteins were
transferred to a nitrocellulose membrane (0.45µm) using a tank transfer system (wet transfer). Nu-
PAGE transfer buffer, containing 20% Methanol, was cooled with an ice block to prevent overheating
during the transfer. The transfer was run for 120min at 90V. Following the transfer, the membrane
was blocked in TBS-T with 5% non-fat, dry milk for 1-2h at RT. After a washing step in TBS-T (3 x
10min), the membrane was placed in the primary antibody solution (TBS-T with 5% non-fat, dry milk)
and gently shaken overnight at 4◦C. The membrane was then washed in TBS-T (3 x 10min) and incu-
bated with the HRP-conjugated secondary antibody (Santa Cruz Biotechnologies) in TBST-T for 2h at
RT. After another washing step in TBS-T (3 x 10min), a chemiluminescence reaction was performed
with Super SignalWest Pico substrate (Pierce). The protein bands were visualized using the Chemo-
Cam detection system (Intas). The following primary antibodies were used: murine CLOCK - rabbit
anti-mCLock (Bethyl Laboratories, A302-618A) , murine BMAL1 - rabbit anti-mBMAL1 (kind gift from
Micheal Brunner) , murine ACTINB - mouse anti mBactin (Sigma, A5441). Secondary antibodies used:
goat anti-mIgG-HRP (SantaCrz Biotechnology, sc-2005), donkey anti-rbIgG-HRP (SantaCrz Biotech-
nology, sc-2005).
Single- and multiplex immunoassays
Singleplex assay: murine IL-6 plasma concentration was determined by ELISA according to manu-
facturer’s protocol (Ebioscience) in a 96-well format (Corning). Plasma samples were diluted 1:200
in supplied assay buffer. Absorption was measured at 470nm. Reference wavelength was mea-
sured at 560nm by Infinite F200Pro plate reader (Tecan). Multiplex assay: 13 cytokines (CCL2/MCP-
1, CCL3/MIP-1α, CCL4/MIP-1β, CCL7/MCP-3, CXCL5, IL-1α, IL-1β, IL-10, IL-12p40, IL-18, Eotaxin,
Rantes/CCL5, TNFα) were assessed using the ProCartaPlex, Mix and Match, Mouse 13-Plex (Affymetrix,
eBioscience). ProCartaPlex was performed as described in manufacturer’s protocol in a 96 well format
(eBioscience). All washing steps were performed using a hand held magnetic washer (eBioscience).
Data were acquired using a MagPix (Luminex) detection device. Data evaluation was performed using
ProcartaPlex Analyst v.1.0 (eBiosciences).
Isolation and quantification of RNA
Total RNA was isolated using the PureLink RNA Mini Kit (Ambion) according to the manufacturer’s man-
ual. In addition, an on-column DNA digestion was performed using PureLink DNase Set (Life Technolo-
gies). RNA was quantified by measuring the absorption at 260nm with NanoDrop 2000C (Thermo
Scientific).
11/15
Quantitative real-time PCR
Total RNA was reverse-transcribed to cDNA using random hexamers to prime reverse transcriptase re-
action. cDNA was diluted 1:10 in H2O for use in qRT-PCR. qRT-PCR was performed using a 2-step pro-
tocol with the following primer-sets: primer-sets for mCry1, mCry2, mDpb, mNr1d1, mNpas2, mPer1,
mPer2 were purchased from Qiagen (QT00117012, QT00168868, QT00103089, QT00164556,
QT00108647, QT00113337, QT00198366, respectively).
mGapdh, fwd: ACGGGAAGCTCACTG-
GCATGGCCTT, rev: CATGAGGTCCACCACCCTGTTGCTG; mBmal1 primers were designed to character-
ize myBmal-KO mice. Forward primer (fwd: GGACACAGACAAAGATGACCC) binds upstream of exon 8
and the reverse primer (rev: TTTTGTCCCGACGCCTCTTT) within exon 8 of Bmal1. Thus after successful
Cre recombination, exon 8 is deleted and no PCR product is detectable. Clock primers were designed
to characterize myClock-KO mice. Primers bind in exon 5 (fwd: ATTGGTGGAAGAAGATGACAAGGA)
and in exon 6 (rev: TACCAGGAAGCATAGACCCC) of clock. As exon 5 and 6 are flanked by loxP sites,
after successful Cre recombination no PCR product is amplified.
Flow cytometry (FACS)
Two panels of antibodies were established to target a broad range of immune cells in various sites of
the organism. Before each experiment, antibody mix of both panels were prepared and kept on 4◦C for
labeling of all samples of the respective experiment in order to minimize intra-experimental variability.
All antibody-mixes were prepared in FACS buffer containing 1:50 FcR blocking reagent. FACS staining
of samples: 100µL of the cell suspensions were transferred into a 96-well plate and spun down for
7min, 300g at 4◦C. Supernatant was carefully discarded and pellet re-suspended in 50µL of master-
mix containing one of two antibody panels. Cells were incubated for at least 30mins at 4◦C in darkness.
Subsequently 200µL FACS buffer was added and cells were centrifuged for 7min, 300g at 4◦C. Cells
were washed twice using 200µL FACS buffer each before fixation in 200µL 4% PFA for 30min at RT.
Finally, cells were spun down and re-suspended in 200µL FACS buffer and stored at 4◦C for up to a
week before FACS data acquisition in a FACS CantoII (BD Biosciences).
Statistical data analysis
Statistical analysis was performed in GraphPad Prism 8 and R. Normality was tested using Shapiro-Wilk
normality test. When data were normally distributed and two parameters were compared, One-way
ANOVA with Dunnett’s multiple comparison as a post hoc test was applied. When comparing more
than two groups and two parameters a two-way ANOVA with Bonferonni’s post hoc test was applied.
Two sample comparison was performed with 2-sided students t-test for normally distributed data
and Mann-Whitney U test for non-normally distributed data. Time-of-day dependent mortality ex-
periments: Mortality data were transformed to probability of death (between 0-1) in order to compute
sine fit using alogistic regression. The confidence interval was derived from sine fit estimation. After sta-
tistical analysis, the mortality data was transformed back to the initial percentages. Cross-correlation
analysis of morality and cytokine data: To correlate mortality rates and cytokine levels across all animal
models, a permutation of sine fit of the mortality data and plasma cytokine data was used. This boot-
strapping (randomization) procedure gives rise to the empirical distribution of correlations. The p-value
is the fraction of randomizations that gave a correlation with the opposite denominator, meaning that
a p=0.05 means that only 5% of correlations crossed zero correlation threshold. For linear regression
analysis of the summary of cytokine and mortality data, the estimated mortality rate determined by
sine fit of mortality data was correlated to mean value of plasma cytokine concentration using Spear-
man’s rank correlation. Bioluminescence recordings were analysed using in-house written software
ChronoStar 3.0 [34]. In brief, raw bioluminescence counts were transformed to log-space and trends
removed by subtracting the 24h running average. Circadian rhythm parameters were estimated by
fitting a damped sine wave to these data. Finally, data were reversely transformed into linear space.
Circadian rhythmicity of cytokine time-series was tested using our in-house written software ChronoL-
yse. In brief, a 24h sine wave was fitted the beforehand log-transformed data and parameters of fit
were used to estimate amplitude, phase and mean levels. Rhythmicity was tested by testing against a
flat line using F-test.
Acknowledgments
This work was supported by the Deutsche Forschungsgemeinschaft (DFG, Grants MA 5108/1-1,
HE2168/11-1, SPP 2041).
12/15
References
1. D. M. Andrews, M. T. Chow, Y. Ma, C. L. Cotterell, S. V. Watt, D. A. Anthony, S. Akira, Y. Iwakura,
J. A. Trapani, L. Zitvogel, and M. J. Smyth. Homeostatic defects in interleukin 18-deficient mice
contribute to protection against the lethal effects of endotoxin. Immunology and cell biology,
89(6):739–46, aug 2011.
2. A. Arjona and D. K. Sarkar. Evidence supporting a circadian control of natural killer cell function.
Brain, behavior, and immunity, 20(5):469–76, sep 2006.
3. J. Bass and M. A. Lazar. Circadian time signatures of fitness and disease. Science (New York, N.Y.),
354(6315):994–999, jul 2016.
4. M. M. Bellet, E. Deriu, J. Z. Liu, B. Grimaldi, C. Blaschitz, M. Zeller, R. a. Edwards, S. Sahar, S. Dan-
dekar, P. Baldi, M. D. George, M. Raffatellu, and P. Sassone-Corsi. Circadian clock regulates the
host response to Salmonella. Proceedings of the National Academy of Sciences of the United States
of America, may 2013.
5. E. Callaway and H. Ledford. Medicine Nobel awarded for work on circadian clocks. Nature,
550(7674):18, 2017.
6. D. E. Carlson and W. C. Chiu. The absence of circadian cues during recovery from sepsis modifies
pituitary-adrenocortical function and impairs survival. Shock (Augusta, Ga.), 29(1):127–32, jan
2008.
7. M. Casanova-Acebes, C. Pitaval, L. a. Weiss, C. Nombela-Arrieta, R. Ch`evre, N. A-Gonz´alez, Y. Ku-
nisaki, D. Zhang, N. van Rooijen, L. E. Silberstein, C. Weber, T. Nagasawa, P. S. Frenette, A. Cas-
trillo, and A. Hidalgo. Rhythmic modulation of the hematopoietic niche through neutrophil clear-
ance. Cell, 153(5):1025–35, may 2013.
8. B. E. Clausen, C. Burkhardt, W. Reith, R. Renkawitz, and I. F¨orster. Conditional gene targeting in
macrophages and granulocytes using LysMcre mice. Transgenic Research, 8(4):265–277, 1999.
9. A. M. Curtis, C. T. Fagundes, G. Yang, E. M. Palsson-McDermott, P. Wochal, A. F. McGettrick,
N. H. Foley, J. O. Early, L. Chen, H. Zhang, C. Xue, S. S. Geiger, K. Hokamp, M. P. Reilly, A. N.
Coogan, E. Vigorito, G. a. FitzGerald, and L. A. J. O’Neill. Circadian control of innate immunity
in macrophages by miR-155 targeting Bmal1. Proceedings of the National Academy of Sciences of
the United States of America, 112(23):7231–6, jun 2015.
10. J. P. Debruyne, E. Noton, C. M. Lambert, E. S. Maywood, D. R. Weaver, and S. M. Reppert. A clock
shock: mouse CLOCK is not required for circadian oscillator function. Neuron, 50(3):465–77,
may 2006.
11. J. P. DeBruyne, D. R. Weaver, and S. M. Reppert. Peripheral circadian oscillators require CLOCK,
2007.
12. W. Deng, S. Zhu, L. Zeng, J. Liu, R. Kang, M. Yang, L. Cao, H. Wang, T. R. Billiar, J. Jiang, M. Xie,
and D. Tang. The Circadian Clock Controls Immune Checkpoint Pathway in Sepsis. Cell Reports,
24(2):366–378, 2018.
13. D. Druzd, A. de Juan, and C. Scheiermann. Circadian rhythms in leukocyte trafficking. Seminars
in immunopathology, 36(2):149–62, mar 2014.
14. D. Druzd, O. Matveeva, L. Ince, U. Harrison, W. He, C. Schmal, H. Herzel, A. H. Tsang,
N. Kawakami, A. Leliavski, O. Uhl, L. Yao, L. E. Sander, C.-S. Chen, K. Kraus, A. de Juan, S. M. Her-
genhan, M. Ehlers, B. Koletzko, R. Haas, W. Solbach, H. Oster, and C. Scheiermann. Lymphocyte
Circadian Clocks Control Lymph Node Trafficking and Adaptive Immune Responses. Immunity,
pages 1–13, 2017.
15. M. Dudek, N. Gossan, N. Yang, H.-J. Im, J. P. D. Ruckshanthi, H. Yoshitane, X. Li, D. Jin, P. Wang,
M. Boudiffa, I. Bellantuono, Y. Fukada, R. P. Boot-Handford, and Q.-J. Meng. The chondrocyte
clock gene Bmal1 controls cartilage homeostasis and integrity. The Journal of clinical investigation,
126(1):365–76, jan 2016.
16. J. C. Ehlen, A. J. Brager, J. Baggs, L. Pinckney, C. L. Gray, J. P. DeBruyne, K. A. Esser, J. S. Takahashi,
and K. N. Paul. Bmal1 function in skeletal muscle regulates sleep. eLife, 6:1–14, 2017.
17. R. D. Feigin, J. N. Middelkamp, and C. Reed. Circadian rhythmicity in susceptibility of mice to
sublethal Coxsackie B3 infection. Nature: New biology, 240(97):57–8, nov 1972.
18. R. D. Feigin, V. H. San Joaquin, M. W. Haymond, and R. G. Wyatt. Daily periodicity of susceptibility
of mice to pneumococcal infection. Nature, 224(5217):379–380, oct 1969.
19. C. Fleischmann, A. Scherag, N. K. J. Adhikari, C. S. Hartog, T. Tsaganos, P. Schlattmann, D. C. An-
gus, K. Reinhart, and International Forum of Acute Care Trialists. Assessment of Global Incidence
and Mortality of Hospital-treated Sepsis. Current Estimates and Limitations. American journal
of respiratory and critical care medicine, 193(3):259–72, feb 2016.
13/15
20. J. Gibbs, L. Ince, L. Matthews, J. Mei, T. Bell, N. Yang, B. Saer, N. Begley, T. Poolman, M. Pariollaud,
S. Farrow, F. DeMayo, T. Hussell, G. S. Worthen, D. Ray, and A. Loudon. An epithelial circadian
clock controls pulmonary inflammation and glucocorticoid action. Nature medicine, 20(8):919–
26, aug 2014.
21. J. E. Gibbs, J. Blaikley, S. Beesley, L. Matthews, K. D. Simpson, S. H. Boyce, S. N. Farrow, K. J. Else,
D. Singh, D. W. Ray, and A. S. I. Loudon. The nuclear receptor REV-ERBα mediates circadian reg-
ulation of innate immunity through selective regulation of inflammatory cytokines. Proceedings
of the National Academy of Sciences of the United States of America, 109(2):582–7, jan 2012.
22. R. N. Gomes, R. T. Figueiredo, F. A. Bozza, P. Pacheco, R. T. Amˆancio, A. P. Laranjeira, H. C. Castro-
Faria-Neto, P. T. Bozza, and M. T. Bozza. Increased susceptibility to septic and endotoxic shock
in monocyte chemoattractant protein 1/cc chemokine ligand 2-deficient mice correlates with
reduced interleukin 10 and enhanced macrophage migration inhibitory factor production. Shock,
26(5):457–63, nov 2006.
23. F. Halberg, E. A. Johnson, B. W. Brown, and J. J. Bittner. Susceptibility rhythm to E. coli endo-
toxin and bioassay. In Proceedings of the Society for Experimental Biology and Medicine. Society for
Experimental Biology and Medicine (New York, NY), volume 103, pages 142–144, jan 1960.
24. E. Haus and M. H. Smolensky. Biologic rhythms in the immune system. Chronobiology interna-
tional, 16(5):581–622, sep 1999.
25. M. Hayashi, S. Shimba, and M. Tezuka. Characterization of the molecular clock in mouse peri-
toneal macrophages. Biological & pharmaceutical bulletin, 30(4):621–6, apr 2007.
26. W. J. Hrushesky, T. Langevin, Y. J. Kim, and P. A. Wood. Circadian dynamics of tumor necro-
sis factor alpha (cachectin) lethality. The Journal of experimental medicine, 180(3):1059–65, sep
1994.
27. H. Huang, T. Liu, J. L. Rose, R. L. Stevens, and D. G. Hoyt. Sensitivity of mice to lipopolysaccha-
ride is increased by a high saturated fat and cholesterol diet. Journal of inflammation (London,
England), 4:22, 2007.
28. M. Keller, J. Mazuch, U. Abraham, G. D. Eom, E. D. Herzog, H.-D. Volk, A. Kramer, and B. Maier.
A circadian clock in macrophages controls inflammatory immune responses. Proceedings of the
National Academy of Sciences of the United States of America, 106(50):21407–12, dec 2009.
29. J. Laermans, C. Broers, K. Beckers, L. Vancleef, S. Steensels, T. Thijs, J. Tack, and I. Depoortere.
Shifting the circadian rhythm of feeding in mice induces gastrointestinal, metabolic and im-
mune alterations which are influenced by ghrelin and the core clock gene Bmal1. PLoS ONE,
2014.
30. J.-E. Lee and I. Edery. Circadian regulation in the ability of Drosophila to combat pathogenic
infections. Current biology : CB, 18(3):195–9, feb 2008.
31. P. Li, H. Allen, S. Banerjee, S. Franklin, L. Herzog, C. Johnston, J. McDowell, M. Paskind, L. Rodman,
and J. Salfeld. Mice deficient in IL-1 beta-converting enzyme are defective in production of
mature IL-1 beta and resistant to endotoxic shock. Cell, 80(3):401–11, feb 1995.
32. J. Liu, G. Malkani, G. Mankani, X. Shi, M. Meyer, S. Cunningham-Runddles, X. Ma, and Z. S. Sun.
The circadian clock Period 2 gene regulates gamma interferon production of NK cells in host
response to lipopolysaccharide-induced endotoxic shock. Infection and immunity, 74(8):4750–
6, aug 2006.
33. D. Luna-Moreno, R. Aguilar-Roblero, and M. D´ıaz-Mu˜noz. Restricted feeding entrains rhythms
of inflammation-related factors without promoting an acute-phase response. Chronobiology
international, 26(7):1409–29, oct 2009.
34. B. Maier, S. Lorenzen, A.-M. Finger, H.-P. Herzel, and A. Kramer. Searching novel clock genes
using RNAi-based screening. Methods in Molecular Biology, in press.
35. M. W. Marino, a. Dunn, D. Grail, M. Inglese, Y. Noguchi, E. Richards, a. Jungbluth, H. Wada,
M. Moore, B. Williamson, S. Basu, and L. J. Old.
Characterization of tumor necrosis factor-
deficient mice. Proceedings of the National Academy of Sciences of the United States of America,
94(15):8093–8, jul 1997.
36. L. Marpegan, M. J. Leone, M. E. Katz, P. M. Sobrero, T. a. Bekinstein, and D. a. Golombek. Diur-
nal variation in endotoxin-induced mortality in mice: correlation with proinflammatory factors.
Chronobiology international, 26(7):1430–42, oct 2009.
37. K. W. McConnell and C. M. Coopersmith. Pathophysiology of septic shock: From bench to bed-
side. Presse medicale (Paris, France : 1983), 45(4 Pt 2):e93–8, apr 2016.
38. S. M´endez-Ferrer, D. Lucas, M. Battista, and P. S. Frenette. Haematopoietic stem cell release is
regulated by circadian oscillations. Nature, 452(7186):442–7, mar 2008.
39. K. D. Nguyen, S. J. Fentress, Y. Qiu, K. Yun, J. S. Cox, and A. Chawla. Circadian Gene Bmal1
Regulates Diurnal Oscillations of Ly6Chi Inflammatory Monocytes.
Science (New York, N.Y.),
1483(August):1483–8, aug 2013.
14/15
40. D. Novick, S. Kim, G. Kaplanski, and C. A. Dinarello. Interleukin-18, more than a Th1 cytokine.
Seminars in Immunology, 25(6):439–448, 2013.
41. I. Sabroe, E. C. Jones, L. R. Usher, M. K. B. Whyte, and S. K. Dower. Toll-Like Receptor (TLR)2
and TLR4 in Human Peripheral Blood Granulocytes: A Critical Role for Monocytes in Leukocyte
Lipopolysaccharide Responses. The Journal of Immunology, 168(9):4701–4710, 2002.
42. Y. Sanz and A. Moya-P´erez. Microbiota, inflammation and obesity. Advances in experimental
medicine and biology, 817:291–317, 2014.
43. C. Scheiermann, Y. Kunisaki, and P. S. Frenette. Circadian control of the immune system. Nature
reviews. Immunology, 16(February):1–9, feb 2013.
44. C. Scheiermann, Y. Kunisaki, D. Lucas, A. Chow, J.-E. Jang, D. Zhang, D. Hashimoto, M. Merad, and
P. S. Frenette. Adrenergic nerves govern circadian leukocyte recruitment to tissues. Immunity,
37(2):290–301, aug 2012.
45. S. Shimba, N. Ishii, Y. Ohta, T. Ohno, Y. Watabe, M. Hayashi, T. Wada, T. Aoyagi, and M. Tezuka.
Brain and muscle Arnt-like protein-1 (BMAL1), a component of the molecular clock, regulates
adipogenesis. Proceedings of the National Academy of Sciences, 102(34):12071–12076, 2005.
46. K.-F. Storch, C. Paz, J. Signorovitch, E. Raviola, B. Pawlyk, T. Li, and C. J. Weitz. Intrinsic circadian
clock of the mammalian retina: importance for retinal processing of visual information. Cell,
130(4):730–41, aug 2007.
47. M. Stratmann and U. Schibler. Properties, entrainment, and physiological functions of mam-
malian peripheral oscillators. Journal of biological rhythms, 21(6):494–506, dec 2006.
48. W. Wang, J. Y. Barnaby, Y. Tada, H. Li, M. T¨or, D. Caldelari, D.-u. Lee, X.-D. Fu, and X. Dong. Timing
of plant immune responses by a central circadian regulator. Nature, 470(7332):110–4, feb 2011.
49. Z. Xing, J. Gauldie, G. Cox, H. Baumann, M. Jordana, X. F. Lei, and M. K. Achong. IL-6 is an anti-
inflammatory cytokine required for controlling local or systemic acute inflammatory responses.
The Journal of clinical investigation, 101(2):311–20, jan 1998.
50. S.-H. Yoo, S. Yamazaki, P. L. Lowrey, K. Shimomura, C. H. Ko, E. D. Buhr, S. M. Siepka, H.-K. Hong,
W. J. Oh, O. J. Yoo, M. Menaker, and J. S. Takahashi. PERIOD2::LUCIFERASE real-time reporting
of circadian dynamics reveals persistent circadian oscillations in mouse peripheral tissues. Pro-
ceedings of the National Academy of Sciences of the United States of America, 101(15):5339–46, apr
2004.
15/15
| 2019 | Susceptibility rhythm to bacterial endotoxin in myeloid clock-knockout mice | 10.1101/766519 | [
"Lang Veronika",
"Ferencik Sebastian",
"Ananthasubramaniam Bharath",
"Kramer Achim",
"Maier Bert"
] | null |
1
Morphofunctional evaluation of the adrenal gland in rats submitted to nutritional
restriction during pregnancy
Bruno dos Santos Telles, Hércules Jonas Rebelato, Marcelo Augusto Marretto
Esquisatto*, Rosana Catisti.
Programa de Pós-graduação em Ciências Biomédicas, Centro Universitário da
Fundação Hermínio Ometto – FHO, Araras, São Paulo, Brasil.
*Correspondence:
Marcelo Augusto Marretto Esquisatto
Programa de Pós-graduação em Ciências Biomédicas
Centro Universitário da Fundação Hermínio Ometto – FHO
Av. Dr. Maximiliano Baruto, 500 - Jardim. Universitário,
13607-339, Araras, São Paulo, Brasil.
Phone.: +55 19 3543-1440; Fax: +55 19 3543-1439
E-mail: marcelosquisatto@fho.edu.br
2
Abstract
Poor nutrition during pregnancy causes permanent metabolic and/or structural
adaptation in offspring. The adrenal gland produces various steroid hormones during
pregnancy. Thus, this study aimed to evaluate the influence of diet during pregnancy on
the adrenal glands of Wistar rats. For this, 10-week-old pregnant Wistar rats (p, n=15)
and non-pregnant rats (np, n=15) were divided into three groups and received a
normoproteic control diet (C, 17% casein, n=5), isocaloric low-protein diet (PR, 6%
casein, n=5), or 50% calorie restriction (CR, 50% of the diet consumed by group C),
over a period of 21 days. On the 21st day of gestation (21dG, p groups) or on the 21st
day of diet (np groups), after anesthetic deepening, the right adrenal gland was
collected, weighed (total mass), and prepared for inclusion in Paraplast® for
histomorphometric and immunohistochemical analysis (Ki-67, glucocorticoid receptors
(GR), and mineralocorticoid receptor (MR)) in the different areas of the gland. Data,
expressed as the mean ± SD, were evaluated by one-way analysis of variance with
Tukey's post-test (p < 0.05). CR in pregnancy increased the amount of GR, MR, and Ki-
67 receptors in the adrenal gland. The npRC group showed highest GR staining
compared to the animals that received a normal diet. Protein restriction in pregnancy
decreases adrenal MR. The results allowed us to conclude that even without altering the
weight of the adrenal glands, the pRC group suffered the most from stress during the
study, suggesting that CR associated with pregnancy can cause morphofunctional
changes in the adrenal glands.
KEYWORDS
caloric
restriction,
protein
restriction,
pregnancy,
adrenal,
glucocorticoid/mineralocorticoid receptors.
3
1 | INTRODUCTION
Adrenal glands (AGs) are bilateral structures located above the upper pole of the
kidneys. In healthy young women, the GA is pink, smooth, and opaque. In men, it is a
smaller and more reddish organ that appears slightly translucent. When sectioned, the
tissue is soft, and the pith, with a rounded to oval appearance, is centrally placed and
dark red in color. In terms of weight, the adrenal gland is approximately 25% heavier in
women than in men, as women have a wider cortex, but the medulla is the same size in
both sexes (Hardy & Cooper, 2010). Histologically, it is divided into the cortex and
medulla and acts as an essential regulator of the stress response (Kanczkowski, Sue &
Bornstein, 2017). In rats, the adrenal glands are paired above the kidneys; the right
adrenal gland is somewhat medial to the superior pole and intimately connected to it,
while the left GA is above this organ (Dunn, 1970).
Calorie restriction (CR) is defined as a reduction in caloric intake below usual ad
libitum without malnutrition, which generally represents a 10–40% decrease in caloric
intake with no reduction in the nutritional content of the diet (Bagherniya et al., 2018).
This results in a delay in aging, prolongation of the maximum and average lifespan in
animals of different species, and a significant decrease in cardiovascular diseases,
diabetes, neurodegenerative diseases, and cancers (Al-Regaiey, 2016).
Protein restriction (PR) can be defined as restriction of amino acid intake
without malnutrition (Youngman, 1993). Studies on yeast and flies have shown that
amino acid restriction promotes longevity and protection. In rodents, protein restriction
prolongs lifespan and alleviates harmful phenotypes associated with aging (Mirzaei;
Raynes & Longo, 2016).
Other significant health benefits of nutritional restriction have also been
demonstrated, including decreased tumor angiogenesis (Hawrylewicz et al., 1982;
Youngman & Campbell, 1992), antioxidant enzymes that enhance defenses (Lammi-
Keefe et al., 1984), improved immunologic responses (Jose & Good, 1973; Bell et al.,
1990), and reduction of total serum cholesterol (Terpstra et al., 1981; Youngman, 1987).
PR animals generally have a smaller body size (Youngman & Campbell, 1992) and
more physically active (Krieger et al., 1988). Furthermore, both PR (Youngman; Park &
Ames, 1992) and RC (Lok et al., 1990) significantly decrease cell division rates in many
tissues (Youngman, 1993).
4
During the gestational period, numerous changes occur in the body to meet the
needs of the mother and fetus. Among these changes are the accumulation of maternal
adipose tissue, increased metabolism rate along with increased cardiac output and
respiratory rate, and a higher calorie intake, which is essential for having a pregnancy
that does not bring risks to the pregnant woman, let alone to the offspring (King, 2000).
During pregnancy, nutrient metabolism undergoes adjustments caused by
hormonal changes, the demand of the fetus, and maternal supply of nutrients, especially
during the last half of pregnancy, which is the period in which the fetus grows the most.
These transitions, along with behavioral habits, changes in the amount of food
consumed or energy expended, food choices, or type of physical activities of mothers,
increase the physiological adjustments necessary during pregnancy. However, when the
physiological adaptation of the body is exceeded during this phase, fetal development
can be harmed (King, 2000).
Deficient nutrition during pregnancy results in permanent metabolic and/or
structural adaptations in the offspring. Females with caloric deficiencies or malnutrition
during pregnancy affect their offspring, increasing the risk of developing pathologies in
adult life, such as metabolic syndrome, obesity, cardiovascular diseases, and type 2
diabetes mellitus (Chango & Pogribny, 2015). It is known that pregnant females can
suffer from increased blood pressure (Gao; Yallampalli & Yallampalli, 2012), along
with changes in the immune system (Thiele; Diao & Ark, 2017), as well as being
subject to changes in their AG during the gestational period.
Although progressive, there has been a reduction in the number of malnourished
individuals in today's society. The dietary patterns of the contemporary society have
undergone changes due to advances in food production and industrialization
technologies. Other factors have also changed, such as the assimilation of cultural
patterns and modification of life habits. Thus, today, important issues related to the
impact of nutritional restriction on metabolism should be evaluated in animal models
with dietary restrictions. The changes in GA during pregnancy in rats are poorly
understood. Given the importance of this gland in the production of hormones in
females, describing the changes induced by food restriction and pregnancy is essential
to understand its physiology. Therefore, this study aimed to evaluate the
morphofunctional organization of the adrenal gland in young adult Wistar rats subjected
to nutritional restriction (CR and PR), regardless of the pregnancy status.
5
2 | MATERIALS AND METHODS
2.1 | Experimental procedure
The study was carried out in accordance with the rules established by the Arouca
Law, approved by the ethical principles of animal research adopted by COBEA and by
the Ethics Committee on Animal Use of the Centro Universitário da Fundação
Hermínio Ometto, FHO, opinion 062/2016. Female Wistar rats (10 weeks old, weighing
approximately 250–300 g) were subjected to mating. Once the presence of spermatozoa
in the vaginal lavage was verified, these animals were called the pregnant group (p, n =
15) and the other group of non-pregnant rats was called the non-pregnant group (np, n =
15). After separating the groups, the rats were divided into three subgroups: those that
received a normoproteic control diet (C, 17% casein, n = 5), an isocaloric low-protein
diet (PR, 6% casein, n = 5), or caloric restriction of 50% (CR, 50% of the diet consumed
by group C, n=5) for a period of 21 days. Diets were calculated daily, considering the
weight of the amount offered and the amount that was left for the controls, that is, the
amount ingested by the control group. From this, 50% was calculated for the RC group.
The rats were kept in individual cages in a temperature-controlled environment (21 ± 1º
C) with a 12 h light/dark cycle and free access to water. On the 21st day of gestation
(21dG, P animals) or on the 21st day of diet (NP groups), after deep anesthesia with
ketamine (100 mg/kg) and xylazine (10 mg/kg), the animals' right adrenals were
collected, weighed, and processed for structural analysis.
2.2 | Body growth and food consumption
Rats were weighed once a week on days 0, 7, 14, and 21 of the experimental (or
gestational) study. The diet was weighed daily for 21 days. In addition to the growth
curve, the mass gain after subtracting the initial masses was determined. Food
consumption was determined by the difference between the weight of the feed added
and the feed remaining in the cages.
2.3 | Processing for the histomorphometric study of adrenal
After removal, the adrenals were weighed and immersed in a fixative solution
containing 10% formaldehyde in Millonig buffer pH 7.4 for 24 h at room temperature.
Then, the pieces were washed in buffer and submitted to standard procedures for
6
embedding in paraffin (Paraplast® - Merck). Cross-sections of 5 µm thick pieces were
subjected to hematoxylin-eosin staining. Three samples were used for each of the five
sections obtained from the median region of each of the three animals in each treatment.
Biopsy images were captured on a Leica DM2000 microscope using Leica Application
Slite software (version 3.3.0).
From the images, the Image J program (National Institutes of Health, Bethesda,
MD, USA) was calibrated to measure the areas of the cortex and medulla, measuring the
scale bar divided by its value (50 µm), which resulted in a value of 8.68 µm/pixel. After
this process, measurements were started by contouring and measuring the total area and
medullary area. The total area was subtracted from the medullary area, which gave the
adrenal cortical area. After the measurement, the following ratios were calculated:
cortical area/total area, medullary area/total area, and adrenal mass/animal mass. The
entire process was performed for all groups, and the results were statistically compared.
2.4 | Quantification of connective tissue (collagen) in the adrenal by Mallory's
trichrome staining
Cross-sections of 5 µm thick pieces were stained with Mallory's trichrome stain.
Three samples were used for each of the five sections obtained from the median region
of each of the three animals in each treatment. Biopsy images were captured on a Leica
DM2000 microscope using Leica Application Slite software (version 3.3.0). The images
were analyzed using Image J software (National Institutes of Health, Bethesda, MD,
USA) by color deconvolution and statistically analyzed.
2.5 | Processing for immunohistochemistry analyses
We evaluated the expression of glucocorticoid receptors (GR), mineralocorticoid
receptors (MR), and Ki-67 antigen in the adrenal glands of pregnant and non-pregnant
young adult rats subjected to different nutritional protocols. All procedures were
performed according to the protocol established by Gianchini et al. (2007). Briefly,
antigen retrieval was performed by immersing the silanized slide in sodium citrate
solution (10 mM, pH 6.0) for 40 min at 95°C. Each step was followed by washing with
PBS. All steps were performed in a humid chamber under care to avoid dehydration of
the sections. Incubation with the primary antibody was performed by incubating the
sections with anti-GR (monoclonal mouse, Santa Cruz, USA), anti-MCR (monoclonal
mouse, Santa Cruz, USA), and anti-Ki-67 (monoclonal mouse, Santa Cruz, USA)
7
antibodies which were diluted 1:200 in PBS containing 3% bovine albumin (v/v)
overnight at 4°C. After the primary antibody reaction, a Novolink Polymer Detection
Systems Kit (RE7280K; Leica Biosystems Newcastle L10, Newcastle Upon Tyne, UK)
containing the secondary antibody was used. After washing in PBS, the peroxidase
reaction was visualized using DAB (3,3'-diaminobenzidine) from the same kit. For each
immunohistochemical reaction, a negative control of the adrenal sections was
performed, omitting the primary antibody. The sections were examined using a Leica
DM2000 Photomicroscope in images digitized with the support of the Sigma Scan Pro
5.0™ program and evaluated by area (µm2) using the Image J software (National
Institutes of Health, Bethesda, MD, USA). Quantification was based on the
decomposition of the immunohistochemical image into three base colors: brown
(immunohistochemistry), purple (Harris hematoxylin), and green (background of glass
slides). Morphometric analysis, corresponding to brown color, was performed using the
threshold function (ImageJ), and antibodies/markers were measured as the percentage of
total pixels in each image (Landini, Martinelli & Piccinini, 2021). Data are reported as
the percentage area of the respective antibody.
2.6 | Statistical analysis
Data were compared using analysis of variance (ANOVA) followed by Tukey's
post-hoc test using GraphPad Prism software (GraphPad Software, Inc. La Jolla, CA,
USA) with a significance level of 5% (p < 0). .05, n = 5). The results were expressed as
mean ± standard deviation (X ± SD) and later represented as a percentage of variation in
relation to controls, to which a value of 100% was assigned.
3 | RESULTS
3.1 | Effect of nutritional restriction on female characteristics
The body mass gain of rats was analyzed by weighing the animals weekly at
time 0 (mating day) and on the 7th, 14th, and 21st gestational days (Figure 1).
Throughout the study period, the pC group of animals presented an ascending weight
curve, indicating that the expected growth occurred during pregnancy, while in the pRC
group, there was weight loss during the first 14 days when compared to the initial
weight, with mass gain only in the last week of pregnancy. The pRP group also showed
a lower body mass gain. There was a gradual decrease in weekly consumption of diet in
8
the groups when evaluated from the 1st to 3rd gestational weeks. Pregnancy and/or diet
did not alter the mass in the right adrenal gland.
3.2 | Effect of caloric restriction on adrenal gland histomorphometry
There was no significant difference between the experimental groups for the
cortical, medullary, and total areas. However, the difference was evident in the adrenal
mass/animal mass ratio, with the npC, pC, and pRP groups showing lower values than
the npRC group. The pRC and npRP groups showed values higher than pRP and pC,
which were statistically lower than those of npRP (Figure 2).
3.3 | Quantification of Connective Tissue (Collagen) by Mallory's Trichrome
In the quantification of connective tissue (collagen) by Mallory's trichrome, the
zona glomerularis of the pRC group showed the smallest amount of connective tissue
(collagen) area compared to the npC group (Figure 3A). In other areas, there were no
differences between the groups (Figure 3).
3.4 | Cell division in the adrenal glands assessed using the Ki-67 marker
The pRP group had a reduced number of dividing cells in the zona glomerularis
compared with the pC and pRC groups. Among the control groups, pC showed greater
cell proliferation than npC. In the zona reticularis, the nutritionally restricted groups,
npRC and npRP, showed higher Ki-67 staining than the npC group. In the reticular and
medullary zones, the npRP group had a greater number of dividing cells than the pRP
group (Figure 4).
3.5 | Effect of diets on GRs in adrenal glands
The zona glomerulosa (Figure 5A) of the npRC group showed the lowest
expression of GR compared to those of the npC, pC, and pRC groups. The npRP and
pRP groups showed a lower amount of labeling for this receptor than the npC, pC, and
pRC groups. GR receptor expression was less marked in the npRP group than in the
npC group. The pRP group showed more GR in the zona glomerulosa and fasciculata
than the pC and pRC groups (Figure 5).
9
3.6 | Effect of diet on mineralocorticoid receptors (MR) in the adrenal glands
There were no differences between the control groups in the four adrenal gland
zones. However, the pRC group had a greater presence of MR markings than the other
five groups. The fasciculate and reticulate zones in the pRP group showed fewer
receptors than those in the npRP group (Figure 6).
4 | DISCUSSION
In this study, we investigated whether PR and CR during pregnancy could
modify the morphology of the adrenal gland as well as the expression of glucocorticoid
and mineralocorticoid receptors. Throughout the experimental period, animals presented
ascending body mass curves, which indicated normal gestational growth.
CR was validated by the lower food intake of the pRC group during the
experimental period. According to the literature, mothers undergoing CR, in addition to
having offspring with lower birth weight, also have lower body mass during pregnancy
(Barker, 2002). Another finding that validates our study is the upward curve of the pC
group, the pregnant group that gained the most mass during pregnancy. A low-protein
maternal diet significantly reduced weight gain throughout pregnancy (Cottrell et al.,
2012).
Several studies in the literature have demonstrated that the mass of the adrenal
gland is restricted, whether in offspring or in mothers, corroborating the data found in
this study (Rosenbrock et al., 2005). Huseby et al. (1945) and Boutwell et al. (1948)
showed that CR in rats resulted in adrenal hypertrophy without necessarily increasing
rat weight (Kritchevsky, 2001). Contrary results have been observed in other studies, in
which an increase in adrenal mass in maternal CR of 50% was observed during the last
week of pregnancy (Eleftheriades; Creatsas & Nicolaides, 2006) and a smaller adrenal
mass was observed (Liang; Zhang & Zhang, 2004).
There have been no specific studies on calorie-restricted or protein-restricted
adrenal mass/animal mass ratios. Interestingly, among the nutritionally restricted
pregnancy groups, the pRP group had a lower adrenal mass/animal mass ratio than the
pRC group. In a study of sodium restriction during pregnancy, there was a 33% increase
in the width of the zona glomerulosa in rats. The combination of dietary sodium
restriction and pregnancy caused a 167% increase in the width of the zona glomerulosa.
There was a direct relationship between the number of cells and the width of the zone,
10
indicating that hyperplasia accompanies hypertrophy of the zona glomerulosa (Pohanka
& Pike, 1970), which was not observed in this study, and the cortical and medullary
areas did not present significant differences between groups. Future studies are
necessary to verify the presence of hyperplasia in a specific zone.
In the presence of inflammation or tissue damage, type I collagen is present in
the remodeling of the damaged site (González et al., 2016) and to verify whether
restriction in conjunction with pregnancy resulted in increased connective tissue
(collagen) deposition in GA, Mallory's trichrome staining was performed, but no
differences were observed between the pregnant and nutritionally restricted groups,
which confirms the previous data on the absence of an increase in the thickness of the
cortical and medullary areas. Only in the zona glomerulosa, the npC group showed a
greater presence of connective tissue (collagen) compared to the pRC groups. Cortisol is
known to be responsible for protein degradation (Silverthor, 2010). We noticed, even if
discreetly, that the nRC group presented a smaller amount of connective tissue
(collagen) in the AG areas, which may indicate that the CR increases stress and
consequently cortisol. It has been suggested that a larger RC or RP may present more
significant results.
Ki-67 is expressed in cell nuclei during proliferation (Sun & Kaufman, 2018), as
a way for tissue to repair damaged cells or meet the organ's demands. Its expression was
higher in the groups with unrestricted pregnancy, with greater cell proliferation in all
four areas of the adrenal gland when compared to non-pregnant rats, corroborating the
data in the literature (Pohanka & Pike, 1970).
Regarding the zona glomerularis, it was observed that pregnancy with control
feeding and with CR had an increase in cell proliferation in the AG in relation to their
non-pregnant peers, suggesting possible hyperplasia and hypertrophy in these groups.
The opposite occurred in pregnancy with RP, which showed lower cell proliferation
than the npRP groups. Although the literature does not report specific studies of KI-67
in RP, some studies on cell proliferation in CR can explain this phenomenon. Studies of
this type have indicated a reduction in cell proliferation in keratinocytes, liver cells,
mammary epithelial cells, splenic T cells, and prostate cells in 30–50% CR (Bruss et al.,
2011; Hsieh et al., 2004; Lok et al., 1990), decreased proliferation of basal cells in the
olfactory mucosa of mice (Iwamura et al., 2019), and decreased cell proliferation and an
increase in apoptotic cell death (Dunn et al., 1997). The opposite result was observed
only in some areas of the brain, such as increased neurogenesis in the dentate gyrus of
11
the hippocampus and the subventricular zone (Kumar et al., 2009; Park et al., 2013;
Iwamura et al., 2019). In studies on sheep in the last trimester of pregnancy, the adrenal
cortex showed occasional scattered mitotic figures in the zona glomerulosa (Hill et al.,
1984).
The only study that addressed the results of cell proliferation in AG was in
conducted with a low-sodium diet, where an increase in zona glomerulosa cells was
observed together with an increase in aldosterone secretion (Ennen; Levay-Young &
Engeland, 2005). Inomata and Sasano (2015) identified greater mitotic activity in the
human AG in the region between the zona glomerulosa and zona fasciculata.
Hill et al. (1984) observed that animals subjected to nutritional restrictions,
without pregnancy, presented greater proliferative activity in the reticular and medullary
zones in relation to the npC group, although in animals fed a normal non-pregnant diet,
mitotic activity was absent throughout the adrenal cortex. This finding indicates that RC
and RP diets can increase the number of cells in these GA zones.
Elevated levels of Ki-67 and circulating aldosterone expression were associated
with treatment with MR antagonists in hypertensive rats, which was observed by
immunohistochemistry of ZG cells. This suggests that the greater the hormonal demand
by the organism, the more the cells multiplied to supply the hormones (Pereira et al.,
2021). This may have occurred in the groups that showed greater cell proliferation,
especially the pRC group.
Synthesized or secreted glucocorticoids may play an important role in the direct
regulation of adrenocortical cell proliferation and function under physiological
conditions (Saito et al., 1979). GR expression in the human adrenal cortex was
originally demonstrated by Loose et al. (1980) and was later confirmed in more recent
investigations (Paust et al., 2006; Asser et al., 2014). In humans, GR is expressed in the
adrenal cortex with functions parallel to those found in other tissues (Briassoulis et al.,
2011; Spiga et al., 2017 ). In addition to its effect on the pituitary and hypothalamus, the
cortisol can affect its own synthesis via a local feedback mechanism within the adrenal
gland (Gjerstad, Lightman & Spiga, 2018).
MR and GR are present in the adrenal gland, zona glomerulosa, fasciculata and
reticulate in humans (Boulkroun et al., 2010), and in sodium-restricted/no MR and GR
mice were highly expressed in ZG and ZF/ZR cells (Chong et al., 2017).
GR and MR receptors were observed in newborn rats whose mothers had
undergone 50% food restriction during the last week of gestation. Food restriction
12
induces a delay in intrauterine growth, disrupts the HPA axis, and decreases adrenal
weight, which was not observed in this study. In addition, newborn mice showed a
reduction in MR and GR mRNA in the hippocampus, reduction of CRH mRNA in the
paraventricular nuclei of the hypothalamus, and reduction in the plasma levels of
adrenocorticotropic hormone (Léonhardt et al., 2002). The same was observed in the
hypothalamus of offspring with maternal RP (Bertram et al., 2001), which confirmed
our results for GR in the pRP group, which was lower than that in the pC group. The
inverse occurred in the expression levels of GR mRNA, which were significantly higher
in the kidney, lung, liver, and hippocampus of fetal and neonatal pups at the end of
gestation (20 days) and in 12-week offspring exposed to maternal RP, suggesting that
this increase is persistent throughout life (Bertram et al., 2001). MR labeling showed
that in the zona glomerulus and reticularis, there was greater labeling in the pRP group
than in the npRP group, which suggests that pregnancy influences the increase in MR,
with glucocorticoids activating MR in most tissues at baseline levels and the GR at
stress levels (Gomez-Sanchez & Gomez-Sanchez, 2014).
GR mRNA expression levels in nutrient-restricted neonatal ewe pups were the
highest in adrenal, kidney, liver, lung, and perirenal adipose tissues, where the
persistence of tissue-specific increases in GR, 11β-hydroxysteroid dehydrogenase type,
was demonstrated. 1-11bHSD1 (which transforms cortisone to cortisol) and angiotensin
II receptor type 1 (AT1) decrease the expression of 11β-hydroxysteroid dehydrogenase
type 2-11bHSD2pa (which oxidizes cortisol) in the adrenals and kidneys of newborn
infants in response to a defined period of maternal nutrient restriction during early
pregnancy. The authors inferred that gene expression is programmed by the availability
of nutrients to the fetus before birth (Bertram et al., 2001). This may explain the large
increase in MR and GR receptors in the pRC group; due to food restriction, there was a
greater production of cortisol and stress due to the greater expression of MR and GR.
Malnutrition in pregnant rats in GA causes a decrease in GR mRNA expression
due to stress, which increases the maternal production of corticosterone. This makes
tissues most sensitive to corticosteroid concentration. This in turn stimulates The
expression of HSD11B1 in rats (Khorram et al ., 2011) causing excess maternal and
fetal plasma corticosterone, downregulating fetal GR and MR, and compromising the
HPA feedback axis in childhood and adulthood (Valsamakis, Chrousos & Mastorakos,
2019). Although the groups did not suffer from malnutrition in the glomerular,
fasciculate, and medullary zones, the pRC group showed greater staining for GR when
13
compared to the pRP group, and the same was repeated for the MR staining in the four
AG zones. The results suggest that the groups may have experienced stress of
restriction, with CR having an impact on rats during pregnancy compared to RP.
However, the results were different from those found in the literature since there was no
negative regulation of the receptors, as the expression levels of MR and GR were higher
in the pRC group.
The negative regulation of the receptors was noted by lower GR labeling in the
npRC group than in the npC group. Although there are no comparative studies between
PR and CR in the literature, we present in this study that in the reticular and medullary
zones, the npRC group had lower markings for GR compared to the npRP group, and
the opposite occurred in the medullary zone when compared to the CR pregnant groups
and the PR groups. MR expression levels were unaffected by the maternal diet in the
kidneys of offspring of maternal RP and were undetectable in the lungs (Bertram et al.,
2001).
In all GA, both GR and MR were expressed more in pRC than in npRC,
suggesting that pregnancy is an additional stressor because in pregnant primates, there
was an increase in maternal cortisol (Recabarren, Valenzuela & Seron-Ferrer, 1997).
However, in a study in adult male rats with perinatal malnutrition, the levels of GR and
MR mRNA expression and the binding capacity or affinity showed no difference
between groups in GA (Dutriez-Casteloot et al., 2008).
In the offspring of pregnant hamsters, the rate of steroidogenesis increased in
malnourished rats (Liang, Zhang & Zhang, 2004) and although we did not measure
cortisol directly, the results of the pregnant group with CR showed an increase in the
number of GR and MR receptors, increased cell proliferation, and indirect features for
an increase in cortisol production by GA cells.
It has been suggested that aldosterone secretion in the rat ZG can be regulated by
MR through ultra-short feedback within the adrenal gland, where aldosterone regulates
its own production, and because it can be activated by the same hormone, GR can also
regulate the production of glucocorticoids in the ZF and ZR (Chong et al., 2017). This
phenomenon may explain our results on the increase in the number of receptors that
may regulate cortisol production without necessarily increasing the number of cells for
this role.
However, there is disagreement as to whether the feedback exerted by the
MR/GR receptors within the adrenal gland is positive or negative. Some studies have
14
stated that the feedback provided is positive. In vitro research with H295R cells (a
human cortisol-secreting adrenocortical cell line) revealed the presence of an intra-
adrenal positive feedback loop that regulates steroid production. These results were
confirmed when GR was inactivated by the pharmacological antagonist RU486 or GR
knockdown by siRNA, which led to the suppression of steroidogenesis, strongly
suggesting an autocrine and GR-mediated ultra-short autocrine positive regulatory loop
(Asser et al., 2014). This type of feedback corresponds to the results of the pRC groups,
in which there were increases in GR and MR receptors in response to past stress during
the pregnancy and CR periods.
Chong et al. (2017) pointed out that this regulation occurs through negative
feedback because MR and GR negatively regulate glucocorticoid production in ZF/ZR
cells (intra-adrenal feedback–short loop). In vitro and in vivo studies have shown that
prior exposure of the adrenal gland to glucocorticoids results in a diminished response
to adrenocorticotropic hormone (ACTH), resulting from an intra-adrenal negative
feedback loop that could constitute an additional GR-regulated control mechanism for
steroidogenesis (Peron et al., 1960; Carsia & Malamed, 1979; Chong et al., 2017;
Gjerstad, Lightman, & Spiga, 2018). This may justify that even with an increase in cell
replication, increased measurements in the adrenal zones in some groups are not
sufficient to increase cortisol production as they are regulated by negative intra-adrenal
feedback by MR and GR.
The results showed that PR and CR during pregnancy did not change the weight
of the adrenal glands when compared to non-pregnant women. RC increased the
expression of GR and MR receptors in GA during pregnancy, whereas RP decreased the
labeling of GR and MR in the zona glomerulosa and fasciculata. We concluded that CR
during pregnancy caused the most stress to the rats, altering the presence of MR and
GR, which may suggest an alteration in the functionality of the GA and, consequently,
in the HPA axis.
ACKNOWLEDGMENTS
This research was supported by the Hermínio Omettto Foundation.
AUTHOR CONTRIBUTIONS
Bruno dos Santos Telles: Carried out experimental work, data collection and data
evaluation. Hércules Jonas Rebelato: Carried out experimental work, data collection
15
and data evaluation. Marcelo Augusto Marretto Esquisatto: Conceptualization,
methodology, validation, formal analysis, investigation and writing. Rosana Catisti:
Conceptualization, methodology, validation, formal analysis, investigation, writing -
review & editing of manuscript and supervision.
ORCID
Marcelo Augusto Marretto Esquisatto https://orcid.org/0000-0002-2588-619X
REFERENCES
Al-Regaiey, K.A. (2016). The effects of calorie restriction on aging: a brief review.
European Review for Medical and Pharmacological Sciences, 20(11), 2468-2473.
Asser, L., Hescot, S., Viengchareun, S., Delemer, B., Trabado, S. & Lombès, M. et al.
(2014). Autocrine positive regulatory feedback of glucocorticoid secretion:
glucocorticoid receptor directly impacts h295r human adrenocortical cell
function. Molecular and Cellular Endocrinology, 395 (1-2), 1-9.
Bagherniya, M., Butler, A. E., Barreto, G. E. & Sahebkar, A. (2018). The effect of
fasting or calorie restriction on autophagy induction: a review of the
literature. Ageing Research Reviews, 47 (1), 183-197.
Barker, D.J. (2002). Fetal programming of coronary heart disease. Cell Press, 13 (9),
364-368.
Bell, R.C. et al. (1990). Natural killer cell activity and tumorigenesis in animals fed low
protein diets. FASEB Journal, 4 (4), 1043-1047.
Bertram, C., Trowern, A. R., Copin, N., Jackson, A.A. & Whorwood, C. B. (2001). The
Maternal Diet during Pregnancy Programs Altered Expression of the
Glucocorticoid Receptor and Type 2 11β-Hydroxysteroid Dehydrogenase:
potential molecular mechanisms underlying the programming of hypertension in
utero. Endocrinology, 142 (7), 2841-2853.
Boulkroun, S., Samson-Couterie, B., Golib, J.F., Lefebvre, H., Louiset, E., Amar, L.,
Plouin, P.F., Lalli, E., Jeunemaitre, X., Benecke, A., Meatchi, T. & Zennaro, M.C.
(2010). Adrenal Cortex Remodeling and Functional Zona Glomerulosa
Hyperplasia in Primary Aldosteronism. Hypertension, 56 (5), 885-892.
Boutwell, R.K., Brush, M.K. & Rusch, H.P. (1948). Some physiological effects
associated with chronic caloric restriction. American Journal of Physiology-
Legacy Content, 154 (3), 517-524.
16
Briassoulis, G., Damianovic, S., Xekouki, P., Lefebvre, H. & Stratakis, C.A. (2011).
The Glucocorticoid Receptor and its Expression in the Anterior Pituitary and the
Adrenal Cortex: a source of variation in hypothalamic-pituitary-adrenal axis
function; implications for pituitary and adrenal tumors. Endocrine Practice, 17
(6), 941-948.
Bruss, M.D., Thompson, A.C.S., Aggarwal, I., Khambatta, C.F. & Hellerstein, M. K.
(2011). The effects of physiological adaptations to calorie restriction on global
cell proliferation rates. American Journal of Physiology-Endocrinology and
Metabolism, 300 (4), 735-745.
Carsia, R.V. & Malamed, S. (1979). Acute Self-Suppression of Corticosteroidogenesis
in Isolated Adrenocortical Cells. Endocrinology, 105 (4), 911-914.
Chango, A. & Pogribny, I. (2015). Considering Maternal Dietary Modulators for
Epigenetic Regulation and Programming of the Fetal Epigenome. Nutrients, 7 (4),
2748-2770.
Chong, C., Hamid, A., Yao, T., Garza, A.E., Pojoga, L. H., Adler, G.K., Romero, J.R. &
Williams, G.H. (2017). Regulation of aldosterone secretion by mineralocorticoid
receptor–mediated signaling. Journal of Endocrinology, 232 (3), 525-534.
Cottrell, E.C., Holmes, M.C., Livingstone, D.E., Kenyon, C.J. & Seckl, J.R. (2012).
Reconciling
the
nutritional
and
glucocorticoid
hypotheses
of
fetal
programming. The Faseb Journal, 26 (5), 1866-1874.
Dunn, T.B. (1970). Normal and Pathologic Anatomy of the Adrenal Gland of the
Mouse, Including Neoplasms. Journal of the National Cancer Institute, 44 (6),
1323-1389.
Dutriez-Casteloot, I., Breton, C., Coupé, B., Hawchar, O., Enache, M., Dickes-
Coopman, A., Keyzer, Y., Deloof, S., Lesage, J. & Vieau, D. (2008). Tissue-
specific Programming Expression of Glucocorticoid Receptors and 11β-HSDs by
Maternal Perinatal Undernutrition in the HPA Axis of Adult Male Rats. Hormone
and Metabolic Research, 40 (4), 257-261.
Eleftheriades, M., Creatsas, G. & Nicolaides, K. (2006). Fetal Growth Restriction and
Postnatal Development. Annals of The New York Academy of Sciences, 1092 (1),
319-330.
Ennen, W.B., Levay-Young, B.K. & Engeland, W.C. (2005) Zone-specific cell
proliferation
during
adrenocortical
regeneration
after
enucleation
in
17
rats. American Journal of Physiology-Endocrinology and Metabolism, 289 (5),
883-891.
Gao, H., Yallampalli, U. & Yallampalli, C. (2012) Protein Restriction to Pregnant Rats
Increases the Plasma Levels of Angiotensin II and Expression of Angiotensin II
Receptors in Uterine Arteries. Biology of Reproduction, 86 (3), 1-8.
Gianchini, F.R.C., Carriel, V., Capelo, L.P., Tostes, R.C., Carvalho, M.H., Fortes, Z.B.,
Zorn, T.M. & San Martin, S. (2007). Maternal diabetes affects specific
extracellular matrix components during placentation. Journal of Anatomy, 212 (1),
31-41.
Gjerstad, J.K., Lightman, S.L. & Spiga, F. (2018). Role of glucocorticoid negative
feedback in the regulation of HPA axis pulsatility. Stress, 21 (5), 403-416.
Gomez‐Sanchez,
E.
&
Gomez‐Sanchez,
C.E.
(2014).
The
Multifaceted
Mineralocorticoid Receptor. Comprehensive Physiology, 4 (3), 965-994.
González, A.C.O., Costa, T.F., Andrade, Z.A. & Medrado, A.R.A.P. (2016). Wound
healing - A literature review. Anais Brasileiros de Dermatologia, 91 (5), 614-620.
Hardy, R. & Cooper, M.S. (2010). Adrenal gland and bone. Archives of Biochemistry
and Biophysics, 503 (1), 137-145.
Hawrylewicz, E.J., Huang, H.H., Kissane, J.Q. & Drab, E.A. (1982). Enhancement of
7,12-dimethylbenz[a]anthracene (DMBA) mammary tumorigenesis by high
dietary protein in rats. Nutrition Reports International, 26 (5), 793-806.
Hill, P.A., Coghlan, J.P., Scoggins, B.A. & Ryan, G.B. (1984). Functional and
morphologic studies of the adrenal cortex and kidney in ovine toxaemia of
pregnancy. The Journal of Pathology, 144 (1), 1-13.
Hsieh, E.A., Chai, C.M., Lumen, B.O., Neese, R. A. & Hellerstein, M.K. (2004).
Dynamics of Keratinocytes in Vivo using 2H2O Labeling: a sensitive marker of
epidermal proliferation state. Journal of Investigative Dermatology, 123 (3), 530-
536.
Huseby, R.A., Ball, Z.B. & Visscher, M.B. (1945). Further observations on the
influence of simple caloric restriction on mammary cancer incidence and related
phenomena in C3H mice. American Association for Cancer Research, 5 (1), 40-
46.
Inomata, A. & Sasano, H. (2015). Practical approaches for evaluating adrenal toxicity in
nonclinical safety assessment. Journal of Toxicologic Pathology, 28 (3), 125-132.
18
Iwamura, H., Kondo, K., Kikuta, S., Nishijima, H., Kagoya, R., Suzukawa, K., Ando,
M., Fujimoto, C., Toma-Hirano, M. & Yamasoba, T. (2019). Caloric restriction
reduces basal cell proliferation and results in the deterioration of neuroepithelial
regeneration following olfactotoxic mucosal damage in mouse olfactory
mucosa. Cell and Tissue Research, 378 (2), 175-193.
Jose, D.G. & Good, R.A. (1973). Quantitative effects of nutritional essential amino acid
deficiency upon immune responses to tumors in mice. Journal of Experimental
Medicine, 137 (1), 1-9.
Kanczkowski, W.; Sue, M. & Bornstein, S.R. (2017) The adrenal gland
microenvironment in health, disease and during regeneration. Hormones, 13 (3),
251-265.
Khorram, N.M., Magee, T.R., Wang, C., Desai, M., Ross, M. & Khorram, O. (2011).
Maternal
Undernutrition
Programs
Offspring
Adrenal
Expression
of
Steroidogenic Enzymes. Reproductive Sciences, 18 (10), 931-940.
King, J.C. (2000). Physiology of pregnancy and nutrient metabolism. The American
Journal of Clinical Nutrition, 71 (5), 1218-1225.
Krieger, E., Youngman, L.D., Campbell, T.C. (1988). The modulation of aflatoxin B1
(AFB1)-induced preneoplastic lesions by dietary protein and voluntary exercise in
Fischer 344 rats. FASEB Journal, 2 (4), 3304.
Kritchevsky, D. (2001). Caloric Restriction and Cancer. Journal of Nutritional Science
and Vitaminology, 47 (1), 13-19.
Kumar, S., Parkash, J., Kataria, H. & Kaur, G. (2009). Interactive effect of excitotoxic
injury and dietary restriction on neurogenesis and neurotrophic factors in adult
male rat brain. Neuroscience Research, 65 (4), 367-374.
Lammi-Keefe, C.J., Swan, P.B. & Hegarty, P.V. (1984). Effect of Level of Dietary
Protein and Total or Partial Starvation on Catalase and Superoxide Dismutase
Activity in Cardiac and Skeletal Muscles in Young Rats. The Journal of Nutrition,
114 (12), 2235-2240.
Landini, G.; Martinelli, G. & Piccinini, F. (2020). Colour deconvolution: stain unmixing
in histological imaging. Bioinformatics, 37 (10), 1485-1487.
Laperle, K.M.D. & Dintzis, S.M. Endocrine System. Comparative Anatomy and
Histology. Elsevier, 2018.
Léonhardt, M., Lesage, J., Dufourny, L. Dickès-Coopman, A., Montel, V. & Dupouy, J-
P. (2002). Perinatal Maternal Food Restriction Induces Alterations in
19
Hypothalamo-Pituitary-Adrenal Axis Activity and in Plasma Corticosterone-
Binding Globulin Capacity of Weaning Rat Pups. Neuroendocrinology, 75 (1),
45-54.
Liang, H., Zhang, J. & Zhang, Z. (2004). Food restriction in pregnant rat-like hamsters
(Cricetulus triton) affects endocrine, immune function and odor attractiveness of
male offspring. Physiology and Behavior, 82 (2-3), 453-458.
Lok, E., Scott, F.W., Mongeau, R., Nera, E.A., Malcolm, S. & Clayson, D.B. (1990).
Calorie restriction and cellular proliferation in various tissues of the female
Swiss Webster mouse. Cancer Letters, 51 (1), 67-73.
Loose, D.S., Do, Y.S., Chen, T.L. & Feldman, D. (1980). Demonstration of
Glucocorticoid Receptors in the Adrenal Cortex: evidence for a direct
dexamethasone suppressive effect on the rat adrenal gland. Endocrinology, 107,
(1), 137-146.
Mirzaei, H., Raynes, R. & Longo, V.D. (2016). The conserved role of protein restriction
in aging and disease. Current Opinion in Clinical Nutrition and Metabolic Care,
19 (1), 74-79.
Park, J.H., Glass, Z., Sayed, K., Michurina, T. V., Lazutkin, A., Mineyeva, O.,
Velmeshev, D., Ward, W. F., Richardson, A & Enikolopov, G. (2013). Calorie
restriction alleviates the age-related decrease in neural progenitor cell division in
the aging brain. European Journal of Neuroscience, 37 (12), 1987-1993.
Paust, H.J., Loeper, S., Else, T., Bamberger, A-M, Papadopoulos, G., Pankoke, D.,
Saeger, W. & Bamberger, C.M. (2006). Expression of the Glucocorticoid
Receptor in the Human Adrenal Cortex. Experimental and Clinical
Endocrinology & Diabetes, 114 (1), 6-10.
Pereira, S.S., Carvalho, L., Costa, M.M., Melo, A., Ferreira, I.M.P.L.V.O., Gomez-
Sanchez, C.E., Vinson, G. & Pignatelli, D. (2021) Mineralocorticoid Receptor
Antagonists
Eplerenone
and
Spironolactone
Modify
Adrenal
Cortex
Morphology and Physiology. Biomedicines, 9 (4), 441.
Péron, F.G., Moncloa, F., Dorfman, R.I., Duclos, R. & Duclos, T. (1960). Studies on
the possible inhibitory effect of corticosterone on corticosteroidogenesis at the
adrenal level in the rat. Endocrinology, 67 (3), 379-388, 1960.
Pohanka, D.G. & Pike, R.L. (1970). Effects of dietary sodium restriction during
pregnancy on the histochemistry of the rat zona glomerulosa. Experimental
Biology and Medicine, 133 (1), 246-251.
20
Recabarren, M.P., Valenzuela, G.J. & Seron-Ferrer, M. (1997) Protein-caloric
restriction during pregnancy affects the adrenal-placental axis and decreases
newborn weight in a primate, the Cebus apella. American Journal of Obstetrics
and Gynecology, 176 (1), 163.
Rosenbrock, H., Koros, E., Bloching, A., Podhorna, J. & Borsini, F. (2005). Effect of
chronic intermittent restraint stress on hippocampal expression of marker
proteins for synaptic plasticity and progenitor cell proliferation in rats. Brain
Research, 1040 (1-2), 55-63.
Saito, E., Mukai, M., Muraki, T., Ichikawa, Y. & Homma, M. (1979). Inhibitory Effects
of Corticosterone on Cell Proliferation and Steroidogenesis in the Mouse
Adrenal Tumor Cell Line Y-l. Endocrinology, 104 (2), 487-492.
Silverthorn, D.U. (2010). Fisiologia Humana: Uma Abordagem Integrada. 5. ed. Porto
Alegre: Artmed.
Spiga, F., Zavala, E., Walker, J. J., Zhao, Z., Terry, J. R. & Lightman, S. L. (2017).
Dynamic
responses
of
the
adrenal
steroidogenic
regulatory
network. Proceedings of the National Academy of Sciences, 114 (31), 1-9.
Terpstra, A.H.M., Harkes, L. & Van der Veen, F.H. (1981). The effect of different
proportions of casein in semipurified diets on the concentration of serum
cholesterol and the lipoprotein composition in rabbits. Lipids, 16 (2), 114-119.
Thiele, K., Diao, L. & Arck, P.C. (2017). Immunometabolism, pregnancy, and
nutrition. Seminars in Immunopathology, 40 (2), 157-174.
Valsamakis, G., Chrousos, G. & Mastorakos, G. (2019). Stress, female reproduction and
pregnancy. Psychoneuroendocrinology, 100 (1), 48-57.
Youngman, L.D. (1987). Recall, memory, persistence, and the sequential modulation of
preneoplastic lesion development by dietary protein. New York: Cornell
University Press.
Youngman, L.D. (1993). Protein restriction (PR) and caloric restriction (CR) compared:
effects on DNA damage, carcinogenesis, and oxidative damage. Mutation
Research, 295 (4-6), 165-179.
Youngman, L.D. & Campbell, T.C. (1992) The sustained development of preneoplastic
lesions depends on high protein intake. Nutrition and Cancer, 18 (2), 131-142.
Youngman, L.D., Park, J.Y. & Ames, B.N. (1992) Protein oxidation associated with
aging is reduced by dietary restriction of protein or calories. Proceedings of the
National Academy of Sciences, 89 (19), 9112-9116.
21
LEGENDS
FIGURE 1 Nutritional restriction in pregnant and non-pregnant rats. A) Body mass of pregnant
and non-pregnant animals. B) Food consumption. C) Total food consumption of the animals. D)
Total weight gain of the animals (n = 5). E) Mass of the right adrenal glands after euthanasia.
Mean ± SD (n = 6; p < 0.05). (ANOVA, post Tukey test).
FIGURE 2 Histomorphometry of the adrenal glands. A) Cortical area. B) Medullary area. C)
Total area. D) Ratio of cortical area to total adrenal area. E) Medullar area to total adrenal area
ratio. F) Ratio of adrenal mass to animal body mass (n = 5). Mean ± SD (n = 6; p < 0.05).
(ANOVA, post Tukey test).
FIGURE 3 Analyzed area of connective tissue (collagen) in the adrenal gland. Quantification of
connective tissue (collagen-in blue), in percentage, in the areas of the adrenal glands using
Mallory's Trichrome staining. Final magnification – 200x. A) Glomerular zone. B) Fasciculate
zone. C) Reticulated zone. D) Medullary zone. Mean ± SD (n = 5; p < 0.05). (ANOVA, post
Tukey test).
FIGURE 4 Immunohistochemistry for Ki-67 antigen levels (in percent). Dark brown (stronger
markings) and light brown (weaker markings) nucleus markings in their respective areas of the
adrenal gland. Final magnification – 200x. A) Glomerular zone. B) Fasciculated zone. C)
Reticular zone, D) Medullary zone. Mean ± SD (n = 5; p < 0.05). (ANOVA, post Tukey test).
FIGURE 5 Immunohistochemistry for glucocorticoid receptor (GR) (in percent). Markings in
both cytoplasm and nucleus: dark brown (stronger markings) and light brown markings (weaker
22
markings) in their respective areas of the adrenal gland. Final magnification – 200x. A)
Glomerular Zone. B) Fasciculate zone. C) Reticular Zone. D) Medullary zone. Mean ± SD (n =
5; p < 0.05). (ANOVA, post Tukey test).
FIGURE 6 Immunohistochemistry for mineralocorticoid receptor (MR) (in percent). Markings
in both cytoplasm and nucleus; dark brown (stronger markings) and light brown (weaker
markings) in their respective areas of the adrenal gland. Final magnification – 200x. A) Zona
glomerulosa. B) Fasciculated zone. C) Reticular zone. D) Medullary zone. Mean ± SD (n = 5; p
< 0.05). (ANOVA, post Tukey test).
boay mass (g)
400
400
350
300
250
200
-* mp
+= pc
+ np'
—+ pC
—* np
-~o- pP
ot eta
npC
pC
npCR
pCR
npPR
pPR
FOOa CONSUMpPUO#N (Q)
200
150
100
50
O0to7
7 to 14
14 to 21
ose ue
npC
pc
npCR
pCR
npPR
pPR
b.f
a,e
bf
ae
b,f
ae
So o fo] °o o
o oO o Oo
o
Oo
wo
6) Uol}dWINSUOD poo] je}JO]L
$88 8 ° 3
- =
(6) ure6 ssew seo;
> ec co co
> Nn =
ju) SSE] JeuUaIpY
Cortical area (um?)
2500000
2000000
1500000
1000000
500000
2*107-
1%10/.
=
x
wo
1.5*10’.
(zu) eae Areyjnpayy
2107
1.5*107
1*107
5x1
(url) ease jeJOL
¢ ® 2° 8 g
XN = = 8 8
oo eo 98 9
Bua pe [e}0}/Base [eI1}105D 013k}
& ¥ £ £ & &
=A OM
- Tr 8 ©
pe jeyoyeere Aej;npey :o1e}
om eg tN
a a
6 co 6 86 6
0.00
q Jewiuesssew jeusipy :oney
oe
(%) esojnsawo|y euoZ —
uaHejjOD) anssi} aANIeUUOD
FS SF
se
g
oe
(%) eyejnoiosey euoZ —
abejoD) anssi} sAlDaUUOD
FS SF SF
& ¢
g
%
Pee
(%) Suejndijay eUoZ —
UaHe}}OD) anssi} aAIBUUOD
Seenn
(%) B1INPaW —
1aHe}jOD) anssi} aAIaUUOD
PEE
(%) esojnsawo|y
Z — Burysew 1129 - 29-1
g
F FF
~
3
(%) eyejnd1I9se4
U0Z — Burysew 1199 - 29-1
Qe
¢
e &
£
&
(%) sueinayey
07 — BUINJRIW 1184 - /9-I1y
nprnr
pPR
ii:
B|INPAW — Buysew [129 - 29-1
ea
PEEEES
(%) Bsojnsawo|y
2U0Z — BulyJeW [12D - YD
npc npCR
pc pCR
APTN nprr
pCR pPR
g
$ HS BW
SSF
&
PEdbas
=
(%) eyejnoissey
10Z — Burysew 1199 - yD
ae = prem
pc pCR pPR
npCR npPR
pCR pPR
Pibii
(%) sueinoney
0Z — Buyew 1199 - YD
(%) eINPaW
— Buppew 1129 - yD
+ &
FF ES
&
P
npCR nprne
pCR pPR
Pid
(%) esojnsawo|y
euo0Z — Hulse [12D - YIN
npc
pc
npCR
pCR
ny
pl
PR
npCR np!
pCR . pPR
e
d
LE EH SHS
CLES F
PEGbis
(%) eyeIndIose4
U0Z — Hulysew [12D - YI
cart ete at
as a
ef
fit
(%) suejnoney
2U07Z — BulyJeW 12D - MIN
npCR
pCR
npPR
pPR
e @
¥ $
%
3 %
a] &.
0,
Dr %
vo 4,
fi
(%) elINPaW
— Burew 1199 - YIN
el
ay ¥
rs
i e.
ran
~ a
com
—
>
| 2022 | Morphofunctional evaluation of the adrenal gland in rats submitted to nutritional restriction during pregnancy | 10.1101/2022.10.16.512413 | [
"Telles Bruno dos Santos",
"Rebelato Hércules Jonas",
"Esquisatto Marcelo Augusto Marretto",
"Catisti Rosana"
] | null |
1
Phylo-geo-network and haplogroup analysis of 611 novel Coronavirus
(nCov-2019) genomes from India
Short running title: Phylogenomic network of nCov-2019 in India
Rezwanuzzaman Laskar1, Safdar Ali1*
1Clinical and Applied Genomics (CAG) Laboratory
Department of Biological Sciences, Aliah University, Kolkata, India
RL: rezwanuzzaman.laskar@gmail.com
*Corresponding author:
Dr Safdar Ali, Assistant Professor, Department of Biological Sciences,
Aliah University, IIA/27, Newtown, Kolkata 700160, India.
E-mail: safdar_mgl@live.in; ali@aliah.ac.in
Telephone No: 91-33-23416479; Fax: 91-33-29860252
2
Abstract
The novel Coronavirus from Wuhan China discovered in December 2019 (nCOV-2019) has
since developed into a global epidemic with major concerns about the possibility of the virus
evolving into something even more sinister. In the present study we constructed the phylo-geo-
network of nCOV-2019 genomes from across India to understand the viral evolution in the
country. A total of 611 genomes full length genomes were extracted from different states of
India from the EpiCov repository of GISAID initiative and NCBI. Their alignment uncovered
270 parsimony informative sites. Further, 339 genomes were divided into 51 haplogroups. The
network revealed the core haplogroup as that of reference sequence NC_045512.2 (Haplogroup
A1) with 157 identical sequences present across 16 states. The rest were having not more than
ten identical sequences across not more than three locations. Interestingly, some locations with
fewer samples have more haplogroups and most haplogroups (41) are localized exclusively to
any one state only, suggesting the local evolution of viruses. The two most common lineages
are B6 and B1 (Pangolin) whereas clade A2a (Covidex) appears to be the most predominant in
India. However, since the pandemic is still emerging, the final outcome will be clear later only.
Keywords
Coronavirus, Phylo-network, Parsimony informative sites, Haplogroups
3
Introduction
Coronaviruses belonging to the family Coronaviridae have been named so owing to the electron
microscopic structure resemblance of their virion structure to that of a crown. The spikes
present on the virion surface provide for the resemblance. Their genome has a positive single
strand RNA of 26 to 32kb in length and are known to infect a wide range of hosts (Cavanagh,
2007; Ismail et al., 2003; Lai and Cavanagh, 1997; Su et al., 2016; Weiss and Navas-Martin,
2005). A novel Coronavirus which has the potential to infect humans has been identified from
Wuhan China in December 2019. It was subsequently referred to as nCOV-2019 (novel
Coronavirus 2019) and since its emergence it has developed into a global epidemic. As of 28th
August 2020, there were 33,10,234 cases and 60,472 deaths in India due to nCOV-2019
(https://www.mygov.in/covid-19). At the same time, as per WHO there have been 24,021,218
cases and 821,462 deaths globally (https://www.worldometers.info/coronavirus/). The nCOV-
2019 is different from earlier Coronavirus outbreaks, severe acute respiratory syndrome
(SARS) coronavirus in 2002 and Middle East respiratory syndrome (MERS) coronavirus in
2012 predominantly due to its extremely high transmission rates. The patients infected with
nCOV-2019 have been observed to have varied symptoms ranging from normal flu like
symptoms to high fever to invasive lesions (Chan et al., 2020; Huang et al., 2020; Peiris et al.,
2004; Zaki et al., 2012; Zhu et al., 2020).
The nCOV-2019 belongs to genus Betacoronavirus and sub genus Sarbecovirus with suggested
origin in bats. Various theories are in discussion about how it reached humans but nothing can
be said with surety just yet (Lu et al., 2020; Zhou et al., 2020). However, the ever-increasing
number of people being infected globally provides for the most conducive environments for
the virus to evolve. The availability of full genome sequences for nCOV-2019 in GISAID has
4
aided the study of these evolving sequences with both global and local perspectives (Shu and
McCauley, 2017a).
At present, we build and analyze the phylo-geo-network of nCOV-2019 in India based on the
publicly available full-length sequences of nCOV-2019 from India. We performed the
haplogroup analysis and phylogenetic lineage study in addition to their defining mutations and
geographical distributions. This assumes significance with India rapidly moving up the ladder
in sharing the global burden of nCOV-2019 cases and is expected to continue so in near future
owing to its demographic and health care structure.
Materials and Methods
Sequence Acquisition
Genome sequences of nCOV-2019 in FASTA format was assessed from the EpiCovTM
repository (www.epicov.org) of GISAID initiative (Shu and McCauley, 2017b) and NCBI
(www.ncbi.nlm.nih.gov).
On 6th June, 2020 we retrieved 611 FASTA sequence congregations along their rational meta
data from GISAID EpiCoV server using the data filter ~ virus name: hCoV-19 - Host: Human
- Location: Asia/India – Complete – High Coverage and use the genome ID by excluding the
first part i.e. “EPI_ISL_” of GISAID accession ID. One sequence from the epicenter, Wuhan,
China (NC_045512.2) was taken as reference. Details of the geographical distribution of the
sequences and their accession numbers are provided in Figure 1 and Supplementary file 1
respectively. Location data of GISAID are used to identify the state of origin in India, and
wherein state name is unavailable, state address of the originating lab has been used.
5
Sequence Alignment
The congregations are aligned with the FFT-NS-fragment method using rapid calculation of
full-length MSA of closely related viral genomes, a light-weight algorithm of MAFFT v7 web-
server (https://mafft.cbrc.jp/alignment/software/closelyrelatedviralgenomes.html) (Katoh et
al., 2018) and keeping alignment size exactly throughout the reference sequence. The
nucleotide transformation sites of the alignment were further studied using MEGA X (Kumar
et al., 2018)
Phylogenetic Network Analysis
Aligned sequences were used to generated parsimony based TCS networks (Clement et al.,
2002) implemented in Population Analysis with Reticulate Trees (PopART v1.7) software
(Leigh and Bryant, 2015) where over 5 percent sites contain undefined states and will be
masked. A map of haplotypes was also drawn using the same software with geotags and traits
label coding.
Genome Annotation
The tool IGLSF (Alam et al., 2019) arranges the location of variable sites according to genes.
Using the software DNAPlotter (Carver et al., 2009) we used the Artemis (Carver et al., 2012)
to annotate the genome and visualized it as a circular plot.
Lineage and Subtyping Analysis
In the predefined cluster, using distinct nomenclature methods only a certain sequence belongs
to the haplogroups have been classified into different lineage and subtype. Lineages that
contribute most of the global spread have been assigned through Pangolin (Phylogenetic
Assignment of Named Global Outbreak Lineages) Web (https://pangolin.cog-uk.io/) , using
6
nomenclature implemented by Rambaut, et al. (Rambaut et al., 2020). Viral subtypes of the
studied Indian population were achieved using ‘SARS Cov 2 Nextstrain’ classification model
of Covidex (https://cacciabue.shinyapps.io/shiny2/), a web-based subtyping tool (Cacciabue et
al., 2020).
Sequence Statistics
Multiple metrics were used to assess the population genetics to decipher the phylogenetic
relationship. We calculated Tajima’s D (Tajima, 1989) statistic to test mutation- drift
equilibrium and Pi value, segregating sites, parsimony-informative sites to measure DNA
polymorphism among sequences using PopART statistics (Leigh and Bryant, 2015).
Results and Discussion
Phylogenetic network analysis
The alignment of genomes and their subsequent analysis revealed a total of 493 segregating
sites of which 270 were parsimony informative (PI) sites. The incidence of sites and their
distribution across gaps and ambiguous sequences and statistical evaluation has been
summarized in Table 1. A negative value of Tajimas D statistic suggests the significance of
these sites in evolution of these genomes. The reported phylo-geo-network herein has been
built using the 270 parsimony informative sites including the gaps and ambiguous sequences.
The phylo-geo-network analysis of the studied genomes has been represented in Figure 1.
Several observations can be drawn from this analysis and the data forming the basis of this
figure. First, the core of the network with maximum genomes (157) is the node of reference
sequence of nCOV-2019 from Wuhan, China with accession no NC_045512.2. The fact that
this accounts for over one fourth (25.7%) of the total studied sequences is a clear indication
7
that in spite of many reported variations, the original nCOV-2019 genome continues to be the
dominantly prevalent form. Interestingly, there was one sequence with genome id 458080 from
Telangana which was hundred percent identical to the Wuhan reference sequence
(Supplementary files 1 and 4). Though the absence of travel history for most of the studied
patients and the sequences only being a partial representation of the patients present makes the
conclusion subjective, but it does indicate about arrival of the virus directly from China to
India. Though the variations are fast accumulating in the virus, it's the original one that still
prevails, at least in the Indian context. Viral evolution is a dynamic and fast process but unless
due selection advantage is offered, a new form wouldn’t take over.
Secondly, the distribution of sequences from across India (Figure 1) don’t corroborate with the
incidence scenarios but are a reflection of the ground level preparations and activity in getting
the genomes sequenced. For instance, the under-representation of Maharashtra and Tamil Nadu
in the present data set in-spite of being the two most affected states. However, assuming that
the virus has an equal chance of evolving anywhere, we believe the number of sequences
analyzed are apt for giving a glimpse of the ongoing viral evolution.
Thirdly, when we analyzed the distribution of PI sites across the genome and found it to be
non-uniform in nature. We studied the distribution in the form of strike-rate of PI sites which
we define as the number of bases after which there will be another PI site. This is to say that a
region with a strike rate of 20 would mean a PI site every 20 bases and so on. Thus, a lower
strike rate will infer a higher density of the PI sites in the region (Table 2). Based on our
analysis, the Envelope and Spike protein have a PI strike rate of 45 and 115 respectively (Table
2). Before drawing any conclusions, we need to understand that a higher incidence of PI sites
doesn’t necessarily corroborate to driving the evolutionary process as their impact on protein
8
functionality needs to be ascertained first. However, it does indicate the potential genomic
regions for the same which herein appear to be Envelope and Spike protein.
Haplogroup analysis and distribution
The network tree construction was accompanied by haplogroup determination of the studied
genomes. The nodes representing haplogroups in phylo-geo-network in Figure 1 have been
named as per accession number of the sequence defining the haplogroup. The nodal haplogroup
represented by the Wuhan reference sequence NC_045512.2 has two maxima associated with
it. The number of sequences therein as represented by the diameter of the circle (157 sequences)
and total number of locations (16 states) in which the sequences are distributed. The details of
distribution of all identical sequences have been summarized in Figure 2a and Supplementary
file 2.
Of the 611 studied genomes, the 51 haplogroups account for 339 genomes. At this juncture,
we would like to note about the sequences left out of haplogroups. They belong to haplotypes
which may converge to an existing haplogroup or emerge as a new one as the pandemic
progresses. Due to the high mutation rate of viruses and with ever increasing incidence of the
diseases the virus is replicating more and more and new polymorphisms are being generated
every day. These variations are changing the haplotype and haplogroup profile on a regular
basis.
We propose the nomenclature of the 51 observed haplogroups as per the path used to construct
the network. We will explain the haplogroup nomenclature by taking a couple of examples.
The haplogroup having NC_045512.2 was named A1 as the core of the network. From this
cluster many haplogroups emerged and so on. The haplogroup A1.1 (420544) is defined by
9
five positions; 241 (C→T), 3037 (C→T), 4809 (C→T), 14408 (C→T) and 23403 (A→G).
However, as we move to haplogroup A1.1.1 (420543), in addition to the above mutations,
another one at position 8782 (C→T) is present which becomes the defining polymorphism for
this haplogroup. Similarly, haplogroup A1.6 (435063) is defined by positions 241(C→T), 1059
(C→T), 3037 (C→T), 14408 (C→T), 23403 (A→G) and 25563 (G→T). Subsequently
haplogroup A1.6.1 (444471) is characterized by mutation at positions 18877 (C→T) and 26735
(C→T) and haplogroup A1.6.1.1 by additional mutations at 22444 (C→T) and 28854 (C→T).
The haplogroup lineage thus defined clearly indicates that A1 is the most prevalent one while
A1.6 is the most evolving one as it has the maximum number of steps going up to A1.6.1.1.1.4
reflective of five steps and stages of mutations/PI sites. The position of all the observed PI sites
has been listed in Table 2/Figure 2b and their details are summarized in Supplementary file 3.
The haplogroup nomenclature has been listed in correlation with their genome IDs and location
in Table 3. If we observe the PI sites reported in the study, it includes most of the commonly
reported sites from across the world besides some novel ones. However, we aren’t emphasizing
on the novelty of sites due to the fast-changing scenario and rapidly emerging data.
The geographical distribution of the haplogroups can be looked at from two different aspects.
To begin with, which haplogroup is found in which location. Herein, A1 (NC_045512.2)
haplogroup as already mentioned was most widely prevalent with 157 sequences distributed
across 16 locations. All other haplogroups had ten or fewer genomes spread across one to three
locations (Figure 2a). The scenario is more interesting if we inverse the analysis as in which
location had how many haplogroups. Gujarat with a maximal representation of 199 genomes
had 27 different haplogroups but this isn’t the norm as in more sequences would mean more
haplogroups. Delhi (63 genomes, 3 haplogroups), Maharashtra (94 genomes, 9 haplogroups)
and West Bengal (40 genomes,7 haplogroups) exhibit the non-linearity of the same. Also, 41
10
haplogroups have a single location only led by Gujarat (21), Maharashtra (6), West Bengal,
Telangana, Tamil Nadu (4 each) and Ladakh, Orissa (1 each). Three states Punjab, Andhra
Pradesh and Kerala don’t have any haplogroup so far. The distribution of haplogroups across
states has been shown in Figure 1 and Supplementary file 2. The fact that some locations with
fewer samples have more haplogroups and most haplogroups are localized exclusively to a
single state is a clear indication about the local evolution of viruses. However, since the
pandemic is still emerging, the final outcome will be clear only at a later stage.
Lineage and Subtype Analysis
We also ascertained the lineage and subtype of the observed sequences through Pangolin and
Covidex respectively. Also, the presence of lineages in India across the world was studied. The
fact that phylogenetic lineage of nCOV-2019 genomes from India exhibits its relation with
diverse countries like USA, Australia, UK, Singapore, China and Turkey is reflective of the
global nature of the pandemic. Most of it can be attributed to international air travel and diverse
regulations across countries. The three most common lineages in India as predicted by Pangolin
are B6, B1 and B1.36 whereas clade A2a appears to be the most predominant one as predicted
by Covidex (Figure 2c, Table 3, Supplementary file 4). These lineages can shift with increasing
incidences and accumulating variations which requires regular monitoring. However, proper
recording of both national and international travel history for all the patients will go a long way
in unveiling the true path of viral evolution.
Conclusions
The understanding of emergence and evolution of nCOV-2019 pandemic in India is an apt set
up to understand viral divergence and evolution due to its huge population and diversity. As of
now, the virus most prevalent in India is of the same haplogroup as the nCOV-2019 reference
11
sequence from Wuhan indicating absence of any significant novel emerging strain. The two
most common lineages are B6 and B1 whereas clade A2a appears to be the most predominant
one in Indian context. However, with ever increasing incidence the situation needs to be
monitored regularly.
Authors' contributions
RL performed the multiple sequence alignment and phylogenomic tree evaluation. SA
supervised the whole study and prepared the manuscript.
Acknowledgements
The authors thank the Department of Biological Sciences, Aliah University, Kolkata, India for
all the financial and infrastructural support provided. Authors acknowledge all the authors
associated with originating and submitting laboratories of the sequences from GISAID’s
EpiFlu™ (www.gisaid.org) Database on which this research is based.
Competing Interests
The authors declare they have no competing interests.
Ethics approval
Not Applicable.
References:
Alam, C.M., Iqbal, A., Sharma, A., Schulman, A.H., Ali, S., 2019. Microsatellite Diversity,
Complexity, and Host Range of Mycobacteriophage Genomes of the Siphoviridae
Family. Frontiers in Genetics 10, 207. https://doi.org/10.3389/fgene.2019.00207
12
Cacciabue, M., Aguilera, P., Gismondi, M.I., Taboga, O., 2020. Covidex: an ultrafast and
accurate
tool
for
virus
subtyping.
bioRxiv
2020.08.21.261347.
https://doi.org/10.1101/2020.08.21.261347
Carver, T., Harris, S.R., Berriman, M., Parkhill, J., McQuillan, J.A., 2012. Artemis: An
integrated platform for visualization and analysis of high-throughput sequence-based
experimental
data.
Bioinformatics
28,
464–469.
https://doi.org/10.1093/bioinformatics/btr703
Carver, T., Thomson, N., Bleasby, A., Berriman, M., Parkhill, J., 2009. DNAPlotter: Circular
and linear interactive genome visualization. Bioinformatics 25, 119–120.
https://doi.org/10.1093/bioinformatics/btn578
Cavanagh, D., 2007. Coronavirus avian infectious bronchitis virus. Veterinary Research 38,
281–297. https://doi.org/10.1051/vetres:2006055
Chan, J.F.W., Yuan, S., Kok, K.H., To, K.K.W., Chu, H., Yang, J., Xing, F., Liu, J., Yip,
C.C.Y., Poon, R.W.S., Tsoi, H.W., Lo, S.K.F., Chan, K.H., Poon, V.K.M., Chan, W.M.,
Ip, J.D., Cai, J.P., Cheng, V.C.C., Chen, H., Hui, C.K.M., Yuen, K.Y., 2020. A familial
cluster of pneumonia associated with the 2019 novel coronavirus indicating person-to-
person transmission: a study of a family cluster. The Lancet 395, 514–523.
https://doi.org/10.1016/S0140-6736(20)30154-9
Clement, M., Snell, Q., Walker, P., Posada, D., Crandall, K., IEEE, 2002. TCS: Estimating
Gene Genealogies, International parallel and distributed processing symposium. IEEE
Computer Society.
Huang, C., Wang, Y., Li, X., Ren, L., Zhao, J., Hu, Y., Zhang, L., Fan, G., Xu, J., Gu, X.,
Cheng, Z., Yu, T., Xia, J., Wei, Y., Wu, W., Xie, X., Yin, W., Li, H., Liu, M., Xiao, Y.,
Gao, H., Guo, L., Xie, J., Wang, G., Jiang, R., Gao, Z., Jin, Q., Wang, J., Cao, B., 2020.
Clinical features of patients infected with 2019 novel coronavirus in Wuhan, China.
The Lancet 395, 497–506. https://doi.org/10.1016/S0140-6736(20)30183-5
Ismail, M.M., Tang, Y., Saif, Y.M., 2003. Pathogenicity of Turkey Coronavirus in Turkeys and
Chickens. Avian Diseases 47, 515–522. https://doi.org/10.1637/5917
Katoh, K., Rozewicki, J., Yamada, K.D., 2018. MAFFT online service: Multiple sequence
alignment, interactive sequence choice and visualization. Briefings in Bioinformatics
20, 1160–1166. https://doi.org/10.1093/bib/bbx108
Kumar, S., Stecher, G., Li, M., Knyaz, C., Tamura, K., 2018. MEGA X: Molecular
evolutionary genetics analysis across computing platforms. Molecular Biology and
Evolution 35, 1547–1549. https://doi.org/10.1093/molbev/msy096
13
Lai, M.M., Cavanagh, D., 1997. The molecular biology of coronaviruses., Advances in virus
research. Academic Press. https://doi.org/10.1016/s0065-3527(08)60286-9
Leigh, J.W., Bryant, D., 2015. POPART: Full-feature software for haplotype network
construction.
Methods
in
Ecology
and
Evolution
6,
1110–1116.
https://doi.org/10.1111/2041-210X.12410
Lu, R., Zhao, X., Li, J., Niu, P., Yang, B., Wu, H., Wang, W., Song, H., Huang, B., Zhu, N.,
Bi, Y., Ma, X., Zhan, F., Wang, L., Hu, T., Zhou, H., Hu, Z., Zhou, W., Zhao, L., Chen,
J., Meng, Y., Wang, J., Lin, Y., Yuan, J., Xie, Z., Ma, J., Liu, W.J., Wang, D., Xu, W.,
Holmes, E.C., Gao, G.F., Wu, G., Chen, W., Shi, W., Tan, W., 2020. Genomic
characterisation and epidemiology of 2019 novel coronavirus: implications for virus
origins and receptor binding. The Lancet 395, 565–574. https://doi.org/10.1016/S0140-
6736(20)30251-8
Peiris, J.S.M., Guan, Y., Yuen, K.Y., 2004. Severe acute respiratory syndrome, Nature
Medicine. Nature Publishing Group. https://doi.org/10.1038/nm1143
Rambaut, A., Holmes, E., Hill, V., O’Toole, Á., McCrone, J., Ruis, C., du Plessis, L., Pybus,
O., 2020. A dynamic nomenclature proposal for SARS-CoV-2 to assist genomic
epidemiology.
bioRxiv
2020.04.17.046086.
https://doi.org/10.1101/2020.04.17.046086
Shu, Y., McCauley, J., 2017. GISAID: Global initiative on sharing all influenza data – from
vision to reality. Eurosurveillance 22, 30494. https://doi.org/10.2807/1560-
7917.ES.2017.22.13.30494
Su, S., Wong, G., Shi, W., Liu, J., Lai, A.C.K., Zhou, J., Liu, W., Bi, Y., Gao, G.F., 2016.
Epidemiology, Genetic Recombination, and Pathogenesis of Coronaviruses, Trends in
Microbiology. Elsevier Ltd. https://doi.org/10.1016/j.tim.2016.03.003
Tajima, F., 1989. Statistical method for testing the neutral mutation hypothesis by DNA
polymorphism. Genetics 123.
Weiss, S.R., Navas-Martin, S., 2005. Coronavirus Pathogenesis and the Emerging Pathogen
Severe Acute Respiratory Syndrome Coronavirus. Microbiology and Molecular
Biology Reviews 69, 635–664. https://doi.org/10.1128/mmbr.69.4.635-664.2005
Zaki, A.M., van Boheemen, S., Bestebroer, T.M., Osterhaus, A.D.M.E., Fouchier, R.A.M.,
2012. Isolation of a Novel Coronavirus from a Man with Pneumonia in Saudi Arabia.
New
England
Journal
of
Medicine
367,
1814–1820.
https://doi.org/10.1056/NEJMoa1211721
14
Zhou, P., Yang, X.L., Wang, X.G., Hu, B., Zhang, L., Zhang, W., Si, H.R., Zhu, Y., Li, B.,
Huang, C.L., Chen, H.D., Chen, J., Luo, Y., Guo, H., Jiang, R.D., Liu, M.Q., Chen, Y.,
Shen, X.R., Wang, X., Zheng, X.S., Zhao, K., Chen, Q.J., Deng, F., Liu, L.L., Yan, B.,
Zhan, F.X., Wang, Y.Y., Xiao, G.F., Shi, Z.L., 2020. A pneumonia outbreak associated
with a new coronavirus of probable bat origin. Nature 579, 270–273.
https://doi.org/10.1038/s41586-020-2012-7
Zhu, N., Zhang, D., Wang, W., Li, X., Yang, B., Song, J., Zhao, X., Huang, B., Shi, W., Lu,
R., Niu, P., Zhan, F., Ma, X., Wang, D., Xu, W., Wu, G., Gao, G.F., Tan, W., 2020. A
novel coronavirus from patients with pneumonia in China, 2019. New England Journal
of Medicine 382, 727–733. https://doi.org/10.1056/NEJMoa2001017
15
Figure Legends
Figure 1: Phylogenomic geographic network of nCOV-2019 genomes from India. The
nodes have been named after the Accession No of the defining sequences representing a
particular cluster. The diameter of the circle represents the number of samples present therein.
more the samples, longer the diameter. The different locations within India have been
represented by color coding and the number of sequences from each state are shown in the
bottom scale of the graph. Also shown are the distribution of haplogroups across different states
in the maps on the periphery. On the right side are haplogroups present only in one state only
whereas others include those present on multiple locations. Maps are generated and powered
by Bing (©Geo Names, Microsoft, TomTom) through MS Excel 2019.
Figure 2: a) Prevalence and geographical distribution of 51 haplogroups of nCOV-2019
genomes in India. The number of identical sequences present in a haplogroup are shown on
primary vertical axis whereas number of locations wherein its distribution is shown on
secondary vertical axes. Note the maximum prevalence and widespread distribution of
NC_045512.2 containing haplogroup (A1). For details of haplogroup IDs, identical sequences
and locations please refer to Supplementary file 2. b) Distribution of parsimony informative
sites across the nCOV-2019 genome. The nCOV-2019 genome has been represented
circularly along with the locations of different genes/ORFs/Non coding regions have been
represented. PI sites are shown as lines traversing the circle. 2c) Lineage and Subtype
Analysis of nCOV-2019 genomes in India. The outermost circle represents haplogroups
reported in the study whereas the middle circle depicts lineage prediction by Pangolin web.
The innermost circle is the clade analysis by Covidex web-tool.
16
Details of Supplementary Files
Supplementary File 1: Details of nCovid 2019 genomes used in the study
Supplementary file 2: Details of identical sequences in the study and their geographical
distribution
Supplementary files 3: Details of parsimony informative sites including gaps and ambiguous
sequences observed in the study
Supplementary file 4: Details of lineage analysis of studied genomes
17
Table 1: Some key statistical parameters observed in the study
S
No
Network
Type
Number of
segregating
sites
Number of parsimony-
informative sites
Nucleotide
diversity
Tajima's D
statistic
Excluding* Including*
1
TCS
493
152
270
pi =
0.00120683
D = -1.82662
p (D >= -1.82662)
= 0.982906
* Gaps and Ambiguous/Missing (Details in Supplementary file 3)
Table 2: Distribution of parsimony informative sites across the genome of nCOV-2019
1
2
S No Genome
Region
Start
position
End
position
Size
(bp)
No of
parsimony sites
Strike-rate of
parsimony sites*
Position of PI Sites
1
5'UTR
1
265
265
9
29.4
22, 55, 56, 94, 106, 218, 219, 222, 241
2
ORF1a
266
13483
1321
8
100
132.2
506,
635, 771, 875, 884, 1059, 1094, 1191, 1218, 1281,
1397, 1589, 1599, 1707, 1820, 1846, 2143, 2368, 2480,
2558, 2632, 2836, 3037, 3039, 3054, 3085, 3426, 3472,
3634, 3686, 3737, 3817, 4067, 4084, 4255, 4354, 4372,
4444, 4679, 4809, 4866, 4893, 4965, 5029, 5062, 5139,
5572, 5700, 5826, 6081, 6310, 6312, 6402, 6466, 6541,
6573, 6616, 6868, 6989, 7319, 7392, 7600, 7945, 8022,
8026, 8080, 8296, 8460, 8653, 8782, 8917, 8950, 9389,
9438, 9628, 9693, 10138, 10277, 10369, 10478, 10479,
10679, 10702, 10771, 10815, 11074, 11083, 11200, 11306,
11335, 11457, 11572, 11620, 12076, 12439, 12616, 12685,
12757, 13458
3
ORF1ab
13468
21555
8088
58
139.4
13585, 13617, 13730, 13859, 14130, 14181, 14274,
14408, 14425, 14673, 14805, 15324, 15435, 15451, 15708,
16017, 16078, 16355, 16393, 16626, 16738, 16852, 16887,
16993, 17135, 17440, 17722, 17747, 17858, 17959, 18052,
18129, 18380, 18395, 18457, 18486, 18511, 18877, 19086,
19185, 19344, 19417, 19524, 19679, 19684, 19816, 19872,
19983, 20006, 20063, 20087, 20151, 20355, 20773, 21004,
21137, 21550, 21551
4
S protein
21563
25384
3822
33
115.8
21575, 21627, 21628, 21646, 21724, 21792, 21795, 21890,
22289, 22343, 22374, 22444, 22468, 22530, 22663, 23120,
23236, 23277, 23111, 23403, 23593, 23638, 23678, 23815,
23821, 23929, 24811, 24933, 25098, 25290, 25314, 25381
5
ORF3a
25393
26220
828
10
82.8
25445, 25461, 25513, 25528, 25563, 25596, 25613, 25855,
25904, 26144
6
NC
26221
26244
24
1
24
26226
7
E
26245
26472
228
5
45.6
26330, 26338, 26375, 26376, 26467
8
M
26523
27191
669
6
111.5
26530, 26681, 26730, 26735, 27110, 27191
9
ORF6
27202
27387
186
5
37.2
27213, 27379, 27382, 27383, 27384
10
ORF7a
27394
27759
366
1
366
27613
11
ORF7b
27756
27887
132
1
132
27874
12
NC
27888
27893
6
1
6
27889
13
ORF8
27894
28259
366
7
52.3
28001, 28077, 28083, 28114, 28221, 28253, 28254
14
N
28274
29533
1260
20
63
28289, 28311, 28312, 28326, 28371, 28396, 28688, 28795,
28854, 28878, 28881, 28882, 28883, 28948, 29039, 29188,
29197, 29236, 29451, 29474
15
NC
29534
29557
24
3
8
29543, 29555, 29557
16
ORF10
29558
29674
117
0
17
3'UTR
29675
29903
229
10
22.9
29722, 29734, 29742, 29743, 29774, 29827, 29829, 29830,
29870, 29874
Total
270
*Calculated by (Size/No of parsimony sites in the region)
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Table 3: Details of Haplogroups: Geographical distribution and Phylogenetic lineage
20
21
Haplogroup
Node
Label/Genome
ID
State
Most Common Countries
Lineage analysis
(Rambaut et al., 2020)
Subtype analysis-SARS
Cov 2 Nextstrain
(Hadfield et al., 2018)
Proposed
Assigned by
(GISAID/NCBI)
Assigned by Database
(GISAID/NCBI)
Assigned by Pangolin Web-server
Prediction by Pangolin
Web-server
Prediction by Covidex
Web-server
A1
NC_045512.2
1.
Assam
2.
Bihar
3.
Delhi
4.
Gujarat
5.
Haryana
6.
Jammu
7.
Karnataka
8.
Madhya Pradesh
9.
Maharashtra
10. Odisha
11. Rajasthan
12. Tamil Nadu
13. Telangana
14. Uttar Pradesh
15. West Bengal
16. Wuhan, China
1. Australia, Singapore, USA
2. India, Singapore, Australia
3. UK, Australia, USA
4. UK, China, USA
5. UK, Spain, Australia
6. UK, USA, Australia
7. UK, USA, China
1. B
2. B.1
3. B.1.1
4. B.1.5
5. B.6
1. A1a
2. A2
3. A2a
4. A3
5. A6
6. A7
A1.1
420544
Maharashtra
UK, USA, Australia
B.1
A2a
A1.1.1
420543
Maharashtra
UK, USA, Australia
B.1
A2a
A1.10
444479
Gujarat
UK, USA, Australia
B.1
A2a
A1.11
444483
Gujarat
UK, USA, Australia
B.1
A2a
A1.12
447584
Tamil Nadu
India, Singapore, Australia
B.6
A3
A1.13
451158
Gujarat
UK, USA, Australia
B.1
A2a
A1.14
452192
1.
Gujarat
2.
Maharashtra
UK, Australia, USA
UK, USA, Australia
1. B.1
2. B.1.1
A2a
A1.14.1
450785
Gujarat
UK, USA, Australia
B.1
A2a
A1.14.2
458059
Telangana
UK, Australia, USA
B.1.1
A2a
A1.15
452213
Maharashtra
Australia, UK, Turkey
B.4
A3
A1.16
452214
1.
Gujarat
2.
Maharashtra
3.
Telangana
UK, USA, Australia
B.1
A2a
A1.17
455660
West Bengal
UK, USA, Australia
B.1
A2a
A1.18
458063
Telangana
India, Singapore, Australia
B.6
A7
A1.19
461490
Gujarat
UK, USA, Australia
B.1
A2a
A1.2
424364
Maharashtra
UK, USA, Australia
B.1
A2a
A1.20
455667
West Bengal
UK, USA, Australia
B.1
A2a
A1.21
458046
Telangana
UK, Australia, Gambia
B.1.1.8
A2a
A1.22
458064
Telangana
UK, USA, Australia
B.1
A2a
A1.23
435101
Ladakh
Australia, UK, Turkey
B.4
A3
A1.24
437442
Gujarat
Australia, Singapore, USA
B.6
A1a
A1.25
447858
Telangana
India, Singapore, Australia
B.6
A3
A1.26
450790
Gujarat
China, South_Korea, USA
A
B4
A1.27
451154
1.
Gujarat
2.
Madhya Pradesh
Australia, Singapore, USA
India, Singapore, Australia
B.6
1. A3
2. A7
A1.28
452204
Maharashtra
China, South_Korea, USA
A
B4
A1.29
452205
Maharashtra
China, South_Korea, USA
A
B4
A1.3
430464
West Bengal
UK, Australia, USA
B.1.1
A2a
A1.30
455653
West Bengal
UK, USA, Australia
B.1
A2a
A1.31
455764
Odisha
China, South_Korea, USA
A
B4
A1.4
430465
1. Tamil Nadu
2. West Bengal
UK, Australia, USA
B.1.1
A2a
A1.4.1
458031
Tamil Nadu
UK, Australia, USA
B.1.1
A2a
A1.4.2
458037
Tamil Nadu
UK, Australia, USA
B.1.1
A2a
A1.4.3
458038
Tamil Nadu
UK, Australia, USA
B.1.1
A2a
A1.5
435056
Gujarat
UK, USA, Australia
B.1
A2a
A1.6
435063
1. Delhi
2. Telangana
UK, USA, Australia
B.1
A2a
A1.6.1
444471
1. Gujarat
2. Odisha
Saudi_Arabia, UK, Turkey
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.1
435065
1. Delhi
2. Gujarat
Saudi_Arabia, UK, Turkey
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.1.1
444461
Gujarat
Saudi_Arabia, UK, Turkey
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.1.1.1
435055
Gujarat
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.1.1.2
444465
Gujarat
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.1.1.3
447033
Gujarat
Saudi_Arabia, UK, Turkey
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.1.1.4
451149
Gujarat
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.1.2
444469
Gujarat
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.2
444456
Gujarat
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.3
444484
Gujarat
Turkey, Finland, UK
B.1.36
A2a
A1.6.1.4
455021
Gujarat
Saudi_Arabia, UK, Turkey
B.1.36
A2a
A1.6.1.5
437449
Gujarat
Turkey, Finland, UK
B.1.36
1. A2
2. A2a
A1.7
436414
1. Assam
2. West Bengal
India, Singapore, Australia
B.6
A1a
A1.8
436426
1. Bihar
2. Delhi
India, Singapore, Australia
B.6
1. A3
2. A7
A1.9
437447
Gujarat
UK, USA, Australia
B.1
A2a
A1.9.1
447549
Gujarat
UK, USA, Australia
B.1
A2a
22
23
24
25
26
27
28
e e * go 38 36 3 “yt
¥ > “a & nen z
& $ eueAiey
et
ysepesg senn
eS
vne 9
Yysopeid “ee ST
pen jwey
eeeelt
je6ueg 3s9m
eeee Ot
euebueja,
eeeeeeegey L6
yeae(n5
eereevererereeereers 661
eqyseseyeW
eeeveeeges 16
140
eegegegc9
eysipo
eee OF
eyeyewey
we OT
“~~ 9
eye
et
_ Z
4ywysey pue nuwer
el
ysepeig esypuy
gl
a
e6u0g som
usepeidsenn |)
eueGueje, e@
npen wer @
vewsefer @
qefung e
eysipo ()
enysereuen @
usopeid edypen @
wierer @
eje29y e
eyerewey |)
snuysey pue nuwec @
eueAseH e@
yang @
wea @
seus ©)
usoperg exypuy @
| 2020 | Phylo-geo-network and haplogroup analysis of 611 novel Coronavirus (nCov-2019) genomes from India | 10.1101/2020.09.03.281774 | [
"Laskar Rezwanuzzaman",
"Ali Safdar"
] | creative-commons |
Quantitative profiling of native RNA modifications and their
dynamics using nanopore sequencing
Oguzhan Begik1,2,3,#, Morghan C Lucas1,4,#, Leszek P Pryszcz1,5, Jose Miguel
Ramirez1, Rebeca Medina1, Ivan Milenkovic1,4, Sonia Cruciani1,4, Huanle Liu1,
Helaine Graziele Santos Vieira1, Aldema Sas-Chen6, John S Mattick3, Schraga
Schwartz6 and Eva Maria Novoa1,2,3,4,7*
1Centre for Genomic Regulation (CRG), The Barcelona Institute of Science and Technology, Dr.
Aiguader 88, Barcelona 08003, Spain
2Garvan Institute of Medical Research, Darlinghurst, NSW, 2010, Australia
3UNSW Sydney, Darlinghurst, NSW, 2052, Australia
4Universitat Pompeu Fabra (UPF), Barcelona, Spain
5International Institute of Molecular and Cell Biology, 4 Ks. Trojdena Street, 02-109 Warsaw, Poland
6 Weizmann Institute of Science, Rehovot, IL
7 Lead Contact
# These authors contributed equally
* Correspondence to: Eva Maria Novoa (eva.novoa@crg.eu)
ABSTRACT
A broad diversity of modifications decorate RNA molecules. Originally conceived as static
components, evidence is accumulating that some RNA modifications may be dynamic, contributing to
cellular responses to external signals and environmental circumstances. A major difficulty in studying
these modifications, however, is the need of tailored protocols to map each modification type
individually. Here, we present a new approach that uses direct RNA nanopore sequencing to identify
and quantify RNA modifications present in native RNA molecules. First, we show that each RNA
modification type results in a distinct and characteristic base-calling ‘error’ signature, which we
validate using a battery of genetic strains lacking either pseudouridine (Y) or 2’-O-methylation (Nm)
modifications. We then demonstrate the value of these signatures for de novo prediction of Y
modifications transcriptome-wide, confirming known Y-modified sites as well as uncovering novel Y
sites in mRNAs, ncRNAs and rRNAs, including a previously unreported Pus4-dependent Y
modification in yeast mitochondrial rRNA, which we validate using orthogonal methods. To explore the
dynamics of pseudouridylation across environmental stresses, we treat the cells with oxidative, cold
and heat stresses, finding that yeast ribosomal rRNA modifications do not change upon environmental
exposures, contrary to the general belief. By contrast, our method reveals many novel heat-sensitive
Y-modified sites in snRNAs, snoRNAs and mRNAs, in addition to recovering previously reported sites.
Finally, we develop a novel software, nanoRMS, which we show can estimate per-site modification
stoichiometries from individual RNA molecules by identifying the reads with altered current intensity
and trace profiles, and quantify the RNA modification stoichiometry changes between two conditions.
Our work demonstrates that Y RNA modifications can be predicted de novo and in a quantitative
manner using native RNA nanopore sequencing.
Keywords : Ribosomal RNA, Non-coding RNA, Messenger RNA, Epitranscriptome, RNA
Modifications, Pseudouridylation, Nanopore, Direct RNA Sequencing, Saccharomyces cerevisiae,
Machine Learning
INTRODUCTION
RNA modifications are chemical moieties that decorate RNA molecules, expanding their lexicon. By
coupling antibody immunoprecipitation or chemical probing with next-generation sequencing (NGS),
transcriptome-wide maps of several RNA modifications have been constructed, including N6-
methyladenosine (m6A) 1,2, pseudouridine (Y) 3–6, 5-methylcytosine (m5C) 7,8, 5-hydroxymethylcytosine
(hm5C) 9, 1-methyladenosine (m1A) 10,11, N3-methylcytosine (m3C) 12, N4-acetylcytosine (ac4C) 13,14
and 7-methylguanosine (m7G) 15,16. These studies have revealed that RNA modifications play a pivotal
role in a large variety of cellular processes, including regulation of cellular fate 17, sex determination 18
and cellular differentiation 19, among others.
Despite these advances, a fundamental challenge in the field is the lack of a generic approach for
mapping diverse RNA modification types simultaneously 20–23. Currently, customized protocols must
be individually set up and optimized for each RNA modification type, leading to experimental designs
in which the RNA modification type to be studied is chosen beforehand, hindering the ability to
characterize the plasticity of the epitranscriptome in a systematic and unbiased manner in response to
different conditions. Moreover, even in those cases where a selective antibody or chemical is
available, NGS-based methods are often not quantitative (i.e. cannot solve the ‘stoichiometry’
problem), have high false positive rates 21, are inconsistent when using distinct antibodies 24, are
unable to produce maps for highly repetitive regions, cannot provide information regarding the co-
occurrence of distant modifications in same transcripts, do not provide isoform-specific information,
and require multiple ligations steps and extensive PCR amplification during the library preparation,
introducing undesired biases in the sequencing data 25.
A promising alternative to NGS-based technologies that can, in principle, overcome these limitations
is the direct RNA sequencing platform developed by Oxford Nanopore Technologies (ONT), which
has the potential to detect virtually any given RNA modification present in native RNA molecules
20,26,27. Algorithms to detect RNA modifications have been made available in the last few months 28–30,
many of which rely on the use of systematic base-calling ‘errors’ caused by the presence of RNA
modifications. However, to date the vast majority of efforts have been devoted to the detection of m6A
modifications 29–33, and it is largely unknown whether other modifications of RNA bases may be
distinguishable from their unmodified counterparts using this technology. Thus, a systematic,
multiplexed and unbiased approach that can map and quantify diverse RNA modifications
simultaneously in full-length molecules is currently missing.
Here, we examine the S. cerevisiae coding and non-coding transcriptome at single molecule
resolution using native RNA nanopore sequencing. We find that most RNA modifications are
characterized by systematic base-calling errors, and that the signature of these base-calling ‘errors’
can be used to identify the underlying RNA modification type. For example, we find that pseudouridine
typically appears in the form of U-to-C mismatches, whereas m5C modifications appear in the form of
insertions. We then exploit the identified signatures to de novo predict RNA modifications in rRNAs,
finding two previously unreported Y modifications in mitochondrial rRNA, which we confirm using
CMC-probing coupled to nanopore sequencing (nanoCMC-seq). We demonstrate that one of these
novel Y modifications (15s:Y854) is placed by the enzyme Pus4, which was previously thought to
pseudouridylate only mRNAs and tRNAs 4. Moreover, we show that once the Y RNA modifications
have been accurately predicted using base-calling ‘errors’, the stoichiometry of a given Y- or Nm-
modified site can be estimated by clustering per-read features (current intensities and trace) of the
modified regions.
We then explore the dynamics of RNA modifications present in non-coding RNAs. It has been
proposed that differential rRNA modifications may constitute a source of ribosomal heterogeneity 34–36,
leading to fine tuning of the ribosomal function and ultimately proteome output. Indeed, previous
studies have shown that temperature changes affect rRNA pseudouridylation levels at specific sites,
suggesting that cells may be able to generate compositionally distinct ribosomes in response to
environmental cues 4,37,38. Similarly, alterations in the stoichiometry of 2’-O-methylation (Am, Cm, Gm,
Um) 39–41 and pseudouridylation (Y) 34–36 have been shown to affect translation initiation of mRNAs
containing internal ribosome entry sites (IRES) 42,43. Here we re-examine this question using direct
RNA sequencing, and characterize the RNA modification dynamics in rRNAs, snRNAs and snoRNAs
upon a battery of environmental cues, translational repertoires and genetic strains. Contrary to
expectations, we find that none of the environmental stresses tested lead to significant changes in the
ribosomal epitranscriptome. By contrast, our method does recapitulate previously reported heat-
dependent Y snRNA modifications, as well as identifies novel heat-sensitive sites in snRNAs and
snoRNAs.
Finally, we develop a novel algorithm, nanoRMS, which we demonstrate can predict Y RNA
modifications de novo, as well as estimate the stoichiometry of modification both in highly-modified
and lowly-modified Y and Nm sites, and illustrate its applicability in vivo across diverse types of RNA
molecules, including rRNAs, sn/snoRNAs and mRNAs. To this end, we first systematically examine
how the choice of distinct per-read features (signal intensity, dwell time and trace) affects our ability to
accurately predict RNA modification stoichiometry from individual read information. Secondly, we
benchmark how the machine learning algorithm choice affects the performance of the predictions.
Thirdly, we assess the robustness of the distinct algorithms and feature combinations upon diverse
ranges of RNA modification stoichiometries. Fourthly, we demonstrate its applicability in vivo, by
showing that it can be applied to highly-modified non-coding RNA molecules such as rRNAs, snRNAs
and snoRNAs, as well as to lowly-modified mRNA molecules. Our approach recapitulates known
Pus1-dependent, Pus4-dependent and heat stress-dependent mRNA sites, as well as reveals novel Y
mRNA sites that had not been previously reported. Altogether, our work establishes a framework for
the study of RNA modification dynamics using direct RNA sequencing, opening novel avenues to
study the plasticity of the epitranscriptome at single molecule resolution.
RESULTS
Detection of RNA modifications in direct RNA sequencing data is strongly dependent on base-
calling and mapping algorithms
Previous studies have shown that N6-methyladenosine (m6A) RNA modifications can be detected in
the form of non-random base-calling ‘errors’ in direct RNA sequencing datasets 29–33. However, it is
unclear how these ‘errors’ may vary with the choice of base-calling and mapping algorithms, and
consequently, affect the ability to detect and identify RNA modifications. To systematically determine
the accuracy of commonly used algorithms for direct RNA base-calling, as well as to assess their
ability to detect RNA modifications in the form of base-calling ‘errors’ 29, we compared their
performance on in vitro transcribed RNA sequences which contained all possible combinations of 5-
mers, referred to as ‘curlcakes’ (CCs) 29, that included: (i) unmodified nucleosides (UNM), (ii) N6-
methyladenosine (m6A), (iii) pseudouridine (Y), (iv) N5-methylcytosine (m5C), and (v) N5-
hydroxymethylcytosine (hm5C) (Figure 1A). In addition, a sixth dataset containing unmodified short
RNAs (UNM-S), with median length of 200 nucleotides, was included in the analysis to assess the
effect of input sequence length in base-calling (see Methods). Each dataset was base-called with two
distinct algorithms (Albacore and Guppy), and using two different versions for each of them, namely:
(i) Albacore version 2.1.7 (AL 2.1.7); (ii) its latest version, Albacore 2.3.4 (AL 2.3.4); (iii) Guppy 2.3.1
(GU 2.3.1); and (iv) a more recent version of the latter base-caller, Guppy 3.0.3 (GU 3.0.3), which
employs a flip-flop algorithm. We found that the latest version of Albacore (2.3.4) base-called 100% of
sequenced reads in all 6 datasets, whereas its previous version did not (average of 90.8%) (Figure
1B). In contrast, both versions of Guppy (2.3.1 and 3.0.3) produced similar results in terms of
percentage of base-called reads (98.71% and 98.75%, respectively) (Table S1).
We then assessed whether the choice of mapper might affect the ability to detect RNA modifications.
To this end, we employed two commonly used long-read mappers, minimap2 44 and GraphMap 45,
using either ‘default’ or ‘sensitive’ parameter settings (see Methods). Strikingly, we found that the
choice of mapper, as well as the parameters used, severely affected the final number of mapped
reads for each dataset (Figure 1C, see also Table S1). The most extreme case was observed with
the Y-modified dataset, where minimap2 was unable to map the majority of the reads (0-0.3%
mapped reads) (Figure 1C,D, see also Figure S1A). By contrast, GraphMap ‘sensitive’ was able to
map 35.5% of Y-modified base-called reads, proving to be a more appropriate choice for highly
modified datasets. To ascertain whether an increase in the number of base-called and mapped reads
was at the expense of decreased accuracy, we assessed the sequence identity percentage (as a
read-out of accuracy), finding that GraphMap outperforms minimap2 with only a minor loss in
accuracy (3%) (Figure S1B, see also Table S2).
Figure 1 (legend in next Page)
Figure 1. Systematic analysis of base-calling and mapping algorithms for the detection of RNA
modifications in direct RNA sequencing datasets (A) Overview of the synthetic constructs used to benchmark
the algorithms, which included both unmodified (UNM and UNM-S) and modified (m6A, m5C, hm5C and Y)
sequences. For each dataset, we performed: i) comparison of base-calling algorithms, ii) comparison of mapping
algorithms, iii) detection of RNA modifications using base-called features and iv) comparative analysis of features
to distinguish similar RNA modifications. (B) Barplots comparing the percentage of base-called reads using 4
different base-calling algorithms in 6 different unmodified and modified datasets. (C) Relative proportion of base-
called and mapped reads using all possible combinations (16) of base-callers and mappers included in this study,
for each of the 6 datasets analyzed. (D) IGV snapshots illustrating the differences in mapping for 3 distinct
datasets: UNM, m6A-modified and Y-modified when base-called with GU 3.0.3. Mismatch frequencies greater
than 0.1 have been colored, grey represents match to reference. (E) Comparison of global mismatch frequencies
using different base-calling algorithms, for the 6 datasets analyzed. Box, first to last quartiles; whiskers, 1.5x
interquartile range; center line, median; points, outliers; violin, distribution of density. (F) Principal Component
Analysis (PCA) using as input the base-calling error features of quality, mismatch frequency and deletion
frequency in positions -2, -1, 0, 1 and 2, for all datasets base-called with GU 3.0.3 and AL 2.1.7 and mapped with
GraphMap and minimap2 on sensitive settings. Only k-mers that contained a modification at position 0 were
included in the analysis, and the equivalent set of unmodified k-mers was used as a control. (G) Mismatch
frequency of each position of the 5-mers centered in the modified position (position 0). Box, first to last quartiles;
whiskers, 1.5x interquartile range; center line, median; points, outliers. See also Figure S1.
_________________________________________________________________________________
Base-calling ‘error’ signatures can be used to predict RNA modification type
While base-calling ‘errors’ can be used to identify m6A RNA modified sites 29,30,32, whether this
approach is applicable for the detection of other RNA modifications, and whether these signatures
could be employed to distinguish among distinct RNA modification types, is largely unknown. To this
end, we systematically characterized the base-calling errors caused by the presence of m6A, Y, m5C
and hm5C. We found that, regardless of the base-caller and mapper settings used, modified RNA
sequences presented decreased quality scores (Figure S1C-E) and higher mismatch frequencies
(Figure 1E), being these differences more prominent in Y-modified datasets. Principal component
analysis of base-calling ‘errors’ of each modified dataset (m6A, Y, m5C and hm5C) -relative to
unmodified- showed that this difference was greatest in Y-modified datasets (Figure 1F), and
maximized in datasets that were base-called with GU 3.0.3. Thus, we find that all four RNA
modifications can be detected in direct RNA sequencing data; however, their detection is severely
affected by the choice of both base-calling and mapping algorithms, and varies depending on the
RNA modification type.
We then examined whether the base-called ‘errors’ observed in modified and unmodified datasets
occured in the modified position. We found that both m6A and Y modifications led to increased
mismatch frequencies at the modified site (Figure 1G), mainly in the form of U-to-C mismatches in the
case of Y modifications (Figure S1F). By contrast, m5C and hm5C modifications did not appear in the
form of increased mismatch frequencies at the modified site; rather, these modifications appeared in
the form of increased mismatch frequencies in the neighboring residues (position -1 and +1 in the
case of m5C modifications; position +1 in hm5C) (Figure 1G). Moreover, the observed base-called
‘error’ signatures of m5C and hm5C were also dependent on the sequence context (Figure S1G).
Altogether, we found that all four RNA modifications studied (m6A, m5C, hm5C and Y) cause base-
calling ‘errors’, and that these ‘errors’ follow specific patterns that depend on the RNA modification
type.
Y modifications can be detected in vivo, in the form of U-to-C mismatches and with single
nucleotide resolution
We then examined whether the results obtained using in vitro transcribed constructs would be
applicable to in vivo RNA sequences. To this end, total RNA from S. cerevisiae was poly(A)-tailed to
allow for ligation between the RNA molecules and the commercial ONT adapters, and then prepared
for direct RNA sequencing (see Methods). Visual inspection of the mapped reads revealed that our
approach captured a high proportion of full-length rRNA molecules, with a high proportion of base-
calling errors present in 25s and 18s rRNAs, as could be expected from sequences that are highly
enriched in RNA modifications (Figure 2A). By contrast, 5s and 5.8s rRNAs did not show such base-
calling errors, in agreement with their low level of modification.
Then, we systematically analyzed base-called features (mismatch, deletion, insertion and per-base
qualities) in rRNAs, comparing the features from rRNA modified sites relative to unmodified ones
(Figure 2B). We found that all rRNA modification types consistently led to decreased per-base
qualities at modified sites, suggesting that per-base qualities can be employed to identify RNA
modifications, but not the underlying RNA modification type. Moreover, we found that Y modifications
caused significant variations in mismatch frequencies, in agreement with our observations using in
vitro constructs. By contrast, other RNA modifications, such as 2’-O-methylcytidine (Cm) or 5-
methylcytosine (m5C) did not appear in the form of increased mismatch frequencies at modified sites,
but rather, in the form of increased insertions. In addition, Y modifications typically appeared in the
form of U-to-C mismatches (Figure 2C, see also Figure S2), in agreement with our in vitro
observations, whereas other RNA modifications such as 2’-O-methyladenosine (Am) did not cause
mismatches with unique directionality. Thus, we conclude that distinct rRNA modification types can be
detected in the form altered base-called features in vivo, and that their base-calling ‘error’ signature is
dependent on the RNA modification type.
To confirm that the detected signal (U-to-C mismatches) in Y positions was caused by the presence of
the Y modification, we compared ribosomal RNA modification profiles from wild type S. cerevisiae to
those from snoRNA-knockout strains (snR3, snR34 and snR36), which lack Y modifications at known
rRNA positions (Figure 3A, see also Table S3). Our results show that changes in rRNA modification
profiles were consistently and exclusively observed in those positions reported as targets of each
snoRNA. Moreover, the remaining Y-modified positions were not significantly altered by the lack of Y
modifications guided by snR3, snR34 or snR36 (Figure 3B), suggesting that the modification status of
Y sites is largely independent from other Y sites.
Figure 2 (legend in next Page)
Figure 2. RNA modifications can be detected in yeast ribosomal RNA in the form of base-calling errors,
and each RNA modification type shows a distinct ‘error’ signature. (A) IGV snapshots of yeast ribosomal
subunits 5s, 5.8s, 18s and 25s. Known modification sites are indicated below each snapshot and nucleotides with
mismatch frequencies greater than >0.1 have been colored and grey represents match to reference or no
mismatch (B) Comparison of base-calling features (base quality, mismatch, deletion and insertion frequency)
from distinct RNA modification types present in yeast ribosomal RNA. The most descriptive base-calling error per
modification is outlined in red. Only RNA modification sites without additional neighboring RNA modifications in
the 5-mer were included in the analysis: Y (n=37), Am (n=14), Cm (n=8), Gm (n=8), Um (n=7), ac4C (n=2), m1A
(n=2), m3U (n=2), m5C (n=2), m1acp3Y (n=1), m5U (n=1), m7G (n=1). Box, first to last quartiles; whiskers, 1.5x
interquartile range; center line, median; dots: individual data points. (C) Ternary plots and barplots depicting the
mismatch directionality for selected rRNA modifications (Y, Am, Cm, Gm). Y rRNA modifications tend towards U-
to-C mismatches while Am, Cm and Gm modifications did not show specific mismatch directionality patterns. See
also Figure S2 and S3.
__________________________________________________________________________________________
2’-O-methylations can be detected in vivo in the form of systematic base-calling ‘errors’, but
their signatures vary across sites
We then sequenced 3 additional S. cerevisiae strains depleted of snoRNAs (snR60, snR61 and
snR62 knockouts) guiding 2’-O-methylation (Nm) at specific positions (Table S3). In contrast to Y
modifications, we found that 2’-O-methylations often caused increased mismatch and deletion
signatures not only at the modified position, but also at neighboring positions (Figure 3C, see also
Figure S3A). These errors disappeared in the knockout strain, suggesting that neighboring base-
calling errors were indeed caused by the 2’-O-methylation (Figure 3C). In contrast to Y modifications,
which mainly affected mismatch frequency, we observed that Nm modifications often affected several
base-called ‘error’ features (mismatch, insertion and deletion frequency) (Figure S3B). Thus, we
reasoned that combining all three features might improve the signal-to-noise ratio for the detection of
2’-O-methylated sites (Figure 3D), and found that the combination of features led to improved
detection of Nm-modified sites, relative to each individual feature. We should note that position
25s:Gm908 was poorly detected in both wild type and snoRNA-depleted strains (Figure S3A,B)
regardless of the feature combination used, likely due to the sequence context in which the site is
embedded -a homopolymeric GGGG sequence-, which is often troublesome for nanopore base-
calling algorithms.
Figure 3. Pseudouridylation and 2’-O-methylations cause systematic base-calling ‘errors’ as well as
altered current intensities, and their signature disappears upon depletion of snoRNAs guiding the
modification. (A) IGV snapshots of wild type and three snoRNA-depleted strains depicting the site-specific loss
of base-called errors at known Y target positions (indicated by asterisks). Nucleotides with mismatch frequencies
greater than 0.1 have been colored. (B) Comparison of snoRNA knockout mismatch frequencies for each base,
relative to wild type, with snoRNA targets sites indicated in red, and non-target sites in gray. (C) IGV snapshots of
wild type and three snoRNA knockout yeast strains depicting the site-specific loss of base-calling errors at known
Nm target positions. Nucleotides with mismatch frequencies greater than 0.1 have been colored. (D) Comparison
of snoRNA knockout summed error frequencies for each base, relative to wild type, with snoRNA targets sites
indicated in red, neighboring sites in blue and non-target sites in gray. (E,F) Distributions of per-read current
intensity at known Y-modified (E), 2’-O-methylated (F) and negative control sites. Current intensities at Y and 2’-
O-methylated positions were altered upon deletion of specific snoRNAs relative to wild type, whereas no shift
was observed in control sites. (G) Current intensity changes along the 25s rRNA molecule upon snR3 depletion,
relative to the wild type strain. In the lower panel, a zoomed subset focusing on the two regions with the most
significant current intensity deviations is shown; the first one comprising the 25s:Y2129 and 25s:Y2133 sites, and
the second one comprising the 25s:Y2264 site. (H) Comparison of current intensities in the 15-mer regions
surrounding Y and 2’-O-methyl knockout sites, for each of the 4 strains. The dotted vertical line indicates the
modified position. See also Figure S4 for current intensity changes in other knockout strains and sites. (I) Per-
read current intensity analysis centered at the 25s:Y2880 site targeted by snR34 (upper panel) and a control site,
25s:Y2880, which is not targeted by any of the knockouts (lower panel). For each site, Principal Component
Analysis was performed using 15-mer current intensity values, and the corresponding scatterplot of the two first
principal components (PC1 and PC2) is shown on the right, using as input the same read populations as in the
left panels. Each dot corresponds to a different read, and is colored according to the strain. (J) Predicted
stoichiometry of Y- and Nm-modified sites using a k-nearest neighbors (KNN) algorithm trained to classify the
reads into 2 classes: modified or unmodified. The features used to predict modifications status of every read from
which stoichiometry was calculated were signal intensity (positions -1,0,+1) and trace (positions -1,0,+1). See
also Figures S4 and S5.
__________________________________________________________________________________________
Current intensity variations can be used to detect Y and Nm RNA modifications, but do not
allow accurate prediction of the modified site
We then wondered whether Y and Nm sites would also be detected at the level of current intensity
changes. We observed that certain Y and Nm-modified sites, such as 25s:Y2129 or 25s:Am1133,
showed drastic alterations of their current intensity values in the snoRNA-depleted strain, while no
significant alteration was observed in control sites (Figure 3E,F). However, the distribution of current
intensities in some sites did not significantly change in the knockout strain (18s:Y1187, Figure 3E
lower panel) or did not differ in their mean (25s:Y2133, Figure S4A).
We hypothesized that deviations in current intensity alterations might not always be maximal in the
modified site, but might sometimes appear in neighbouring sites. To test this, we examined the
difference in current intensity values along the rRNA molecules for each wild type-knockout pair
(Figure 3G, see also Figure S4B). As expected, we found that the depletion of snR3 led to two
regions with altered current intensity values along the 25s rRNA - one comprising the 25s:Y2129 and
25s:Y2133 sites, and the second comprising the 25s:Y2264 site. However, the highest deviations in
current intensity were not observed at the modified site (Figure 3G lower panel). From all 6 Y sites
that were depleted in the 3 knockout strains studied, only 2 of them (25s:Y2826 and 25s:Y2880)
showed a maximal deviation in current intensity in the modified site (Figure 3H, see also Figure
S4C). Similarly, depletion of Nm sites led to changes in current intensity values, but the largest
deviations were not observed at the modified site. Thus, we conclude that current intensity-based
methods can detect both Y and Nm RNA modifications; however, base-calling errors are a better
choice to achieve single nucleotide resolution, at least in the case of Y RNA modifications.
Per-read current intensity analysis of Y- and Nm-modified sites allows binning of individual
reads based on their modification status
Direct RNA sequencing produces current intensity measurements for each individual native RNA
molecule. Thus, native RNA sequencing can in principle estimate modification stoichiometries by
identifying the proportion of reads with altered current intensity at a given site. To reveal whether
current intensity alone would be sufficient to bin the reads into modified and unmodified populations,
we first examined the per-read current intensity values of wild type and knockout strains at the Y- and
Nm-depleted sites. We found that there was a significant variability across reads, even when 100% of
the positions are unmodified, however, we were able to observe robust differences in current
intensities across strains at the per-read level (Figure 3I, upper panel). As a control, we performed
the same analysis in Y sites unaffected by snoRNA depletion, finding no differences between wild
type and knockout strains at these positions (Figure 3I, lower panel). However, in some sites such as
18s:1187, the per-read shifts in current intensity between the wild type and knockout strain were far
more modest (Figure S4D).
We then performed Principal Component Analysis (PCA) of the current intensity values corresponding
to the 15-mer regions that contained the modified site, for all snoRNA-depleted strains affecting Y
(snR3, snR34, snR36) and Nm modifications (snR60, snR61, snR62), as well as for the wild type
strain (Figure 3I right panels, see also Figure S4E). As could be expected based on the per-read
current intensity plots, we observed that the reads clustered into two distinct populations: the first
cluster mainly comprised unmodified reads from the snoRNA-depleted strain, whereas the second
comprised reads from the 3 other strains, which are mostly modified.
To our surprise, we observed that Nanopolish software did not resquiggle the reads evenly across
sites. For example, it failed at resquiggling the majority of reads in the region surrounding 25s:Y2264
(Figure S4D). Thus, we examined whether the Tombo resquiggling algorithm, which uses global
resquiggling instead of local resquiggling, might overcome this limitation, finding that Tombo
resquiggling led to a global increase in the proportion of resquiggled reads (Figure S5A). Moreover,
Tombo was equally effective at resquiggling both modified and unmodified reads, whereas
Nanopolish preferentially resquiggled unmodified reads relative to modified ones, biasing the
unmodified:modified proportion up to 7:1 (Figure S5B). This uneven resquiggling from Nanopolish
implies that using Nanopolish for predicting RNA modification levels at individual sites may cause a
dramatic bias in the predicted stoichiometry of individual sites, especially in scenarios where RNA
modifications are substoichiometric, such as mRNAs. Thus, based on these results, we decided to
adopt Tombo resquiggling instead of Nanopolish resquiggling for the prediction of RNA modification
stoichiometries from individual RNA reads in all our downstream analyses.
Stoichiometry prediction of Y and Nm-modified sites using signal intensity, dwell time and
trace
Our results show that the presence of Y and Nm modifications lead to significant alterations in the
current intensity profiles at the modified region (e.g. 25s:Y2880, Figure 3H-I). However, in other sites
such as 18s:Y1187, current intensity alone was insufficient to bin the reads into two separate clusters
(Figure S4D,E), suggesting that, in addition to current intensity, other features might be needed to
distinguish modified from unmodified reads.
Previous works predicting DNA modifications from individual nanopore reads have typically relied on
features such as signal intensity or dwell time to distinguish modified and unmodified read populations
46–49. Here, in addition to these two features, we explored whether the use of ‘trace’ (also termed ‘base
probability’), which is reported directly by Guppy into the base-called FAST5 files, would improve our
ability to predict RNA modification stoichiometry. To this end, we first examined how the presence of
Y and Nm modifications altered each of the features (signal intensity, dwell time and trace) in Y and
Nm modified sites by comparing the observed features in wild type and snoRNA-deficient strains, both
at snoRNA-targeted positions and control sites (Figure S6). Our results show that in addition to signal
intensity, base probability (trace) was significantly different in the snoRNA-deficient strains in all
examined sites. Moreover, in some sites such as 25s:Y2264, trace was the most altered feature from
those examined. By contrast, we found that dwell time was not consistently different in snoRNA-
targeted sites relative to wild type (e.g. 25s:Y2264, 25s:Y2826, 18s:Y1187).
We then proceeded to systematically benchmark the use of distinct features for RNA modification
stoichiometry. To this end, we built nanoRMS, a software that extracts the distinct features (signal
intensity, trace and dwell time) from individual reads, and then predicts RNA modification
stoichiometry by using distinct feature combinations as well as various machine learning algorithms.
Firstly, we generated different mixes of modified (wild type) and unmodified (knockout) reads to
simulate varying read stoichiometry (0, 20, 40, 60, 80 and 100%), for each of the Y and Nm positions
for which knockouts were available (Table S3). Then, we examined how different supervised and
unsupervised algorithms would predict the stoichiometry of each of the sites, and using distinct
combinations of the 3 features (signal intensity, trace and dwell time) for each individual site (Figure
S5C). Our results show that the combination of signal intensity and trace outperformed all the other
feature combinations for predicting both Y and Nm modification stoichiometry, and that the supervised
k-nearest neighbor (KNN) was the best performing algorithm. The k-means clustering algorithm
(KMEANS) was the best-performing algorithm among the unsupervised clustering methods tested,
although its the performance in predicting Y modification stoichiometry was slightly better than in the
case of Nm modification stoichiometry predictions. Overall, we find that nanoRMS can accurately
predict Y and Nm RNA modification stoichiometry from individual RNA reads (Figure 3J), with
predicted stoichiometry values that are similar to those that have been previously reported by Mass
Spectrometry 50 (Table S4).
De novo prediction of Y modifications reveals a novel Pus4-dependent mitochondrial rRNA
modification
The identification of RNA modification-specific signatures allows us to perform de novo prediction of Y
RNA modifications transcriptome-wide using direct RNA sequencing. In this regard, S. cerevisiae
mitochondrial rRNAs remains much less characterized than cytosolic rRNAs, with only 3 modified
sites identified so far in S.cerevisiae LSU (21s) 51, and none in SSU (15s) rRNAs. Thus, we
hypothesized that direct RNA might reveal previously uncharacterized Y-modified sites in
mitochondrial rRNAs. To this end, we first determined the ‘error’-based thresholds (mismatch
frequency and C mismatch frequency) that would distinguish unmodified uridines from pseudouridines
in cytosolic rRNAs (Figure 4A). We then applied this filter to predict Y modifications on 15s rRNA and
21s rRNA, identifying two novel candidate Y sites (15s:854 and 15s:579) that displayed high
modification frequency as well as U-to-C mismatch signature (Figure 4B,C).
To further confirm that the two predicted 15s rRNA sites are pseudouridylated, we developed
nanoCMC-seq, a novel protocol that identifies Y modifications by coupling CMC probing with
nanopore cDNA sequencing. This method allows capturing reverse-transcription drop-off information
by sequencing only the first-strand cDNA molecules of CMC-probed RNAs using a customized direct
cDNA sequencing protocol (Figure 4D, see also Methods). We found that NanoCMC-seq captured
known sites in cytoplasmic rRNA with a very high signal-to-noise ratio, as well as confirmed the
existence of Y in position 854 and 579 of 15s rRNA, validating our de novo predictions using direct
RNA sequencing (Figure 4E, see also Figure S7A).
We then examined the sequence context of these two novel 15s rRNA modifications. We observed
that 15s:Y854 was embedded in a similar sequence context and structure as the t-arm of tRNAs,
which contains a pseudouridylated (Y55) position placed by Pus4 (Figure 4F). Given the resemblance
between these two sequences and structures, we hypothesized that Pus4 might be responsible for
this modification. To validate our hypothesis, we sequenced total RNA from a S. cerevisiae Pus4
knockout strain, finding that the 15s:854 position loses its mismatch signature upon deleting Pus4
gene without altering the base-called feature of any other position on the ribosomal RNAs, confirming
that not only this site is pseudouridylated, but also that it is Pus4-dependent (Figure 4G, see also
Figure
S7B).
Additionally,
we
observed
that
previously
reported
Pus4
target
sites
(TEF1:239,TEF2:239) 3–5 completely lost their mismatch signature in Pus4 knockout cells (Figure
S7B,C), confirming that our method is able to capture previously reported Pus4-dependent Y sites, in
addition to novel ones.
Figure 4. De novo prediction of Y modifications reveals a novel Pus4-dependent mitochondrial rRNA
modification. (A) Density distributions of mismatch frequency and C mismatch frequency in unmodified uridine
positions (red) and pseudouridine positions (cyan). The dashed lines represent the optimal cutpoints between two
groups determined by maximizing the Youden-Index. In the right panel, the ROC curve illustrates the sensitivity
and specificity at these two cutpoints. (B) IGV coverage tracks of the 15s mitochondrial rRNA, including a
zoomed version showing the tracks centered at the 15s:854 and 15s:579 sites, in two biological replicates.
Nucleotides with mismatch frequencies greater than 0.15 have been colored. (C) Location of the putative Y854
modified site in the yeast mitochondrial ribosome. The LSU has been colored in cyan, whereas the SSU has
been colored in gray. The tRNA is located in the P-site of the ribosome. The PDB structure shown corresponds to
5MRC. (D) Validation of the 15s:579 and 15s:Y854 with nanoCMC-Seq, which combines CMC treatment with
Nanopore cDNA sequencing in order to capture RT-drops that occur at Y-modified sites upon CMC probing. RT-
drops are defined by counting the number of reads ending (3’) at a given position. CMC-probed samples will
cause accumulation of reads with the same 3’ ends at positions neighboring the Y site (red), whereas untreated
samples will show random distribution of 3’ ends of their reads (teal). (E) Predicted Y sites U854 and U579
(orange) in the 15s rRNA are validated using nanoCMC-seq (upper panel). Dashed lines indicate the CMC-score
threshold used for determining the positive sites (upper panel). As a control, we analyzed the nanoCMC-seq
results in other rRNAs (lower panel), finding that all positions with a significant CMC Score (>25) correspond to
known Y rRNA modification sites (blue). See also Figure S7A for CMC scores in additional rRNA transcripts. (F)
The candidate Y854 site is located at the 852-860 loop of the 15s rRNA, which resembles the t-arm of the tRNAs
that is modified by Pus4. The binding motif of Pus4 (RRUUCNA) matches the motif surrounding the 854U site 4.
(G) Scatterplot of mismatch frequencies in WT and Pus4KO cells, showing that the only significant position
affected by the knockout of Pus4 is 15s:U854 (left panel). IGV coverage tracks showing that Pus4 knockout leads
to depletion of the mismatch signature in the 15s:854 position (right panel), but not at the 15s:579 position.
__________________________________________________________________________________________
rRNA modification profiles do not vary upon exposure to oxidative or thermal stress, whereas
Y modification levels in several snRNAs and snoRNAs significantly change upon heat stress
Ribosomal RNAs are extensively modified as part of their normal maturation, and their modification
landscape is relatively well-defined for a series of organisms 38,52–55. Typically placed by either stand-
alone enzymes or snoRNA-guided mechanisms, rRNA modifications tend to cluster in functionally
important sites of the ribosome, stabilizing its structure and fine-tuning its decoding capacities 56.
Despite the central role that rRNA molecules play in protein translation, recent evidence has shown
that rRNA modifications are in fact dynamically regulated 57,58, and that their alterations can lead to
disease states 40,41,59–65. Moreover, it has been shown that some pseudouridylated and 2′-O-
methylated rRNA sites are only partially modified, and that their stoichiometry is cell-type dependent,
suggesting that rRNAs modifications may be an important source of ribosomal heterogeneity 42,50,53,66–
68. However, a systematic and comprehensive analysis of which environmental cues may lead to
changes in rRNA modification stoichiometries, which RNA modifications may be subject to this tuning,
and to which extent, is largely missing.
To assess whether rRNA modification profiles change in response to environmental stimuli, we
treated S. cerevisiae cells with diverse environmental cues (oxidative, cold and heat stress) and
sequenced their RNA, in biological duplicates, using direct RNA sequencing. Firstly, we examined the
reproducibility across biological replicates, finding that the rRNA modification profiles from
independent biological replicates were highly reproducible (pearson r2=0.976-0.996). Then, we
examined whether exposure to stress (oxidative, cold and heat stress) would lead to significant
changes in base-calling ‘errors’ in rRNA molecules, finding no significant differences in rRNA
modification profiles between normal and stress conditions (Figure 5A). By contrast, we recapitulated
previously reported changes in snRNA Y modifications upon exposure to environmental cues4 (Figure
5B, see also Figure S7D), as well as identified 8 additional Y modification sites in snRNAs and
snoRNAs whose stoichiometry varies upon heat exposure, which had not been previously described
(Figure 5, see also Figure S7E and Table S5) 3,4,37,69. Overall, our approach confirmed previous
reports and predicted novel Y sites in ncRNAs whose modification levels vary upon heat shock
exposure (Figure 5B-D, see also S7D-E), but did not identify any rRNA modified site to be varying in
its stoichiometry upon any of the tested stress conditions.
Figure 5. Comparative analysis of yeast rRNA and snRNA Y modifications upon distinct environmental
stresses identifies previously known and novel heat-sensitive snRNA and snoRNA Y modifications. (A)
Comparison of mismatch frequencies for all rRNA bases from untreated or yeast exposed to oxidative stress
(H2O2, left panel), cold stress (4ºC, middle panel) or heat stress (45ºC, right panel). Each dot represents a uridine
base. All rRNA bases from cytosolic rRNAs were included in the analyses. (B) Comparison of mismatch
frequencies in untreated versus stressed-exposed yeast cells (oxidative, cold or heat), in previously reported
ncRNA Y sites 3,4. (C) Stress scores in sn/snoRNA Y sites calculated by ∆ mismatch frequency between heat
shock and WT. (D) IGV snapshots of normal condition (rep1 and rep2) and heat shock condition (rep1 and rep2)
yeast cells zoomed into the known sn/snoRNA Y positions (indicated by an asterisk). Nucleotides with mismatch
frequencies greater than 0.1 have been colored. Coverage for each position/condition is given on the top left of
each row. (E) Profiles of ribosomal fractions isolated from yeast grown under normal conditions, using sucrose
gradient fractionation, including free rRNAs which are not assembled into ribosomal subunits (F1), rRNAs from
40s and 60s subunits (F2), rRNAs extracted from monosomal fractions (F3) and polysome fractions (F4). (F) IGV
snapshots of the two Y sites that change stoichiometry between translational fractions and four representative Y
sites that show no significant change. Nucleotides with mismatch frequencies greater than 0.1 have been
colored. See also Figure S7.
rRNA modification profiles do not vary across translational repertoires
Next we questioned whether pseudouridylation changes in distinct translational repertoires may be
more nuanced, in that Y levels may differ between rRNAs present in different translational fractions
along a polysome gradient, which would not be detected when examining rRNAs as a whole. To test
this, we sequenced both total (input) and polysomal rRNAs from untreated and H2O2-treated yeast
cells (Figure S7F). However, we observed no significant changes in Y rRNA modification profiles
when comparing rRNAs from actively translating ribosomes in untreated versus H2O2-treated cells
(Figure S7G).
In an attempt to further dissect the different translational repertoires into a higher number of rRNA
pools, we sequenced: i) rRNAs from unassembled free rRNA fractions (F1), ii) rRNAs from 40s and
60s subunits (F2), iii) rRNAs from monosomal fractions (F3) and iv) rRNAs from polysomal fractions
(F4) (Figure 5E). While two positions showed slightly decreased levels of Y (5.8s:Y73 and 25s:Y776)
in the free rRNA fraction (F1) compared to assembled ribosomes and/or subunits, no significant
changes were observed across the other translational fractions (Figure 5F, see also Figure S7H).
Globally, these results indicate that differential rRNA modification is likely not a mechanism employed
by yeast cells to adapt to environmental stress conditions, in agreement with previous observations 3.
De novo prediction of Y modifications in mRNAs using direct RNA sequencing reveals novel Y
sites that are Pus1, Pus4 and heat stress-dependent
Ribosomal RNAs are modified at very high stoichiometries 50,53. By contrast, other RNA molecules
such as mRNAs are considered to be modified at much lower stoichiometries, making the detection of
their RNA modifications a much more challenging task 21. To ascertain whether our methodology
would be applicable to lowly modified RNA sites, such as those present in mRNAs, we first assessed
the performance of nanoRMS in RNA molecules that contained Y RNA modifications at low RNA
modification stoichiometries (0, 3, 7 and 20%) (Figure 6A, see also Methods). These synthetic RNA
molecules were produced by in vitro transcription, and their relative incorporation of Y RNA
modifications was validated using Mass Spectrometry. We then examined the quantitative
performance of nanoRMS under low stoichiometry conditions using both KNN and k-means, finding
that the combination of signal intensity and trace features yielded the most accurate results in terms of
stoichiometry prediction (Figure 6B), in agreement with our previous results (Figure S5C).
Next, we sequenced polyA(+)-selected RNA from S. cerevisiae wild type, Pus1 knockout, Pus4
knockout and heat stress-exposed strains using direct RNA sequencing, in biological duplicates.
Considering that mRNA sites are lowly modified, we restricted our de novo identification of mRNA Y
sites to those whose base-calling ‘error’ features significantly changed between pairwise conditions
(Figure 6C, see also Methods), met the pseudouridine ‘error’ signature, and had a minimum coverage
of 30 reads in both conditions and biological replicates (Table S6, see also Methods). Through this
approach, we predicted 13 Pus1-dependent Y mRNA modifications, 14 Pus4-dependent Y mRNA
modifications, 17 heat stress-dependent Y mRNA modifications and 16 heat stress-dependent Y
ncRNA modifications, respectively (Figure 6D-G left panels, see also Tables S7-10), some of which
were not previously reported to be Y-modified.
NanoRMS recovered 11% of previously reported Pus1-dependent Y sites as well as 75% Pus4-
dependent Y sites, in addition to predicting 10 novel Pus1 and 11 novel Pus4-dependent mRNA Y-
modified sites (Table S7 and S8). These novel predicted Y mRNA sites displayed similar mismatch
signatures to those observed in previously reported Y sites (Figure 6D-E, right panels), were highly
replicable across biological replicates, and their signature disappeared in Pus1 or Pus4 knockout
strains. Similarly, nanoRMS was able to capture previously reported heat-responsive Y sites present
in mRNAs and ncRNAs, which resulted in predicting 17 heat-responsive Y mRNAs sites, among
which 6 of them were previously reported Y sites (Figure 6F, see also Table S9), as well as 16 heat-
responsive Y ncRNAs sites, from which 10 were previously reported Y sites (Figure 6G, see also
Table S10).
Surprised by the relatively poor overlap between our predictions and previously reported Pus1 mRNA
Y-modified sites (3 out of 16 sites), as well as poor overlap between predicted and previously reported
heat stress-dependent sites (7 out of 128 sites), we inspected the individual per-read features at
previously reported Pus1- and heat stress-dependent sites (Figure S8A,B). Indeed, the Y sites that
nanoRMS did not report as Pus1 or heat stress-dependent were not significantly different for any of
the features examined (current intensity, dwell time or trace). Thus, we wondered whether some of
these sites might have been misassigned as Pus1 or heat stress-dependent by previous works.
Indeed, if we examine the overlap between mRNA and ncRNA Y sites predicted by the two previously
published studies using CMC probing coupled to Illumina sequencing 3,4, which we used to define the
set of ‘previously reported Pus1-, Pus4- and heat stress-dependent Y sites’, we observed that the
overlap between the two studies was in fact very poor (Figure S8C), both when examining the set of
predicted mRNA and ncRNA Y sites (7% and 17%, respectively), as well as when examining the sets
of predicted Pus1- and Pus4-dependent mRNA and ncRNA Y sites (6% and 50%, respectively).
Altogether, our approach detected 100% of Pus1- and Pus4-dependent sites that were identified by
both studies, but very few of those that were identified by only one of the studies. Thus, we conclude
that the poor overlap between our results and previously reported Y sites is in fact a direct
consequence of the poor overlap between the set of predicted Pus1-, Pus4- and heat stress-
dependent mRNA and ncRNA Y sites by the two previous studies (Figure S8C).
Finally, we applied nanoRMS to predict the modification stoichiometry of all previously reported and
novel Y sites predicted in mRNAs and ncRNAs. To this end, reads were classified based on the per-
read signal intensity and trace features from positions -1, 0, and +1 using the k-means unsupervised
clustering algorithm (Figure 6H-K). As expected, we observed that per-read stoichiometry predictions
were low in non-targeted Y sites. By contrast, predicted Y Pus1/Pus4/heat stress-dependent sites
(which included both previously reported and novel Y sites) typically showed significant RNA
modification stoichiometry changes, ranging from 5 to 50% change in their Y modification
stoichiometries between the two conditions.
Altogether, our results show that the use or differential ‘error’ Y signatures are a useful approach to
identify dynamic Y RNA modifications across two conditions even at low stoichiometry sites, and that
nanoRMS can be used to de novo predict and quantify the RNA modification stoichiometry dynamics
at these sites from their per-read features, both in previously reported Y sites, as well as in de novo
predicted Y sites.
Figure 6 (legend in next Page)
Figure 6. Quantitative prediction of pseudouridine stoichiometry transcriptome-wide and systematic
benchmarking of nanoRMS using RNA molecules with diverse modification stoichiometries. (A) LC-MS
validation of pseudouridine incorporation at different proportions (0%, 3%, 20%, 100%) in the in vitro transcribed
products, relative to the expected incorporation (% YTP relative to UTP) (left panel). In the right panel, the dotplot
illustrates the mismatch frequency distribution of the uridine positions in the in vitro transcribed products
incorporated with different concentrations of Y. Each dot represents one uridine position. (B) Stoichiometry
predictions of the Y incorporated in vitro transcription products using two different algorithms (KNN and k-means)
with different current information (middle right and right panels). (C) Conditions and strains used to predict Y
mRNA modifications transcriptome-wide. (D-K) Transcriptome-wide Y RNA modification predictions and
predicted stoichiometries in mRNAs and ncRNAs, for Pus1-dependent mRNA Y sites (D,H), Pus4-dependent
mRNA Y sites (E,I), heat stress-dependent mRNA Y sites (F,J) and heat stress-dependent ncRNA Y sites (G,K).
(D-G) Venn diagrams depict the overlap between Y sites predicted by our analysis and the previously reported
pseudouridine sites. IGV snapshots of reported and novel predicted sites illustrate the absence of the mismatch
signature in the Pus1 (D) or Pus4 (E) knockout samples as well as under normal conditions, relative to heat
stress conditions in mRNA (F) and ncRNA (G) are also shown. The reported or predicted Y site is indicated by an
asterisk. Nucleotides with mismatch frequencies greater than 0.15 have been colored. We should note that IGV
snapshots that show a reference “A” with mismatch signature to G are genes that are in the minus strand (and
thus are in reality positions showing U-to-C mismatch signatures). (H-K) Quantitative analysis of previously
reported and de novo predicted Y sites in mRNAs and ncRNAs. In the left panels, comparative scatterplots of
mismatch frequency illustrate differentially modified sites of reported and de novo predicted Y sites. In the right
panels, stoichiometry prediction differences between WT and knockout strains (H-I) or between normal and heat
stress conditions (J-K) are depicted in the form of boxplots. Each dot represents a Y site.
__________________________________________________________________________________________
DISCUSSION
RNA modifications are key regulators of a wide range of biological processes 70–72. They can
modulate the fate of RNA molecules, such as mRNA splicing 73–75 or mRNA decay 76,77, as well as
affect major cell and organism-level decisions, such as cellular differentiation 78,79 and sex
determination 18,80,81. While the biological relevance of RNA modifications is out of question, a major
difficulty in studying them has been the need for tailored protocols to map each modification
individually 20,82. In this context, direct RNA nanopore sequencing has emerged as a promising
platform that can overcome many of the limitations that NGS-based methods suffer from, as it can
sequence full-length native RNA molecules, including their RNA modifications.
In the last few years, direct RNA nanopore sequencing has successfully been applied to reveal long-
read native RNA transcriptomes from a wide variety of organisms 29–31,83–86. However, the detection of
distinct RNA modification types in individual native RNA molecules is still an unsolved challenge. The
ideal solution would be that direct RNA base-calling algorithms, such as Guppy or Albacore, would
predict RNA modifications on-the-fly during the base-calling step, in a similar fashion to what Guppy
3.5.+ and later versions do in genomic DNA runs, where the base-calling algorithm can identify 6
different DNA nucleotides within the reads: A, G, C, T, m6A and m5C. However, this is not yet the
reality for direct RNA sequencing, partly due to the higher noise-to-signal ratio of RNA nanopore
reads. Consequently, solutions to identify RNA modifications in direct RNA sequencing data have so
far relied on the use of post-processing software 28–30,32,33,87.
While both current intensity-based and ‘error’-based methods have proven useful strategies to detect
RNA modifications, these methods have been mainly focused on the detection of m6A 29–31,33-, and
are typically unable to predict which RNA modification type they are in fact detecting (e.g. m6A, Y, Am
or m5C) 28,49. Moreover, current algorithms to study RNA modifications using direct RNA sequencing
are not quantitative. To overcome these limitations, here we first explored how distinct RNA
modifications may differentially affect direct RNA nanopore signals and base-calling ‘errors’. We find
that different RNA modification types (e.g. Y versus m5C) produce distinct yet characteristic base-
calling ‘error’ signatures, both in vitro (Figure 1E-G, S1F) as well as in vivo (Figure 2). Consequently,
base-calling errors can be used not only to predict whether a given site is modified or not, but also to
identify the underlying RNA modification type. While we should note that base-calling signatures
depend to some extent on the surrounding sequence context, we find that Y modifications lead to
robust U-to-C mismatch signatures, which can be exploited for de novo prediction of Y modifications
(Figure 4). Through this approach, we identified two novel Y modifications in yeast 15s mitochondrial
rRNA (15s:579 and 15s:Y854) that were not reported to date, as well as confirmed previously
reported Y-modified sites in rRNAs, snRNAs and mRNAs (Figures 3-6). Moreover, we revealed that
Pus4, which was previously thought to modify only tRNAs and mRNAs, is the enzyme responsible for
placing Y854 in mitochondrial rRNA. These findings were further validated using nanoCMC-seq, a
novel orthogonal method that can detect Y modifications with single nucleotide resolution by coupling
CMC probing to nanopore cDNA sequencing (Figure 4D).
While we find that Y modifications can be detected both in the form of base-calling ‘errors’ and altered
current intensities (Figures 3), we observe that the latter does not provide single nucleotide
resolution, with maximal current intensity shifts are often seen a few nucleotides away from the real
modified site, and that these shifts will also depend on the resquiggling algorithm used. Thus, current
intensity-based methods alone may suffer from imprecisions in the assignment of the RNA-modified
site. Here we propose that the combination of both approaches, i.e. base-called features and current
intensity/trace features, is the optimal design to obtain stoichiometric information of Y-modified sites
with single nucleotide resolution. Specifically, we show that once the site has been located using
base-calling error features, per-read features (current intensity, trace and dwell time) from the regions
surrounding Y or Nm-modified site are sufficient to robustly bin the reads into two separate clusters
(modified and unmodified), and provide good estimates of Y and Nm modification stoichiometries
(Figure 3J and 6B).
One surprising feature of base-calling ‘errors’ is that fully modified sites do not always lead to same
mismatch frequencies, suggesting that mismatch frequencies alone cannot be used per se as an
estimation of the stoichiometry of the site (Figure 2B). While it is true that within the same sequence
context, higher mismatch frequencies correspond to higher modification levels, this same rule cannot
be used to compare across distinct RNA-modified sites. We speculate that the differences observed in
mismatch frequency across different sites might be in fact a consequence of the deviation in current
intensity of the modified k-mer relative to unmodified counterparts. For example, in the case of Y, the
current intensity distribution of the Y-centered k-mers is shifted towards C-centered k-mers, and
consequently, leading to U-to-C mismatch signatures (Figure S8D). However, the shift in current
intensity may vary depending on the sequence context, leading to differences in mismatch
frequencies across Y-modified sites (e.g. 25s:Y2826 compared to 25s:Y2880), despite having similar
modification stoichiometries 50.
Finally, we should note that while nanoRMS allows predicting and studying the dynamics of diverse
RNA modifications in a quantitative manner, there are caveats and limitations, leaving ample room for
future improvements. First, not all RNA modifications lead to strong alterations in the base-calling
features and/or current intensity patterns, such as 2'-O-methylcytosine (Cm), which is poorly detected
in direct RNA sequencing datasets, compared to other RNA modifications (Figure 2C). Newer
versions of protein nanopores, which are actively being developed, might lead to increased
differences in current intensities when these RNA modifications pass through the nanopores. Second,
the detection of RNA modifications is partly dependent on the sequence context; for example, we
were unable to detect 25s:Gm908 (Figure S3). Similarly, some Y-modified sites, such as 18s:Y1187,
cause weaker alterations in base-calling features and current intensity shifts than other Y-modified
positions (Figure 3), although this limitation can be alleviated by the incorporation of additional
features into the model (Figure S5C). Third, not all RNA modifications lead to base-calling errors with
single nucleotide resolution, as with pseudouridine. For example, 2'-O-methylations often affect
neighboring bases (Figure 3C and S4A), making it challenging to de novo predict modified sites
without any prior information. Fourth, stoichiometry prediction is heavily affected by the choice of
resquiggling algorithms (Figure S5). For example, we were unable to predict stoichiometry in
25s:Y2264 when using resquiggling due to the low number of reads that the Nanopolish algorithm
was able to resquiggle (Figure S4E); however, this limitation could be overcome when using Tombo
resquiggling, leading to stoichiometry predictions similar to those observed using Mass Spectrometry
(Figure 3J). Future algorithms that improve the current intensity-to-base relationship will likely
maximize our ability to extract modification information from direct RNA nanopore sequencing
datasets. Finally, we should note that while nanoRMS was successful at detecting RNA modification
stoichiometry changes as low as 5-10% (Figure 6), the detection of RNA modification changes in
sites that show low modification stoichiometry was only possible when using comparison of pairwise
conditions.
Despite these challenges and limitations, our work provides a novel framework for the systematic and
comprehensive analysis of the epitranscriptome with single molecule resolution, showing that direct
RNA sequencing can be employed to estimate Y and Nm modification stoichiometry as well as to de
novo predict Y RNA modifications transcriptome-wide, in rRNAs, ncRNAs and mRNAs. Future work
will be needed to functionally dissect the biological roles and dynamics of RNA modifications across
further biological conditions and in disease states, to better comprehend how and when the
epitranscriptome is tuned to regulate diverse cellular functions.
ONLINE METHODS
Yeast culturing
Saccharomyces cerevisiae (strain BY4741) was grown at 30ºC in standard YPD medium (1% yeast
extract, 2% Bacto Peptone and 2% dextrose). The deletion strains snR3Δ, snR34Δ and snR36Δ were
generated on the background of the BY4741 strain by replacing the genomic snoRNA sequence with
a kanMX4 cassette as detailed in Parker et al. 91. Cells were then quickly transferred into 50 mL pre-
chilled falcon tubes, and centrifuged for 5 minutes at 3,000 g in a 4ºC pre-chilled centrifuge.
Supernatant was discarded, and cells were flash frozen. For thermal stress, Saccharomyces
cerevisiae BY4741 cultures were grown in 4 mL of YPD overnight at 30ºC. The next day, cultures
were diluted to 0.0001 OD600 in 200 mL of YPD and grown overnight at 30ºC shaking (250 rpm).
When the cultures reached an OD600 of 0.4-0.5, the cultures were divided into 3 x 50 mL subcultures,
which were then incubated at 30ºC (control), 45ºC (heat shock) or 4ºC (cold shock) for 1 hour. Cells
were collected by pelleting and snap freezing. For the analysis of rRNAs modifications across
polysomal fractions, yeast BY4741 starter cultures were grown in 6 mL YPD medium at 30ºC with
shaking (250 rpm) overnight. 100 mL of fresh YPD medium was inoculated with 10 µL of the
stationary culture in a 250 mL erlenmeyer flask, in biological duplicates. Cells were incubated at 30ºC
with shaking (250 rpm) until the cultures reached mid-exponential growth phase (O.D660.~ 0.4-0.6).
Yeast cells were then treated with 1 mM H202 or left without treatment (control) for 30 minutes. 1 mL
of cycloheximide stock solution (10 mg/mL) was added to each culture. Pus4 knockout strains
(BY4741 MATa pus4::KAN) and its parental strain were obtained from the Yeast Knockout Collection
(Dharmacon) and grown under standard conditions in YPD (1% [w/v] yeast extract, 2% [w/v] peptone
supplemented with 2% glucose) at 30°C unless stated otherwise.
Total RNA extraction from yeast cultures
Saccharomyces cerevisiae BY4741 cells (strains: snR3Δ, snR34Δ snR36Δ, snR60∆, snR61∆,
snR62∆ and WT) were harvested via centrifugation at 3000 rpm for 1 minute, followed by two washes
with water. RNA was purified from pelleted cells using a MasterPure Yeast RNA extraction kit
(Lucigen, MPY03100), according to manufacturer’s instructions. Total RNA was then treated with
Turbo DNase (Thermo, #AM2238) with a subsequent RNAClean XP bead cleanup prior to starting the
library preparation. For stress conditions and the Pus4KO strain, flash frozen pellets were
resuspended in 700 µL Trizol with 350 µL acid washed and autoclaved glass beads (425-600 µm,
Sigma G8772). The cells were disrupted using a vortex on top speed for 7 cycles of 15 seconds (the
samples were chilled on ice for 30 seconds between cycles). Afterwards, the samples were incubated
at room temperature for 5 minutes and 200 µL chloroform was added. After briefly vortexing the
suspension, the samples were incubated for 5 minutes at room temperature. Then they were
centrifuged at 14,000 g for 15 minutes at 4ºC and the upper aqueous phase was transferred to a new
tube. RNA was precipitated with 2X volume Molecular Grade Absolute ethanol and 0.1X volume
Sodium Acetate. The samples were then incubated for 1 hour at -20ºC and centrifuged at 14,000 g for
15 minutes at 4ºC. The pellet was then washed with 70% ethanol and resuspended with nuclease-
free water after air drying for 5 minutes on the benchtop. Purity of the total RNA was measured with
the NanoDrop 2000 Spectrophotometer. Total RNA was then treated with Turbo DNase (Thermo,
#AM2238) with a subsequent RNAClean XP bead cleanup.
mRNA extraction from yeast cultures
Saccharomyces cerevisiae BY4741 (strains: BY4741 MATa pus4::KAN, BY4741 MATa pus1::KAN
and BY4741 MATa) were cultured up to log phase at 30ºC. The cultures were then divided into two
flasks and cultivated at 30ºC or 45ºC for 1 hour. The cells were harvested via centrifugation at 3,000
rpm for 5 minutes and snap frozen. Total RNA was purified from pelleted cells using a MasterPure
Yeast RNA extraction kit (Lucigen, MPY03100), according to manufacturer’s instructions. Total RNA
was then DNAse-treated (Ambion, AM2239) at 37ºC for 20 minutes with a subsequent clean up using
RNeasy MinElute Cleanup Kit (Qiagen, 74204). 70-100 ug of total RNA was subjected to double
polyA-selection using Dynabeads Oligo(dT)25 (Invitrogen, 61002) and finally eluted in ice-cold 10 mM
Tris pH 7.5.
Polysome gradient fractionation and rRNA extraction
Yeast pellets from 100 mL cultures were washed with 6 mL of ice-cold Polysome Extraction Buffer
(PEB), which contained 20 mM Tris-HCl pH 7.4, 100 mM KCl, 10 mM MgCl2, 0.5 mM DTT, 0.1 mg/mL
cycloheximide and 100 U/mL RNAse inhibitors (RNaseOUT, Invitrogen, #18080051). Cells were
centrifuged for 5 minutes at 3,000 g at 4ºC. Washing was repeated by adding 6 mL of ice-cold PEB,
followed by centrifugation. Cells were then resuspended in 700 µL of ice-cold PEB, and transferred
into pre-chilled 2 mL Eppendorf tubes containing 450 µL of pre-chilled RNAse-free 425-600 µm
diameter glass beads (Sigma G8772). Cells were lysed by vortexing at maximum speed for 5 minutes
at 4ºC, followed by centrifugation also at maximum speed at bench centrifuge for 5 minutes at 4ºC.
10% of the supernatant was aliquoted into Trizol for total RNA isolation, and kept at -80ºC, which was
later used as input. The remaining volume, corresponding approximately to 8 x 108 cells, was
subsequently loaded onto the sucrose gradient. Linear sucrose gradients of 10-50% were prepared
using the Gradient Station (BioComp). Briefly, SW41 centrifugation tubes (Beckman, Ultra-ClearTM
344059) were filled with Gradient Solution 1 (GS1), which consisted of 20 mM Tris-HCl pH 7.4, 100
mM KCl, 10 mM MgCl2, 0.5 mM DTT, 0.1 mg/mL cycloheximide and 10% w/v RNAse-free sucrose.
Solutions GS1 and GS2 were prepared with RNase-DNase free UltraPure water and filtered with a
0.22 µM filter. The tube was then filled with 6.3 mL of Gradient Solution 2 (GS2) layered at the bottom
of the tube, which consisted of 20 mM Tris-HCl pH 7.4, 100 mM KCl, 10 mM MgCl2, 0.5 mM DTT, 0.1
mg/mL cycloheximide and 50% w/v RNAse-free sucrose. The linear gradient was formed using the
tilted methodology, with the Gradient Station Maker (Biocomp). Once the gradients were formed, 350
µL of each lysate was carefully loaded on top of the gradients, and tubes were balanced in pairs,
placed into pre-chilled SW41Ti buckets and centrifuged at 4ºC for 150 minutes at 35,000 rpm.
Gradients were then immediately fractionated using the Gradient Station, and 20 x 500 µL fractions
were collected in 1.5 mL Eppendorf tubes, while absorbance was monitored at 260 nm continuously.
Fractions were combined in the following way: the free rRNA (F1, fractions 1 and 2), the unassembled
subunits (F2, fractions 3-6), the lowly-translating monosomes (F3, fractions 7-10) and the highly-
translating polysomes (F4, fractions 12-17). The pooled fractions were then concentrated using
Amicon-Ultra 100K columns (Millipore), and washed two times with cold PEB. The final volume was
brought down to 200 µL, and RNA was extracted using TRIzol reagent. Purity of the RNA was
measured with NanoDrop 2000 Spectrophotometer.
In vitro transcription of modified and unmodified RNAs
The synthetic ‘curlcake’ sequences 29 used in this study are designed to include all possible 5-mers
while minimizing the secondary RNA structure, and consist in 4 in vitro transcribed constructs: (i)
Curlcake 1, 2244 bp; (ii) Curlcake 2, 2459 bp; (iii) Curlcake 3, 2595 bp, and (iv) Curlcake 4, 2709. The
curlcake constructs were in vitro transcribed using Ampliscribe™ T7-Flash™ Transcription Kit
(Lucigen-ASF3507) with either unmodified rNTPs (UNM), N6-methyladenosine triphosphate (m6ATP),
5-methylcytosine triphosphate (m5CTP), 5-hydroxymethylcytosine triphosphate (hm5CTP) or
pseudouridine triphosphate (YTP). All modified NTPs were purchased from TriLink. The sequences
included in the short unmodified dataset (UNM-S), which included B. subtilis guanine riboswitch, B.
subtilis lysine riboswitch and Tetrahymena ribozyme, were also produced by in vitro transcription
using Ampliscribe™ T7-Flash™ Transcription Kit (Lucigen-ASF3507). All constructs were 5’ capped
using vaccinia capping enzyme (NEB-M2080S) and polyadenylated using E. coli Poly(A) Polymerase
(NEB-M0276S). Poly(A)-tailed RNAs were purified using RNAClean XP beads, and the addition of
poly(A)-tail was confirmed using Agilent 4200 Tapestation. Concentration was determined using Qubit
Fluorometric Quantitation. Purity of the IVT product was measured with NanoDrop 2000
Spectrophotometer.
Direct RNA library preparation and sequencing of in vitro transcribed constructs
The RNA libraries for direct RNA Sequencing (SQK-RNA001) were prepared following the ONT Direct
RNA Sequencing protocol version DRS_9026_v1_revP_15Dec2016, which corresponds to the
flowcell FLO-MIN106. Briefly, 800 ng of Poly(A)-tailed and capped RNA (200 ng per construct) was
ligated to ONT RT Adaptor (RTA) using concentrated T4 DNA Ligase (NEB-M0202T), and was
reverse transcribed using SuperScript III RT (Thermo Fisher Scientific-18080044). The products were
purified using 1.8X Agencourt RNAClean XP beads (Fisher Scientific-NC0068576), washing with 70%
freshly prepared ethanol. RNA Adapter (RMX) was ligated onto the RNA:DNA hybrid, and the mix was
purified using 1X Agencourt RNAClean XP beads, washing with Wash buffer (WSB) twice. The
sample was then eluted in Elution Buffer (ELB) and mixed with RNA running buffer (RRB) prior to
loading onto a primed R9.4.1 flowcell, and ran on a MinION sequencer with MinKNOW acquisition
software version 1.15.1. The sequencing was performed in independent days and using a different
flowcell for each sample (UNM, m6A, m5C, hm5C, Y, UNM-S).
Direct RNA library preparation and sequencing of yeast total RNAs and mRNAs
Here we performed direct RNA sequencing of two types of S. cerevisiae RNA inputs: i) total RNA from
S. cerevisiae, and ii) polyA-selected RNA from S. cerevisiae. Yeast total RNAs were polyadenylated
using E. coli Poly(A) Polymerase (NEB, M0276S), following the commercial protocol, prior to starting
the library prep. Yeast polyA-selected RNA was directly used as input to start the libraries since they
already contain poly(A) tail. Four different direct RNA libraries were barcoded according to the recent
protocol that we recently published 92. Custom RT adaptors (IDT) were annealed using following
conditions: custom Oligo A and B (Table S11) were mixed in annealing buffer (0.01 M Tris-Cl pH 7.5,
0.05M NaCl) to the final concentration of 1.4 µM each in a total volume of 75 µL. The mixture was
incubated at 94°C for 5 minutes and slowly cooled down (-0.1°C/s) to room temperature. RNA library
for direct RNA Sequencing (SQK-RNA002) was prepared following the ONT Direct RNA Sequencing
protocol version DRS_9080_v2_revI_14Aug2019 with half reaction for each library until the RNA
Adapter (RMX) ligation step. Per reaction (half), 250 ng total of yeast RNAs were ligated to pre-
annealed custom RT adaptors (IDT) 92 using concentrated T4 DNA Ligase (NEB-M0202T), and was
reverse transcribed using Maxima H Minus RT (Thermo Scientific, EP0752), without the heat
inactivation step. The products were purified using 1.8X Agencourt RNAClean XP beads (Fisher
Scientific-NC0068576) and washed with 70% freshly prepared ethanol. 50 ng of reverse transcribed
RNA from each reaction was pooled and RMX adapter, composed of sequencing adapters with motor
protein, was ligated onto the RNA:DNA hybrid and the mix was purified using 1X Agencourt
RNAClean XP beads, washing with Wash Buffer (WSB) twice. The sample was then eluted in Elution
Buffer (EB) and mixed with RNA Running Buffer (RRB) prior to loading onto a primed R9.4.1 flowcell,
and ran on a MinION sequencer with MinKNOW acquisition software version v.3.5.5.
NanoCMC-seq
CMC treatment was adapted from Schwartz et al 4 with minor changes. Briefly, 20 ug total RNA was
incubated in NEBNext® Magnesium RNA Fragmentation Module at 94°C for 1.5 minutes. The
fragmented RNA was then incubated with either 0.3 M CMC dissolved in 100 µL TEU buffer (50 mM
Tris pH 8.5, 4 mM EDTA, 7 M Urea) or 100 µL TEU buffer (no CMC) for 20 minutes at 37°C. Reaction
was stopped with 100 µL of Buffer A (0.3 M NaOAc and 0.1 mM EDTA, pH 5.6), 700 µL absolute
ethanol, and 1 µL GlycoBlue (Thermo Scientific, AM9515). RNA in the stop solution was chilled on dry
ice for 5 minutes, and then centrifuged at maximum speed for 15 minutes at 4°C. Supernatant was
removed and the pellet was washed with 70% ethanol. After air drying for a few minutes, the pellet
was dissolved in 100 µL Buffer A and mixed with 300 µL absolute ethanol and 1 µL GlycoBlue. After
chilling on dry ice for 5 minutes, the solution was then centrifuged at maximum speed for 15 minutes
at 4°C. Supernatant was removed, and the pellet was washed with 70% ethanol. After washing, the
pellet was air dried, and resuspended in 40 µL of 50 mM sodium bicarbonate, pH 10.4, and incubated
at 37°C for 3 hours. Furthermore, RNA was mixed with 100 µL Buffer A, 700 µl ethanol, and 1 µL
Glycoblue overnight at -20°C. The next day, the solution was centrifuged at maximum speed for 15
minutes at 4°C and the pellet was washed with 70% ethanol and dissolved in the appropriate amount
of water after air drying. Unprobed and probed RNAs were treated with T4 Polynucleotide Kinase
(PNK) (NEB, M0201S) as described above before proceeding with ONT Direct cDNA sequencing.
Before starting the library preparation, 9 µL of 100 µM Reverse-transcription primer (Original ONT
VNP: 5’ /5Phos/ACTTGCCTGTCGCTCTATCTTCTTTTTTTTTTTTTTTTTTTTVN 3’) and 9 µL of 100
µM complementary oligo (CompA: 5’ GAAGATAGAGCGACAGGCAAGTA 3’ ) were mixed with 1 µL
0.2 M Tris pH 7.5 and 1 µL 1 M NaCl. The mix was incubated at 94°C for 1 minute and the
temperature was ramped down to 25°C (-0.1°C/s) in order to pre-anneal the oligos. Then, 100 ng
polyA-tailed RNA was mixed with 1 µL pre-annealed VNP+CompA, 1 µL 10 mM dNTP mix, 4 µL 5X
RT Buffer, 1 µL RNasin® Ribonuclease Inhibitor (Promega, N2511), 1 µL Maxima H Minus RT
(Thermo Scientific. EP0742) and nuclease-free water up to 20 µL. The reverse-transcription mix was
incubated at 60°C for 60 minutes and inactivated by heating at 85°C for 5 minutes before moving
ontoice. Furthermore, RNAse Cocktail (Thermo Scientific, AM2286) was added to the mix in order to
digest the RNA and the mix was incubated at 37°C for 10 minutes. Then the reaction was cleaned up
using 1.2X AMPure XP Beads (Agencourt, A63881). In order to be able to ligate the sequencing
adapters the the first strand, 1 µL 100 µM CompA was again annealed to the 15 µL cDNA in a tube
with 2.25 µL 0.1 M Tris pH 7.5, 2.25 µL 0.5 M NaCl and 2 µL nuclease-free water. The mix was
incubated at 94°C for 1 minute and the temperature was ramped down to 25 °C (-0.1°C/s) in order to
anneal the complementary to the first strand cDNA. Furthermore, 22.5 µL first strand cDNA was mixed
with 2.5 µL Native Barcode (EXP-NBD104) and 25 µL Blunt/TA Ligase Mix (NEB, M0367S) and
incubated in room temperature for 10 minutes. The reaction was cleaned up using 1X AMPure XP
beads and the libraries were pooled into one tube that finally contains 200 fmol library. The pooled
library was then ligated to the sequencing adapter (AMII) using Quick T4 DNA Ligase (NEB, M2200S)
in room temperature for 10 minutes, followed with 0.65X AMPure XP Bead cleanup using ABB Buffer
for washing. The sample was then eluted in Elution Buffer (EB) and mixed with Sequencing Buffer
(SQB) and Loading Beads (LB) prior to loading onto a primed R9.4.1 flowcell, and ran on a MinION
sequencer with MinKNOW acquisition software version v.3.5.5.
Analysis of nanoCMC-seq
Reads were base-called with stand-alone Guppy version 3.6.1 with default parameters running in
GPU, with built-in demultiplexing tool of Guppy. Unclassified reads were then demultiplexed further
using Porechop with --barcode_threshold 50 option (https://github.com/rrwick/Porechop). Then all the
merged classified reads were mapped to cytosolic and mitochondrial ribosomal RNA sequences in S.
cerevisiae using minimap2 default. Furthermore, a custom script was used to extract RT-drop
signatures and the RT-drop scores were plotted using ggplot2. All scripts used to process nanoCMC-
seq
data
with
RT-Drop
information
have
been
made
available
in
GitHub
(https://github.com/novoalab/yeast_RNA_Mod). Notably, due to the 5’ end truncation of the nanopore
sequencing reads by ~13 nt, RT-drop positions were shifted by 13 nt to accurately determine the exact
RT-drop positions. To identify significant RT drops in a given transcript, we first computed RT-drop
scores at each site, which took the difference in the coverage at a given position (0) relative to the
previous position (-1). We then computed the difference (delta RT drop-off score) in RT-drop scores
between CMC-probed and unprobed conditions. Lastly, we normalized the delta RT drop-off score at
each position by the median RT drop-off per transcript, leading to final CMC-Scores, which can be
compared across transcripts. Positions with CMC-Score greater than 25 were considered significant,
i.e. to contain a pseudouridine. We should note that the nanoCMC-seq signal-to-noise ratio is
dependent on the coverage of the individual transcript.
Demultiplexing direct RNA sequencing
Demultiplexing of the barcoded direct RNA sequencing libraries was performed using DeePlexiCon
with default parameters 92. Reads with demultiplexing confidence scores greater than 0.95 were kept
for downstream analyses. We used a lower score in the case of polysomal fractions and mRNA runs
(0.8), due to the low read coverage of some fractions and/or genes. We should note that the dataset
was also analyzed using 0.95 threshold, and results and conclusions of the analysis did not change,
compared to those obtained using 0.80 threshold.
Base-calling direct RNA sequencing
Reads were base-called with stand-alone Albacore versions 2.1.7 and 2.3.4 with the --disable_filtering
parameter, and stand-alone Guppy versions 2.3.1 and 3.0.3 with default parameters running in CPU.
In-house scripts were used for computing the number of unique and common base-called reads
between the different approaches, as well as to compare the tendency of each base-caller regarding
read lengths and qualities. Both Albacore and Guppy are available to ONT customers via their
community site (https://community.nanoporetech.com/). Differences between the base-called features
using distinct base-callers were determined using Kruskal-Wallis test with Bonferroni correction for
pairwise comparisons, whereas differences between unmodified and modified sites were assessed
using Mann-Whitney-Wilcoxon test.
Mapping algorithms and parameters
Reads were mapped using either Minimap2 44 or GraphMap 45. Minimap2 version 2.14 was run with
two different parameter settings: (i) minimap2 -ax map-ont, which is the recommended setting for
direct RNA sequencing mapping, and thus we refer to as ‘default’, and (ii) minimap2 -ax map-ont -k 5,
which we refer to as ‘sensitive’. GraphMap version 0.5.2 was also run with two different parameter
settings, for comparison, (i) graphmap align, using ‘default’ parameters, and (ii) graphmap align --
rebuild-index -v 1 --double-index --mapq -1 -x sensitive -z 1 -K fastq --min-read-len 0 -A 7 -k 5, which
is expected to increase the tolerance to errors that may occur under the presence of RNA
modifications, and thus we refer to as ‘sensitive’. Yeast total RNA runs were mapped to ribosomal
RNAs and non-coding RNA transcripts using graphmap with default settings. Yeast poly(A)-selected
runs were mapped to the yeast genome (SacCer3) using minimap2 with -ax splice -k14 -uf
parameters.
The
scripts
can
be
found
in
the
GitHub
repository
https://github.com/novoalab/yeast_RNA_Mod. Sequencing, base-calling and mapping statistics for all
yeast sequencing runs (total RNA and polyA-selected RNA) can be found in Tables S12 and 13.
Analysis of base-called features in curlcakes
Sam files were transformed into bam files using Samtools version 1.9 93, and were then sorted and
indexed in order to visualize the data using the Integrative Genomics Viewer (IGV) version 2.4.16 94.
Base-called features were extracted with EpiNano version 1.1 (https://github.com/enovoa/EpiNano).
Principal Component Analysis (PCA) was used to reduce the dimensionality of the base-calling error
data to visually inspect for base-calling differences, using as input the base-called features (mismatch
frequency, deletion frequency and per-base quality) from all 5 positions of each k-mer. Only k-mers
that contained a given modification once in the 5-mer were included in the analysis. All scripts used to
analyze in vitro transcribed sequences using different base-calling algorithms and mappers, as well as
to
generate
the
Figures
related
to
their
analysis
are
available
in
https://github.com/novoalab/Best_Practices_dRNAseq_analysis.
Analysis of base-called features in yeast RNAs
Sam files were transformed into bam files using Samtools version 1.9 93, then sorted and indexed in
order to visualize the data using the Integrative Genomics Viewer (IGV) version 2.4.16 94. Base-called
features were extracted using EpiNano version 1.1 with minor modifications, which consisted in
including in the output csv file the directionality of mismatched bases (C_frequency, G_frequency,
A_frequency,
U_frequency).
The
modified
EpiNano
script
can
be
found
at
https://github.com/novoalab/yeast_RNA_Mod. Scripts for the analysis and visualization of base-called
features are also included in the same GitHub repository.
Visualization per-read current intensities using Nanopolish
Nanopolish eventalign output was processed to extract the current intensity values corresponding to
the 15-mer regions centered in the modified sites, for the following sites: (i) 6 Y rRNA sites for which
knockout data was available (25s:2133, 25s:2129, 25s:2826, 25s:2880, 25s:2264, 18s:1187), for all 4
sequencing datasets (wild type, snR3-KO, snR34-KO, snR36-KO); (ii) 4 Nm sites for which knockout
data was available (25s:817, 25s:908, 25s:1133, 25s:1888), for all 4 sequencing datasets (wild type,
snR60-KO, snR61-KO, snR62-KO); (iii) 7 Y snRNA/snoRNA sites which were identified as heat-
sensitive, for which there was a minimum of 100 reads of coverage. Reads with empty values in the
15-mer region in the Nanopolish eventalign output were omitted from the analysis.
Analysis of current intensity, dwell time and trace
In this work, we used two different softwares to extract current intensity: Nanopolish 95 and Tombo 49.
Nanpolish was used to extract the aligned current intensity values per read and position, using the
option --scale-events. Mean current intensity per-position was computed by summing the current
intensities of all reads aligned to the same position, divided by the total number of reads mapping at a
given position. All scripts used to process Nanopolish event align output, including scripts to display
mean current intensity values along transcripts have been made available in GitHub
(https://github.com/novoalab/nanoRMS).
Signal intensity, dwell time and trace were retrieved using get_features.py script, which is available as
part of nanoRMS. This program internally uses: minimap2 (read alignment), Tombo (calculation of
signal intensity and dwell time) and ont-fast5-api (retrieval of trace). Trace represents the probability
that a given signal intensity chunk may be originating from each of the 4 canonical bases (A, C, G and
T/U), and it is reported relative to the reference base. For example, in a T reference position that is
incorrectly reported as C (common base-calling error observed for Y sites), the trace value will be
reported for the reference base (T in this case). Then, the final read alignment and all the features are
stored into sorted BAM files. All scripts necessary to retrieve and store per-read, per-position features
and
plot/calculate
results
are
available
within
the
nanoRMS
GitHub
repository
(https://github.com/novoalab/nanoRMS).
De novo prediction of pseudouridine modifications on yeast mitochondrial rRNAs
To systematically identify Y sites de novo based on the Y base-calling signatures, we first extracted
the mismatch frequency and per-base mismatch frequency (C_freq, A_freq, U_freq, G_freq) from both
unmodified (U) and modified (Y) sites from cytosolic ribosomal RNAs, from three biological replicates.
As expected, C mismatch frequency (C_freq) and global mismatch frequency (mis_freq) showed
clearly distinct distributions when comparing unmodified and Y-modified sites (Figure 4A). We then
determined the optimal cut-points for these two features using the cutpointr package in R with
oc_youden_kernel method, which applies Kernel smoothing and maximizes the Youden-Indexing.
This approach predicted C_freq=0.137 and mis_freq= 0.587 as optimal cut-offs. For the mitochondrial
ribosomal RNA, we filtered the uridine sites based on the selected features and assigned those that
are replicable in three biological replicates as “candidate” pseudouridine sites.
De novo prediction of pseudouridine modifications in yeast mRNAs and non-coding RNAs
Due to the lower stoichiometry of modification of noncoding RNAs (snRNA and snoRNAs) and
mRNAs, we focused on analysis of the de novo detection of Y sites whose pseudouridylation levels
would be changing between two conditions, either by comparing normal and stress (heat-shock)
conditions, or by comparing the base-calling ‘error’ patterns of wild type strains and Pus1 or Pus4-
deficient strains. Only sites which passed the coverage filter (n>30 reads) in both biological replicates
from both conditions were considered in the analysis (Table S6). Sites with minimal mismatch
frequency difference of 0.1 between the two conditions in both replicates that met the identified Y
signature (C_freq=0.137 and mis_freq= 0.587) were considered as true Y sites that were either heat-
sensitive, Pus1-dependent, or Pus4-dependent, respectively. The individual candidate Y mRNA and
ncRNA sites identified using nanopore sequencing, as well as the previously reported Y mRNA and
ncRNA sites (using CMC probing coupled to Illumina sequencing) can be found in Tables S7-S10.
Prediction of RNA modification stoichiometry using nanoRMS
Per-position
features
from
individual
reads
were
stored
in
BAM
files
using
pysam
(https://github.com/pysam-developers/pysam)
and
stored
them
either
in
Numpy
arrays
(https://numpy.org/)
or
Pandas
DataFrames
(https://pandas.pydata.org/)
using
the
script
get_features.py, which is available as part of nanoRMS. Models were trained with combinations of
features with diverse ranges of sequence contexts surrounding the modified sites (k=1-15). Features
used to predict stoichiometry included: (i) current intensity (SI), (ii) dwell time in the centre of the pore
(at position 0, DT/DT0), (iii) dwell time at helicase centre (shifted by 10 positions, DT10) and (iv) base
probability (trace, TR). Estimation of modification frequency was performed using unsupervised
(GMM, KMEANS, IsolationForest, OneClassSVM) and supervised (KNN, RandomForest) machine
learning methods implemented in sklearn (https://sklearn.org/). Plots were built using matplotlib and
seaborn (https://seaborn.pydata.org/).
Trained models were first benchmarked with unmodified (KO) and modified (WT) reads from rRNA
mutants dataset, to identify which machine learning methods and which combination of features
discriminated between modified and unmodified reads. Then, we tested how the diverse models
would perform at diverse stoichiometries of modification. To this end, we simulated samples with
varying levels of modification: 0%, 20%, 40%, 60%, 80% and 100% (using mixes of KO and WT
reads) and estimated the modification level in those simulated samples by comparing them to KO
(Figure S5C).
NanoRMS performed best when trained with signal intensity (SI) + trace (TR) as features, and when
using KNN supervised models or KMEANS unsupervised models, both for Y and Nm-modified sites.
Predictions by each clustering algorithm, and for each individual rRNA modified site, are shown in
Table S4. For mRNA and ncRNA analysis, only sites with more than 30 reads of coverage in all
conditions and replicates were included for predicting RNA modification stoichiometry. Prediction of
RNA modification stoichiometry in mRNAs and non-coding RNAs was performed using signal intensity
+ trace as features, and k-means as classification algorithm. Stoichiometry changes were reported as
the difference in predicted stoichiometry between the two conditions. All code and examples to predict
RNA modification stoichiometry are available as part of the nanoRMS GitHub repository
(https://github.com/novoalab/nanoRMS).
DATA AVAILABILITY
For in vitro transcribed datasets, FAST5 files used in this work were already publicly available (UNM
and m6A: PRJNA521324), or have been made publicly available in SRA (m5C:PRJNA563591; hm5C:
PRJNA548268; Y:PRJNA511582, UNM-S: PRJNA575545). Base-called and demultiplexed FASTQ
from all yeast RNA direct RNA sequencing data runs have been made publicly available in GEO,
under the accession number GSE148603, including processed EpiNano outputs and Nanopolish
outputs. FAST5 files for all yeast RNA direct RNA sequencing are available in ENA under accession
PRJEB37798 and PRJEB41495 . A detailed description of the datasets used and sequenced in this
work, with their corresponding GEO and ENA/SRA IDs can be found in Table S14. BAM files with
extracted features for the rRNA mapped reads in WT and snoRNA-depleted strains are available
through the nanoRMS GitHub repository (https://github.com/novoalab/nanoRMS).
CODE AVAILABILITY
All
scripts
and
code
used
in
this
work
have
been
made
available
in
GitHub:
https://github.com/novoalab/Best_Practices_dRNAseq_analysis
(analysis
of
in
vitro
curlcake
datasets),
https://github.com/novoalab/yeast_RNA_Mod
(analysis
of
in
vivo
datasets)
and
https://github.com/novoalab/nanoRMS (prediction of RNA modifications and estimation of RNA
modification stoichiometries).
ACKNOWLEDGEMENTS
We thank all the members of the Novoa lab for their valuable insights and discussion. We thank Vivek
Malhotra for sharing the Pus1 and Pus4 knockout strains. OB is supported by a UNSW International
PhD fellowship. MCL is supported by an FPI Severo-Ochoa fellowship by the Spanish Ministry of
Economy, Industry and Competitiveness (MEIC). IM and SC are supported by ”la Caixa” INPhINIT
PhD fellowships (LCF/BQ/DI18/11660028 and LCF/BQ/DI19/11730036, respectively). This project
has received funding from the European Union’s Horizon 2020 research and innovation programme
under the Marie Skodowska-Curie grant agreement No. 713673. This work was supported by the
Australian Research Council (DP180103571 to EMN) and the Spanish Ministry of Economy, Industry
and Competitiveness (MEIC) (PGC2018-098152-A-100 to EMN). We acknowledge the support of the
MEIC
to
the
EMBL
partnership,
Centro
de
Excelencia
Severo
Ochoa
and
CERCA
Programme/Generalitat de Catalunya.
AUTHOR CONTRIBUTIONS
OB and MCL performed the majority of wet lab experiments, including RNA extraction and nanopore
library preparation. OB and LPP performed bioinformatic analysis of the data, together with JMR and
EMN. OB conceived and performed nanoCMC-Seq experiments. MCL produced the in vitro
transcribed sequences with modifications and their corresponding nanopore libraries. OB produced
the in vitro transcribed sequences with different pseudouridine stoichiometry and performed their
corresponding nanopore library. LPP benchmarked and wrote the nanoRMS code, together with OB
and EMN. JMR performed bioinformatic analyses on in vitro transcribed constructs and compared
base-calling and mapping algorithms. IM built polysome gradients and helped with their corresponding
nanopore libraries. SC and IM prepared and sequenced the 2'-O-methylation mutant strains. HGSV
and RM cultured the S. cerevisiae strains under different stress conditions. HL contributed with code
for the analysis of current intensity values. ASC cultured all snoRNA-depleted yeast mutant strains
and extracted their total RNA. EMN conceived the project. EMN supervised the work, with the
assistance of SS and JSM. MCL, OB and EMN built the figures. OB, MCL and EMN wrote the paper,
with contributions from all authors.
DECLARATIONS OF INTERESTS
The authors declare that they have no competing interests.
REFERENCES
1.
Dominissini, D. et al. Topology of the human and mouse m6A RNA methylomes revealed by
m6A-seq. Nature 485, 201–206 (2012).
2.
Meyer, K. D. et al. Comprehensive Analysis of mRNA Methylation Reveals Enrichment in 3′
UTRs and near Stop Codons. Cell vol. 149 1635–1646 (2012).
3.
Carlile, T. M. et al. Pseudouridine profiling reveals regulated mRNA pseudouridylation in yeast
and human cells. Nature 515, 143–146 (2014).
4.
Schwartz, S. et al. Transcriptome-wide mapping reveals widespread dynamic-regulated
pseudouridylation of ncRNA and mRNA. Cell 159, 148–162 (2014).
5.
Lovejoy, A. F., Riordan, D. P. & Brown, P. O. Transcriptome-wide mapping of pseudouridines:
pseudouridine synthases modify specific mRNAs in S. cerevisiae. PLoS One 9, e110799 (2014).
6.
Li, X. et al. Chemical pulldown reveals dynamic pseudouridylation of the mammalian
transcriptome. Nat. Chem. Biol. 11, 592–597 (2015).
7.
Hussain, S., Aleksic, J., Blanco, S., Dietmann, S. & Frye, M. Characterizing 5-methylcytosine in
the mammalian epitranscriptome. Genome Biol. 14, 215 (2013).
8.
Huang, T., Chen, W., Liu, J., Gu, N. & Zhang, R. Genome-wide identification of mRNA 5-
methylcytosine in mammals. Nat. Struct. Mol. Biol. 26, 380–388 (2019).
9.
Delatte, B. et al. RNA biochemistry. Transcriptome-wide distribution and function of RNA
hydroxymethylcytosine. Science 351, 282–285 (2016).
10. Safra, M. et al. The m1A landscape on cytosolic and mitochondrial mRNA at single-base
resolution. Nature 551, 251–255 (2017).
11. Li, X. et al. Base-Resolution Mapping Reveals Distinct mA Methylome in Nuclear- and
Mitochondrial-Encoded Transcripts. Mol. Cell 68, 993–1005.e9 (2017).
12. Marchand, V. et al. AlkAniline-Seq: Profiling of m7G and m3C RNA Modifications at Single
Nucleotide Resolution. Angew. Chem. Int. Ed. 57, 16785–16790 (2018).
13. Arango, D. et al. Acetylation of Cytidine in mRNA Promotes Translation Efficiency. Cell 175,
1872–1886.e24 (2018).
14. Sas-Chen, A. et al. Dynamic RNA acetylation revealed by quantitative cross-evolutionary
mapping. Nature (2020) doi:10.1038/s41586-020-2418-2.
15. Zhang, L.-S. et al. Transcriptome-wide Mapping of Internal N7-Methylguanosine Methylome in
Mammalian mRNA. Mol. Cell 74, 1304–1316.e8 (2019).
16. Pandolfini, L. et al. METTL1 Promotes let-7 MicroRNA Processing via m7G Methylation. Mol. Cell
74, 1278–1290.e9 (2019).
17. Delaunay, S. & Frye, M. RNA modifications regulating cell fate in cancer. Nat. Cell Biol. 21, 552–
559 (2019).
18. Haussmann, I. U. et al. m6A potentiates Sxl alternative pre-mRNA splicing for robust Drosophila
sex determination. Nature vol. 540 301–304 (2016).
19. Vu, L. P. et al. The N6-methyladenosine (m6A)-forming enzyme METTL3 controls myeloid
differentiation of normal hematopoietic and leukemia cells. Nature Medicine vol. 23 1369–1376
(2017).
20. Novoa, E. M., Mason, C. E. & Mattick, J. S. Charting the unknown epitranscriptome. Nat. Rev.
Mol. Cell Biol. 18, 339–340 (2017).
21. Anreiter, I., Mir, Q., Simpson, J. T., Janga, S. C. & Soller, M. New Twists in Detecting mRNA
Modification Dynamics. Trends Biotechnol. 0, (2020).
22. Li, X., Xiong, X. & Yi, C. Epitranscriptome sequencing technologies: decoding RNA modifications.
Nat. Methods 14, 23–31 (2016).
23. Motorin, Y. & Helm, M. Methods for RNA Modification Mapping Using Deep Sequencing:
Established and New Emerging Technologies. Genes 10, (2019).
24. Grozhik, A. V. et al. Antibody cross-reactivity accounts for widespread appearance of m 1 A in
5’UTRs. Nat. Commun. 10, 1–13 (2019).
25. Lahens, N. F. et al. IVT-seq reveals extreme bias in RNA sequencing. Genome Biol. 15, R86
(2014).
26. Garalde, D. R. et al. Highly parallel direct RNA sequencing on an array of nanopores. Nat.
Methods 15, 201–206 (2018).
27. Jonkhout, N. et al. The RNA modification landscape in human disease. RNA 23, 1754–1769
(2017).
28. Leger, A., Amaral, P. P., Pandolfini, L. & Capitanchik, C. RNA modifications detection by
comparative Nanopore direct RNA sequencing. BioRxiv (2019).
29. Liu, H. et al. Accurate detection of m6A RNA modifications in native RNA sequences. Nat.
Commun. 10, 4079 (2019).
30. Parker, M. T. et al. Nanopore direct RNA sequencing maps the complexity of Arabidopsis mRNA
processing and m6A modification. eLife vol. 9 (2020).
31. Price, A. M. et al. Direct RNA sequencing reveals m6A modifications on adenovirus RNA are
necessary for efficient splicing. bioRxiv 865485 (2019) doi:10.1101/865485.
32. Wongsurawat, T. et al. Decoding the Epitranscriptional Landscape from Native RNA Sequences.
bioRxiv 487819 (2018) doi:10.1101/487819.
33. Lorenz, D. A., Sathe, S., Einstein, J. M. & Yeo, G. W. Direct RNA sequencing enables m6A
detection in endogenous transcript isoforms at base-specific resolution. RNA 26, 19–28 (2020).
34. Jack, K. et al. rRNA pseudouridylation defects affect ribosomal ligand binding and translational
fidelity from yeast to human cells. Mol. Cell 44, 660–666 (2011).
35. Yoon, A. et al. Impaired control of IRES-mediated translation in X-linked dyskeratosis congenita.
Science 312, 902–906 (2006).
36. Bellodi, C. et al. Loss of Function of the Tumor Suppressor DKC1 Perturbs p27 Translation
Control and Contributes to Pituitary Tumorigenesis. Cancer Research vol. 70 6026–6035 (2010).
37. Wu, G., Xiao, M., Yang, C. & Yu, Y.-T. U2 snRNA is inducibly pseudouridylated at novel sites by
Pus7p and snR81 RNP. EMBO J. 30, 79–89 (2011).
38. Taoka, M., Nobe, Y., Hori, M. & Takeuchi, A. A mass spectrometry-based method for
comprehensive quantitative determination of post-transcriptional RNA modifications: the complete
chemical structure of …. Nucleic acids (2015).
39. Basu, A. et al. Requirement of rRNA methylation for 80S ribosome assembly on a cohort of
cellular internal ribosome entry sites. Mol. Cell. Biol. 31, 4482–4499 (2011).
40. Marcel, V. et al. p53 acts as a safeguard of translational control by regulating fibrillarin and rRNA
methylation in cancer. Cancer Cell 24, 318–330 (2013).
41. Belin, S. et al. Dysregulation of Ribosome Biogenesis and Translational Capacity Is Associated
with Tumor Progression of Human Breast Cancer Cells. PLoS ONE vol. 4 e7147 (2009).
42. Buchhaupt, M. et al. Partial methylation at Am100 in 18S rRNA of baker’s yeast reveals ribosome
heterogeneity on the level of eukaryotic rRNA modification. PLoS One 9, e89640 (2014).
43. Chen, H. et al. METTL5, an 18S rRNA-specific m6A methyltransferase, modulates expression of
stress response genes. bioRxiv 2020.04.27.064162 (2020) doi:10.1101/2020.04.27.064162.
44. Li, H. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics vol. 34 3094–3100
(2018).
45. Sović, I. et al. Fast and sensitive mapping of nanopore sequencing reads with GraphMap. Nat.
Commun. 7, 11307 (2016).
46. Liu, Q. et al. Detection of DNA base modifications by deep recurrent neural network on Oxford
Nanopore sequencing data. Nat. Commun. 10, 2449 (2019).
47. McIntyre, A. B. R. et al. Single-molecule sequencing detection of N6-methyladenine in microbial
reference materials. Nat Commun 10: 579. (2019).
48. De Coster, W., Stovner, E. B. & Strazisar, M. Methplotlib: analysis of modified nucleotides from
nanopore sequencing. Bioinformatics 36, 3236–3238 (2020).
49. Stoiber, M. et al. De novo Identification of DNA Modifications Enabled by Genome-Guided
Nanopore Signal Processing. bioRxiv 094672 (2017) doi:10.1101/094672.
50. Taoka, M. et al. The complete chemical structure of Saccharomyces cerevisiae rRNA: partial
pseudouridylation of U2345 in 25S rRNA by snoRNA snR9. Nucleic Acids Res. 44, 8951–8961
(2016).
51. Pintard, L., Bujnicki, J. M., Lapeyre, B. & Bonnerot, C. MRM2 encodes a novel yeast
mitochondrial 21S rRNA methyltransferase. EMBO J. 21, 1139–1147 (2002).
52. Sharma, S. & Lafontaine, D. L. J. ‘View From A Bridge’: A New Perspective on Eukaryotic rRNA
Base Modification. Trends Biochem. Sci. 40, 560–575 (2015).
53. Taoka, M. et al. Landscape of the complete RNA chemical modifications in the human 80S
ribosome. Nucleic Acids Res. 46, 9289–9298 (2018).
54. Fischer, N. et al. Structure of the E. coli ribosome--EF-Tu complex at< 3 Å resolution by C s-
corrected cryo-EM. Nature 520, 567–570 (2015).
55. Sergeeva, O. V., Bogdanov, A. A. & Sergiev, P. V. What do we know about ribosomal RNA
methylation in Escherichia coli? Biochimie 117, 110–118 (2015).
56. Sloan, K. E. et al. Tuning the ribosome: The influence of rRNA modification on eukaryotic
ribosome biogenesis and function. RNA Biol. 14, 1138–1152 (2017).
57. Hebras, J., Krogh, N., Marty, V., Nielsen, H. & Cavaillé, J. Developmental changes of rRNA
ribose methylations in the mouse. RNA Biol. 1–15 (2019).
58. Higa-Nakamine, S. et al. Loss of ribosomal RNA modification causes developmental defects in
zebrafish. Nucleic Acids Res. 40, 391–398 (2012).
59. Sahoo, T. et al. Prader-Willi phenotype caused by paternal deficiency for the HBII-85 C/D box
small nucleolar RNA cluster. Nat. Genet. 40, 719–721 (2008).
60. Heiss, N. S. et al. X-linked dyskeratosis congenita is caused by mutations in a highly conserved
gene with putative nucleolar functions. Nat. Genet. 19, 32–38 (1998).
61. Knight, S. W. et al. X-Linked Dyskeratosis Congenita Is Predominantly Caused by Missense
Mutations in the DKC1 Gene. The American Journal of Human Genetics vol. 65 50–58 (1999).
62. Liao, J. et al. Small nucleolar RNA signatures as biomarkers for non-small-cell lung cancer. Mol.
Cancer 9, 198 (2010).
63. Mei, Y.-P. et al. Small nucleolar RNA 42 acts as an oncogene in lung tumorigenesis. Oncogene
31, 2794–2804 (2012).
64. Bortolin-Cavaille, M.-L., -L. Bortolin-Cavaille, M. & Cavaille, J. The SNORD115 (H/MBII-52) and
SNORD116 (H/MBII-85) gene clusters at the imprinted Prader-Willi locus generate canonical box
C/D snoRNAs. Nucleic Acids Research vol. 40 6800–6807 (2012).
65. Erales, J. et al. Evidence for rRNA 2’-O-methylation plasticity: Control of intrinsic translational
capabilities of human ribosomes. Proc. Natl. Acad. Sci. U. S. A. 114, 12934–12939 (2017).
66. Krogh, N. et al. Profiling of 2′-O-Me in human rRNA reveals a subset of fractionally modified
positions and provides evidence for ribosome heterogeneity. Nucleic Acids Research vol. 44
7884–7895 (2016).
67. Birkedal, U. et al. Profiling of ribose methylations in RNA by high-throughput sequencing. Angew.
Chem. Int. Ed Engl. 54, 451–455 (2015).
68. Natchiar, S. K., Myasnikov, A. G., Kratzat, H., Hazemann, I. & Klaholz, B. P. Visualization of
chemical modifications in the human 80S ribosome structure. Nature 551, 472–477 (2017).
69. van der Feltz, C., DeHaven, A. C. & Hoskins, A. A. Stress-induced Pseudouridylation Alters the
Structural Equilibrium of Yeast U2 snRNA Stem II. J. Mol. Biol. 430, 524–536 (2018).
70. Frye, M., Harada, B. T., Behm, M. & He, C. RNA modifications modulate gene expression during
development. Science 361, 1346–1349 (2018).
71. Roundtree, I. A., Evans, M. E., Pan, T. & He, C. Dynamic RNA Modifications in Gene Expression
Regulation. Cell 169, 1187–1200 (2017).
72. Li, S. & Mason, C. E. The pivotal regulatory landscape of RNA modifications. Annu. Rev.
Genomics Hum. Genet. 15, 127–150 (2014).
73. Wang, X. et al. LARP7-Mediated U6 snRNA Modification Ensures Splicing Fidelity and
Spermatogenesis in Mice. Mol. Cell 77, 999–1013.e6 (2020).
74. Louloupi, A., Ntini, E., Conrad, T. & Ørom, U. A. V. Transient N-6-Methyladenosine
Transcriptome Sequencing Reveals a Regulatory Role of m6A in Splicing Efficiency. Cell Reports
vol. 23 3429–3437 (2018).
75. Zhou, K. I. et al. Regulation of Co-transcriptional Pre-mRNA Splicing by m6A through the Low-
Complexity Protein hnRNPG. Mol. Cell 76, 70–81.e9 (2019).
76. Lee, Y., Choe, J., Park, O. H. & Kim, Y. K. Molecular Mechanisms Driving mRNA Degradation by
m6A Modification. Trends Genet. 36, 177–188 (2020).
77. Guo, M., Liu, X., Zheng, X., Huang, Y. & Chen, X. m6A RNA Modification Determines Cell Fate
by Regulating mRNA Degradation. Cellular Reprogramming vol. 19 225–231 (2017).
78. Geula, S. et al. Stem cells. m6A mRNA methylation facilitates resolution of naïve pluripotency
toward differentiation. Science 347, 1002–1006 (2015).
79. Weng, H. et al. METTL14 Inhibits Hematopoietic Stem/Progenitor Differentiation and Promotes
Leukemogenesis via mRNA m6A Modification. Cell Stem Cell 22, 191–205.e9 (2018).
80. Lence, T. et al. m6A modulates neuronal functions and sex determination in Drosophila. Nature
540, 242–247 (2016).
81. Kan, L. et al. The m6A pathway facilitates sex determination in Drosophila. Nat. Commun. 8,
15737 (2017).
82. Schaefer, M., Kapoor, U. & Jantsch, M. F. Understanding RNA modifications: the promises and
technological bottlenecks of the ‘epitranscriptome’. Open Biology vol. 7 170077 (2017).
83. Depledge, D. P. et al. Direct RNA sequencing on nanopore arrays redefines the transcriptional
complexity of a viral pathogen. Nat. Commun. 10, 754 (2019).
84. Roach, N. P. et al. The full-length transcriptome of C. elegans using direct RNA sequencing.
Genome Res. 30, 299–312 (2020).
85. Workman, R. E. et al. Nanopore native RNA sequencing of a human poly(A) transcriptome.
doi:10.1101/459529.
86. Kim, D. et al. The Architecture of SARS-CoV-2 Transcriptome. Cell 181, 914–921.e10 (2020).
87. Pratanwanich, P. N. et al. Detection of differential RNA modifications from direct RNA sequencing
of human cell lines. bioRxiv 2020.06.18.160010 (2020) doi:10.1101/2020.06.18.160010.
88. Bokar, J. A., Shambaugh, M. E., Polayes, D., Matera, A. G. & Rottman, F. M. Purification and
cDNA cloning of the AdoMet-binding subunit of the human mRNA (N6-adenosine)-
methyltransferase. RNA 3, 1233–1247 (1997).
89. Torchet, C. et al. The complete set of H/ACA snoRNAs that guide rRNA pseudouridylations in
Saccharomyces cerevisiae. RNA 11, 928–938 (2005).
90. Hamma, T. & Ferré-D’Amaré, A. R. Pseudouridine synthases. Chem. Biol. 13, 1125–1135 (2006).
91. Parker, S. et al. A resource for functional profiling of noncoding RNA in the yeast Saccharomyces
cerevisiae. RNA 23, 1166–1171 (2017).
92. Smith, M. A. et al. Barcoding and demultiplexing Oxford Nanopore native RNA sequencing reads
with deep residual learning. bioRxiv 864322 (2019) doi:10.1101/864322.
93. Li, H. et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics vol. 25 2078–
2079 (2009).
94. Robinson, J. T. et al. Integrative genomics viewer. Nat. Biotechnol. 29, 24–26 (2011).
95. Loman, N. J., Quick, J. & Simpson, J. T. A complete bacterial genome assembled de novo using
only nanopore sequencing data. Nat. Methods 12, 733–735 (2015).
| 2021 | Quantitative profiling of native RNA modifications and their dynamics using nanopore sequencing | 10.1101/2020.07.06.189969 | [
"Begik Oguzhan",
"Lucas Morghan C",
"Pryszcz Leszek P",
"Ramirez Jose Miguel",
"Medina Rebeca",
"Milenkovic Ivan",
"Cruciani Sonia",
"Liu Huanle",
"Vieira Helaine Graziele Santos",
"Sas-Chen Aldema",
"Mattick John S",
"Schwartz Schraga",
"Novoa Eva Maria"
] | null |
1
CD4-binding site immunogens elicit heterologous anti-HIV-1 neutralizing antibodies in
transgenic and wildtype animals
Harry B. Gristick1,‡, Harald Hartweger2,‡,, Maximilian Loewe2, Jelle van Schooten1,#, Victor Ramos2,
Thiago Y. Oliviera2, Yoshiaki Nishimura3, Nicholas S. Koranda1, Abigail Wall4,5,¶, Kai-Hui Yao2, Daniel
Poston2,§, Anna Gazumyan2, Marie Wiatr2, Marcel Horning2, Jennifer R. Keeffe1, Magnus A.G.
Hoffmann1, Zhi Yang1, Morgan E. Abernathy1,¢, Kim-Marie A. Dam1, Han Gao1, Priyanthi N.P.
Gnanapragasam1, Leesa M. Kakutani1, Ana Jimena Pavlovitch-Bedzyk1,†, Michael S. Seaman6, Mark
Howarth7, Andrew T. McGuire4,5, Leonidas Stamatatos4,5, Malcolm A. Martin3, Anthony P. West, Jr.1,
Michel C. Nussenzweig2,8,*, Pamela J. Bjorkman1,*
1Division of Biology and Biological Engineering, California Institute of Technology, Pasadena, CA 91125, USA
2Laboratory of Molecular Immunology, The Rockefeller University, New York, NY 10065, USA
3Laboratory
of
Molecular
Microbiology,
National
Institute
of
Allergy
and
Infectious
Diseases,
National Institutes of Health, Bethesda, MD 20892, USA.
4Vaccine and Infectious Disease Division, Fred Hutchinson Cancer Research Center, Seattle, WA
5Department of Global Health, University of Washington, Seattle, WA
6Center for Virology and Vaccine Research, Beth Israel Deaconess Medical Center, Boston, MA 02215, USA.
7Department of Biochemistry, University of Oxford, South Parks Road, Oxford, OX1 3QU, UK
8Howard Hughes Medical Institute, The Rockefeller University, New York, NY 10065, USA
*Corresponding authors: bjorkman@caltech.edu, nussen@rockefeller.edu
‡These authors contributed equally
#Present Address: Department of Medical Microbiology, Amsterdam UMC, University of Amsterdam,
Amsterdam, the Netherlands.
¶Present Address: Sage Bionetworks, Seattle, WA 98121.
§Present Address: Laboratory of Retrovirology, The Rockefeller University, New York, NY 10065, USA
¢Present Address: 290 Jane Stanford Way, ChEM-H/Neuro Building, Stanford, CA 94305, USA
†Present Address: Department of Microbiology and Immunology, Stanford University School of Medicine,
Stanford, CA 94305, USA.
2
Summary
Passive transfer of broadly neutralizing anti-HIV-1 antibodies (bNAbs) protects against infection,
and therefore eliciting bNAbs by vaccination is a major goal of HIV-1 vaccine efforts. bNAbs that target
the CD4-binding site (CD4bs) on HIV-1 Env are among the most broadly active, but to date, responses
elicited against this epitope in vaccinated animals have lacked potency and breadth. We hypothesized
that CD4bs bNAbs resembling the antibody IOMA might be easier to elicit than other CD4bs antibodies
that exhibit higher somatic mutation rates, a difficult-to-achieve mechanism to accommodate Env’s
N276gp120 N-glycan, and rare 5-residue light chain complementarity determining region 3s (CDRL3s).
As an initial test of this idea, we developed IOMA germline-targeting Env immunogens and evaluated
a sequential immunization regimen in transgenic mice expressing germline-reverted IOMA. These mice
developed CD4bs epitope-specific responses with heterologous neutralization, and cloned antibodies
overcame neutralization roadblocks including accommodating the N276gp120 glycan, with some
neutralizing selected HIV-1 strains more potently than IOMA. The immunization regimen also elicited
CD4bs-specific responses in animals containing polyclonal antibody repertoires. Thus, germline-
targeting of IOMA-class antibody precursors represents a potential vaccine strategy to induce CD4bs
bNAbs.
Introduction
A successful vaccine against HIV-1 would be the most effective way to contain the AIDS pandemic,
which so far is responsible for > 36 million deaths in total and 1 - 2 million new infections each year
(https://www.unaids.org/en/resources/fact-sheet). Clinical trials of vaccine candidates have revealed
disappointing outcomes, and as a result, there is no currently available protective vaccine against HIV-
1 (1), in part due to the large number of circulating HIV-1 strains (2). For the last decade, a major focus
of HIV-1 vaccine design has been on eliciting broadly neutralizing antibodies (bNAbs), which neutralize
a majority of HIV-1 strains in vitro at low concentrations (1). Multiple studies have demonstrated that
passively administered bNAbs can prevent HIV-1 or simian/human immunodeficiency virus (SHIV)
infection (3-15), suggesting a vaccination regimen that elicits bNAbs at neutralizing concentrations
would be protective.
The HIV-1 Envelope protein (Env), a trimeric membrane glycoprotein comprising gp120 and gp41
subunits that is found on the surface of the virus, is the sole antigenic target of neutralizing antibodies
(16). An impediment to HIV-1 vaccine design is that most inferred germline (iGL) precursors of known
bNAbs do not bind with detectable affinity to native Envs on circulating HIV-1 strains (17-28). As a
result, potential Env immunogens must be modified to bind and select for bNAb precursors in vivo
3
during immunization (i.e., a “germline-targeting” approach). This approach has been used to activate
precursors of the VRC01-class of bNAbs that target the CD4 binding site (CD4bs) on gp120 (25, 29).
Eliciting VRC01-class bNAbs that target the CD4bs would be desirable due to their breadth and potency
(30). However, the VRC01-class of bNAbs may be difficult to elicit due to their requirement for rare
short light chain complementarity region 3 (CDRL3) loops of 5 residues (present in only ~1% of human
antibodies) (31) and many somatic hypermutations (SHMs), including a difficult-to achieve sequence
of mutations to sterically accommodate the highly-conserved N276gp120 glycan (32).
Crystal structures of a natively glycosylated HIV-1 soluble Env trimer derived from the clade A
BG505 strain (BG505 SOSIP.664) (33) complexed with the antibody IOMA, revealed that this CD4bs
bNAb exhibits distinct properties from VRC01-class bNAbs (34). In common with VRC01-class bNAbs,
IOMA is derived from the VH1-2 immunoglobulin heavy chain (HC) gene segment, and it binds Env
with a similar overall pose as other VH1-2–derived CD4bs bNAbs, but it is not as potent or broad as
many of the VRC01-class antibodies (34). However, unlike VRC01-class bNAbs, IOMA includes a
normal-length (8 residues) CDRL3 (34) and is less mutated with 9.5% HC and 7% light chain (LC)
nucleotide mutations to its iGL compared to VRC01 with 30% HC and 19% LC nucleotide mutations
(35, 36). In addition, IOMA accommodates the N276gp120 glycan, a roadblock for raising VRC01-class
bNAbs (32), using a relatively easy-to-achieve mechanism involving a short helical CDRL1, and four
amino acid changes (including a single mutated glycine) that each require single nucleotide
substitutions. By contrast, the CDRL1s of VRC01-class bNAbs include either a 3 - 6 residue deletion
or large numbers of SHMs that introduce multiple glycines and/or other insertions to create flexible
CDRL1 loops (32). Thus, IOMA-like antibodies likely represent an easier pathway for vaccine induced
maturation of CD4bs precursors to mature CD4bs bNAbs.
Here we report immunogens engineered to elicit IOMA and other CD4bs bNAbs. Using these
immunogens, we devised a sequential immunization strategy that elicited broad heterologous serum
neutralization in both IOMA iGL knock-in and wildtype (wt) mouse models. Notably, this was achieved
using fewer than half of the immunizations in other studies (37, 38). Moreover, IOMA-like bNAbs elicited
in knock-in mice were more potent than IOMA against some strains. Finally, the immunization regimen
developed in knock-in mice also elicited CD4bs-specific responses in multiple wt animals including
mice, rabbits, and rhesus macaques providing a rationale for using the IOMA-targeting immunogens
described here as part of an effective HIV vaccine.
Results
Design of IOMA-targeting immunogens
4
To create the IOMA iGL antibody, we reverted the HC and LC sequences of mature IOMA (34) to
their presumptive germline sequences. The IOMA iGL HC sequence was based on human IGHV1-
2*02, IGHD3-22*01 and IGHJ6*02 and contained 22 amino acid changes compared to the HC of IOMA,
all within the V gene. The CDRH3 was unaltered due to uncertainty with respect to D gene alignment
and potential P and N nucleotides – the IOMA iGL HC sequence maintains one gp120-contacting
residue (W100F; Kabat numbering (39)) found in mature IOMA (34). The sequence of the IOMA iGL
LC was derived from human IGLV2-23*02 and IGLJ2*01 containing 16 amino acid changes compared
with mature IOMA, including 3 SHMs in CDRL3. Of the 3 mutations in CDRL3, two non-contact amino
acids (V96, A97; Kabat numbering (39)) at the V-J junction were left as in mature IOMA (Figure S1A,
Table S1).
No currently available germline-targeting CD4bs immunogens bind IOMA iGL with detectable affinity
(Figure S1B), and IOMA iGL does not neutralize primary HIV-1 strains (Figure S1C). We therefore used
in vitro selection methods to identify potential IOMA-targeting immunogens (Figure 1A). We chose S.
cerevisiae yeast display for selecting an IOMA-targeting immunogen for two reasons: (i) Yeast libraries
can contain up to 1 x 109 variants (40, 41) and therefore allow screening a large number of immunogen
constructs, a necessity since we were starting with no detectable binding of IOMA iGL to any CD4bs-
targeting immunogens, and (ii) S. cerevisiae attach different forms of N-linked glycans to glycoproteins
than mammalian cells; e.g., yeast can add up to 50 mannoses to Man8-9GlcNAc2 (42). Such glycan
differences may be an advantage because N-glycosylated immunogens selected in a yeast library to
bind IOMA iGL and increasingly mature forms of IOMA might stimulate an antibody maturation pathway
that is relatively insensitive to the form of N-glycan at any potential N-linked glycosylation site (PNGS)
on HIV-1 Env. Promiscuous glycan recognition is desired because Env trimers on viruses exhibit
heterogeneous glycosylation at single PNGSs, even within one HIV-1 strain (43-45). Using yeast
display for immunogen selection, we sought to achieve promiscuous N-glycan accommodation through
recognition of a glycan’s core pentasaccharide, a common feature of both complex-type and high-
mannose glycans, which we observed as being recognized at some N-glycan sites in structures of
antibody Fab-Env complexes (21, 46, 47).
Yeast display libraries were produced using variants of the 426c.NLGS.TM4∆V1-3 monomeric
gp120 immunogen (hereafter referred to as 426c.TM4 gp120), a modified clade C gp120 that was
designed to engage VRC01-class precursor antibodies (25, 48, 49). We started with a gp120-based
immunogen instead of the engineered outer domain immunogens (eODs) previously used to select for
VRC01-class bNAb precursors (27, 29, 50) because, unlike certain other CD4bs bNAbs, IOMA contacts
the inner domain of gp120 (34), which is absent in the eOD constructs (27, 29, 50). To aid in determining
5
which immunogen residues should be varied to achieve IOMA iGL binding, we solved a 2.07 Å crystal
structure of IOMA iGL Fab (Figure S1D, Table S2), which was nearly identical (root mean square
deviation, RMSD, of 0.64 Å for 209 Ca atoms) to the mature IOMA Fab structure complexed with BG505
Env trimer (34) (Figure S1E). Based on modeling the IOMA iGL Fab structure (Figure 1B-C, Figure
S1D) into the mature IOMA Fab-Env structure (34), we varied 7 positions in 426c.TM4 gp120. A library
with ~108 variants was produced using degenerate codons so that all possible amino acids were
incorporated at the selected positions (see methods). R278gp120 was varied because this position might
select for IOMA iGL’s unique CDRL1 conformation (34). In addition, we introduced a D279Ngp120
substitution because IOMA is ~2-3-fold more potent against HIV-1 viruses that have an N at this position
(34). Next, V430gp120 was varied to increase the interaction with the HC of IOMA iGL. Lastly, residues
460gp120-464gp120 were varied in the V5-loop of 426c gp120 to accommodate and select for IOMA iGL’s
normal length CDRL3.
Following three rounds of fluorescence-activated cell sorting (FACS) using one fluorophore for
IOMA iGL and another against a C-terminal Myc tag to monitor gp120 expression, there was a > 100-
fold enrichment for gp120 variants that bound IOMA iGL (Figure S1F, middle panel), demonstrated by
increased staining for IOMA iGL compared to the starting 426c.TM4 gp120 (Figure 1B-C, Figure S1F,
left and middle panels). Two clones (from ~100 sequenced after the third sort) accounted for 50% of
the sequences, suggesting that IOMA iGL-binding activity was enriched. IGT1, the best variant
identified by the initial yeast display library, had an affinity of ~30 µM for IOMA iGL, as determined by
a surface plasmon resonance (SPR)-based binding assay (Figure 1C, right panel). IGT1 was then used
as a guide to construct a second yeast library to select for an immunogen with higher affinity to IOMA
iGL (Figure 1D). Based on their selection in IGT1, we maintained residues R278gp120, N279gp120, and
P430gp120, while allowing amino acids R/N/K/S to be sampled at position 460. In addition, residues 461-
464 and 471 were allowed to be fully degenerate and sample all possible amino acids. Following 7
rounds of sorting, multiple clones were selected including IGT2, which bound to IOMA iGL with a 0.5
µM affinity (Figure 1D, right panel and Figure S1F, right panel).
IOMA-targeting mutations selected by yeast display were transferred onto a 426c soluble native-
like Env trimer (a SOSIP.664 construct (33)) to hide potentially immunodominant off-target epitopes
within the Env trimer core that are exposed in a monomeric gp120 protein. The SOSIP versions of IGT1
and IGT2 were well behaved in size-exclusion chromatography and SDS-PAGE (Figure S1G-H). IGT1
and IGT2 SOSIPs bound to IOMA iGL IgG with higher apparent affinities than IGT1 and IGT2 gp120s
due to avidity effects (Figure 1C-D). IGT1 and IGT2 SOSIP- and gp120-based immunogens were also
evaluated for binding to a panel of VRC01-class iGL antibodies (27) (VRC01, 3BNC60, BG24). IGT2
6
bound all the iGLs tested, making it the only reported immunogen that binds to iGLs from both IOMA-
and VRC01-class CD4bs bNAbs (Figure 1E, Figure S1I-J). Finally, using the SpyCatcher-SpyTag
system (51), we covalently linked our SpyTagged SOSIP-based immunogens to the designed 60-mer
nanoparticle SpyCatcher003-mi3 (52) (Figure 1A, 1F), thereby enhancing antigenicity and
immunogencity through avidity effects from multimerization (53, 54) (Figure S1I), while also reducing
the exposure of undesired epitopes at the base of soluble Env trimers (55-57). Efficient covalent
coupling of the immunogens to SpyCatcher003-mi3 was demonstrated by SDS-PAGE (Figure S1H),
and negative stain electron microscopy (EM) showed that these nanoparticles were densely conjugated
and uniform in size and shape (Figure 1F).
Sequential immunization of transgenic IOMA iGL knock-in mice elicits broad heterologous
neutralizing serum responses
To evaluate whether our immunogens induced IOMA-like antibody responses, we generated
transgenic mouse models expressing the full, rearranged IOMA iGL VH or VL genes in the mouse Igh
(IghIOMAiGL) and Igk loci (IgkIOMAiGL) (Figure S2A-B). Mice homozygous for both chains, termed IOMAgl
mice, showed overall normal B cell development with reduced numbers of pre-B cells and late
upregulation of CD2 (suggesting accelerated B cell development due to the already rearranged VDJ
and VJ genes), a preference for the IOMA iGL Igκ as seen by a reduction of mouse Igλ-expressing
cells, and a reduction in IgD expression indicative of low autoreactivity (58) (Figure S2C-H). Total B cell
numbers in IOMAgl mice were grossly normal, making them suitable to test IOMA germline-targeting
immunogens (Figure S2D, S2G).
We primed the IOMAgl mice using mi3 nanoparticles coupled with the SOSIP version of the
immunogen with the highest affinity to IOMA iGL (Figure 2A, IGT2-mi3) adjuvanted with the SMNP
adjuvant (59), and compared binding by ELISA to IGT2 and a CD4bs knockout mutant IGT2 (CD4bs
KO: G366R/D368R/D279N/A281T). Priming the IOMAgl mice with IGT2-mi3 elicited only weak
responses to the priming and boosting (IGT1-mi3) immunogens (Figure 2B-C). However, boosting with
mi3 nanoparticles coupled with IGT1, which bound IOMA iGL with a lower affinity than IGT2 (Figure
1C, middle panel), increased the magnitude and specificity of the serum responses, as demonstrated
by an increase in binding to IGT2 and IGT1 compared to IGT2- and IGT1-CD4bs KO (Figure 2B-C). A
comparable level of differential binding was preserved throughout the remaining immunizations (Group
1) following boosting with 426c degly2 D279N (degly2: removal of N460gp120 and N462gp120 PNGSs)
followed by mosaic8-mi3, a nanoparticle coupled with 8 different wt SOSIPs chosen from a global HIV-
1 reference panel used to screen bNAbs (60) (Table S1). Serum binding also increased throughout the
7
immunization regimen for 426c and 426c D279N, a mutation preferred by IOMA, compared to 426c-
CD4bs KO (Figure 2D). Terminal bleed sera showed binding to a panel of heterologous wt and N276A
Env SOSIPs in (Figure 2E-F) and when screened against a panel of IOMA-sensitive HIV-1 strains, 8 of
12 IOMA iGL knock-in animals neutralized up to 9 of 15 strains (Figure 2G, Figure S3A-M, Table S3).
However, one of these mice (ET34) also neutralized the MuLV control virus, suggesting that the
neutralization activity from this mouse is at least partially non-specific for HIV.
To determine whether a shorter immunization regimen could elicit heterologous neutralizing
responses, we tested 7 other immunization regimens in IOMA iGL knock-in mice (Figure S4, groups 2-
8). ELISA binding titers against 426c degly2 and 426c SOSIPs using serum from group 1, which was
primed with IGT2-mi3 and sequentially boosted with IGT1-mi3, 426c degly2 D279N-mi3, and mosaic8-
mi3, were significantly higher than binding titers from the other groups (p: <0.0001) (Figure S4B). These
results demonstrate the requirement for germline targeting through sequential immunization to induce
IOMA-like antibodies.
bNAbs isolated from IOMA iGL knock-in mice
To analyze immunization-induced antibodies, we isolated B cells from spleen and mesenteric lymph
nodes of three IOMA iGL knock-in mice of group 1 (ES30, HP1, and HP3) following the final boost
(Week 18 or 23, Figure 2G). We sorted immunization-induced germinal center B cells or used antigen-
bait combinations of 426c degly2 D279N or CNE8 N276A together with 426c degly2 D279N-CD4bs KO
(Table S1) to sort epitope-specific B cells (Figure S5). Among the identified HC and LC sequences, we
noted a correlation (R2 = 0.78 for HCs and R2 = 0.62 for LCs) between the total number of V region
amino acid mutations and V region mutations with identical or chemical similarity to IOMA. We
compared this to unbiased VH1-2*01 or VL2-23*02 sequences derived from peripheral blood of HIV-
negative human donors which showed both a lower rate and correlation (R2 = 0.52 for HCs and R2 =
0.55 for LCs) of IOMA-like mutations indicating that the immunization regimen induced maturation of
IOMA iGL towards IOMA (61), particularly of the heavy chain which constitutes the majority of contact
surface between IOMA and Env-based immunogens (34) (Figure 3A). 55 paired sequences were
selected for antibody production based on mutation load and similarity to mature IOMA (Figure S6). In
addition, 10x Genomics VDJ analysis of germinal center B cells revealed 5207 paired HC and LC
sequences, of which another 12 were chosen for recombinant antibody production (Figure S5, S6, S7A-
B).
The selected monoclonal antibodies were tested for binding to a panel of heterologous Envs by
ELISA (Figure S8A). Isolated IOMA-like antibodies that demonstrated binding to the Envs were then
8
evaluated in pseudotyped in vitro neutralization assays (62), and several exhibited similar neutralization
potencies as mature IOMA on a small panel of heterologous HIV-1 strains. Some antibodies neutralized
the tier 2 strain 25710, which IOMA does not neutralize, and IO-010 neutralized Q842.D12 better than
IOMA (Figure 3B). We also noted that among the Env-binding monoclonal antibodies, stronger
neutralization activity tended to occur with antibodies that shared a larger number of critical residues
with IOMA (Figure 3C).
Two mature IOMA residues, CDRH2 residues F53HC and R54HC, interact with the CD4bs Phe43
binding pocket (63) on gp120 and are critical for Env recognition (34) (Figure 3D-G, Figure S6). 29 of
67 clones chosen for antibody production (Figure S6) contained both mutations and another 15
contained R54HC, 5 of which in combination with Y53HC, which is chemically similar to F53HC. N53FHC
is a rare mutation that is found in only ~0.13% of VH1-2*02-derived antibodies (64, 65). In contrast, our
immunization regimen elicited this mutation in ~45% of antibodies, a ~350-fold increase. S54R is
elicited at slightly higher frequencies in VH1-2*02-derived antibodies (~2.7%). However, our
immunization regimen elicited this mutation at an ~24-fold higher rate compared to the random
frequency of this mutation in VH1-2*02-derived antibodies (Table S4) (64). In addition, our sequential
immunization regimen selected for a negatively-charged DDE motif in CDRH3 (replacing the IOMA
sequence of S100, A100A and D110B) in 23 of 63 sequences and another 27 sequences with at least
1 of the 3 mutations, which was likely selected for by a highly conserved patch of positively-charged
residues found at the IOMA-contacting interface of the Envs used in our immunization regimen [K97gp120
(90% conserved), R476 gp120 (R - 64% conserved, R/K - 98% conserved), and R480gp120 (99%
conserved)] (Figure 3D-G, Figure S6). To accommodate the N276gp120 glycan, IOMA acquired 3
mutations in CDRL1 (S29GLC, Y30FLC, N31DLC). The group 1 immunization regimen elicited all 3 of
these substitutions; however, none of the clones contained all these mutations. Of 63 antibodies 7
contained two and another 25 contained one of these mutations (Figure 3D-E, Figure S6). Two of the
most potent antibodies elicited by our immunization regimen, IO-010 and IO-017, acquired the
S31Ggp120 mutation, suggesting this mutation is more critical to accommodate the N276gp120 glycan
(Figure 3D-G) and generating antibody breadth and potency. While accommodation of the N276 glycan
is critical for CD4bs bNAbs to develop breadth and potency, CD4bs bNAbs must also acquire mutations
to better interact with the N197 glycan, such as K19THC in FR1. Our immunization strategy elicited the
K19THC mutation in 31 of 67 monoclonal antibodies (~46%), which is ~20-fold higher compared to the
random frequency of this mutation in VH1-2*02-derived antibodies (Table S4) (64). Within the CDRL3,
VRC01-class bNAbs acquire a G96ELC mutation that enables interactions with the CD4bs loop, while
IOMA acquires a similar G95DLC mutation. Once again, this mutation was elicited in 22 of 67 antibodies
9
(~33%) by our immunization regimen (Figure 3D-G, Figure S6). An essential interaction of VRC01-
class bNAbs involves the germline-encoded N58 residue in FR3HC, which makes backbone contacts to
the highly (~95%) conserved R456gp120. Due to a shift away from gp120 in CDRL2, IOMA acquires an
N58KHC substitution such that the longer lysine sidechain can access R456gp120 (34). Our immunization
regimen elicited substitutions at N58HC to amino acids with longer sidechains in 39 of 67 (~58%)
antibodies and was mutated to N58KHC in 17 of 67 (~24%) antibodies, a ~1.5-fold increase over the
random frequency of the N58KHC mutation (Figure 3D-G, Figure S6, Table S4). Our immunization
regimen elicited additional IOMA-like mutations within CDRH2: G56AHC (~30%) and T57VHC (~31%),
~3-fold and ~74-fold increases over the random frequency in other VH1-2*02-derived antibodies (Table
S4). The 10x Genomics VDJ analysis produced an unbiased view of the extent of SHM elicited in the
germinal center over the course of the immunization regimen, which, excluding frame shifts, reached
up to 26 amino acid mutations in the HC exceeding the number of mutations of the IOMA HC and up
to 10 mutations in the LC (Figure S7C-E).
Sera from prime-boosted wt mice targeted the CD4bs and displayed heterologous neutralizing
activity
We next investigated the same immunization regimen in wt mice (Figure 4A). Since IOMA does not
have the same sequence requirements as VRC01-class bNAbs (34), we hypothesized that a prime-
boost with IGT2-IGT1 could induce IOMA-like antibodies (which we define as recognizing the CD4bs
and including a normal-length CDRL3 (34)) in wt mice, even though these mice do not contain the VH1-
2 germline gene segment. Priming with IGT2-mi3 in wt mice elicited strong serum binding responses
that were CD4bs-specific (Figure 4B, p ≤ 0.05), compared to the IOMA iGL knock-in mice, which only
responded robustly after boost with IGT1-mi3 (Figure 2B-C). As in the IOMA iGL knock-in mice, the
magnitude of these responses increased after boosting with IGT1-mi3, and importantly, a significant
fraction of the response was still epitope-specific (p ≤ 0.001). To characterize antibodies in immunized
serum, we measured binding to anti-idiotypic monoclonal antibodies raised against IOMA iGL. While
naïve serum did not react with either of the anti-idiotypic antibodies, priming with IGT2-mi3 elicited
serum responses that bound both anti-idiotypic antibodies, and boosting with IGT1-mi3 increased these
responses (Figure 4C). After further boosting with 426c-mi3 and mosaic8-mi3 (Figure 4A), we
measured binding to heterologous wt Envs. Our immunization regimen elicited significantly increased
binding responses to all 9 Envs (Figure 4D-E, p ≤ 0.05 to < 0.001) in the majority of mice. Importantly,
serum binding to CNE8 N276Agp120 and CNE20 N276Agp120 was significantly higher compared to CNE8
and CNE20 (p ≤ 0.05), suggesting that these responses were at least partially specific to the CD4bs
10
(Figure 4E). Finally, we evaluated neutralization activity against a panel of heterologous HIV-1 strains
and detected weak heterologous neutralization in the serum of 7 of 16 wt animals (Figure 4F, Figure
S3 N-X, Table S5).
Immunization of rabbits and rhesus macaques elicited CD4bs-specific responses
To evaluate this immunization regimen in other wt animals with more potential relevance to humans,
we started by immunizing rabbits and rhesus macaques with IGT2-mi3 followed by IGT1-mi3 (Figure
5A). For these experiments, we assayed only for binding antibody responses since we did not achieve
heterologous neutralization after a prime or a prime/single boost of a different HIV-1 immunogen in
rabbits or non-human primates (NHPs) (57, 66). As with the wt mouse immunizations, the IGT2-mi3
immunization elicited robust responses that were partially epitope-specific as evaluated by comparing
binding to IGT1 versus to IGT1-CD4bs KO (Figure 5B). When boosted with IGT1-mi3, the responses
showed significant increases in epitope specificity to the CD4bs in both rabbits and non-human
primates (NHPs) (p ≤ 0.05) (Figure 5B). In addition, post-prime and post-boost sera exhibited potent
neutralization of pseudoviruses generated from the IGT2 and IGT1 immunogens (Figure 5C). As stated
above, we did not evaluate neutralization of heterologous pseudoviruses since our previous results
using a different HIV-1 immunogen in rabbits and NHPs showed heterologous neutralization only after
a second boost (57). The increase in epitope specificity and serum neutralization titers following
boosting with IGT1 suggests that our immunization strategy is well optimized to elicit CD4bs antibody
responses.
Discussion
Here we describe an immunization regimen to elicit antibodies to the CD4bs epitope on HIV Env
using engineered immunogens targeting IOMA-like CD4bs antibody precursors. The ultimate goal of
the germline-targeting approach is the induction of bNAbs at protective concentrations (1), but to date,
no study has been able to accomplish this feat, although a recent study involving mRNA delivery of
HIV-1 Env and gag genes reported reduced risk of SHIV infection in immunized NHPs (67). A previous
study using a transgenic mouse expressing diverse VRC01 germline precursors demonstrated that
priming with eOD-GT8 followed by sequential boosting with more native-like Envs elicited VRC01-like
bNAbs (38). However, that study required 9 immunizations over 81 weeks to elicit VRC01-class
antibodies with heterologous neutralization. By comparison, our study elicited bNAbs with similar
breadth and potency using only 4-5 immunizations in 18-23 weeks. In addition, sequence analysis of
the monoclonal antibodies elicited in the IOMA iGL transgenic mice revealed that our immunization
11
regimen was much more efficient at eliciting critical mutations required for bNAb development
compared to the immunogens used in the attempts to elicit VRC01-class bNAbs (38). Finally, the
neutralization profiles of monoclonal antibodies often correlated with serum neutralization from the
mouse they were isolated from. For example, IO-010 and IO-017, which neutralized PVO.4 and Q23.17,
were isolated from HP3 and HP1, whose serum also demonstrated neutralization activity against these
strains (Figure 2G and Figure S3A-M).
Accommodation of the N276gp120 glycan is considered the major impediment to the elicitation of
bNAbs targeting the CD4bs (32). To accommodate the N276 gp120 glycan, VRC01-class bNAbs require
a 2 - 6 residue deletion or the selection of multiple glycines within CDRL1 (32). IOMA requires simpler
substitution of 4 residues in CDRL1 (S27ARLC, S29GLC, Y30FLC, N31DLC) (34). These mutations were
elicited in our immunization regimen, although no single clone contained all 4 of these residues. The
two most potent monoclonal antibodies isolated from immunized iGL mice, IO-010 and IO-017,
contained the S31G mutation, suggesting this residue is most critical for accommodating the N276gp120
glycan in IOMA-like antibodies and to the development of bNAbs capable of potent heterologous
neutralization. Although these antibodies were cloned from mice following the 4th or 5th immunization,
sera from week 8 of our immunization regimen displayed significant binding to 426c Envs containing
the N276gp120 glycan (Figure 2D), suggesting these mutations were elicited following only two
immunizations. In contrast, in the same study noted above (38), mutations within CDRL1 of VRC01
required to accommodate the N276gp120 glycan occurred only after the ninth immunization at 81 weeks
(Figure S8B). Additional mutations known to be important for binding to the CD4bs were also elicited
earlier and at higher efficiencies in our immunization regimen compared to previous studies (Figure
S8B). Importantly, no other reported vaccination regimen to elicit CD4bs antibodies has elicited all of
the required SHMs to accommodate the N276gp120 glycan (23, 37, 38, 68-70), making our results an
important achievement in the pursuit to elicit CD4bs bNAbs, although these mutations need to be
elicited more efficiently and at higher frequencies in a protective vaccine. Since the CDRL1 of IOMA
iGL was already in a helical conformation, the CDRL1 of the IOMA precursor cells selected by priming
and boosting with IGT2 and IGT1 might have been in a conformation that allowed it to accommodate
the N276gp120 glycan and therefore not required additional SHMs to accommodate the N276gp120 glycan
introduced in the third immunization using 426c. Thus, boosting with Envs that incorporate only high-
mannose glycans at N276gp120 followed by boosting with Envs that only incorporate complex-type
glycans at N276gp120 starting at the second or third immunizations might force IOMA precursor cells to
adapt to more diverse and branching glycan moieties and acquire these critical SHMs.
12
Utilizing the strategy that we developed in IOMA iGL knock-in mice, we immunized wt mice with the
same immunization regimen (Figure 4A). A prime-boost sequence with IGT2-mi3 (prime) and IGT1-mi3
(boost) elicited robust CD4bs-specific responses. Importantly, the antibodies elicited by these
immunogens resembled IOMA based on binding to an anti-idiotypic antibody raised against IOMA iGL
using previously described methods (71, 72). Subsequent immunization with more native-like Envs,
426c degly2 and mosaic8, generated serum responses capable of neutralizing heterologous HIV
strains. Importantly, serum neutralization correlated with ELISA binding titers; e.g., mice that elicited
the highest serum binding titers against CNE8 (M21, M28, and M29) also elicited heterologous
neutralizing activity against CNE8 pseudovirus. To our knowledge, these results represent the first time
CD4bs-specific responses and heterologous neutralization were elicited in wt mice, thereby setting a
new standard by which to evaluate HIV immunogens in wt mice, although additional work is required
to determine whether antibodies targeting the CD4bs were responsible for the neutralizing responses.
Due to the success of our immunogens in wt mice, we tested them in additional animals with polyclonal
antibody repertoires - rabbits and rhesus macaques. Once again, our priming immunogens elicited
CD4bs-specifc responses in both animal models, representing the first time a germline-targeting
immunogen designed to target CD4bs Abs elicited epitope-specific responses in rabbits and rhesus
macaques.
As a final boost, we used a mosaic8 nanoparticle presenting eight different wt Envs on the surface,
with the intention of more efficiently selecting cross-reactive B cells and increasing neutralization
breadth, a strategy that was employed to elicit cross-neutralizing responses to influenza or to zoonotic
coronaviruses of potential pandemic interest (73, 74). Indeed, serum isolated from both wt and
transgenic mice after a mosaic8-mi3 boost bound to heterologous Envs in ELISAs and neutralized a
panel of heterologous HIV pseudoviruses, although additional experiments need to be completed to
determine whether the cross-neutralization was due to boosting with mosaic8-mi3.
Although a previous study suggested using gp120 cores as an important intermediate immunization
step (38), our approach resulted in heterologously-neutralizing antibodies using trimeric SOSIP-based
Envs for all immunizations. This is an important distinction, since using trimeric Envs provides the
additional benefit of simultaneous targeting of multiple bNAb epitopes. Indeed, a protective HIV-1
vaccine will most likely require the elicitation of bNAbs to multiple epitopes to prevent escape from the
host immune response during early infection to enable clearing of the virus. Thus, our immunogens
provide a scaffold upon which to engineer other epitopes to initiate germline-targeting of additional
bNAb precursors.
13
IOMA’s relatively lower number of SHMs and normal-length CDRL3 (34) suggest that eliciting
IOMA-like bNAbs by vaccination might be easier to achieve, compared with eliciting VRC01-class
bNAbs. Indeed, the fact our IOMA immunogens elicited CD4bs-specific responses in four animal
models suggests that germline-targeting immunogens designed to elicit IOMA-like antibodies are an
attractive route to generate an HIV-1 vaccine, which is supported by our engineered immunogens
eliciting epitope-specific responses in wt animals and by a commonality of the mutations that were
induced across individual transgenic mice. Furthermore, IOMA-like bNAbs have been isolated from
multiple patients (34, 75), suggesting an immunization regimen targeting this class of bNAbs could be
universally effective in a global population. Although IOMA’s neutralization breadth is smaller than that
of other bNAbs, the fact that some vaccine-elicited IOMA-like antibodies neutralized strains that IOMA
neutralizes less potently or does not neutralize at all suggests that it is possible to create polyclonal
serum responses that include individual antibodies with more breadth than IOMA. If elicited at sufficient
levels, such antibodies could mediate protection from more strains than predicted by the original IOMA
antibody. This is an important property of a potential active vaccine, since clinical trials to evaluate
protection from HIV-1 infection by passive administration of VRC01 in humans demonstrated a lack of
protection from infection by HIV-1 strains against which the VRC01 exhibited weak in vitro potencies
(28). Although polyclonal antibodies raised against the CD4bs may be more protective than a single
administered monoclonal anti-CD4bs antibody, a successful HIV-1 vaccine will likely require broader
and more potent responses to the CD4bs and other epitopes on HIV-1 Env. Our results provide new
germline-targeting immunogens to build upon, demonstrate that IOMA-like precursors provide a new
starting point to elicit CD4bs bNAbs and suggest that eliciting this class of bNAbs should be further
pursued as a possible strategy to generate a protective HIV-1 vaccine.
Data and materials availability
The structure of IOMA iGL Fab is available in the Protein Data Bank under accession code 7TQG. 10x
Genomics VDJ sequencing data is available from Gene Expression Omnibus Accession number
GSE197951. All other data, mice and reagents used in this study are available from the corresponding
authors upon reasonable request. Antibody HC and LC genes were analyzed using our previously
described
IgPipeline
(76,
77).
The
code
for
the
IgPipeline
is
available
at
https://github.com/stratust/igpipeline/tree/igpipeline2_timepoint_v2.
Further
information
and
reasonable requests for reagents and resources should be directed to Pamela J. Bjorkman
(bjorkman@caltech.edu).
14
Acknowledgements
We thank J. Moore (Weill Cornell Medical College), R.W. Sanders and M.J. van Gils (Amsterdam UMC)
for SOSIP expression plasmids; M. Silva, M. B. Melo and D. J. Irvine (MIT) for providing SMNP adjuvant;
Lotta von Boehmer (Stanford University) for discussion; J. Vielmetter, P. Hoffman, and the Protein
Expression Center in the Beckman Institute at Caltech for expression assistance; T. Eisenreich and S.
Tittley for animal husbandry and K. Gordon and K. Chosphel for fluorescence-activated cell sorting at
Rockefeller University. Electron microscopy was performed in the Caltech Cryo-EM Center with
assistance from S. Chen and A. Malyutin. This work was supported by the National Institute of Allergy
and Infectious Diseases (NIAID) Grants HIVRAD P01 AI100148 (to P.J.B. and M.C.N.), HIVRAD P01
AI138212 (to L.S., A.T.M., and M.N.), and R21 AI127249 (to A.T.M.), the Bill and Melinda Gates
Foundation Collaboration for AIDS Vaccine Discovery (CAVD) grant INV-002143 (P.J.B., M.C.N., and
M.A.M.), a Bill and Melinda Gates Foundation grant # OPP1146996 (to M.S.S.), the Intramural
Research Program of the NIAID (to M.A.M. and Y.N.), and NIH P50 AI150464 (P.J.B.). A.T.D. and
M.E.A. were supported by NSF Graduate Research Fellowships. M.C.N. is an HHMI investigator. Under
the grant conditions of the Bill and Melinda Gates Foundation Collaboration, a Creative Commons
Attribution 4.0 Generic License has already been assigned to the Author Accepted Manuscript version
that might arise from this submission.
METHODS
Antibody, gp120, and Env trimer expression and purification
Env immunogens were expressed as soluble SOSIP.664 native-like gp140 trimers (33) as described
(66). For SpyTagged trimers, either SpyTag (13 residues) (78) or SpyTag003 (16 residues) (79) was
added to the C-terminus to allow formation of an irreversible isopeptide bond to SpyCatcher003
moieties. All soluble SOSIP Envs were expressed by transient transfection in HEK293-6E cells
(National Research Council of Canada) or Expi293 cells (Life Technologies) and purified from
transfected cell supernatants by 2G12 affinity chromatography. Soluble Envs were stored at 4˚C in 20
mM Tris pH 8.0, 150 mM sodium chloride (TBS) (untagged and AviTagged versions) or 20 mM sodium
phosphate pH 7.5, 150 mM NaCl (PBS) (SpyTagged versions). We also expressed untagged gp120
proteins as cores with N/C termini and V1/V2/V3 loop truncations as described (63) by transient
transfection of suspension-adapted HEK293-S cells. gp120s were purified using Ni-NTA affinity
chromatography and Superdex 200 16/60 SEC. Proteins were stored in 20 mM Tris, pH 8.0, 150 mM
sodium chloride.
15
The iGL sequences of IOMA was derived as described in the main text. The iGL sequences of
VRC01 and 3BNC60 were derived as described (27, 80). The iGL of BG24, a VRC01-class bNAb with
relatively few SHMs (81), was derived as described (82). IgGs were expressed by transient transfection
in Expi293 cells or HEK293-6E cells and purified from cell supernatants using MabSelect SURE
(Cytiva) columns followed by SEC purification using a 10/300 or 16/600 Superdex 200 (GE Healthcare)
column equilibrated with PBS (20 mM sodium phosphate pH 7.4, 150 mM NaCl). His-tagged Fabs were
prepared by transient transfection of truncated heavy chain genes encoding a C-terminal 6x-His tag
with a light chain expression vector and purified from supernatants using a 5 mL HisTrap colum (GE
Healthcare) followed by SEC as described above.
Generation of anti-idiotypic monoclonal antibodies
Mice were injected three times with purified IOMA iGL. 3 days after the final injection spleens were
harvested and used to generate hybridomas at the Fred Hutchinson Antibody Technology Center.
Hybridoma supernatants were initially screened against IOMA iGL to identify antigen-specific
hybridomas. Supernatants from positive wells were then screened against a panel of monoclonal
antibodies that included IOMA, IOMA iGL, and inferred germlines of other anti-HIV-1 antibodies that
served as isotype controls using a high throughput bead array. We identified two hybridomas of interest;
3D3, which bound specifically to IOMA iGL, and 3D7, which bound to IOMA and IOMA iGL, which were
subcloned from single cells. To produce recombinant anti-idiotypes, RNA was extracted from 1 ×
106 cells using the RNeasy kit (Qiagen), and the heavy and light chain sequences of the murine
hybridomas were by obtained using the mouse Ig-primer set (69831; EMD Millipore) as described (83).
Sequences were codon optimized, cloned into pTT3-based IgG expression vectors with human
constant regions (84) using In-Fusion cloning (Clontech), expressed in 293 cells, and purified using
Protein A chromatography.
X-ray crystallography
Crystallization screens for IOMA iGL Fab were performed using the sitting drop vapor diffusion
method at room temperature (RT) by mixing 0.2 µL Fabs with 0.2 µL of reservoir solution (Hampton
Research) using a TTP Labtech Mosquito automatic microliter pipetting robot. IOMA iGL Fab crystals
were obtained in 20% (v/v) PEG 2000, 0.1 M Sodium Acetate (pH 4.6). Crystals were looped and
cryopreserved in reservoir solution supplemented with 20% glycerol and flash frozen in liquid nitrogen.
The crystal structure of IOMA iGL Fab was solved with data sets. A 1.9 Å-resolution structure of
IOMA – 10-1074 – BG505 was solved with a single data set collected at 100 K and 1 Å resolution on
16
Beamline 12-2 at the Stanford Synchrotron Radiation Lightsource (SSRL) with a Pilatus 6M pixel
detector (Dectris) that was indexed and integrated with iMosflm v7.4, and then merged with AIMLESS
in the CCP4 software package v7.1.018. The structure was determined by molecular replacement using
Phaser with one copy of IOMA Fab (PDB 5T3Z). Coordinates were refined with PHENIX v1.19.2-4158
(85) with group B factor and TLS restraints. Manual rebuilding was performed iteratively with Coot
v1.0.0 (86). Data refinement statistics are shown in Table S2, with > 98% of the residues in the favored
region of the Ramachandran plot and < 1% in the disallowed regions.
Cloning yeast libraries
Crystal structures of IOMA in complex with BG505 SOSIP.664 (PDB ID 5T3X and 5T3Z) were
analyzed to determine mutations on gp120 that potentially could be beneficial for IOMA iGL binding. In
addition, we modeled the crystal structure of IOMA iGL (PDB ID 7TQG) onto 426c.TM4ΔV1-3
(426c.TM4) gp120 (PDB ID 5FEC) and selected positions within gp120 that we predicted to be
favorable for IOMA iGL binding. We chose 426c.TM4ΔV1-3 (426c TM4), an engineered clade C Env
previously shown to activate B cell precursors of HIV-1 bNAbs targeting the CD4bs (25) as the starting
point for our library design.
Yeast libraries were generated as described (87). Specifically, to generate the libraries of 426c
gp120 variants we used degenerate oligos in conjunction with an overlap assembly polymerase chain
reaction (PCR) method. Overlapping primers for the PCR assembly reactions were designed using
Primerize (88) and shown in Table S6. NNK codons (where N = A/C/G/T and K = G/T) were utilized
that encode for all 20 amino acids but decrease the chances of introducing a premature stop codon.
Two different DNA fragments (426c library fragment 1 and 2) were synthesized first and then linearized
in a final PCR step to generate the full-length 426c gp120 library used in yeast transformation. To obtain
the full-length 426c gp120, a final PCR reaction was performed in which the PCR products of the 426c
Library Fragment 1 and 2 were used as a template. Primers were used with overhangs complementary
to the yeast display vector pCTCON-2 necessary for the homologous recombination in yeast. Library 2
was cloned in a similar manner as Library 1, but using a different set of primers as shown in Table S6
based on results from Library 1.
Yeast transformation
The yeast display vector pCTCON-2 was used for cell surface display of the 426c gp120 proteins
in Saccharomyces cerevisiae (S. cerevisiae) strain EBY100. A primary culture of 5 mL 2x YPD (40 g/L
glucose, 20 g/L peptone, 20 g/L yeast extract) media was inoculated with a single S. cerevisiae EBY100
17
colony (freshly streaked on a YPD plate) and incubated overnight in a shaker at 30 °C and 250 rpm.
100 μL of the overnight yeast S. cerevisiae EBY100 cultures was transferred into 5 mL 2x YPD media
and incubated overnight at 30 °C, 250 rpm. The following day, 300 mL 2x YPD media was inoculated
with the overnight precultures to an OD600 ~0.3 and was grown until an OD600 ~1.6. 3 mL of sterile
filtered Tris/DTT (0.462 g 1,4-dithiothreitol in 3 mL 1 M Tris, pH 8.0) and 15 mL sterile filtered 2 M
LiAc/TE (1.98 g LiAc in 10 mL of TE (10 mM Tris, 1 mM EDTA) was added and the culture incubated
for 15 min at 30 °C and 250 rpm. Yeast cells were then pelleted at 3,500 g for 3 min and washed with
50 mL ice-cold sterile filtered NewE buffer (0.6 g Tris base, 91.09 g Sorbitol (1 M), 73.50 mg CaCl2 in
ddH2O to a final volume of 500 mL, pH 7.5). After two additional wash steps, the pellet was re-
suspended in 3 mL NewE buffer and 50 μg 426c library DNA insert and 10 μg pCTCON-2 vector
(digested with NheI and BamHI) was added. 200 μL of this transformation mix was then aliquoted into
pre-chilled 2 mm electroporation cuvettes (Bio-Rad) and electroporated at 1500 V with an average time
constant of ~4.5 ms using a Gene Pulser Xcell Electroporation System (Bio-Rad), which was repeated
for the entire transformation mix. After electroporation, yeast cells were directly recovered with 2 mL
2x YPD media and transferred into 50 mL cold 2x YPD media (final volume up to 200 mL 2x YPD
media) and grown for 1 h at 30 °C and 250 rpm. Serial dilutions of the freshly transformed yeast culture
were plated on SDCAA (20 g/L glucose, 6.7 g/L Difco yeast nitrogen base, 1.4 g/L Yeast Synthetic
Drop-out Medium Supplements without histidine, leucine, tryptophan and uracil, 20 mg/L uracil,
50 mg/L histidine, 100 mg/L leucine) agarose plates to test the viability and size of the library. After 1 h,
the culture was removed and the cells were pelleted and resuspended in 500 mL SDCAA media +
carbenicillin (100 μg/mL final concentration) and grown for two days at 30 °C and 250 rpm. To confirm
the genetic diversity of the library, a yeast colony PCR was performed on the liquid culture and the PCR
product was sequenced. Sequencing reactions were performed at Laragen Inc (Culver City, CA). The
sequence data was analyzed using SeqMan Pro (DNASTAR, v13.02). After two days, cells were
pelleted and glycerol stocks were made by suspending ~109 yeast cells in 1 mL of freezing buffer
(0.335 g Yeast Nitrogen Base, 1 mL glycerol in 50 mL H2O, sterilized by filtration). Aliquots were flash
frozen in liquid nitrogen and stored at -80 °C.
Magnetic-activated cell sorting
Magnetic-activated cell sorting (MACS) was used to remove transformants containing stop codons.
After growing up the freshly transformed cells for two days in SDCAA, cells were pelleted and induced
at an OD600 ~1.0 in 100 mL SGCAA-carb (SDCAA prepared with 20 g/L galactose instead of glucose
and supplemented with 100 µg/mL carbenicillin final concentration) for 20 h at 20 °C and 250 rpm.
18
Yeast cells were washed 5 times with PBSF (PBS + 0.1% bovine serum albumin (BSA)) and 108 cells
were incubated with 400 μL PBSF and 100 μL μMACS™ anti-c-Myc MicroBeads (Miltenyi Biotec) for
45 min on a rotator at 4 °C. Cells were then pelleted and resuspended in 5 mL PBSF and sorted using
a MidiMACS Separator magnet (Miltenyi Biotec) in combination with an LS column (Miltenyi Biotec)
equilibrated in PBSF. Isolated cells were then grown for 2 days in 100 mL SDCAA-carb at 30 °C and
250 rpm and then induced again with SGCAA-carb for 20 h at 20 °C and 250 rpm.
Yeast flow cytometry and cell sorting
To prepare the yeast library for FACS analysis, cells were pelleted at 3000 rpm for 2 min and
washed 5 times with PBSF. Cells were then stained at a density of 107 cells/mL with 1:500 anti-c-Myc
antibody conjugated to AlexaFluor488 (Abcam, ab190026) and 1 µM IOMA iGL and incubated for
1 – 2 h on a rotator at 4 °C. Cells were then washed twice with PBSF and resuspended in 200 μL PBSF
with 1:1000 goat anti-human antibody conjugated to AlexaFluor647 (Abcam, ab190560) and incubated
for 30 min at 4 °C. Cells were then analyzed on a MACSQuant Analyzer (Miltenyi Biotec) or sorted
using an SY3200 cell sorter system (Sony). In either case, non-transformed yeast cells and single-
stained transformed samples stained with either anti-cMyc or IOMA iGL IgG were used to set the gates
for analysis and collection. Cells that stained double-positive for both c-Myc and IOMA iGL were
collected and grown in 5 mL SDCAA-carb for 1 - 2 days at 30 °C and 250 rpm and then transferred to
100 mL SDCAA-carb for an additional 1 - 2 days at 30 °C and 250 rpm. Cells were then pelleted and
resuspended in H2O and plated onto SDCAA-carb for 2 - 3 days at 30 °C. After multiple iterative rounds
of sorting (three rounds for Library 1 and seven rounds for Library 2), sequences were recovered by
colony PCR and sequence confirmed (Laragen). Primers were used with specific complementary
regions to enable ligation of the linear product into the expression vector pTT5 using the Gibson
assembly method for protein production. After construction, plasmids were isolated from E.coli using
the QIAprep Miniprep kit (Qiagen) and confirmed by Sanger sequencing (Laragen).
ELISAs
Serum ELISAs were performed using randomly biotinylated SOSIP trimers using the EZ-Link NHS-
PEG4-Biotin kit (Thermo Fisher Scientific) according to the manufacturer’s guidelines. Based on the
Pierce Biotin Quantitation kit (Thermo Fisher Scientific), the number of biotin molecules per protomer
was estimated to be ~1 - 4. Biotinylated SOSIP timers were immobilized on Streptavidin-coated 96-well
plates (Thermo Fisher Scientific) at a concentration of 2 - 5 µg/mL in blocking buffer (1% BSA in TBS-
T: 20 mM Tris pH 8.0, 150 mM NaCl, 0.1% Tween 20) for 1 h at RT. After washing plates in TBS-T,
19
plates were incubated with a 3-fold concentration series of mouse, rabbit, or rhesus macaque serum at
a top dilution of 1:100 in blocking buffer for 2-3 h at RT. After washing plates with TBS-T, HRP-
conjugated goat anti-mouse Fc antibody (Southern Biotech, #1033-05) or HRP-conjugated goat anti-
rabbit IgG Fc antibody (Abcam, ab98467) or HRP-conjugated goat anti-human multi-species IgG
antibody (Southern Biotech, #2014-05) was added at a dilution of 1:8,000 in blocking buffer for 1 h at
RT. After washing plates with TBS-T, 1-Step Ultra TMB substrate (Thermo Fisher Scientific) was added
for ~3 min. Reactions were quenched by addition of 1 N HCl and absorbance at 450 nm were analyzed
using a plate reader (BioTek). ELISAs with gp120s and anti-idiotype monoclonal antibodies were
performed as above except these proteins were immobilized directly onto high-binding 96-well assay
plates (Costar) in 0.1 M sodium bicarbonate buffer (pH 9.8) at a concentration of 2 – 5 µg/mL in
blocking buffer (1% BSA in TBS-T) for 2 h at RT. ELISAs with IgGs instead of serum were performed
as above with a top IgG concentration of 100 µg/mL. All reported values represent the average of at
least two independent experiments.
SPR binding studies
All SPR measurements were performed on a Biacore T200 (GE Healthcare) at 20 °C in HBS-EP+
(GE Healthcare) running buffer. IgGs were directly immobilized onto a CM5 chip (GE Healthcare) to
~3000 resonance units (RUs) using primary amine chemistry. A concentration series of monomeric
gp120 core constructs (IGT2, IGT1, 426c TM4) were injected over the flow cells at increasing
concentrations (top concentrations ranging from 600 µM to 10 µM) at a flow rate of 60 µL/min for 60 s
and allowed to dissociate for 300 s. Regeneration of flow cells was achieved by injecting one pulse
each of 10 mM glycine pH 2.0 at a flow rate of 90 µL/min. Kinetic analyses were used after subtraction
of reference curves to derive on/off rates (ka/kd) and binding constants (KDs) using a 1:1 binding model
with or without bulk refractive index change (RI) correction as appropriate (Biacore T200 Evaluation
software v3.0). Reported affinities represent the average of two independent experiments. SPR
experiments that were not used to derive binding affinities or kinetic constants were done using a single
high concentration (1 µM) to qualitatively determine binding versus no binding.
Preparation of SOSIP-mi3 nanoparticles
SpyCatcher003-mi3 particles were prepared by purification from BL21 (DE3)-RIPL E. coli (Agilent)
transformed with a pET28a SpyCatcher003-mi3 gene (89) (including an N-terminal 6x-His tag) as
described (74, 90). Briefly, cell pellets from transformed bacterial were lysed with a cell disruptor in the
presence of 2.0 mM PMSF (Sigma). Lysates were spun at 21,000 g for 30 min, filtered with a 0.2 µm
20
filter, and mi3 particles were isolated by ammonium sulfate precipitation followed by SEC purification
using a HiLoad 16/600 Superdex 200 (GE Healthcare) column equilibrated with 25 mM Tris-HCl pH 8.0,
150 mM NaCl, 0.02% NaN3 (TBS). SpyCatcher003-mi3 particles were stored at 4 °C and used for
conjugations for up to 1 month after filtering with a 0.2 µm filter and spinning for 30 min at 4 °C and
14,000 g.
Purified SpyCatcher003-mi3 was incubated with a 2-fold molar excess (SOSIP to mi3 subunit) of
purified SpyTagged SOSIP (either a single SOSIP or an equimolar mixture of eight SOSIPs for making
mosaic8 particles) overnight at RT in PBS. Conjugated SOSIP-mi3 particles were separated from free
SOSIPs by SEC on a Superose 6 10/300 column (GE Healthcare) equilibrated with PBS. Fractions
corresponding to conjugated mi3 particles were collected and analyzed by SDS-PAGE. Concentrations
of conjugated mi3 particles were determined using the absorbance at 280 nm as measured on a
Nanodrop spectrophotometer (Thermo Scientific).
Electron microscopy of SOSIP-mi3 nanoparticles
SOSIP-mi3 particles were characterized using negative stain electron microscopy (EM) to confirm
stability and the presence of conjugated SOSIPs on the mi3 surface. Briefly, SOSIP-mi3 particles were
diluted to 20 µg/mL in 20 mM Tris (pH 8.0), 150 mM NaCl and 3 µL of sample was applied onto freshly
glow-discharged 300-mesh copper grids. Sample was incubated on the grid for 40 s and excess sample
was then blotted away with filter paper (Whatman). 3 µL uranyl acetate was added for 40 s and excess
stain was then blotted off with filter paper. Prepared grids were imaged on a Talos Arctica
(ThermoFisher Scientific) transmission electron microscope at 200 keV using a Falcon III 4k × 4k
(ThermoFisher Scientific) direct electron detector at 13,500x magnification.
Generation of IOMA-expressing RAMOS cells by CRISPR/Cas9 gene editing
A targeting vector was constructed using the NEB Hifi DNA assembly kit to clone a gBlock (IDT)
into pUCmu (91). The gBlock (IDT) contained ~0.5 kb homology arms to the human IgH locus which
flanked an expression cassette consisting of the Cµ splice acceptor, the entire IOMA LC gene, a furin-
GSG-P2A sequence (92) followed by the IOMA HC Leader-VDJ and the JH4 splice donor based on
previously-described designs (93) (Figure S5C). Vectors were maxi-prepped (Machery-Nagle) for
transfection.
RAMOS (RA 1) cells were purchased from ATCC (CRL-1596) and maintained in RPMI-1640
supplemented with 10 FCS, 1x antibiotic/antimycotic, 2 mM glutamine, 1 mM sodium pyruvate, 10 mM
HEPES and 55 µM β-mercaptoethanol. Before transfection, cells were harvested, washed once in PBS
21
and resuspended at 6x107 cells/mL in Neon kit buffer T (ThermoFisher). Three ribonucleoprotein
complexes (RNPs) were prepared using 3 different sgRNAs. AGGCATCGGAAAATCCACAG was used
to target the IgH locus in the intron 3’ of IGHJ6 to integrate the sequence flanked by the appropriate
homology arms from the targeting vector; CTGGGAGTTACCCGATTGGA was used to ablate the
human IGKC exon and CACGCATGAAGGGAGCACCG was used to ablate all functional IGLC genes
(IgLC1, IGL2, IGLC3 and IGLC7). Complexes were prepared by mixing 1.875 µL of 100 µM sgRNA
with 1 µL of 61 µM Cas9 (all IDT) for a molar ratio of ~3:1 followed by incubation for 20 min at RT.
IGH:IGK:IGL RNPs were then mixed at a 2:1:1 v/v/v ratio. 2.6 µg targeting vector (at 4 mg/mL) were
mixed with 1.5 µL IGH:IGK:IGL RNP mix and 11 µL RAMOS cells in buffer T. 10 µL of the final mix
were transfected in a 10 µL Neon tip in a Neon device at 1350 V 30 ms 1 pulse. Cells were immediately
transferred into 50 µL RAMOS medium without 1x antibiotic/antimycotic in a 48-well plate and 2 h later
450 µL full RAMOS medium was added. Cells were then cultured as before. Edited IOMA-expressing
cells were bulk sorted by flow cytometry as live, singlet, CD19+, RC1 antigenhi, IgL+, IgK+ IgM+ (Table
S7) and cultured as before. IOMA-expression was further verified by staining with 426c-, CNE8- and
CNE20-derived SOSIPs and 426c-CD4bs-KO proteins to show specificity.
Mice
C57BL/6J and B6(Cg)-Tyrc-2J/J (B6 albino) mice were purchased from Jackson Laboratories.
IghIOMAiGL and IgkIOMAiGL mice were generated with the Rockefeller University CRISPR and Genome
Editing Center and Transgenic and Reproductive Technology Center in CY2.4 albino C57BL/6J-Tyrc-
2J-derived embryonic stem cells. Chimeras were crossed to B6(Cg)-Tyrc-2J/J for germline
transmission. IghIOMAiGL and IgkIOMAiGL mice carry the IGV(D)J genes encoding the IOMA iGL HC and
LC respectively. IOMA iGL LC was targeted into the Igk locus deleting the endogenous mouse Igkj1 to
Igkj5 gene segments. IOMA iGL HC was targeted into the Igh locus and deleting the endogenous
mouse Ighd4-1 to Ighj4 gene segments thereby minimizing rearrangement of the locus (Figure S2A-B)
(94, 95). The constant regions of Igh and Igk remain of mouse origin. Mice were only crossed to
C57BL/6J or B6(Cg)-Tyrc-2J/J or themselves and maintained at Rockefeller University and all
experiments shown used double homozygous animals for IghIOMAiGL and IgkIOMAiGL abbreviated IOMAgl
mice. These mice are available upon request. Mice were housed at a temperature of 22 °C and humidity
of 30 – 70% in a 12 h light/dark cycle with ad libitum access to food and water. Male and female mice
aged 6 – 12 weeks at the start of the experiment were used throughout. All experiments were
conducted with approval from the institutional review board and the institutional animal care and use
committee at the Rockefeller University. Sample sizes were not calculated a priori. Given the nature of
22
the comparisons, mice were not randomized into each experimental group and investigators were not
blinded to group allocation. Instead, experimental groups were age- and sex-matched.
Animal immunizations and sampling
Mice were immunized intraperitoneally with 10 µg conjugated mi3-SOSIP in 100 µL PBS with 1 U
SMNP adjuvant (59) (kindly provided by Murillo Silva, Mariane B. Melo and Darrell J. Irvine, MIT).
Serum samples were collected throughout the experiment by submandibular bleeding and animals
were terminally bled under isoflurane anesthesia first submandibularly followed by cardiac puncture.
Spleen and mesenteric lymph nodes were dissected, mashed though a 70 µm cell strainer and frozen
in FCS with 10% DMSO in a gradual freezing (~1 °C/min) container, followed by transfer to liquid N2
for long-term storage.
Eight six-month-old New Zealand White rabbits (LabCorp) were used for immunizations. Rabbits
were immunized subcutaneously with 50 µg of a SOSIP-mi3 in SMNP adjuvant (375 U/animal) as
described (66, 96). Serum samples were collected from rabbits at the time points indicated in Figure
5A. Procedures in rabbits were approved by the Denver PA IACUC Committee.
Five rhesus macaques (Macaca mulatta) of Indian genetic origin were housed in a biosafety level 2
NIAID facility and cared for in accordance with Guide for Care and Use of Laboratory Animals Report
number NIH 82-53 (Department of Health and Human Services, Bethesda, 1985). All animal
procedures and experiments were performed according to protocols approved by the IACUC of NIAID,
NIH. The NHPs used in this study did not express the MHC class I Mamu-A*01, Mamu-B*08 and Mamu-
B*17 alleles. NHPs were immunized subcutaneously in the medial inner forelegs and hind legs (total of
4 sites per animal) with 200 μg of the indicated SOSIP-mi3 adjuvated in SMNP (375 U/animal) as
described (66). Immunizations and blood samples were obtained from naïve and immunized macaques
at the time points indicated in Figure 5A.
Flow Cytometry and cell sorting
Fresh bone marrow was flushed out of 1 femur and 1 tibia per mouse. Fresh mouse spleens were
forced through a 70 µm mesh into FACS buffer (PBS containing 2% heat-inactivated FBS and 2 mM
EDTA), and red blood cells of fresh spleens or bone marrow were lysed in ammonium-chloride-
potassium buffer lysing buffer (Gibco) for 3 min. Frozen cells were thawed in a 37 °C water bath and
immediately transferred to prewarmed mouse B cell medium consisting of RPMI-1640, supplemented
with 10% heat-inactivated FBS, 10 mM HEPES, 1× antibiotic-antimycotic, 1 mM sodium pyruvate,
2 mM L-glutamine, and 53 µM 2-mercaptoethanol (all from Gibco).
23
Bait proteins were randomly conjugated to biotin and free biotin removed using EZ-Link Micro NHS-
PEG4-Biotinylation Kit (ThermoFisher # 21955) according to the manufacturer’s instructions.
Fluorophore conjugated bait and bait-KO antigen tetramers were prepared by mixing a 5 µg/mL solution
of a single randomly-biotinylated bait protein with fluorophore-conjugated streptavidin (Table S7) at a
1:200 to 1:600 dilution in PBS for 30 min on ice. Conjugates were then mixed equivolumetrically.
RAMOS cells were harvested, washed in FACS buffer and stained with human FC-blocking reagent,
biotinylated bait antigen-streptavidin tetramers (PE, AF647 and sometimes PECy7) and Zombie-NIR
Live/Dead cell marker for 15 min before addition of anti-human antibodies to IgL-APC, IgK-BV421, IgM-
FITC, and for some experiments, CD19-PECy7 (Table S7).
Mouse cells and controls (see below) were washed and resuspended in a solution of mouse Fc-
receptor blocking antibody, fluorophore-conjugated antigen tetramers and Zombie-NIR Live/Dead cell
marker for 15 min on ice. A mastermix of other antibodies was then added and cells stained for another
20 min on ice. Antibodies and reagents are listed in Table S7. All cells were analyzed on an
LSRFortessa or cells were sorted on a FACS Aria III (both Becton Dickinson) using IOMA-expressing
RAMOS cells as an antigen-binding positive control and splenocytes from naïve IOMAgl mice as
negative controls (Figure S5). To derive absolute cell numbers, a master mix of AccuCheck counting
beads (ThermoFisher #PCB100) in FACS buffer was prepared and 104 beads/sample were added
before acquisition. Absolute numbers of cells were calculated as:
[𝑐𝑜𝑢𝑛𝑡 𝑜𝑓 𝑎𝑐𝑞𝑢𝑖𝑟𝑒𝑑 𝑏𝑒𝑎𝑑𝑠]
[𝑡𝑜𝑡𝑎𝑙 𝑏𝑒𝑎𝑑𝑠 𝑝𝑒𝑟 𝑠𝑎𝑚𝑝𝑙𝑒] [𝑓𝑟𝑎𝑐𝑡𝑖𝑜𝑛 𝑜𝑓 𝑜𝑟𝑔𝑎𝑛 𝑢𝑠𝑒𝑑 𝑖𝑛 𝑠𝑡𝑎𝑖𝑛] × [𝑐𝑜𝑢𝑛𝑡 𝑜𝑓 𝑐𝑒𝑙𝑙 𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛]
IOMA-expressing RAMOS cells were separated from unedited cells by sorting into RAMOS medium,
and then washed and cultured as described above.
1838 single, mouse B cells from spleen and mesenteric lymph nodes of 3 IOMA iGL knock-in mice
(ES30, HP1, and HP3) following the final boost (Week 18 or 23) were sorted into individual wells of a
96-well plate containing 5 μL of lysis buffer (TCL buffer (Qiagen, 1031576) with 1% of 2-β-
mercaptoethanol). Plates were immediately frozen on dry ice and stored at −80 °C. Singlet, live Zombie-
NIR− CD4− CD8− F4/80− NK1.1− CD11b− CD11c− B220+ double Bait+ BaitKO− lymphocytes were sorted
unless GC B cells were sorted, which were gated as single, live Zombie-NIR− CD4− CD8− F4/80− NK1.1−
CD11b− CD11c− B220+ CD38− FAS+ lymphocytes (see Figure S5).
24
Mouse GC B cells for 10x Genomics single cell analysis were processed in PBS with 0.5% BSA
instead of FACS buffer and 31,450 cells sorted into 5 µL of 0.05% BSA in PBS. Cells were spun down
400 g 6 min at 4 °C and volume adjusted to 22 µL before further processing.
10x Genomics single cell processing and next generation V(D)J sequencing
Cells were counted in the final injection volume, and 18,000 cells loaded onto a Chromium Controller
(10x Genomics). Single-cell RNA-seq libraries were prepared using the Chromium Single Cell 5 v2
Reagent Kit (PN-1000265) according to manufacturer’s protocol. Chromium Single Cell Mouse BCR
Amplification Kit (PN-1000255) was used for VDJ cDNA amplification. After QC, 5’ expression and VDJ
Libraries were pooled 1:1 and sequenced on an Illumina NOVAseq S1 flowcell at the Rockefeller
University Genomics Core.
Computational Analyses of V(D)J sequences derived from IOMAgl mice by next generation
sequencing
The single-cell V(D)J assembly was carried out by Cell Ranger 6.0.1. A customized reference was
created by adding the knocked-in IOMA iGL V(D)J genes to the mouse GRCm38 V(D)J reference so
Cell Ranger could recognize and assemble the human/mouse chimera transcripts. Contigs associated
with a valid cell barcode according to Cell Ranger were selected for downstream processing using
seqtk version 1.3-r106 (https://github.com/lh3/seqtk).
IgBlast standalone version 1.14 (97) was used to annotate the immunoglobulin sequences based
on a custom database with mouse and human V(D)J genes. Productive IG sequences with more than
20 reads of coverage and with any identified isotype were selected for downstream processing.
Unexpectedly, although the IgBlast algorithm identified the V and J genes for 8010 LC sequences, it
failed to annotate the CDR3, and consequently, the information regarding their functionality was
missing. We extracted and submitted 7782 (97.15%) sequences corresponding to the knock-in LC to
IMGT/V-QUEST (98), which successfully identified the CDR3 and provided the productivity information.
Cell barcodes associated with sequences coded by different V genes for either HC or LC were
considered doublets and were subsequently removed from downstream analysis. HCs and LCs derived
from the same cell were paired, and clones were assigned using our previously-described IgPipeline
(76, 77) (https://github.com/stratust/igpipeline/tree/igpipeline2_timepoint_v2).
Single cell antibody cloning
25
Sequencing and cloning of mouse monoclonal antibodies from single cell-sorted B cells were performed
as described (99) with the following modifications. Briefly, single cell RNA in 96-well plates was purified
using magnetic beads (RNAClean XP, Beckman Coulter, Cat # A63987). RNA was eluted from the
magnetic beads with 11 μL of a solution containing 14.5 ng/μL of random primers (Invitrogen, Cat #
48190011), 0.5% of Igepal Ca-630 (type NP-40, 10% in dH2O, MP Biomedicals, Cat # 198596) and
0.6 U/μL of RNase inhibitor (Promega, Cat# N2615) in nuclease-free water (Qiagen, Cat # 129117),
and incubated at 65 °C for 3 min. cDNA was synthesized by reverse transcription (SuperScript™ III
Reverse Transcriptase 10,000 U, Invitrogen, Cat# 18080-044). cDNA was stored at −80 °C or used for
antibody gene amplification by nested polymerase chain reaction (PCR) after addition of 10 μL of
nuclease-free water.
Mouse antibody genes were amplified using HotstarTaq DNA polymerase (Qiagen Cat # 203209)
with the primer sets specific for the IghIOMAiGL and IgkIOMAiGL transgenes. Primer sequences and reaction
mixes are provided in Table S8. Thermocycler conditions were as follows for annealing (°C)/elongation
(s)/number of cycles: PCR1 (IgG, IgM and IgK): 51/55/50; PCR2 (IgG and IgM): 54/55/50; PCR2 (IgK):
50/55/50.
PCR products of antibody HC and LC genes were purified and Sanger-sequenced (Genewiz) and
*ab1
files
analyzed
using
our
previously
described
IgPipeline
(https://github.com/stratust/igpipeline/tree/igpipeline2_timepoint_v2) (76, 77). V(D)J sequences were
ordered as eBlocks (IDT) with short homologies for Gibson assembly and cloned into human IgG1 or
human IgL2 expression vectors using the NEB Hifi DNA Assembly mix (NEB, Cat#E2621L). Plasmid
sequences were verified by Sanger sequencing (Genewiz).
Mutation analysis
All HC and LC V(D)J sequences were translated and the CDR3 region was trimmed. The resulting
V region was aligned against the IOMA iGL and IOMA using MAFFT (100). Indels were ignored for
downstream analysis. All mismatches to IOMA iGL were counted as total mismatches (Figure 3A). Only
mismatches shared with IOMA mature when compared to IOMA iGL were used to assess chemical
equivalence and calculate IOMA-like mutations (Figure 3A). Chemical equivalence was as follows:
Group 1: G/A/V/L/I; Group 2: S/T; Group 3: C/M; Group 4: D/N/E/Q; Group 5: R/K/H; Group 6: F/Y/W;
Group 7: P. The baseline was calculated using extracted IGHV1-2, IGHJ5 (318,769) and IGLV2-23,
IGLJ2 (1,790,961) sequences from healthy, HIV-negative donors generated by Soto et al. (61) and
26
downloaded from cAb-Rep (64), a database of human shared BCR clonotypes available at https://cab-
rep.c2b2.columbia.edu/.
3D neutralization plot shows the total number of V(D)J amino acid mutations (untrimmed) of each
antibody vs the number of these mutations that are chemically equivalent to IOMA (Figure 3C).
Chemical equivalence defined as above.
In vitro neutralization assays
Pseudovirus neutralization assays were conducted as described (62, 101), either in house (Figure 2G,
Figure 3A, Figure 4F, Figure 5B) or at the Collaboration for AIDS Vaccine Discovery (CAVD) core
neutralization facility (Figure S1C). Monoclonal antibody IgGs were evaluated in duplicate with an 8-
point, 3-fold dilution series starting at a top concentration of ~100 µg/mL. All pseudovirus assays using
monoclonal antibody IgGs were repeated at least twice for each value reported here. For polyclonal
neutralizations, serum samples were heat inactivated at 56 ºC for 30 min before being added to the
neutralization assays, and then neutralization was evaluated in duplicate with an 8-point, 4-fold dilution
series starting at a dilution of 1:60. The percent of neutralization at a 1:100 dilution (% 1:100) are
reported for all serum samples. Tiers for viral strains were obtained from ref. (102) Antibody
neutralization score was calculated as
[𝑛𝑒𝑢𝑡𝑟𝑎𝑙𝑖𝑧𝑎𝑡𝑖𝑜𝑛 𝑠𝑐𝑜𝑟𝑒] = 9 5 − 𝑙𝑜𝑔!"(10# × [𝐼𝐶$"]%)
𝑛
&
%'!
were 𝑛 is number of different HIV pseudoviruses 𝑣 tested for that antibody and [𝐼𝐶$"]% is the IC50 of
pseudovirus 𝑣 in µg/mL.
Statistical analysis
Comparisons between groups for ELISAs and neutralization assays were calculated using an
unpaired or paired t-test in Prism 9.0 (Graphpad). Differences were considered significant when p
values were less than 0.05. Exact p values are in the relevant figure at the top of the plot, with asterisks
denoting level of significance (* denotes 0.01 < p ≤ 0.05, ** denotes 0.001 < p ≤ 0.01, *** denotes
0.0001 < p ≤ 0.001, and **** denotes p ≤ 0.0001). Comparisons between total amino acid mutations
and IOMA-like mutations in antibodies cloned from IOMA iGL mice (Figure 3) were performed using a
Pearson correlation and R2 values are presented.
Analysis Software
Unless stated otherwise, Geneious Prime 2021.2.2, MacVector 18.2.0 and DNAStar SeqMan Pro
17.1.1 were used for sequence analysis and graphs were created using R language. Flow cytometry
27
data were processed using Mac versions of FlowJo 10.7.2. and GraphPad Prism 9.3 and Microsoft
Excel for Mac 16.54 were used for data analysis. Structural figures were made using PyMOL
(Schrödinger, LLC) or ChimeraX (103). V(D)J gene assignments of NHP and murine antibodies were
done using IMGT/V-QUEST (98). Sequence alignments were done using Clustal Omega (104).
References
1.
Z. Mu, B. F. Haynes, D. W. Cain, Strategies for eliciting multiple lineages of broadly neutralizing
antibodies to HIV by vaccination. Current opinion in virology 51, 172-178 (2021).
2.
R. P. Smyth, M. P. Davenport, J. Mak, The origin of genetic diversity in HIV-1. Virus Res 169, 415-429
(2012).
3.
T. W. Baba et al., Human neutralizing monoclonal antibodies of the IgG1 subtype protect against
mucosal simian-human immunodeficiency virus infection. Nat Med 6, 200-206 (2000).
4.
A. B. Balazs et al., Antibody-based protection against HIV infection by vectored immunoprophylaxis.
Nature 48, 81-88 (2012).
5.
J. W. Eichberg, K. K. Murthy, R. H. Ward, A. M. Prince, Prevention of HIV infection by passive
immunization with HIVIG or CD4-IgG. AIDS Res Hum Retroviruses 8, 1515 (1992).
6.
E. A. Emini et al., Prevention of HIV-1 infection in chimpanzees by gp120 V3 domain-specific
monoclonal antibody. Nature 355, 728-730 (1992).
7.
A. J. Hessell et al., Broadly neutralizing monoclonal antibodies 2F5 and 4E10 directed against the
human immunodeficiency virus type 1 gp41 membrane-proximal external region protect against
mucosal challenge by simian-human immunodeficiency virus SHIVBa-L. J Virol 84, 1302-1313 (2010).
8.
A. J. Hessell et al., Effective, low-titer antibody protection against low-dose repeated mucosal SHIV
challenge in macaques. Nat Med 15, 951-954 (2009).
9.
A. J. Hessell et al., Broadly neutralizing human anti-HIV antibody 2G12 is effective in protection against
mucosal SHIV challenge even at low serum neutralizing titers. PLoS Pathog 5, e1000433 (2009).
10.
J. R. Mascola et al., Protection of macaques against vaginal transmission of a pathogenic HIV-1/SIV
chimeric virus by passive infusion of neutralizing antibodies. Nat Med 6, 207-210 (2000).
11.
J. R. Mascola et al., Protection of Macaques against pathogenic simian/human immunodeficiency virus
89.6PD by passive transfer of neutralizing antibodies. J Virol 73, 4009-4018 (1999).
12.
B. Moldt et al., Highly potent HIV-specific antibody neutralization in vitro translates into effective
protection against mucosal SHIV challenge in vivo. Proc Natl Acad Sci U S A 109, 18921-18925 (2012).
13.
P. W. Parren et al., Antibody protects macaques against vaginal challenge with a pathogenic R5
simian/human immunodeficiency virus at serum levels giving complete neutralization in vitro. J Virol 75,
8340-8347 (2001).
14.
J. Pietzsch et al., A mouse model for HIV-1 entry. Proc Natl Acad Sci U S A 109, 15859-15864 (2012).
15.
R. Shibata et al., Neutralizing antibody directed against the HIV-1 envelope glycoprotein can completely
block HIV-1/SIV chimeric virus infections of macaque monkeys. Nat Med 5, 204-210 (1999).
16.
L. E. McCoy, D. R. Burton, Identification and specificity of broadly neutralizing antibodies against HIV.
Immunol Rev 275, 11-20 (2017).
17.
X. Xiao et al., Germline-like predecessors of broadly neutralizing antibodies lack measurable binding to
HIV-1 envelope glycoproteins: implications for evasion of immune responses and design of vaccine
immunogens. Biochem Biophys Res Commun 390, 404-409 (2009).
18.
M. Bonsignori et al., Analysis of a clonal lineage of HIV-1 envelope V2/V3 conformational epitope-
specific broadly neutralizing antibodies and their inferred unmutated common ancestors. J Virol 85,
9998-10009 (2011).
19.
J. F. Scheid et al., Sequence and Structural Convergence of Broad and Potent HIV Antibodies That
Mimic CD4 Binding. Science 333, 1633-1637 (2011).
20.
F. Klein et al., Somatic mutations of the immunoglobulin framework are generally required for broad
and potent HIV-1 neutralization. Cell 153, 126-138 (2013).
28
21.
H. Mouquet et al., Complex-type N-glycan recognition by potent broadly neutralizing HIV antibodies.
Proc Natl Acad Sci U S A 109, E3268-3277 (2012).
22.
T. Zhou et al., Structural basis for broad and potent neutralization of HIV-1 by antibody VRC01. Science
329, 811-817 (2010).
23.
P. Dosenovic et al., Immunization for HIV-1 Broadly Neutralizing Antibodies in Human Ig Knockin Mice.
Cell 161, 1505-1515 (2015).
24.
N. A. Doria-Rose et al., Developmental pathway for potent V1V2-directed HIV-neutralizing antibodies.
Nature 509, 55-62 (2014).
25.
A. T. McGuire et al., Engineering HIV envelope protein to activate germline B cell receptors of broadly
neutralizing anti-CD4 binding site antibodies. J Exp Med 210, 655-663 (2013).
26.
S. Hoot et al., Recombinant HIV envelope proteins fail to engage germline versions of anti-CD4bs
bNAbs. PLoS Pathog 9, e1003106 (2013).
27.
J. Jardine et al., Rational HIV immunogen design to target specific germline B cell receptors. Science
340, 711-716 (2013).
28.
L. Corey et al., Two Randomized Trials of Neutralizing Antibodies to Prevent HIV-1 Acquisition. N Engl
J Med 384, 1003-1014 (2021).
29.
J. G. Jardine et al., HIV-1 broadly neutralizing antibody precursor B cells revealed by germline-targeting
immunogen. Science 351, 1458-1463 (2016).
30.
A. P. West, Jr. et al., Structural Insights on the Role of Antibodies in HIV-1 Vaccine and Therapy. Cell
156, 633-648 (2014).
31.
A. P. West, Jr., R. Diskin, M. C. Nussenzweig, P. J. Bjorkman, Structural basis for germline gene usage
of a potent class of antibodies targeting the CD4 binding site of HIV-1 gp120. Proc Natl Acad Sci U S A
Plus 109, E2083-2090 (2012).
32.
T. Zhou et al., Multidonor analysis reveals structural elements, genetic determinants, and maturation
pathway for HIV-1 neutralization by VRC01-class antibodies. Immunity 39, 245-258 (2013).
33.
R. W. Sanders et al., A next-generation cleaved, soluble HIV-1 Env Trimer, BG505 SOSIP.664 gp140,
expresses multiple epitopes for broadly neutralizing but not non-neutralizing antibodies. PLoS Pathog
9, e1003618 (2013).
34.
H. B. Gristick et al., Natively glycosylated HIV-1 Env structure reveals new mode for antibody
recognition of the CD4-binding site. Nat Struct Mol Biol 23, 906-915 (2016).
35.
X. Wu et al., Rational design of envelope identifies broadly neutralizing human monoclonal antibodies
to HIV-1. Science 329, 856-861 (2010).
36.
M. Bonsignori et al., Inference of the HIV-1 VRC01 Antibody Lineage Unmutated Common Ancestor
Reveals Alternative Pathways to Overcome a Key Glycan Barrier. Immunity 49, 1162-1174 e1168
(2018).
37.
M. Tian et al., Induction of HIV Neutralizing Antibody Lineages in Mice with Diverse Precursor
Repertoires. Cell 166, 1471-1484 e1418 (2016).
38.
X. Chen et al., Vaccination induces maturation in a mouse model of diverse unmutated VRC01-class
precursors to HIV-neutralizing antibodies with >50% breadth. Immunity 54, 324-339 e328 (2021).
39.
G. Johnson, T. T. Wu, Kabat database and its applications: 30 years after the first variability plot.
Nucleic Acids Res 28, 214-218 (2000).
40.
M. S. Packer, D. R. Liu, Methods for the directed evolution of proteins. Nat Rev Genet 16, 379-394
(2015).
41.
G. Chao et al., Isolating and engineering human antibodies using yeast surface display. Nat Protoc 1,
755-768 (2006).
42.
M. J. Betenbaugh, N. Tomiya, S. Narang, J. T. Hsu, Y. C. Lee, Biosynthesis of human-type N-glycans
in heterologous systems. Curr Opin Struct Biol 14, 601-606 (2004).
43.
A. J. Behrens et al., Composition and Antigenic Effects of Individual Glycan Sites of a Trimeric HIV-1
Envelope Glycoprotein. Cell reports 14, 2695-2706 (2016).
44.
E. P. Go et al., Characterization of glycosylation profiles of HIV-1 transmitted/founder envelopes by
mass spectrometry. J Virol 85, 8270-8284 (2011).
45.
C. B. Karsten, G. Alter, The HIV-1 Glycan Shield: Strategically Placed Kinks in the Armor Improve
Antigen Design. Cell reports 19, 669-670 (2017).
29
46.
L. Scharf et al., Antibody 8ANC195 reveals a site of broad vulnerability on the HIV-1 envelope spike.
Cell reports 7, 785-795 (2014).
47.
L. Scharf et al., Broadly Neutralizing Antibody 8ANC195 Recognizes Closed and Open States of HIV-1
Env. Cell 162, 1379-1390 (2015).
48.
A. T. McGuire et al., HIV antibodies. Antigen modification regulates competition of broad and narrow
neutralizing HIV antibodies. Science 346, 1380-1383 (2014).
49.
A. T. McGuire et al., Specifically designed immunogens select and activate B cells expressing
precursors of broadly neutralizing human antibodies to HIV-1 in knock-in mice. Nature Comm 7, 10618
(2016).
50.
J. G. Jardine et al., HIV-1 VACCINES. Priming a broadly neutralizing antibody response to HIV-1 using
a germline-targeting immunogen. Science 349, 156-161 (2015).
51.
K. D. Brune et al., Plug-and-Display: decoration of Virus-Like Particles via isopeptide bonds for modular
immunization. Scientific reports 6, 19234 (2016).
52.
T. U. J. Bruun, A. C. Andersson, S. J. Draper, M. Howarth, Engineering a Rugged Nanoscaffold To
Enhance Plug-and-Display Vaccination. ACS Nano 12, 8855-8866 (2018).
53.
J. López-Sagaseta, E. Malito, R. Rappuoli, M. J. Bottomley, Self-assembling protein nanoparticles in
the design of vaccines. Computational and Structural Biotechnology Journal 14, 58-68 (2016).
54.
M. K. Slifka, I. J. Amanna, Role of Multivalency and Antigenic Threshold in Generating Protective
Antibody Responses. Frontiers in Immunology 10, (2019).
55.
M. Bianchi et al., Electron-Microscopy-Based Epitope Mapping Defines Specificities of Polyclonal
Antibodies Elicited during HIV-1 BG505 Envelope Trimer Immunization. Immunity 49, 288-300 e288
(2018).
56.
L. E. McCoy et al., Holes in the Glycan Shield of the Native HIV Envelope Are a Target of Trimer-
Elicited Neutralizing Antibodies. Cell reports 16, 2327-2338 (2016).
57.
A. Escolano et al., Sequential immunization of macaques elicits heterologous neutralizing antibodies
targeting the V3-glycan patch of HIV-1 Env. Science translational medicine 13, (2021).
58.
T. Amendt et al., Primary Immune Responses and Affinity Maturation Are Controlled by IgD. Front
Immunol 12, 709240 (2021).
59.
M. Silva et al., A particulate saponin/TLR agonist vaccine adjuvant alters lymph flow and modulates
adaptive immunity. Sci Immunol 6, eabf1152 (2021).
60.
A. deCamp et al., Global panel of HIV-1 Env reference strains for standardized assessments of
vaccine-elicited neutralizing antibodies. J Virol 88, 2489-2507 (2014).
61.
C. Soto et al., High frequency of shared clonotypes in human B cell receptor repertoires. Nature 566,
398-402 (2019).
62.
D. C. Montefiori, Measuring HIV neutralization in a luciferase reporter gene assay. Methods Mol Biol
485, 395-405 (2009).
63.
P. D. Kwong et al., Structure of an HIV gp120 envelope glycoprotein in complex with the CD4 receptor
and a neutralizing human antibody. Nature 393, 648-659 (1998).
64.
Y. Guo, K. Chen, P. D. Kwong, L. Shapiro, Z. Sheng, cAb-Rep: A Database of Curated Antibody
Repertoires for Exploring Antibody Diversity and Predicting Antibody Prevalence. Front Immunol 10,
2365 (2019).
65.
Z. Sheng et al., Gene-Specific Substitution Profiles Describe the Types and Frequencies of Amino Acid
Changes during Antibody Somatic Hypermutation. Front Immunol 8, 537 (2017).
66.
A. Escolano et al., Immunization expands B cells specific to HIV-1 V3 glycan in mice and macaques.
Nature 570, 468-473 (2019).
67.
P. Zhang et al., A multiclade env-gag VLP mRNA vaccine elicits tier-2 HIV-1-neutralizing antibodies
and reduces the risk of heterologous SHIV infection in macaques. Nat Med 27, 2234-2245 (2021).
68.
B. Briney et al., Tailored Immunogens Direct Affinity Maturation toward HIV Neutralizing Antibodies.
Cell 166, 1459-1470 e1411 (2016).
69.
K. R. Parks et al., Overcoming Steric Restrictions of VRC01 HIV-1 Neutralizing Antibodies through
Immunization. Cell reports 29, 3060-3072 e3067 (2019).
70.
H. Duan et al., Glycan Masking Focuses Immune Responses to the HIV-1 CD4-Binding Site and
Enhances Elicitation of VRC01-Class Precursor Antibodies. Immunity 49, 301-311 e305 (2018).
30
71.
E. Seydoux et al., Development of a VRC01-class germline targeting immunogen derived from anti-
idiotypic antibodies. Cell reports 35, 109084 (2021).
72.
P. Dosenovic et al., Anti-idiotypic antibodies elicit anti-HIV-1-specific B cell responses. J Exp Med 216,
2316-2330 (2019).
73.
M. Kanekiyo et al., Mosaic nanoparticle display of diverse influenza virus hemagglutinins elicits broad B
cell responses. Nat Immunol 20, 362-372 (2019).
74.
A. A. Cohen et al., Mosaic nanoparticles elicit cross-reactive immune responses to zoonotic
coronaviruses in mice. Science 371, 735-741 (2021).
75.
J. van Schooten et al., Identification of IOMA-class neutralizing antibodies targeting the CD4-binding
site on the HIV-1 envelope glycoprotein. Nat Commun 13, 4515 (2022).
76.
Z. Wang et al., Naturally enhanced neutralizing breadth against SARS-CoV-2 one year after infection.
Nature 595, 426-431 (2021).
77.
Z. Wang et al., Enhanced SARS-CoV-2 neutralization by dimeric IgA. Science translational medicine
13, (2021).
78.
B. Zakeri et al., Peptide tag forming a rapid covalent bond to a protein, through engineering a bacterial
adhesin. Proc Natl Acad Sci U S A 109, E690-697 (2012).
79.
A. H. Keeble et al., Approaching infinite affinity through engineering of peptide–protein interaction.
Proceedings of the National Academy of Sciences 116, 26523-26533 (2019).
80.
L. Scharf et al., Structural basis for germline antibody recognition of HIV-1 immunogens. Elife 5,
(2016).
81.
C. O. Barnes et al., A naturally arising broad and potent CD4-binding site antibody with low somatic
mutation. bioRxiv, (2022).
82.
K.-M. A. Dam et al., HIV-1 CD4-binding site germline antibody–Env structures inform vaccine design.
bioRxiv, (2022).
83.
R. W. Siegel, Antibody affinity optimization using yeast cell surface display. Methods Mol Biol 504, 351-
383 (2009).
84.
J. Snijder et al., An Antibody Targeting the Fusion Machinery Neutralizes Dual-Tropic Infection and
Defines a Site of Vulnerability on Epstein-Barr Virus. Immunity 48, 799-811 e799 (2018).
85.
P. D. Adams et al., PHENIX: a comprehensive Python-based system for macromolecular structure
solution. Acta Crystallogr D Biol Crystallogr 66, 213-221 (2010).
86.
P. Emsley, B. Lohkamp, W. G. Scott, K. Cowtan, Features and development of Coot. Acta Crystallogr D
Biol Crystallogr 66, 486-501 (2010).
87.
S. A. Gai, K. D. Wittrup, Yeast surface display for protein engineering and characterization. Curr Opin
Struct Biol 17, 467-473 (2007).
88.
S. Tian, R. Das, Primerize-2D: automated primer design for RNA multidimensional chemical mapping.
Bioinformatics, (2017).
89.
R. Rahikainen et al., Overcoming Symmetry Mismatch in Vaccine Nanoassembly through Spontaneous
Amidation. Angew Chem Int Ed Engl 60, 321-330 (2021).
90.
A. A. Cohen et al., Construction, characterization, and immunization of nanoparticles that display a
diverse array of influenza HA trimers. PLoS One 16, e0247963 (2021).
91.
J. Staal, K. Alci, W. De Schamphelaire, M. Vanhoucke, R. Beyaert, Engineering a minimal cloning
vector from a pUC18 plasmid backbone with an extended multiple cloning site. Biotechniques 66, 254-
259 (2019).
92.
M. L. Donnelly et al., Analysis of the aphthovirus 2A/2B polyprotein 'cleavage' mechanism indicates not
a proteolytic reaction, but a novel translational effect: a putative ribosomal 'skip'. The Journal of general
virology 82, 1013-1025 (2001).
93.
H. Hartweger et al., HIV-specific humoral immune responses by CRISPR/Cas9-edited B cells. J Exp
Med 216, 1301-1310 (2019).
94.
R. Pelanda et al., Receptor Editing in a Transgenic Mouse Model: Site, Efficiency, and Role in B Cell
Tolerance and Antibody Diversification. Immunity 7, 765-775 (1997).
95.
T. A. Shih, M. Roederer, M. C. Nussenzweig, Role of antigen receptor affinity in T cell-independent
antibody responses in vivo. Nat Immunol 3, 399-406 (2002).
96.
K. Lövgren-Bengtsson, B. Morein, in Vaccine Adjuvants: Preparation Methods and Research Protocols,
D. T. O'Hagan, Ed. (Humana, 2000), pp. 239-258.
31
97.
J. Ye, N. Ma, T. L. Madden, J. M. Ostell, IgBLAST: an immunoglobulin variable domain sequence
analysis tool. Nucleic Acids Res 41, W34-40 (2013).
98.
X. Brochet, M. P. Lefranc, V. Giudicelli, IMGT/V-QUEST: the highly customized and integrated system
for IG and TR standardized V-J and V-D-J sequence analysis. Nucleic Acids Res 36, W503-508 (2008).
99.
C. Viant, A. Escolano, S. T. Chen, M. C. Nussenzweig, Sequencing, cloning, and antigen binding
analysis of monoclonal antibodies isolated from single mouse B cells. STAR Protoc 2, 100389 (2021).
100.
K. Katoh, D. M. Standley, MAFFT multiple sequence alignment software version 7: improvements in
performance and usability. Mol Biol Evol 30, 772-780 (2013).
101.
D. C. Montefiori, Evaluating neutralizing antibodies against HIV, SIV, and SHIV in luciferase reporter
gene assays. Current protocols in immunology / edited by John E. Coligan ... [et al Chapter 12, Unit 12
11 (2005).
102.
M. M. Sajadi et al., Identification of Near-Pan-neutralizing Antibodies against HIV-1 by Deconvolution of
Plasma Humoral Responses. Cell 173, 1783-1795 e1714 (2018).
103.
T. D. Goddard et al., UCSF ChimeraX: Meeting modern challenges in visualization and analysis.
Protein Sci 27, 14-25 (2018).
104.
F. Sievers et al., Fast, scalable generation of high‐quality protein multiple sequence alignments using
Clustal Omega. Molecular Systems Biology 7, 539 (2011).
32
Figure 1. Design and characterization of IOMA-iGL targeting immunogens. (A) Overview of
strategy to engineer and test immunogens designed to elicit IOMA-like antibodies: (B-D) Residues
selected from the unmutated starting protein 426c.TM4 gp120 to mutate in yeast display library (left
panel), FACS summary (second and third panels), and SPR data for highest affinity immunogen
selected from each library as gp120 (fourth panel) and SOSIP (fifth panel) are shown (B), Library 1 (C),
and Library 2 (D). Residues shown in red represent degenerate positions in the library, while residues
shown in green represent point mutations from 426c.TM4 gp120. Representative sensorgrams are
shown in red with the 1:1 binding model fits shown in black. IgG was immobilized to the CM5 chip and
gp120 at varying concentrations was flowed over the chip surface (IGT2 gp120: 4.9 nM – 5,000 nM;
IGT1 gp120: 2.3 nM – 150,000 nM; 426c.TM4 gp120: 7,000 nM – 609,000 nM; IGT2 SOSIP: 31 nM –
2,000 nM; IGT1 SOSIP: 78 nM – 10,000 nM; 426c degly2 SOSIP: 313 nM – 40,000 nM). (E) ELISA
demonstrating binding of CD4bs IgGs to various Env proteins. Bars indicate mean and 95% confidence
interval. (F) Representative negative stain EM micrographs of unconjugated SpyCatcher003-mi3
nanoparticles (left) and IGT2-SpyTag SOSIP conjugated to SpyCatcher003-mi3 nanoparticles (right).
Scale bar is 50 nm.
33
Figure 2. Sequential immunization with IOMA-targeting immunogens elicits heterologous
neutralizing serum responses in IOMA iGL transgenic mice. (A) Schematic and timeline of
immunization regimen for IOMA iGL knock-in mice. (B-F) Serum ELISA binding at the indicated time
points for IGT2 and IGT2 KO (B), IGT1 and IGT1 KO (C), 426c D279N, 426c, and 426c KO (D), and
to a panel of wt and N276A-versions of SOSIP-based Envs (E-F). (G) Serum neutralization activity
against a panel of 18 HIV pseudoviruses and an murine leukemia virus (MLV) control after terminal
bleed. Animal immunization studies were performed as 3 independent experiments. Each dot
represents results from one mouse. Bars indicate mean and 95% confidence interval. AUC, area
under the curve. n, number of animals.
34
Figure 3. Monoclonal antibodies cloned from IOMA iGL transgenic mice neutralize
heterologous HIV strains. (A) Graphs show the total number of V region (excluding CDR3) amino
acid mutations in HC (top) and LC (bottom) of all antibody sequences (x-axis) vs. the number of
mutations that are identical or chemically equivalent to mutations in IOMA for aa positions where
IOMA and IOMA iGL differ (y-axis). Sequences derived from IOMAgl mice HP1, HP3 and ES30 from
immunization group 1 (red) and baseline human VH1-2*01 or VL2-23*02 sequences from peripheral
blood of HIV-negative human donors (gray). The size of the dot is proportional to the number of
sequences. Number of sequences (n), determination coefficient (Pearson, R2) and linear regression
lines are indicated. Chemical equivalence classified in 7 groups as follows: (1) G=A=V=L=I; (2) S=T;
(3) C=M; (4) D=N=E=Q; (5)_R=K=H; (6) F=Y=W; (7) P. (B) Neutralization titers (IC50s) of nine
representative monoclonal antibodies isolated from IOMA iGL transgenic mice against a panel of 14
viruses and an MLV control. IC50s for IOMA are shown on the far left. (C) 3D plot showing
neutralization activity (color coded), total number of amino acid mutations in both HC and LC V(D)Js
(x-axis), and the number of mutations that are identical or chemically equivalent to mutations in the
IOMA (y-axis) for all Env-binding monoclonal antibodies from IOMAgl mice HP1, HP3, and ES30 from
immunization group 1. Chemical equivalence is as in (A). For each antibody a neutralization score
was calculated (see Methods). Red indicates higher neutralization activity and score. Number of
sequences (n) are indicated. (D) Residues mutated from IOMA iGL are shown as red spheres
mapped onto the crystal structure of mature IOMA (shown in cartoon representation) bound to BG505
gp120 (depicted in surface representation) (PDB 5T3Z). SHMs are depicted for mature IOMA (left
panel) as well as two antibodies isolated from IOMAgl mice: the more potent IO-010 (middle panel)
and weaker IO-040 (right panel). (E) Total SHMs for mature IOMA (left panel) or SHMs found in the
IOMA-gp120 interface (right panel) are colored according to their percentages of occurrence from
green to magenta (left panel). Structures are depicted as in (D). (F) Key mutations essential for IOMA
binding to Env that were elicited in our immunization strategy are mapped onto antibody IO-010 and
highlighted in each inset box. IO-010 depicted as in (D). Each inset represents a different interaction
between IOMA and gp120. (G) Amino acid sequence alignment of IOMA VH and VL and monoclonal
antibodies from (B) with IOMA iGL as a reference.
35
Figure 4. Sequential immunization with IOMA-targeting immunogens elicits CD4bs-specific
responses and heterologous neutralizing serum responses in wildtype mice. (A) Schematic and
timeline of immunization regimen for wt mice. (B) Serum ELISA binding at the indicated time points to
IGT1 or IGT1 CD4bs-KO (KO). (C-D) Serum ELISA binding to anti-idiotypic monoclonal antibodies
raised against IOMA iGL (left, 3D3) and IOMA iGL + mature IOMA (right, 3D7). Mean ± SEM of 9 to 16
mice per time point are depicted. (D-E) Serum ELISA binding at the indicated time points to a panel of
WT and N276A-versions of SOSIP-based Envs. (F) Serum neutralization against a panel of 18 viruses
and an MLV control at week 23 of wt mice. Animal immunization studies were performed as 3
independent experiments. Each dot represents results from one mouse. Bars indicate mean and 95%
confidence interval. AUC, area under the curve.
36
Figure 5. Prime-boost with IGT2-IGT1 elicits CD4bs-specific responses and potent autologous
neutralization in rabbits and rhesus macaques. (A) Schematic and timeline of immunization regimen
for rabbits and rhesus macaques. (B) Serum ELISA binding to IGT1 and IGT1 KO for rabbits (left) and
rhesus macaques (right). (C) Serum neutralization ID50s of IGT2 and IGT1 pseudoviruses for rabbits
(left) and rhesus macaques (right). The dotted line at y = 102 indicates the lowest dilution evaluated.
Significance was demonstrated using a paired t test (p ≤ 0.05).
37
Figure S1. Development and characterization of IGT1 and IGT2 immunogens. (A) Amino acid
alignment of IOMA and VRC01 to their respective germline V genes. (B) Representative SPR
sensorgrams demonstrating no detectable binding of IOMA iGL to previously described immunogens
(eOD-GT8, 426c.TM4, BG505.v4.1-GT1). This experiment was performed to qualitatively evaluate
binding of IGT2 and previously described CD4bs immunogens to IOMA iGL rather than to derive
affinity or kinetic constants. (C) Neutralization titers (IC50s) of IOMA and IOMA iGL against a panel of
38 viruses and an MLV control. (D) 2.07 Å crystal structure of IOMA iGL Fab shown in two views. (E)
Structural overlay of IOMA iGL Fab and IOMA Fab from BG505-bound structure (PDB 5T3Z).(F) Flow
cytometric analysis of yeast cells expressing 426c.TM4 starting protein (left), Library 1 (middle), or
Library 2 (right) stained with IOMA iGL IgG/anti-IgG AF647 (x-axis) and anti-cMyc AF488 (y-axis). (G)
Representative size exclusion chromatography profiles and Coomassie-stained SDS-PAGE analysis
for 426c.TM4 gp120, IGT1 gp120, and IGT2 gp120, 426c SOSIP, IGT1 SOSIP, and IGT2 SOSIP
demonstrating that all of these proteins are monodispersed samples and that the selected mutations
do not alter the stability or behavior of the immunogens compared to the starting proteins. (H)
Coomassie-stained SDS–PAGE analysis for mi3, IGT2, IGT2-mi3, IGT1, and IGT1-mi3 under non-
reducing and reducing conditions. (I) SPR sensorgrams demonstrating binding of IGT2 (dashed line)
and IGT2-mi3 (solid line) to IOMA iGL IgG (red), VRC01 iGL IgG (purple), 3BNC60 iGL IgG (green),
and BG24 iGL IgG (orange). IgG was immobilized to the CM5 chip and 1 µM SOSIP or 1 µM SOSIP-
mi3 was flowed over the chip surface. (J) Representative ELISA binding curves measuring binding of
426c.TM4 gp120, IGT1 gp120, and IGT2 gp120 to the same iGL IgGs as in (I). Dots indicate mean
and error bars indicate 95% confidence interval.
38
Figure S2. Targeting strategy and characterization of IOMAgl mice (A) In IghIOMAiGL mice Ighd4-1
to Ighj4 are replaced by a self-excising Neomycin cassette followed by the mouse Ighv9-4 promoter,
a leader sequence (L) followed by the iGL version of the IOMA HC VDJ sequence and a Ighj1 splice
donor sequence. (B) In IgkIOMAiGL mice Igkj1 to Igkj5 are replaced by a self-excising Neomycin
cassette followed by a mouse Igkv3-12 promoter, a leader sequence followed by the iGL version of
the IOMA lambda LC VDJ sequence and a Igkj5 splice donor sequence. DTA, diphtheria toxin A (C)
Flow cytometric analysis of B cell development in the bone marrow of control (C57BL/6J) or IOMAgl
(IghIOMAiLG/IOMiGL IgkIOMAiG/IOMAiGLL) mice. (D) Absolute cell number quantification from (C). (E)
Geometric mean fluorescence intensity (gMFI) of IgD in mature recirculating B cells from the bone
marrow. (F) Flow cytometric analysis of peripheral B cell development in the spleens of control
(C57BL/6J) or IOMAgl mice. (G) Absolute cell number quantification from (F). (H) gMFI of IgD in
marginal zone and follicular B cell. MZ, marginal zone B cells; MZP, marginal zone precursors; FOB,
follicular B cells. Data from 1 of 2 independent experiments, each dot represents a data from 1
mouse. Bars represent mean ± SEM. Statistical analysis used unpaired t test.
39
Figure S3. Serum neutralization from immunized mice. Neutralization curves of serum isolated
from IOMA iGL transgenic mice (A-M) or C57BL/6J wildtype mice (N-X) against the following HIV
strains or control MuLV: (A,N) CNE8, (B,O) CNE8 N276A, (C,P) CNE20, (D,Q) CNE20 N276A, (E,R)
PVO.4, (F,U) Q23.17, (G,T ) WITO4160.33, (H) YU2, (I) JRCSF, (J, V) 6535.5, (K) 3415_V1_C1, (L)
CAAN5342.A2, (M,X) MuLV, (S) Q842.D12 and (W) BG505. Naïve serum was also tested against the
same strains when available. Note that sera which showed neutralization activity of < 40% as listed in
Table S3 are presented in Figure 2G as white rectangles; several of these sera neutralized strains
above background including ET33 against PVO.4; ET34 against CNE20 N276A and Q23.17; HP1
against CNE8 N276A, CNE20, and WITO4160.33; HP2 against Q23.17; HP3 against Q23.17 and
PVO.4; HP4 against CNE8 N276A, CNE20 N276A, and PVO.4.
40
Figure S4. Screening immunization regimens to determine the optimal boosting strategy. (A)
Schematic and timeline of immunization strategies to determine the optimal regimen to elicit IOMA-like
bNAbs. (B) Serum ELISA binding to 426c and 426c degly2 represented as AUC using serum samples
isolated from mice at the end of the regimen.
41
Figure S5. Cell sorting strategies and sorting controls. (A) Representative full gating of cell sorts
for single cell Bait++ BaitKO- B cell cloning and 10x Genomics next generation VDJ sequencing of bulk-
sorted GC B cells from splenic and mesenteric lymph nodes. Baits used were 426c degly2 D279N or
CNE8 N276A with 426c degly2 D279N-CD4bs KO, the former is shown. (B) Induction of germinal
center response and wt SOSIP-binding cells by immunization regimen (group 1). Naïve IOMAgl mouse
splenocytes and IOMA-expressing RAMOS cells served as negative and positive control, respectively.
(C) Gene editing strategy to generate IOMA-expressing RAMOS cells. Simultaneous targeting of IgH,
IgK and IgL loci with CRISPR/Cas9 to delete endogenous LCs and edit a promoterless tricistronic
expression cassette into the IgH locus to express IOMA on the surface of RAMOS cells. A polycistronic
mRNA was created using T2A and P2A sequences to induce ribosomal skipping (92).
42
Figure S6. Amino acid alignments of selected IOMAgl mouse-derived antibodies. (A) VH
alignment of cloned antibodies IO-001 to IO-067 that were expressed and tested for Env binding.
IOMA iGL and IOMA sequence at the bottom as reference. Mouse ID and population sorted are
indicated. Differences to IOMA iGL are highlighted using chemically similar color coding; dots indicate
identical residues to IOMA iGL. Kabat numbering and percent identity of residues are indicated on
top. Domains and residues of structural importance are annotated below. (B) as above but
corresponding VL alignment.
43
Figure S7. Next generation single cell VDJ analysis determines the extent of mutations in
germinal centers of IOMAgl mice. (A) Clonal analysis of paired HC and LC sequences from splenic
and mesenteric lymph node germinal center B cells of IOMAgl mouse HP3. (B) Isotype distribution
among these cells. (C) Frequency distribution of the number of amino acid mutations to IOMA iGL in
the HC sequences of these cells. (D) Frequency distribution of the number of amino acid mutations to
IOMA iGL in the LC sequences of these cells. (E) Frequency distribution of the number of amino acid
mutations to IOMA iGL in the paired HC and LC sequences of these cells.
44
Figure S8. Monoclonal antibodies cloned from IOMA iGL transgenic mice bind to heterologous
Envs. (A) AUC of ELISA binding curves of selected monoclonal antibodies isolated from IOMA iGL
knock-in mice to BG505, CE0217, CNE20 and CNE20 N276A SOSIPs. (B) Comparison of the
occurrence frequency of key mutations among IOMA-like antibody sequences selected for cloning
and VRC01-class antibody sequences from reference 37 at different time points throughout the
respective sequential immunization regimen. Mutations essential for IOMA-class antibody binding to
gp120 are listed first, while mutations essential for VRC01-class antibody binding to gp120 are listed
second in brackets. Values for each residue represent the percentage of antibodies containing one of
the essential mutations at that position.
45
Table S1: Amino acid sequences for HIV Envs and antibodies used in this study.
Protein
Name
Sequence
IGT2 gp120
VWKEAKTTLFCASDAKAYEKECHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMVDQMQEDVISIWDQ
CLKPCVKLTNTSTLTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGKGPCNNVSTVQCTHGIKPVVSTQ
LLLNGSLAEEEIVIRSKNLRNNAKIIIVQLNKSVEIVCTRPNNGGSGSGGDIRQAYCNISGRNWSEAVNQV
KKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGEFFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEPG
KAIYAPPIKGNITCKSDITGLLLLRDGGNALRPTEIFRPSGGDMRDNWRSELYKYKVVEIKPLHHHHHH
IGT1 gp120
VWKEAKTTLFCASDAKAYEKECHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMVDQMQEDVISIWDQ
CLKPCVKLTNTSTLTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGKGPCNNVSTVQCTHGIKPVVSTQ
LLLNGSLAEEEIVIRSKNLRNNAKIIIVQLNKSVEIVCTRPNNGGSGSGGDIRQAYCNISGRNWSEAVNQV
KKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGEFFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEPG
KAIYAPPIKGNITCKSDITGLLLLRDGGNSQRETEIFRPSGGDMRDNWRSELYKYKVVEIKPLHHHHHH
426c.TM4 gp120
VWKEAKTTLFCASDAKAYEKECHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMVDQMQEDVISIWDQ
CLKPCVKLTNTSTLTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGKGPCNNVSTVQCTHGIKPVVSTQ
LLLNGSLAEEEIVIRSKNLRDNAKIIIVQLNKSVEIVCTRPNNGGSGSGGDIRQAYCNISGRNWSEAVNQV
KKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGEFFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVG
KAIYAPPIKGNITCKSDITGLLLLRDGGDTTDNTEIFRPSGGDMRDNWRSELYKYKVVEIKPLHHHHHH
IGT2 SOSIP
GSNLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKNLRNNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNALRPTE
IFRPSGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
IGT1 SOSIP
GSNLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKNLRNNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNSQRETE
IFRPSGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
426c SOSIP
AENLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKNLSDNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNTTNNTE
IFRPGGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
426c D279N
SOSIP
AENLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKNLSNNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNTTNNTE
IFRPGGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
46
426c degly2
SOSIP
AENLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKNLTDNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNTANNAE
IFRPGGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
426c degly2
D279N SOSIP
AENLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKNLTNNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNTANNAE
IFRPGGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
426c degly3
SOSIP
AENLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKALTDNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNTANNAE
IFRPGGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
BG505.v4.1-GT1
SOSIP
AENLWVTVYYGVPVWKDAETTLFCASDAKAYETKKHNVWATHACVPTDPNPQEIHLENVTEEFNMWKNNMV
EQMHTDIISLWDQSLKPCVKLTPLCVTLQCTNVTNAITDDMRGELKNCSFNMTTELRDKRQKVHALFYKLD
IVPINENQNTSYRLINCNTAAITQACPKVSFEPIPIHYCAPAGFAILKCKDKKFNGTGPCPSVSTVQCTHG
IKPVVSTQLLLNGSLAEEEVMIRSEDIRNNAKNILVQFNTPVQINCTRPNNNTRKSIRIGPGQWFYATGDI
IGDIRQAHCNVSKATWNETLGKVVKQLRKHFGNNTIIRFANSSGGDLEVTTHSFNCGGEFFYCDTSGLFNS
TWISNTSVQGSNSTGSNDSITLPCRIKQIINMWQRIGQAMYAPPIQGVIRCVSNITGLILTRDGGSTDSTT
ETFRPSGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAAS
MTLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICC
TNVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
eOD-GT8
DTITLPCRPAPPPHCSSNITGLILTRQGGYSNANTVIFRPSGGDWRDIARCQIAGTVVSTQLFLNGSLAEE
EVVIRSEDWRDNAKSICVQLATSVEIACTGAGHCAISRAKWANTLKQIASKLREQYGAKTIIFKPSSGGDP
EFVNHSFNCGGEFFYCASTQLFASTWFASTGTGTK
IGT2 SOSIP
SpyTag
GSNLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKNLRNNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNALRPTE
IFRPSGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGGGSGGGS
GSGAHIVMVDAYKPTK
IGT1 SOSIP
SpyTag
GSNLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKNLRNNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNSQRETE
IFRPSGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGGGSGGGS
GSGAHIVMVDAYKPTK
47
426c degly2
D279N SOSIP
SpyTag
AENLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEVVLENVTENFNMWKNDMV
DQMQEDVISIWDQSLKPCVKLTPLCVTLNCTNVNVTSNSTNVNSSSTDNTTLGEIKNCSFDITTEIRDKTR
KEYALFYRLDIVPLDNSSNPNSSNTYRLINCNTSTCTQACPKVTFDPIPIHYCAPAGYAILKCNNKTFNGK
GPCNNVSTVQCTHGIKPVVSTQLLLNGSLAEEEIVIRSKNLTNNAKIIIVQLNKSVEIVCTRPNNNTRRSI
RIGPGQTFYATDIIGDIRQAYCNISGRNWSEAVNQVKKKLKEHFPHKNISFQSSSGGDLEITTHSFNCGGE
FFYCNTSGLFNDTISNATIMLPCRIKQIINMWQEVGKCIYAPPIKGNITCKSDITGLLLLRDGGNTANNAE
IFRPGGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVFLGFLGAAGSTMGAASM
TLTVQARNLLSGIVQQQSNLLRAPEAQQHLLKLTVWGIKQLQARVLAVERYLRDQQLLGIWGCSGKLICCT
NVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGGGSGGGS
GSGRGVPHIVMVDAYKRYK
398F1 SOSIP
SpyTag
AENLWVTVYYGVPVWKDAETTLFCASDAKAYHTEVHNVWATHACVPTDPNPQEINLENVTEEFNMWKNKMV
EQMHTDIISLWDQSLKPCVQLTPLCVTLDCQYNVTNINSTSDMAREINNCSYNITTELRDREQKVYSLFYR
SDIVQMNSDNSSKYRLINCNTSACKQACPKVTFEPIPIHYCAPAGFAILKCKDKEFNGTGPCKNVSTVQCT
HGIKPVVSTQLLLNGSLAEEKVMIRSENITDNAKNIIVQFKEPVKINCTRPNNNTRKSVRIGPGQTFYATG
EIIGDIRQAHCNVSKAHWENTLQEVANQLKLMIHSNKTIIFANSSGGDLEITTHSFNCGGEFFYCYTSGLF
NYTFNDTSTNSTESKSNDTITLQCRIKQIINMWQRAGQCVYAPPIPGIIRCESNITGLILTRDGGNNNSNT
NETFRPGGGDMRDNWRSELYRYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAA
SMTLTVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLIC
CTNVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGGGSGG
GSGSGAHIVMVDAYKPTK
BJOX2000 SOSIP
SpyTag
AENLWVTVYYGVPVWKEATTTLFCASDAKAYDTEVHNVWATHACVPTDPDPQEMFLENVTENFNMWKNNMV
DQMHEDVISLWDQSLKPCVKLTPLCVTLECKNVNSSSSDTKNGTDPEMKNCSFNATTELRDRKQKVYALFY
KLDIVPLNEKNSSEYRLINCNTSTCTQACPKVTFDPIPIHYCTPAGYAILKCNDEKFNGTGPCSNVSTVQC
THGIKPVVSTQLLLNGSLAEKGIVIRSENLTNNVKTIIVHLNQSVEILCIRPNNNTRKSIRIGPGQTFYAT
GEIIGDIRQAHCNISGKVWNETLQRVGEKLAEYFPNKTIKFNSSSGGDLEITTHSFNCGGEFFYCNTSKLF
NGTFNGTYMPNVTEGNSTISIPCRIKQIINMWQKVGRCMYAPPIEGNITCKSKITGLLLERDGGPENDTEI
FRPGGGDMRNNWRSELYKYKVVEIKPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAASMT
LTVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLICCTN
VPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGGGSGGGSG
SGAHIVMVDAYKPTK
CE1176 SOSIP
SpyTag
AENLWVTVYYGVPVWKEAKTTLFCASDAKAYEKEVHNVWATHACVPTDPNPQEMVLENVTENFNMWKNDMV
DQMHEDVISLWDQSLKPCVKLTPLCVTLTCTNTTVSNGSSNSNANFEEMKNCSFNATTEIKDKKKNEYALF
YKLDIVPLNNSSGKYRLINCNTSACAQACPKVTFEPIPIHYCAPAGYAILKCNNKTFNGTGPCNNVSTVQC
THGIKPVVSTQLLLNGSLAEKEIIIRSENLTNNAKTIIIHFNESVGIVCTRPSNNTRKSIRIGPGQTFYAT
GDIIGDIRQAHCNVSKQNWNRTLQQVGRKLAEHFPNRNITFNHSSGGDLEITTHSFNCRGEFFYCNTSGLF
NGTYHPNGTYNETAVNSSDTITLQCRIKQIINMWQEVGRCMYAPPIAGNITCNSTITGLLLTRDGGINQTG
EEIFRPGGGDMRDNWRNELYKYKVVEIKPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAA
SMTLTVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLIC
CTNVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGGGSGG
GSGSGAHIVMVDAYKPTK
CE0217 SOSIP
SpyTag
AENLWVTVYYGVPVWREAKTTLFCASDAKAYEREVHNVWATHACVPTDPNPQERVLENVTENFNMWKNNMV
DQMHEDIISLWDESLKPCIKLTPLCVTLNCGNAIVNESTIEGMKNCSFNVTTELKDKKKKEYALFYKLDVV
PLNGENNNSNSKNFSEYRLINCNTSTCTQACPKVSFDPIPIHYCAPAGFAILKCNNETFNGTGPCNNVSTV
QCTHGIKPVVSTQLLLNGSLAEKEIIIRSENLTNNAKIIIVHLNNPVKIICTRPGNNTRKSMRIGPGQTFY
ATGDIIGDIRRAYCNISEKTWYDTLKNVSDKFQEHFPNASIEFKPSAGGDLEITTHSFNCRGEFFYCDTSE
LFNGTYNNSTYNSSNNITLQCKIKQIINMWQGVGRCMYAPPIAGNITCESNITGLLLTRDGGNNKSTPETF
RPGGGDMRDNWRSELYKYKVVEIKPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAASMTL
TVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLICCTNV
PWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGGGSGGGSGS
GAHIVMVDAYKPTK
CNE55 SOSIP
SpyTag
AENLWVTVYYGVPVWRDADTTLFCASDAKAHETEVHNVWATHACVPTDPNPQEIHLVNVTENFNMWKNKMV
EQMQEDVISLWDESLKPCVKLTPLCVTLNCTTANTNETKNNTTDDNIKDEMKNCTFNMTTEIRDKKQRVSA
LFYKLDIVPIDDSKNNSEYRLINCNTSVCKQACPKVSFDPIPIHYCTPAGYVILKCNDKNFNGTGPCKNVS
SVQCTHGIKPVVSTQLLLNGSLAEEEIIIRSENLTDNAKNIIVHLNKSVEINCTRPSNNTRTSVRIGPGQV
FYRTGDITGDIRKAYCNISGTEWNKTLTQVAEKLKEHFNKTIVYQPPSGGDLEITMHHFNCRGEFFYCNTT
QLFNNSVGNSTIKLPCRIKQIINMWQGVGQCMYAPPISGAINCLSNITGILLTRDGGGNNRSNETFRPGGG
NIKDNWRSELYKYKVVEIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAASMTLTVQAR
NLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLICCTNVPWNSS
WSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGGGSGGGSGSGAHIV
MVDAYKPTK
48
Tro11 SOSIP
SpyTag
AENLWVTVYYGVPVWKDASTTLFCASDAKAYDTEVHNVWATHACVPTDPNPQEVVLGNVTENFNMWKNNMV
DQMHEDIISLWDQSLKPCVKLTPLCVTLNCTDNITNTNTNSSKNSSTHSYNNSLEGEMKNCSFNITAGIRD
KVKKEYALFYKLDVVPIEEDKDTNKTTYRLRSCNTSVCTQACPKVTFEPIPIHYCAPAGFAILKCNDKKFN
GTGPCTNVSTVQCTHGIRPVVSTQLLLNGSLAEEEVVIRSENFTNNAKTIIVQLNESIAINCTRPNNNTRR
SIHIGPGRAFYATGDIIGDIRQAHCNISRTEWNSTLRQIVTKLREQLGDPNKTIIFNQSSGGDTEITMHSF
NCGGEFFYCNTTKLFNSTWNGNNTTESDSTGENITLPCRIKQIINLWQEVGKCMYAPPIKGQISCSSNITG
LLLTRDGGNNNSSGPETFRPGGGNMKDNWRSELYKYKVIKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVS
LGFLGAAGSTMGAASMTLTVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQ
QLLGIWGCSGKLICCTNVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLL
ALDGGGGSGGGSGGGSGSGAHIVMVDAYKPTK
X1632 SOSIP
SpyTag
AENLWVTVYYGVPVWEDADTTLFCASDAKAYSTESHNVWATHACVPTDPNPQEIYLENVTEDFNMWENNMV
EQMQEDIISLWDESLKPCVKLTPLCVTLTCTNVTNVTDSVGTNSRLKGYKEELKNCSFNTTTEIRDKKKQE
YALFYKLDIVPINDNSNNSNGYRLINCNVSTCKQACPKVSFDPIPIHYCAPAGFAILKCRDKEFNGTGTCR
NVSTVQCTHGIKPVVSTQLLLNGSLAEGDIVIRSENITDNAKTIIVHLNKTVSITCTRPNNNTRKSIRIGP
GQALYATGAIIGDTRQAHCNISGSEWYEMIQNVKNKLNETFKKNITFNPSSGGDLEITTHSFNCRGEFFYC
NTSELFNSSHLFNGSTLSTNGTITLPCRIKQIVRMWQRVGQCMYAPPIAGNITCRSNITGLLLTRDGGTNK
DTNEAETFRPGGGDMRDNWRSELYKYKVVKIKPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGST
MGAASMTLTVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSG
KLICCTNVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGG
GSGGGSGSGAHIVMVDAYKPTK
X2278 SOSIP
SpyTag
AENLWVTVYYGVPVWKEATTTLFCASEAKAYDTEVHNIWATHACVPTDPNPQEMELKNVTENFNMWKNNMV
EQMHQDIISLWDQSLKPCVKLTPLCVTLDCTNINSTNSTNNTSSNSKMEETIGVIKNCSFNVTTNIRDKVK
KENALFYSLDLVSIGNSNTSYRLISCNTSICTQACPKVSFDPIPIHYCAPAGFAILKCRDKKFNGTGPCRN
VSSVQCTHGIRPVVSTQLLLNGSLAEEEIVIRSANLTDNAKTIIIQLNETIQINCTRPNNNTRRSIPIGPG
RTFYATGDIIGDIRKAYCNISATKWNNTLRQIAEKLREKFNKTIIFNQSSGGDPEVVRHTFNCGGEFFYCN
SSQLFNSTWYSNGTSNGGLNNSANITLPCRIKQIINLWQEVGKCMYAPPIKGVINCLSNITGIILTRDGGE
NNGTTETFRPGGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGST
MGAASMTLTVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSG
KLICCTNVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALDGGGGSGG
GSGGGSGSGAHIVMVDAYKPTK
BG505 SOSIP
NLWVTVYYGVPVWKDAETTLFCASDAKAYETEKHNVWATHACVPTDPNPQEIHLENVTEEFNMWKNNMVEQ
MHTDIISLWDQSLKPCVKLTPLCVTLQCTNVTNNITDDMRGELKNCSFNMTTELRDKKQKVYSLFYRLDVV
QINENQGNRSNNSNKEYRLINCNTSAITQACPKVSFEPIPIHYCAPAGFAILKCKDKKFNGTGPCPSVSTV
QCTHGIKPVVSTQLLLNGSLAEEEVMIRSENITNNAKNILVQFNTPVQINCTRPNNNTRKSIRIGPGQAFY
ATGDIIGDIRQAHCNVSKATWNETLGKVVKQLRKHFGNNTIIRFANSSGGDLEVTTHSFNCGGEFFYCNTS
GLFNSTWISNTSVQGSNSTGSNDSITLPCRIKQIINMWQRIGQAMYAPPIQGVIRCVSNITGLILTRDGGS
TNSTTETFRPGGGDMRDNWRSELYKYKVVKIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGST
MGAASMTLTVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSG
KLICCTNVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
AMC011 SOSIP
AEQLWVTVYYGVPVWKEATTTLFCASDARAYDTEVRNVWATHCCVPTDPNPQEVVLENVTENFNMWKNNMV
EQMHEDIISLWDQSLKPCVKLTPLCVTLNCTDLRNATNTNATNTTSSSRGTMEGGEIKNCSFNITTSMRDK
VQKEYALFYKLDVVPIKNDNTSYRLISCNTSVITQACPKVSFEPIPIHYCAPAGFAILKCNNKTFNGTGPC
TNVSTVQCTHGIRPVVSTQLLLNGSLAEEEVVIRSANFTDNAKIIIVQLNKSVEINCTRPNNNTRKSIHIG
PGRWFYTTGEIIGDIRQAHCNISGTKWNDTLKQIVVKLKEQFGNKTIVFNHSSGGDPEIVMHSFNCGGEFF
YCNSTQLFNSTWNDTTGSNYTGTIVLPCRIKQIVNMWQEVGKAMYAPPIKGQIRCSSNITGLILIRDGGKN
RSENTEIFRPGGGDMRDNWRSELYKYKVVKIEPLGIAPTKCKRRVVQRRRRRRAVGIGAVFLGFLGAAGST
MGAASMTLTVQARQLLSGIVQQQNNLLRAPECQQHLLKLTVWGIKQLQARVLAVERYLKDQQLLGIWGCSG
KLICCTAVPWNTSWSNKSYNQIWNNMTWMEWEREIDNYTSLIYTLIEDSQNQQEKNEQELLELD
B41 SOSIP
AAKKWVTVYYGVPVWKEATTTLFCASDAKAYDTEVHNVWATHACVPTDPNPQEIVLGNVTENFNMWKNNMV
EQMHEDIISLWDQSLKPCVKLTPLCVTLNCNNVNTNNTNNSTNATISDWEKMETGEMKNCSFNVTTSIRDK
IKKEYALFYKLDVVPLENKNNINNTNITNYRLINCNTSVITQACPKVSFEPIPIHYCAPAGFAILKCNSKT
FNGSGPCTNVSTVQCTHGIRPVVSTQLLLNGSLAEEEIVIRSENITDNAKTIIVQLNEAVEINCTRPNNNT
RKSIHIGPGRWFYATGDIIGNIRQAHCNISKARWNETLGQIVAKLEEQFPNKTIIFNHSSGGDPEIVTHSF
NCGGEFFYCNTTPLFNSTWNNTRTDDYPTGGEQNITLQCRIKQIINMWQGVGKAMYAPPIRGQIRCSSNIT
GLLLTRDGGRDQNGTETFRPGGGNMRDNWRSELYKYKVVKIEPLGIAPTACKRRVVQRRRRRRAVGLGAFI
LGFLGAAGSTMGAASMALTVQARLLLSGIVQQQNNLLRAPEAQQHMLQLTVWGIKQLQARVLAVERYLRDQ
QLLGIWGCSGKIICCTNVPWNDSWSNKTINEIWDNMTWMQWEKEIDNYTQHIYTLLEVSQIQQEKNEQELL
ELD
49
CH119 SOSIP
AENLWVTVYYGVPVWKEATTTLFCASDAKAYDTEVHNVWATHACVPTDPSPQELVLENVTENFNMWKNEMV
NQMHEDVISLWDQSLKPCVKLTPLCVTLECSKVSNNETDKYNGTEEMKNCSFNATTVVRDRQQKVYALFYR
LDIVPLTEKNSSENSSKYYRLINCNTSACTQACPKVSFEPIPIHYCTPAGYAILKCNDKTFNGTGPCHNVS
TVQCTHGIKPVVSTQLLLNGSLAEGEIIIRSENLTNNVKTILVHLNQSVEIVCTRPNNNTRKSIRIGPGQT
FYATGDIIGDIRQAHCNISKWHETLKRVSEKLAEHFPNKTINFTSSSGGDLEITTHSFTCRGEFFYCNTSG
LFNSTYMPNGTYLHGDTNSNSSITIPCRIKQIINMWQEVGRCMYAPPIEGNITCKSNITGLLLVRDGGTES
NNTETNNTEIFRPGGGDMRDNWRSELYKYKVVEIKPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAA
GSTMGAASMTLTVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWG
CSGKLICCTNVPWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
CE0217 SOSIP
AENLWVTVYYGVPVWREAKTTLFCASDAKAYEREVHNVWATHACVPTDPNPQERVLENVTENFNMWKNNMV
DQMHEDIISLWDESLKPCIKLTPLCVTLNCGNAIVNESTIEGMKNCSFNVTTELKDKKKKEYALFYKLDVV
PLNGENNNSNSKNFSEYRLINCNTSTCTQACPKVSFDPIPIHYCAPAGFAILKCNNETFNGTGPCNNVSTV
QCTHGIKPVVSTQLLLNGSLAEKEIIIRSENLTNNAKIIIVHLNNPVKIICTRPGNNTRKSMRIGPGQTFY
ATGDIIGDIRRAYCNISEKTWYDTLKNVSDKFQEHFPNASIEFKPSAGGDLEITTHSFNCRGEFFYCDTSE
LFNGTYNNSTYNSSNNITLQCKIKQIINMWQGVGRCMYAPPIAGNITCESNITGLLLTRDGGNNKSTPETF
RPGGGDMRDNWRSELYKYKVVEIKPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAASMTL
TVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLICCTNV
PWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
CNE8 SOSIP
AENLWVTVYYGVPVWRDADTTLFCASDAKAYDTEVHNVWATHACVPTDPNPQEIHLENVTENFNMWKNKMA
EQMQEDVISLWDESLKPCVQLTPLCVTLNCTNANLNATVNASTTIGNITDEVRNCSFNTTTELRDKKQNVY
ALFYKLDIVPINNNSEYRLINCNTSVCKQACPKVSFDPIPIHYCAPAGYAILRCNDKNFNGTGPCKNVSSV
QCTHGIKPVVSTQLLLNGSLAEDEIIIRSENLTDNVKTIIVHLNKSVEINCTRPSNNTRTSVRIGPGQVFY
RTGDIIGDIRKAYCNISRTKWHETLKQVATKLREHFNKTIIFQPPSGGDIEITMHHFNCRGEFFYCNTTKL
FNSTWGENTTMEGHNDTIVLPCRIKQIVNMWQGVGQCMYAPPIRGSINCVSNITGILLTRDGGTNMSNETF
RPGGGNIKDNWRSELYKYKVVEIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAASMTL
TVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLICCTNV
PWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
CNE8 N276A
SOSIP
AENLWVTVYYGVPVWRDADTTLFCASDAKAYDTEVHNVWATHACVPTDPNPQEIHLENVTENFNMWKNKMA
EQMQEDVISLWDESLKPCVQLTPLCVTLNCTNANLNATVNASTTIGNITDEVRNCSFNTTTELRDKKQNVY
ALFYKLDIVPINNNSEYRLINCNTSVCKQACPKVSFDPIPIHYCAPAGYAILRCNDKNFNGTGPCKNVSSV
QCTHGIKPVVSTQLLLNGSLAEDEIIIRSEALTDNVKTIIVHLNKSVEINCTRPSNNTRTSVRIGPGQVFY
RTGDIIGDIRKAYCNISRTKWHETLKQVATKLREHFNKTIIFQPPSGGDIEITMHHFNCRGEFFYCNTTKL
FNSTWGENTTMEGHNDTIVLPCRIKQIVNMWQGVGQCMYAPPIRGSINCVSNITGILLTRDGGTNMSNETF
RPGGGNIKDNWRSELYKYKVVEIEPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAASMTL
TVQARNLLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLICCTNV
PWNSSWSNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
CNE20 SOSIP
NLWVTVYYGVPVWKEATTTLFCASDAKAYDTEVHNVWATHACVPTDPNPHELVLENVTENFNMWKNEMVNQ
MHEDVISLWDQSLKPCVKLTPLCVTLECGNITTRKESMTEMKNCSFNATTVVKDRKQTVYALFYKLDIVPL
SGKNSSGYYRLINCNTSACTQACPKVNFDPIPIHYCTPAGYAILKCNDKTFNGTGPCHNVSTVQCTHGIKP
VISTQLLLNGSLAEGEIVIRSENLTNNAKIIIVHLNQTVEIVCTRPGNNTRKSIRIGPGQTFYATGEIIGN
IRQAHCNISENQWHKTLQNVSKKLAEHFQNKTITFASSSGGDLEITTHSFNCRGEFFYCNTSGLFNGTYMS
NNTEGNSSSIITIPCRIKQIINMWQEVGRCIYAPPIEGNITCKSNITGLLLERDGGTESNDTEIFRPGGGD
MRNNWRSELYKYKVVEIKPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAASMTLTVQARN
LLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLICCTNVPWNSSW
SNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
CNE20 N276A
SOSIP
NLWVTVYYGVPVWKEATTTLFCASDAKAYDTEVHNVWATHACVPTDPNPHELVLENVTENFNMWKNEMVNQ
MHEDVISLWDQSLKPCVKLTPLCVTLECGNITTRKESMTEMKNCSFNATTVVKDRKQTVYALFYKLDIVPL
SGKNSSGYYRLINCNTSACTQACPKVNFDPIPIHYCTPAGYAILKCNDKTFNGTGPCHNVSTVQCTHGIKP
VISTQLLLNGSLAEGEIVIRSEALTNNAKIIIVHLNQTVEIVCTRPGNNTRKSIRIGPGQTFYATGEIIGN
IRQAHCNISENQWHKTLQNVSKKLAEHFQNKTITFASSSGGDLEITTHSFNCRGEFFYCNTSGLFNGTYMS
NNTEGNSSSIITIPCRIKQIINMWQEVGRCIYAPPIEGNITCKSNITGLLLERDGGTESNDTEIFRPGGGD
MRNNWRSELYKYKVVEIKPLGVAPTRCKRRVVGRRRRRRAVGIGAVSLGFLGAAGSTMGAASMTLTVQARN
LLSGIVQQQSNLLRAPEPQQHLLKDTHWGIKQLQARVLAVEHYLRDQQLLGIWGCSGKLICCTNVPWNSSW
SNRNLSEIWDNMTWLQWDKEISNYTQIIYGLLEESQNQQEKNEQDLLALD
IOMA HC Fab
EVQLVESGAQVKKPGASVTVSCTASGYKFTGYHMHWVRQAPGRGLEWMGWINPFRGAVKYPQNFRGRVSMT
RDTSMEIFYMELSRLTSDDTAVYYCAREMFDSSADWSPWRGMVAWGQGTLVTVSSASTKGPSVFPLAPSSK
STSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNH
KPSNTKVDKRVEPKSCDKT
50
IOMA HC
EVQLVESGAQVKKPGASVTVSCTASGYKFTGYHMHWVRQAPGRGLEWMGWINPFRGAVKYPQNFRGRVSMT
RDTSMEIFYMELSRLTSDDTAVYYCAREMFDSSADWSPWRGMVAWGQGTLVTVSSASTKGPSVFPLAPSSK
STSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNH
KPSNTKVDKRVEPKSCDKTHTCPPCPAPELLGGPSVFLFPPKPKDTLMISRTPEVTCVVVDVSHEDPEVKF
NWYVDGVEVHNAKTKPREEQYNSTYRVVSVLTVLHQDWLNGKEYKCKVSNKALPAPIEKTISKAKGQPREP
QVYTLPPSREEMTKNQVSLTCLVKGFYPSDIAVEWESNGQPENNYKTTPPVLDSDGSFFLYSKLTVDKSRW
QQGNVFSCSVMHEALHNHYTQKSLSLSPGK
IOMA LC
QSALTQPASVSGSPGQSITISCAGSSRDVGGFDLVSWYQQHPGKAPKLIIYEVNKRPSGISSRFSASKSGN
TASLTISGLQEEDEAHYYCYSYADGVAFGGGTKLTVLGQPKAAPSVTLFPPSSEELQANKATLVCLISDFY
PGAVTVAWKADSSPVKAGVETTTPSKQSNNKYAASSYLSLTPEQWKSHRSYSCQVTHEGSTVEKTVAPTEC
S
IOMA iGL HC Fab
QVQLVQSGAEVKKPGASVKVSCKASGYTFTGYYMHWVRQAPGQGLEWMGWINPNSGGTNYAQKFQGRVTMT
RDTSISTAYMELSRLRSDDTAVYYCARDFTSSYDSSGYYHEGYWGQGTLVTVSSASTKGPSVFPLAPSSKS
TSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHK
PSNTKVDKRVEPKSCDKT
IOMA iGL HC
QVQLVQSGAEVKKPGASVKVSCKASGYTFTGYYMHWVRQAPGQGLEWMGWINPNSGGTNYAQKFQGRVTMT
RDTSISTAYMELSRLRSDDTAVYYCARDFTSSYDSSGYYHEGYWGQGTLVTVSSASTKGPSVFPLAPSSKS
TSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHK
PSNTKVDKRVEPKSCDKTHTCPPCPAPELLGGPSVFLFPPKPKDTLMISRTPEVTCVVVDVSHEDPEVKFN
WYVDGVEVHNAKTKPREEQYNSTYRVVSVLTVLHQDWLNGKEYKCKVSNKALPAPIEKTISKAKGQPREPQ
VYTLPPSREEMTKNQVSLTCLVKGFYPSDIAVEWESNGQPENNYKTTPPVLDSDGSFFLYSKLTVDKSRWQ
QGNVFSCSVMHEALHNHYTQKSLSLSPGK
IOMA iGL LC
QSALTQPASVSGSPGQSITISCTGTSSDVGSYNLVSWYQQHPGKAPKLMIYEVSKRPSGVSNRFSGSKSGN
TASLTISGLQAEDEADYYCCSYAGSVAFGGGTKLTVLGQPKAAPSVTLFPPSSEELQANKATLVCLISDFY
PGAVTVAWKADSSPVKAGVETTTPSKQSNNKYAASSYLSLTPEQWKSHRSYSCQVTHEGSTVEKTVAPTEC
S
VRC01 iGL HC
QVQLVQSGAEVKKPGASVKVSCKASGYTFTGYYMHWVRQAPGQGLEWMGWINPNSGGTNYAQKFQGRVTMT
RDTSISTAYMELSRLRSDDTAVYYCARGKNSDYNWDFQHWGQGTLVTVSSASTKGPSVFPLAPSSKSTSGG
TAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNT
KVDKRVEPKSCDKTHTCPPCPAPELLGGPSVFLFPPKPKDTLMISRTPEVTCVVVDVSHEDPEVKFNWYVD
GVEVHNAKTKPREEQYNSTYRVVSVLTVLHQDWLNGKEYKCKVSNKALPAPIEKTISKAKGQPREPQVYTL
PPSREEMTKNQVSLTCLVKGFYPSDIAVEWESNGQPENNYKTTPPVLDSDGSFFLYSKLTVDKSRWQQGNV
FSCSVMHEALHNHYTQKSLSLSPGK
VRC01 iGL LC
EIVLTQSPATLSLSPGERATLSCRASQSVSSYLAWYQQKPGQAPRLLIYDASNRATGIPARFSGSGSGTDF
TLTISSLEPEDFAVYYCQQYEFFGQGTKLEIKRTVAAPSVFIFPPSDEQLKSGTASVVCLLNNFYPREAKV
QWKVDNALQSGNSQESVTEQDSKDSTYSLSSTLTLSKADYEKHKVYACEVTHQGLSSPVTKSFNRGEC
3BNC60 iGL HC
QVQLVQSGAEVKKPGASVKVSCKASGYTFTGYYMHWVRQAPGQGLEWMGWINPNSGGTNYAQKFQGRVTMT
RDTSISTAYMELSRLRSDDTAVYYCARERSDFWDFDLWGRGTLVTVSSASTKGPSVFPLAPSSKSTSGGTA
ALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNTKV
DKRVEPKSCDKTHTCPPCPAPELLGGPSVFLFPPKPKDTLMISRTPEVTCVVVDVSHEDPEVKFNWYVDGV
EVHNAKTKPREEQYNSTYRVVSVLTVLHQDWLNGKEYKCKVSNKALPAPIEKTISKAKGQPREPQVYTLPP
SREEMTKNQVSLTCLVKGFYPSDIAVEWESNGQPENNYKTTPPVLDSDGSFFLYSKLTVDKSRWQQGNVFS
CSVMHEALHNHYTQKSLSLSPGK
3BNC60 iGL LC
DIQMTQSPSSLSASVGDRVTITCQASQDISNYLNWYQQKPGKAPKLLIYDASNLETGVPSRFSGSGSGTDF
TFTISSLQPEDIATYYCQQYEFIGPGTKVDIKRTVAAPSVFIFPPSDEQLKSGTASVVCLLNNFYPREAKV
QWKVDNALQSGNSQESVTEQDSKDSTYSLSSTLTLSKADYEKHKVYACEVTHQGLSSPVTKSFNRGEC
BG24 iGL HC
QVQLVQSGAEVKKPGASVKVSCKASGYTFTGYYMHWVRQAPGQGLEWMGWINPNSGGTNYAQKFQGRVTMT
RDTSISTAYMELSRLRSDDTAVYYCATQLELDSSAGYAFDIWGQGTMVTVSSASTKGPSVFPLAPSSKSTS
GGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPS
NTKVDKRVEPKSCDKTHTCPPCPAPELLGGPSVFLFPPKPKDTLMISRTPEVTCVVVDVSHEDPEVKFNWY
VDGVEVHNAKTKPREEQYNSTYRVVSVLTVLHQDWLNGKEYKCKVSNKALPAPIEKTISKAKGQPREPQVY
TLPPSREEMTKNQVSLTCLVKGFYPSDIAVEWESNGQPENNYKTTPPVLDSDGSFFLYSKLTVDKSRWQQG
NVFSCSVMHEALHNHYTQKSLSLSPGK
51
BG24 iGL LC
QSALTQPRSVSGSPGQSVTISCTGTSSDVGGYNYVSWYQQHPGKAPKLMIYDVSKRPSGVPDRFSGSKSGN
TASLTISGLQAEDEADYYCSSYEYFGGGTKLTVLSQPKAAPSVTLFPPSSEELQANKATLVCLISDFYPGA
VTVAWKADSSPVKAGVETTTPSKQSNNKYAASSYLSLTPEQWKSHRSYSCQVTHEGSTVEKTVAPTECS
52
Table S2: X-ray data collection for IOMA iGL Fab crystals
Space group
P 21 21 21
Cell dimensions
a, b, c (Å)
57.7, 66.7, 166.3
α, β, γ (°)
90, 90, 90
Resolution (Å)
38.6–2.07 (2.15–2.07)a
R merge
0.08 (0.58)
R pim
0.05 (0.36)
I/σ(I)
9.7 (2.5)
CC 1/2
0.99 (0.92)
Completeness (%)
99 (99)
Redundancy
6.3 (6.6)
Refinement
Resolution (Å)
38.6–2.07
No. reflections
39,372
Rwork / Rfree
0.224 / 0.257
No. atoms
Protein
3,241
Ligand/ion
N/A
B factors (Å2)
Protein
46.7
Ligand/ion
N/A
R.m.s. deviations
Bond lengths (Å)
0.008
Bond angles (°)
1.00
a Values in parentheses are for the highest-resolution shell.
53
Table S3: Serum neutralization data for IOMA iGL transgenic mice
ES30
ES32
ES34
ES37
ET33
ET34
B3 (week 18)
B3 (week 18)
B3 (week 18)
B3 (week 18)
B3 (week 18)
B3 (week 18)
Virus
Clade
Tier
ID50
%
1:100
ID50
%
1:100
ID50
%
1:100
ID50
%
1:100
ID50
%
1:100
ID50
%
1:100
426c
C
2
–
–
–
–
–
–
–
–
<100
0
<100
0
25710
B
2
–
–
–
–
–
–
–
–
<100
0
<100
0
CNE8
AE
1
<100
0
<100
0
<100
0
<100
0
<100
0
104
54
CNE8
N276A
AE
1
463
71
<100
0
<100
0
<100
0
<100
0
<100
40
CNE20
BC
2
<100
49
<100
0
<100
0
<100
0
<100
0
<100
0
CNE20
N276A
BC
2
14,922
95
<100
0
<100
0
<100
0
<100
0
<100
36
JRCSF
B
2
136
65
<100
0
<100
0
<100
0
<100
0
<100
0
Q23.17
A
1
100
51
<100
0
<100
0
<100
0
<100
0
<100
26
YU2
B
2
571
86
<100
0
<100
0
<100
0
<100
0
112
56
BG505
T332N
A
2
–
–
–
–
–
–
–
–
<100
0
<100
0
6535.5
B
1
–
–
–
–
–
–
–
–
<100
0
104
53
3415_V1_C1
A
2
–
–
–
–
–
–
–
–
154
53
102
59
CAAN5342.
A2
B
2
–
–
–
–
–
–
–
–
<100
0
<100
41
PVO.4
B
3
113
52
<100
0
<100
0
<100
0
<100
22
<100
57
Q842.D12
A
2
–
–
–
–
–
–
–
–
<100
0
<100
0
RHPA4259.7
B
2
–
–
–
–
–
–
–
–
<100
0
<100
43
WITO4160.3
3
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
45
ZM214M.PL
15
C
2
–
–
–
–
–
–
–
–
<100
0
<100
0
MuLV
<100
0
<100
0
<100
0
<100
0
<100
0
162
65
54
HP1
HP2
HP3
HP4
HP6
HP7
HQ4
B4
(week 23)
B4
(week 23)
B4
(week 23)
B4
(week 23)
B4
(week 23)
B4
(week 23)
B4
(week 23)
Virus
Cla
de
Tier
ID50
%
1:100
ID50
%
1:100
ID50
%
1:100
ID50
%
1:100
ID50
%
1:100
ID50
%
1:100
ID50
%
1:100
426c
C
2
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
25710
B
2
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
CNE8
AE
1
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
CNE8
N276A
AE
1
<100
26
<100
0
<100
0
<100
12
–
–
<100
0
<100
0
CNE20
BC
2
<100
23
<100
0
<100
40
<100
0
–
–
<100
0
<100
0
CNE20
N276A
BC
2
338
83
610
79
903
88
<100
16
–
–
2,017
98
<100
0
JRCSF
B
2
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
Q23.17
A
1
120
57
<100
30
<100
28
<100
0
–
–
<100
0
<100
0
YU2
B
2
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
BG505
T332N
A
2
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
6535.5
B
1
165
67
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
3415_V
1_C1
A
2
<100
0
<100
40
<100
0
<100
0
–
–
<100
0
108
50
CAAN5
342.A2
B
2
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
PVO.4
B
3
<100
43
<100
45
<100
12
<100
35
–
–
115
50
<100
0
Q842.D
12
A
2
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
RHPA4
259.7
B
2
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
WITO4
160.33
B
2
<100
28
145
53
<100
0
<100
48
–
–
<100
40
<100
0
ZM214
M.PL15
C
2
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
MuLV
<100
0
<100
0
<100
0
<100
0
–
–
<100
0
<100
0
55
Table S4: Mutational analysis of antibodies isolated from IOMA iGL transgenic mice.
VH1-2*02
amino
acid
VH
Position
amino acid
substitution
Random
Frequency
# of IGT2-induced mAbs with this
SHM
IGT2
Frequency
IOMA
Substitution
Critical
Interaction
VRC01
Substitution
Q
1
–
67
100.0
E
E
0.0
0
0.0
V
2
67
100.0
V
Q
3
67
100.0
Q
L
4
67
100.0
L
V
5
67
100.0
V
Q
6
67
100.0
E
E
0.1
0
0.0
S
7
67
100.0
S
G
8
67
100.0
G
A
9
67
100.0
A
G
E
10
67
100.0
Q
Q
Q
0.2
0
0.0
V
11
66
98.5
V
M
M
3.3
1
1.5
K
12
56
83.6
K
R
6.9
11
16.4
K
13
67
100.0
K
P
14
67
100.0
P
G
15
67
100.0
G
E
A
16
67
100.0
A
S
17
67
100.0
S
V
18
67
100.0
V
M
K
19
22
32.8
T
YES
R
T
2.3
31
46.3
R
9.6
14
20.9
V
20
67
100.0
V
I
S
21
67
100.0
S
C
22
67
100.0
C
K
23
55
82.1
T
R
A
0.2
1
1.5
T
2.3
0
0.0
E
3.3
5
7.5
R
4.6
6
9.0
A
24
61
91.0
A
T
13.5
6
9.0
S
25
67
100.0
S
G
26
67
100.0
G
56
Y
27
67
100.0
Y
T
28
65
97.0
K
E
K
0.6
0
0.0
N
2.0
2
3.0
F
29
66
98.5
F
L
3.8
1
1.5
T
30
58
86.6
T
I
A
1.7
1
1.5
I
7.2
8
11.9
G
31
32
47.8
G
D
E
0.8
1
1.5
A
11.7
6
9.0
D
34.3
28
41.8
Y
32
65
97.0
Y
C
H
8.5
2
3.0
Y
33
27
40.3
H
YES
T
E
0.1
14
20.9
D
0.6
9
13.4
S
1.9
2
3.0
H
4.5
8
11.9
F
8.4
7
10.4
M
34
33
49.3
M
L
L
13.8
7
10.4
I
48.4
27
40.3
H
35
61
91.0
H
N
Q
2.2
1
1.5
Y
3.5
5
7.5
W
36
67
100.0
W
V
37
67
100.0
V
I
R
38
67
100.0
R
Q
39
66
98.5
Q
L
R
1.1
1
1.5
A
40
65
97.0
A
V
2.0
2
3.0
P
41
67
100.0
P
G
42
67
100.0
G
Q
43
66
98.5
R
K
R
1.5
1
1.5
G
44
67
100.0
G
R
L
45
64
95.5
L
P
F
1.9
3
4.5
E
46
66
98.5
E
57
D
0.5
1
1.5
W
47
67
100.0
W
M
48
62
92.5
M
L
4.9
4
6.0
V
6.4
1
1.5
G
49
67
100.0
G
W
50
52
77.6
W
R
0.0
14
20.9
L
0.8
1
1.5
I
51
66
98.5
I
L
S
0.5
1
1.5
N
52
58
86.6
N
K
H
2.2
3
4.5
S
4.4
6
9.0
P
(52A)
67
100.0
P
N
53
15
22.4
F
YES
R
F
0.1
30
44.8
E
1.0
2
3.0
T
1.4
5
7.5
R
1.8
7
10.4
Y
3.5
5
7.5
D
8.0
1
1.5
K
13.5
2
3.0
S
54
12
17.9
R
YES
G
F
0.2
7
10.4
R
2.7
45
67.2
N
11.9
1
1.5
T
14.8
2
3.0
G
55
67
100.0
G
G
56
31
46.3
A
YES
A
R
0.9
2
3.0
N
0.9
8
11.9
V
6.3
5
7.5
A
11.3
20
29.9
D
22.4
1
1.5
T
57
21
31.3
V
YES
V
V
0.4
21
31.3
R
0.9
3
4.5
P
1.3
2
3.0
I
1.4
18
26.9
S
1.9
2
3.0
N
58
18
26.9
K
YES
58
G
0.7
10
14.9
E
1.4
15
22.4
D
7.2
8
11.9
K
16.5
16
23.9
Y
59
52
77.6
Y
C
0.7
2
3.0
S
5.6
13
19.4
A
60
42
62.7
P
R
0.1
3
4.5
E
1.6
3
4.5
T
1.9
6
9.0
P
2.1
0
0.0
V
2.2
8
11.9
S
3.0
5
7.5
Q
61
56
83.6
Q
R
R
2.8
3
4.5
E
4.3
8
11.9
K
62
64
95.5
N
P
R
7.5
1
1.5
N
9.4
2
3.0
F
63
66
98.5
F
L
L
2.4
1
1.5
Q
64
56
83.6
R
R
4.8
11
16.4
G
65
67
100.0
G
R
66
67
100.0
R
V
67
66
98.5
V
L
2.3
1
1.5
T
68
64
95.5
S
I
2.4
2
3.0
S
4.4
1
1.5
M
69
64
95.5
M
L
12.3
3
4.5
T
70
67
100.0
T
R
71
67
100.0
R
YES
D
72
67
100.0
D
T
73
66
98.5
T
V
P
0.8
1
1.5
S
74
66
98.5
S
Y
T
0.7
1
1.5
I
75
67
100.0
M
S
M
1.5
0
0.0
59
S
76
37
55.2
E
D
E
0.1
0
0.0
I
0.3
1
1.5
K
0.6
1
1.5
R
3.1
1
1.5
T
13.3
20
29.9
N
18.0
7
10.4
T
77
64
95.5
I
I
0.6
3
4.5
A
78
60
89.6
F
F
0.5
0
0.0
T
2.1
1
1.5
V
15.3
6
9.0
Y
79
67
100.0
Y
F
M
80
64
95.5
M
L
L
9.1
3
4.5
E
81
65
97.0
E
V
0.4
2
3.0
L
82
66
98.5
L
M
2.3
1
1.5
S
(82A)
45
67.2
S
R
K
1.2
1
1.5
R
8.1
2
3.0
N
8.9
16
23.9
T
12.8
3
4.5
R
(82B)
59
88.1
R
S
G
13.4
8
11.9
L
(82C)
66
98.5
L
V
0.9
1
1.5
R
83
60
89.6
T
T
N
0.7
1
1.5
I
1.7
1
1.5
K
7.9
2
3.0
T
29.6
3
4.5
S
84
66
98.5
S
V
Y
2.3
1
1.5
D
85
66
98.5
D
N
1.0
1
1.5
D
86
67
100.0
D
T
87
67
100.0
T
A
88
67
100.0
A
V
89
60
89.6
V
60
R
0.1
1
1.5
A
0.3
1
1.5
M
4.6
1
1.5
I
10.8
4
6.0
Y
90
67
100.0
Y
Y
91
48
71.6
Y
F
N
0.0
2
3.0
F
14.6
17
25.4
C
92
67
100.0
C
A
93
65
97.0
A
T
T
3.5
1
1.5
V
5.1
1
1.5
R
94
67
100.0
R
61
VL2-
23*0
2
amini
o
acid
VL
Positio
n
amino
acid
substituti
on
Random
Frequen
cy
# of
IGT2
induce
d
mAbs
with
this
SHM
IGT2
Frequen
cy
IOMA
Substituti
on
Critical
Interactio
n
VK3-
20*01
Residu
e
VL
Positio
n
VRC01
Substituti
on
Random
Frequen
cy
Critical
Interactio
n
Q
1
67
100.0
Q
Q
1
Q
100.0
S
2
66
98.5
S
S
2
S
98.5
F
0.1
1
1.5
1.5
A
3
67
100.0
A
A
3
A
100.0
L
4
67
100.0
L
L
4
L
100.0
T
5
67
100.0
T
T
5
T
100.0
Q
6
67
100.0
Q
Q
6
Q
100.0
P
7
67
100.0
P
P
7
P
100.0
A
8
67
100.0
A
A
8
A
100.0
S
9
67
100.0
S
S
9
S
100.0
V
11
67
100.0
V
V
11
V
100.0
S
12
66
98.5
S
S
12
S
98.5
F
0.1
1
1.5
1.5
G
13
67
100.0
G
G
13
G
100.0
S
14
67
100.0
S
S
14
S
100.0
P
15
67
100.0
P
P
15
P
100.0
G
16
62
92.5
G
G
16
G
92.5
E
0.1
5
7.5
7.5
Q
17
67
100.0
Q
Q
17
Q
100.0
S
18
67
100.0
S
S
18
S
100.0
I
19
65
97.0
I
I
19
I
97.0
S
0.1
1
1.5
1.5
T
0.1
1
1.5
1.5
T
20
67
100.0
T
T
20
T
100.0
I
21
67
100.0
I
I
21
I
100.0
S
22
67
100.0
S
S
22
S
100.0
C
23
67
100.0
C
C
23
C
100.0
T
24
67
100.0
A
T
24
A
100.0
A
2.3
0
0.0
0.0
G
25
63
94.0
G
G
25
G
94.0
V
0.1
4
6.0
6.0
T
26
65
97.0
S
T
26
S
97.0
P
0.6
1
1.5
1.5
A
2.6
1
1.5
1.5
S
7.2
1
1.5
1.5
S
27
65
97.0
S
S
27
S
97.0
G
2.2
1
1.5
1.5
62
R
2.5
1
1.5
1.5
S
(27A)
63
94.0
R
YES
S
(27A)
R
94.0
YES
R
2.4
2
3.0
3.0
N
6.1
2
3.0
3.0
D
(27B)
67
100.0
D
D
(27B)
D
100.0
V
(27C)
53
79.1
V
V
(27C)
V
79.1
F
1.9
1
1.5
1.5
I
17.0
13
19.4
19.4
G
28
67
100.0
G
G
28
G
100.0
S
29
51
76.1
G
YES
S
29
G
76.1
YES
I
2.0
1
1.5
1.5
R
3.2
4
6.0
6.0
G
5.3
4
6.0
6.0
T
14.3
1
1.5
1.5
N
14.5
6
9.0
9.0
Y
30
56
83.6
F
YES
Y
30
F
83.6
YES
S
3.0
4
6.0
6.0
F
4.1
7
10.4
10.4
N
31
35
52.2
D
YES
N
31
D
52.2
YES
Y
1.3
4
6.0
6.0
D
10.2
28
41.8
41.8
L
32
65
97.0
L
L
32
L
97.0
F
8.2
2
3.0
3.0
V
33
67
100.0
V
V
33
V
100.0
S
34
66
98.5
S
S
34
S
98.5
P
0.0
1
1.5
1.5
W
35
67
100.0
W
W
35
W
100.0
Y
36
67
100.0
Y
Y
36
Y
100.0
Q
37
67
100.0
Q
Q
37
Q
100.0
Q
38
67
100.0
Q
Q
38
Q
100.0
H
39
67
100.0
H
H
39
H
100.0
P
40
67
100.0
P
P
40
P
100.0
G
41
67
100.0
G
G
41
G
100.0
K
42
66
98.5
K
K
42
K
98.5
N
0.7
1
1.5
1.5
A
43
61
91.0
A
A
43
A
91.0
T
0.8
5
7.5
7.5
V
7.6
1
1.5
1.5
P
44
67
100.0
P
P
44
P
100.0
K
45
67
100.0
K
K
45
K
100.0
L
46
66
98.5
L
L
46
L
98.5
F
2.4
1
1.5
1.5
63
M
47
63
94.0
I
M
47
I
94.0
L
10.7
2
3.0
3.0
I
34.4
2
3.0
3.0
I
48
66
98.5
I
I
48
I
98.5
L
2.6
1
1.5
1.5
Y
49
66
98.5
Y
Y
49
Y
98.5
H
1.8
1
1.5
1.5
E
50
51
76.1
E
E
50
E
76.1
K
0.3
4
6.0
6.0
D
6.5
12
17.9
17.9
V
51
67
100.0
V
V
51
V
100.0
S
52
46
68.7
N
S
52
N
68.7
I
3.2
1
1.5
1.5
N
19.8
14
20.9
20.9
T
29.2
6
9.0
9.0
K
53
38
56.7
K
K
53
K
56.7
A
0.2
1
1.5
1.5
R
3.8
23
34.3
34.3
Q
4.9
4
6.0
6.0
E
7.1
1
1.5
1.5
R
54
67
100.0
R
R
54
R
100.0
P
55
67
100.0
P
P
55
P
100.0
S
56
67
100.0
S
S
56
S
100.0
G
57
67
100.0
G
G
57
G
100.0
V
58
61
91.0
I
V
58
I
91.0
I
10.9
6
9.0
9.0
S
59
66
98.5
S
S
59
S
98.5
Y
0.0
1
1.5
1.5
N
60
65
97.0
S
N
60
S
97.0
S
7.3
1
1.5
1.5
D
17.3
1
1.5
1.5
R
61
67
100.0
R
R
61
R
100.0
F
62
67
100.0
F
F
62
F
100.0
S
63
65
97.0
S
S
63
S
97.0
A
0.4
2
3.0
3.0
G
64
65
97.0
A
G
64
A
97.0
D
0.1
1
1.5
1.5
A
5.3
1
1.5
1.5
S
65
67
100.0
S
S
65
S
100.0
K
66
67
100.0
K
K
66
K
100.0
S
67
65
97.0
S
S
67
S
97.0
C
0.0
1
1.5
1.5
64
A
0.5
1
1.5
1.5
G
68
65
97.0
G
G
68
G
97.0
D
3.3
2
3.0
3.0
N
69
66
98.5
N
N
69
N
98.5
K
0.8
1
1.5
1.5
T
70
64
95.5
T
T
70
T
95.5
M
0.8
2
3.0
3.0
S
0.8
1
1.5
1.5
A
71
67
100.0
A
A
71
A
100.0
S
72
67
100.0
S
S
72
S
100.0
L
73
67
100.0
L
L
73
L
100.0
T
74
56
83.6
T
T
74
T
83.6
P
0.0
2
3.0
3.0
I
0.5
9
13.4
13.4
I
75
67
100.0
I
I
75
I
100.0
S
76
67
100.0
S
S
76
S
100.0
G
77
66
98.5
G
G
77
G
98.5
D
0.5
1
1.5
1.5
L
78
32
47.8
L
L
78
L
47.8
F
0.0
35
52.2
52.2
Q
79
65
97.0
Q
Q
79
Q
97.0
R
2.5
2
3.0
3.0
A
80
59
88.1
E
A
80
E
88.1
E
0.1
0
0.0
0.0
D
0.3
1
1.5
1.5
T
4.9
6
9.0
9.0
P
5.1
1
1.5
1.5
E
81
67
100.0
E
E
81
E
100.0
D
82
67
100.0
D
D
82
D
100.0
E
83
57
85.1
E
E
83
E
85.1
V
0.0
3
4.5
4.5
K
0.1
2
3.0
3.0
G
0.2
5
7.5
7.5
A
84
66
98.5
A
A
84
A
98.5
G
4.2
1
1.5
1.5
D
85
61
91.0
H
D
85
H
91.0
G
0.2
2
3.0
3.0
Y
0.9
1
1.5
1.5
N
2.1
1
1.5
1.5
H
2.6
0
0.0
0.0
E
4.1
2
3.0
3.0
Y
86
65
97.0
Y
Y
86
Y
97.0
65
N
2
3.0
3.0
Y
87
45
67.2
Y
Y
87
Y
67.2
E
0.0
1
1.5
1.5
D
0.0
5
7.5
7.5
H
4.3
12
17.9
17.9
F
6.2
4
6.0
6.0
C
88
67
100.0
C
C
88
C
100.0
C
89
57
85.1
Y
C
89
Y
85.1
W
0.8
1
1.5
1.5
Y
1.8
4
6.0
6.0
S
10.3
5
7.5
7.5
S
90
60
89.6
S
S
90
S
89.6
L
0.7
6
9.0
9.0
T
1.0
1
1.5
1.5
Y
96
61
91.0
Y
Y
96
Y
91.0
C
0.7
1
1.5
1.5
S
2.1
3
4.5
4.5
F
7.4
2
3.0
3.0
A
97
48
71.6
A
A
97
A
71.6
E
1.4
1
1.5
1.5
T
5.8
3
4.5
4.5
G
7.2
9
13.4
13.4
V
8.8
6
9.0
9.0
66
Table S5: Serum neutralization in wildtype mice
M1
M3
M4
M5
M13
M14
M15
M21
B3 (week
18)
B3 (week
18)
B3 (week
18)
B3 (week
18)
B3 (week
18)
B3 (week
18)
B3 (week
18)
B4 (week
23)
Virus
Cla
de
Ti
er
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:
10
0
426c
C
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
25710
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
CNE8
AE
1
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
225
56
CNE8
N276A
AE
1
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
CNE20
BC
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
40
<100
0
CNE20
N276A
BC
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
JRCSF
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
Q23.17
A
1
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
40
<100
0
YU2
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
BG505
T332N
A
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
6535.5
B
1
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
47
<100
0
3415_
V1_C1
A
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
CAAN534
2.A2
B
2
<100
0
–
–
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
PVO.4
B
3
113
57
–
–
<100
0
<100
0
<100
41
<100
0
145
58
<100
0
Q842.D12
A
2
<100
43
–
–
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
RHPA425
9.7
B
2
<100
0
–
–
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
WITO416
0.33
B
2
<100
47
–
–
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
ZM214M.
PL15
C
2
<100
0
–
–
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
MuLV
<100
0
–
–
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
M22
M23
M24
M25
M26
M27
M28
M29
B4 (week
23)
B4 (week
23)
B4 (week
23)
B4 (week
23)
B4 (week
23)
B4 (week
23)
B4 (week
23)
B4 (week
23)
Virus
Cla
de
Ti
er
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:1
00
ID50
%
1:
10
0
426c
C
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
25710
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
CNE8
AE
1
<100
0
<100
0
100
50
<100
0
729
70
<100
0
841
61
242
71
CNE8
N276A
AE
1
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
CNE20
BC
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
CNE20
N276A
BC
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
67
JRCSF
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
Q23.17
A
1
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
YU2
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
BG505
T332N
A
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
41
<100
0
6535.5
B
1
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
3415_V1_
C1
A
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
CAAN534
2.A2
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
PVO.4
B
3
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
43
<100
0
Q842.D12
A
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
RHPA425
9.7
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
WITO416
0.33
B
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
ZM214M.
PL15
C
2
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
MuLV
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
<100
0
68
Table S6: Oligonucleotides used to generate yeast display gp120 libraries.
Oligo Name
Fragment
Sequence
426c Library 1 For
1
GTCTGGAAAGAGGCTAAGACCACACTG
426c Library 1 Rev
1
CAGGTTTTTTGATCTGATCACAATCTCTTC
426c Library 1 - 1 For
2
GAAGAGATTGTGATCAGATCAAAAAACCTGNNKAACAATGCCAAGATCATTATCGTGC
426c Library 1 - 2 Rev
2
ATCTCCACACTCTTATTCAGCTGCACGATAATGATCTTGGCATT
426c Library 1 - 3 For
2
AGCTGAATAAGAGTGTGGAGATCGTCTGCACACGACCTAACA
426c Library 1 - 4 Rev
2
GCCTGCCGAATATCTCCCCCAGATCCGCTGCCGCCATTGTTAGGTCGTGTGCAGACG
426c Library 1 - 5 For
2
GGAGATATTCGGCAGGCTTATTGTAACATCAGTGGCAGAAATTGGTCAGAAGCCGTGAA
426c Library 1 - 6 Rev
2
TGGGGGAAGTGCTCTTTCAGCTTTTTCTTGACCTGGTTCACGGCTTCTGACCAATTT
426c Library 1 - 7 For
2
AAAGAGCACTTCCCCCATAAGAATATTAGCTTTCAGTCTAGTTCAGGCGGGGAC
426c Library 1 - 8 Rev
2
TCGCCTCCGCAGTTGAAGGAGTGTGTGGTGATTTCCAGGTCCCCGCCTGAACTA
426c Library 1 - 9 For
2
ACTGCGGAGGCGAGTTCTTTTACTGTAATACATCCGGCCTGTTTAACG
426c Library 1 - 10 Rev
2
CCGGCAAGGCAGCATGATTGTGGCATTAGAAATGGTATCGTTAAACAGGCCGGATGTA
426c Library 1 - 11 For
2
GCTGCCTTGCCGGATCAAGCAGATTATCAACATGTGGCAGGAA
426c Library 1 - 12 Rev
2
TGCCCTTGATGGGTGGTGCATAGATAGCCTTTCCMNNTTCCTGCCACATGTTGATAATC
426c Library 1 - 13 For
2
CCACCCATCAAGGGCAATATCACCTGTAAGAGTGACATTACAGGGCTGCTGCTGCTGAGA
426c Library 1 - 14 Rev
2
GCCGGAAAATCTCGGTmnnmnnmnnmnnmnnTCCCCCATCTCTCAGCAGCAGCAGCC
426c Library 1 - 15 For
2
ACCGAGATTTTCCGGCCTAGCGGAGGAGACATGCGAGATAATTGGCGGTCTGAACTG
426c Library 1 - 16 Rev
2
GGATCCCAGAGGCTTGATCTCGACCACCTTATATTTGTACAGTTCAGACCGCCAATTA
426c Library 1 For
3
CTGGTGGAGGCGGTAGCGGAGGCGGAGGGTCGGCTAGCGTCTGGAAAGAGGCTAAGACCA
426c Library 1 Rev
3
TTACAAGTCCTCTTCAGAAATAAGCTTTTGTTCGGATCCCAGAGGCTTGATCTCGACCAC
426c Library 2 For
1
GTCTGGAAAGAGGCTAAGACCACACTG
426c Library 2 Rev
1
CAGGTTTTTTGATCTGATCACAATCTCTTC
426c Library 2 - 1 For
2
GAAGAGATTGTGATCAGATCAAAAAACCTGNNKAACAATGCCAAGATCATTATCGTGC
426c Library 2 - 2 Rev
2
ATCTCCACACTCTTATTCAGCTGCACGATAATGATCTTGGCATT
426c Library 2 - 3 For
2
AGCTGAATAAGAGTGTGGAGATCGTCTGCACACGACCTAACA
426c Library 2 - 4 Rev
2
GCCTGCCGAATATCTCCCCCAGATCCGCTGCCGCCATTGTTAGGTCGTGTGCAGACG
426c Library 2 - 5 For
2
GGAGATATTCGGCAGGCTTATTGTAACATCAGTGGCAGAAATTGGTCAGAAGCCGTGAA
426c Library 2 - 6 Rev
2
TGGGGGAAGTGCTCTTTCAGCTTTTTCTTGACCTGGTTCACGGCTTCTGACCAATTT
426c Library 2 - 7 For
2
AAAGAGCACTTCCCCCATAAGAATATTAGCTTTCAGTCTAGTTCAGGCGGGGAC
426c Library 2 - 8 Rev
2
TCGCCTCCGCAGTTGAAGGAGTGTGTGGTGATTTCCAGGTCCCCGCCTGAACTA
426c Library 2 - 9 For
2
ACTGCGGAGGCGAGTTCTTTTACTGTAATACATCCGGCCTGTTTAACG
426c Library 2 - 10 Rev
2
CCGGCAAGGCAGCATGATTGTGGCATTAGAAATGGTATCGTTAAACAGGCCGGATGTA
426c Library 2 - 11 For
2
GCTGCCTTGCCGGATCAAGCAGATTATCAACATGTGGCAGGAA
426c Library 2 - 12 Rev
2
TGCCCTTGATGGGTGGTGCATAGATAGCCTTTCCMNNTTCCTGCCACATGTTGATAATC
426c Library 2 - 13 For
2
CCACCCATCAAGGGCAATATCACCTGTAAGAGTGACATTACAGGGCTGCTGCTGCTGAGA
426c Library 2 - 14 Rev
2
GCCGGAAAATCTCGGTmnnmnnmnnmnnmnnTCCCCCATCTCTCAGCAGCAGCAGCC
426c Library 2 - 15 For
2
ACCGAGATTTTCCGGCCTAGCGGAGGAGACATGCGAGATAATTGGCGGTCTGAACTG
426c Library 2 - 16 Rev
2
GGATCCCAGAGGCTTGATCTCGACCACCTTATATTTGTACAGTTCAGACCGCCAATTA
426c Library 2 For
3
CTGGTGGAGGCGGTAGCGGAGGCGGAGGGTCGGCTAGCGTCTGGAAAGAGGCTAAGACCA
426c Library 2 Rev
3
TTACAAGTCCTCTTCAGAAATAAGCTTTTGTTCGGATCCCAGAGGCTTGATCTCGACCAC
69
Table S7: Flow cytometric reagents.
Reagent
Target
species
Antibody
clone
Company / Source
Cat.#
CD16/32
mouse
2.4G2
BD Biosciences
7248907
CD4-APCeF780
mouse
RM4-5
Thermo Fisher
47-0042-82
CD8a-APCeF780
mouse
53-6.7
Thermo Fisher
47-0081-82
NK1.1-APCeF780
mouse
PK136
Thermo Fisher
47-5941-82
F4/80-APCeF780
mouse
BM8
Thermo Fisher
47-4801-82
Ly-6G/C (Gr1)-APCeF780
mouse
RB6-8C5
Thermo Fisher
47-5931-82
CD11b-APCeF780
mouse
M1/70
Thermo Fisher
47-0112-82
CD11c-APCeF780
mouse
N418
Thermo Fisher
47-0114-82
CD93-APC
mouse
AA4.1
Thermo Fisher
17-5892-82
TER-119-APCCy0
mouse
TER-119
BD Pharmingen
560509
CD95 (FAS)-FITC
mouse
SA367H8
BioLegend
152606
CD38-AF700
mouse
90
Thermo Fisher
56-0381-82
CD45R/B220-BV421
mouse /
human
RA3-6B2
BD Horizon
562922
CD45R/B220-BV605
mouse /
human
RA3-6B2
BioLegend
103244
IgD-BV786
mouse
11-26c.2a
BD Horizon
563618
CD19-PECy7
mouse
6D5
BioLegend
115520
CD2-PE
mouse
RM2-5
BioLegend
100108
CD23-PE
mouse
B3B4
BioLegend
101607
Ig light chain lambda-APC
mouse
RML-42
BioLegend
407306
Ig light chain kappa-BV421
mouse
187.1
BD Horizon
562888
CD21/CD35
mouse
7G6
BD Horizon
562756
IgM Fab-FITC
mouse
polyclonal
Jackson
Immunoresearch
115-097-
020
Zombie NIR
N/A*
N/A
BioLegend
423105
Streptavidin-PE
N/A
N/A
BD Pharmingen
554061
Streptavidin-AF647
N/A
N/A
BioLegend
405237
Streptavidin-PECy7
N/A
N/A
BioLegend
405206
RC1-biotin
N/A
N/A
in house
N/A
CNE8 N276A-biotin
N/A
N/A
in house
N/A
426c degly2 D279N-biotin
N/A
N/A
in house
N/A
426c degly2 D279N
CD4bs-KO -biotin
N/A
N/A
in house
N/A
Human Fc Block
human
N/A
BD Horizon
564220
Ig light chain lambda-APC
human
MHL38
BioLegend
316610
CD19-PECy7
human
SJ25C1
BioLegend
363012
IgM-FITC
human
MHM88
BioLegend
314506
Ig light chain kappa-BV421
human
MHK-49
BioLegend
316518
*N/A – not applicable
70
Table S8: Single cell antibody cloning reaction conditions.
PCR1 IgH
Primer sequence
PCR1 mastermix
HH_1FL (forward,
leader)
CCATGGGATGGTCATGTATCA
Reagent
Volume/plate (µL)
Concentration
HH_1RG (reverse,
IgG)
GGACAGGGATCCAGAGTTCC
nuclease free water
3328
HH_1RM (reverse,
IgM)
CCCATGGCCACCAGATTCTT
10x buffer
384
1x
dNTP (25 mM)
48
0.3 mM
PCR1 IgK
Primer sequence
5' forward Primer (50 µM)
HC 15; LC 19
HC 0.25 µM; LC
0.25 µM
HH_1FL (forward,
leader)
CCATGGGATGGTCATGTATCA
3' reverse Primer (50 µM)
HC 23 (IgG/IgM 1:1);
LC 19
HC 0.30 µM; LC
0.25 µM
HH_1RK (reverse,
IgK)
GACTGAGGCACCTCCAGATG
HotStar DNA Polymerase (5
U/µL)
42
0.055 U/µL
total
3840
PCR2 IgH
Primer sequence
PCR2 mastermix
HH_2FL (forward,
leader)
GTAGCAACTGCAACCGGTGTACATTCT Reagent
Volume/plate (µL)
Concentration
HH_2RG (reverse,
IgG)
GCTCAGGGAARTAGCCCTTGAC
nuclease free water
2536
HH_2RM (reverse,
IgM)
AGGGGGAAGACATTTGGGAAGGAC
loading buffer*
800
10x buffer
384
1x
dNTP (25 mM)
48
0.3 mM
PCR2 IgK
Primer sequence
5' forward Primer (50 µM)
HC 12; LC 15
HC 0.16 µM; LC
0.2 µM
HH_2FL (forward,
leader)
GTAGCAACTGCAACCGGTGTACATTCT 3' reverse Primer (50 µM)
HC 18 (IgG/IgM 1:1);
LC 15
HC 0.23 µM; LC
0.2 µM
HH_2RK (reverse,
IgK)
AACTGCTCACTGGATGGTGG
HotStar DNA Polymerase (5
U/µL)
42
0.055 U/µL
total
3840
*loading buffer: 40% (w/v) sucrose in nuclease free water with cresol red
added to dark red color.
A
B
D
C
F
E
49.5
0.046
0.15
50.3
43.1
0.053
0.11
56.7
32.6
21
3.15
43.3
34.2
0.029
0.017
65.7
426c
gp120
IOMA iGL LC
IOMA iGL HC
0
100
200
300
400
0
200
400
600
800
1000
1200
Time (s)
0
100
200
300
400
Time (s)
426c.TM4 gp120
KD >350 µM
KD ~30 µM
IGT1 gp120
IOMA iGL IgG
0
100
200
300
400
0
50
100
150
200
0
100
200
300
400
0
50
100
150
200
Time (s)
0
100
200
300
400
0
100
200
300
400
500
Time (s)
IOMA iGL + α-IgG-AF647
cMyc-AF488
Pre-sort
Pre-sort
426c.TM4
Library 1
Library 2
collected
collected
426c
gp120
IOMA iGL LC
IOMA iGL HC
460-464
(V5 Loop)
426c
gp120
IOMA iGL LC
IOMA iGL HC
R278
V430P
V430P
S471
D279N
3rd sort
59.1
0.055
0.18
40.7
Pre-sort
collected
3rd sort
5th sort
Unconjugated mi3
IGT2-mi3
IOMA iGL IgG Concentration (µg/mL)
A450
10
1
1000
100
0.1
0.01
0.0
0.5
1.0
1.5
2.0
2.5
426c.TM4 gp120
IGT1 gp120
IGT2 gp120
ELISA: Immunogens
0
10
20
30
40
50
0
100
200
300
400
Time (s)
IOMA iGL IgG
0
50
100
150
200
R278
D279N
Synthesize 426c
gp120 library
FACS selection of
IOMA iGL-binding clones
IGT1 SOSIP
IOMA iGL IgG
IGT2 SOSIP
IOMA iGL IgG
(1)
(2)
Select positions to
mutate on gp120
Figure 1
KD ~0.5 µM
IGT2 gp120
IOMA iGL IgG
38.9
9.17
39.9
12
(7)
(6)
Sort B cells and clone monoclonal antibodies
Analyze serum reponses and antibodies
Yeast display
gp120 library
(3)
Biochemical and biophysical characterization of
gp120 variants (SEC, SDS-PAGE, ELISA, SPR)
gp120 mammalian
expression
(4)
Generate SOSIP versions of immunogens;
multimerize on mi3
IGT2-SpyTag
SpyCatcher003-mi3
IGT2-mi3
pH 7.4 @ RT,
~16 hrs
Test immunization regimens in animal models
(5)
Neutralization Assay
ELISA
pAb
Immunize
Immunize
Sort
Express Ab
Clone
Sort
460-464
(V5 Loop)
50 nm
50 nm
Time (s)
426c degly3 SOSIP
IOMA iGL IgG
Response (RU)
Response (RU)
Response (RU)
Response (RU)
Response (RU)
Response (RU)
IOMA iGL
Mice
Week
0
5
3
8
Prime
IGT2-mi3
Boost 1
IGT1-mi3
10
13
Boost 2
426c-mi3
15
18
Boost 3
mosaic8-mi3
20
23
Boost 4
mosaic8-mi3
BG505
0
23
AMC011
0
23
B41
0
23
CH119
0
23
CE0217
(Autologous)
0
23
Wk
Wk
CNE8
CNE8
N276A
CNE20
CNE20
N276A
0
23
0
23
0
23
0
23
Env ELISA
Env ELISA
*
*
*
*
*
*
IGT1
KO
Wk 0
IGT1
KO
Wk 3
IGT1
KO
Wk 8
IGT1
KO
Wk 13
IGT1
KO
Wk 18
IGT1
KO
Wk 23
0
1
2
3
4
5
15
30
45
Serum Titers (AUC x 103)
Serum Titers (AUC x 103)
IGT2
KO
Wk 0
IGT2
KO
Wk 3
IGT2
KO
Wk 8
IGT2
KO
Wk 13
IGT2
KO
Wk 18
IGT2
KO
Wk 23
0
1
2
3
4
5
10
15
Figure 2
*
**
*
**
* *
** *
*
* *
* *
*
**
*
**
0
1
2
5
10
15
20
25
D279N
WT
KO
Wk 0
D279N
WT
KO
Wk 3
D279N
WT
KO
Wk 8
D279N
WT
KO
Wk 13
D279N
WT
KO
Wk 18
D279N
WT
KO
Wk 23
0
1
2
3
4
5
15
30
45
Serum Titers (AUC x 103)
Serum Titers (AUC x 103)
0
1
2
3
4
5
10
15
Serum Titers (AUC x 103)
A
E
D
G
F
B
C
ES30
HP3
ES30
HP2
HP1
HP7
ES30
HP3
(n=13)
(n=13)
(n=13)
(n=13)
(n=7)
IGT2 ELISA
IGT1 ELISA
426c ELISA
*
B
D
E
G
F
CDRH2
R54HC
F53HC
N276gp120
glycan
G31LC
CDRL1
T19HC
N197gp120
glycan
–––––––––––FR1–––––––– –CDRL1– –––––FR2–––– –––––CDRL2––––– –––––––––––FR3––––––––––– ––CDRL3–– ––FR4––
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 105
• • • • • • • • • • • • • • • • • • • • •
IOMA_iGL QSALTQPASVSGSPGQSITISCTGTSSDVGSYNLVSWYQQHPGKAPKLMIYEVSKRPSGVSNRFSGSKSGNTASLTISGLQAEDEADYYCCSYAGSVAFGGGTKLTVL
IOMA ......................A.S.R...GFD...............I....N.....I.S...A...............E....H...Y...DG............
IO-003 ..........................R.I..F...................D.................D.M...I...F......E.F...F...L...........
IO-008 ................................D.............................................................D.LV......V.G.
IO-010 ............................I.G....................D.N.........................F..............DTLV..........
IO-017 ............................I.G.D..................D.N.........................F..............DTLV..........
IO-018 ............................I...D..................D...........................F..............DTLV..........
IO-040 ............................................V..........................S.........T..V........GDNLV.....R....
IO-044 .................S........N.I...D..................D.T................K........F.T........W...D...........G.
IO-049 ........................P...I.T......................NR.............C.........DF..........S..E.TL...........
IO-050 ......................................................R........................F..............DTLI..........
–––––––––––––FR1––––––––– –CDRH1–– ––––FR2–––– –––––CDRH2–––– –––––––––––––FR3––––––––––––– –––––––CDRH3––––––– –––FR4–––
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 105 110
• • • • • • • • • • • • • • • • • • • • • •
IOMA_iGL QVQLVQSGAEVKKPGASVKVSCKASGYTFTGYYMHWVRQAPGQGLEWMGWINPNSGGTNYAQKFQGRVTMTRDTSISTAYMELSRLRSDDTAVYYCAREMFDSSADWSPWRGMVAWGQGTLVTVSS
IOMA E....E...Q........T...T....K....H.........R..........FR.AVK.P.N.R...S......MEIF.......T.......................................
IO-003 ..................T...........D.FI...................FR.AVD.....R...........T.....................A.....DE....H.L.............
IO-008 ..................T.............EL...................YR.AVK......................V.NG.............D.....DE..............P.....
IO-010 ..................T.............HL...................FR.AIG.....R...........I......RG.........F........DDE....................
IO-017 ..................T.............HL...................FR.AIG.....R...........N......RG.........F........DD.....................
IO-018 ..................T..........ID..I...................YR..PG.....R....L......N......N..K.......F....L..RDD.....................
IO-040 ..................T...........D.FI.....V.............RF.V.DS....R...........T.....................T....D......C.L.............
IO-044 ..........M.......T...A......ADHFI...................RF.N.DS....R...........T...................T..VS...........L.............
IO-049 ..................T...R.........D................R...FR..IE........................................I.....E....................
IO-050 ..................R...........D.EI...................FR.AVK.R............P..T...............R..........D.E....................
HC
LC
IOMA_iGL QVQLVQSGAEVKKPGASVKVSCKASGYTFTGYYMHWVRQAPGQGLEWMGWINPNSGGTNYAQKFQGRVTMTRDTSISTAYMELSRLRSDDTAVYYCAREMFDSSADWSPWRGMVAWG
IOMA E....E...Q........T...T....K....H.........R..........FR.AVK.P.N.R...S......MEIF.......T..............................
IO-003 ..................T...........D.FI...................FR.AVD.....R...........T.....................A.....DE....H.L....
IO-008 ..................T.............EL...................YR.AVK......................V.NG.............D.....DE...........
IO-010 ..................T.............HL...................FR.AIG.....R...........I......RG.........F........DDE...........
IO-017 ..................T.............HL...................FR.AIG.....R...........N......RG.........F........DD............
IO-018 ..................T..........ID..I...................YR..PG.....R....L......N......N..K.......F....L..RDD............
IO-040 ..................T...........D.FI.....V.............RF.V.DS....R...........T.....................T....D......C.L....
IO-044 ..........M.......T...A......ADHFI...................RF.N.DS....R...........T...................T..VS...........L....
IO-049 ..................T...R.........D................R...FR..IE........................................I.....E...........
IO-050 ..................R...........D.EI...................FR.AVK.R............P..T...............R..........D.E...........
IOMA_iGL QVQLVQSGAEVKKPGASVKVSCKASGYTFTGYYMHWVRQAPGQGLEWMGWINPNSGGTNYAQKFQGRVTMTRDTSISTAYMELSRLRSDDTAVYYCAREMFDSSADWSPWRGMVAWG
IOMA E....E...Q........T...T....K....H.........R..........FR.AVK.P.N.R...S......MEIF.......T..............................
IO-003 ..................T...........D.FI...................FR.AVD.....R...........T.....................A.....DE....H.L....
IO-008 ..................T.............EL...................YR.AVK......................V.NG.............D.....DE...........
IO-010 ..................T.............HL...................FR.AIG.....R...........I......RG.........F........DDE...........
IO-017 ..................T.............HL...................FR.AIG.....R...........N......RG.........F........DD............
IO-018 ..................T..........ID..I...................YR..PG.....R....L......N......N..K.......F....L..RDD............
IO-040 ..................T...........D.FI.....V.............RF.V.DS....R...........T.....................T....D......C.L....
IO-044 ..........M.......T...A......ADHFI...................RF.N.DS....R...........T...................T..VS...........L....
IO-049 ..................T...R.........D................R...FR..IE........................................I.....E...........
IO-050 ..................R...........D.EI...................FR.AVK.R............P..T...............R..........D.E...........
IOMA_iGL QSALTQPASVSGSPGQSITISCTGTSSDVGSYNLVSWYQQHPGKAPKLMIYEVSKRPSGVSNRFSGSKSGNTASLTISGLQAEDEADYYCCSYAGSVAFGGGTKLTVL
IOMA ......................A.S.R...GFD...............I....N.....I.S...A...............E....H...Y...DG............
IO-003 ..........................R.I..F...................D.................D.M...I...F......E.F...F...L...........
IO-008 ................................D.............................................................D.LV......V.G.
IO-010 ............................I.G....................D.N.........................F..............DTLV..........
IO-017 ............................I.G.D..................D.N.........................F..............DTLV..........
IO-018 ............................I...D..................D...........................F..............DTLV..........
IO-040 ............................................V..........................S.........T..V........GDNLV.....R....
IO-044 .................S........N.I...D..................D.T................K........F.T........W...D...........G.
IO-049 ........................P...I.T......................NR.............C.........DF..........S..E.TL...........
IO-050 ......................................................R........................F..............DTLI..........
IOMA_iGL QVQLVQSGAEVKKPGASVKVSCKASGYTFTGYYMHWVRQAPGQGLEWMGWINPNSGGTNYAQKFQGRVTMTRDTSISTAYMELSRLRSDDTAVYYCAREMFDSSADWSPWRGMVAWG
IOMA E....E...Q........T...T....K....H.........R..........FR.AVK.P.N.R...S......MEIF.......T..............................
IO-003 ..................T...........D.FI...................FR.AVD.....R...........T.....................A.....DE....H.L....
IO-008 ..................T.............EL...................YR.AVK......................V.NG.............D.....DE...........
IO-010 ..................T.............HL...................FR.AIG.....R...........I......RG.........F........DDE...........
IO-017 ..................T.............HL...................FR.AIG.....R...........N......RG.........F........DD............
IO-018 ..................T..........ID..I...................YR..PG.....R....L......N......N..K.......F....L..RDD............
IO-040 ..................T...........D.FI.....V.............RF.V.DS....R...........T.....................T....D......C.L....
IO-044 ..........M.......T...A......ADHFI...................RF.N.DS....R...........T...................T..VS...........L....
IO-049 ..................T...R.........D................R...FR..IE........................................I.....E...........
IO-050 ..................R...........D.EI...................FR.AVK.R............P..T...............R..........D.E...........
IOMA_iGL QSALTQPASVSGSPGQSITISCTGTSSDVGSYNLVSWYQQHPGKAPKLMIYEVSKRPSGVSNRFSGSKSGNTASLTISGLQAEDEADYYCCSYAGSVAFGGGTKLTVL
IOMA ......................A.S.R...GFD...............I....N.....I.S...A...............E....H...Y...DG............
IO-003 ..........................R.I..F...................D.................D.M...I...F......E.F...F...L...........
IO-008 ................................D.............................................................D.LV......V.G.
IO-010 ............................I.G....................D.N.........................F..............DTLV..........
IO-017 ............................I.G.D..................D.N.........................F..............DTLV..........
IO-018 ............................I...D..................D...........................F..............DTLV..........
IO-040 ............................................V..........................S.........T..V........GDNLV.....R....
IO-044 .................S........N.I...D..................D.T................K........F.T........W...D...........G.
IO-049 ........................P...I.T......................NR.............C.........DF..........S..E.TL...........
IO-050 ......................................................R........................F..............DTLI..........
IO-010
CDRH3
A100AHC
S100HC
R476gp120
R480gp120
K97gp120
D100BHC
IOMA
SHMs elicited from IOMA iGL
Occurrence of SHMs
IO-010
IO-040
IOMA
CDRH1
IOMA (Interface)
IOMA (Interface)
Figure 3
100 % occurrence
60
40
10
0
CDRL3
D95LC
N280gp120
N279gp120
R456gp120
CDRL1
N276gp120
glycan
N197gp120
glycan
BG505
gp120
LC
HC
4
5
6
7
8
9
10
11
12
13
14
15
16
17
8
10 12 14 16 18 20 22 24 26 28 30 32 34 36
Total aa mutations
IOMA-like aa mutations
(n = 25)
Neutralization
score
Monoclonal antibodies
Heavy + light chains
C
A
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
R2 = 0.78
n = 759
R2 = 0.52
n = 318,769
0
2
4
6
8
10
12
14
16
18
0
5
10
15
20
25
30
35
40
45
50
# of aa residues different from IOMA iGL
# of IOMA−like aa residues
type
IOMAgl
baseline
type
●
●
IOMAgl
baseline
n
Heavy chains
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
R2 = 0.62
n = 924
R2 = 0.55
n = 1,790,961
0
1
2
3
4
5
6
7
8
9
0
5
10
15
20
25
30
35
40
# of aa residues different from IOMA iGL
# of IOMA−like aa residues
n
2×105
1×105
Light chains
3×104
2×104
1×104
●
●
B
A
WT
Mice
Week
0
5
3
8
10
13
15
18
20
23
Serum Titers (AUC x 103)
IGT1
KO
IGT1
KO
IGT1
KO
Wk 3
Wk 0
Wk 8
0
2
4
15
30
45
**
*
Figure 4
Prime
IGT2-mi3
Boost 1
IGT1-mi3
Boost 2
426c-mi3
Boost 3
mosaic8-mi3
Boost 4
mosaic8-mi3
(n=16)
(n=16)
(n=16)
(n=16)
(n=9)
C
3D3 ELISA
101
102
103
104
105
106
0.00
0.25
0.50
0.75
RLU
Naive (n=16)
Prime (n=16)
Boost 1 (n=16)
Boost 2 (n=16)
Boost 3 (n=16)
Boost 4 (n=9)
3D7 ELISA
101
102
103
104
105
106
0.00
0.25
0.50
0.75
RLU
Naive (n=16)
Prime (n=16)
Boost 1 (n=16)
Boost 2 (n=16)
Boost 3 (n=16)
Boost 4 (n=9)
Serum Titers (AUC x 103)
Serum Titers (AUC x 103)
Env ELISA
E
D
F
Wk
BG505
0
23
AMC011
0
23
B41
0
23
CH119
0
23
CE0217
(Autologous)
0
23
Wk
CNE8
CNE8
N276A
CNE20
CNE20
N276A
Env ELISA
Env ELISA
0
23
0
23
Wk
CNE8 CNE8
N276A
23
23
0
23
0
23
0
1
3
2
4
5
10
15
M21
20
Serum Titers (AUC x 103)
****
****
****
****
**
****
****
**
****
****
**
**
CNE20CNE20
N276A
23
23
**
M29
M28
0
1
3
2
4
5
10
15
20
1
3
2
4
5
10
15
20
0
IGT1 ELISA
Reciprocal Serum Dilution
Reciprocal Serum Dilution
A
B
C
Week
0
5
3
8
10
Week
0
3
6
Naive
Wk 3
Figure 5
ID50
101
102
103
104
105
106
107
IGT1
KO
IGT1
KO
IGT1
KO
ID50
Wk 0
Wk 3
Wk 6
IGT2
101
102
103
104
105
106
107
Wk 0
Wk 3
Wk 6
IGT1
IGT1 ELISA
0
2
4
6
25
50
75
100
Serum Titers (AUC x 103)
Wk 8
IGT1 ELISA
0
2
4
6
20
30
40
50
Serum Titers (AUC x 103)
Naive
Wk 3
IGT1
KO
IGT1
KO
IGT1
KO
Wk 6
*
*
****
***
***
**
*
*
*
Wk 0
Wk 5
Wk 10
IGT2
Wk 0
Wk 5
Wk 10
IGT1
NHPs
Neutralization Assay
Neutralization Assay
Rabbits
Prime
IGT2-mi3
Boost 1
IGT1-mi3
Prime
IGT2-mi3
Boost 1
IGT1-mi3
(n=8)
(n=5)
–––––––––––––FR1––––––––––– ––CDRH1–– ––––FR2–––– –––––CDRH2–––––– ––––––––––––––––FR3––––––––––––––– ––––––––CDRH3–––––––– –––FR4––––
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 105 110
∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙
VH1-2*02 QVQLVQSGAEVKKPGASVKVSCKASGYTF-TGYYMHWVRQAPGQGLEWMGWINPNSGGTNY-AQKFQGRVTMTRDTSI----STAYMELSRLRSDDTAVYYCAR------------------------------
IOMA E....E...Q........T...T....K.....H.........R..........FR.AVK..P.N.R...S......M....EIF.......T...........EMFDS..SADWSPWRGMVAWGQGTLVTVSS
VRC01 ........GQM....E.MRI..R....E..IDCTLN.I.L...KRP.....LK.RG.AV....RPL.........VYS....D..FL..RS.TV......F.T.GKNCD.......YNWDFEHWGRGTPVIVSS
–––––––––––––FR1––––––––– –CDRL1–– –––––FR2–––– –––––CDRL2–––––– –––––––––––FR3––––––––––– –––CDRL3––– –––––FR4–––––––
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 105 110
∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙ ∙
VL2-23*02 QSALTQ-PASVSGSPGQSITISCTGTSSDVGSYNLVSWYQQHPGKAPKLMIYEVSKRPSGVSNRFSGSKSGNTASLTISGLQAEDEADYYCCSYAGS---STF---------------
IOMA ......-................A.S.R...GFD...............I....N.....I.S...A...............E....H...Y...DG---VA.GGGTKLT-VLGQPKA
VK3-20*01 EIVLTQSPGTLSLSPGERATLSCRASQSVSSSYLAWYQQKPGQAPRLLIYGASSRATGIPDRFSGSGSGTDFTLTISRLEPEDFAVYYCQQYG---SS------------------
VRC01 .................T.II...T..Y---GS......R.......V..SG.T..A.........RW.P.YN....N..SG..G.......E------FFGQGTKVQVDIKR---
Supplemental Figure 1
B
A
C
D
IOMA iGL
LC
LC
HC
IOMA iGL
HC
IOMA iGL
LC
IOMA iGL
HC
2.07Å IOMA iGL crystal structure
90o
E
RMSD – 0.64Å
IOMA iGL
LC
IOMA iGL
HC
IOMA iGL
LC
IOMA iGL
HC
IOMA
LC
IOMA HC
IOMA HC
IOMA LC
Overlay of IOMA iGL & IOMA (PDB 5T3Z)
180o
0
100
200
300
400
0
25
50
75
100
125
150
Time (s)
Response (RU)
IGT2 SOSIP
426c.TM4 gp120
426c degly3 SOSIP
426c degly3 SOSIP-mi3
BG505.v4.1-GT1 SOSIP
eOD-GT8
IOMA iGL IgG
1 µM Immunogen
Library 1
Pre-Sort
426c
59.1
0.055
0.18
40.7
collected
49.5
0.046
0.15
50.3
collected
1st Sort
3rd Sort
32.6
21
3.15
43.3
34.2
0.029
0.017
65.7
4th Sort
5th Sort
35.9
0.26
9.12
54.8
31.6
0.019
0.09
68.3
23.6
2.44
0.52
73.5
39
0.1
0.059
60.9
2nd Sort
82.3
0.026
0.68
16.8
74.8
0.046
0.069
25.1
Library 2
43.1
0.053
0.11
56.7
collected
53.8
8.97
1.27
35.9
58
1.03
0.2
40.7
61.7
2.58
13.5
22.2
49.8
6.22
26.3
17.6
38.9
9.17
39.9
12
IOMA iGL � α-IgG AF647
α-cMyc-AF488
I
G
H
A280
Volume (mL)
IOMA iGL
426c SOSIP
MW
(kDa)
Immunogen
IOMA iGL IgG
Immunogen
3BNC60 iGL IgG
Immunogen
BG24 iGL IgG
3BNC60 iGL
BG24 iGL
IGT1 SOSIP
IGT2 SOSIP
0
100
200
300
400
Time (s)
SOSIP
SOSIP–mi3
SOSIP
SOSIP–mi3
SOSIP
SOSIP–mi3
Normalized Response (RU)
0
100
200
300
400
Time (s)
Normalized Response (RU)
0
100
200
300
400
Time (s)
Normalized Response (RU)
J
mi3
IGT1
mi3–IGT2
IGT2
mi3–IGT1
IGT1
mi3–IGT2
mi3
mi3–IGT1
IGT2
Unconjugated
mi3
Conjugated
mi3–gp41
gp120
gp140
Non-reduced
Reduced
250
150
100
75
50
37
25
20
426c.TM4 gp120
IGT1 gp120
IGT2 gp120
4
6
8
10
12
14
16
18
20
A450
10
1
1000
100
0.1
0.01
0.0
0.5
1.0
1.5
2.0
2.5
A450
10
1
1000
100
0.1
0.01
0.0
0.5
1.0
1.5
2.0
2.5
426c.TM4 gp120
IGT1 gp120
IGT2 gp120
Immunogen
VRC01 iGL IgG
VRC01 iGL
SOSIP
SOSIP–mi3
0
100
200
300
400
Time (s)
Normalized Response (RU)
A450
10
1
1000
100
0.1
0.01
0.0
0.5
1.0
1.5
2.0
2.5
IgG Concentration (µg/mL)
IgG Concentration (µg/mL)
IgG Concentration (µg/mL)
IgG Concentration (µg/mL)
A450
10
1
1000
100
0.1
0.01
0.0
0.5
1.0
1.5
2.0
2.5
MW
(kDa)
IGT2
IGT1
426c degly2
426c.TM4
IGT2
IGT1
gp120
gp140
SOSIP
gp120
150
100
75
50
37
25
20
15
F
C
D
F
G
E
0.35
66.9
32.7
CD2-PE
4.72
86.5
8.63
95.0
3.99
0.38
25.0
72.4
Pro
Mature
Immature
Pre
Mature
IgL+
Pro
Mature
Immature
Pre
Mature
IgL+
Mature
Immature
Pre
Mature
IgL+
2.52
Mature
Immature
Pre
Mature
IgL+
IgM-FITC
34.3
62.8
Mature
Immature
Pre
2.35
IgM-FITC
IgD-BV786
92.7
0.62
6.45
IgK-BV421
IgL-APC
C57BL/6J
IOMAgl
Live, singlet
Dump- B220+ CD19+
lymphocytes
Live, singlet
Dump- B220+ CD19+
lymphocytes
Bone Marrow
96.0
3.95
81.5
13.2
48.0
34.1
9.48
83.1
11.9
94.9
89.6
10.4
CD93-APC
B220-BV605
80.0
10.9
IgM-FITC
CD21/35-BV421
56.4
42.4
0.71
CD23-PE
IgM-FITC
72.7
19.6
CD23-PE
B220-BV605
1.46
IgM-FITC
IgD-BV786
C57BL/6J
IOMAgl
Live, singlet
Dump- B220+ CD19+
lymphocytes
Spleen
Bone marrow
Mature
0
2500
5000
7500
10000
12500
gMFI IgD
Bone marrow
C57BL/6J
IOMAgl
MZ
FOB
0
2500
5000
7500
10000
12500
gMFI IgD
Spleen
C57BL/6J
IOMAgl
Mature
IgL+
T3
T2
T1
FOB
MZP
MZ
B220+
CD19+
T1
T2
T3
MZ
MZP
FOB
104
105
106
107
108
absolute cell number / spleen
Spleen
B220+
CD19+
Pro
Pre
Imma-
ture
Ma-
ture
Mature
IgL+
103
104
105
106
107
absolute cell number / leg
H
Supplemental Figure 2
A
B
LVDJ
loxP
Promoter
J intron
JH4
JH3
JH2
LVDJ
of IOMA iGL
loxP
loxP
Promoter
tAce-Cre/NeoR
homology arm
homology arm
homology arm
homology arm
homology arm
homology arm
homology arm
homology arm
DH4-1 JH1
DTA
J intron
self-excising Cre
IghIOMAiGL
targeting vector
wildtype Igh
LVJ
loxP
Promoter
Ck
LVJ
of IOMA iGL
loxP
loxP
Promoter
tAce-Cre/NeoR
Ck
Jk5
Jk4
Jk3
Jk2
Jk1
DTA
self-excising Cre
IgkIOMAiGL
targeting vector
wildtype Igk
Heavy chain
IghIOMAiGL
Light chain
IgkIOMAiGL
homology arm
homology arm
homology arm
homology arm
CNE8A
Supplemental Figure 3
ET34
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HQ4
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP7
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP3
HP4
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-7
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
HP1
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
HP2
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
ET34
HQ4
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP7
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
HP3
HP4
10-4
100
10-1
10-6
20
0
-20
-60
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-40
-20
-60
40
60
80
100
10-2
10-3
10-5
HP1
HP2
-60
20
0
-20
-40
40
60
80
100
ES30
CNE8
N276A
B
10-4
100
10-1
10-2
10-3
10-6
10-5
HQ4
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP7
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ES30
CNE20
C
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ET34
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP1
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
ET34
HQ4
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP7
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
HP3
HP4
10-4
10-1
10-6
20
0
-20
-40
-60
-80
-100
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
HP1
HP2
-80
20
0
-20
-40
-60
40
60
80
100
ES30
CNE20
N276A
D
10-4
100
10-1
10-2
10-3
10-6
10-5
100
ES30
PVO.4
E
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ET34
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
HP4
10-4
100
10-1
10-6
-60
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-80
20
0
-20
-60
-40
40
60
80
100
10-2
10-3
10-6
10-5
ET34
ES30
Q23.17
F
YU2H
JRCSF
I
HP3
ET33
WITO
4160.
33
G
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ET34
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
HP4
10-4
100
10-1
10-6
-80
20
0
-20
-40
-60
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP3
HP2
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP1
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
6535.5
J
3415K
MuLVM
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
ES30
HP2
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP3
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP4
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HP7
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
HQ4
10-4
100
10-1
10-6
-60
20
0
-40
-20
40
60
80
100
10-2
10-3
10-5
HP1
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
HP2
10-4
100
10-1
10-6
-60
20
0
-40
-20
40
60
80
100
10-2
10-3
10-5
HP7
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
HQ4
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
HP4
HP7
10-4
100
10-1
10-6
20
0
-20
-40
-60
-80
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
HP1
HP2
10-4
100
10-1
10-6
-100
20
0
-20
-80
-60
-40
40
60
80
100
10-2
10-3
10-5
HP3
% Neutralization
10-4
100
10-1
10-6
-40
20
0
40
60
80
100
Reciprocal Serum Dilution
10-2
10-3
10-5
-20
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
HQ4
ES37
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ES34
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ES32
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ES30
HP7
HQ4
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
10-2
10-3
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
HP3
HP4
-40
20
0
-20
40
60
80
100
10-4
100
10-1
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
HP2
HP1
ET34
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ET33
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ES30
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ET34
HP1
HP3
HP1
HP3
10-4
100
10-1
10-6
-40
-20
20
0
40
60
80
100
10-2
10-3
10-5
ET34
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
ES30
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
-60
20
0
-20
-40
40
60
80
100
10-4
100
10-1
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
ET33
ET34
10-4
100
10-1
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
HP1
HP3
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
-40
20
0
-20
40
60
80
100
10-4
100
10-1
10-2
10-3
10-6
10-5
10-4
100
10-1
-60
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
10-6
20
0
-20
-60
-40
40
60
80
100
10-2
10-3
10-5
ET33
ET34
HP1
HP3
HQ4
Terminal serum (week 18 or 23)
Naive serum (week 0)
CAAN
L
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-20
-60
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
10-2
10-3
10-5
ET33
ET34
HP1
HP3
M1
CNE8N
10-4
100
10-1
10-7
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
M15
10-4
100
10-1
10-7
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
M29
10-4
100
10-1
10-7
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
M28
10-4
100
10-1
10-7
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
M24
M26
10-4
100
10-1
10-7
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
10-7
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
M21
10-4
100
10-1
10-7
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
M22
10-4
100
10-1
10-7
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
M15
M29
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M28
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
M24
M26
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
M21
M22
-60
-80
-100
20
0
-20
-40
40
60
80
100
M1
CNE8
N276A
O
10-4
100
10-1
10-2
10-3
10-6
10-5
M14
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M13
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
M29
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M28
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
M22
M23
-40
20
0
-20
40
60
80
100
10-4
100
10-1
10-2
10-3
10-6
10-5
M1
CNE20
P
CNE20
N276A
Q
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M15
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M15
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M24
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
M15
M29
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M28
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
M13
M21
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-20
-60
-40
40
60
80
100
10-2
10-3
10-5
M24
M5
-40
20
0
-20
40
60
80
100
M1
PVO.4
R
10-4
100
10-1
10-2
10-3
10-6
10-5
M1
10-4
100
10-1
10-6
10-2
10-3
10-5
M1
Q842.
D12
S
WITO
4160.
33
T
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M24
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
M24
10-4
100
10-1
10-6
-80
-60
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-60
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
-80
20
0
-20
-40
-60
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-80
20
0
-20
-40
-60
40
60
80
100
10-2
10-3
10-6
10-5
M24
M15
M14
10-4
100
10-1
10-6
20
0
-20
-40
-60
-80
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
M15
Q23.17
U
6535.5
V
M1
10-4
100
10-1
10-6
10-2
10-3
10-5
M28
BG505
W
MuLVX
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M14
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
M5
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
-40
20
0
-20
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
10-4
100
10-1
-40
20
0
-20
40
60
80
100
10-2
10-3
10-6
10-5
M21
M24
M26
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
10-4
100
10-1
10-6
20
0
-20
-40
40
60
80
100
10-2
10-3
10-5
M15
M1
10-4
100
10-1
10-6
10-2
10-3
10-5
A
IOMAgl
C57BL/6J
A
B
1
2
3
4
Group
5
6
7
8
0
2
4
6
8
10
Serum Titers (AUC x 103)
426c ELISA
****
****
****
****
****
****
****
1
2
3
4
Group
5
6
7
8
0
5
10
15
Serum Titers (AUC x 103)
426c degly2 ELISA
*
***
***
***
**
***
***
Group 1
Group 2
Group 3
Group 4
Group 5
Group 6
Group 7
Group 8
Mice
Week
0
5
3
8
Prime
IGT2-mi3
Boost 1
IGT1-mi3
10
13
Boost 2
426c-mi3
15
18
Boost 3
mosaic8-mi3
Mice
Week
0
5
3
8
Prime
IGT2-mi3
Boost 1
IGT2-mi3
10
13
Boost 2
IGT2-mi3
15
18
Boost 3
IGT2-mi3
Mice
Week
0
5
3
8
Prime
IGT1-mi3
Boost 1
IGT1-mi3
10
13
Boost 2
426c-mi3
15
18
Boost 3
mosaic8-mi3
Mice
Week
0
5
3
8
Prime
IGT2-mi3
Boost 1
426c-mi3
10
13
Boost 2
mosaic8-mi3
15
18
Boost 3
mosaic8-mi3
Mice
Week
0
5
3
8
Prime
IGT2-mi3
Boost 1
IGT1-mi3
10
13
Boost 2
426c-mi3
Mice
Week
0
5
3
8
Prime
IGT2-mi3
Boost 1
IGT1-mi3
Mice
Week
0
5
3
8
Prime
426c-mi3
Boost 1
mosaic8-mi3
Mice
Week
0
5
3
8
Prime
mosaic8-mi3
Boost 1
mosaic8-mi3
Supplemental Figure 4
IOMAgl
IOMAgl
IOMAgl
IOMAgl
IOMAgl
IOMAgl
IOMAgl
IOMAgl
(n=13)
(n=8)
(n=8)
(n=8)
(n=8)
(n=8)
(n=8)
(n=8)
A
B
82.1
FSC-A
SSC-A
98.4
FSC-A
FSC-H
99.1
SSC-A
SSC-W
25.9
B220-BV421
Dump-APCeF780 +
Live/Dead-NIR
99.9
BaitKO-PECy7
Bait-PE
0.21
Bait-AF647
Bait-PE
2.00
CD38-AF700
CD95-FITC
Sort GC B cells
Sort
Bait++ BaitKO-
HP3 Spleen
+ mLN
All events
0.26
1.68
8.35E-3
0.29
80.7
426c degly2 D279N-AF647
426c degly2 D279N-PE
B cells
B cells
naive
IOMAgl
mouse
Group 1 HP3
day 175:
IGT2>IGT1>426c>
mosaic8>mosaic8
IOMA-expressing
RAMOS cells
Supplemental Figure 5
CD95-FITC
CD38-AF700
C�7 exon
C�3 exon
C�2 exon
C�1 exon
IgL loci
C� exon
IgK locus
C�
E�
enhancer
CRISPR
JH
VH
VH
JH
Endogenous
rearranged
variable
(VDJ)
splice
acceptor
targeting vector
T2A
variable
variable constant
IOMA
heavy chain
IOMA
light chain
homology
homology
P2A
IgH locus
splice
donor
C
Sorted
Identity [%]
Mouse
Population Antibody ID
HP1
GC
IO-001
HP1
GC
IO-002
HP1
426c+/KO-
IO-003
HP1
426c+/KO-
IO-004
HP1
GC
IO-005
HP1
GC
IO-006
HP1
GC
IO-007
HP1
GC
IO-008
HP1
GC
IO-009
HP3
426c+/KO-
IO-010
HP3
426c+/KO-
IO-011
HP1
GC
IO-012
HP1
GC
IO-013
HP1
GC
IO-014
HP1
426c+/KO-
IO-015
HP1
426c+/KO-
IO-016
HP3
426c+/KO-
IO-017
HP3
426c+/KO-
IO-018
HP3
426c+/KO-
IO-019
ES30
426c+/KO-
IO-020
ES30
426c+/KO-
IO-021
ES30
426c+/KO-
IO-022
ES30
426c+/KO-
IO-023
ES30
426c+/KO-
IO-024
ES30
426c+/KO-
IO-025
ES30
426c+/KO-
IO-026
ES30
426c+/KO-
IO-027
ES30
426c+/KO-
IO-028
ES30
426c+/KO-
IO-029
ES30
426c+/KO-
IO-030
ES30
426c+/KO-
IO-031
ES30
CNE8 N276A+
IO-032
ES30
426c+/KO-
IO-033
ES30
426c+/KO-
IO-034
ES30
426c+/KO-
IO-035
ES30
426c+/KO-
IO-036
ES30
426c+/KO-
IO-037
HP3
426c+/KO-
IO-038
HP3
426c+/KO-
IO-039
HP3
426c+/KO-
IO-040
HP3
426c+/KO-
IO-041
HP3
426c+/KO-
IO-042
HP3
426c+/KO-
IO-043
HP3
426c+/KO-
IO-044
HP3
426c+/KO-
IO-045
HP3
426c+/KO-
IO-046
HP3
426c+/KO-
IO-047
ES30
426c+/KO-
IO-048
ES30
426c+/KO-
IO-049
ES30
426c+/KO-
IO-050
ES30
426c+/KO-
IO-051
ES30
426c+/KO-
IO-052
ES30
426c+/KO-
IO-053
ES30
426c+/KO-
IO-054
ES30
426c+/KO-
IO-055
HP3
10x GCB
IO-056
HP3
10x GCB
IO-057
HP3
10x GCB
IO-058
HP3
10x GCB
IO-059
HP3
10x GCB
IO-060
HP3
10x GCB
IO-061
HP3
10x GCB
IO-062
HP3
10x GCB
IO-063
HP3
10x GCB
IO-064
HP3
10x GCB
IO-065
HP3
10x GCB
IO-066
HP3
10x GCB
IO-067
IOMA
IOMA iGL
Domains
Structurally important residues
Color code
52
A
1
10
20
30
40
52
50
60
70
70
82
80
90
<30 %
<100 - 30 %
Identity
100 %
110
113
100
82
A B C
100
A B C D E F G H I
Kabat #
*
*
*
**
*
*
*
*
Supplemental Figure 6
Sorted
Identity [%]
Mouse
Population Antibody ID
HP1
GC
IO-001
HP1
GC
IO-002
HP1
426c+/KO-
IO-003
HP1
426c+/KO-
IO-004
HP1
GC
IO-005
HP1
GC
IO-006
HP1
GC
IO-007
HP1
GC
IO-008
HP1
GC
IO-009
HP3
426c+/KO-
IO-010
HP3
426c+/KO-
IO-011
HP1
GC
IO-012
HP1
GC
IO-013
HP1
GC
IO-014
HP1
426c+/KO-
IO-015
HP1
426c+/KO-
IO-016
HP3
426c+/KO-
IO-017
HP3
426c+/KO-
IO-018
HP3
426c+/KO-
IO-019
ES30
426c+/KO-
IO-020
ES30
426c+/KO-
IO-021
ES30
426c+/KO-
IO-022
ES30
426c+/KO-
IO-023
ES30
426c+/KO-
IO-024
ES30
426c+/KO-
IO-025
ES30
426c+/KO-
IO-026
ES30
426c+/KO-
IO-027
ES30
426c+/KO-
IO-028
ES30
426c+/KO-
IO-029
ES30
426c+/KO-
IO-030
ES30
426c+/KO-
IO-031
ES30
CNE8 N276A+
IO-032
ES30
426c+/KO-
IO-033
ES30
426c+/KO-
IO-034
ES30
426c+/KO-
IO-035
ES30
426c+/KO-
IO-036
ES30
426c+/KO-
IO-037
HP3
426c+/KO-
IO-038
HP3
426c+/KO-
IO-039
HP3
426c+/KO-
IO-040
HP3
426c+/KO-
IO-041
HP3
426c+/KO-
IO-042
HP3
426c+/KO-
IO-043
HP3
426c+/KO-
IO-044
HP3
426c+/KO-
IO-045
HP3
426c+/KO-
IO-046
HP3
426c+/KO-
IO-047
ES30
426c+/KO-
IO-048
ES30
426c+/KO-
IO-049
ES30
426c+/KO-
IO-050
ES30
426c+/KO-
IO-051
ES30
426c+/KO-
IO-052
ES30
426c+/KO-
IO-053
ES30
426c+/KO-
IO-054
ES30
426c+/KO-
IO-055
HP3
10x GCB
IO-056
HP3
10x GCB
IO-057
HP3
10x GCB
IO-058
HP3
10x GCB
IO-059
HP3
10x GCB
IO-060
HP3
10x GCB
IO-061
HP3
10x GCB
IO-062
HP3
10x GCB
IO-063
HP3
10x GCB
IO-064
HP3
10x GCB
IO-065
HP3
10x GCB
IO-066
HP3
10x GCB
IO-067
IOMA
IOMA iGL
Domains
Structurally important residues
27
27
A B C
96
94
107
CDRL2
Kabat #
1
9
100
90
80
70
60
50
40
30
20
11
*
***
*
A
B
A
B
C
E
D
Paired
HC+LC
n = 5207
IOMAgl (HC+LC) = 97%
2
3
4
5
6
7
8
9
10
11
13
14
12
Clones
IGHV rearrangment
with IOMAgl LC
IGHV rearrangment + endogenous IgL
HP3 IOMAgl GCBs
from spleen and mLN
singles
n = 5207
IgA
IgM
IgG1
IgG2b
IgG2c
IgG3
0
5
10
15
20
25
0
5
10
15
20
25
25
50
75
100
125
125
1000
2000
3000
4000
# of aa mutations to IOMA iGL HC
# of IOMAgl clone GC B cells with paired HC & LC
Heavy chain
0
5
10
15
0
5
10
15
20
20
40
60
80
100
125
120
1000
2000
3000
4000
5000
# of aa mutations to IOMA iGL LC
# of IOMAgl clone GC B cells with paired HC & LC
Light chain
Isotype
0
5
10
15
20
25
30
35
0
5
10
15
20
25
25
50
75
100
125
1000
2000
3000
4000
# of IOMAgl clone GC B cells with paired HC & LC
Heavy + light chain
Supplemental Figure 7
# of aa mutations to IOMA iGL HC + LC
IOMA
IOMA
IOMA
IOMA iGL
IOMA
10-1074
IO-003
IO-004
IO-005
IO-007
IO-008
IO-009
IO-010
IO-011
IO-012
IO-014
IO-017
IO-018
IO-037
IO-039
IO-040
IO-042
IO-044
IO-050
IO-051
IO-052
IO-055
Supplemental Figure 8
IgG Titers (AUC x 103)
0
100
200
300
400
BG505
CNE20
CNE20
N276A
Env ELISA
CNE20
CNE20
N276A
Env ELISA
A
B
CE0217
(Autologous)
IgG Titers (AUC x 103)
0
100
200
300
400
**
| 2022 | CD4-binding site immunogens elicit heterologous anti-HIV-1 neutralizing antibodies in transgenic and wildtype animals | 10.1101/2022.09.08.507086 | [
"Gristick Harry B.",
"Hartweger Harald",
"Loewe Maximilian",
"van Schooten Jelle",
"Ramos Victor",
"Oliviera Thiago Y.",
"Nishimura Yoshiaki",
"Koranda Nicholas S.",
"Wall Abigail",
"Yao Kai-Hui",
"Poston Daniel",
"Gazumyan Anna",
"Wiatr Marie",
"Horning Marcel",
"Keeffe Jennifer R.",
... | null |
1
1
Long title: The positive role of noise for information acquisition in biological
2
signaling pathways
3
Short title: Noise can increase information acquisition in signaling pathways
4
Eugenio Azpeitia1,2,3, Andreas Wagner1,2,4
5
1. Department of Evolutionary Biology and Environmental Studies, University of Zürich,
6
Zürich, Switzerland.
7
2. Swiss Institute of Bioinformatics, Lausanne, Switzerland.
8
3. Centro de Ciencias Matemáticas, UNAM, Morelia, México.
9
4. The Santa Fe Institute, Santa Fe, NM, USA.
10
11
Abstract
12
All living systems acquire information about their environment. At the cellular level,
13
they do so through signaling pathways, which rely on interactions between molecules
14
that detect and transmit the presence of an extracellular cue or signal to the cell’s
15
interior. Such interactions are inherently stochastic and thus noisy. In classical
16
information theory, a noisy communication channel degrades the amount of
17
transmissible information relative to a noise-free channel. For this reason, one would
18
expect that the kinetic parameters that determine a pathway’s operation minimize
19
noise. We show that this is not the case under a wide range of biologically sensible
20
parameter values. Specifically, we perform computational simulations of simple
21
signaling systems, which show that a noisy molecular interaction dynamics is a
22
necessary condition for information acquisition. Moreover, we show that optimal
23
information acquisition, where a system reacts most sensitively to changes in the
24
environment, can be obtained close to the maximal attainable level of noise in the
2
25
system. Our work highlights the positive role that noise can have in biological
26
information processing.
27
28
Author summary
29
The acquisition of information is fundamental for living systems, because the decisions
30
they take based on such information directly affect survival and reproduction. The
31
molecular mechanisms used by cells to acquire information are signaling pathways. The
32
molecular interactions of signaling pathways, such as the binding of a signal to a
33
receptor, are by nature noisy. This is important, because noise disrupts information.
34
Hence, to maximize the acquisition of information, signaling pathways should minimize
35
the noise of their molecular interactions. Here we show that a noisy dynamic of the
36
molecular interactions can improve the acquisition of information, and that the maximal
37
capacity to acquire information can be obtained with a close-to-maximal level of noise in
38
a signaling pathway. Thus, contrary to expectations, noise can improve the acquisition of
39
information in living systems.
40
41
Introduction
42
Information about the environment is fundamental when living organisms make
43
decisions that affect their survival and reproduction (1). For example, microbes adjust
44
their growth in response to nutrient concentrations, animals flee in response to
45
predators, and plants synthesize defense chemicals in response to herbivores.
46
47
At the cellular level, signaling pathways are the main molecular mechanism by which
48
organisms acquire information. They typically detect the presence of a molecular signal
49
or cue (2) about the environment through the binding of this molecule to a receptor.
3
50
Once the signal has been detected, a chain of intermediary events transmits this
51
information to the cell’s interior, where it ultimately regulates gene expression.
52
Signaling pathways vary widely, including in their number of molecular interactions,
53
signal and receptor affinities, the presence of feedback and feed-forward interactions,
54
and the number of regulated genes (3,4). However, they all share some elementary
55
processes, such as the reversible binding of molecules, which is necessary to detect a
56
signal by a receptor, transmit its presence via effector molecules, for example through
57
allosteric control of these molecules, and regulate gene expression through the binding
58
of transcription factors to DNA.
59
60
Noise is present at all spatial and temporal scales of biological organization, from
61
population dynamics to molecular interactions, including signaling pathways (5,6). The
62
sources of noise in signaling pathways include random fluctuations in the concentration,
63
movement, activity, and interactions of molecules (7–12). Classical information theory
64
predicts that noise degrades the capability of a communication channel to transmit
65
information. For example, in simple systems such as a binary or a symmetric
66
information transmission channel, the maximum capacity of the channel to transmit
67
information can only be realized in the absence of noise (13). In signaling pathways,
68
noise transforms a stimulus, such as the concentration of a nutrient, into a distribution
69
of outputs or responses. Any overlap between the response distributions produced by
70
two different stimuli, such as two different signal concentrations, creates uncertainty
71
about which stimuli produced which output (Fig 1a; 14). For this reason, noise also
72
decreases the information acquisition capacity of signaling pathways.
73
4
74
Fig 1. Relationship of acquired information with both noise and output range. Example of two hypothetical stimuli
75
that produce different but overlapping response distributions (green and blue distributions). The amount of
76
information acquired at different levels of noise (y axis) and with different output ranges (x axis) is indicated by the
77
color bar. The black dashed line in (a) is a schematic representation of noise (i.e., the standard deviation of the
78
response distributions, green and blue). The square bracket above the response distributions in (a) indicates the
79
output range (i.e., the maximal difference of the mean values of the response distributions). Increasing the output
80
range (b) and reducing the noise level (c) decrease the overlap (uncertainty) between response distributions
81
observed in (a). (d) Acquired information is largest (maximal) when noise is minimized and the output range is
82
maximized.
83
84
The overlap between response distributions can be reduced in two ways (11). First, the
85
response distributions can be made more distinct by separating their means while
86
preserving their dispersion (e.g., their standard deviation). This implies that the range of
87
outputs produced by the stimuli will increase (Fig 1b). The second way to decrease the
88
overlap between the response distributions is to decrease their dispersion, while
89
keeping the mean constant. This is equivalent to reducing noise, which allows detecting
90
the signal with increasing precision (Fig 1c). Thus, information acquisition is maximized
91
when the output range is maximized and noise is minimized (Fig 1d).
92
93
Various mechanisms can either reduce noise or increase the output range to improve
94
information acquisition (10,11,15–17). These include feedback loops and protein-
95
protein interactions that reduce the level of noise (12,18,19), and increasing the number
96
of molecules, which increase the output range (20,21). As a result, we know that noise –
97
and thus also information acquisition – can be tuned within some limits (20,22–26).
98
However, it is less clear how signaling pathways adjust their kinetic parameters, such as
99
the association and dissociation rate of binding molecules, to minimize noise or increase
100
the output range to respond efficiently to an environmental signal.
5
101
102
While a few studies have explored the effect of kinetic parameters on information
103
transmission in signaling pathways, small gene networks, and gene expression systems
104
(10,11,16,26), most of these studies did not explicitly model the molecular interactions
105
involved in signaling. Therefore, they provide little intuition about why and how the
106
kinetic properties of molecular processes affect information acquisition. To overcome
107
this limitation, we use models that explicitly include all relevant molecular interactions
108
and that do not make any a priori assumptions about statistical properties of noise. With
109
these models, we analyze how the kinetic properties of the reversible binding
110
interactions used by signaling pathways affect the relationship between noise, output
111
range and information acquisition. First, we study the relationship between noise,
112
output range and information in the reversible binding of two molecules that represent
113
a signal and a receptor. We then analyze how information is transmitted in a chain of
114
consecutive binding interactions. We then focus on information acquisition in gene
115
regulation by the reversible binding of a TF to the DNA. Finally, we assemble all these
116
components to help us understand information acquisition in a simple model of a linear
117
signaling pathway. Our results show that, contrary to what is expected, under a broad
118
range of biochemically sensible parameters, a noisy dynamic of the molecular
119
interactions increases information acquisition in signaling pathways.
120
121
Results
122
The models
123
We study multiple models that represent either different fundamental steps of a
124
signaling pathway or a complete pathway. All these models include an input or signal
125
molecule S and an output O that conveys information about the signal’s value. We
6
126
quantified (1) noise as the average standard deviation of the response distributions, (2)
127
the output range as the maximal difference of the means of the response distributions,
128
and (3) information as the mutual information between the signal and the output (see
129
Methods). We estimate these quantities through at least 1000 stochastic simulations for
130
each of n evenly distributed values of the number of signal molecules (NS) within the
131
interval [NSmax/n,NSmax].
132
133
In all our models, the signal is detected by reversibly binding of a molecule to either a
134
receptor R or to a DNA binding site (DNAbs). Hence, all models contain at least one
135
reversible binding interaction between molecules. We describe the affinity of two
136
reversibly binding molecules with the equilibrium constant Keq(M)=kd/ka, where kd and
137
ka represent the dissociation and association rate, respectively. The equilibrium
138
constant represents the concentration of free signal molecules at which half of the
139
receptors are bound to a signal molecule. As the equilibrium constant decreases, the
140
concentration of signal molecules required to occupy 50% of the receptors decreases
141
too. Hence, smaller Keq means higher affinity.
142
143
Throughout this paper, we will refer to low, intermediate and high affinities in the
144
following sense. A low affinity refers to an equilibrium constant that is much higher than
145
the maximal concentration of the signal. A high affinity refers to an equilibrium constant
146
that is much lower than the maximal concentration of the signal. Finally, an intermediate
147
affinity refers to an equilibrium constant that is between the minimal and the maximal
148
concentration of the signal. In all our models we considered biologically meaningful
149
values of all biochemical parameters (See Methods and S1-4 Tables).
150
7
151
Reversible binding of molecules
152
We first study the reversible binding between two types of molecules, S and R that form
153
RS complexes (Fig 2a). In this highly simplified model of an information transmission
154
system, we consider the number of RS complexes as the output or response that conveys
155
information about the presence of the signal S. Although this notation is suggestive of
156
interactions between a signal (S) and a receptor (R), our framework below applies to
157
any other reversible binding of two molecules that form a complex. However, for
158
simplicity, we will refer to R molecules as receptors, and to S molecules as signal
159
molecules.
160
161
Fig 2. Noise, output range and information in the reversible binding of molecules. (a) Schematic representation
162
of reversible binding involving a receptor and a signal as examples. ka and kd correspond to the association and
163
dissociation rate, respectively. (b) Acquired information, output range, and noise for receptor-ligand binding at
164
different affinity values (Keq). The red circle denotes the affinity at which mutual information between signal and
165
output is optimized. Information, noise, and output range are normalized by their respective maximal values. Further
166
panels show the system’s behavior at (c,d) low affinity (Keq=10-5), (e,f) high affinity (Keq=10-9), and (g,h) intermediate
167
affinity (Keq=10-7;g-h). (c, e, g) show the temporal dynamic of the receptor-signal complexes (NRS) at three different
168
concentrations of the signal S. (d, f, h) show response distributions at these signal concentrations (see color legend at
169
the bottom of the figure).
170
171
We asked how noise, output range and information change with the affinity between
172
receptor and signal molecules. For this analysis, we assumed that the concentration of
173
the receptors is 10-8M and that the concentration of signal molecules lies within the
174
range [10-8M,10-6M]. This means that the maximal number of signal molecules is greater
175
than the total number of receptors. Our simulations allow us to distinguish three
176
regimes as a function of affinity. First, when affinity is low, noise, output range and
177
information are close to zero (Fig 2b). The reason is that few receptor-signal complexes
8
178
form at low affinity, regardless of the signal concentration (S1 Fig). Consequently, the
179
number of receptor-signal complexes (NRS) is close to zero for all values of the signal
180
concentration (i.e., NRS0 for all NS as Keq∞; Fig 2c). For this reason, response
181
distributions overlap greatly (Fig 2d), causing the output range and information to
182
approach zero (Fig 2b). In addition, noise is also close to zero because the number of
183
receptor-signal complexes fluctuates little (Fig 2c).
184
185
Second, when affinity is high, receptors are saturated at most or all signal concentrations
186
(S1 Fig), such that the number of receptor-signal complexes is equal to the total number
187
of receptors (i.e., NRSNR for all NS as Keq0; Fig 2e), and the output range approaches
188
zero (Fig 2b). Noise approaches zero as well (Fig 2b), because the number of receptor-
189
signal complexes barely fluctuates from its large value (Fig 2e), and because the overlap
190
between response distributions is large (Fig 2f), acquired information approaches zero
191
as well (Fig. 2b).
192
193
All this changes at intermediate affinities, where receptors can acquire information
194
about the number of signal molecules, because receptors are no longer mainly saturated
195
or unoccupied. Instead, the number of receptor-signal complexes fluctuates (Fig 2g).
196
These fluctuations increase noise, but at the same time they permit that the mean
197
number of receptor-signal complexes differs for different number of signal molecules. As
198
a result, the output range increases (Fig 2b), which decreases the overlap between
199
output distributions (Fig 2h), increasing the acquired information (Fig 2b). These
200
observations show that a noisy signal-receptor binding dynamics can be beneficial when
201
a receptor is to acquire information about a signal. Remarkably, the amount of acquired
202
information is maximal when noise is close to its maximally possible value (Fig 2b). In
9
203
sum, if information is acquired through reversible binding interactions, the binding
204
kinetics that yield close to maximal noise also yield close to maximal information.
205
206
Next, we wondered if one can preserve the low noise of high and low affinity binding,
207
while increasing the output range to maximize information acquisition. In doing so, we
208
studied how changes in signal and receptor concentrations affect information, noise and
209
output range. These concentrations, together with the affinity, completely determine the
210
system’s behavior. We varied these concentrations in two different ways. First, we
211
varied the concentrations of the receptors and signal molecules by identical amounts,
212
which keep the ratio of receptors to signal molecules constant. Second, we only varied
213
the concentration of the signal, which changes this ratio. In both cases, we found the
214
same qualitative relationship between noise, output range, and information as before, as
215
long as the maximal number of signal molecules is in excess of the number of receptors.
216
In other words, efficient information acquisition requires high levels of noise and a high
217
output range (Fig 3a-c; S2 Fig). The higher the signal and receptor concentrations are,
218
the lower are the affinity values required for efficient information acquisition (Fig 3a).
219
The reason is that a receptor’s affinity to its signal needs to decrease as signal
220
concentration increases; otherwise receptors become saturated and no longer detect
221
signal changes effectively.
222
223
Fig 3. Acquired information, output range and noise at different signal concentrations. Contour plots of (a)
224
information, (b) output range, and (c) noise at different receptor signal ratios NRT/NSmax (x axis), and at different
225
affinities (y axis). (a-c) The large red-dashed rectangles circumscribe biologically common ranges of receptor-signal
226
affinities ([10-6M,10-9M]) and NRT/NSmax ratios (NRT=10-8 and 10-7M≤NSmax≤10-5M). The small red-dashed rectangle
227
circumscribes the region where NSmax=NRT and where the system is noise-free, reaches the maximally possible output
228
range, and where information acquisition is ‘perfect’. Acquired information, output range and noise are plotted from
229
minimally to maximally observed values, color-coded as indicated by the color bar.
10
230
231
The only scenario where low noise allows maximal information acquisition requires
232
fewer signal molecules than receptors (Fig 3a-c, lower right corners small red
233
rectangle). As an extreme case, one can think of a system with an infinitely large number
234
of receptors, a finite number of signal molecules, and extremely high receptor-signal
235
affinity. In such a system all signal molecules are bound to receptors. Because there are
236
fewer signal molecules than receptors, the system effectively ‘counts’ the number of
237
signal molecules through the number of receptor-signal complexes. Notice that
238
experimentally measured affinity values between receptors and signals, are not
239
extremely high. Instead, they are of the same order of magnitude as signal and receptor
240
concentrations (11,28). In our simulations, these are the affinity values where high
241
information acquisition entails high noise (Fig 3a-c big red rectangle), suggesting that
242
biological system operate in the noisy regime. In sum, under biologically feasible
243
conditions, high noise is necessary for (maximal) information acquisition.
244
245
Consecutive reversible binding interactions
246
In a signaling pathway, the binding of a signal to a receptor is usually the first of a chain
247
of reversible events. These events include the reversible modification of one or more
248
intermediary signaling molecules, and they usually terminate in the reversible binding
249
of transcriptional regulators, such as a transcription factors (TF), to DNA. TF-DNA
250
binding differs from other signaling binding interactions because regulated genes have
251
one or few copies in any one genome, and any one regulated gene harbors few – usually
252
fewer than ten – TF-binding sites (29,30). In the simplest signaling pathways, signal-
253
bound receptors can directly regulate transcription without intervening signaling steps
254
(31).
11
255
256
To study how TF-DNA binding might affect information acquisition in such a pathway,
257
we model two consecutive reversible binding interactions. They represent the formation
258
of a receptor-signal complex, and the binding of this complex to a DNA binding site
259
(DNAbs). We assumed that the concentration of receptor molecules is 10-8M, that the
260
concentration of signal molecules lies in the interval [10-8M, 10-6M] and that a single
261
DNA binding site mediates transcriptional regulation. The receptor-signal-DNAbs
262
complex represents the ultimate output of the system that harbors information about
263
the signal.
264
265
We analyzed how the affinities of both the receptor to the signal (KeqR,S) and of the
266
receptor-signal complex to DNA (KeqRS,D) affect the acquisition of information, output
267
range and noise. As in the simpler two-molecule system, the receptor is able to detect
268
different signal concentration at intermediary affinity values between the receptor and
269
the signal, where the largest output ranges, which are necessary for the receptor to
270
sense the signal, are produced with high levels of noise (Fig 4a-c).
271
272
Fig 4. Information, and noise in a pair of reversible binding interactions. Contour plots of (a and d) noise, output
273
range (b and e) and information acquisition (c and f) in the receptor-signal complex (RS; a-c) and in the receptor-
274
signal-DNAbs complex (RSD; d-f) as a function of the affinities between both the receptor and the signal (KeqR,S), and the
275
receptor-signal complex with the downstream molecule (KeqRS,D). Red-dashed rectangles circumscribe biologically
276
sensible receptor-signal DNA affinities ([10-8M,10-13M]) and receptor signal affinities ([10-6M,10-9M]). White-dashed
277
rectangles delineate the region of maximal information acquisition at the receptor-signal-DNAbs level. Acquired
278
information, output range and noise are plotted from minimally to maximally observed values, color-coded as
279
indicated by the color bar.
280
12
281
To subsequently transmit the information acquired by the receptor-signal complex to
282
the receptor-signal-DNAbs complex, DNA binding needs to be subject to the same kind of
283
fluctuations. Such fluctuations only occur at intermediary affinity values between the
284
receptor-signal complex and the DNA, otherwise the DNA binding site is either almost
285
always bound (saturated) or unbound by the receptor-signal molecules. The fluctuations
286
in receptor-signal-DNA binding increase noise (Fig 4d), but they also lead to different
287
probabilities of DNA binding for different concentrations of the receptor-signal complex,
288
which increases the output range (Fig 4e). As a result, the acquisition of information
289
increases (Fig 4f). In sum, information about a signal is obtained at intermediate values
290
of both affinities (compare the white rectangles, indicating the region with maximal
291
information at the receptor-signal-DNAbs level in Fig 4). We also note that the affinities
292
leading to high information acquisition and high noise in our model are similar to
293
experimentally measured affinities between receptors and signals, as well as between
294
transcriptional regulators to DNA (Fig 4, large red rectangles). Repeating our analysis
295
with up to ten DNA binding sites leads to the same conclusions (S3 Fig): A noisy dynamic
296
is essential to acquire information.
297
298
Gene expression system
299
At the end of signaling pathways stands the regulation of gene expression, which usually
300
requires reversible binding (of a transcription factor to DNA), and additionally involves
301
the synthesis and degradation of mRNA and protein. To find out whether the observed
302
relationship between information, output range, and noise is similar in the presence of
303
synthesis and degradation, we modeled the regulation of gene expression mediated by a
304
transcription factor that reversibly binds to DNA. We assumed that a gene with a single
305
DNA binding site drives transcription initiation, which occurs only when the binding site
13
306
is bound by a transcription factor. In this case, mRNA molecules are transcribed at rate
307
k1, and proteins are translated from the mRNA molecules at a rate k2. Both mRNA and
308
protein molecules become degraded at rates, d1 and d2, respectively (Fig 5a). We
309
considered the number of TF molecules as the signal, and the number of protein
310
molecules NP as the output or response.
311
312
Fig 5. Noise, output range and information in gene regulation. (a) Schematic representation of our model of gene
313
regulation. ka and kd correspond to the association and dissociation rate, respectively of a TF with its DNA binding
314
site; k1 and k2 correspond to the mRNA and protein synthesis rate, respectively; d1 and d2 correspond to the mRNA
315
and protein degradation rates, respectively. (b) Information, output range and noise observed in numerical
316
simulations of the system at different TF-DNA affinities (Keq). Information, noise, and output range are normalized by
317
their respective maximal values. The red circle denotes the affinity at which maximal information is acquired. System
318
behavior at low (Keq=10-9; c and d), intermediate (Keq=10-11; e and f), and high (Keq=10-13; g and h) affinities. Temporal
319
protein dynamics at three different TF concentrations are shown in c, e and g. Response distribution for the same
320
simulations are shown in d, f and h. The blue dashed line in c-h marks the expected mean protein value for
321
constitutive (unregulated, always-on) expression.
322
323
We started by analyzing how a TF’s affinity to its DNAbs affects the relationships between
324
information, output range, and noise (Fig 5b). As in receptor-signal binding, at the
325
lowest affinities, the DNA binding site is almost never bound by TF molecules, regardless
326
of the number of TF molecules (S4a Fig). Hence, little mRNA and protein is produced,
327
independently of the number of TF molecules (Fig 5c and S4 Fig). Response
328
distributions are insensitive to the TF concentration, with a mean close to zero and
329
almost no variation (Fig 5d). For this reason, both noise and output range approach zero
330
as the affinity approaches zero, and so does the acquired information (Fig 5b).
331
332
At the highest affinities, noise shows one noticeable difference to the reversible binding
333
of molecules (Fig 2b): it does not decrease to zero (Fig 5b). The reason is that mRNA and
14
334
protein production have ‘bursty’ dynamics with large excursions from a base line. This
335
bursty dynamics comes from the stochastic nature of mRNA and protein production,
336
which causes fluctuations in the concentration of both kinds of molecules (23). For this
337
reason, gene expression is intrinsically noisy. In particular, for a gene with constitutive
338
expression, the expected number of protein molecules NP and its expected noise
339
(standard
deviation)
are
equal
to
and
𝐸(𝑁𝑃) = (𝑘1/𝑑2)(𝑘2/𝑑1)
𝐸(𝜎(𝑁𝑃)) =
340
, respectively (24). At the highest affinities, the system behaves like a
(𝑘1/𝑑2)(𝑘2/𝑑1)2
341
constitutive gene, because TF molecules are almost always bound to the DNAbs (S4a Fig),
342
and the regulated gene is thus constantly transcribed. Accordingly, in our simulations,
343
the mean number of expressed protein molecules and its standard deviation are close to
344
the expected values for a constitutive gene independently of the number of TF molecules
345
(Fig 5e; S4c Fig; S5 Fig). Because the mean is similar for all number of TF molecules, the
346
output range tends to zero (Fig 5b). However, the amount of noise is higher than zero
347
(Fig 5b) and similar to that expected for a constitutive gene (S5 Fig). As a result, all
348
response distributions are similar to those for the highest concentration of the
349
transcription factor (Fig 5f), and the amount of acquired information about this
350
concentration is small (Fig 5b).
351
352
At intermediary affinities, the DNA binding site is not always bound by a TF. It fluctuates
353
between a bound (active) state, when protein molecules are synthesized, and an
354
unbound (inactive) state, when previously synthesized proteins are degraded (Fig 5g,
355
S4a Fig). These fluctuations increase the noise in protein concentrations at intermediate
356
affinities relative to low and high affinities (Fig 5b). They also increase the output range
357
of the system (Fig. 5b). Most importantly, the probability that a binding site is bound by
358
a TF changes with the number of TF molecules, which renders the system’s output – the
15
359
number of synthesized proteins – sensitive to its input (Fig 5g). Hence, the amount of
360
information acquired about this input increases too (Fig 5b). In other words, noise can
361
increase the acquisition of information also in this gene expression system. Moreover,
362
once again, acquired information is maximal when noise is close to its maximal value
363
(Fig 5b).
364
365
Information, noise and output range in a complete signaling pathway
366
367
In a final analysis, we assembled all of the above elements – receptor-signal binding, TF-
368
DNA binding, and gene regulation – into a model of a simple complete signaling
369
pathway. This pathway is akin to a nuclear hormone receptor pathway, such as the
370
signaling pathway of estrogen, progesterone, and various other lipid-soluble signals
371
(31). In this pathway, we quantified the amount of information about the concentration
372
of the input (hormone) signal that is contained in the number of expressed protein
373
molecules. This analysis confirmed our previous results. As in the simpler systems,
374
maximal information acquisition requires high noise, which increases the pathway’s
375
sensitivity to variations in the signal (Supplementary text; S6 Fig).
376
377
Discussion
378
A fundamental step in signaling pathways is the reversible binding of molecules, which
379
is necessary for the detection of a signal by receptors and for the acquisition of
380
information about this signal. Previous experimental and theoretical work has
381
demonstrated that biological processes, including signaling pathways and their binding
382
interactions, are inherently noisy (23). One would thus expect that the kinetic
383
parameters of binding interactions have evolved to minimize noise, because noise is
16
384
detrimental for the acquisition of information (13,14,20). However, we find the
385
opposite. The kinetic parameters of signaling pathways must produce noisy binding
386
dynamics or a signaling pathway will acquire little or no information. This is due to the
387
nature of reversible binding interactions. Under biologically sensible parameter values
388
and realistic concentrations of ligands and receptors, binding of molecules is noise-free
389
only when a receptor is completely saturated with its ligand, or if it is unable to bind the
390
ligand. In either case, information acquisition is impossible. Hence, noise in molecular
391
binding is not just unavoidable but necessary for information acquisition in signaling
392
pathways. Importantly, the positive role of noise for information acquisition is not
393
limited to individual binding interactions, but also occurs in more complex systems that
394
include gene expression regulation and more than one binding interaction.
395
396
In our models, we observe only one condition where noise is not required for
397
information acquisition. At high signal-receptor affinity, a noise-free ‘perfect’ detection
398
of a signal is possible when the number of receptors is greater than the number of signal
399
molecules. However, producing more receptors than signaling molecules would incur
400
enormous energetic costs. Relatedly, transcriptional regulation generally involves fewer
401
than ten TF binding sites per regulated gene – the analog of a receptor in such a system
402
(29,30) – a number that is much smaller than the average number of transcription
403
factors per cell, which are usually in the hundreds for bacteria and in the thousands for
404
yeast and mammal cells (32,33). Hence, a perfect detection of the number of TFs or
405
signal molecules is not biologically plausible.
406
407
Some previous work hinted at a positive role of noise for information acquisition
408
(11,25,27), but this work was not ideally suited to understand the mechanisms by which
17
409
noise helps increase information acquisition: It did not focus on signaling pathways, did
410
not model molecular interactions explicitly, or it assumed that noise comes from an
411
external source and can be made arbitrarily small. In contrast, our models represent
412
molecular interactions explicitly, which causes noise to emerge naturally from them. In
413
doing so, they also provide a mechanistic explanation of the relationship between noise
414
and information acquisition. However, our models focus on the simplest molecular
415
interactions, and they do not exhaust all possible signaling interactions. Whether other
416
properties of signaling pathways change the way kinetic parameters affect noise and
417
information acquisition is an important task for future work.
418
419
Our models include multiple simplifying assumptions. For example, we assumed that the
420
numbers of signaling molecules, receptors, and transcriptional regulators are constant,
421
whereas they may change dynamically in cells. We also considered a simple linear
422
pathway, whereas signaling pathways usually contain regulatory motifs, such as
423
feedback circuits and feed-forward loops (34). In addition, we did not consider
424
molecular interactions such as dimerization (18,19). Similarly, we did not consider the
425
costs of expressing an information processing machinery (20). Because these factors do
426
not affect the nature of reversible binding, we suspect that they might also not reduce
427
the positive role of noisy binding dynamics for information acquisition. However, some
428
of them might increase information acquisition at low noise by other means. For
429
example, some signaling mechanisms increase the amount of information acquired
430
while decreasing noise (11,12,15–19). In contrast to the molecular interactions we
431
study, where noise increases information acquisition by increasing a system’s output
432
range, these mechanisms maintain the output range while decreasing noise (14). It
18
433
remains to be seen how such different mechanisms interact and jointly affect how
434
biological systems acquire information.
435
436
Methods
437
Reversible and consecutive molecular binding models
438
We consider two kinds of molecules, S (signal) and R (receptor), which can associate
439
reversibly into receptor-signal complexes at an association rate ka (M−1s−1), and a
440
dissociation rate kd (s−1).
441
442
To model consecutive reversible binding steps, we assume that, first, a signal (S) and a
443
receptor (R) reversibly associate into a receptor-signal (RS) complex. Second, this
444
complex binds reversibly to a downstream molecule (D), such as DNA. We denote the
445
rate of association between the signal and the receptor by kaR,S (M−1s−1), and that of
446
dissociation by kdRS (s−1). Similarly, we denote the rate of association between the
447
receptor-signal complex and the downstream molecule by kaRS,D (M−1s−1), and that of
448
dissociation by kdRSD (s−1).
449
450
Gene expression system
451
We model a gene expression system where one chemical species, denoted as TF
452
(transcription factor), binds to a DNA binding site (DNAbs) to regulate the expression of a
453
nearby gene. TF molecules associate with the DNAbs at a rate ka (M−1s−1). The
454
dissociation of TF-DNAbs complexes happens at a rate kd (s−1). In the disassociated state,
455
no transcription occurs, and in the associated state transcription occurs at a rate k1 (s−1).
456
Transcribed mRNA molecules are degraded at a rate d1 (s−1). Finally, proteins are
457
translated from mRNA molecules at a rate k2 (s−1), and degraded at a rate d2 (s−1).
19
458
459
Complete linear signaling pathway
460
Our model considers the reversible receptor-ligand complex formation and gene
461
expression activation, which is mediated by the receptor-signal complex. Consequently,
462
the parameters that govern the behavior of such a pathway are similar to those
463
described so far, namely: 1) an association rate (kaR,S) between the signal and receptor
464
(R) and a dissociation rate(kdRS) of the receptor-signal complexes (RS), 2) an association
465
(kaRS,D) and a dissociation (kdRSD) rate between RS and a DNA binding site (DNAbs), and 3) a
466
rate of gene transcription (mRNA synthesis, k1), mRNA degradation (d1), protein
467
synthesis (k2), and protein degradation (d2).
468
469
Stochastic simulations
470
We simulated the behavior of the models described above using Gillespie’s discrete
471
stochastic simulation algorithm (35), using the numpy python package for scientific
472
computing (http://www.numpy.org/). Gillespie’s algorithm captures the stochastic
473
nature of chemical systems. It assumes a well-stirred and thermally equilibrated system
474
with constant volume and temperature. The algorithm requires the probability pj that a
475
chemical reaction Rj occurs in a given time interval [t,t+). Any such probability pj is
476
proportional to both the reaction rate and the number of reacting molecules. Notice that
477
for first-order reactions, such as the dissociation of a molecular complex into its
478
constituent molecules, pj is independent of the volume in which the reaction takes place.
479
In contrast, pj is inversely proportional to the volume in second-order reactions, such as
480
the association of two molecules. For the reversible molecular binding modeled here,
481
the probabilities pa and pd that two molecules associate and dissociate, respectively, are
482
proportional to
20
483
𝑝𝑎 =
𝑘𝑎
𝑉𝑁𝐴𝑁𝑎𝑁𝑏
484
𝑝𝑑 = 𝑘𝑑𝑁𝑐
485
where V is the reaction volume, NA is Avogadro’s number, and Na, Nb, and Nc are the
486
numbers of molecules of the two chemical species a and b and of the complex c.
487
488
The probabilities pmRs, pmRd, pPs and pPd of a mRNA transcription, mRNA degradation,
489
protein synthesis, and protein degradation event are given by
490
𝑝𝑚𝑅𝑠 = 𝑘1𝑁𝑐
491
𝑝𝑚𝑅𝑑 = 𝑑1𝑁𝑚𝑅
492
𝑝𝑃𝑠 = 𝑘2𝑁𝑚𝑅
493
,
𝑝𝑃𝑑 = 𝑑2𝑁𝑃
494
respectively. In these expressions, the quantities NmR and NP are the numbers of mRNA
495
molecules and of protein molecules, respectively. We model a haploid organism with
496
only a single DNA binding site, corresponding to a single regulated gene. In this case, the
497
probability of mRNA synthesis can be reduced to
498
𝑝𝑚𝑅𝑠 = 𝑘1
499
when the binding site is bound by transcription factor (Nc=1) and to
500
𝑝𝑚𝑅𝑠 = 0
501
when the binding site is unbound (Nc=0).
502
503
Initial conditions for simulations
504
To determine the initial conditions of the system, we first calculated the expected
505
number of complexes formed as
506
𝑁𝐶 = 𝑁𝐵𝑇
𝑁𝐴
𝐾𝑒𝑞 + 𝑁𝐴
21
507
The quantity NA is the number of signal or TF molecules and NBT is the total number of
508
receptor molecules or DNA binding sites. Notice that
is a real number, and for the
𝑁𝐶
509
specific case of the TF-DNAbs interaction it can only take a value between 0 and 1. Thus,
510
equals to the probability that the DNA binding site is bound by a transcription
𝑁𝑅𝑆
511
factor. However, for the receptor signal complexes, it represents the number of
512
complexes formed. We selected the initial state of the number of complexes (
) at
𝑁𝐶𝑖
513
random with binomial probability for the binding site. However, we selected the closest
514
integer to
for the receptor signal case. Then we defined
𝑁𝐶
515
𝑁𝐴𝑖 = 𝑁𝐴 ‒ 𝑁𝐶𝑖
516
𝑁𝐵𝑖 = 𝑁𝐵𝑇 ‒ 𝑁𝐶𝑖
517
as the initial state of the number of free signal or TF molecules (
) and of the receptors
𝑁𝐴𝑖
518
or binding sites (
). Finally, as the initial state of the number of mRNA and protein
𝑁𝐵𝑖
519
molecules we used
520
𝑁𝑚𝑅𝑖 = 𝑁𝐶
𝑘1
𝑑1
521
𝑁𝑃𝑖 = 𝑁𝐶
𝑘1
𝑑1
𝑘2
𝑑2
522
which is the expected average number of mRNA and protein molecules for a
523
constitutively expressed gene (23), multiplied by the probability that the DNA binding
524
site is bound by a TF molecule.
525
526
Information quantification
527
The number of molecules of any chemical species in a cell or in a unit volume fluctuates,
528
because molecules are produced and decay stochastically, and because they undergo
22
529
random Brownian motion caused by thermal vibrations. We use Shannon’s entropy to
530
quantify the unpredictability caused by such stochastic fluctuations in a signal as
531
𝐻(Pr(𝑆)) =‒
𝑁𝑆𝑚𝑎𝑥
∑
𝑁𝑆 =
𝑁𝑆𝑚𝑎𝑥
𝑛
𝑝(𝑁𝑆)log2𝑝(𝑁𝑆),
532
where Pr(S) is the probability distribution of the signal, and p(NS) is the probability that
533
the system contains N molecules of the signal. In our models the signal represents either
534
a molecular signal or cue (2).
535
536
For all our analyses, we performed at least 1000 simulations for each of n different
537
numbers of signal molecules that were evenly distributed within the interval
538
[NSmax/n,NSmax] (n≤NSmax). For this reason
539
𝐻(Pr(𝑆)) = log2𝑛
540
Signals trigger changes in a cell’s state that produce a response or output (O) of the
541
system, such as the production of molecules. A cell acquires information when the
542
output O reflects (fully or partially) the value of S. This information can be quantified via
543
the mutual information:
544
,
𝐼(𝑆;𝑂) = 𝐻(Pr(𝑆)) ‒ 𝐻(Pr(𝑆∣𝑂))
545
which is equal to the entropy H(Pr(S)) minus the conditional entropy H(Pr(S|O)), which
546
represents the entropy of S given that O is known (13). In other words, the mutual
547
information I quantifies the acquired information as the amount of information that an
548
output chemical species O harbors about a signal S.
549
550
Noise quantification and output range
23
551
The systems we model produce a probabilistic response for any given quantity NS of the
552
signal. This response can thus be represented as a conditional probability distribution:
553
Pr (0 < 𝑁𝑂 < 𝑁𝑂𝑚𝑎𝑥|𝑆 = 𝑁𝑆),
554
where NO and NOmax are the number and maximal number of output molecules,
555
respectively. We estimated this response distribution through at least 1000 replicate
556
simulations of the system for each value of NS. We then quantified noise as the standard
557
deviation of the response distributions, averaged over all n possible values of NS:
558
σ =
1
𝑛
𝑁𝑆𝑚𝑎𝑥
∑
𝑁𝑆 = 𝑁𝑆𝑚𝑖𝑛
σ(Pr(𝑂|𝑆 = 𝑁𝑠)),
559
We define the output range as the difference between the maximal and minimal mean
560
value of all response distributions.
561
562
Parameter values
563
Our simulations considered the following biologically sensible parameter ranges. The
564
association and dissociation constants ka and kd of reversible complex formation define
565
the equilibrium constant Keq=kd/ka (M), which we used in our simulations. The smaller
566
Keq becomes, the more association becomes favored over dissociation (28). In particular,
567
for the binding between ligands and (nuclear) receptors, we used values of Keq within
568
the interval [10-6M,10-9M], because the micromolar to nanomolar range is common for
569
such complexes (28,32,36–38). For TF-DNA binding, empirical data suggests that usually
570
Keq<10-8 and can reach picomolar (10-12M) or even smaller values (28,32,36–38). Thus,
571
we used values in the interval [10-8M,10-12M].
572
573
For mRNA, experimentally measured half-lives usually lie in the range of seconds to
574
hours (39–42). Protein half-lives typically lie between hours and days (41,43). Taking all
24
575
this information into consideration, we chose mRNA half-lives within the interval
576
[1min,30min], and protein half-lives where within [15min, 3h]. We assume that the ratio
577
k2/k1, which describes the speed of the protein synthesis rate relative to the mRNA
578
synthesis rate, exceeds 1.0 (44). Because the residence time of transcription factors on
579
DNA lies within seconds to hours (45,46), we assumed a residence time within this
580
interval [10sec,2h]
581
582
Finally, we always considered concentrations of molecules to lie within the interval [10-
583
9M,10-6M], because these are typical concentration of most molecules within a cell or a
584
nucleus (47). Notice that for some of our simulation we also needed to explore values
585
outside these ranges. Specific parameter values used for each simulation are listed in
586
supplementary tables 1-4.
587
588
Acknowledgments
589
We acknowledge support from the European Research Council under Grant Agreement
590
No. 739874, by Swiss National Science Foundation grant 31003A_172887, as well as by
591
the University Priority Research Program in Evolutionary Biology.
592
593
References
594
1.
Bowsher CG, Swain PS. Environmental sensing, information transfer, and cellular
595
decision-making. Current Opinion in Biotechnology. 2014 Aug;28:149–55.
596
2.
Laidre ME, Johnstone RA. Animal signals. Current Biology. 2013
597
Sep;23(18):R829–33.
598
3.
Boyer PD, Krebs EG, Tamanoi F, editors. The Enzymes. 3d ed. New York:
599
Academic Press; 1970. 1 p.
600
4.
Cantley LC. Signal transduction: principles, pathways, and processes. Cold Spring
601
Harbor, N.Y.: Cold Spring Harbor Laboratory; 2014.
602
5.
Tsimring LS. Noise in biology. Reports on Progress in Physics. 2014 Feb
603
1;77(2):026601.
604
6.
Pilpel Y. Noise in Biological Systems: Pros, Cons, and Mechanisms of Control. In:
25
605
Castrillo JI, Oliver SG, editors. Yeast Systems Biology [Internet]. Totowa, NJ: Humana
606
Press; 2011 [cited 2019 Aug 9]. p. 407–25. Available from:
607
http://link.springer.com/10.1007/978-1-61779-173-4_23
608
7.
Selimkhanov J, Taylor B, Yao J, Pilko A, Albeck J, Hoffmann A, et al. Accurate
609
information transmission through dynamic biochemical signaling networks. Science.
610
2014 Dec 12;346(6215):1370–3.
611
8.
Uda S, Saito TH, Kudo T, Kokaji T, Tsuchiya T, Kubota H, et al. Robustness and
612
Compensation of Information Transmission of Signaling Pathways. Science. 2013 Aug
613
2;341(6145):558–61.
614
9.
Cheong R, Rhee A, Wang CJ, Nemenman I, Levchenko A. Information Transduction
615
Capacity of Noisy Biochemical Signaling Networks. Science. 2011 Oct
616
21;334(6054):354–8.
617
10.
Tkacik G, Callan CG, Bialek W. Information flow and optimization in
618
transcriptional regulation. Proceedings of the National Academy of Sciences. 2008 Aug
619
26;105(34):12265–70.
620
11.
Tkačik G, Walczak AM, Bialek W. Optimizing information flow in small genetic
621
networks. Physical Review E [Internet]. 2009 Sep 29 [cited 2019 Aug 9];80(3). Available
622
from: https://link.aps.org/doi/10.1103/PhysRevE.80.031920
623
12.
Thattai M, van Oudenaarden A. Intrinsic noise in gene regulatory networks.
624
Proceedings of the National Academy of Sciences. 2001 Jul 17;98(15):8614–9.
625
13.
Cover TM, Thomas JA. Elements of information theory. 2nd ed. Hoboken, N.J:
626
Wiley-Interscience; 2006. 748 p.
627
14.
Rhee A, Cheong R, Levchenko A. The application of information theory to
628
biochemical signaling systems. Physical Biology. 2012 Aug 1;9(4):045011.
629
15.
Rieckh G, Tkačik G. Noise and Information Transmission in Promoters with
630
Multiple Internal States. Biophysical Journal. 2014 Mar;106(5):1194–204.
631
16.
Tkačik G, Walczak AM, Bialek W. Optimizing information flow in small genetic
632
networks. III. A self-interacting gene. Physical Review E [Internet]. 2012 Apr 5 [cited
633
2019 Aug 9];85(4). Available from:
634
https://link.aps.org/doi/10.1103/PhysRevE.85.041903
635
17.
Walczak AM, Tkačik G, Bialek W. Optimizing information flow in small genetic
636
networks. II. Feed-forward interactions. Physical Review E [Internet]. 2010 Apr 6 [cited
637
2019 Aug 9];81(4). Available from:
638
https://link.aps.org/doi/10.1103/PhysRevE.81.041905
639
18.
Liu F-Y, Lo S-C, Shu C-C. The Reaction of Dimerization by Itself Reduces the Noise
640
Intensity of the Protein Monomer. Scientific Reports [Internet]. 2019 Dec [cited 2019
641
Aug 9];9(1). Available from: http://www.nature.com/articles/s41598-019-39611-6
642
19.
Ghim C-M, Almaas E. Genetic noise control via protein oligomerization. BMC
643
Systems Biology [Internet]. 2008 Dec [cited 2019 Aug 9];2(1). Available from:
644
https://bmcsystbiol.biomedcentral.com/articles/10.1186/1752-0509-2-94
645
20.
ten Wolde PR, Becker NB, Ouldridge TE, Mugler A. Fundamental Limits to Cellular
646
Sensing. Journal of Statistical Physics. 2016 Mar;162(5):1395–424.
647
21.
Pahle J, Green AK, Dixon CJ, Kummer U. Information transfer in signaling
648
pathways: A study using coupled simulated and experimental data. BMC Bioinformatics.
649
2008 Dec;9(1):139.
650
22.
Richard M, Yvert G. How does evolution tune biological noise? Frontiers in
651
Genetics [Internet]. 2014 Oct 28 [cited 2019 Aug 9];5. Available from:
652
http://journal.frontiersin.org/article/10.3389/fgene.2014.00374/abstract
653
23.
Raj A, van Oudenaarden A. Single-Molecule Approaches to Stochastic Gene
26
654
Expression. Annual Review of Biophysics. 2009 Jun;38(1):255–70.
655
24.
Kepler TB, Elston TC. Stochasticity in Transcriptional Regulation: Origins,
656
Consequences, and Mathematical Representations. Biophysical Journal. 2001
657
Dec;81(6):3116–36.
658
25.
Lestas I, Vinnicombe G, Paulsson J. Fundamental limits on the suppression of
659
molecular fluctuations. Nature. 2010 Sep;467(7312):174–8.
660
26.
Bialek W, Setayeshgar S. Physical limits to biochemical signaling. Proc Natl Acad
661
Sci U S A. 2005 Jul 19;102(29):10040.
662
27.
Bowsher CG, Voliotis M, Swain PS. The Fidelity of Dynamic Signaling by Noisy
663
Biomolecular Networks. Haugh JM, editor. PLoS Comput Biol. 2013 Mar
664
28;9(3):e1002965.
665
28.
Corzo J. Time, the forgotten dimension of ligand binding teaching. Biochemistry
666
and Molecular Biology Education. 2006 Nov;34(6):413–6.
667
29.
Lusk RW, Eisen MB. Evolutionary Mirages: Selection on Binding Site Composition
668
Creates the Illusion of Conserved Grammars in Drosophila Enhancers. Barsh GS, editor.
669
PLoS Genetics. 2010 Jan 22;6(1):e1000829.
670
30.
Bilu Y, Barkai N. The design of transcription-factor binding sites is affected by
671
combinatorial regulation. Genome Biology. 2005 Dec 2;6(12):R103.
672
31.
Sever R, Glass CK. Signaling by Nuclear Receptors. Cold Spring Harbor
673
Perspectives in Biology. 2013 Mar 1;5(3):a016709–a016709.
674
32.
Biggin MD. Animal Transcription Networks as Highly Connected, Quantitative
675
Continua. Developmental Cell. 2011 Oct;21(4):611–26.
676
33.
Ho B, Baryshnikova A, Brown GW. Unification of Protein Abundance Datasets
677
Yields a Quantitative Saccharomyces cerevisiae Proteome. Cell Systems. 2018
678
Feb;6(2):192-205.e3.
679
34.
Alon U. Network motifs: theory and experimental approaches. Nature Reviews
680
Genetics. 2007 Jun 1;8:450.
681
35.
Gillespie DT. Exact stochastic simulation of coupled chemical reactions. The
682
Journal of Physical Chemistry. 1977 Dec;81(25):2340–61.
683
36.
Belikov S, Berg OG, Wrange Ö. Quantification of transcription factor-DNA binding
684
affinity in a living cell. Nucleic Acids Research. 2015 Dec 10;44(7):3045–58.
685
37.
Fisher WW, Li JJ, Hammonds AS, Brown JB, Pfeiffer BD, Weiszmann R, et al. DNA
686
regions bound at low occupancy by transcription factors do not drive patterned reporter
687
gene expression in Drosophila. Proceedings of the National Academy of Sciences. 2012
688
Dec 26;109(52):21330–5.
689
38.
Li X-Y, Thomas S, Sabo PJ, Eisen MB, Stamatoyannopoulos JA, Biggin MD. The role
690
of chromatin accessibility in directing the widespread, overlapping patterns of
691
Drosophila transcription factor binding. Genome Biology. 2011;12(4):R34.
692
39.
Lugowski A, Nicholson B, Rissland OS. Determining mRNA half-lives on a
693
transcriptome-wide scale. Methods. 2018 Mar;137:90–8.
694
40.
Laalami S, Zig L, Putzer H. Initiation of mRNA decay in bacteria. Cellular and
695
Molecular Life Sciences. 2014 May;71(10):1799–828.
696
41.
Schwanhäusser B, Busse D, Li N, Dittmar G, Schuchhardt J, Wolf J, et al. Global
697
quantification of mammalian gene expression control. Nature. 2011
698
May;473(7347):337–42.
699
42.
Sharova LV, Sharov AA, Nedorezov T, Piao Y, Shaik N, Ko MSH. Database for
700
mRNA Half-Life of 19 977 Genes Obtained by DNA Microarray Analysis of Pluripotent
701
and Differentiating Mouse Embryonic Stem Cells. DNA Research. 2009 Jan 8;16(1):45–
702
58.
27
703
43.
Toyama BH, Hetzer MW. Protein homeostasis: live long, won’t prosper. Nature
704
Reviews Molecular Cell Biology. 2013 Jan;14(1):55–61.
705
44.
Hausser J, Mayo A, Keren L, Alon U. Central dogma rates and the trade-off
706
between precision and economy in gene expression. Nature Communications [Internet].
707
2019 Dec [cited 2019 Aug 9];10(1). Available from:
708
http://www.nature.com/articles/s41467-018-07391-8
709
45.
Morisaki T, Müller WG, Golob N, Mazza D, McNally JG. Single-molecule analysis of
710
transcription factor binding at transcription sites in live cells. Nature Communications
711
[Internet]. 2014 Dec [cited 2019 Aug 9];5(1). Available from:
712
http://www.nature.com/articles/ncomms5456
713
46.
Mueller F, Stasevich TJ, Mazza D, McNally JG. Quantifying transcription factor
714
kinetics: At work or at play? Critical Reviews in Biochemistry and Molecular Biology.
715
2013 Sep;48(5):492–514.
716
47.
Legewie S, Herzel H, Westerhoff HV, Blüthgen N. Recurrent design patterns in the
717
feedback regulation of the mammalian signalling network. Molecular Systems Biology.
718
2008 Jan;4(1):190.
719
720
Supporting information captions
721
S1 Fig. Mean number of receptor-signal complexes (
) formed at different affinity values (Keq). The maximum
𝑁𝑅𝑆
722
number of receptor-signal complexes for this simulation is 50 (see S1 Table).
723
724
S2 Fig. Noise, output range and information observed in numerical simulations of the receptor-signal system at
725
different affinities (Keq) and with different concentrations of R and S. a) R=10-9M and S=[10-9M,10-7M]. b) R=10-8M and
726
S=[10-8M,10-6M]. a) R=10-7M and S=[10-7M,10-5M]. Information, noise, and output range are normalized by their
727
respective maximal values.
728
729
S3 Fig. Information and noise in two consecutive binding interactions with ten DNAbs. Contour plots of (a and d)
730
noise, output range (b and e) and information acquisition (c and f) in the receptor-signal complex (RS; a-c) and in the
731
receptor-signal-DNAbs complex (RSD; d-f) as a function of the affinities between both the receptor and the signal
732
(KeqR,S), and the receptor-signal complex with the downstream molecule (KeqRS,D). Red-dashed rectangles circumscribe
733
biologically sensible receptor-signal DNA affinities ([10-8M,10-13M]) and receptor signal affinities ([10-6M,10-9M]).
734
White-dashed rectangles delineate the region of maximal information acquisition at the receptor-signal-DNAbs level.
735
Acquired information, output range and noise are plotted from minimally to maximally observed values, color-coded
736
as indicated by the color bar.
737
738
S4 Fig. (a) Probability of TF-DNA binding, (b) mean number of mRNA molecules (
), and (c) mean number of
𝑁𝑚𝑅𝑁𝐴
739
protein molecules (
), as a function of TF-DNA affinity (Keq). For the parameters we used (S3 Table), the expected
𝑁𝑃
28
740
mean number of mRNA and protein molecules produced from a constitutively expressed gene is 2 and 100,
741
respectively.
742
743
S5 Fig. Non-normalized values of noise in the number of proteins. The black line is the standard deviation
744
expected for the parameters used in this simulation (S3 Table).
745
746
S6 Fig. Contour plot of acquired information and noise in a complete signaling pathway. (a) Information and (b)
747
noise for different affinities between receptor and signal molecules (KeqR,S, y axis), and receptor-signal complexes with
748
a DNA binding site (KeqRS,D, x axis) in a model of a simple lineal signaling pathway where a signal bound receptor can
749
bind DNA and regulate gene expression. The number of signal molecules constitute the input into this pathway, and
750
the number of protein molecules expressed from the regulated gene constitute the output. The red-dashed rectangle
751
show experimentally measured affinity values between receptors and signals ([10-6M,10-9M]) and between
752
transcriptional regulators and DNA ([10-8M,10-13M]). The amount of acquired information and noise level are
753
indicated in the color bar.
Stimulus 1 Stimulus 2
Output range
Noise level
Output range +
Trnfarmatian
10* 107 10° 10° 10% 10°
Key
10" 10°
time
Probability
Probabili
Response distributions
time
Response distributions
Pe ee ed
Probabili
time
Low S concentration
(10°M)
Intermediate S
WS concentration (107M)
Response distributions
High S concentration
iz (10°M)
1.0 — Information
. vo — Noise
Signal Zs 0.8 — Output range
Ss
>
k,
—> Z 06
XS
—_— 2 0.4
Receptor-Signal 5
0.2
Receptor Complex Zz
0.0:
10" 10° 10% 107 10° 10° 10% 10°
d Koy
100
2 60
20
Do
100-—Y =
2 60 4
= 2 2
°
100 a
2 60
20
time f Response distributions
100 ---------_------__---_-------—_—
2 60
= 20
lity
Receptor-Signal
Receptor Complex
Information Output range
RS output range
10” 10** 10°
Kegrsp
RSD noise RSD output range
200
~
DNAow “o™ DNAon ¥ ¥
b 1.0 @ O
— Information
0.8 — Noise
3 —— Output range
a
>
3 0.6
2
E 04
°
Zz
02
0.0
10 “To? 10" To? “lo 07
200
WOO | -------2--neeneeeeeeeeeeeeeeeeeeeee
oh =
200 =
te po i
200, £
Pai
OO | -------2-2-222-2eeeneeeeeeeee ee goo :
0 PR :
time Response distributions
+r 0 oT saaae oT 1
10" 10? 10%" 107° 10° 10° 10”
Probability
Response distributions
~
Probability
cll
h Response distributions
Probability
hi neon orn
Response distributions
| 2019 | The positive role of noise for information acquisition in biological signaling pathways | 10.1101/762989 | [
"Azpeitia Eugenio",
"Wagner Andreas"
] | creative-commons |
Title: Intravital imaging of real-time endogenous actin dysregulation in proximal and distal tubules
at the onset of severe ischemia-reperfusion injury
Running Title: Intravital imaging of endogenous actin dysregulation
Authors: Peter R. Corridon,1,2,3 Shurooq H. Karam,1 Ali A. Khraibi,1 Anousha A. Khan,1 and
Mohamed A. Alhashmi1
Affiliations:
1Department of Immunology and Physiology, College of Medicine and Health Sciences, Khalifa
University of Science and Technology, PO Box 127788, Abu Dhabi, UAE
2Department of Biomedical Engineering, Healthcare Engineering Innovation Center, Khalifa
University of Science and Technology, PO Box 127788, Abu Dhabi, UAE
3Indiana Center for Biological Microscopy, Indiana University School of Medicine, Indianapolis,
Indiana
Corresponding Author Contact Information:
Peter R. Corridon, PhD
Assistant Professor, Department of Immunology and Physiology
College of Medicine and Health Sciences
Director, Pre-Medicine Bridge Program, College of Arts and Sciences
Khalifa University of Science and Technology, PO Box 127788, Abu Dhabi, UAE
Office email: peter.corridon@ku.ac.ae
Office phone: +971 2 401 8128; Office Fax: +971 2 810 1999
ORCID iD: 0000-0002-6796-4301
Co-author Email Contact Information:
Shurooq H. Karam; email: 100041430@ku.ac.ae
Ali A. Khraibi; email: ali.khraibi@ku.ac.ae
Anousha A. Khan; email: 100045026@ku.ac.ae
Mohamed A. Alhashmi; email: 100053507@ku.ac.ae
Abstract
Severe renal ischemia-reperfusion injury (IRI) can lead to acute and chronic kidney dysfunction.
Cytoskeletal modifications are among the main effects of this condition. The majority of studies
that have contributed to the current understanding of IRI have relied on histological analyses
using exogenous probes after the fact. Here we report the successful real-time visualization of
actin cytoskeletal alterations in live proximal and distal tubules that arise at the onset of severe
IRI. To achieve this, we induced fluorescent actin expression in these segments in rats with
hydrodynamic gene delivery (HGD). Using intravital two-photon microscopy we then tracked and
quantified endogenous actin dysregulation that occurred by subjecting these animals to 60
minutes of bilateral renal ischemia. Rapid (by 1-hour post-reperfusion) and significant (up to 50%)
declines in actin content were observed. The decline in fluorescence within proximal tubules was
significantly greater than that observed in distal tubules. Actin-based fluorescence was not
recovered during the measurement period extending 24 hours post-reperfusion. Such injury
decimated the renal architecture, in particular, actin brush borders, and hampered the
reabsorptive and filtrative capacities of these tubular compartments. Thus, for the first time, we
show that the combination of HGD and intravital microscopy can serve as an experimental tool to
better understand how IRI modifies the cytoskeleton in vivo and provide an extension to current
histopathological techniques.
Keywords: actin cytoskeleton; hydrodynamic gene delivery; ischemia-reperfusion injury; acute
kidney injury and chronic kidney disease; intravital two-photon fluorescence microscopy
INTRODUCTION
Ischemia-reperfusion injury (IRI) is a complex cascade of events that support structural and
functional losses in renal tubular segments. Sudden and temporary restrictions to blood flow
induce oxidative stress and inflammatory responses. These responses adversely affect the
vascular endothelium and tubular epithelium, hampering renal reabsorption and filtration.
Depending on the severity of the IRI, this condition can be reversed to reinstate normal function
or be sustained for the subsequent loss of function [1]. Severe IRI is a common cause of acute
kidney injury (AKI) [2], and produces irreversible damage that supports the progression of AKI to
chronic kidney disease (CKD) and, ultimately, end-stage renal failure [3]. This disease
progression is a growing global health problem with no current specific treatment that has been
the focus of research for several decades [3].
Animal models have been pivotal for investigating the cascade of events that support the
development of irreversible damage from IRI, and have identified the proximal [4] and distal [5]
tubules as major sites of injury with this condition. Severe renal IRI is characterized by losses in
brush border components and polarity in proximal tubule epithelial cells, tubular occlusions, as
well as cytoskeletal dysregulation [6]. Modifications in the cytoskeleton are among the main
effects of IRI. However, to a lesser extent, distal segments succumb to the effects of IRI, and
research has highlighted the formation of casts within the lumen of these tubules as a major
manifestation of the insult. Specifically, it has been shown that cellular blebs aggregate with other
intraluminal materials to form casts, which are either excreted into the urine or stay logged in the
lumen and create substantial obstructions within these tubules [5]. Irrespective of the tubular
segment, these changes occur rapidly and correlate with the severity and duration of insult, and
affect the overall structural and functional integrity of renal tubules [7]. However, the cellular and
subcellular mechanisms responsible for this cascade and resulting tubular damage, are not fully
understood.
Pioneering works that have contributed to the current understanding have relied on histological
analyses after the fact [8]. Furthermore, the ability to conduct these studies in vivo was
traditionally dependent on the use of exogenous probes that provide indirect measures of live
processes [9] and techniques that could only facilitate the imaging of subcellular structures in
confined regions within the kidney [10]. For decades, numerous investigators have outlined the
value of gene delivery to the kidney in attempts to expand the expression of trackable
endogenous probes [11]. Yet, the complex nature of the renal system has provided significant
challenges to progress in this field.
Recent advances in renal gene delivery have provided a much-anticipated option to address this
challenge. By altering hydrodynamic fluid pressures, it has been shown that it is possible to
transiently increase intravascular pressure within peritubular capillary networks and induce
exogenous gene expression in the surrounding tubular epithelium [12]. This method, which is
known as hydrodynamic gene delivery (HGD), is capable of producing widespread genetic
alterations in various segments of the kidney, namely the proximal and distal convoluted tubules,
with minimal effect to the organ [13, 14]. Therefore, to improve the current understanding of renal
IRI, we extended this approach to investigate real-time changes in tubular structures that are
altered in response to IRI, using a sophisticated imaging tool.
In this study, we visualized changes that occur in live proximal and distal tubules by inducing
fluorescent actin protein expression in these segments in rats using HGD, subjected these
animals to 60 minutes of bilateral ischemic injury, and monitored the immediate changes that
occurred within the tubules after blood flow was reinstated to the kidney. Accordingly, we
described and validated the combinative use of HGD and intravital two-photon microscopy as an
experimental tool to track and quantify actin cytoskeletal dysregulation that occurs at the onset of
severe IRI in vivo.
RESULTS
Comparative View of the Actin-Rich Brush Border Ex Vivo Using Exogenous and
Endogenous Probes
Brightfield images were obtained from cortical sections of normal rat kidneys that were
counterstained with hematoxylin and eosin (H&E) (Fig. 1A). These images outlined the intrinsic
actin brush border by the localization of the exogenous eosin fluorophore (pink fluorescence),
which is a hallmark of the proximal tubule that is used to differentiate it from other portions of the
renal tubule in standard histological analyses. Confocal laser scanning micrographs (Fig. 1B) also
revealed the actin-rich brush border using another exogenous fluorophore, Texas-red-labeled
phalloidin (red fluorescence). In comparison, images collected from kidneys that received HGD to
show the expression of an endogenous EGFP-actin (green fluorescence) again highlighted actin
localization along the brush border in cortical sections (Fig. 1C) that correlate well with standard
histological findings.
Figure 1. Brightfield and confocal microscopic images highlight the presence of actin in the renal
brush border ex vivo. Images were taken with 60X objectives (2x digital zoom) using brightfield
(image A) and confocal (images B and C) microscopes. These images of the proximal tubule (PT)
in cortical kidney sections outline the innate actin localization (identified by arrows) along brush
borders using exogenous probes (H&E, image A, and Texas-red phalloidin, image B). Similarly,
image C highlights the intense presence of actin along the brush border in the PT of cortical
kidney sections obtained from rats that expressed EGFP-actin fusion proteins using HGD. Image
B was taken using only the red-pseudo-color channel, and image C was taken using only the
green-pseudo-color channel. Scale bars represent 10 µm.
Endogenous Fluorescent Actin Expression, and Verification of Normal Renal Morphology
and Function Before Injury In Vivo
Fluorescent images were acquired from live kidneys and provided the ability to distinguish
between proximal and distal tubules in vivo based on their relative levels of innate
autofluorescence (Fig. 2A). Furthermore, an enhanced level of contrast was generated by the
expression of EGFP-actin, primarily along the brush border (Fig. 2B through Fig. 2E) and
correlate with the distribution of actin observed using conventional histological techniques (Fig.
1). Such fluorescent protein expression provided an additional way to distinguish between
proximal and distal segments, which are routinely visualized in vivo using this imaging technique
and allowed us to monitor the actin cytoskeleton. Moreover, continuous imaging over a period of
60 minutes did not appear to induce photobleaching (Fig. 3).
EGFP-actin expression also outlined standard morphology that would support innate functions,
such as patent lumens of proximal and distal tubules. The venous introduction of Hoechst 33342
and 150-kDa TRITC-dextran dyes also supported the histological verification of inherent
functional morphology in the rat kidney before IRI (Fig. 2C and 2F). Hoechst 33342 stained the
nuclei of proximal tubular epithelial cells to display their typical appearance, and the high-
molecular-weight dextran molecules were confined to the lumen of peritubular capillaries and
confirmed normal vascular architecture.
Figure 2. HGD allowed the visualization of the actin-rich renal brush border in vivo. Image A
(taken at 2X optical zoom) shows innate autofluorescent patterns that are used to routinely
distinguish the proximal tubule (PT) from the distal tubule (DT) segments but were unable to
outline brush border segments. In comparison, images B and C (taken at 2X optical zoom), as
well as D and E (taken at 1X optical zoom), highlight the presence of the actin-rich brush border
in vivo (identified by arrows) in the proximal tubule. The region outlined in image D (dashed-line)
is presented as image C, to focus on the brush border as we did in Fig. B. Images A and B were
formed by merging the green- and red-pseudo-color channels, while images C and D were
formed by merging the blue-, green- and red-pseudo-color channels. We presented different
combinations of the pseudo-channels shown in image D to create images E and F, to better
highlight EGFP-actin expression in the tubules. Specifically, Image E was created by merging the
green- and blue-pseudo-colors, and image F was created by merging the blue- and red-pseudo-
colors. Overall, the presence of Hoechst 33342 and 150-kDa TRITC-dextran dyes in images C
through F delineated the tubular and supporting vasculature architectures. Scale bars represent
20 µm.
Figure 3. Intravital two-photon micrographs were taken with a X60 objective from a live rat that
received HGD. All images were formed by merging the green-and red-pseudo-color channels and
shows the homogenous distribution of EGFP-actin expression in both proximal and distal tubules
that did not appear to be affected by continuous imaging over a 1-hour period. Scale bar
represents 20 µm.
Impact on In Vivo Tubular Structure and Function with Severe Ischemia-Reperfusion Injury
This form of injury decimated the renal architecture, in particular, the actin brush border, and
hampered the reabsorptive and filtrative capacities of these tubular compartments. After 1 hour of
reperfusion, live imaging provided evidence of alterations to normal tubular structure and function
(Fig. 4B and 4D). The disruptions to tubular EGFP-actin fluorescence, as well as innate
autofluorescence, made it difficult to distinguish proximal from distal tubular segments.
Moreover, after 24 hours of reperfusion, the introduction of fluorescently labeled low-molecular-
weight (4-kDa FITC) and high-molecular-weight (150-kDa TRITC) dextran markers provided
further evidence of distorted renal function (Fig. 5). For instance, the combined presence of the
FITC and TRITC dextrans within the lumen of the tubules outlined that both types of molecules
could have been simultaneously filtered by glomeruli. The combined presence of these dyes
within the lumen highlighted the possible impairment of normal filtrative capacities, as the
molecular-weight should have been confined to the vasculature, and not enter the filtrate.
Imaging of these regions over a subsequent period of 1 hour confirmed the severity of the
induced renal injury (Supplemental Video 1). We also observed aggregated red blood cells
(rouleaux) within vasculature, sluggish blood flow and narrowed peritubular capillaries.
Analogously, there was little evidence to support the entry of low-molecular-weight dextran
molecules within the proximal tubules. This process may indicate the impairment of innate tubular
endocytic capacities and correlate with the dysregulation of the actin cytoskeleton and induced
injury.
Figure 4. Extensive alterations to tubular structure that occurred after one hour of reperfusion.
Intravital two-photon micrographs taken with a 60X objective show the effects of severe IRI in
vivo. Animals that received sham injuries maintained intact tubular structure (images A and C).
Image A was obtained from an animal that did not receive HGD (group 1), while image C (which
displays the actin-rich brush border) was obtained from an animal that received HGD (group 3).
In comparison, we also observed substantial damage to both proximal and distal tubules in
images B and D, which were taken from animals that were subjected to severe IRI (animals in
group 2 did not receive HDG, image B, and animals in group 4 received HGD, image D). Images
C and D can also be found in Fig. 2 (as image F) and Fig. 6 (as image F) respectively. Such
injury dysregulated the actin cytoskeleton, and specifically, stripped the proximal tubule of its
characteristic brush border that was visible in (image C). Scale bars represent 20 µm.
Figure 5. Time-lapse images outline disruptions to normal renal filtrative and endocytic capacities
that resulted from severe IRI. Intravital two-photon micrographs taken with a 60X objective from a
live rat in group 4, which received HGD and was subjected to IRI. After reinstating blood flow to
the kidney, we observed a substantial injury 24 hours after reperfusion. At that time point, we
infused of a mixture of 4-kDa FITC and 150-kDa TRITC dextrans, along with Hoechst 33342, via
the jugular vein of the animal, to track renal dynamics. Images A through I illustrate the loss of
EGFP-actin expression, reductions in the thicknesses of the vasculature (V), aggregated red
blood cells (rouleaux) in the peritubular capillaries (dashed line in image D), absence of endocytic
uptake of low-molecular-weight FITC dextran molecules by the proximal tubules, and
simultaneous entry of both FITC and TRITC dyes in the lumen. Moreover, these images illustrate
the initial presence of the TRITC dye entering the peritubular vasculature (image B) and then the
entry of the FITC dye (image C). After that, there was a reduced level of fluorescence within the
vasculature observed in image I. Scale bar represents 20 µm. The time-lapse video for this event
is presented in Supplemental Video 1.
Real-Time In Vivo Imaging of Actin Dysregulation at the Onset of Reperfusion
We first examined the changes in fluorescence intensity within individual groups across the first
60 minutes of reperfusion in the sham injury and IRI models. There were significant differences in
fluorescence intensity between proximal and distal tubular segments in rats in group 1 (no gene
delivery, sham injury), based on native autofluorescence (p = 0.015), and those in group 4 (gene
delivery, IRI), based on EGFP-actin fluorescence (p=0.023). Whereas analyses performed on
animals in groups 2 (no gene delivery, IRI) and 3 (gene delivery, sham injury) showed that the
analogous reductions in fluorescence were not significant, (p = 0.078) and (p = 0.428),
respectively.
Using data recorded from the two groups of animals that received sham injuries (group 1 and
group 3), the Student’s t-test identified a significant difference (p = 0.006) in the reductions of
proximal tubular fluorescence intensity, but not in the decreases in distal tubular fluorescence
intensity (p = 0.237), between these groups. In comparison, data obtained from animals
subjected to IRI (group 2 and group 4) revealed significant differences between the loss in
fluorescence intensity in proximal tubules in group 2 and those in group 4 (p = 0.004). We also
observed significant differences between the loss in fluorescence intensity in distal tubules in
group 2 and those in group 4 (p = 0.003). Additionally, the decline in actin-based fluorescence
intensity in proximal tubules was significantly greater than that observed in distal tubules among
rats in group 4 (p = 0.027). We also compared data among the four groups and assessed the
effect of injury using the ANOVA test. The observed F value 14.436 is larger than the critical
value of 3.098 and may be interpreted as statistically significant difference among the means of
the groups at the α error level 0.05 (p = 3.029 x 10-5).
Fluorescent actin protein expression allowed us to characterize general, as well as specific actin-
based, alterations that were observed in rat proximal and distal tubules (Fig. 6). Using the data,
we quantified the time-dependent variations in fluorescence intensity (Fig. 7). Imaging was
conducted in a manner previously utilized to limit the occurrence of phototoxicity [15], and was
confirmed in control studies presented in Fig. 3. Within 10 minutes of reperfusion, tubular lumens
were narrowed, and normal actin-rich brush border patterns were replaced by coalesced masses
that blocked the lumen of 30-40% of the tubules that were imaged. EGFP-actin appeared more
heterogeneously distributed and clumped at that time, yet it was still possible to differentiate
between distal and proximal tubules then (Fig. 6A).
As time progressed, fluorescent clumps were dislodged from the tubule into the lumen. This
sloughing process continued, and fluorescent actin-derived structures amalgamated into free-
floating blebs and casts of various sizes within the lumen (Fig. 6B and 6C). Antegrade flow within
the tubules supported the movement of the fluorescent cell/tissue debris through the lumen
(Supplemental Video 2). This normal flow pattern was intermittently replaced by retrograde flow
that accompanied further abnormal tubular narrowing.
By the 20-minute mark (Fig. 6B), there was an intense, approximately 30%, reduction of EGFP-
actin fluorescence in both proximal and distal tubular components. At that time, it was difficult to
find signs of intact brush borders, as the majority of these components were shed into the lumina
leaving behind greater heterogeneity in fluorescent actin localization. Moreover, it was possible to
witness entire groups of cells, within the cuboidal epithelia, dislocate from their tubular linings.
Fluorescent debris was restricted to the lumen, as there were no signs of EGFP-actin in regions
that corresponded to the neighboring vasculature. These changes supported the development of
ghost tubules (tubular segments devoid of living cells, previously identified by Hall et al. using
intravital multiphoton microscopy [16]), and, in some instances, we observed drastic
improvements in the patency of the tubular lumen, as fluorescent debris was seen to be
transported swiftly and bidirectionally within the lumen (Supplemental Video 2).
Furthermore, after 40-minutes of reperfusion (Fig. 6D), we observed the decimation of 20-30% of
all imaged tubular segments, and it became difficult to locate and even identified the lineage of
various tubules after 50 minutes of IRI (Fig. 6B). The progressive loss of fluorescence continued,
and resulted in substantial declines in EGFP-actin fluorescence within the first hour of reperfusion
(Fig. 6A through 6F). Finally, we estimated as much as 60% reductions in EGFP-actin
fluorescence occurred in both proximal and distal segments after 60 minutes reperfusion.
Interestingly, actin-based fluorescence was not recovered during our measurement period that
extended to roughly 24 hours after reperfusion.
Figure 6. Time-lapse images tracked alterations in actin-based fluorescence observed during the
first 60 minutes of reperfusion. Intravital two-photon micrographs taken with a 60X objective from
a live rat in group 4 across 60 minutes (this is the same imaging field that is previously presented
in Fig. 3D). This animal received HGD and was subjected to ischemia-reperfusion injury (IRI). All
images were formed by merging the green-and red-pseudo-color channels and shows the
expression of EGFP actin in both proximal and distal tubules. This fluorescent protein expression
allowed us to visualize the live and real-time changes in tubular structure and function that
resulted from IRI (arrows identified changes in proximal tubules, and arrowheads identified
changes in distal tubules). At the 10-minute mark, EGFP-actin appeared more heterogeneously
distributed and clumped in tubular segments. The dashed ovals in images B and E track the
outlined region and show how cells have sloughed off the proximal tubule segment and migrated
into the lumen to generate ghost tubules (tubules mostly devoid of living cells) by the 50-minute
mark. It should be noted that there were minor shifts in the field during the 60-minute imaging
period that resulted from the vibration caused by respiration. Scale bar represents 20 µm. A time-
lapse video showing portions of this event is presented in Supplemental Video 2.
Figure 7. In vivo changes in mean fluorescence intensities obtained from proximal and distal
tubular segments. There were no considerable differences in fluorescence intensity recorded
from proximal and distal tubular segments from animals that did not receive HGD (group 1), but
there were larger variations in autofluorescence that resulted from ischemia-reperfusion injury
(group 2) across the 60-minute measurement period. In comparison, we observed substantial
decreases in fluorescence intensities in proximal and distal tubular segments recorded from
animals that received HGD (groups 3 and 4).
DISCUSSION
Fluorescent probes and animal models have been used extensively to investigate mechanisms
that generate irreversible damage from IRI. Gaining a better understanding of the disease
etiology can help devise novel strategies to prevent the progression of AKI to CKD and,
ultimately, kidney failure. Traditional light and electron microscopy have provided significant
insight into the cascade of events that occur with such pathologies [12]. Yet, consolidated and
unified descriptions of the associated cellular and sub-cellular mechanisms are needed. A major
technical drawback that has limited progress in this area lies in the ability to perform these
investigations in vivo.
Recent advances in imaging technologies and genetic engineering have provided a means to
perform such studies in real-time. Powerful imaging tools, like intravital two-photon microscopy,
have contributed to the present understanding of the functional morphology in the live kidney,
deviations that occur with damage, and ways to better manage these conditions [16]. Likewise,
techniques like HGD can help change the genetic makeup of cells within the kidney, and thus
offer a newfound way to examine in vivo processes using endogenous markers [17]. As a result,
in this study, for the first time we show that the combination of intravital two-photon microscopy
and HGD can be used to visualize and measure the rate of degradation of the actin cytoskeleton
at location known to be targeted by IRI.
Our ex vivo findings illustrate that HGD can facilitate the expression of genetically altered forms
of actin within proximal and distal tubules. The fluorescence patterns collected from cortical
kidney sections, using confocal microscopy, confirm the intrinsic localization of EGFP-actin fusion
proteins, particularly along the brush border. This is an important histological characteristic that
has been relied on for decades [8], and that has been previously recorded ex vivo using
micropuncture gene delivery [9].
In comparison, for in vivo imaging studies, the early proximal and distal tubules in rodents can be
easily accessed for live imaging using intravital two-photon microscopy, and routine distinctions
are made between these tubules based on their relative levels of autofluorescence [18].
Furthermore, the fluorescent images acquired from live kidneys highlighted the different renal
structural patterns observed in vivo and ex vivo as previously reported, while confirming the
enhanced presence of actin along the brush border [19]. Proximal segments have higher
autofluorescent signatures than their distal counterparts. However, it is difficult to differentiate
between the individual segments of the proximal convoluted tubule based only on innate tissue
autofluorescence. We thus sought to determine whether HGD could provide a better means for
tubular segment differentiation. We observed that hydrodynamic-based renal gene transfer
facilitated the endogenous expression of actin fusion proteins within the renal tubules, and thus
enhanced the contrast between distal segments and proximal tubules.
Overall, EGFP-actin expression helped outline normal renal morphology and function along with
the nuclear and vascular probes in live rodent kidneys. To further underscore the utility of the
model, live EGFP-actin expression was visualized within and along the brush border of proximal
tubular epithelial cells at various levels, and thus the varied degrees of actin localization can
provide a means to distinguish between S1 and S2 segments of the proximal tubule, based on
the relative thicknesses of the actin brush border, similar to pioneering studies conducted that
provided this distinction in cortical segments [8]. Our studies used plasmid transgene vectors that
expressed both fluorescent filamentous (F-actin) and monomeric globular (G-actin) proteins, and
thus additional segment-specific markers will be needed to support this claim. In the future, we
can also consider the use of plasmid vectors that support the fluorescent expression of only F-
actin fusion proteins [20] to refine the way actin cytoskeletal components can be tracked in vivo.
However, these investigations may be limited by the resolution of the imaging system.
Also, in some instances there was a higher fluorescent signal from non-filamentous actin, which
is consistent with previous research conducted in cell culture [20]. These ex vivo studies have
also determined that the expression of eGFP-actin can affect cell behaviour. Fortunately, this
concern as the plasmid titer and period of expression were previously shown to support the
stable expression of fluorescent/exogenous proteins that did not significantly alter cellular
function in vivo [13, 14]. It was suitable to consider this type of expression vector for our initial
studies, as both forms of actin are essential cytoskeletal components.
Once the focus was shifted to investigating the impact of injury, it was evident that this severe
form of IRI had a devastating effect on tubular structure and function. Bilateral renal ischemia for
the 60-minute period would have supported severe and sustained reductions in blood pressure to
induce tubular necrosis [21]. Pathological changes that occur with this condition include reduced
filtrative capacities of the tubules that results from hypoperfusion. Cellular debris and casts also
amalgamate to obstruct the lumina and hinder the movement of the filtrate through the nephron.
The damage to the tubular epithelium that stems from ischemia has conventionally been
considered as a consequence of cellular necrosis [22], and these irreversible effects, which were
visualized in real-time within the first 60 minutes of our reperfusion study, demonstrate the proof
of concept. However, there is growing evidence to suggest that apoptosis has a significant
contribution to the acute injury. Reductions in renal apoptosis antagonizing transcription factor
have been shown to result from IRI, hampering intrinsic activation of antiapoptotic pathways
and/or inhibition of proapoptotic pathways [23]. Further investigations that can combine this
imaging technique and gene delivery may be used to identify the potential therapeutic application
of this transcription factor in IRI, and potentially extend the value of the presented model.
Meanwhile, it is well known that the proximal tubules rely mainly on mitochondrial metabolism for
ATP synthesis due to their limited glycolytic capacities and are thus particularly susceptible to IRI.
The rapid and significant recorded reductions in actin-based fluorescence allowed us to visualize
such proximal tubular damage, which included brush border losses that would have resulted from
profound decreases in intracellular ATP. Drops in ATP levels would have occurred early after
onset of ischemia and driven actin cytoskeletal derangements that favor the non-filamentous form
of actin, as the cytoskeleton requires ATP to remain in a filamentous form [16].
Cytoskeletal dysregulation would have, in turn, led to the redistribution of integrins and Na+-K+-
ATPase from the basal membrane [24]. This process would have resulted in impaired cellular
transport mechanisms that ultimately support cellular death and sloughing from tubular basement
membranes. Comparatively, the distal tubule epithelium would have succumbed to less damage
based on their relatively lower dependence on mitochondrial metabolism but would have been
drastically impacted by intraluminal obstructions generated from damage to proximal tubular
segments [5, 21]. Paradoxically, reinstating blood flow would have supported additional tissue
damage. Emerging evidence suggests that the mitochondrial production of reactive oxygen
species, which occurs during the reperfusion phase of IRI, has a critical role in destroying cellular
components, as well as initiating apoptosis and necroptosis [25]. Monitoring this process thus
allowed us to quantify the cumulative damage by estimating the reductions in EGFP-actin
fluorescence that occurred within the first hour of reperfusion. Furthermore, the ability to examine
live events with this novel approach, like antegrade flow patterns that occurred within the tubular
lumen, as well as the dynamic changes in tubular diameter extend beyond the limits of
established histopathological techniques.
In summary, due to the complex nature of the kidney, in vivo studies have relied on exogenous
probes to investigate the underlying nature of renal tubular morphological and functional
processes [26, 27]. To extend this approach, this study demonstrates the utility of the
combinative use of HGD and intravital two-photon microscopy to track the dynamic remodeling of
the actin cytoskeleton. Importantly, this method signifies a way to monitor intrinsic cellular and
molecular mechanisms involved in the generation of irreversible kidney injury that results from
IRI. Future studies can be employed to compare the levels of alterations in and potential recovery
of actin content as a function of injury severity, tubular complexity and renal function.
Furthermore, such studies can reinforce the combined use of HGD and intravital imaging. This
combination may provide a powerful tool to examine therapeutic targets that can limit the
progression of renal injuries associated with IRI [28].
METHODS
Fluorescent Plasmids and Dyes
Plasmid DNA encoding enhanced green fluorescent (EGFP)-actin (Takara Bio USA, Mountain
View, CA) facilitated exogenous gene expression in rodent kidneys. The following dyes were
used for intravital two-photon imaging and bolus injected intravenously in a volume of 0.5 ml: 50
µl of 150-kDa tetramethyl rhodamine isothiocyanate (TRITC) and/or 4-kDa fluorescein
isothiocyanate (FITC) dextrans (TdB Consultancy, Uppsala, Sweden) and 30-50 µl of Hoechst
33342 (Invitrogen, Carlsbad, CA). Texas red-phalloidin (Invitrogen Corporation, Mountain View,
CA) was used for ex vivo actin staining.
Hydrodynamic Gene Delivery
All experiments were perfromed on 200 to 400 g male Sprague-Dawley rats (Harlan Laboratories,
Indianapolis, IN). The experimental were approved by the Indiana University School of Medicine
Institutional Animal Care and Use Committee, and Animal Research Oversight Committee at
Khalifa University of Science and Technology, and the study was carried out in compliance with
the ARRIVE guidelines. Animals were anesthetized with inhaled isoflurane (5% in oxygen,
Webster Veterinary Supply, Devens, MA) and then given intraperitoneal injections of 50 mg/kg of
pentobarbital (Hospira, Inc., Lake Forest, IL). The details of the HGD process are outlined in the
literature [13, 14]. Briefly, for this process, 1-3 µg of EGFP-actin plasmid DNA, per gram of body
weight were suspended in 0.5 ml of saline for retrograde renal vein injections. Animals were
allowed 14 days to recover before further experimentation.
Brightfield Imaging
Kidneys were fixed with 4% paraformaldehyde for 24 hours at 4°C, and immersed in 4%
phosphate-buffered formalin, again for a minimum of 24 hours at room temperature. Specimens
were rinsed in distilled H2O and stored in 70% ethanol. Specimens were dehydrated through a
graded series of ethanol (70%; 80%, 95%, 100%), cleared in xylene, infiltrated with 4 changes of
paraffin (under vacuum at 59°C; 45 minutes each), and embedded in fresh paraffin. After which,
4-5 μm thick sections were collected with a Reichert-Jung 820 microtome (Depew, NY), flattened
on a warm water bath and mounted on glass slides, and stained with H&E. A Nikon Microphot SA
Upright Microscope with a 60X objective and sensitive Diagnostic Instruments SPOT RT Slider
color camera (Nikon, Tokyo, Japan) used to collect images.
Confocal Imaging
Whole kidneys were harvested from control rats and those that received HGD. Kidneys were
immersion fixed with 4% paraformaldehyde, 100-200 μm thick cortical sections were obtained,
and incubated overnight in a phalloidin staining solution. This solution was prepared by diluting
Texas-red-phalloidin in a blocking buffer (2% bovine serum albumin and 0.1% Triton X-100,
diluted in phosphate-buffered saline) at a ratio of 1:200 for roughly 24 hours. The tissues then
were rinsed three times for two hours in PBS and mounted onto slides. Images were collected
with a 60X objective.
Intravital Two-Photon Imaging
While under sedation, vertical flank incisions were made to externalize left kidneys for imaging.10
In some cases, the internal jugular vein was cannulated for intravenous infusions of dyes. Body
temperature was controlled, as exteriorized kidneys were positioned inside a glass-bottom dish
containing saline, which was set above a 60X water-immersion objective. Fluorescent
micrographs were collected using an Olympus (Center Valley, PA) FV 1000-MPE Microscope
equipped with a Spectra-Physics (Santa Clara, CA) MaiTai Deep See laser, with dispersion
compensation for two-photon microscopy, tuned to 770-860 nm excitation wavelengths. The
system was mounted on an Olympus IX81 inverted microscope, was also equipped with dichroic
mirrors to collect blue, green, and red emissions and two external detectors for two-photon
imaging.
Bilateral Renal Ischemia-Reperfusion Injury
Two weeks after recovering from the HGD, transfected animals, along with others that did not
receive gene transfer, were separated into four groups (n=3 for all groups). Groups 1 and 2 did
not receive gene transfer, while groups 3 and 4 received HGD. All animals were anesthetized for
median laparotomies that allowed blunt dissection of renal pedicles. For rats in groups 2 and 4,
non-traumatic vascular clamps were applied to bilateral renal pedicles simultaneously for
60 minutes. After the clamps were removed, reperfusion was confirmed visually. Whereas rats in
groups 1 and 3 received sham injuries. Midline incisions were closed, and the animals were
prepared for intravital imaging.
Investigation of Changes in Fluorescence and Tubular Function
Two-photon fluorescent micrographs were collected to analyze immediate structural and
functional changes in live kidneys directly after reperfusion. For morphological changes, we
estimated relative variations in autofluorescence or EGFP-actin fluorescence in proximal and
distal tubular segments during the first 60 minutes of reperfusion. Four equal and adjacent
regions were randomly chosen on proximal and distal tubular segments to record changes in
mean fluorescence intensities at 10-minute intervals. The data was averaged across each group
to track time-based losses in actin fluorescence post reperfusion. Changes in tubular
reabsorptive and filtrative capacities were also analyzed at the 60-minute mark using fluorescent
dextrans [10].
Statistical Analysis of Data
Statistical data are presented as the mean ± SE. Differences in fluorescence intensities were
investigated among study groups using one-way analysis of variance (ANOVA) and Student t-
tests were applied with p < 0.05 level of significance as appropriate.
ACKNOWLEDGMENTS
The authors would like to acknowledge George J. Rhodes, MD (Indiana University) for
contributions and guidance in developing the live injury model.
GRANTS
This publication is based upon work supported by the Khalifa University of Science and
Technology under Award No. RC2-2018-022 (HEIC) and Research Fund FSU-2020-25 granted
to P.R.C. Support for this study was also provided by NIH P-30 O’Brien Center (DK 079312)
DISCLOSURES
No conflicts of interest, financial or otherwise, are declared by the authors.
AUTHOR CONTRIBUTIONS
Author contributions: P.R.C. conceived and designed research; P.R.C. performed experiments;
P.R.C. and S.H.K. analyzed data; P.R.C. interpreted results of experiments; P.R.C., S.H.K.,
A.A.K., A.A.K., and M.A.A. prepared figures; P.R.C., S.H.K., A.A.K., A.A.K., and M.A.A. drafted
the manuscript; P.R.C., S.H.K, A.A.K., A.A.K., and M.A.A. edited and revised manuscript; and
P.R.C., S.H.K., A.A.K., A.A.K., and M.A.A. approved final version of manuscript.
SUPPLEMENTARY MATERIAL
Video 1. Renal tubular filtrative and endocytic capacities impaired by severe ischemia-reperfusion
injury. Supplemental Video 1 available at URL: https://figshare.com/s/9b9624b41b2ef7c0c751
DOI: 10.6084/m9.figshare.13615889
Video 2. Actin dysregulation at the onset of severe ischemia-reperfusion injury.
Supplemental Video 2 available at URL: https://figshare.com/s/02d7148b26446224c0e3
DOI: 10.6084/m9.figshare.14130146
REFERENCES
1.
Situmorang, G.R. and N.S. Sheerin, Ischaemia reperfusion injury: mechanisms of
progression to chronic graft dysfunction. Pediatr Nephrol, 2019. 34(6): p. 951-963.
2.
Wang, X., et al., Effect of apigenin on apoptosis induced by renal ischemia/reperfusion
injury in vivo and in vitro. Ren Fail, 2018. 40(1): p. 498-505.
3.
Devarajan, P. and J.L. Jefferies, Progression of Chronic Kidney Disease after Acute
Kidney Injury. Prog Pediatr Cardiol, 2016. 41: p. 33-40.
4.
Chevalier, R.L., The proximal tubule is the primary target of injury and progression of
kidney disease: role of the glomerulotubular junction. Am J Physiol Renal Physiol, 2016.
311(1): p. F145-61.
5.
Molitoris, B.A., Actin cytoskeleton in ischemic acute renal failure. Kidney Int, 2004. 66(2):
p. 871-83.
6.
Wang, X., et al., Kidney protection against ischemia/reperfusion injury by
myofibrillogenesis regulator-1. Am J Nephrol, 2014. 39(4): p. 279-87.
7.
Kalogeris, T., et al., Cell biology of ischemia/reperfusion injury. Int Rev Cell Mol Biol,
2012. 298: p. 229-317.
8.
Venkatachalam, M.A., et al., Ischemic damage and repair in the rat proximal tubule:
differences among the S1, S2, and S3 segments. Kidney Int, 1978. 14(1): p. 31-49.
9.
Tanner, G.A., et al., Micropuncture gene delivery and intravital two-photon visualization of
protein expression in rat kidney. Am J Physiol Renal Physiol, 2005. 289(3): p. F638-43.
10.
Ashworth, S.L., et al., Two-photon microscopy: visualization of kidney dynamics. Kidney
Int, 2007. 72(4): p. 416-21.
11.
Imai, E. and Y. Isaka, Strategies of gene transfer to the kidney. Kidney Int, 1998. 53(2): p.
264-72.
12.
Collett, J.A., et al., Hydrodynamic isotonic fluid delivery ameliorates moderate-to-severe
ischemia-reperfusion injury in rat kidneys. Journal of the American Society of Nephrology,
2017. 28(7): p. 2081-2092.
13.
Corridon, P.R., et al., A method to facilitate and monitor expression of exogenous genes
in the rat kidney using plasmid and viral vectors. American Journal of Physiology-Renal
Physiology, 2013. 304(9): p. F1217-F1229.
14.
Kolb, A.L., et al., Exogenous gene transmission of isocitrate dehydrogenase 2 mimics
ischemic preconditioning protection. Journal of the American Society of Nephrology, 2018.
29(4): p. 1154-1164.
15.
Hall, A.M., et al., In vivo multiphoton imaging of mitochondrial structure and function
during acute kidney injury. Kidney international, 2013. 83(1): p. 72-83.
16.
Hall, A.M. and B.A. Molitoris, Dynamic multiphoton microscopy: focusing light on acute
kidney injury. Physiology (Bethesda), 2014. 29(5): p. 334-42.
17.
Woodard, L.E., et al., Kidney-specific transposon-mediated gene transfer in vivo. Sci Rep,
2017. 7: p. 44904.
18.
Dunn, K.W., et al., Functional studies of the kidney of living animals using multicolor two-
photon microscopy. Am J Physiol Cell Physiol, 2002. 283(3): p. C905-16.
19.
Corridon, P.R., et al., A method to facilitate and monitor expression of exogenous genes
in the rat kidney using plasmid and viral vectors. Am J Physiol Renal Physiol, 2013.
304(9): p. F1217-29.
20.
Lopata, A., et al., Affimer proteins for F-actin: novel affinity reagents that label F-actin in
live and fixed cells. Sci Rep, 2018. 8(1): p. 6572.
21.
Holderied, A., et al., "Point of no return" in unilateral renal ischemia reperfusion injury in
mice. J Biomed Sci, 2020. 27(1): p. 34.
22.
Kaushal, G.P., A.G. Basnakian, and S.V. Shah, Apoptotic pathways in ischemic acute
renal failure. Kidney Int, 2004. 66(2): p. 500-6.
23.
Xie, J. and Q. Guo, Apoptosis antagonizing transcription factor protects renal tubule cells
against oxidative damage and apoptosis induced by ischemia-reperfusion. J Am Soc
Nephrol, 2006. 17(12): p. 3336-46.
24.
Ratliff, B.B., et al., Oxidant Mechanisms in Renal Injury and Disease. Antioxid Redox
Signal, 2016. 25(3): p. 119-46.
25.
Kalogeris, T., Y. Bao, and R.J. Korthuis, Mitochondrial reactive oxygen species: a double
edged sword in ischemia/reperfusion vs preconditioning. Redox Biol, 2014. 2: p. 702-14.
26.
Schiessl, I.M., et al., Just Look! Intravital Microscopy as the Best Means to Study Kidney
Cell Death Dynamics. Semin Nephrol, 2016. 36(3): p. 220-36.
27.
Schiessl, I.M., et al., Long-Term Cell Fate Tracking of Individual Renal Cells Using Serial
Intravital Microscopy. Methods Mol Biol, 2020. 2150: p. 25-44.
28.
Corridon, P.R., et al., Bioartificial kidneys. Current Stem Cell Reports, 2017. 3(2): p. 68-
76.
| 2021 | Intravital imaging of real-time endogenous actin dysregulation in proximal and distal tubules at the onset of severe ischemia-reperfusion injury | 10.1101/2021.03.01.433337 | [
"Corridon Peter R.",
"Karam Shurooq H.",
"Khraibi Ali A.",
"Khan Anousha A.",
"Alhashmi Mohamed A."
] | creative-commons |
1
Influence of prior beliefs on perception in early psychosis:
effects of illness stage and hierarchical level of belief
J. Haarsma1, F. Knolle1, J.D. Griffin1, H. Taverne1, M. Mada2, I.M. Goodyer1, the NSPN
Consortium, P.C Fletcher1,3,4, G.K. Murray1,4
1Department of Psychiatry, University of Cambridge, United Kingdom,
2Cognition and Brain Sciences Unit, University of Cambridge, United Kingdom
3Wellcome Trust MRC Institute of Metabolic Science, Cambridge Biomedical Campus,
United Kingdom,
4Cambridgeshire and Peterborough NHS Foundation Trust, United Kingdom
Correspondence to:
Dr Murray Department of Psychiatry, University of Cambridge, United Kingdom
(gm285@cam.ac.uk)
Disclosures:
P.C.F. has received payments in the past for ad hoc consultancy services to
GlaxoSmithKline All other authors declare no competing interests.
Keywords:
Perceptual priors, Cognitive priors Psychosis, Glutamate, ARMS, MRS
Funding:
This work was supported by the Neuroscience in Psychiatry Network, a strategic
award from the Wellcome Trust to the University of Cambridge and University
College London (095844/Z/11/Z), Wellcome Trust (093270) Bernard Wolfe Health
Neuroscience Fund (P.C.F.), and the Cambridge NIHR Biomedical Research Centre.
2
General scientific summary:
What we perceive and believe on any given moment will allow us to form
expectations about what we will experience in the next. In psychosis, it is believed
that the influence of these so-called perceptual and cognitive ‘prior’ expectations on
perception is altered, thereby giving rise to the symptoms seen in psychosis.
However, research thus far has found mixed evidence, some suggesting an increase
in the influence of priors and some finding a decrease. Here we test the hypothesis
that perceptual and cognitive priors are differentially affected in individuals at-risk
for psychosis and individuals with a first episode of psychosis, thereby partially
explaining the mixed findings in the literature. We indeed found evidence in favour
of this hypothesis, finding weaker perceptual priors in individuals at-risk, but
stronger cognitive priors in individuals with first episode psychosis.
3
Abstract
Alterations in the balance between prior expectations and sensory evidence may
account for faulty perceptions and inferences leading to psychosis. However,
uncertainties remain about the nature of altered prior expectations and the degree
to which they vary with the emergence of psychosis. We explored how expectations
arising at two different levels – cognitive and perceptual – influenced processing of
sensory information and whether relative influences of higher and lower level priors
differed across people with prodromal symptoms and those with psychotic illness. In
two complementary auditory perception experiments, 91 participants (30 with first
episode psychosis, 29 at clinical risk for psychosis, and 32 controls) were required to
decipher a phoneme within ambiguous auditory input. Expectations were generated
in two ways: an accompanying visual input of lip movements observed during
auditory presentation, or through written presentation of a phoneme provided prior
to auditory presentation. We determined how these different types of information
shaped auditory perceptual experience, how this was altered across the prodromal
and established phases of psychosis, and how this relates to cingulate glutamate
levels assessed by magnetic resonance spectroscopy. The psychosis group relied
more on high level cognitive priors compared to both healthy controls and those at
clinical risk for psychosis, and more on low level perceptual priors than the clinical
risk group. The risk group were marginally less reliant on low level perceptual priors
than controls. The results are consistent with previous theory that influences of prior
expectations in psychosis in perception differ according to level of prior and illness
phase.
4
1.1.Background
It has been hypothesized that the brain forms a model of the world by actively trying
to predict it and to update these predictions iteratively by function of the prediction
error, a hierarchical computational framework usually referred to as predictive
coding (Rao & Ballard et al., 1999; Bar, 2009; Friston, 2005 & 2009; Bastos et al.,
2012; Clark et al., 2013 & 2015; Hohwy et al., 2013; Knill et al., 2004). In this
framework, the formation of delusional beliefs and hallucinatory experiences are
proposed to be due to alterations in the cognitive and biological mechanisms of
predictive coding (Fletcher & Frith, 2009; Adams et al., 2013).
Whilst initial clinical studies documenting alterations in the way the expectation
influences perception in psychosis are promising in demonstrating case-control
alterations in various behavioural measures of predictive coding (eg Shergill et al
2005, Teufel et al., 2010; Powers et al 2017), it is already clear that there will be no
straightforward unifying explanation of psychosis in simple terms of priors being
“too strong” or “too weak” in general. Predictive processing theory envisions a
highly interlinked (cortical) cognitive hierarchy, where different layers aim to predict
the incoming input from lower-layers (Rao & Ballard et al., 1999; Bar et al., 2009;
Friston, 2005 & 2009; Bastos et al., 2012; Clark et al., 2013 & 2015; Hohwy et al.,
2013; Knill et al., 2004). Moving up the hierarchy, the predictions become more
abstract, ranging from lower-level sensory prediction to higher-order beliefs about
the environment. It therefore does not suffice to ask the question whether prior
expectations are stronger or weaker in psychosis. Instead in order to form a
complete picture of the underlying mechanisms of psychosis, we need to look at the
contribution of different types of prior expectations, including both sensory
expectations and higher-level beliefs about the environment.
Recent influential predictive coding accounts of psychosis have emphasized that
priors at low and high hierarchical levels may be differentially affected in psychotic
illness. For example, Sterzer et al (2018) conclude that “In contrast to weak low-level
priors, the effects of more abstract high-level priors may be abnormally strong” in
5
psychosis. This postulate is mainly drawn through a combination of theoretical
arguments and synthesis across diverse studies. To our knowledge no single study
has yet demonstrated a combination of weak low-level perceptual priors and strong
high-level cognitive priors in patients with psychosis, although Schmack (2013) and
colleagues provided supportive evidence in a study of individual differences in
healthy individuals. Those authors delineated priors at different hierarchical levels by
manipulating what they referred to as perceptual priors and cognitive priors in two
related experiments; they found that delusional ideation in health (sometimes
termed delusion proneness) was associated with a decrease in the contribution of
perceptual priors, and an increase in the contribution of cognitive priors, highlighting
the importance to separate the two (Schmack et al., 2013). Clearly, clinical studies
are required testing the hypothesis of simultaneous weak low-level and strong high-
level priors in psychotic illness, yet few have been attempted. One exception was
another study from Schmack and colleagues, who found evidence against differential
strengths of sensory and cognitive priors in schizophrenia (Schmack et al 2017).
A further complexity is that cognitive and biological mechanisms of psychosis may be
markedly different at different illness stages, adding nuance to the attractive, yet
arguably overly simplistic, continuum model of psychosis. Previous reviews
acknowledge that there may be evolving patterns of cognitive and/or physiological
disturbances over time as psychotic illness develops (Fletcher & Frith, 2009; Adams
et al., 2013; Heinz et al., 2018). In many cases psychotic illness is heralded by the
development of delusions (often delusional interpretations of hallucinations) after a
prodromal period of hallucinatory experiences without delusional interpretation
and/or delusional mood. In the context of weak low level (sensory) priors and high
precision of sensory prediction errors, delusions may emerge as result of
compensatory increases in the precision of high-level beliefs (i.e. enhanced high
level, cognitive priors) (Adams et al 2013, Sterzer et al 2018, Heinz et al 2018). It
follows then that in the very early phases of psychosis, prior to the development of
delusions, such compensatory increases in the precision of high level beliefs may be
yet to emerge. Although one previous study found alterations in the utilisation of
priors in individuals at clinical risk for psychosis (putatively in the prodrome)
6
compared to controls (Teufel et al 2015), this study did not include any patients with
established psychotic illness, and thus none of the sample had developed delusions
at the time of the experiment. It thus remains unclear whether, or how, alterations
in the use of higher or lower level priors changes as psychotic illness emerges.
We acknowledge the vital importance of the range of previous studies exploring the
contribution of prior expectation in perception in psychosis. However, here we argue
that two important aspects of the predictive coding account have been largely
neglected in empirical clinical studies: the contribution of different disease stages to
the effect of prior expectations, and the type of prior expectation. It is the aim of the
present study to bring these two together, by studying how different prior
expectations are affected throughout individuals at risk for psychosis and individuals
who recently had an episode of psychosis.
In order to test the hypothesis that sensory and cognitive priors are differently used
depending on the stage of psychosis, we designed two novel auditory perception
paradigms, one testing the influence of lip-movements on auditory perception
(perceptual priors) and a second testing the influence of learned written-word-sound
associations on auditory perception (cognitive priors); and we gathered data on
these two paradigms in two patient groups – individuals at elevated clinical risk for
psychosis, and individuals who recently had their first episode of psychosis, and
compared them to a group of healthy controls. Help-seeking individuals who are at-
risk for psychosis usually have sub-clinical psychotic symptoms that are not severe or
frequent enough to warrant a clinical diagnosis, but are at considerably raised risk of
developing a psychotic illness in the short to medium term (Yung et al., 2003).
Studying these early stages of illness may help us to understand the mechanisms
underlying the emergence of a psychosis by examining which aberrancies precede
psychosis and might therefore be predictive of developing psychosis.
The first paradigm (from now on ‘perceptual priors task’) assesses the influence of
lip-movements on auditory perception. Lip-movements have been shown to
influence auditory perception. McGurk and MacDonald (1976) showed that when
7
individuals where presented with an auditory /Ba phoneme in combination with lip-
movements pronouncing /Ga, most individuals perceive a mixture between the two,
i.e. /Da. This effect has become known as the McGurk illusion (McGurk &
MacDonald, 1976). Studies of the neural mechanisms underlying the influence of lip-
movements on auditory perception provide support for the Bayesian framework, in
that lip-movements are suggested to constitute a prior expectation with respect to
the incoming auditory signal (Arnal et al., 2012; Blank & Davis, 2016). One previous
study of mainly male, middle-aged adults with chronic schizophrenia documented a
diminishment in perceiving the McGurk illusion, relying more on the auditory input;
the finding that was associated with illness chronicity (White et al., 2014). Pearl et al
(2009) also studied the McGurk illusion in schizophrenia, finding mixed results:
adolescents with schizophrenia, but not adults with schizophrenia, showed a
diminished illusory effect. Schizophrenia has been associated with a diminished
ability in using lip-movements in aiding auditory discrimination, suggesting
aberrancy in the ability to integrate the two sources of information (Myslobodsky et
al., 1992; de Gelder et al., 2002; Ross et al., 2007; Szycik et al., 2013). However, it
remains unclear whether the influence of prior information in auditory perception is
altered in the early stages of psychosis, as no previous first episode psychosis study
or study of people with prodromal symptoms of psychosis has been conducted. The
purpose of the perceptual priors task was to measure precisely how much lip-
movements influence what participants hear by using a staircase procedure
(Cornsweet, 1962), in which the balance between two sounds was changed in
predefined steps, providing a more fine-grained measures of individual susceptibility
to the illusion than in previous clinical studies.
The second paradigm (from now on ‘cognitive priors task’), assesses the influence of
learned written-word-sound associations on auditory perception. The impact of
learned associations on auditory perception has been shown in sensory conditioning,
where one stimulus functions as a predictor for an auditory stimulus that is
otherwise difficult to detect. In these early experiments, participants were asked to
identify auditory stimuli on the basis of a visual cue. Sometimes the participants
reported perceiving an auditory stimulus when only presented with the visual cue, as
8
the brain predicted an auditory stimulus on the basis of the cue (Ellson, 1941; Kot et
al., 2002; Warburton et al., 1985; Agathon et al., 1973; Brogden et al., 1947; Powers
et al., 2017). Previous research found that this omission effect is stronger in
individuals with hallucinations (Kot et al., 2002; Powers et al., 2017), suggesting an
increase in the influence of learned ‘cognitive’ expectation on auditory perception in
psychosis, in contrast to the diminishment in the influence of ‘sensory’ expectations
in schizophrenia discussed above. However, up to date, no study has explored the
influence of learned cognitive expectations in individuals at-risk for psychosis and
compared it to the influence of sensory expectations on perception.
We recognize that the sensory and cognitive priors tasks are strictly speaking not
able to estimate the relative precision and mean of the prior expectations and
sensory evidence for each participant directly. Instead we make the assumption
based on Bayesian theories of the brain that perception is a function of the precision
and mean of the prior and sensory evidence. Therefore rather then estimating the
precision and mean for the prior and sensory evidence separately, we infer the
relative contribution of prior information and sensory evidence, and term this for the
remainder of this paper the relative strength of the sensory and cognitive prior.
Reconciling the exact level of priors used in the current experiment in relation to the
exact level of priors used in previous experiments in schizophrenia spectrum patients
is not trivial. However, this is not central to our experiment. Our aim is to examine
the effects of two different levels of priors on a given process at different stages of
psychosis.
Another issue currently understudied relates to the neurobiological underpinnings of
alterations in the contribution of prior expectations in perception. Changes in
glutamate levels have been associated with schizophrenia (Marsman et al., 2011;
Merritt et al., 2016; Treen et al., 2016), including in the cingulate cortex, where there
is evidence of excessive glutamate in early illness stages, possibly progressing to
reductions in later stages (Merritt et al 2016, Kumar et al 2018). It remains unclear to
what extent glutamate levels in the brain relate to predictive coding mechanisms
putatively mediating psychosis, in spite of various theoretical arguments and
9
extrapolations from preclinical experiments (Corlett et al., 2011; Sterzer et al 2018).
Notably the anterior cingulate cortex (ACC) has been associated with processing
uncertainty (Rushworth et al., 2008) and precision-weighting of information in
health and psychosis (Cassidy et al., 2017; Katthagen, et al., 2018; Haarsma et al.,
2019). Thus alterations in glutamate levels in the ACC might alter the precision of
prior information, thereby changing the degree to which priors influence perception.
We therefore explored this issue by measuring magnetic resonance spectroscopy
(MRS) glutamate levels in the anterior cingulate cortex and relating these
measurements to the contribution of prior expectations in the different
experimental groups. Our study is not powered to provide definitive results relating
glutamate measures to our predictive coding measures, the latter being of primary
interest here. Nevertheless, we report preliminary, exploratory analyses that may be
hypothesis generating and could provide the basis for power calculations for future
studies combining MRS with behavioural data in patients.
In summary, we use a cross-sectional design to study altered use of prior
expectations in auditory perception in individuals at-risk for psychosis, first episode
psychosis and controls. We expect to find differences in the balance between the
use of prior expectations and sensory input depending on the origin of the prior
expectation (sensory vs. cognitive) and disease stage (at-risk vs. first episode
psychosis). Specifically, we expect that at early stages of psychosis (clinical risk),
patients make relatively stronger use of sensory input then prior expectations
relative to controls and individuals with a full manifestation of illness (first episode
psychosis), but that in those with first episode psychosis, patients would rely more
on cognitive priors relative to sensory input compared to controls and individuals at
risk for psychosis. A secondary hypothesis is that cortical glutamate levels will be
related to changes in the usage of sensory and cognitive priors.
10
1.2.Method
1.2.1. Participants
Participants with first episode psychosis (FEP, n=30, average 24.8 years, 6 female) or
at-risk mental state patients (ARMS, n=29, average 21.5 years, 8 female) were
recruited from the Cambridge Early intervention service North and South. In
addition, ARMS patients were recruited from a help-seeking, low-mood, high
schizotypy sub-group following a latent class analysis on the (Neuroscience in
Psychiatry Network (NSPN) cohort (Davis et al., 2017) or through advertisement via
posters displayed at the Cambridge University counselling services. Individuals with
FEP or at-risk mental states for psychosis met FEP or ARMS criteria on the CAARMS
interview. All FEP participants had current delusions or previous delusions in the
case of those with partial or recent recovery. Healthy volunteers (Healthy control
sample HCS, n=32, average 22.6 years, 15 female) without a history of psychiatric
illness or brain injury were recruited as control subjects. Healthy volunteers did not
report any personal or family history of neurological, psychiatric or medical
disorders. All participants had normal hearing and normal or corrected to normal
vision. All participants gave informed consent. The study was part of the NCAAPS
study (Neuroscience Clinical Adolescent and Adult Psychiatry Study), which was
approved by the West of Scotland (REC 3) ethical committee. See Table 1 for details
on demographics and symptom scores. 3 ARMS patients and 17 FEP patients were
receiving anti-psychotic medication.
1.2.2. Questionnaires and interviews
We used the Cardiff Abnormal Perceptions scale (CAPS, Bell et al., 2006), Peters
Delusion Index scale (PDI, Peters et al., 1999), Comprehensive Assessment for the At-
risk Mental State interview (CAARMS, Yung et al., 2003) and Positive and Negative
Symptoms Scale (PANSS, Kay et al., 1989) to assess “caseness”, symptom severity
and frequency. Both the total scores for the CAPS and PDI and the subscales of the
CAPS and PDI are reported in table 1. For the PDI and CAPS the participants were
required to give a yes or a no answer to a particular question. In case of a yes
answer, 3 subscales were filled in which utilised a 5-point Likert scale. The CAARMS
and PANSS are semi-structured interviews, where the interviewer rates severity of
various types of psychotic and other psychiatric symptoms.
1.2.3. Magnetic Resonance Spectroscopy
A subset of participants was scanned on a Siemens Prisma 3T scanner at the
cognition brain sciences unit in Cambridge. The spectroscopy scan was part of a
larger MRI protocol which contained in addition 2 fMRI protocols and a structural
scan totalling 90 minutes. The structural scan was used to plan the MRS voxel. A
15mm isotropic voxel was placed carefully in the anterior cingulate cortex. A PRESS
sequence was used to assess glutamate levels, with a TR of 1880ms and TE of 30ms.
150 water-suppressed acquisitions were collected in addition to 16 unsuppressed
acquisitions. Data was analysed in LCModel. MRS data was successfully collected
from 18 healthy controls 19 ARMS, and 14 FEP patients.
1.2.4. Experiment 1 – providing perceptual priors
In the present study auditory stimuli were presented that contained varying
proportions of the phoneme /Ba or /Da (Figure 1). The balance between the two
stimuli always adds up to one. The contribution of the stimulus /Ba is denoted as
ωBa, which stands for “the weight of /Ba”. The proportion of ωDa can be derived from
ωBa as 1- ωBa = ωDa. From henceforth the notation ωBa be used to indicate what
exactly was presented to participants in terms of auditory stimulus.
12
Figure 1: Procedure of the sensory prior task. The participant was presented between a mixture of the phonemes
/Ba and /Da (above) which co-occurred with either a still face (reference condition) or lip-movements
pronouncing /Ba or /Da
Training phase
The task started with a training phase. The purpose of which was to familiarize the
participants with the auditory stimuli. Here they were presented with a still face in
combination with an auditory stimulus consisting of a stimulus ωBa= .8 or ωBa= .2.
They were then asked to report which sound they believed was dominant, after
which they received feedback (correct/incorrect). The training was completed as
soon as participants reported the correct answer 4 times for each stimulus. All
participants identified the phonemes correctly.
Testing phase
During the testing phase, the participants were presented with an auditory stimulus
consisting of a mix between the sound /Ba and /Da (as described above), which
simultaneously occurred with a visual stimulus consisting of a black and white male
face. The face would pronounce either /Ba or /Da (lip-movement condition), or the
face would remain still (the reference condition). All three conditions were
presented in a pseudo-randomised order such that all three conditions were
presented in a random order before one of the conditions is presented again. The
participants were instructed to keep looking at the lips of the face throughout the
task, but asked to report what phoneme was dominant in the auditory stimulus by
pressing one of four buttons indicating the level of certainty and the perceived
phoneme.
During the main task, the balance between the /Ba and /Da phoneme was changed
in a stepwise fashion. That is, when the participant reported the sound /Ba to be
dominant in for example the reference condition, then the next time that condition
came up, the balance between the sound /Ba and /Da would have been shifted in
favour of the non-reported phoneme, in this case: /Da. By following this procedure,
the task would converge towards a point where the participant would find it difficult
to distinguish which of the phonemes is dominant in the auditory stimulus. This
point is referred to as the perceptual indifference point. In the reference condition,
where no lip-movements were presented, we expected the perceptual indifference
point to converge on a stimulus which contains .5 of /Ba and .5 of /Da. However,
when lip-movements, for example pronouncing /Ba were presented to bias
perception towards the prior expectation, we expected that the task converged
upon an indifference point that contained less auditory /Ba, and more auditory /Da.
In other words, more auditory /Da was needed to overcome the influence that the
/Ba lip-movements had (see Figure 2, top panel, for a schematic representation of
the perceptual staircase experiment and figure 3 for an example of a staircase).
Figure 2: Schematic representation of a staircase in the perceptual priors task (upper panel) and cognitive priors
task (lower). The experiment adjusted the balance between /Ba and /Da during the experiment in favour of the
non-reported stimulus (slope line), ensuring convergence to a subject threshold (flat line). The distance A indicates
the strength of the Da prior, whereas B indicates the strength the Ba prior. C is a total measure of prior strength
irrespective of the specific prior presented.
For each of the three conditions (Reference, /Da and /Ba), the perceptual
indifference point was assessed twice: Once where the auditory stimulus started
14
with a dominant /Ba stimulus (ωBa= .7, ωDa= .3) and once where /Da was dominant
(ωBa= .3, ωDa= .7). This created 6 conditions, which were presented to the participant
in pseudorandom order. A condition was completed when either one of two criteria
was met. First, in the majority of cases, a perceptual indifference point was reached
which was defined as having made 6 switches in perceiving one stimulus over the
other (e.g. previously perceiving /Ba on trial t-1 and perceiving /Da on t0, indicating
the balance between the two auditory stimuli is close to the participants perceptual
indifference point). Second, a condition was completed when the participant
indicated that the sound /Ba or /Da is 100% dominant in the auditory stimulus (e.g. a
participant perceived /Da, even though the stimulus is 100% /Ba/ which could
happen when the visual priors are dominating perception). In the second case this
would technically not be an indifference point. However, for the remainder of this
study we will refer to it as such for the sake of simplicity. The priors dominated
perception only in a small minority of cases (see results). A condition was aborted
when 30 trials had been presented avoiding the task from taking too long. This did
not change the way the effect of the prior was calculated. In order to test for group
and condition differences in the amount of trials needed to reach an indifference
point and a possible interaction, we used a mixed-ANOVA with group as between
subject factor and visual condition as within subject factor.
Figure 3: Example of the staircase procedure. All 6 of the conditions are represented here. The top figure shows
the 3 visual conditions where the staircase started at ωBa=.3, whereas the bottom figure shows the 3 visual
conditions where the staircase started at ωBa=.7.
At the beginning of the staircase, the balance between /Ba and /Da was changed in
steps of .05. After the first switch, the balance was changed in steps of .015. This
15
procedure ensured that the first switch was reached quickly. Thereafter the staircase
became more sensitive so that the perceptual indifference point could be assessed
more precisely. The strength of each of the visual priors was calculated separately by
taking the difference between the perceptual indifference point of the visual prior
condition and the reference condition (see Figure 2 upper panel: A and B). The total
strength of the visual priors was calculated by taking the distance between the
indifference points of both sensory prior conditions (see Figure 2 upper panel: C).
1.2.5. Experiment 2 – providing cognitive priors
Training phase
The cognitive priors tasks was designed to measure how much a learned cue would
influence what participants hear. During the training phase participants learned the
association between the letters BA and the phoneme /Ba, and vice versa for DA. In
75% of the training trials the letters BA or DA were presented 500ms prior to hearing
the auditory stimulus which consisted of ωBa= .3 and ωDa= .7 when preceded by the
letters DA or ωBa= .7 and ωDa= .3 when preceded by the letters BA, making the letters
predictive of the auditory stimuli. In the other 25% of the trials, no sound was
presented following the letters. Here the participants were asked to report what
they expected to hear. The training was complete as soon as the participants
indicated 8 times that they expected to hear the /Ba following the letters BA and /Da
following the letters DA.
Figure 4: Procedure of the experimental phase of the cognitive prior task. A: First one of the three sets of letters
was presented to the participant to indicate what sound was most likely to occur according to the training phase.
B: participants were required to indicate which phoneme they believed to be most likely presented. C: The
A
B
C
D
16
participant was again presented with one of the three letters (the same as in A) and 500ms later was presented
with the mixed phoneme. D: After the presentation of the sound the stimuli were removed from the screen and
the participant was required to indicate what phoneme they perceived to be dominant.
Testing phase
The cognitive priors task is similar to the perceptual priors task, in that participants
were instructed to report which sound they believed to be dominant under different
prior expectations. However, this time the prior expectations came from learned
written word-sound associations. Again, the main task consisted of 3 conditions, a
cognitive prior BA and DA condition, and a reference condition, which consisted of
the letter ‘?A’. Each trial started with the presentation of the letters ‘BA’, ‘DA’ or
‘?A’. After seeing ‘BA’ or ‘DA’, participants were asked which phoneme they
expected to perceive, which they indicated using one of 4 buttons indicating the
perceived phoneme and certainty like in the perceptual priors task. The participants
were only asked to indicate their prediction following seeing the letters ‘BA’ and
‘DA’, but not after seeing ‘?A’. By making a conscious prediction regarding the
upcoming stimulus, the use of the cognitive prior could be validated. In the
reference condition, no reliable prediction could be generated as both options were
equally likely. 500ms after they made a decision or the reference stimulus had been
presented, the auditory stimulus was presented. Subsequently, participants
indicated what they perceived to be the dominant stimulus (see Figure 4).
Again, the balance between the auditory phoneme /Ba and /Da was shifted in favour
of the non-reported stimulus in a step-wise fashion. However, in contrast to the
perceptual priors task, each condition was presented once for each cognitive prior
BA and DA, instead of twice. Within the cognitive BA prior condition, the staircase
started at ωBa = .7 and ωDa = .3, meaning the auditory stimulus was relatively clearly
a /Ba sound. The same is true for the cognitive DA prior condition, where the
staircase started at ωBa = .3 ωDa = .7, meaning the auditory stimulus was relatively
clearly a /Da sound. This matching of the auditory stimulus to the cognitive prior
condition at the beginning of the staircase was done to reaffirm the association
between the prior and the sound, otherwise the association between the cue and
sound could have been lost immediately in the beginning of the staircase. Note that
17
if we would compare the difference in perceptual indifference points in the two
cognitive prior conditions, we would have a confound, as the staircases for the two
cognitive prior conditions started at different intensities, explaining any differences
between the two conditions. Therefore, we introduced two reference conditions to
which the prior conditions can be compared, getting rid of the confound. These
consisted of the letters ‘?A’, one of which had a staircase starting at ωBa = .7 and ωDa
= .3 so it could be directly compared to the cognitive BA prior, the other starting at
ωBa = .3 ωDa = .7, so it could be directly compared to the cognitive DA prior. As in the
first task, at the beginning of the staircase procedure, the balance between /Ba and
/Da was again changed in steps of .05. Then, after the first switch, the balance was
changed in steps of .015.
In total, the cognitive priors task consisted of 4 conditions: a BA and a DA condition,
a reference condition for BA, and a reference condition for DA. The order of the
condition per participants was pseudorandomised. In each condition, a perceptual
indifference point was assessed.
The perceptual indifference point for each condition was quantified by taking the
average of ωBa at the last two switches. We also briefly rapport the results for taking
the final four switches to demonstrate this does not influence the results
substantially. In order to quantify the strength of each prior, these perceptual
indifference points were subtracted from their reference condition, and the total
cognitive prior strength was calculated by adding the strength of separate priors (see
Figure 2 lower panel).
1.2.6. Stimuli, Apparatus and Procedure
Participants completed two tasks: the perceptual priors task first and the cognitive
priors task second. Each task was performed on a MacBook Pro, Retina, 13-Inch,
Early 2013, and each lasted on average about 10 minutes. Participants wore
Sennheisser Headphones to ensure optimal hearing. Both the Ba and the Da stimuli
had an intensity of 68dB. All participants reported perceiving the auditory stimuli
18
clearly. The experiment was conducted in an environment with minimal background
noise, ensuring minimal distraction of the participant (<15dB).
Psychtoolbox-3 was used to design the experiment. The auditory stimulus in both
the perceptual priors task and the cognitive priors task consisted of a mixture of a
natural speech male voice /Ba phoneme and a /Da phoneme. The auditory stimulus
was created by multiplying the auditory spectrum of the /Ba stimulus by a weighting
factor ωBa. This was then added to a weighted auditory spectrum of /Da (where ωDa=
1-ωBa) ensuring the total of auditory stimulus to always be 1 (stimulus = (ωBa x Ba) +
(ωDa x Da)).
1.2.7. Analyses
Since this is a novel paradigm, we first wanted to establish whether the variables of
interest were reliable in the sense that two separate measurements of the variable
were highly correlated. Since we assessed the perceptual indifference points twice in
each condition, we were able to test the correlation between two separately
obtained measurements, giving an indication of their reliability. We tested the
reliability of two separate variables. First, we tested the reliability of the indifference
points in the condition without a perceptual prior, which should give an indication of
the reliability of the staircase method. Second, we tested the reliability of the
strength of the perceptual priors, which give an indication of the reliability of the
method to measure the influence of lip-movements on auditory perception.
Furthermore we tested whether the perceptual and cognitive priors were correlated
with each other. Due to non-normality of the cognitive priors task, a Spearman
correlation was used to assess this.
One tailed paired T tests were used to test for a main effect of whether the lip-
movements shifted the perceptual indifference points in the expected direction
compared to the reference condition. This was done for both the sensory and
cognitive prior tasks.
In order to test the hypothesis that perceptual priors and cognitive priors were
different across groups, we computed the influence of the prior for each individual
as described above, and used a one-way ANOVA with two-tailed post-hoc Bonferroni
19
corrected t-tests if applicable. Furthermore, a Kruskal Wallis non-parametric ANOVA
was used with cognitive prior data, with Bonferroni corrected non-parametric post-
hoc t-tests. We also report the results of Bayesian statistical tests in relation to the
group differences using JASP. We report effect sizes for the key statistical tests, i.e.
effect of group on prior strength. We report Cohen’s d for T-tests, and η2 for the
one-way ANOVA’s. All effect sizes are calculated on the basis of parametric tests.
1.3.Results
Table 1: Demographics and symptom scores participants in the study
HCS
ARMS
FEP
p-value
32
29
30
PANSS
13.1(4.6)
26.7(12.1)
31.6(12.3)
<.001
Positive
6.5(2.3)
13.6(5.7)
18.0(6.9)
<.001
Negative
6.6(2.4)
13.1(7.5)
13.6(7.7)
<.001
MFQ
8.5(5.1)
33.2(17.4)
31.8(23.6)
<.001
CAPS
32.9(1.4)
44.1(7.0)
43.6(9.7)
<.001
Distress
1.6(3.0)
29.8(20.9)
32.1(33.9)
<.001
Intrusive
2.2(3.7)
34.9(22.8)
38.5(37.4)
<.001
Frequency
1.3(2.3)
28.3(17.8)
29.7(31.1)
<.001
PDI total
22.4(1.5)
29.3(4.5)
31.1(6.5)
<.001
Distress
2.4(2.8)
24.1(16.9)
28.0(23.9)
<.001
Intrusive
2.4(2.7)
23.6(17.4)
29.5(22.9)
<.001
Conviction
3.6(4.0)
24.9(15.9)
31.0(25.3)
<.001
Age
22.4(3.7)
21.8(3.5)
25.1(4.8)
<.01
N Males
17
21
24
>.05
20
1.3.1. Perceptual priors task
1.3.1.1.
No difference between groups in the amount of trials needed to assess
perceptual indifference point
On average participants required 18.9 trials to reach a perceptual indifference point
across all conditions. We found no overall effect of group on the trials needed to
reach a perceptual indifference point (F{2,87}=.262, p=.77) (HCS: 19.1, SE: 0.5;
ARMS: 19.1, SE:0.6; FEP: 18.6, SE: 0.4). However, we did find an effect of prior
condition (F{2,174}=17.1, p<.001): needing fewer trials in the visual reference
condition (17.3, SE: 0.3) than in the visual BA (18.9 SE: 0.4) and visual DA condition
(20.7, SE: 0.54). Importantly, we found no group by condition interaction
(F{4,174}=.456, p=.77). Thus, the patient groups did not differ in terms of the trials
needed to reach indifference points.
1.3.1.2.
Individual perceptual indifference points can be estimated reliably
The perceptual indifference point for each visual condition was assessed twice in the
perceptual priors task. As this is a novel task, we tested whether these
simultaneously assessed indifference points correlated strongly, as that would give
us an indication of the reliability of the measurement. First, we correlated the
indifference points in the condition where no priors were presented (the reference
condition). Across groups the correlation was r=.73. Separately it was r=.83 for HCS,
r=.76 for ARMS and r=.55 for FEP (all p<.01). The correlation between the two
reference points was significantly higher in the HCS group compared to the FEP
group (Fisher r-to-z transformation: p= .033), but not between other groups all p> .2.
Second, in a similar fashion, we assessed how strongly the effect of the perceptual
priors was correlated across the two simultaneously assessed staircases. The
reliability of the strength of the perceptual priors across groups was r=.78.
Separately, it was r=.88 for HCS, r=.79 for ARMS and .69 for FEP (all p<.01) (Figure 5).
The differences in correlations between perceptual priors were not significantly
different p>.2. For the remainder of the analyses we averaged for each visual
condition (Ba Da and reference) the perceptual indifference points, and the
estimation of the sensory prior strength (Figure 2).
21
Figure 5: Correlations testing the reliability of the experiment are presented here. A: reliability of the perceptual
indifference point in the reference condition. B: reliability of the strength of perceptual priors. C: Correlation
between the effect of cognitive Ba stimulus and the cognitive Da stimulus. D: correlation between sensory and
cognitive priors. E-G: relationship between cognitive and sensory priors for each experimental group. Whereas
healthy controls and FEP show a positive correlation, ARMS shows a negative correlation. We calculate Spearman
correlations but include linear fit lines for display purposes.
22
Figure 6: Main effects of the sensory and cognitive priors are presented here. A: relative shift in perceptual
indifference points under different sensory prior conditions (lip movements pronouncing /Ba or /Da) compared to
reference condition (still lips). B: relative shift in perceptual indifference points under different cognitive prior
conditions (the letters ‘BA’ and ‘DA’) compared to reference condition (letters ‘?A’). C: relative strength of
perceptual priors and cognitive priors. D: the perceptual indifference points in the reference conditions per group
(effect of no interest). Error bars represent standard error of the mean.
1.3.1.3.
Perceptual priors shifted the perceptual indifference points in the
expected direction
We tested whether the perceptual priors shift the perceptual indifference points in
the expected directions compared to the reference condition. On average, across all
groups taken together, Ba lip-movements lowered the value of ωBa in the perceptual
indifference point by .21 (95% ci: .18-.23, T{89}=14.0, p<.0001). In contrast, Da lip-
movements increased the value of ωBa in the perceptual indifference point by .16
(95% ci: .14-.18, T{89}=13.2, p<.0001) on average. When comparing the relative
strength of the Ba and Da lip-movements, we found a significant difference
(T{178}=2.29, p=.022), indicating a slightly stronger effect of Ba lip-movements then
Da (Figure 6A).
23
1.3.1.4.
The perceptual indifference point in the reference condition was equal
across groups
Analysing group differences, the perceptual indifference point in the reference
condition was a variable of no interest, as it merely reflects a personal preference for
either the auditory /Ba or /Da stimulus. Indeed, the average perceptual indifference
point in the reference condition across groups in reference groups was equal
(MHCS=.48 SEHCS=.02, MARMS=.49 SEARM=.01, MFEP=.51 SEFEP=.01; F{2,88}=1.02, p=.36)
(Figure 6D).
1.3.1.5.
Perceptual priors were significantly lower in ARMS compared to FEP
To test whether the perceptual priors were significantly different across groups, we
conducted a one-way ANOVA. We indeed found evidence for a difference across
groups (F{2,88} = 5.32, p=.007, effect size η2=.11; Figure 7A, 7C). Bonferroni
corrected post-hoc T-tests revealed a significant difference between ARMS
(MARMS=.28 SEARMS=.03) and FEP (MFEP=.44 SEFEP=.04) (p=.005, effect size d=.89, ci=
.46-1.32), but not between healthy controls (MHCS=.37 SEHCS=.04) and ARMS (p=.20,
effect size d= -.51, ci=.01-1.01) or FEP (p=.44, effect size d=.34, ci=-.13-.85). We
tested whether changing the amount of switch points that were used to calculate
the indifference point changed the results. When we change this from two to four,
we find the same (slightly stronger) effect: F(2,88) =5.72, p=.005, ARMS vs FEP:
p=.002, ARMS vs HCS: p=.12, HCS vs FEP: p=.24).
Figure 7: The effects per group are presented here in boxplot A: The effect of perceptual priors across groups. B:
The effect of cognitive priors across groups. * = p<.05
24
We furthermore analysed the perceptual prior data in a Bayesian fashion. For this
section we use Jeffreys’s (1961) suggested evidence categories for the Bayes factor.
We found that an ANOVA revealed moderate evidence in support for a difference
across groups (BF=6.3). Independent-sample t-tests revealed anecdotal evidence in
favour of a difference between ARMS and healthy controls (BF=1.4), but anecdotal
evidence in favour of no difference between healthy controls and FEP (BF=1.8).
There was strong evidence for a difference between ARMS and FEP (BF=26.1) (Figure
7A, 7C).
The /Ba perceptual prior dominated perception completely in 4/32 HCS, 0/29 ARMS
and 5/31 FEP participants, whereas the /Da perceptual prior dominated perception
completely in 5/32 HCS, 2/29 ARMS and 11/31 FEP. In one FEP participant the both
the /Da and /Ba lip-movements completely dominated perception.
25
1.3.2. Cognitive priors task
1.3.2.1.
FEP needed on average an extra trial to finish the training phase
We first tested whether the different experimental groups differed in the amount of
trials needed to end the training using an ANOVA. The groups differed significantly in
the number of trials needed (F{2,88}=3.34, p=.040). The HCS group and the ARMS
group required on average 8.7 trials and 8.8 trials respectively before the training
was finished, whereas the FEP required on average 9.9 trials.
1.3.2.2.
No difference between groups in the amount of trials needed to assess
perceptual indifference point
During the actual experiment, the participants generally required 18.5 trials to reach
a perceptual indifference point across all conditions. We found no overall effect of
group on the trials needed to reach a perceptual indifference point (F{2,88}=.44,
p=.64) (HCS: 18.5, SE: 0.6; ARMS: 18.9, SE:0.6; FEP: 18.1, SE: 0.6). However, we did
find an effect of prior condition (F{2,88}=3.56, p=.033). Needing significantly fewer
trials in the DA condition (17.6, SE: 0.5) then in the visual BA (19.5 SE: 0.5) (T
{180}=2.63, p=. 018) but not the reference condition (18.3, SE: 0.5) (T{180}=1.08,
p=.56, Bonferroni corrected). Importantly, we found no group by condition
interaction (F{4,176}=.27, p=.90). Thus, the patient groups did not differ in terms of
the trials needed to reach indifference points.
1.3.2.3.
Cognitive priors shifted the perceptual indifference points in the
expected direction
In order to assess the main effect of cognitive priors, each perceptual indifference
point of the two cognitive prior conditions was subtracted from its own reference
condition. We found that the cognitive BA prior lowered the value of ωBa by .042
(zval = -5.2, p< .0001), and for the cognitive DA prior the value of ωBa was increased
by .027 (zval = 3.7, p= .0002). This shows that there was indeed a main effect of
cognitive priors on perceptual indifference points. The relationship between effect
of BA and DA priors is shown in Figure 5C. For the remainder of the analyses, the
26
degree of influence of the BA and DA cognitive priors were added together and
averaged in order to create a single measure of cognitive prior strength (see Figure
6B).
1.3.2.4.
Effect of cognitive priors in the FEP group was significantly higher than
the ARMS and controls
We used a non-parametric ANOVA that is robust against Type I errors in non-
normally distributed data. The differences between the average strength of the
cognitive priors was significant (Independent-Samples Kruskal-Wallis Test: p=.023,
effect size η2=.11). Using a post-hoc Bonferroni corrected Wilcoxon rank sum test,
we found stronger usage of cognitive priors in the FEP group compared to both the
HCS group (zval: 2.35, ranksum: 840, p=.037, effect size d=.64, ci=.11-1.17), and the
ARMS group (zval:2.35, ranksum: 714, p=.037, effect size d=.62, ci=.10-1.14), but
between the HCS group and the ARMS group p>.5. We tested whether changing the
amount of switch points that were used to calculate the indifference point changed
the results. When we change this from two to four, we find the same (slightly
stronger) effect: FEP vs HCS: p=.015, FEP vs ARMS: p=.016, HCS vs ARMS: p>.5).
We also analysed the cognitive prior data in a Bayesian fashion, and found that an
ANOVA revealed moderate evidence in support for a difference across groups
(BF=7.5). Independent-sample t-tests revealed moderate evidence in favour of no
difference between ARMS and healthy controls (BF=3.5), but moderate evidence in
favour of a difference between healthy controls and FEP (BF=3.5). There was also
anecdotal evidence for a difference between ARMS and FEP (BF=2.8) (Figure 7B, 7D).
Although we had no evidence that the extreme values represent measurement
error, we analysed the results having excluded outliers in all three experimental
groups (1 HCS, 1 ARMS, 3 FEP). We found similar results (two sample t-test adjusted
for multiple comparisons: averaging over final 2 switch points: HCS vs FEP: p=.035,
ARMS vs FEP: p=.038. Final 4 switch points: HCS vs FEP p= .050, ARMS vs FEP p=
.051).
For the cognitive prior experiment there was one FEP participant for whom the BA
prior completely dominated perception, and 2 other FEP participants for whom the
27
DA prior completely dominated perception, with no occurrences in ARMS or HCS.
There were no participants for whom both the BA and DA cue completely dominated
perception.
1.3.3. Perceptual priors had a stronger effect on perception than cognitive priors
and were differently correlated across groups
Finally, we analysed whether the strength of the priors was different between tasks.
This was indeed the case, showing a stronger effect of perceptual priors (.37)
compared to the cognitive priors across all groups (.07) (T{90}=-14.34, p<.0001,
effect size d=1.5, ci= 1.8-1.2) (Figures 5D, 6C). Subsequently, we tested whether the
strength of cognitive and perceptual priors was correlated using a Spearman
correlation. This was indeed the case (Rho=.24, p<.02). When exploring the
correlations separate for each group, we found a negative (trend-level) correlation in
the ARMS group (Rho=-.33, p=.08), and positive correlations in the HCS (Rho=.52,
p=.002) and (trend-level) in the FEP group (Rho=.30, p=.10). Using a Fisher r-to-z
transformation We found that the relationship was significantly different for the
ARMS group compared to the healthy control group (Z=-3.28, p=.001), and FEP group
(Z=-2.25, p=.024). The correlation between healthy controls and FEP was not
significantly different (Z=1.0, p=.31). As these findings constituted secondary
analyses, they are not properly controlled for multiple comparisons. When
controlling for multiple tests, only the relationship in the healthy control group
remains significant.
1.3.4. Glutamate levels correlate with cognitive priors in HCS and perceptual
priors in FEP
Correlations with glutamate were tested in a subset of participants, namely 18
healthy controls, 19 ARMS, and 14 FEP patients. We looked for a correlation across
all participants between glutamate levels and the strength of the perceptual and
cognitive priors, but found no significant correlation (perceptual: Rho=.18, p=.21,
cognitive: Rho=.17, p=.23). When exploring the correlations in the separate patient
groups, we found that there is a significant positive relationship between glutamate
levels and cognitive priors in the control group (Rho=.53, p=.023), but not with
28
perceptual priors (Rho=.294, p=.24). In the ARMS group no significant correlations
were found for either cognitive (Rho=.0, p=1) or perceptual priors (Rho=.07, p=.78).
In the FEP group a significant correlation was found with perceptual (Rho=.57,
p=.035) but not cognitive priors (Rho=.43, p=.128). As these findings were secondary
to the core hypothesis in the present chapter, they were not corrected for multiple
comparisons. The effects do not remain significant when they are controlled for
multiple comparisons (See Figures 8 and 9).
29
30
Figure 8: Correlations between perceptual prior strength and glutamate levels for all groups. We report Spearman’s correlations but plot linear fits for display purposes.
31
32
Figure 9: Correlations between cognitive prior strength and glutamate levels for all groups. We report Spearman’s correlations but plot linear fits for display purposes.
33
1.3.5. Stronger cognitive priors are associated with delusion ideation in ARMS
and weaker perceptual priors is associated with delusion ideation and
hallucinations in FEP
To explore the relationship between the usage of sensory and cognitive priors and
the relation with symptoms, we computed Spearmen correlations within the
different experimental groups (Table 2).
In brief, we found that an increase in cognitive prior use was associated with
delusion ideation in ARMS, whereas a decrease in the usage of perceptual priors was
associated with perceptual abnormalities and delusion ideation in the FEP group.
Table 2: Correlations between abnormal perception and belief and usage if sensory
and cognitive priors across all groups.
ARMS
FEP
ARMS+FEP
HCS
Sensory
Cognitive
Sensory
Cognitive Sensory
Cognitive Sensory
Cognitive
PDI
p=.44
Rho=-.16
p=.030
Rho=.44
p=.023
Rho=.-48
p=.19
Rho=-.29
p=.09
Rho=-.25
p=.28
Rho=-.16
p=.92
Rho=.018
p=.27
Rho=.21
CAPS
p=.84
Rho=.044
p=.87
Rho=.037
p=.008
Rho=.-55
P=.16
Rho=-.31
p=.10
Rho=-.24
p=.86
Rho=.03
p=.06
Rho=.34
p=.56
Rho=.11
34
1.4.Discussion
In the present study we found that whether prior expectations have a stronger or
weaker effect on perception in psychosis depends on the origin of the prior
expectation and the disease stage. We found strong evidence of weakened
perceptual priors in the ARMS group compared to the FEP group, and some evidence
of ARMS versus controls differences. In contrast, when comparing cognitive priors
we found that the FEP group had stronger priors compared to the ARMS and healthy
control group, whereas the healthy controls and ARMS group did not differ from
each other.
The present findings can be interpreted in the hierarchical predictive coding
framework. This framework suggests that the brain models the world by making
predictions about upcoming sensory input, that are subsequently updated by
discrepancies between the predictions regarding the sensory input and the actual
sensory input, termed the prediction error (Knill et al., 2004; Friston, et al., 2005 &
2009; Rao et al., 1999; Bastos et al., 2012; Clark et al., 2013; Hohwy, 2014). In these
models, abnormal perception and delusional beliefs can be expected to occur when
the balance between the prior expectations and sensory input is shifted (Fletcher &
Frith, 2009), as was found in the present experiment. That is, sensory input can come
to dominate perception, likely resulting in the subjective experience of being
overwhelmed by their sensory environment and attributing importance to otherwise
irrelevant stimuli, as is sometimes reported in the early, including prodromal, stages
of psychosis (Corlett et al., 2010, McGhie and Chapman, 1961; Bowers and
Freedman, 1966; Freedman, 1974; Matussek, 1952).
Our results of abnormally strong high-level priors in first episode psychosis, all of
whom had either current or recent delusions, are in accordance with previous
postulates (e.g. Adams et al 2013, Sterzer et al 2018). We further note that high
level, cognitive, priors were stronger in established psychosis compared to the
ARMS, consistent with previous theory that strong high-level priors may develop
subsequent to weak low-level priors (Adams et al 2013, Sterzer et al 2018). As Heinz
35
et al (2018) reason, “reduced precision of perceptual beliefs encoded at low levels,
e.g. in sensory cortices, may be compensated by increased precision of more
abstract conceptual beliefs encoded in higher-level brain circuits.” However,
previous theories have not described on what time scale this compensation
happens, and no previous studies have examined over what time scale or at what
stages in psychotic illness this may occur. Our data suggest that this compensation
may not necessarily be instant, but might develop over time, possibly in the
transition from the prodromal stage (ARMS) to frank psychosis (FEP).
A recent study examining the influence of prior expectations on auditory perception
used a conditioning paradigm to study aberrancies in healthy voice hearers, voice
hearers with psychotic illness, and psychotic illness without voice hearing (Powers et
al., 2017). Individuals who heard voices were susceptible to report hearing a sound
when none was present following a previously associated cue. Computational
modelling showed that individuals with psychotic illness had difficulties learning that
a cue failed to predict a sound, sticking to their prior expectations, whereas
individuals who heard voices but did not have psychotic illness did recognise
volatility and were able to alter high-level beliefs. This might in part explain why we
only see an effect of the cognitive priors in the psychosis group, but not in the at-risk
mental state group, who, although help-seeking, do not (yet) have psychotic illness.
Because the current paradigm involves a staircase experiment, we only pick up
strong effects of prior expectations in individuals who remain influenced by the
priors towards the end of the experiment. The individuals at-risk for psychosis might
have been influenced in the task early on, but changed their expectations regarding
the validity of the cue later on. Since our key-variable is the influence of the priors at
convergence, we might have been unable to pick this up.
It should be pointed out that there are a number of outliers in the first episode
psychosis group. Although our statistical tests are robust against Type I errors in a
data set with outliers (Zimmerman, 1994), and the results hold when removing these
outliers, it still raises the question what the nature of the outliers is. One possibility
is that there is a subset of individuals that is exceptionally strongly influenced by
36
prior expectations. Indeed previous studies have reported non-normal data on such
variables (see Powers et al., 2017 Fig 1E). However, there is also the possibility that
these participants performed the task differently or misunderstood the instructions,
although we have no evidence that these outliers were caused by experimental
measurement error. The reliability of the perceptual priors was slightly less in the
FEP group compared to the other groups, but it should be noted that this difference
was not significant, and there was still a reliable correlation between the
independently assessed prior strengths (correlation 0.7 for use of sensory priors in
FEP). In addition an average was taken from the two independently assessed priors,
likely increasing the reliability further. Furthermore, since the present task does not
measure performance, but rather a perceptual bias, an increase in noisy decision
making will not bias the results in one way or the other.
It has been argued that there might be a relationship between early sensory
processing deficits and high level deficits in schizophrenia (Leitman et al., 2009). This
raises the question what the exact nature of this exact relationship is and whether it
might be relevant in understanding the development of psychosis. Whilst in the
sample as a whole, cognitive and perceptual prior strengths were weakly (rho=0.24),
but significantly, correlated, the strengths of the correlations were significantly
different across groups. Although we acknowledge the caveat that within group
correlations were of secondary interest in this study, and not well powered, the fact
that the group comparisons in strengths of priors were sensitive to whether priors
were high or low level provides supporting evidence that level of priors does matter
in this research context. Perceptual priors in ARMS were negatively correlated with
cognitive priors, whereas in FEP and healthy controls, and the sample as a whole, the
correlation was positive. We speculate on the possibility that an increase in the
influence of cognitive priors on perception in the FEP group is an adaptation to early
visual processing deficits in the earlier stages of psychosis as seen in the at-risk
group. This increase in cognitive priors subsequently could potentially act to counter
the decrease in diminishment of perceptual priors explaining the positive correlation
that is observed in the FEP group. This increase in cognitive priors may manifest
themselves as delusions on the phenomenological level as can be seen in both the
37
strong cognitive priors in FEP, and in the correlation with symptom severity in the
ARMS group. Subsequently, if perceptual priors remain low in the FEP stage, this is
correlated to worse symptomology, suggesting a failure for the brain to deal with a
change in the perceptual system may be important for psychopathology severity in
this stage of the illness. Interestingly, in the FEP stage there is no correlation
between cognitive priors and symptoms, possibly due to noise added to the data
through the effects of treatment, recovery in some, and delusional belief formation
being an attempt at making sense of a changing sensory world (Mishara & Corlett,
2009). Overall, our data emphasize the importance of distinguishing between priors
at high and low levels of the cognitive hierarchy (Schmack et al 2013).
We conclude that the initial stages of psychosis may be characterised by a
weakening of lower-level perceptual priors. Compensatory neural systems changes
may lead to deploying stronger higher-level priors in order to deal with the increased
strong drive on perceptual input. These changes might be associated with formation
of delusional beliefs (as supported by the correlations with symptoms). If this
compensatory strategy is effective, the weakened perceptual priors may be restored
throughout development. If ineffective, the perceptual priors remain weak and
psychotic symptoms maintain (as supported by the correlations with symptoms).
This model is described in Figure 10, where in red and blue the strength of
perceptual and cognitive priors are depicted respectively over time in psychosis, in
which the dotted line indicates worse clinical outcome in some patients. This model
can be tested in longitudinal designs to clarify the temporal and causal relationship
between the different priors.
38
Figure 10: A proposed model for the interaction between different levels of prior over time in psychosis. The early
stages of psychosis might be characterized by a weakening of lower-level perceptual priors as indicated by a fall
in the lower red line. This causes a shift in the strength of cognitive priors as an attempt to explain the abnormal
perceptual experiences, causing positive symptoms of psychosis. This will counter the weakening of lower-level
priors A failure to attenuate the weakening of lower-level priors may result in more severe, sustained symptoms
as indicated by the dashed lines.
Two previous studies have looked at the McGurk effect in schizophrenia. White et al
(2009) found that patients were, on average, less vulnerable to the illusion than
controls, with a strong relationship with duration of illness, such that individuals who
have been ill for longer were less likely to report a McGurk effect (White et al.,
2014). Pearl et al (2009) used a more complex recruitment design and had more
mixed results that interacted in a complex fashion with age; the interpretation of
their patient results are made challenging given that results in controls interacted
with age in an unexpected manner. In these previous studies participants were
required to report binary choices on whether they perceived the McGurk effect,
whereas we used a staircase procedure to examine the degree of influence that lip-
movements have on auditory perception. We did not find a diminishment in the
degree that lip-movements influenced auditory perception in psychosis patients.
This might relate to differences in methodology, or perhaps to the age difference
between our study (mean age 24.9 years) and White’s study (mean age 39.0 years),
given that the absence of illusory effect was more marked in White et al’s older
patients with longer disease duration. Further studies looked at the ability for
schizophrenia patients to use lip-movements to understand written speech, which
found aberrancies in schizophrenia, while general lip-reading capabilities remained
39
intact (Myslobodsky et al., 1992; de Gelder et al., 2002; Ross et al., 2007; Pearl et al.,
2009; Szycik et al., 2013). Again, the patient groups in these studies consisted of
schizophrenia patients who were older and in a more chronic phase than in the
present sample, potentially explaining the discrepancy with the present study.
In the present study we have described our effects in terms of an increase or
decrease in the influence of prior expectations. However, it should be noted that the
present paradigm is not able to directly discern whether a stronger influence of prior
expectations in auditory perception is due to a change in the strength of prior or a
weakening in the strength of the sensory input. Indeed previous studies have shown
impairments in the ability to do simple auditory discrimination tasks in schizophrenia
(Javitt et al., 2015). Future studies could utilise simple auditory discrimination tasks
to explore whether these effects are driven by these deficiencies or whether they
can be separated.
It has been proposed that glutamatergic abnormalities may be prominent in the
early stages of psychotic illness (Merritt et al 2016, Kumar et al 2018), and that these
may be key in driving pathophysiology of illness, predictive processing dysfunction,
and psychopathology (Corlett et al., 2009 & 2011; Sterzer et al 2018). We did not
find a significant relationship between glutamate levels in the anterior cingulate
cortex and the strength of the perceptual and cognitive priors across all participants.
However, in an exploratory analyses, we analysed the groups separately, and here
we did find that in the healthy group there was a significant positive relationship
between anterior cingulate glutamate levels and cognitive priors, and in the FEP
group a significant relationship between glutamate levels and perceptual priors. This
relationship between anterior cingulate glutamate levels and perceptual priors in the
FEP group is interesting as the correlations suggest that a (sustained) weakening of
perceptual priors is particularly relevant to FEP symptomology, and thus glutamate
might play a role in having sustained weakened perceptual priors. The absence of a
correlation with the cognitive priors might be due to a lack of power, as a successful
glutamate scan was only acquired from 14 individuals who had first episode
psychosis. We report MRS results uncorrected for multiple comparisons, which
40
should currently be viewed as preliminary. Larger sample size studies on glutamate
levels, the strength of perceptual priors in psychosis, and their inter-relation, will be
required to confirm (or refute) our results, which should currently be interpreted
with caution. A further limitation of our MRS work is the use of a single region,
located in the anterior cingulate cortex, from which our glutamate measure is
drawn. We do not mean to imply that this is the only region influencing the role of
priors in decisions, but until MRS technology matures to allow simultaneous
acquisition of neurochemistry measures across the whole brain, a priori region of
interest selection will remain the norm.
As with all studies that use the at-risk-mental-state construct, there is an inherent
limitation in terms of the inability to prospectively determine whether an at-risk
individual will develop a first episode of psychosis. Therefore, future studies would
benefit from following up individuals determined to be in the at-risk group, and so
explore the predictive validity of a change in the usage of priors. Indeed, longitudinal
studies will be required for definitive conclusions about how use of priors relates to
illness stage.
Extending this research beyond the field of psychosis, we note that autism has been
suggested to also be associated with a weakening of priors, but which usually does
not develop into psychotic symptoms (Pellicano et al., 2012; van Boxtel et al., 2013;
Lawson et al., 2014), although there are increased rates of psychotic symptoms in
autism and other neurodevelopmental disorders (Hussain and Murray 2015; Larson
et al., 2017). The difference between schizophrenia spectrum psychosis and autism
may lie in the fact that autism presents itself in early childhood, whereas
schizophrenia spectrum illness typically develops later in adolescence. The
consequence of this is that during the emergence of schizophrenia spectrum
psychosis the brain has to explain a changing world, whereas the sensory driven
world autism is characterized by presents itself at birth, requiring no changes in the
model of the world to form (i.e. no formation of delusional beliefs), yet the
experience of being overwhelmed by sensory experiences remains. Future
experiments would need to use a longitudinal approach to support this hypothesis,
41
namely that psychosis is preceded by a decrease in the influence of perceptual priors
on perception, followed by a normalization accompanied by an increase in higher-
level cognitive priors, whereas autism has weakened priors from birth. In order to
test such hypotheses, longitudinal paradigms are preferred which require potentially
large groups of people. In order to acquire such amounts of data, the possibility of
online testing could be considered, for which the present experiments are well
adapted too, due to the simplicity of the paradigm and the brief duration of the
experiments (10 minutes each).
In conclusion, we found that the influence of perceptual priors might be weakened
in the early stages of psychosis but not in the later stages, whereas cognitive priors
are strengthened in the later stages but not early stages. We therefore suggest that
previous reported inconsistencies in the literature regarding the influence of prior
expectations on sensory processing might be due to differences in the origin of the
prior expectation and the disease stage. Furthermore, changes in perceptual and
cognitive priors might interact with each other throughout the development of
psychosis and glutamate might play a mediating role in the process.
42
Author roles: JH: conceptualization (lead role), methodology (lead role), software,
formal analysis, investigation, writing (original draft preparation, review and editing).
FK: formal analysis, supervision, writing (original draft preparation, review and
editing). JG: conceptualization, methodology, investigation, formal analysis, writing
(review and editing). HT: investigation, formal analysis, writing (review and editing).
MM: investigation, methods, formal analysis, writing (review and editing). IG:
supervision, project administration, funding acquisition. PCF: conceptualization,
project administration, writing (review and editing). GKM, conceptualization, project
administration, methodology, supervision (lead role), writing (original draft
preparation, review and editing)
Conflicts of Interest: P.C.F. has received payments in the past for ad hoc consultancy
services to GlaxoSmithKline All other authors declare no competing interests.
Funding: This work was supported by the Neuroscience in Psychiatry Network, a
strategic award from the Wellcome Trust to the University of Cambridge and
University College London (095844/Z/11/Z), Wellcome Trust (093270) Bernard Wolfe
Health Neuroscience Fund (P.C.F.), and the Cambridge NIHR Biomedical Research
Centre.
Acknowledgements
We would like to thank Owen Parsons with his help in designing the paradigm,
Rachel Anderson and Eleanor van Sprang for their help in data collection, CAMEO
staff for help with recruitment, and the participants.
43
References
Adams, R. A., Stephan, K. E., Brown, H. R., Frith, C. D., & Friston, K. J. (2013). The Computational Anatomy of Psychosis.
Frontiers in Psychiatry, 4.
Agathon M, Roussel A (1973): [Use of a sensory conditioning test in patients treated with psychotropic drugs]. Int
Pharmacopsychiatry 8:221–233.
Arnal, L. H., Wyart, V., & Giraud, A.-L. (2011). Transitions in neural oscillations reflect prediction errors generated in audiovisual
speech. Nature Neuroscience, 14(6), 797–801.
Bastos, A. M., Usrey, W. M., Adams, R. A., Mangun, G. R., Fries, P., & Friston, K. J. (2012). Canonical Microcircuits for Predictive
Coding. Neuron, 76(4), 695–711.
Bell, V., Halligan, P. W., & Ellis, H. D. (2006). The Cardiff Anomalous Perceptions Scale (CAPS): A new validated measure of
anomalous perceptual experience. Schizophrenia Bulletin, 32(2), 366–377. http://doi.org/10.1093/schbul/sbj014
Blank, H., & Davis, M. H. (2016). Prediction Errors but Not Sharpened Signals Simulate Multivoxel fMRI Patterns during Speech
Perception. PLoS Biology, 14(11), e1002577.
Bowers Jr., M. B. (1968). Pathogenesis of acute schizophrenic psychosis. An experimental approach. Archives of General
Psychiatry, 19(3), 348–355
Van Boxtel, J. J., & Lu, H. (2013). A predictive coding perspective on autism spectrum disorders. Frontiers in psychology, 4, 19.
Brogden WJ (1947): Sensory pre-conditioning of human subjects. J Exp Psychol 37:527–539
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain
Sciences, 36(3), 181–204.
Clark, A. (2015). Surfing uncertainty: Prediction, action, and the embodied mind. Oxford University Press.
Corlett, P. R., Frith, C. D., & Fletcher, P. C. (2009). From drugs to deprivation: A Bayesian framework for understanding models
of psychosis. Psychopharmacology.
Corlett, P. R., Honey, G. D., Krystal, J. H., & Fletcher, P. C. (2011). Glutamatergic model psychoses: prediction error, learning,
and inference. Neuropsychopharmacology, 36(1), 294.
Cornsweet, T. N. (1962). The staircase-method in psychophysics. The American journal of psychology, 75(3), 485-491.
Davies, D.J. (2017). Latent class analysis of healthy cohort reveals different subgroups of psychosis vulnerability. (Unpublished).
Egerton, A., Stone, J. M., Chaddock, C. A., Barker, G. J., Bonoldi, I., Howard, R. M., ... & McLean, M. A. (2014). Relationship
between
brain
glutamate
levels
and
clinical
outcome
in
individuals
at
ultra
high
risk
of
psychosis. Neuropsychopharmacology, 39(12), 2891-2899.
Ellson, D. G. (1941). Hallucinations produced by sensory conditioning. Journal of Experimental Psychology, 28(1), 1.
44
Fletcher, P. C., & Frith, C. D. (2009). Perceiving is believing: a Bayesian approach to explaining the positive symptoms of
schizophrenia. Nature Reviews Neuroscience, 10(1), 48–58.
Freedman, B., & Chapman, L. J. (1973). Early subjective experiences in schizophrenic episodes. Journal of abnormal
psychology, 82(1), 46.
Friston, K. (2005). A theory of cortical responses. Philosophical Transactions of the Royal Society of London. Series B, Biological
Sciences, 360(1456), 815–36.
Friston, K. (2009). The free-energy principle: a unified brain theory? Nature Reviews. Neuroscience, 11(2), 127–138.
de Gelder, B., Vroomen, J., Annen, L., Masthof, E., & Hodiamont, P. (2003). Audio-visual integration in
schizophrenia. Schizophrenia research, 59(2), 211-218.
Haarsma, J., Fletcher, P., Griffin, J., Taverne, H., Ziauddeen, H., Spencer, T., ... & Murray, G. (2019). Precision-weighting of
superior frontal cortex unsigned prediction error signals benefits learning, is mediated by dopamine, and is impaired in
psychosis. BioRxiv, 558478.
Heinz, A., Murray, G.K., Schlagenhauf, F., Sterzer, P., Grace, A.A., Waltz, J.A., (2018) Towards a Unifying Cognitive and
Computational Neuroscience Account Of Psychotic Experience. Biorxiv
Hohwy, J. (2013). The predictive mind. Oxford University Press.
Hussain, H., & Murray, G. K. (2015). Common childhood neurodevelopmental disorders are associated with increased risk of
psychotic experiences in early adolescence. Evidence-based mental health, 18(2), 51.
Katthagen, T., Mathys, C., Deserno, L., Walter, H., Kathmann, N., Heinz, A., & Schlagenhauf, F. (2018). Modeling subjective
relevance in schizophrenia and its relation to aberrant salience. PLoS computational biology, 14(8), e1006319.
Kay, S. R., Opler, L. A., & Lindenmayer, J. P. (1989). The Positive and Negative Syndrome Scale (PANSS): Rationale and
standardisation. British Journal of Psychiatry.
Knill, D. C., & Pouget, A. (2004). The Bayesian brain: The role of uncertainty in neural coding and computation. Trends in
Neurosciences, 27(12), 712–719.
Kot, T., & Serper, M. (2002). Increased susceptibility to auditory conditioning in hallucinating schizophrenic patients: a
preliminary investigation. The Journal of nervous and mental disease, 190(5), 282-288.
Kumar, J., Liddle, E. B., Fernandes, C. C., Palaniyappan, L., Hall, E. L., Robson, S. E., ... & Skelton, M. (2018). Glutathione and
glutamate in schizophrenia: a 7T MRS study. Molecular Psychiatry, 1.
Larson, F. V., Wagner, A. P., Jones, P. B., Tantam, D., Lai, M. C., Baron-Cohen, S., & Holland, A. J. (2017). Psychosis in autism:
comparison of the features of both conditions in a dually affected cohort. The British Journal of Psychiatry, 210(4), 269-
275.
Lawson, R. P., Rees, G., & Friston, K. J. (2014). An aberrant precision account of autism. Frontiers in human neuroscience, 8, 302.
45
Leitman, D. I., Sehatpour, P., Higgins, B. A., Foxe, J. J., Silipo, G., & Javitt, D. C. (2010). Sensory deficits and distributed
hierarchical dysfunction in schizophrenia. American Journal of Psychiatry, 167(7), 818-827.
Marsman, A., van den Heuvel, M. P., Klomp, D. W., Kahn, R. S., Luijten, P. R., & Hulshoff Pol, H. E. (2011). Glutamate in
schizophrenia: a focused review and meta-analysis of 1H-MRS studies. Schizophrenia bulletin, 39(1), 120-129.
McGhie, A. N. D. R. E. W., & Chapman, J. (1961). Disorders of attention and perception in early schizophrenia. Psychology and
Psychotherapy: Theory, Research and Practice, 34(2), 103-116.
McGurk, H., & MacDonald, J. (1976). Hearing lips and seeing voices. Nature, 264(5588), 746-748.
Merritt, K., Egerton, A., Kempton, M. J., Taylor, M. J., & McGuire, P. K. (2016). Nature of glutamate alterations in schizophrenia:
a meta-analysis of proton magnetic resonance spectroscopy studies. JAMA psychiatry, 73(7), 665-674.
Mishara, A. L., & Corlett, P. (2009). Are delusions biologically adaptive? Salvaging the doxastic shear pin. Behavioral and Brain
Sciences, 32(6), 530-531.
Myslobodsky, M. S., Goldberg, T., Johnson, F., Hicks, L., & Weinberger, D. R. (1992). Lipreading in patients with
schizophrenia. The Journal of nervous and mental disease, 180(3), 168-171.
Pearl, D., Yodashkin-Porat, D., Katz, N., Valevski, A., Aizenberg, D., Sigler, M., … Kikinzon, L. (2009). Differences in audiovisual
integration, as measured by McGurk phenomenon, among adult and adolescent patients with schizophrenia and age-
matched healthy control groups. Comprehensive Psychiatry, 50(2), 186–192.
Peters, E. R., Joseph, S. A., & Garety, P. A. (1999). Measurement of delusional ideation in the normal population: introducing
the PDI (Peters et al. Delusions Inventory). Schizophrenia bulletin, 25(3), 553-576.
Powers, A. R., Mathys, C., & Corlett, P. R. (2017). Pavlovian conditioning–induced hallucinations result from overweighting of
perceptual priors. Science, 357(6351), 596-600.
Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical
receptive-field effects. Nature Neuroscience, 2(1), 79–87.
Ripke, S., Neale, B. M., Corvin, A., Walters, J. T., Farh, K. H., Holmans, P. A., ... & Pers, T. H. (2014). Biological insights from 108
schizophrenia-associated genetic loci. Nature, 511(7510), 421.
Pellicano, E., & Burr, D. (2012). When the world becomes ‘too real’: a Bayesian explanation of autistic perception. Trends in
cognitive sciences, 16(10), 504-510.
Ross, L. A., Saint-Amour, D., Leavitt, V. M., Molholm, S., Javitt, D. C., & Foxe, J. J. (2007). Impaired multisensory processing in
schizophrenia: deficits in the visual enhancement of speech comprehension under noisy environmental
conditions. Schizophrenia research, 97(1), 173-183.
Schmack, K., de Castro, A. G. C., Rothkirch, M., Sekutowicz, M., Rössler, H., Haynes, J. D., ... & Sterzer, P. (2013). Delusions and
the role of beliefs in perceptual inference. Journal of Neuroscience, 33(34), 13701-13712.
Schmack, K., Schnack, A., Priller, J., & Sterzer, P. (2015). Perceptual instability in schizophrenia: Probing predictive coding
accounts of delusions with ambiguous stimuli. Schizophrenia Research: Cognition, 2(2), 72-77.
46
Schmack, K., Rothkirch, M., Priller, J., & Sterzer, P. (2017). Enhanced predictive signalling in schizophrenia. Human brain
mapping, 38(4), 1767-1779.
Shergill, S. S., Samson, G., Bays, P. M., Frith, C. D., & Wolpert, D. M. (2005). Evidence for sensory prediction deficits in
schizophrenia. American Journal of Psychiatry, 162(12), 2384-
Sterzer, P., Adams, R.A., Fletcher, P., Frith, C., Lawrie, S.M., Muckli, L., Petrovic, P., Uhlhaas, P., Voss, M. and Corlett, P.R. (2018).
The predictive coding account of psychosis. Biological Psychiatry.
Szycik, G. R., Ye, Z., Mohammadi, B., Dillo, W., Wildt, B. T. T. E., Samii, A., … Gelder, D. (2013). MALADAPTIVE CONNECTIVITY OF
BROCA ’ S AREA IN SCHIZOPHRENIA DURING AUDIOVISUAL SPEECH PERCEPTIONM: AN FMRI STUDY, 253, 274–282.
Teufel, C., Kingdon, A., Ingram, J. N., Wolpert, D. M., & Fletcher, P. C. (2010). Deficits in sensory prediction are related to
delusional
ideation
in
healthy
individuals.
Neuropsychologia,
48(14),
4169–4172.
http://doi.org/10.1016/j.neuropsychologia.2010.10.024
Teufel, C., Subramaniam, N., Dobler, V., Perez, J., Finnemann, J., Mehta, P. R., … Fletcher, P. C. (2015). Shift toward prior
knowledge confers a perceptual advantage in early psychosis and psychosis-prone healthy individuals. Proceedings of
the National Academy of Sciences, 112(43), 13401–13406.
Treen, D., Batlle, S., Mollà, L., Forcadell, E., Chamorro, J., Bulbena, A., & Perez, V. (2016). Are there glutamate abnormalities in
subjects at high risk mental state for psychosis? A review of the evidence. Schizophrenia research, 171(1), 166-175.
Warburton, D. M., Wesnes, K., Edwards, J., & Larrad, D. (1985). Scopolamine and the sensory conditioning of
hallucinations. Neuropsychobiology, 14(4), 198-202.
White, T. P., Wigton, R. L., Joyce, D. W., Bobin, T., Ferragamo, C., Wasim, N., … Shergill, S. S. (2014). Eluding the illusion?
Schizophrenia, dopamine and the McGurk effect. Frontiers in Human Neuroscience, 8(August), 565.
Yung, A. R., Yuen, H. P., Phillips, L. J., Francey, S., & McGorry, P. D. (2003). Mapping the onset of psychosis: The comprehensive
assessment of at-risk mental states (CAARMS). Schizophrenia Research, 60(1), 30–31.
Zimmerman, D. W. (1994). A note on the influence of outliers on parametric and nonparametric tests. The journal of general
psychology, 121(4), 391-401.
| 2019 | Influence of prior beliefs on perception in early psychosis: effects of illness stage and hierarchical level of belief | 10.1101/421891 | null | creative-commons |
Climate risk to European fisheries and coastal communities
1
2
Mark R. Payne1*, Manja Kudahl1, Georg H. Engelhard2,3 , Myron A. Peck4,** and John K. Pinnegar2,3
3
4
1. National Institute of Aquatic Resources (DTU-Aqua), Technical University of Denmark, 2800 Kgs.
5
Lyngby, Denmark.
6
2. Centre for Environment, Fisheries & Aquaculture Science (Cefas), Pakefield Road, Lowestoft, United
7
Kingdom
8
3. School of Environmental Sciences, University of East Anglia (UEA), Norwich, United Kingdom
9
4. Department of Coastal Systems, Royal Netherlands Institute for Sea Research (NIOZ), PO Box 59,
10
1790 AB Den Burg (Texel), the Netherlands
11
12
13
*Corresponding author.
14
Mark R. Payne
15
National Institute of Aquatic Resources (DTU-Aqua)
16
Technical University of Denmark
17
2800 Kgs. Lyngby
18
Denmark
19
Tel.: +45 3396 3455
20
ORCID: 0000-0001-5795-2481
21
E-mail address: mpay@aqua.dtu.dk
22
23
**Present Address: Royal Netherlands Institute for Sea Research, PO Box 59, 1790 AB Den Burg, Texel,
24
The Netherlands
25
26
27
Keywords: hazard, exposure, vulnerability, risk, climate change, fleets
28
Abstract
29
With the majority of the global human population living in coastal regions, correctly characterising the
30
climate risk that ocean-dependent communities and businesses are exposed to is key to prioritising where
31
the finite resources available to support adaptation should be deployed. We apply a climate risk analysis
32
across the European fisheries sector for the first time to identify the most at-risk fleets and sub-national
33
regions and then link the two analyses together. We combine a trait-based approach with physiological
34
metrics to differentiate climate hazards between 556 populations of fish and use these to assess the relative
35
climate risk for 380 fishing fleets and 105 coastal regions in Europe. Countries in southeast Europe as well
36
as the UK have the highest risks to both fishing fleets and coastal regions overall, while in other countries,
37
the risk-profile is greatest at either the fleet or at the regional level. European fisheries face a diversity of
38
challenges posed by climate change: climate adaptation, therefore, needs to be tailored to each country,
39
region and fleet’s specific situation. Our analysis supports this process by highlighting where and what
40
adaptation measures might be needed and informing where policy and business responses could have the
41
greatest impact.
42
Main Text
43
The ocean provides human societies with a wide variety of goods and services, ranging from food and
44
employment to climate regulation and cultural nourishment (1). Climate change is already shifting the
45
abundance, distribution, productivity and phenology of living marine resources (2–4) and, thereby,
46
impacting many of the ecosystem services upon which society depends (5). These impacts, however, are
47
not being experienced uniformly by human society but depend on the characteristics and context of the
48
community or business affected. Raising awareness and understanding the risk to human systems is,
49
therefore, a critical key first step (6) to developing and prioritising appropriate adaptation options in
50
response to the challenges of the climate crisis (7).
51
Over the past decades, climate risk assessments (CRAs) and climate vulnerability assessments (CVAs) have
52
been developed to support the prioritisation of adaptation options. The approach, developed by the
53
Intergovernmental Panel on Climate Change (IPCC), has shifted over time from a focus on “vulnerability”
54
to a focus on “risk” (8), in part due to criticisms of the negative framing that “vulnerability” implies (9).
55
The modern CRA framework (10) considers risk as the intersection of hazard, exposure and vulnerability
56
(Table 1). CVAs, and more recently CRAs, have been applied widely in the marine realm, for example in
57
coastal communities in northern Vietnam (11), Kenya (12) and the USA (13), at the national level across
58
coastal areas of the USA (14, 15) and Australia (16, 17), across regions such as Pacific island nations (18,
59
19) and globally (6, 20, 21). Several ‘best practice’ guides have also been developed (7, 22).
60
Table 1 Definitions of terms, as used in the context of this Climate Risk Analysis. These definitions are
61
adapted for the present study from those used in the most recent IPCC report (5).
62
Term
Definition used here
Climate risk
The potential for climate change to have adverse consequences for human systems,
specifically for European coastal regions and fishing fleets.
Hazard
The potential for, and severity of, climate change impacts on the unit of interest (i.e.
fish and shellfish populations). Here we focus explicitly on negative impacts, following
from the definition of risk as being an adverse consequence.
Exposure
The sensitivity of a region or fishing fleet to changes in the living marine resources that
it depends on.
Vulnerability
The ability of a region or fleet to cope with or adapt to the hazards presented by climate
change. High adaptive capacity gives low vulnerability.
63
CRAs and CVAs covering European waters, are however, notable by their absence. The lack of attention
64
to climate risk in European fisheries may arise, in part, from the previous results of global CVAs (6) that
65
ranked European countries as having low vulnerabilities due to their affluence and, therefore, high ‘adaptive
66
capacity’. Yet the European region poses unique challenges when assessing climate risks due to its wide
67
range of species, biogeographical zones and habitats. Fishing techniques and the scale of fisheries vary
68
widely, from large fleets of small vessels in the Mediterranean Sea (23) to some of the largest fishing vessels
69
in the world (e.g. the 144-m long Annelies Ilena). Furthermore, although fisheries contribute very little to
70
national GDP, food or income-security for most European countries (24), in specific communities and
71
regions fishing is the mainstay of employment (25). Adapting European fisheries to a changing climate
72
therefore requires the development of robust analyses capable of assessing the climate risk across this
73
extremely diverse continent.
74
We conducted a detailed CRA across the European marine fisheries sector, estimating the climate risk of i)
75
coastal regions and ii) fishing fleets in linked analyses. Our analyses span more than 50 degrees of latitude
76
from the Black Sea to the Arctic and encompass the United Kingdom, Norway, Iceland and Turkey in
77
addition to the 22 coastal nations of the European Union. We apply an approach that incorporates fine-scale
78
geographical differences in the climate hazard of fish and shellfish populations and then assess the climate
79
risk of both European coastal regions and fishing fleets. Since both CRAs are based on the same underlying
80
climate hazard assumptions, these analyses can be combined to compare the relative importance of the
81
climate hazard to fleets and coastal regions within a country.
82
Coastal-Region Climate Risk Analysis
83
Our index of climate hazard is derived from the biological traits of the species being harvested, together
84
with modelled distribution data. Species trait data were gathered for 157 fish and shellfish species harvested
85
in European waters, representing 90.3% of the total value of landings in Europe and at least 78% (and
86
typically more than 90%) of national value. We accounted for the expected large differences in climate
87
hazard throughout a species range (i.e. from the cold to warm edges of the distribution) by focusing on
88
“populations” (i.e. a single species in a single FAO subarea). Population-level climate hazards were then
89
defined based on the thermal-safety margin (TSM) between the temperature in that subregion and the upper
90
thermal tolerance of the species (26, 27). Climate hazards were calculated for 556 “populations” in 23 FAO
91
subareas, based on the TSM of the population and the inherent traits of the species (15, 28, 29).
92
We then calculated the climate risk for 105 coastal regions across 26 countries in the European continent
93
(Figure 1). Population-level climate hazards of fish were integrated to regions, weighted by the relative
94
value of landings in that region. We defined exposure metrics based on the diversity and dominance (30,
95
31) of these landings, and vulnerability based on regional socio-economic metrics (6). We focused our
96
analysis on coastal regions, as these are the communities most directly dependent on the ocean: regions far
97
from the sea but within a coastal nation were explicitly excluded (e.g. Bavaria in Germany).
98
The analysis reveals appreciable variation in the climate risk within the European continent and even within
99
a single country (Figure 1a). In the United Kingdom, for example, climate risk is greatest in the north of
100
England, while Scotland and the south of England show the least risk. Indeed, six of the 10 regions with
101
the highest climate risk, including the overall top region (Tees Valley & Durham), are in the UK (Table
102
S8). These results are strongly influenced by high hazard scores for the species landed in these regions
103
(Figure 1b), combined with high vulnerability due to low GDP per capita in some of these regions.
104
Larger-scale patterns in climate risk are also apparent. South-east Europe stands out with consistently high
105
climate risk, with coastal Romania and Croatia in the top five. Both countries have high vulnerability scores
106
due to low GDP per capita of their coastal regions, and high exposure scores due to fisheries that target
107
only a few species (e.g. the value of Romania’s fisheries is more than 70% veined rapa whelk, Rapana
108
venosa). Many northern European countries, including Belgium, the Netherlands and Scandinavian nations
109
have relatively low climate risks due to their wealth (high GDP per capita), diverse fisheries and the
110
relatively low climate hazard of the fish populations targeted.
111
These overall climate-risk scores are heavily influenced by the relative importance of the elements (hazard,
112
exposure or vulnerability) that dominate the risk profile (Figure 1b). The climate risk profiles of south-east
113
Europe, the Iberian peninsula and some regions on the south coast of the Baltic Sea are dominated by the
114
vulnerability dimension, reflecting the low GDP per capita of these regions. For the most part, exposure
115
scores are important in Northern Europe and in Scandinavia, reflecting the narrower range of species landed
116
compared to the Mediterranean region. The climate risk of Iceland, the UK, and parts of France or northern
117
Italy, on the other hand, are dominated by the climate hazard component, i.e. the traits and thermal
118
preferences of the species targeted. The relative contributions of the individual components are critical to
119
understanding the climate risk of each country and the suitability of particular adaptation responses.
120
121
Figure 1 Climate risk of European regions. Maps show a) the combined climate risk for each region and b) the individual
122
component (blue: hazard, green: exposure, purple: vulnerability) making the largest contribution to the combined risk and its value.
123
Colour scales on both panels are linear in the value of the corresponding score, but are presented without values, as they have little
124
direct meaning. National borders are also shown for reference. Insets at bottom-left of each panel show small regions.
125
Fleet Segment Climate Risk Analysis
126
The risks associated with climate change will also be felt by the fishing vessels and fleets that heavily
127
depend on living marine resources. We, therefore, performed a second CRA to examine the climate risk of
128
European fishing fleets. As the basis for this analysis, we followed the EU definition of a “fleet segment”
129
based on the size classes of the vessels, the country of registration, the gear being used and the geographical
130
region being fished (Atlantic or Mediterranean) (23). We integrated climate hazards at the fish population
131
level up to the fleet segment level, based on the composition of landings by value of that fleet, while we
132
based exposure on the diversity and dominance of landings and vulnerability on the net profitability of the
133
fleet. Coverage of our analysis at this fleet segment level was poorer than at the national level: nevertheless,
134
we still cover 75% or more of total fishery catch value for more than 70% of the 380 fleet segments within
135
the EU and UK.
136
The smallest class of vessels (0-6m) had an appreciably higher climate risk than all other size classes (Figure
137
2a). For the most part, these fleets operated in the Mediterranean region, particularly in Croatia, Bulgaria,
138
France, Malta and Greece (Table S9). This result reflects, in part, the higher climate risk of stocks in this
139
area, but is also driven by the poor profitability (and therefore higher vulnerability) of these fleets. On the
140
other hand, the high catch diversity of these fleets reduces exposure and helps to reduce their net climate
141
risk.
142
143
Figure 2 Climate risk of European fleet segments. The climate risk across 380 fleet segments is plotted as a function of a) the
144
size range of the vessels (m), b) the gear type employed (sorted by median risk) and c) the country of origin of the fleet (sorted by
145
median risk). Risk is represented on a linear scale from highest to lowest: the absolute values are not shown, as they have little
146
direct meaning. The distribution of risk is shown as a boxplot, where the vertical line is the median, the box corresponds to the
147
interquartile range (IQR), and the whiskers cover all points less than 1.5 times the IQR from the box. Outliers are plotted as points.
148
Boxes are coloured based on the median climate risk for that category. The number of fleet segments in each class is shown at right.
149
Note that EU definitions of small length classes (less than 12m) vary between individual countries and therefore have a degree of
150
overlap. Specific gear codes are aggregated here to broader-scale categories of “Gear Types” to ensure comparability between
151
Atlantic and Mediterranean fisheries (Table S4).
152
Systematic differences in climate risk are seen among gear types (Figure 2b), with dredgers having the
153
highest climate risk. These fleets generally target populations with high climate hazards and have low
154
species diversity in their catches (giving high exposure): good profitability, on the other hand, lowers their
155
vulnerability and somewhat reduces overall risk (Table S9). Fleets using pelagic and demersal trawls
156
together with purse seine fleets have the lowest climate risks, primarily due to the low hazard associated
157
with the species on which they fish.
158
The strongest differentiation in climate risk between fleet segments is at the national level (Figure 2c). A
159
clear cluster of high climate risk fleet segments can be seen in south-east Europe, particularly in Croatia,
160
Greece, Bulgaria, Cyprus and Romania (Figure S1). The risk profiles underlying each of these cases,
161
however, are quite different, emphasising the need to understand the components in detail. Greek and
162
Cypriot fleets have high climate risks due to poor profitability and, therefore, high vulnerability, while
163
Bulgarian and Romanian fleets active in the Black Sea have extremely low catch diversities, giving them
164
unusually high exposures (Table S9). It is also important to note that there is substantial variation among
165
fleets within a country. For example, two of the five most at-risk fleets (including the most at risk) are
166
Spanish (Table S9), even though the national level median for Spain is amongst the lowest in Europe. A
167
detailed examination of the individual elements of the risk-profile is, therefore, critical to understanding the
168
underlying factors responsible for these results.
169
Comparative Analysis
170
A strength of the analysis performed here is that the results of the region and fleet CRAs can be directly
171
compared. While the regions and fleets are exposed to the same base set of hazards, the relative importance
172
of each fish or shellfish population (and therefore hazard) differs. Each region and fleet also has its own
173
intrinsic exposure and vulnerability profiles, further modulating the overall climate risk. However, as the
174
base set of hazards is the same in both CRAs, a direct comparison of the two cases is possible, allowing the
175
relative climate risk to regions and fleets to be gauged.
176
Systematic differences in risk between fleets and coastal communities can be seen among European
177
countries (Figure 3) and several characteristic types of responses are apparent. Countries in south-eastern
178
Europe, together with the United Kingdom, have the highest risk across both fleets and coastal regions. The
179
climate risk scores of regions on the south coast of the Baltic Sea (Latvia, Lithuania, Estonia and Poland)
180
are typically higher than their fleet level scores, while the high fleet risk of NW European states is
181
moderated by their relative affluence and therefore low risk to regions. Spain and Sweden are characterised
182
by generally low climate risks in both coastal regions and fleets.
183
184
185
Figure 3 Comparison of the median fleet- and region-based risks for European countries. Labels indicate the country code.
186
In addition, France (FR) and Spain (ES) are split into their Atlantic (-A suffix) and Mediterranean (-M suffix) seaboards. As the
187
fleet-segment analysis only covers fleets from the EU and UK, no data are available for Turkey, Norway and Iceland: their regional
188
risk results are plotted in the horizontal margin. Dashed lines divide the coordinate system into quarters. Country codes: BE:
189
Belgium. BG: Bulgaria. CY: Cyprus. DE: Germany. DK: Denmark. EE: Estonia. EL: Greece. ES: Spain. FI: Finland. FR: France.
190
HR: Croatia. IE: Ireland. IS: Iceland. IT: Italy. LT: Lithuania. LV: Latvia. MT: Malta. NL: Netherlands. NO: Norway. PL: Poland.
191
PT: Portugal. RO: Romania. SE: Sweden. SI: Slovenia. TR: Turkey. UK: United Kingdom.
192
Discussion and Conclusions
193
Our analysis highlights the wide variety of challenges facing European countries with adapting their
194
fisheries sectors to a changing climate. In some cases, such as in the southern-Baltic states, a focus on
195
strengthening the resilience of coastal regions would be of most benefit e.g. by creating alternative
196
employment opportunities or providing an economic ‘safety net’ through wider social measures. In other
197
regions, fleet risks dominate and, therefore, increasing the efficiency, resilience and diversity of the fleet
198
would appear to be a priority. Some areas, such as the UK and south-east Europe appear to require both
199
types of intervention and, therefore, present the greatest adaptation challenges. Thus, there is no “one-size-
200
fits-all” solution that can be applied across all European waters or even, in some cases, across a country
201
(e.g. the UK): climate adaptation plans therefore need to be tailored to these realities.
202
Climate risk and vulnerability analyses can play a key role to play in shaping adaptation plans. By increasing
203
awareness of the elements that contribute to a fleet or coastal region’s risk (6), CVAs and CRAs can help
204
maximise the effectiveness of interventions given limited resources (32). Previous socio-economic linked
205
analyses have focused on adaptive capacity (in the CVA framework) as a focal point for action (6, 12).
206
However, the diversity of European risk profiles found here highlights the need and potential for adaptation
207
actions across all components of the risk portfolio.
208
Ensuring sustainable management of the living marine resources upon which the sector rests is a key action
209
for the European fisheries sector. The impacts of over-exploitation can be more important than those
210
stemming from climate change, particularly in the heavily fished North Atlantic region (33). Maintaining
211
stocks at a higher abundance leads to increases in genetic diversity, meta-population complexity, and age
212
structure, all of which make stocks more resilient to the challenges of a changing environment (34, 35). The
213
ensuing increase in productivity and incomes also simultaneously benefits both fishing fleets and regions,
214
generating a “win-win” effect (36). Fisheries scientists already have many of the tools necessary to ensure
215
that management systems are robust to climate change and climate variability (37), while new tools, such
216
as seasonal-to-decadal marine ecological forecasts and early-warning systems (38), can potentially provide
217
the basis for further coping strategies (39).
218
Diversification is a second key action to reduce climate risk. Fishing fleets and coastal regions relying on
219
only a few species have an elevated risk of climate impacts: increasing this spread reduces (by definition)
220
exposure and buffers fleets and regions against climate risk (31, 40, 41). Diversification of catches and
221
landings can take place autonomously as fishers respond to changes in the abundance and distribution of
222
the fish they depend on (32, 37). For example, changes in the distribution of fish species in surrounding
223
waters (42–44) have led to the development of new fisheries in the UK for squid, seabass and red mullet,
224
amongst others (45). CRAs such as this can also have an important role in this process by highlighting
225
alternative species or populations with a lower climate hazard that can be targeted, thereby further reducing
226
risk. Alternatively, diversification of income sources by, for example, participating in multiple fisheries or
227
in tourism and recreation has also been shown to reduce variability in income and thereby risk (46).
228
There are, however, barriers to diversification (31, 41), including knowledge, economic and governance
229
barriers. For example, the ability to catch new species may be limited by existing quota agreements (47), a
230
particularly challenging issue under the “relative stability” agreements of the EU Common Fisheries Policy.
231
Ecology can also be constraining: the limited catch diversity and therefore high exposure of fleets and
232
coastal regions adjoining the Black and Baltic Seas, for example, arises at least in part from the naturally
233
low biodiversity of these seas. Changing target species or fishing technologies can also be costly, creating
234
financial barriers to diversification (46).
235
Governance has a key role to play in coordinating and driving actions to reduce the vulnerability of fleets
236
and regions. Investments and support for developing new, and switching between, fishing, storage, transport
237
and processing technologies can increase the efficiency of fleet operations and, therefore, reduce
238
vulnerability (18, 37, 48). Increasing regional development, including employment opportunities outside
239
the fisheries sector, reduces regional vulnerability and risk (6, 49). Furthermore, both fishing fleets and
240
coastal regions can also potentially benefit from governance-led actions that increase the flexibility, ability
241
to learn, social organisation and the power and freedom to respond to challenges (50). Regional, national
242
and European governments therefore have a critical role to play in helping fisheries and ocean-dependent
243
regions to adapt to the risks presented by climate change.
244
Several key caveats of our results need to be highlighted. Our analysis focused solely on the sensitivity to
245
ocean warming, ignoring other climate-driven processes, such as ocean acidification, deoxygenation, and
246
changes in storminess or circulation patterns (5, 30) that, while important, we view as second order effects.
247
Spatial differences in the rates of warming across European regional seas were also not accounted for here
248
but the range of these rates (up to 2°C by 2050) is much smaller compared to the variability in thermal
249
safety margins across the range of some species (range up to 15°C) (Figure S3). The treatment of
250
uncertainty in CVAs and CRAs varies greatly between studies (15, 51) but in such a semi-quantitative
251
analysis, the choice of metrics is usually the most important aspect (52). We believe that this “structural
252
uncertainty” (53) is best addressed by focusing on a limited, but transparent and readily interpretable set of
253
indicators, rather than by quantifying uncertainties or increasing complexity. Finally, while we have
254
considered European fisheries targeting fish stocks that span the Mediterranean Sea, we have not
255
incorporated coastal communities in African countries that also fish on these same stocks. The relatively
256
low GDP per capita of these communities suggests that they would have correspondingly high regional
257
vulnerabilities and therefore correspondingly high climate risk profiles but it is not possible to draw robust
258
conclusions in the absence of appropriate data sets: the population-level hazards generated here (Table S7)
259
could be readily applied to aid such analyses in the future.
260
This study has shown that even though climate risk to European countries is, on average, moderate
261
compared to many other countries across the globe (6, 21), major differences exist across the European
262
continent. This corroborates with fine-scale spatial differences among fishing communities documented in
263
eastern North America (13, 54) and the Caribbean (30, 55), each requiring very different adaptation actions.
264
Our detailed analyses allow a distinction between climate hazard, exposure and vulnerability as key sources
265
of climate risk to fleets and coastal regions, and highlight where (and which) adaptation measures can have
266
greatest impact in increasing resilience, given limited financial resources.
267
Acknowledgements
268
This project received funding from the European Union’s Horizon 2020 research and innovation
269
programme under grant agreement No 678193 (CERES – Climate change and European Aquatic
270
Resources). The results generated by this analysis can be explored using an online tool available at
271
https://markpayne.shinyapps.io/CERES_climate_risk/ Source code is available at
272
https://github.com/markpayneatwork/CERES_vulnerability. "Fishing Boat", "Urban" and "Thermometer"
273
icons in Figure 4 by smalllikeart from www.flaticon.com.
274
Methods
275
General approach
276
We have applied an integrated approach to a climate risk assessment (CRA) across the European fisheries
277
sector. The CRA has three major components (Figure 4; Figure S2). The first and most fundamental of
278
these is the population hazard component, where the hazard associated with climate change impacts on
279
individual fish populations is quantified. We then use these hazard metrics as inputs into two parallel
280
climate risk assessments focussing on coastal regions and fishing fleets in turn. In each of these cases, the
281
population hazard is integrated up to the region or fleet level based on information about the relative
282
importance of each fish population to that unit to form the region- or fleet-specific hazards. These hazard
283
data are then complemented with region- and fleet-focused exposure and vulnerability metrics to produce
284
a climate risk for each. Finally, we combine the risks from each component into a comparative analysis
285
across nations.
286
287
Figure 4 Schematic diagram illustrating the approach used here to estimate climate risk in European fishery-dependent coastal
288
regions and fishing fleets. Species traits and population specific analyses of the thermal safety margin are combined to give a
289
population-specific climate hazard. This hazard then forms the basis for the region and fleet level CRAs, based on the combination
290
of hazard, exposure and vulnerability. Finally, the region and fleet risks are combined again into a comparative analysis. A detailed
291
flow diagram is presented in the supplementary material (Figure S2).
292
Scope and Data Sources
293
We aimed to assess the climate risk for the European marine fisheries sector, including all 22 EU countries
294
with marine borders, the United Kingdom, Norway, Iceland and Turkey. We based our analysis primarily
295
on catch data from FAO Areas 21, 27, 34 and 37 held in the EUROSTAT database (Table S1), excluding
296
distant water fleets. While this database covers more than 1200 species, many of these are economically
297
minor. We therefore aimed to cover the largest 90% of the value of the marine fish and shellfish sector in
298
each country and across Europe as a whole. Two species predominately inhabiting freshwater, European
299
perch (Perca fluviatilis) and pike-perch (Sander lucioperca), were removed from the database. Alternative
300
(or misspelled) scientific names were corrected where we could identify these (following World Register
301
of Marine Species, WoRMS) (Table S3).
302
Regional analyses were performed for European coastal regions based on NUTS2 statistical units. Sub-
303
national indicators of landings composition were derived from monthly harbour-level “first-sales” data
304
from the EU Market Observatory for Fisheries and Aquaculture (EUMOFA) (Table S1). In cases where
305
this data covered more than one NUTS2 unit within a country (10 countries), the harbour data was
306
aggregated up to NUTS2 units based on the geographical coordinates of the harbours. Where EUMOFA
307
data coverage was insufficient, the coastal NUTS2 units of that country were merged into one “region”
308
(Table S5) and EUROSTAT national landings data were assigned to it (Table S1). Socio-economic data for
309
the NUTS2 units was also obtained from EUROSTAT and integrated up to our “regions”, if relevant.
310
The Annual Economic Report (AER) provided by the EU Scientific, Technical and Economic Committee
311
for Fisheries (STECF) (23) formed the basis of the fishing fleet analysis (Table S1). This dataset has the
312
advantage of providing a single coherent source for fleet segments (the combination of fishing technique
313
and a vessel length category) across all of the European Union and United Kingdom: however, it does not
314
include data on fleets from Norway, Iceland or Turkey, and in the absence of comparable datasets, these
315
countries were not included in this part of the analysis.
316
All data was averaged over the period 2010-2018, where available.
317
Hazard Metrics
318
The hazard dimension of our CRA measures the strength and severity of climate change on the unit of
319
interest: in this case, fish populations in European waters. Many previous CVAs and CRAs do not
320
distinguish between the positive and negative effects of climate change, and simply highlight elements of
321
their study system that will change, making interpretation difficult. In contrast, and following the IPCC’s
322
definition of risk in relation to an “adverse event” (5), we focus explicitly on “negative” impacts in order
323
to have an unambiguous interpretation. We consider the hazard due to climate change impacts on living
324
marine resources as being the combination of both species-specific and population-specific processes as
325
follows.
326
Species-specific processes
327
A trait-based approach was employed to characterise the hazard of a species to climate change. Such an
328
approach is well established in climate risk and vulnerability analyses (15, 17, 28), due to its ability to draw
329
on general understanding of the response of species to climate change. Trait data was collated from
330
previously published databases (56–59) and complemented with data from Fishbase (60) and Sealifebase
331
(61) (accessed April-July 2019) (Table S1). Of the original set of species from EUROSTAT, 24 taxa were
332
only at the genus level, and appropriate trait sets were therefore identified based on ‘exemplar species’: in
333
some cases different exemplar species were used for the North Atlantic (FAO Area 27) and Mediterranean
334
regions (FAO Area 37) (Table S2). Barnacles (Pollicipes pollicipes) and solen razor clams (Solen spp.)
335
were also removed owing to a lack of biological traits data and difficulties identifying suitable exemplar
336
species.
337
Trait selection aimed to avoid double-counting information due to inclusion of correlated traits, a commonly
338
overlooked issue (57) that impacts many published analyses (15, 28, 29, 33). For example, smaller fish are
339
typically planktivorous, live shorter and grow faster, giving a high correlation between maximum length,
340
lifespan, growth rates and trophic level. Lifespan is the most commonly available of these metrics and was
341
therefore chosen as an exemplar for this set of traits. Shorter lifespans are associated with seasonal and
342
variable environments (57), implying robustness to change and variability, paralleling the approach used in
343
other studies (15, 28, 29, 33).
344
A “habitat specificity” metric was also developed. Species with spatially restricted habitat requirements
345
during part or all of their life-history are recognised as being more sensitive to disruption (62, 63). In
346
addition, mobile species have the ability to move rapidly to avoid unfavourable conditions in a way that
347
sedentary species do not, and therefore have a lower climate hazard (30). Traits defining the mobility, and
348
vertical and horizontal habitats were therefore collated into a single “habitat-specificity score” (Table 2).
349
The final set of traits is included as supplementary material (Table S6).
350
Table 2 Combination of mobility, vertical and horizontal habitat traits to generate a habitat specificity score. Trait categories
351
follow the scheme of Engelhard et al (56).
352
Habitat Specificity Mobility
Vertical habitat
Horizontal habitat
Low (0.00)
Highly migratory species
Any
Any
Mobile
Any
Oceanic
Mobile
Bathydemersal
Mesopelagic
Slope
Medium (0.33)
Mobile
Unknown
Benthopelagic
Demersal
Pelagic
Epipelagic
Slope
Shelf
Outer shelf
Unknown
Bathydemersal
Slope
Mobile
Bathydemersal
Outer shelf
High (0.67)
Mobile
(catadromous/anadromous)
Pelagic
Any
Mobile
Demersal
Inner shelf
Mobile
Benthopelagic
Coastal
Very high (1.00)
Sedentary
Any
Any
Mobile
Reef-associated
Any
353
Population-specific processes
354
The stress a fish population experiences as the ocean warms depends on the amount of warming, a
355
commonly employed metric of exposure in CVAs (6, 15). However, the physiological context of this
356
warming is also critical but often overlooked. For example, cod (Gadus morhua) in the North Sea are close
357
to their upper thermal limit, and will therefore experience negative impacts of warming, while cod in the
358
Barents Sea are far from this limit and will experience little or no negative effects of the same amount of
359
warming (64). Such a spatial and physiological context of warming is often overlooked in many CRAs and
360
CVAs, yet is critical to differentiate the climate hazard between different populations of the same species.
361
We resolve this problem in two ways. We first perform our analysis at the “population” level, defined as
362
the combination of species and FAO subarea e.g., cod in subarea 27.4 (North Sea). Note that while this
363
approach is similar to that used to manage many European fish stocks, we explicitly avoid the use of the
364
term “stock” to refer to this unit of analysis, as it has clear implications in fisheries management but is not
365
always the same as our definition “population”. Populations comprising less than 5% of the total catch of
366
the species were excluded from the analysis. Secondly, we place the degree of warming experienced by
367
these populations in a physiological context using the thermal-safety margin (TSM) (26, 27, 65, 66). TSM
368
is defined as the difference between the maximum temperature that the species can sustain and the
369
temperature of the environment: high TSMs indicate a high capacity to tolerate warming. Population-
370
specific TSMs therefore permit a fine-grained measure of the warming-related hazard.
371
We derived population-specific TSM metrics from the habitat models, parameters and maps provided by
372
Aquamaps www.aquamaps.org (67) (Table S1). We downloaded “native distribution maps” from the
373
Aquamaps website for the species selected above: where multiple maps were available, choice was guided
374
by the internal map quality ranking system. For the invasive species purple whelk (Rapana venosa),
375
originally from waters around Japan, Korea and China but now supporting a large fishery in the Black Sea,
376
the “Suitable Habitat map” was used. From each species’ map we used the “90th percentile” parameter for
377
the temperature response as an estimate of its upper thermal tolerance. Temperatures in a subarea were
378
based on the data underpinning the Aquamaps model (NOAA NCEP Climatology, 1982-1999) (67),
379
ensuring congruence between the tolerance parameters and the temperature data. Sea-surface or -bottom
380
temperature data, as used in generating the species’ Aquamap, were masked using the habitat model to
381
eliminate unsuitable habitat for each individual species (Figure 5). Projected temperatures changes from
382
1999 to 2050 under the SRES A2 scenario were also available in this dataset and extracted for each
383
population in the same manner for use in supporting analyses (Figure S3). Population-specific TSM was
384
calculated as the median difference between the species’ “90th percentile” parameter and temperature across
385
all valid pixels in that subarea.
386
387
Figure 5 Use of Aquamaps to calculate TSM metrics. Atlantic cod (Gadus morhua) as an example. Environmental data and
388
species thermal tolerance data from Aquamaps are used to calculate the thermal safety margin (TSM) for this species (coloured
389
pixels) and masked using the habitat model to limit data to modelled regions of occurrence. Median TSM values are then calculated
390
within each FAO subarea defining a population (grey polygons, blue labels).
391
Population-level hazard
392
Hazard metrics were combined based on their relative ranking for each population. We chose to give equal
393
weight to the species (lifespan, habitat-specificity) and population-level (TSM) aspects of the analysis when
394
combining the metrics: after converting to a rank percentile, a weight of 0.25 was given to the species’
395
lifespan (shorter-lifespans give a low hazard), 0.25 for the species’ habitat-specificity (low specificity gives
396
a low hazard) and 0.5 to the population TSM (high TSMs give a low hazard). Equal weighting of the metrics
397
(0.33 / 0.33 / 0.33) was also considered (68) but the resulting hazard metrics were found to be strongly
398
correlated with the original (0.25 / 0.25 / 0.50) weighting (Spearman correlation coefficient of 0.95; Figure
399
S4), indicating that the relative hazard ranking of individual populations under the two schemes is very
400
similar.
401
Population-level hazard scores were integrated up to coastal region and fishing-fleet levels. In the case of
402
the fleet analysis, this was based on the relative composition (by value) of the populations that each fleet
403
fishes on, while in the case of the coastal region analysis it was based on the composition (by value) of
404
landings in that region (Figure 4, Figure S2).
405
Exposure metrics
406
We define exposure as an indicator of how sensitive a coastal region or fishing fleet is to changes in the
407
fish populations it is dependent on. Fleets or coastal regions have lower exposure (higher resilience) if they
408
catch a wide range of different fish species, rather than concentrating on a specific resource (30, 31, 41). If
409
one species is reduced or lost due to the effects of climate change, the impact of that loss is relatively less
410
severe for fleets and coastal regions that are dependent on a broad portfolio of species. We therefore defined
411
our exposure metrics following this logic, using two different metrics to characterise diversity of catch or
412
landings: i) the Shannon diversity index, one of the most commonly used diversity indices in ecology and
413
ii) Simpson’s dominance index, a statistic that emphasizes the relative abundance of the most common
414
species in the sample (30).
415
For coastal regions, exposure metrics were based on the value of landings data from EUMOFA and
416
EUROSTAT (Table S1; Figure S2). While EUROSTAT data is species resolved, EUMOFA data is
417
organised in approximately 100 “main commercial species” (MCS) groupings: we therefore harmonised
418
the two datasets by aggregating EUROSTAT data to the MCS groupings based on correlation keys provided
419
by EUMOFA. The Shannon and Simpson metrics were then calculated to estimate the diversity of MCS
420
groups.
421
For fleet segments, the value of landings is available by species code from the STECF Annual Economic
422
Report (23). The two diversity indices could therefore be calculated directly to quantify the diversity of
423
species.
424
In both cases, the exposure index was produced as a composite index of the two indices described above
425
by averaging the percentile ranks and then re-calculating percentile ranks again.
426
Vulnerability metrics
427
Vulnerability in this setting refers to the resilience of the analysis unit (either a coastal region or a fleet) and
428
its ability to mitigate the hazard via adaptation.
429
The region vulnerability metric was based on the gross-domestic product per capita of the region, as
430
calculated from EUROSTAT data at the NUTS2 level (Table S1). Regions with high GDP per capita were
431
viewed as having a high adaptive capacity and therefore low vulnerability. Regional vulnerability was
432
calculated as the percentile rank of this statistic.
433
Fleet segment vulnerability was based on the net profit margin (NPM). This is a standard economic metric,
434
defined as net profit (i.e. revenue minus variable, fixed and opportunity costs) divided by the total revenue:
435
it therefore represents how much of the total income generated by the fleet is net profit (23). NPM has the
436
feature of taking into account many of the different factors that influence the profitability of the fleet, and
437
is also scale independent (as profitability is divided by the revenue), allowing comparison of both large and
438
small segments. NPM was calculated for each fleet segment based on economic data from the STECF
439
Annual Economic Report (23) (Table S1), and the vulnerability score generated based on percentile rank.
440
Fleet segments with high profitability were viewed as being less vulnerable to the effects of climate change,
441
as they could absorb the potential loss associated with a climate change having a negative impact on their
442
target species.
443
Climate risk metrics
444
For each of the coastal regions, and for each of the fleet segments, the overall climate risk was calculated
445
as the unweighted mean of the hazard, exposure and vulnerability percentile ranks.
446
447
References
448
1.
R. M. Hassan, R. Scholes, N. Ash, Ecosystems and Human Well-being - Current State and Trends:
449
Findings of the Condition and Trends Working Group of the Millennium Ecosystem Assessment.
450
Millenn. Ecosyst. Assess. Ser. (v. 1), xxi, 917 (2005).
451
2.
E. S. Poloczanska, et al., Global imprint of climate change on marine life. Nat. Clim. Chang. 3, 919–
452
925 (2013).
453
3.
B. Phillips, M. Pérez-Ramírez, Climate Change Impacts on Fisheries and Aquaculture: A Global
454
Analysis (John Wiley and Sons Inc., 2018).
455
4.
FAO, Impacts of climate change on fisheries and aquaculture (2018).
456
5.
IPCC, IPCC Special Report on the Ocean and Cryosphere in a Changing Climate (2019)
457
https:/doi.org/https://www.ipcc.ch/report/srocc/.
458
6.
E. H. Allison, et al., Vulnerability of national economies to the impacts of climate change on
459
fisheries. Fish Fish. 10, 173–196 (2009).
460
7.
J. E. Johnson, et al., Assessing and reducing vulnerability to climate change: Moving from theory
461
to practical decision-support. Mar. Policy 74, 220–229 (2016).
462
8.
M. Oppenheimer, et al., “Emergent risks and key vulnerabilities” in Climate Change 2014: Impacts,
463
Adaptation, and Vulnerability. Part A: Global and Sectoral Aspects. Contribution of Working Group
464
II to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, C. Field, et
465
al., Eds. (Cambridge University Press, 2014), pp. 1039–1099.
466
9.
A. Connelly, J. Carter, J. Handley, S. Hincks, Enhancing the practical utility of risk assessments in
467
climate change adaptation. Sustain. 10, 1–12 (2018).
468
10.
O. D. Cardona, et al., “Determinants of risk: Exposure and vulnerability” in Managing the Risks of
469
Extreme Events and Disasters to Advance Climate Change Adaptation: A Special Report of Working
470
Groups I and II of the Intergovernmental Panel on Climate Change (IPCC)., C. B. Field, et al., Eds.
471
(Cambridge University Press, 2012), pp. 65–108.
472
11.
N. W. Adger, Social Vulnerability to Climate Change and Extremes in Coastal Vietnam. World Dev.
473
27, 249–269 (1999).
474
12.
J. E. Cinner, et al., Evaluating Social and Ecological Vulnerability of Coral Reef Fisheries to Climate
475
Change. PLoS One 8, e74321 (2013).
476
13.
L. L. Colburn, et al., Indicators of climate change and social vulnerability in fishing dependent
477
communities along the Eastern and Gulf Coasts of the United States. Mar. Policy 74, 323–333
478
(2016).
479
14.
J. A. Ekstrom, et al., Vulnerability and adaptation of US shellfisheries to ocean acidification. Nat.
480
Clim. Chang. 5, 207–214 (2015).
481
15.
J. A. Hare, et al., A Vulnerability Assessment of Fish and Invertebrates to Climate Change on the
482
Northeast U.S. Continental Shelf. PLoS One 11, e0146756 (2016).
483
16.
E. A. Fulton, et al., Decadal scale projection of changes in Australian fisheries stocks under climate
484
change. (CSIRO, Australia, 2017).
485
17.
G. T. Pecl, et al., Rapid assessment of fisheries species sensitivity to climate change. Clim. Change
486
127, 505–520 (2014).
487
18.
J. D. Bell, J. E. Johnson, A. J. Hobday, Vulnerability of tropical pacific fisheries and aquaculture to
488
climate change (Secretariat of the Pacific Community, 2011).
489
19.
W. Barsley, C. De Young, C. Brugere, Vulnerability assessment methodologies: an annotated
490
bibliography for climate change and the fisheries and aquaculture sector. FAO Fish. Aquac. Circ.
491
No. 1083 (2013).
492
20.
M. Barange, et al., Impacts of climate change on marine ecosystem production in societies
493
dependent on fisheries. Nat. Clim. Chang. 4, 211–216 (2014).
494
21.
R. Blasiak, et al., Climate change and marine fisheries: Least developed countries top global index
495
of vulnerability. PLoS One 12, 1–15 (2017).
496
22.
C. Brugere, C. De Young, Assessing climate change vulnerability in fisheries and aquaculture
497
Available methodologies and their relevance for the sector (FAO Fisheries and Aquaculture
498
Technical Paper 597. FAO, 2015).
499
23.
STECF, The 2018 Annual Economic Report on the EU Fishing Fleet (STECF 18-07), N. Carvalho,
500
M. Keatinge, J. Guillen, Eds. (Publications Office of the European Union, 2018)
501
https:/doi.org/10.2760/56158.
502
24.
M. A. Peck, J. K. Pinnegar, “Chapter 5: Climate change impacts, vulnerabilities and adaptations:
503
North Atlantic and Atlantic Arctic marine fisheries” in Impacts of Climate Change on Fisheries and
504
Aquaculture: Synthesis of Current Knowledge, Adaptation and Mitigation Options., (FAO Fisheries
505
and Aquaculture Technical Paper No. 627. FAO, 2018), pp. 87–111.
506
25.
F. Natale, N. Carvalho, M. Harrop, J. Guillen, K. Frangoudes, Identifying fisheries dependent
507
communities in EU coastal areas. Mar. Policy 42, 245–252 (2013).
508
26.
C. A. Deutsch, et al., Impacts of climate warming on terrestrial ectotherms across latitude. Proc.
509
Natl. Acad. Sci. 105, 6668–6672 (2008).
510
27.
J. M. Sunday, et al., Thermal-safety margins and the necessity of thermoregulatory behavior across
511
latitude and elevation. Proc. Natl. Acad. Sci. U. S. A. 111, 5610–5615 (2014).
512
28.
M. C. Jones, W. W. L. Cheung, Using fuzzy logic to determine the vulnerability of marine species
513
to climate change. Glob. Chang. Biol. 24, e719–e731 (2018).
514
29.
W. W. L. Cheung, T. J. Pitcher, D. Pauly, A fuzzy logic expert system to estimate intrinsic extinction
515
vulnerabilities of marine fishes to fishing. Biol. Conserv. 124, 97–111 (2005).
516
30.
J. K. Pinnegar, G. H. Engelhard, N. J. Norris, D. Theophille, R. D. Sebastien, Assessing vulnerability
517
and adaptive capacity of the fisheries sector in Dominica: long-term climate change and catastrophic
518
hurricanes. ICES J. Mar. Sci. 76, 1353–1367 (2019).
519
31.
T. J. Cline, D. E. Schindler, R. Hilborn, Fisheries portfolio diversification and turnover buffer
520
Alaskan fishing communities from abrupt resource and market changes. Nat. Commun. 8, 1–7
521
(2017).
522
32.
M. Lindegren, K. Brander, Adapting Fisheries and Their Management To Climate Change: A
523
Review of Concepts, Tools, Frameworks, and Current Progress Toward Implementation. Rev. Fish.
524
Sci. Aquac. 26, 400–415 (2018).
525
33.
W. W. L. Cheung, M. C. Jones, G. Reygondeau, T. L. Frölicher, Opportunities for climate-risk
526
reduction through effective fisheries management. Glob. Chang. Biol. 24, 5149–5163 (2018).
527
34.
K. F. Drinkwater, et al., On the processes linking climate to ecosystem changes. J. Mar. Syst. 79,
528
374–388 (2010).
529
35.
B. Planque, et al., How does fishing alter marine populations and ecosystems sensitivity to climate?
530
J. Mar. Syst. 79, 403–417 (2010).
531
36.
J. D. Bell, et al., Effects of climate change on oceanic fisheries in the tropical Pacific: implications
532
for economic development and food security. Clim. Change 119, 199–212 (2013).
533
37.
M. L. Pinsky, N. Mantua, Emerging Adaptation Approaches for Climate-Ready Fisheries
534
Management. Oceanography 27, 146–159 (2014).
535
38.
M. R. Payne, et al., Lessons from the First Generation of Marine Ecological Forecast Products.
536
Front. Mar. Sci. 4 (2017).
537
39.
A. J. Hobday, et al., A Framework for Combining Seasonal Forecasts and Climate Projections to
538
Aid Risk Management for Fisheries and Aquaculture. Front. Mar. Sci. 5, 1–9 (2018).
539
40.
M. L. Pinsky, et al., Fish and fisheries in hot water: What is happening and how do we adapt? Popul.
540
Ecol. (2020) https:/doi.org/10.1002/1438-390X.12050.
541
41.
T. Young, et al., Adaptation strategies of coastal fishing communities as species shift poleward. ICES
542
J. Mar. Sci. 76, 93–103 (2019).
543
42.
A. L. Perry, Climate Change and Distribution Shifts in Marine Fishes. Science (80-. ). 308, 1912–
544
1915 (2005).
545
43.
S. D. Simpson, et al., Continental shelf-wide response of a fish assemblage to rapid warming of the
546
sea. Curr. Biol. 21, 1565–70 (2011).
547
44.
A. R. Baudron, et al., Changing fish distributions challenge the effective management of European
548
fisheries. Ecography (Cop.). 43, 494–505 (2020).
549
45.
J. K. Pinnegar, P. J. Wright, K. Maltby, A. Garrett, The impacts of climate change on fisheries,
550
relevant to the coastal and marine environment around the UK. MCCIP Sci. Rev. 2020, 456–481
551
(2020).
552
46.
S. Kasperski, D. S. Holland, Income diversification and risk for fishermen. Proc. Natl. Acad. Sci. U.
553
S. A. 110, 2076–2081 (2013).
554
47.
M. L. Pinsky, et al., Preparing ocean governance for species on the move. Science (80-. ). 360, 1189–
555
1191 (2018).
556
48.
A. McIlgorm, et al., How will climate change alter fishery governance? Insights from seven
557
international case studies. Mar. Policy 34, 170–177 (2010).
558
49.
M. C. Badjeck, E. H. Allison, A. S. Halls, N. K. Dulvy, Impacts of climate variability and change
559
on fishery-based livelihoods. Mar. Policy 34, 375–383 (2010).
560
50.
J. E. Cinner, et al., Building adaptive capacity to climate change in tropical coastal communities.
561
Nat. Clim. Chang. 8, 117–123 (2018).
562
51.
P. D. Spencer, A. B. Hollowed, M. F. Sigler, A. J. Hermann, M. W. Nelson, Trait‐based climate
563
vulnerability assessments in data‐rich systems: An application to eastern Bering Sea fish and
564
invertebrate stocks. Glob. Chang. Biol. 25, 3954–3971 (2019).
565
52.
I. Monnereau, et al., The impact of methodological choices on the outcome of national-level climate
566
change vulnerability assessments: An example from the global fisheries sector. Fish Fish. 18, 717–
567
731 (2017).
568
53.
M. R. Payne, et al., Uncertainties in projecting climate-change impacts in marine ecosystems. ICES
569
J. Mar. Sci. 73, 1272–1282 (2016).
570
54.
L. A. Rogers, et al., Shifting habitats expose fishing communities to risk under climate change. Nat.
571
Clim. Chang. 9, 512–516 (2019).
572
55.
I. Monnereau, et al., Vulnerability of the fisheries sector to climate change impacts in Small Island
573
Developing States and the Wider Caribbean (CERMES Technical Report No77. University of the
574
West Indies. Barbados, 2015).
575
56.
G. H. Engelhard, J. R. Ellis, M. R. Payne, R. ter Hofstede, J. K. Pinnegar, Ecotypes as a concept for
576
exploring responses to climate change in fish assemblages. ICES J. Mar. Sci. 68, 580–591 (2011).
577
57.
L. Pecuchet, et al., From traits to life-history strategies: Deconstructing fish community composition
578
across European seas. Glob. Ecol. Biogeogr. 26, 812–822 (2017).
579
58.
E. Beukhof, et al., Marine fish traits follow fast-slow continuum across oceans. Sci. Rep. 9, 17878
580
(2019).
581
59.
E. Beukhof, T. S. Dencker, M. L. D. Palomares, A. Maureaud, A trait collection of marine fish
582
species from North Atlantic and Northeast Pacific continental shelf seas. Pangaea, 1–12 (2019).
583
60.
R. Froese, D. Pauly, Fishbase (2019).
584
61.
M. L. D. Palomares, D. Pauly, Sealifebase (2019).
585
62.
P. Petitgas, et al., Impacts of climate change on the complex life cycles of fish. Fish. Oceanogr. 22,
586
121–139 (2013).
587
63.
A. D. Rijnsdorp, M. A. Peck, G. H. Engelhard, C. Möllmann, J. K. Pinnegar, Resolving the effect
588
of climate change on fish populations. ICES J. Mar. Sci. 66, 1570–1583 (2009).
589
64.
K. F. Drinkwater, The response of Atlantic cod (Gadus morhua) to future climate change. ICES J.
590
Mar. Sci. 62, 1327–1337 (2005).
591
65.
F. T. Dahlke, S. Wohlrab, M. Butzin, H.-O. Pörtner, Thermal bottlenecks in the life cycle define
592
climate vulnerability of fish. Science (80-. ). 369, 65–70 (2020).
593
66.
M. L. Pinsky, A. M. Eikeset, D. J. McCauley, J. L. Payne, J. M. Sunday, Greater vulnerability to
594
warming of marine versus terrestrial ectotherms. Nature (2019) https:/doi.org/10.1038/s41586-019-
595
1132-4.
596
67.
K. Kesner-Reyes, K., et al., AquaMaps: algorithm and data sources for aquatic organisms. FishBase.
597
World Wide Web Electron. Publ. (2012) (May 1, 2020).
598
68.
N. E. Hamilton, M. Ferry, ggtern: Ternary diagrams using ggplot2. J. Stat. Softw. 87 (2018).
599
600
| 2021 | Climate risk to European fisheries and coastal communities | 10.1101/2020.08.03.234401 | [
"Payne Mark R.",
"Kudahl Manja",
"Engelhard Georg H.",
"Peck Myron A.",
"Pinnegar John K."
] | creative-commons |
1
Molecular signatures of resource competition: clonal interference
1
drives the emergence of ecotypes
2
Massimo Amicone+, * and Isabel Gordo+, *
3
+Instituto Gulbenkian de Ciência (IGC)
4
*corresponding authors: mamicone@igc.gulbenkian.pt, igordo@igc.gulbenkian.pt
5
6
Short running title: Clonal interactions drive ecotypes formation.
7
Keywords: Eco-evolutionary dynamics, rapid adaptation, clonal interference, resource
8
competition, competitive exclusion, diversification, community assembly.
9
Type of article: Letters.
10
Content (number of words/items): Abstract (141), Main text (4990), References (70),
11
Figure legends (6).
12
13
Author contributions
14
MA and IG designed the study, all authors wrote the manuscript and provided final
15
approval for publication.
16
Conflict of interest
17
The authors declare no conflict of interest.
18
Data accessibility
19
Should the manuscript be accepted, the data supporting the results will be archived in
20
an appropriate public repository (Figshare) and the data DOI will be included at the end
21
of the article.
22
2
Abstract
23
Microbial ecosystems harbor an astonishing diversity that can persist for long times. To
24
understand how such diversity is generated and maintained, ecological and
25
evolutionary processes need to be integrated at similar timescales, but this remains a
26
difficult challenge. Here, we extend an ecological model of resource competition to
27
allow for evolution via de novo mutation, focusing on large and rapidly adapting asexual
28
populations. Through numerical and analytical approaches, we characterize adaptation
29
and diversity at different levels and show how clonal interference – the interaction
30
between simultaneously emerging lineages – shapes the eco-evolutionary dynamics. We
31
find that large mutational inputs can foster diversification under sympatry, increasing
32
the probability that phenotypically and genetically distinct clusters arise and stably
33
coexist, constituting an initial form of community. Our findings have implications
34
beyond microbial populations, providing novel insights about the interplay between
35
ecology and evolution in clonal populations.
36
37
3
Introduction
38
Understanding the mechanisms behind the evolution of biodiversity and the formation
39
of communities remains a difficult challenge. One must integrate ecology and evolution
40
over similar timescales, as taken together, they can give rise to phenomena that could
41
not be explained by either alone (Schoener, 2011). The competitive exclusion principle
42
(Hardin, 1960) theoretically bounds the number of species by the number of limiting
43
resources. This principle generated an apparent contradiction between theoretical
44
expectations and observations, as ecosystems can be replete of diversity even in
45
limiting environments, both in nature (Hutchinson, 1961; Tilman, 1982; Huston, 1994)
46
and in more controlled laboratory conditions (Maharjan et al., 2006; Gresham et al.,
47
2008; Kinnersley, Holben and Rosenzweig, 2009; Herron and Doebeli, 2013). Much
48
theoretical work has been done to resolve such controversy, often referred to as the
49
“paradox of the plankton” (Hutchinson, 1961). Different ecological mechanisms were
50
proposed to maintain diversity, including trade-offs on the species’ traits (Posfai,
51
Taillefumier and Wingreen, 2017), heterogeneity in space (Abrams, 1988) and time
52
(Litchman and Klausmeier, 2001) or gene regulation (Pacciani-Mori et al., 2020).
53
However, solutions often rely on specific conditions, where a small change in the
54
parameters could cause a collapse of the species richness. What amount of diversity can
55
be maintained if the species’ traits are not constant but rather evolve? is a critical
56
question that remains to be answered.
57
Ecological theories have been challenged by evolutionary processes (Geritz et al., 1998).
58
In eco-evolutionary frameworks, mutations generate new genetic variants whose fate
59
depends on the state of the ecosystem and, in turn, their increase in frequency can alter
60
the populations. A common outcome of such eco-evolutionary feedbacks is that
61
4
evolution limits diversity by reducing the effectiveness of coexistence mechanisms
62
(Edwards et al., 2018). The diversity that could be possible by ecological principles
63
alone, is reduced by selection of the fittest and competitive exclusion. Although several
64
studies have produced novel understanding on the evolution of diversity (Dieckmann
65
and Doebeli, 1999; Shoresh, Hegreness and Kishony, 2008; Doebeli, 2011; Kremer and
66
Klausmeier, 2017), the majority of them rely on a timescale separation between
67
ecological and evolutionary processes, i.e. on the strong-selection-weak-mutation
68
assumption. The emergence of beneficial mutations is assumed to be much slower than
69
the ecological dynamics, thus, before a new lineage arises, the population has already
70
reached ecological equilibrium. While allowing for analytical tractability this
71
assumption comes at a cost: it neglects the overlap between multiple evolving lineages –
72
clonal interference – a phenomena that has been extensively observed in microbial
73
communities in vitro and in vivo (Desai, Fisher and Murray, 2007; Barroso-Batista et al.,
74
2014). Population genetics models incorporating clonal interference have generated
75
predictions for the adaptation rate, fixation probabilities and genetic diversity of a
76
population (Gerrish and Lenski, 1998; Park and Krug, 2007; Good et al., 2012; de Sousa
77
et al., 2016), yet excluding ecological interactions. When the population size (N) and/or
78
the mutation rate (U) are not small (NU >>1), multiple lineages can emerge
79
simultaneously, ecologically interact with each other and evolve in non-trivial ways.
80
Although these processes are inevitably intertwined (Lawrence et al., 2012; Barroso-
81
Batista et al., 2014, 2020; Garud et al., 2019), little was done to investigate how they act
82
in chorus.
83
Here, we study how clonal interference affects the interplay between ecological and
84
evolutionary processes and structures communities. Following a model of competition
85
5
for resources (Posfai, Taillefumier and Wingreen, 2017), we build an eco-evolutionary
86
framework, where no restrictive assumption on the timescales is made, as common in
87
other models (Geritz et al., 1998; Shoresh, Hegreness and Kishony, 2008; Good, Martis
88
and Hallatschek, 2018). We follow an initially isogenic population throughout time and
89
show how different patterns of adaptation and diversity emerge. Every generation,
90
multiple mutations arise simultaneously in the population and interactions between
91
clones, which occur via competition, affect both the sign and the strength of selection.
92
To our knowledge this is the first study to investigate adaptation under competition for
93
resources across phenotypic and genetic diversity and to characterize the probability of
94
diversification under different molecular evolution parameters. We describe how a
95
large amount of variation can be maintained via a balance between selection and
96
mutation, and how ecotypes - functionally distinct clusters of genotypes - can emerge,
97
coexist and establish a first form of community.
98
99
Model and Methods
100
Eco-Evolutionary model
101
We model the dynamics of a single clonal lineage evolving to consume a set of different
102
substitutable resources, constantly replenished in a well-mixed environment.
103
Individuals mutate at a rate U and the fate of the emerging mutations depends on their
104
phenotypic effects, on the resource concentration, on the other individuals present in
105
the environment and on drift.
106
The underlying dynamics are determined by the MacArthur’s consumer resource model
107
(Mac Arthur, 1969), recently formalized by Posfai et al. to explain high levels of diversity
108
6
in a purely ecological context, i.e. in the absence of mutations (Posfai, Taillefumier and
109
Wingreen, 2017). Briefly, let M be the number of types present at time t with densities
110
(#cells/V) ni, (i=1…M) and R the number of substitutable resources with input
111
concentrations rj, (j=1…R). The expected density dynamics of each type are:
112
���
�� � ������∑
��
�����
∑
�����·��
���
����
���
� �
�
���
�
(1)
113
where ��
��� represents the consumption rate of resource j by type i and δ is the death
114
rate. The resource amounts are constant in this model since, as Posfai et al., we assume
115
that metabolic reactions occur much faster than cell division (Posfai, Taillefumier and
116
Wingreen, 2017).
117
We also assume a finite amount of energy available for each cell and limit their ability of
118
consuming resources by an energetic constraint (E):
119
0 � ∑
��
��� � �
�
���
, �� � 1, … , �
(2)
120
A trade-off has been considered in previous studies as a fixed energy budget (Posfai,
121
Taillefumier and Wingreen, 2017; de Oliveira, Amado and Campos, 2018; Amado and
122
Campos, 2019), but in our model it acts as an upper limit. Assuming equally supplied
123
resources ��� � � ��� and unitary energy, volume and death rate (E, V, δ=1), the
124
population size is N=Rr.
125
We model an initial isogenic population (M(t0) =1) with given traits ����� and allow for
126
mutations that change the heritable traits and give rise to new genotypes. Every
127
generation, each genotype i (i=1…M(t)) will generate a Poisson number of mutants with
128
expected value
129
��#����� � ����� � �
7
with U being the per-genome, per-generation rate of non-lethal mutations. Assuming an
130
infinite site model, a mutation on genotype i will result into a new individual with
131
unique genotype (i') whose phenotypes differ from the parental traits as:
132
������ � ����� !�,
!� " #�
For simplicity, we assume the Δj to be normally distributed. The mutation effects are
133
defined as follows:
134
!�: %!� & '�0, (�, � sampled from 31,… , 45
!� & '�0,6 � (�, for all 7 8 �
9
If ρ =1, a mutation changes all the traits with equal probability. If 0< ρ <1, mutations
135
target one trait (randomly sampled with probability 1/R), but also alter partially the
136
others. If ρ =0, a mutation only changes one trait. The parameter ρ modulates different
137
degrees of trait interdependence or equivalently the pleiotropic effect of mutations,
138
while the parameter σ modulates the magnitude of the mutation effects (see Fig. 1B).
139
In order to respect the boundary condition (2), we assumed that: i) mutations leading to
140
negative values of αj are loss of function and thus assigned αj =0; ii) mutations that do
141
not respect the energy constraint cannot exist, therefore Δj are drawn until the upper
142
limit of (2) is satisfied.
143
In the limit of discrete time steps, we define the selection acting on genotype i at time t,
144
si(t), as the expected increase in abundance in the absence of drift, such that:
145
������ 1�� � ������1 :�����
and from (1):
146
:���� � ∑
��
�����
∑
�����·��
���
����
���
� �
�
���
.
(3)
147
8
Thus, the fate of each genotype depends on its ability to consume each of the resources
148
and on the ecosystem’s ecology.
149
To simulate drift, we draw the final abundances via multinomial sampling with
150
probabilities
����������
����
∀i=1…M(t). Every generation, the number of genotypes M(t+1) is
151
updated together with their abundances and traits. Along the simulations, we record the
152
genotypic and phenotypic composition of the population, from which classical statistics
153
used in population genetics are calculated as discussed below.
154
Phenotypic and genotypic diversity
155
In the two-resources case, within a population, each type i is characterized by a vector
156
of its consumption traits ���
���, ��
���� and a vector of the mutations that gave rise to it,
157
each with a unique identifier (e.g. the vector [1,2,7,10] represents genotype 10 whose
158
ancestors are, in order, genotypes 7,2 and 1, and 1 is the ancestor common to every
159
type). From this implementation we can reconstruct the entire phylogenetic tree of a
160
population at any time point and map it on the phenotypic space. To measure the
161
genetic diversity of a population we quantify the average pairwise genetic distance πG in
162
a sample of m individuals: ;���� �
∑
����,��
��,��
��
��
, where m=100 and dG(i,j) is the number of
163
mutations that separate genotype i from j. From πG and the number of segregating sites
164
in the sample, we further compute another population genetics statistic: Tajima’s D,
165
whose expectation under the simplest population genetics neutral model is zero
166
(Tajima, 1989). At the functional level, we compute the average pairwise phenotypic
167
distance πP, defined as: ;���� �
∑
����,��
��,��
��
��
�
�
� , where dE is the classical Euclidean
168
distance.
169
Neutral mutation model
170
9
To understand the dynamics in the absence of selection, we run simulations with only
171
drift. Genotypes acquire mutations with the same trait effects and probability as
172
described before, but their growth probabilities are equal and do not depend on the
173
phenotypes. Modelling the explicit αj under neutrality, instead of assuming that the
174
mutations have no effect (i.e. Δj=0), allows for a better comparison with the model of
175
selection. These neutral simulations were run for the same time as the selection case.
176
Even though this time is much smaller than that needed to accumulate neutral changes
177
in a classical Wright-Fisher neutral model (Wright, 1930; Fisher, 1958), the outcome
178
informs on how drift contributes to the patterns of molecular and phenotypic evolution
179
measured here.
180
Parameters of the numerical simulations
181
In each simulation, an initially maladapted (��
��� � ��
��� � 0.05) monomorphic
182
population undergoes the eco-evolutionary process described above for 10000
183
generations. Its genotypic and phenotypic compositions are analyzed along time, for the
184
following set of parameters: N=107, U: {10-8, 5⋅ 10-8, 10-7, 5⋅ 10-7, 10-6, 5⋅ 10-6, 10-5,5⋅ 10-5}, σ:
185
{0.0125,0.025,0.05}, ρ: {0,0.5,1}. The parameter combinations are studied under
186
selection or neutrality, with the exception of fixing σ=0.05 and ρ=0 in the latter, as by
187
changing these values, the outcome would be equivalent. Of these, each set was
188
simulated in 100 independent replicates to obtain the statistics of diversity. In order to
189
disentangle the role of the energetic constraint assumption, we simulated adaptation
190
under selection but without any boundary condition (i.e. E=+∞ in (2)), with parameters:
191
N=107, U: {10-8, 10-7, 10-6, 10-5}, σ: {0.0125,0.025,0.05}, ρ=0. When we tested additional
192
parameters, this is specified in the text. The algorithm was written in R (version 3.6.1)
193
and the results analyzed in RStudio.
194
10
Results
195
Competition-driven diminishing return and the rate of adaptation
196
The initially monomorphic population, which is poorly adapted, is expected to acquire
197
mutations that improve the ability to consume the available resources and to advance in
198
the phenotypic space towards better adapted states. Our aim is to identify what
199
influences the speed of this adaptive process.
200
We first obtain some analytical approximations on how the selective pressure changes
201
in time and with the genetic composition of the population, under the competition for
202
resources set by (1). Selection acting on an emerging genotype i is given by (3) and
203
depends on the population investment on each resource �: >�
������ ? ∑
����� · ��
���
����
���
.
204
Let us first simplify the problem by considering a monomorphic (M=1) population
205
whose phenotype (��) mirrors the resource input proportions:
��
∑ ��
�
�
�
��
∑ ��
�
�
, �� � 1,… , 4
206
which consists of the local optimal strategy for a given energetic investment (∑ ��
�
�
). In
207
the absence of mutations, such population at equilibrium satisfies:
208
>�
��� � �� · ∑
��
�
���
�� � 1, … , 4.
(4)
209
Now consider a mutant that emerges from this population with phenotypes
210
����� � �� !�. From (3) and (4) it follows that the selection acting on such mutant is:
211
:������, ∆BB�� �
∑ ∆�
�
�
∑ ��
�
�
.
(5)
212
While bigger steps result in stronger selection, equation (5) also implies that the same
213
mutations are subject to weaker selection when emerging on a better adapted
214
background. Thus, this system exhibits diminishing returns epistasis that emerges from
215
the competition dynamics. Because we sample the phenotypic changes from a normal
216
11
distribution (see Methods), their additive effect will also follow a normal distribution
217
∑ ∆�
�
�
& '�0, ( � where ( � (C�4 � 1� � 6� 1 ; therefore, stronger pleiotropic effects
218
lead to stronger selection. From the continuous univariate distribution theory (Johnson
219
et al., 1994) we can retrieve the expected beneficial mutation effect ���∑ ∆�
�
�
�� and,
220
from (5), compute the corresponding expected selection coefficients (��:��) for varying
221
values of ∑ ��
�
�
(see Appendix S1 in Supporting Information). Figure 2A shows how the
222
selection strength decreases for better adapted genetic backgrounds across different σ
223
and ρ conditions. It is important to note that the diminishing return epistasis in this
224
model is not due to the energy constraint; nonetheless, such boundary condition
225
truncates the distribution of mutation effects (see Fig. S1A and Appendix S1) and
226
further slows down the rate of adaptation of well adapted populations (inset in Fig. 2A).
227
Following (5), the proportion of beneficial mutations, computed as
228
D�:�� �
!
"
���∑
��
�
�
�
�
�#,� �
!
"
���∑
��
�
�
�
��
�#,� �
229
drops rapidly in well adapted populations as the energetic trade-off makes the
230
deleterious mutations more common (Fig. S1). Next, we tested how well the analytical
231
approximations – obtained by assuming monomorphic populations at consecutive
232
equilibria – predict the regimes of extensive clonal interference, where the adapting
233
populations are polymorphic and out of equilibrium. To do so, after running simulations
234
with NU=100 and two resources, we calculate the average population trait sum
235
�E��� : �
∑
$��
������
���%
����
�
&�����
"
, compute the expected beneficial selection coefficient as
236
��:�|�E� �
��'��'�|�)�
�)
and compare it with the mean beneficial selection observed in the
237
simulations, during the first 300 generations. The strength of selection acting on
238
12
polymorphic populations follows the predicted diminishing return pattern, but is often
239
underestimated (Fig. 2B). In fact, it decreases with the mean population phenotype �E,
240
but it increases with the population phenotype variance (Fig. S2).
241
Numerical simulations further allow us to link the strength of selection with the speed
242
of phenotypic adaptation: larger phenotypic changes imply stronger selection, resulting
243
in faster adaptation (Fig. 2C). We find that, when time is scaled by G
��:�|�E�H�E � (
�
�)��#�
,
244
the populations’ mean phenotype moves with similar velocity, demonstrating that both
245
the complex form of selection and the mutation type mediate the speed of adaptation
246
(see Fig. 2D and Appendix S1). Finally, the simulations show how larger mutation rates
247
further accelerate adaptation. Specifically, populations under intense clonal
248
interference (NUR1) can approach the phenotypic optimum rapidly (Fig. 2E).
249
In summary, these results show how the availability and the effect of beneficial
250
mutations, together with the complex form of selection, dictate the rate of adaptation
251
that slows down over time and drives the populations from an initially strong to a
252
finally weak selection regime.
253
Number of coexisting genotypes
254
During adaptation of an initially monomorphic population, de novo mutations can
255
generate polymorphism but at the same time purifying selection tends to reduce such
256
diversity. How many genotypes are generated and maintained under competition for
257
resources? The simulations show that, after a first burst of diversity, the mean number
258
of genotypes first declines and later plateaus (Fig. 3A). This reduction in the mean
259
number of genotypes is due to the energetic constraint; In fact, populations evolving
260
under neutrality or without such boundary do not suffer any decline (Fig. S3). In
261
13
contrast, when the populations’ phenotypes approach the energetic constraint,
262
beneficial mutations become rarer and selection reduces the number of coexisting
263
genotypes (Fig. S3). Despite the more abundant deleterious mutations, the populations
264
can maintain a dynamic balance between the lineages that are purged and the newly
265
emerging ones (inset of Fig. 3A). The long-term number of genotypes, M*, averaged
266
across the last 1000 generations, can be fitted by a linear function of the population
267
mutation rate whose slope decreases for stronger selection: M*RaNU+1, with an
268
inferred a= {8.25±0.06, 6.83±0.02, 5.48±0.02} for σ= {0.0125,0.025,0.05}, respectively
269
(Fig. 3B). Additional simulations, run over longer times (50000 generations), confirmed
270
the long-term plateau of the number of genotypes (M), which can be fitted to a power
271
law decay M(t)=ct-β+d, with asymptote dR1 (see Fig. S4, for parameter values N=107,
272
U=10-5, σ=0.05, ρ=0).
273
The number of genotypes, M*, deviates from NU/σ - the expected mean number of
274
deleterious mutations under mutation-selection balance in a simple model of constant
275
negative selection (Haigh, 1978) (see Fig. S5).
276
Population diversification into ecotypes
277
In this model, adapting populations consist of a cloud of many genotypes (Fig. 3) and we
278
now characterize the phenotypic structures of these clouds. In trait space, the
279
population optimum Ω (represented by the star in Fig. 1B) has ∑ ��
�*� � 1
�
�
and all
280
individuals’ phenotypes mirroring the resource supply proportions I��
�*� �
��
���� , �� �
281
1, … , 4J. Such state cannot be invaded by any mutant, thus excluding diversity.
282
However, Posfai et al. have shown that, in the absence of de novo mutations, large
283
collections of phenotypes can stably coexist when they are distributed around the
284
optimum, if polymorphism and metabolic trade-off are the initial conditions (Posfai,
285
14
Taillefumier and Wingreen, 2017). Thus, we now ask: if mutation is the only source of
286
variation, will an initially maladapted isogenic population evolve towards a single
287
optimal state or towards distinct states?
288
The simulations show that, due to the stochastic nature of mutation, populations
289
adapting under exactly the same conditions can evolve either one or multiple states
290
(Fig. 4A). Remarkably, the same ancestral genotype can give rise to a better adapted
291
population with many genotypes, all functionally similar (Fig. 4A, left panel), or
292
diversify into different ecotypes: clusters of genotypes with distinct metabolic
293
preferences, capable of coexisting indefinitely (Fig. 4A, right panel).
294
We now investigate the conditions favoring such ecotype diversification, in the regime
295
of strong selection (NsR1). Using the mean shift clustering algorithm (Cheng, 1995),
296
which tests for multimodality, one can group each adapted population into functional
297
clusters (see details in Appendix S2). We find that the proportion of populations that
298
evolved into 1, 2 or more clusters, changes dramatically with NU. Under regimes of
299
more intense clonal interference, distinct ecotypes more probably emerge and coexist.
300
Importantly, when NU is very large, the number of ecotypes may exceed the number of
301
limiting resources (Fig. 4B). The probability of diversification (P) – computed as the
302
proportion of populations that evolved more than one cluster – is close to zero when
303
NU<1 but significantly increases when NU≥1. Such increase can be fitted by a logistic
304
function: D �
�
��+����������� (see Fig. 4C and Tables S1-2). In contrast, the populations do
305
not diversify into clusters under neutrality (Fig. 4C).
306
Both the rate and the type of mutations influence the diversification process. Under
307
intense clonal interference, larger mutation effects (σ) and/or smaller pleiotropic
308
effects (ρ) promote the formation of multiple ecotypes (Fig. 4C, Fig S6 and S7).
309
15
In summary, during the process of adaptation to improve consumption of the available
310
resources, different outcomes can evolve: i) if the input of new beneficial mutations is
311
low, the recurrently fittest takes over as a cloud of genotypes until the population forms
312
a single generalist ecotype (e.g. Video S1); ii) if NU is large, the availability of many
313
beneficial mutations provides the potential for several genotypes to coevolve and
314
distinct ecotypes to stably coexist (e.g. Video S2).
315
Phenotypic and genotypic diversity within populations
316
Can we predict ecotypes from the genotypes in the population? Populations composed
317
by a single functional cluster can have considerable genetic diversity if multiple lineages
318
have converged to similar phenotypes (see examples in Fig. S8). We characterized the
319
adapting populations by calculating their average pairwise genetic (πG) and phenotypic
320
(πP) distances. Both measures of diversity increase with NU and always exceed the
321
neutral case (Fig. 5A). The increase of πP, but not of πG, is negatively affected by larger
322
pleiotropic effect (ρ) (ANOVA p-value = 3.2x10-10 and 0.09, respectively). Correlating πP
323
and πG in the evolved populations, we confirmed that more intense clonal interference
324
fosters functional diversification but stronger pleiotropy constrains it (Fig. S9A-B).
325
We next attempted to predict the existence of multiple ecotypes using the genotypic
326
diversity. The populations that evolved more than one cluster have on average larger
327
πG. However, it is difficult to establish a threshold above which multiple ecotypes can be
328
predicted (Fig. 5B). The Receiver Operating Characteristic (ROC) curves – obtained via
329
varying thresholds – show the predictive power of πG and confirm that this improves for
330
those populations that evolved under weaker pleiotropy (Fig. 5C). For example, the
331
thresholds needed to predict ecotypes with 95% specificity are: πG> {2.7, 3.6, 4.5} for ρ=
332
{0, 0.5, 1}, respectively (as shown by the dotted lines in Fig. 5B and the dots in Fig. 5C).
333
16
A common statistic used to infer population genetic structure is the Tajima’s D (Tajima,
334
1989), which compares the number of segregating mutations with the average pairwise
335
genetic distance. The expected value of D is zero under an equilibrium model of
336
mutation and drift, without selection. We note that the time simulated (104 generations)
337
is much shorter than that required to achieve a neutral equilibrium at these population
338
sizes (N=107), so as expected our results in the absence of selection show a negative
339
Tajima’s D, whose mean decreases with NU due to larger numbers of newly emerging
340
segregating sites. Under selection, populations with the same NU may present different
341
genetic structures (see three examples in Fig. S8) and Tajima’s D ranges from very
342
negative to very positive values (Fig. 6). Interestingly, the mean Tajima’s D is maximal
343
for intermediate NU and it can be close to zero under both low or very large NU
344
(compare the black and the gold distributions in Fig. 6). These patterns suggest that
345
other statistics beyond Tajima’s D are needed to understand the structure of non-
346
equilibrium populations undergoing strong selection.
347
348
Discussion
349
Microbial communities are vital for humans and many other host species (Nicholson et
350
al., 2012; Sunagawa et al., 2015). Emerging observations of evolution in such
351
ecosystems (Barroso-Batista et al., 2014; Garud et al., 2019; Zhao et al., 2019) motivate
352
new theories where the mechanisms that generate diversity involve complex forms of
353
selection and clonal interference (Gordo, 2019). We propose that an eco-evolutionary
354
model of resource competition, describing the mechanisms behind ecological
355
divergence, can help understand diversification within ecosystems. Our framework can
356
be generalized to incorporate other evolutionary mechanisms, including transmission
357
17
and horizontal gene transfer, and can serve as a bridge between ecology and population
358
genetics.
359
Population genetics models of clonal interference have greatly advanced our
360
understanding on adaptation under this condition (Gerrish and Lenski, 1998; Park and
361
Krug, 2007; Good et al., 2012; de Sousa et al., 2016). However, clonal interference is
362
rarely considered in theoretical studies of ecosystems (Farahpour et al., 2018), even
363
though it greatly impacts the evolution of microbes within communities (Barroso-
364
Batista et al., 2014) and may be relevant in key ecosystems such as the human
365
microbiota (Zhao et al., 2019). Commensal species in the gut have large population sizes
366
~108 cells/g; If each bacterium mutates in the gut as it does in the laboratory (Drake,
367
1991), then each gram of material will host around 105 new mutant cells every
368
generation. Even if only 0.1% brings up a benefit (Perfeito et al., 2007), clonal
369
interference still extensively affects the gut microbiota dynamics.
370
Here we have studied an ecological model where clones do not compete for fixation but
371
for resources. Modeling competition explicitly allows to make testable predictions
372
about diversity as both the traits and genomes can now easily be studied. We show that
373
clonal interactions can drive an initial monomorphic population to polymorphism with
374
distinct ecotypes, deviating from the simple expectation of adapting to a single optimum
375
phenotype (Fig. 4).
376
Taking the MacArthur model, Posfai et al. have demonstrated that metabolic trade-offs
377
promote coexistence of more species than resource types, thus overcoming the
378
competitive exclusion principle in the absence of evolution (Posfai, Taillefumier and
379
Wingreen, 2017). Extensions of this framework already demonstrated its power in
380
recapitulating experimental results from studies of soil, plant (Goldford et al., 2018) or
381
18
mammalian gut microbiotas (Leónidas Cardoso et al., 2020). In addition, a similar
382
framework provided analytical descriptions of how populations adapt under
383
competition for resources and demonstrated that directional selection can limit
384
ecological diversification (Good, Martis and Hallatschek, 2018). This pattern is also
385
observed in our simulations; In fact, stronger pleiotropy, which causes stronger
386
directional selection, limits the emergence of clusters (Fig. 4C). Our work differs from
387
the latter by focusing on clonal interference and by introducing an energetic trade-off.
388
We show that under clonal interference, the outcome of phenotypic adaptation is
389
probabilistic, whereby the populations can evolve paths that would be impossible to
390
observe through equilibrium assumptions.
391
Trade-offs are commonly assumed and expected to affect evolutionary trajectories
392
(Farahpour et al., 2018; Amado and Campos, 2019), but this is not always observed.
393
While many empirical results have confirmed the role of trade-offs during adaptation
394
(Bell and Reboud, 1997; Bull, Badgett and Wichman, 2000; Turner and Elena, 2000;
395
Dykhuizen and Dean, 2004; Greene et al., 2005; Duffy, Turner and Burch, 2006; Coffey et
396
al., 2008; Ward, Perron and MacLean, 2009; Bailey and Kassen, 2012; Li, Petrov and
397
Sherlock, 2019), others did not find evidence for any (Reboud and Bell, 1997; Kassen
398
and Bell, 1998; Turner and Elena, 2000; Trindade et al., 2009; Bedhomme, Lafforgue
399
and Elena, 2012). Here we assumed a trade-off in the form of an energetic constraint,
400
which only affects well adapted genotypes. Thus, in our model the observation of a
401
trade-off depends on the time at which it is measured. It would be interesting to test for
402
trade-offs at different times during adaptation, as this could explain some of the
403
contrasting findings outlined above. Compatible with this hypothesis, a trade-off in
404
Escherichia coli ability to grow in the presence of both glucose and lactose was found,
405
19
but it only emerged after a period of constraint-free adaptation (Satterwhite and
406
Cooper, 2015). We find that the metabolic trade-off in the resource consumption is not
407
required for the formation of distinct ecotypes but it promotes their stable coexistence
408
(Fig. 4, S8 and Video S2).
409
We introduced mutation in a model of competition for resources to couple ecology and
410
evolution. This framework allowed some analytical approximations of the selection that
411
underlies adaptation. Consistent with empirical data, competition-driven selection is
412
characterized by diminishing return epistasis - the benefit decline in populations with
413
higher mean fitness (Chou et al., 2011; Kryazhimskiy et al., 2014; Schoustra et al., 2016;
414
Wünsche et al., 2017).
415
In many ecosystems coexisting types seem to outnumber the limiting resources, and,
416
solving this contention has motivated numerous studies. Previous eco-evolutionary
417
analysis suggest that adding evolutionary changes confirms (Edwards et al., 2018) or
418
even exacerbates (Shoresh, Hegreness and Kishony, 2008) this paradox. Perhaps
419
surprisingly, our simulations show that large mutational inputs maintain a dynamically
420
stable number of types that overcome the competitive exclusion (Fig. 3). And at the
421
functional level, we show that diversity generally respects the exclusion principle
422
(number of types ≤ number of resources) but with exceptions: in a regime of strong
423
clonal interference, the number of extant ecotypes can even exceed the number of
424
limiting resources (Fig. 4B). The process leading to such ecological diversification is
425
strongly influenced by the underlying molecular parameters: regimes of low clonal
426
interference (NU <1) lead to the evolution of a single generalist population, but large
427
mutational inputs (NU≥1) can lead to the formation of two or more differentially
428
specialized ecotypes (Fig. 4 and Video S1-S2).
429
20
Our analysis demonstrates that both large mutation effects and weak pleiotropy foster
430
ecological diversification (Fig. 4C). Different pleiotropic effects are meant to represent
431
different interactions between the traits under selection. If the available resources are
432
similar (e.g. chemical composition) and/or the metabolic processes involved in their
433
consumption share many genes, this could increase the chances that a mutation affects
434
the two traits simultaneously leading to large pleiotropy. Contrarily, less related
435
resources could involve more independent effects leading to smaller pleiotropy,
436
promoting diversification (Fig. 4C). This interpretation could explain why adaptive
437
diversification occurred in some experimental evolution setups (Friesen et al., 2004;
438
Sandberg et al. 2017) but was not observed in others (Satterwhite and Cooper, 2015;
439
Sandberg et al. 2017). In agreement with this hypothesis, Sandberg and colleagues
440
showed that evolving on less metabolically related resources promoted ecological
441
diversification (Sandberg et al. 2017).
442
Sympatric diversification can be observed experimentally and predicted by theoretical
443
models (Friesen et al., 2004). The framework of adaptive dynamics has been extensively
444
used in this context as it describes evolution on fitness landscapes that change
445
dynamically due to frequency-dependent interactions (Geritz et al., 1998; Doebeli,
446
2011). While very useful to understand diversification, adaptive dynamics are based on
447
equilibrium assumptions: the populations first evolve to an equilibrium state before
448
diversification occurs, as explained by the concept of evolutionary branching points.
449
Our approach in contrast follows adaptation through an individual-based model that
450
undergoes strong non-equilibrium dynamics. Previous studies (Rosindell, Harmon and
451
Etienne, 2015; Ispolatov, Madhok and Doebeli, 2016) highlighted how considering
452
evolution at the individual level is necessary to fully understand the adaptation process.
453
21
We find that the genetic diversity of a population can be used to predict the underlying
454
phenotypic structure, but with limitations. The accuracy of such prediction decreases if
455
mutations are not constrained and can affect equally all the traits under selection (Fig.
456
5). Other statistics based on genetic data (such as the Tajima’s D) can strongly deviate
457
from the expected values if the populations are not yet at (evolutionary) equilibrium
458
and can be hard to interpret if the underlying populations’ parameters are unknown.
459
The model studied here can be considered a first step of integration. It predicts that the
460
typical high mutational input of bacterial species and cancer cells, coupled with an
461
energetic constraint, is a mechanism capable of generating functionally diverse clonal
462
communities. Future frameworks addressing microbial ecology and evolution will need
463
to address how space, migration and/or fluctuating conditions affect the patterns of
464
diversity observed here. In addition, cooperation between genotypes (such as cross-
465
feeding or use of costly public goods) have the potential to shape diversity levels and
466
should be investigated in future studies.
467
468
Acknowledgements
469
The authors acknowledge C Bank and the members of the Evolutionary Dynamics and
470
Evolutionary Biology labs of the IGC for their assistance throughout the development of
471
this work; PR Campos, T Paixão, S Miller, G Sgarlata and R Ramiro for their comments
472
on the manuscript.
473
This work was supported by Portuguese Science Foundation (FCT) Grant
474
PTDC/BIA-EVL/31528/2017; Deutsche Forschungs-gemeinschaft (DFG) Grant
475
SFB 1310; and a cooperation agreement between University of
476
22
Cologne and Gulbenkian Institute. MA was supported by FCT Grant
477
PD/BD/138735/2018.
478
479
References
480
Abrams, P. A. (1988) ‘How should resources be counted?’, Theoretical Population Biology. doi:
481
10.1016/0040-5809(88)90014-7.
482
Amado, A. and Campos, P. R. A. (2019) ‘Ecological specialization under multidimensional trade-offs’,
483
Evolutionary Ecology. Springer International Publishing, 33(6), pp. 769–789. doi: 10.1007/s10682-019-
484
10013-4.
485
Mac Arthur, R. (1969) ‘Species Packing, and What Competition Minimizes’, Proceedings of the National
486
Academy of Sciences, 64(4), pp. 1369–1371. doi: 10.1073/pnas.64.4.1369.
487
Bailey, S. F. and Kassen, R. (2012) ‘Spatial structure of ecological opportunity drives adaptation in a
488
bacterium’, American Naturalist. doi: 10.1086/666609.
489
Barroso-Batista, J. et al. (2014) ‘The First Steps of Adaptation of Escherichia coli to the Gut Are Dominated
490
by Soft Sweeps’, PLoS Genetics. doi: 10.1371/journal.pgen.1004182.
491
Barroso-Batista, J. et al. (2020) ‘Specific Eco-evolutionary Contexts in the Mouse Gut Reveal Escherichia
492
coli Metabolic Versatility’, Current Biology. doi: 10.1016/j.cub.2020.01.050.
493
Bedhomme, S., Lafforgue, G. and Elena, S. F. (2012) ‘Multihost experimental evolution of a plant RNA virus
494
reveals local adaptation and host-specific mutations’, Molecular Biology and Evolution. doi:
495
10.1093/molbev/msr314.
496
Bell, G. and Reboud, X. (1997) ‘Experimental evolution in Chlamydomonas II. Genetic variation in strongly
497
contrasted environments’, Heredity. doi: 10.1038/hdy.1997.78.
498
Bull, J. J., Badgett, M. R. and Wichman, H. A. (2000) ‘Big-benefit mutations in a bacteriophage inhibited
499
with heat’, Molecular Biology and Evolution. doi: 10.1093/oxfordjournals.molbev.a026375.
500
Cheng, Y. (1995) ‘Mean Shift, Mode Seeking, and Clustering’, IEEE Transactions on Pattern Analysis and
501
Machine Intelligence. doi: 10.1109/34.400568.
502
Chou, H. H. et al. (2011) ‘Diminishing returns epistasis among beneficial mutations decelerates
503
adaptation’, Science, 332(6034), pp. 1190–1192. doi: 10.1126/science.1203799.
504
Coffey, L. L. et al. (2008) ‘Arbovirus evolution in vivo is constrained by host alternation’, Proceedings of
505
the National Academy of Sciences of the United States of America. doi: 10.1073/pnas.0712130105.
506
Desai, M. M., Fisher, D. S. and Murray, A. W. (2007) ‘The Speed of Evolution and Maintenance of Variation
507
in Asexual Populations’, Current Biology. doi: 10.1016/j.cub.2007.01.072.
508
Dieckmann, U. and Doebeli, M. (1999) ‘On the origin of speciesby sympatric speciation’, Science, 400(July),
509
pp. 0–3. doi: 10.1038/22521.
510
Doebeli, M. (2011) Adaptive diversification (MPB-48), Adaptive Diversification (MPB-48). doi:
511
10.23943/princeton/9780691128931.001.0001.
512
Drake, J. W. (1991) ‘A constant rate of spontaneous mutation in DNA-based microbes’, Proceedings of the
513
National Academy of Sciences of the United States of America. doi: 10.1073/pnas.88.16.7160.
514
Duffy, S., Turner, P. E. and Burch, C. L. (2006) ‘Pleiotropic costs of niche expansion in the RNA
515
bacteriophage Φ6’, Genetics. doi: 10.1534/genetics.105.051136.
516
23
Dykhuizen, D. E. and Dean, A. M. (2004) ‘Evolution of specialists in an experimental microcosm’, Genetics.
517
doi: 10.1534/genetics.103.025205.
518
Edwards, K. F. et al. (2018) ‘Evolutionarily stable communities: a framework for understanding the role of
519
trait evolution in the maintenance of diversity’, Ecology Letters, 21(12), pp. 1853–1868. doi:
520
10.1111/ele.13142.
521
Farahpour, F. et al. (2018) ‘Trade-off shapes diversity in ecoevolutionary dynamics’, eLife, 7. doi:
522
10.7554/eLife.36273.
523
Fisher, R. A. (1958) ‘The genetical theory of natural selection, 2nd edn.’, Dover Publication, New York. doi:
524
10.1111/jeb.12566.
525
Friesen, M. L. et al. (2004) ‘Experimental evidence for sympatric ecological diversification due to
526
frequency-dependent competition in Escherichia coli’, Evolution, 58(2), pp. 245–260. doi:
527
10.1111/ele.12430.
528
Garud, N. R. et al. (2019) ‘Evolutionary dynamics of bacteria in the gut microbiome within and across
529
hosts’, PLoS biology, 17(1), p. e3000102. doi: 10.1371/journal.pbio.3000102.
530
Geritz, S. A. H. et al. (1998) ‘Evolutionary singular strategies and the adaptive growth and branching of the
531
evolutionary tree’, Evolutionary Ecology, 12, pp. 35–57.
532
Goldford, J. E. et al. (2018) ‘Emergent simplicity in microbial community assembly’, Science. doi:
533
10.1126/science.aat1168.
534
Good, B. H. et al. (2012) ‘Distribution of fixed beneficial mutations and the rate of adaptation in asexual
535
populations’, Proceedings of the National Academy of Sciences of the United States of America, 109(13), pp.
536
4950–4955. doi: 10.1073/pnas.1119910109.
537
Good, B. H., Martis, S. and Hallatschek, O. (2018) ‘Adaptation limits ecological diversification and
538
promotes ecological tinkering during the competition for substitutable resources’, Proceedings of the
539
National Academy of Sciences of the United States of America, 115(44), pp. E10407–E10416. doi:
540
10.1073/pnas.1807530115.
541
Gordo, I. (2019) ‘Evolutionary change in the human gut microbiome: From a static to a dynamic view’,
542
PLoS Biology, 17(2), pp. 1–5. doi: 10.1371/journal.pbio.3000126.
543
Greene, I. P. et al. (2005) ‘Effect of Alternating Passage on Adaptation of Sindbis Virus to Vertebrate and
544
Invertebrate Cells’, Journal of Virology. doi: 10.1128/jvi.79.22.14253-14260.2005.
545
Gresham, D. et al. (2008) ‘The repertoire and dynamics of evolutionary adaptations to controlled
546
nutrient-limited environments in yeast’, PLoS Genetics. doi: 10.1371/journal.pgen.1000303.
547
Haigh, J. (1978) ‘The accumulation of deleterious genes in a population-Muller’s Ratchet’, Theoretical
548
Population Biology, 14(2), pp. 251–267. doi: 10.1016/0040-5809(78)90027-8.
549
Hardin, G. (1960) ‘The Competitive Exclusion Principle’, Science, 131(November).
550
Herron, M. D. and Doebeli, M. (2013) ‘Parallel Evolutionary Dynamics of Adaptive Diversification in
551
Escherichia coli’, PLoS Biology. doi: 10.1371/journal.pbio.1001490.
552
Huston, M. A. (1994) ‘Biological diversity: the coexistence of species on changing landscapes’, Biological
553
diversity: the coexistence of species on changing landscapes. doi: 10.2307/2404662.
554
Hutchinson, G. E. (1961) ‘THE PARADOX OF THE PLANKTON * New Haven , Connecticut Osborn
555
Zoological Laboratory , The problem that I wish to discuss in the present contribution is raised by the
556
very paradoxical situation of the plankton , particularly the phyto- plankton , of relativ’, XCV(882), pp.
557
137–145.
558
Ispolatov, I., Madhok, V. and Doebeli, M. (2016) ‘Individual-based models for adaptive diversification in
559
high-dimensional phenotype spaces’, Journal of Theoretical Biology. Elsevier, 390, pp. 97–105. doi:
560
10.1016/j.jtbi.2015.10.009.
561
24
Johnson, Norman L et al. (1994) ‘Continuous Univariate Distributions’, Wiley, 1, p. Section 10.1.
562
Kassen, R. and Bell, G. (1998) ‘Experimental evolution in Chlamydomonas. IV. Selection in environments
563
that vary through time at different scales’, Heredity. doi: 10.1046/j.1365-2540.1998.00329.x.
564
Kinnersley, M. A., Holben, W. E. and Rosenzweig, F. (2009) ‘E unibus plurum: Genomic analysis of an
565
experimentally evolved polymorphism in Escherichia coli’, PLoS Genetics. doi:
566
10.1371/journal.pgen.1000713.
567
Kremer, C. T. and Klausmeier, C. A. (2017) ‘Species packing in eco-evolutionary models of seasonally
568
fluctuating environments’, Ecology Letters. doi: 10.1111/ele.12813.
569
Kryazhimskiy, S. et al. (2014) ‘Global epistasis makes adaptation predictable despite sequence-level
570
stochasticity’, Science, 344(6191), pp. 1519–1522. doi: 10.1126/science.1250939.
571
Lawrence, D. et al. (2012) ‘Species interactions alter evolutionary responses to a novel environment’,
572
PLoS Biology. doi: 10.1371/journal.pbio.1001330.
573
Leónidas Cardoso, L, Durão, P, Amicone, M and Gordo, I (2020) ‘Dysbiosis individualizes the fitness effect
574
of antibiotic resistance in the mammalian gut’, Nature Ecology and Evolution. Springer US, 4(9), pp. 1268–
575
1278. doi: 10.1038/s41559-020-1235-1.
576
Li, Y., Petrov, D. A. and Sherlock, G. (2019) ‘Single nucleotide mapping of trait space reveals Pareto fronts
577
that constrain adaptation’, Nature Ecology and Evolution. Springer US, 3(11), pp. 1539–1551. doi:
578
10.1038/s41559-019-0993-0.
579
Litchman, E. and Klausmeier, C. A. (2001) ‘Competition of phytoplankton under fluctuating light’,
580
American Naturalist. doi: 10.1086/318628.
581
Maharjan, R. et al. (2006) ‘Clonal adaptive radiation in a constant environment’, Science. doi:
582
10.1126/science.1129865.
583
Nicholson, J. K. et al. (2012) ‘Host-gut microbiota metabolic interactions’, Science. doi:
584
10.1126/science.1223813.
585
de Oliveira, V. M., Amado, A. and Campos, P. R. A. (2018) ‘The interplay of tradeoffs within the framework
586
of a resource-based modelling’, Ecological Modelling. Elsevier, 384(June), pp. 249–260. doi:
587
10.1016/j.ecolmodel.2018.06.026.
588
Pacciani-Mori, L. et al. (2020) ‘Dynamic metabolic adaptation can promote species coexistence in
589
competitive microbial communities’, PLoS Computational Biology, 16(5), pp. 1–18. doi:
590
10.1371/journal.pcbi.1007896.
591
Park, S. C. and Krug, J. (2007) ‘Clonal interference in large populations’, Proceedings of the National
592
Academy of Sciences of the United States of America, 104(46), pp. 18135–18140. doi:
593
10.1073/pnas.0705778104.
594
Perfeito, L. et al. (2007) ‘Adaptive mutations in bacteria: High rate and small effects’, Science, 317(5839),
595
pp. 813–815. doi: 10.1126/science.1142284.
596
Philip J. Gerrish & Richard E. Lenski (1998) ‘The fate of competing beneficial mutations in an asexual
597
population’, Genetica, 102/103, pp. 127–144,.
598
Posfai, A., Taillefumier, T. and Wingreen, N. S. (2017) ‘Metabolic Trade-Offs Promote Diversity in a Model
599
Ecosystem’, Physical Review Letters, 118(2), pp. 1–5. doi: 10.1103/PhysRevLett.118.028103.
600
Reboud, X. and Bell, G. (1997) ‘Experimental evolution in Chlamydomonas III. Evolution of specialist and
601
generalist types in environments that vary in space and time’, Heredity. doi: 10.1038/hdy.1997.79.
602
Rosindell, J., Harmon, L. J. and Etienne, R. S. (2015) ‘Unifying ecology and macroevolution with individual-
603
based theory’, Ecology Letters, 18(5), pp. 472–482. doi: 10.1111/ele.12430.
604
Satterwhite, R. S. and Cooper, T. F. (2015) ‘Constraints on adaptation of Escherichia coli to mixed-
605
25
resource environments increase over time’, Evolution, 69(8), pp. 2067–2078. doi: 10.1111/evo.12710.
606
Schoener, T. W. (2011) ‘The Newest SynthesisO: Understanding Ecological Dynamics’, Science,
607
331(January), pp. 426–429. doi: 10.1126/science.1193954.
608
Schoustra, S. et al. (2016) ‘Diminishing-returns epistasis among random beneficial mutations in a
609
multicellular fungus’, Proceedings of the Royal Society B: Biological Sciences, 283(1837). doi:
610
10.1098/rspb.2016.1376.
611
Shoresh, N., Hegreness, M. and Kishony, R. (2008) ‘Evolution exacerbates the paradox of the plankton’,
612
Proceedings of the National Academy of Sciences of the United States of America, 105(34), pp. 12365–
613
12369. doi: 10.1073/pnas.0803032105.
614
de Sousa, J. A. M. et al. (2016) ‘Competition and fixation of cohorts of adaptive mutations under Fisher
615
geometrical model’, PeerJ, 2016(2), pp. 1–17. doi: 10.7717/PEERJ.2256.
616
Sunagawa, S. et al. (2015) ‘Structure and function of the global ocean microbiome’, Science. doi:
617
10.1126/science.1261359.
618
T. E. Sandberg, C. J. Lloyd, B..O. Palsson, A. M. F. (2017) ‘Laboratory Evolution to Alternating Substrate
619
Environments Yields Distinct Phenotypic and Genetic Adaptive Strategies’, 83(13), pp. 1–15.
620
Tajima, F. (1989) ‘Statistical method for testing the neutral mutation hypothesis by DNA polymorphism’,
621
Genetics.
622
Tilman, D. (1982) ‘Resource competition and community structure.’, Monographs in population biology.
623
doi: 10.2307/4549.
624
Trindade, S. et al. (2009) ‘Positive epistasis drives the acquisition of multidrug resistance’, PLoS Genetics.
625
doi: 10.1371/journal.pgen.1000578.
626
Turner, P. E. and Elena, S. F. (2000) ‘Cost of host radiation in an RNA virus’, Genetics.
627
Ward, H., Perron, G. G. and MacLean, R. C. (2009) ‘The cost of multiple drug resistance in Pseudomonas
628
aeruginosa’, Journal of Evolutionary Biology. doi: 10.1111/j.1420-9101.2009.01712.x.
629
Wright, S. (1930) ‘The genetical theory of natural selection: A review’, Journal of Heredity. doi:
630
10.1093/oxfordjournals.jhered.a103361.
631
Wünsche, A. et al. (2017) ‘Diminishing-returns epistasis decreases adaptability along an evolutionary
632
trajectory’, Nature Ecology and Evolution, 1(4), pp. 1–6. doi: 10.1038/s41559-016-0061.
633
Zhao, S. et al. (2019) ‘Adaptive Evolution within Gut Microbiomes of Healthy People’, Cell Host and
634
Microbe. Elsevier Inc., 25(5), pp. 656-667.e8. doi: 10.1016/j.chom.2019.03.007.
635
636
Figure legends
637
Figure 1. Ecological dynamics and individual-based evolutionary processes. A) Illustration of the
638
eco-evolutionary dynamics. Bacterial genotypes, represented by circles of different colors, grow
639
according to the resources’ concentration (squares and semicircles) and their own phenotypic traits,
640
represented by the enzyme-like structures on the cell wall. Mutation events (red arrow) can generate
641
new types whose fate will depend on drift and selection. B) Phenotypic space and mutation process. An
642
initially maladapted monomorphic population (green circle) can acquire de-novo mutations according to
643
26
the given assumptions, as explained in the inset. C) Examples of different clonal interference regimes, NU
644
<< 0.1 above, NU >>1 below. Each color represents a different mutant type.
645
646
Figure 2. Diminishing return epistasis and the adaptation rate. A) Analytical predictions of the
647
diminishing return epistasis in monomorphic populations. The inset shows the effect of the energy
648
constraint. B) Dots represent the mean of the observed positive selection during the first 300 generations
649
of adaptation. The dotted line represents the analytical expectation at each generation �, given the
650
average population trait sum ������ �
∑
������ �
���! �
����
����
�
"
�. Each color represents an independent
651
simulated population, which adapted with σ � 0.05, ρ � 0.5, � � 10# and U� 10$%. C) The population
652
average trait sum ����� is shown as proxy of adaptation under different σ and � conditions. Other
653
parameters: � � 2, � � 10#, � � 10$&. Lines are the averages over 100 simulations and the shaded
654
areas represent the confidence interval. D) Same dynamics as in C, but on a different scale: Generations ·
655
I · σ', where I and σ' are defined in the figure. E) Adaptation across different mutational inputs (color
656
coded) and the expected beneficial effect of mutations at the end of the adaptation process (generation
657
10000). The inset panel highlights the dynamics over the last 100 generations. Points in the right panel
658
are represented by the mean +/- one standard deviation in log-log scale. Other parameters for panel E:
659
� � 10#, " � 0.05,ρ � 0.5.
660
661
Figure 3. Genotypes’ dynamics and the mutation-selection balance under competition for two
662
resources. A) Number of genotypes present in the environment over time, under neutrality (diamonds)
663
or under selection (lines). Lines represent the average across 100 populations and the shaded area their
664
confidence interval. The populations eventually plateau and dynamically maintain large number of
665
genotypes (#'). Other parameters: " � 0.05,� � 0 and � � 10#. B) Long-lasting number of genotypes,
666
computed as the average over the last 1000 generations (e.g. dotted line in the right panel of A). The lines
667
represent the linear regressions with model: #' � $�� % 1 and ��: '0.1,0.5,1,5,10,50,100,500( where
668
$ � '8.25 * 0.06,6.83 * 0.02,5.48 * 0.02( for " � '0.0125,0.025,0.05(, respectively. Both axes are
669
represented in log scale with ticks every '1,…,9( · 10(. Other parameters: � � 0 and � � 10#.
670
27
671
Figure 4. Ecological diversification under competition for two resources. A) Two example
672
populations evolving under the same conditions (� � 10#, � � 10$%,� � 0.5," � 0.05,R � 2). The
673
phenotypes and the preference distributions show one population that has evolved into a single optimal
674
cluster (squares) and another population that gave rise to a stable diverse community composed by two
675
clusters (circles). Lines connecting the shapes represent mutations. B) Counts of populations that evolved
676
into 1,2 or more phenotypic clusters. Here, � � 10#, σ � 0.05,R � 2,ρ � 0. C) Populations diversify
677
with a probability that increases with ln���� and σ but decreases with ρ. The lines represent the fit of the
678
data to the logistic function: 2 �
)
)!*����������� , ��:'0.1,0.2,0.3,0.4,0.5,0.6,1,5,10,50,100,500(. The
679
inferred parameters a and 3 are reported in Tables S1-2 and the full set of data is shown in Fig. S6. In the
680
left plot: ρ � 0, while in the right one: σ � 0.05. The probabilities were computed as proportions out of
681
100 independent populations and their 95% confidence interval by normal approximation: 2 * 45+�)$+�
),, ,
682
4 � 1.96.
683
684
Figure 5. Diversity within and between populations and ecotypes’ prediction. A) Average pairwise
685
genotypic (π-) and phenotypic (π+) diversity were measured within each population, as defined in the
686
Methods. 100 independent populations were simulated under the conditions specified on the x axis and
687
by the colors. Other parameters: σ = 0.05, R=2 and � � 10#. B) π- distributions of the populations that
688
evolved in more than one cluster (dark blue) or in a single cluster (light blue). Data form panel A with
689
different NU were pulled together for a total of 800 populations per condition (ρ). The violin plots in A
690
and B show the distribution of the data and their median. The dotted lines represent the π- threshold that
691
would ensure 95% specificity in ecotype prediction (i.e. 5% false positive) and correspond to the circles
692
in panel C. C) π- was used to predict whether each population is composed by one or multiple ecotypes.
693
The ROC curves represent the sensitivity over one minus specificity of the prediction outcomes, for
694
varying thresholds. The dotted line and the circles represent the thresholds that ensure 5% false positive
695
rate and these are: π- 7 2.7,3.6 or 4.5 for ρ=0, 0.5 or 1, respectively. The area under the curve (AUC) is
696
reported on the figure. * or *** indicate p-value <0.01 or <0.0001, respectively.
697
28
698
Figure 6. Tajima’s D within and between populations. The Tajima’s D statistics was computed for each
699
population at the end of the adaptation (generation 10000) from a sample of 9 � 100 genotypes.
700
Samples under selection (empty circles) are compared with samples under neutrality (full diamonds).
701
Each data point is an independent population. Those populations whose sample did not present
702
polymorphism, are not represented (e.g. black diamonds are missing) because the Tajima’s D is not
703
defined as the number of segregating sites would be zero. The corresponding count of monomorphic
704
samples is given in the lower panel. Other parameters: � � 10#,� � 0, " � 0.05,� � 2. The violin plots
705
show the distributions of the data and their median.
706
707
Pleiotropic effect
ρ = 0
ρ = 1
...
...
∼ σ
Optimal
phenotype
(α2)
(α1)
1
0
1
0
Consumption rate of r1
Consumption rate of r2
1
200
Genotypes’ abundance
NU << 1
1
200
Generations
NU >> 1
N: population size
U: mutation rate
C
A
B
t + 1
Resources:
r1
r2
Genotypes’ abundance
ρ:
0
0.5
1
σ:
0.0125
0.025
0.05
0.1
1
0
0.3
0.6
Expected [s+]
0.9
1
0
0.05
0.9
NU
0.1
1
10
100
0
2000
4000
6000
8000
10000
Generations
^
Average trait sum ( )
Expected [s+]
10-2
10-3
10-4
0.99
1
9900
10000
E
A
B
C
1
10
100
0.1
NU:
●
●
●
●
ρ:
0
0.5
1
0
0.5
1
0
2000
4000
6000
8000
10000
Generations
^
Average trait sum ( )
Generations Ι σ∗
. .
0
90
σ∗
Ι
Analitical approximation E [ s+ | α ]
^
0.9
0.1
0.9
1
Mean (s+)
^
Average trait sum ( )
1
Observed
0
0.3
0.6
0
0.05
0
0.5
1
0
0.5
1
D
σ:
0.0125
0.025
0.05
Generations
1
10
100
1000
0
2000
4000
6000
8000
10000
545
555
565
9000
10000
Number of genotypes (M)
Neutral
NU:
0.1
1
10
100
10
100
1000
10000
Mean number of genotypes (M*)
NU
5 10
50 100
500
σ = 0.05
σ = 0.025
σ = 0.125
B
A
M*
1
0.5 1
0.1
0.0
0.5
1.0
Probability (Clusters > 1)
0.0
0.5
1.0
-2
6
4
0
2
ln(NU)
Neutral
σ:
0.05
0.025
0.0125
ρ:
0
0.5
1
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
α1
α1
α2
α2
0.0
0.5
1.0
α1
α1 + α2
0.0
0.5
1.0
α1
α1 + α2
Optimal phenotype
Ancestral phenotype
A
B
C
-2
6
4
0
2
ln(NU)
0
25
50
75
100
0.1
0.5
1
5
10
50
100
500
NU
Population counts
Clusters:
3
2
1
4
●
●● ●
●●●
●
●
● ●
●
●
●
●●
● ●●
●●
●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ●
●
●
●
●
●
●
●
● ●
● ●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●
●● ●
●
●
● ●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
● ●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
● ●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●●
●●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
● ●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●
●
●●●●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●●
●●
●
●●
● ●
●●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
● ● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
2.5
5.0
7.5
10.0
12.5
ρ = 0
ρ = 0.5
ρ = 1
Neutral
πP
Mean pairwise phenotypic distance
ρ = 0
ρ = 0.5
ρ = 1
Neutral
0.0
2.5
5.0
7.5
10.0
πG
Mean pairwise genetic distance
0.5
1
5
0.1
NU:
50
100
500
10
A
B
ρ = 0
ρ = 0.5
ρ = 1
0.0
2.5
5.0
7.5
10.0
πG
1 cluster
>1 clusters
False positive percentage
True positive percentage
0
20
40
60
80
100
0
20
40
60
80
100
ρ : 0
0.5
1
***
***
C
95% specificity
AUC: 83.6%
●
AUC: 91.3%
AUC: 94.8%
●
●
*
−2.5
0.0
2.5
S
N
S
N
S
N
S
N
S
N
S
N
S
N
S
N
Tajima’s D
0
25
50
75
100
S
N
Monomorphic samples
S
N
S
N
S
N
S
N
S
N
S
N
S
N
Selection (S)
Neutral (N)
0.5
1
5
0.1
NU:
50
100
500
10
A
B
| 2020 | Molecular signatures of resource competition: clonal interference drives the emergence of ecotypes | 10.1101/2020.11.20.391151 | [
"Amicone Massimo",
"Gordo Isabel"
] | creative-commons |
1
Title: A machine-learning approach to human footprint index estimation with applications to
sustainable development
Authors: Patrick W. Keys1, Elizabeth A. Barnes2, and Neil H. Carter3
Affiliations:
1School of Global Environmental Sustainability, Colorado State University, Fort Collins, CO
2Department of Atmospheric Science, Colorado State University, Fort Collins, CO
3School for Environment and Sustainability, University of Michigan, Ann Arbor, MI
Abstract
The human footprint index is an extensively used tool for interpreting the accelerating pressure
of humanity on Earth. Up to now, the process of creating the human footprint index has required
significant data and modeling, and updated versions of the index often lag the present day by
many years. Here we introduce a near-present, global-scale machine learning-based human
footprint index (ml-HFI) which is capable of routine update using satellite imagery alone. We
present the most up-to-date map of the human footprint index, and document changes in human
pressure during the past 20 years (2000 to 2019). Moreover, we demonstrate its utility as a
monitoring tool for the United Nations Sustainable Development Goal 15 (SDG15), "Life on
Land", which aims to foster sustainable development while conserving biodiversity. We identify
43 countries that are making progress toward SDG15 while also experiencing increases in their
ml-HFI. We examine a subset of these in the context of conservation policies that may or may
not enable continued progress toward SDG15. This has immediate policy relevance, since the
majority of countries globally are not on track to achieve Goal 15 by the declared deadline of
2030. Moving forward, the ml-HFI may be used for ongoing monitoring and evaluation support
toward the twin goals of fostering a thriving society and global Earth system.
Introduction
The human footprint index (HFI) represents one of the most important tools for interpreting
human pressure on the landscape (Venter et al 2016b). A dimensionless metric which captures
the extent of human influence on the terrestrial surface (Fig. 1a), the HFI is distinct from many
2
land-use metrics in that it captures the total influence of human existence on a given location in a
single index, rather than categorizing individual land use types (Riggio et al 2020). The
applications to which the HFI is put to use are manifold (Di Marco et al 2018, Watson et al
2018, Belote et al 2020, Beyer et al 2020), with enormous demand for information about human
pressure on the land surface to support policies related to land use change, biodiversity
conservation, and climate action (Pascual et al 2017, Beyer et al 2020, Ruckelshaus et al 2020).
A key challenge for expanded use of the HFI for operational efforts, however, is that new
updates typically lag the present day by seven or more years (Venter et al 2016a, Williams et al
2020). This time lag means that large-scale or substantial changes to the land surface caused by
human activities can occur well before we detect them, hampering our ability to monitor and
respond to their effects on biodiversity. This temporal mismatch pertains directly to the
achievement of the United Nations Sustainable Development Goals (SDG) since the deadline for
their achievement is 2030 (Sachs et al 2020b, Naidoo and Fisher 2020, Anon 2020), with some
specific targets within SDG15, “Life on Land”, set for even sooner (Sachs et al 2020a). While
international (Sachs et al 2020a) and voluntary national-scale assessments of various biodiversity
metrics (Government of the Co-operative Republic of Guyana 2019) are available for monitoring
progress toward SDG15, comparing those SDG15 assessments with an operational interpretation
of human influence on the terrestrial land surface - one that can be easily and regularly updated -
would provide synchronous insights on the social-ecological outcomes and processes of
landscape change.
The conventional approach for creating the HFI (Fig 1a) requires a harmonization of eight
different sub-indices representing different aspects of human pressures on the terrestrial surface
of the Earth (Williams et al 2020). These sub-indices include built infrastructure, population
density, and a variety of land use and land cover data. Quantification of the global HFI in this
way is a time intensive process and the most recent version of the HFI (published in 2020
(Williams et al 2020)) covers years up to 2013.
Machine learning, specifically convolutional neural networks (CNN), are well suited to the task
of identifying patterns in imagery (Krizhevsky et al 2012, Jean et al 2016, Xie et al 2015) , and
have been leveraged extensively to identify various patterns of human activity (Hoffman et al
3
2011, Kumar et al 2019, Qiu et al 2020). Here, we train a CNN to ingest space-borne imagery of
the earth’s surface and predict the human footprint index. Specifically, the Hansen Global Forest
Change imagery version 1.7 (GFCv1.7) (Hansen et al 2013) includes global, cloud-free, growing
season Landsat imagery for the years 2000 and 2019 (Hansen et al 2013). We exploit the overlap
between the year-2000 Williams-HFI and the year-2000 GFCv1.7 imagery for training the CNN.
We then use the trained CNN to predict the 2019 ml-HFI based on 2019 GFCv1.7 imagery. The
result is the first near-present, global, machine learning-based human footprint index (ml-HFI)
for the years 2000 and 2019 (Fig 1).
While a near-present, global ml-HFI opens many avenues for research, we demonstrate its
application to global conservation policy, by comparing it against country-level progress being
made toward SDG15. Specifically, using the most up-to-date trends from the United Nations, we
find that 43 countries are documented as making progress toward SDG15 and yet also have
experienced increases in ml-HFI between 2000 and 2019. Progress toward SDG15 can come
about in multiple ways, so we analyze 8 of the 43 countries to visualize how the geographic
distribution of changes in human pressure coincide with specific indicators of conservation
success. Given that the near-present ml-HFI could be used to motivate specific types of
conservation activities, we further explore the policy mechanisms that may support the
coexistence of SDG15 progress and increased human footprint. We anticipate future updates to
the ml-HFI as soon as updated cloud-free remote sensing imagery is made available, which
should enable this research to provide continuing support for biodiversity conservation and
sustainable development.
Data
Human Footprint Index
We quantify human pressure on the landscape using the human footprint index (HFI), which is a
global, dimensionless index of human pressure on the land surface (Sanderson et al 2002, Venter
et al 2016a). We employ the most up-to-date version of the HFI (Williams et al 2020), which we
denote as “Williams-HFI” hereafter in this text. The Williams-HFI is comprised of eight
different sub indices representing different aspects of human pressures to the terrestrial surface
of the Earth, including 1) extent of the built environment, 2) population density, 3) electric
4
infrastructure, 4) agricultural lands, 5) pasture lands, 6) roadways, 7) railways, and 8) navigable
waterways. Here, we rescale the original index, which ranges from 0 to 50, to a range from 0 to 1
(from low to high human pressure). The Williams-HFI data is output in a gridded format, where
each gridpoint is 0.00989273 degrees latitude by 0.00989273 degrees longitude.
Landsat imagery
We use the Global Forest Change dataset, a global analysis of forest cover change based on
Landsat imagery (Hansen et al 2013). The data consist of processed Landsat imagery from the
bands 3, 4, 5, and 7. The global processing identifies growing-season imagery and only includes
cloud-free images. We specifically use bands 3, 5, and 7 in our analysis (corresponding to the
“red” band, and two bands of “near infrared”). We chose these bands based on preliminary
sensitivity tests, which revealed that these bands provided sufficient spectral specificity for
identifying terrestrial changes. The latest version of the Global Forest Change product includes
cloud-free processed Landsat imagery for the years 2000 and 2019 (GFCv1.7). However, when
cloud-free imagery was not available for the specific year, imagery was taken from the closest
year with cloud-free data (for year 2000, data could come from 1999-2012; for 2019 could come
from 2010-2015).
Sustainable Development Goal trends
Data indicating the country-level progress toward Sustainable Development Goal (SDG) 15,
‘Life on Land’ is taken from the latest SDG report (Sachs et al 2020a). Broadly, the SDG report
serves as the most up-to-data compilation of progress toward all SDGs, drawing from both
official sources (e.g.,the United Nations, the World Bank) and from unofficial sources (e.g.
research institutions and non-governmental organizations). The three types of data that were used
to create the SDG15 trend for each country includes published research on the fraction of
protected forest-habitat (Curtis et al 2018), fraction of protected freshwater-habitat (Lenzen et al
2012, 2013), and threatened “red list” species (Butchart et al 2007). The report includes the most
current assessment of the trend in SDG15 progress, in terms of ‘on track to achieve the SDG’,
‘moderately improving’, ‘stagnating’, or ‘decreasing’. In this work, we denote ‘on track to
achieve the SDG’ and ‘moderately improving’, as both indicative of progress toward SDG15.
5
The convolutional neural network method
We train a neural network to take a single Landsat GFCv1.7 image as input and output a single
value which represents the network’s prediction of the human pressure on the land surface. For
this task, we employ convolutional neural networks (CNNs), a class of neural networks often
used in image applications due to their ability to identify specific shapes (or features) in the data
that assist in making accurate predictions. We refer the reader to (LeCun et al 2015, Géron 2019,
Yamashita et al 2018) for detailed descriptions of CNNs for image applications of machine
learning. The specific CNN design and sampling technique utilized here is outlined in detail in
the Supplementary Material for transparency and reproducibility.
The input into the CNN is an array with shape (120,120,3) representing 3 Landsat channels of
120x120 pixels each. A Williams-HFI pixel is approximately 40x40 Landsat pixels. We found
that the model performed best when the input included surrounding Landsat pixels which gives
the CNN more context to predict the ml-HFI in the center of the image. Thus, we input 3
120x120 pixel Landsat images and task the CNN to output a single value between 0 and 1. This
value represents the ml-HFI in the center (see white boxes in Supp. Fig. 2). The CNN is trained
to minimize the mean squared error of the predicted ml-HFI compared to the Williams-HFI for
the year 2000 when the Williams-HFI and the GFCv1.7 imagery overlap. The model is then
frozen (i.e., the CNN weights are fixed) and used to predict the ml-HFI for the year 2019 based
on the 2019 GFCv1.7 imagery. In doing this, we make the assumption that the spectral
characteristics of individual land-use in the year 2019 are similar to the spectral signatures of the
same land-uses in the year 2000, a well-tested method in machine learning-based analyses of
land-cover change (Curtis et al 2018, Omeiza 2019). Moreover, we have performed manual
quality control on 2000 locations globally via Google Earth Pro’s time lapse feature (see
Results).
We train a separate CNN for each region shown in Supp. Fig. 1 between 70S and 70N. Different
biogeographic areas across the globe provide different challenges (e.g., rocky terrain, deserts) for
training the CNN. In addition, the types of human impacts vary across regions as well. We found
that training separate CNNs for each region improved the accuracy of the predictions compared
to requiring one CNN to describe all possibilities. It is very possible that increasing the amount
6
of training data (see discussion below), and utilizing a more complex CNN architecture, would
alleviate this step and allow one to train a single CNN for the entire globe. The CNN trained on
Outer Asia was used to evaluate pixels within Oceania due to the small sample size of the region.
The model used to evaluate Northern Africa was trained on all of Africa but only evaluated over
Northern Africa. A separate CNN was trained over Southern Africa. Pixels used for training are
shown in Supp. Fig. 3 (purple shading), and were chosen to optimize computation time while
maximizing the representation of unique geophysical features and human impacts.
Over all trained CNNs, we train on 3.8 million Williams-HFI pixels and predict 1.4e8 ml-HFI
values for 2000 and 2019, each. That is, we only train on approximately 3% of the non-water
data available due to computational and storage limitations, but expect many more samples could
be included to further improve the ml-HFI accuracy.
Results
Human pressure increasing globally
Global, regional, and local patterns of human footprint from the Williams-HFI (Fig 1a) are
identified in the year 2000 ml-HFI with high fidelity (Fig 1b). High accuracy is realized over all
levels of human footprint by training on only 3% of the terrestrial locations (Supp Fig 3), with a
mean absolute error on unseen testing data of 0.07 (Supp Fig 6). With that said, there are regions
of the planet that are systematically more challenging for the CNN to reproduce (Supp Fig 8),
including rocky areas of remote deserts and high latitude mountain ranges. Training the model
on a larger portion of locations is expected to overcome this limitation in future iterations. Once
trained, the CNN is then tasked with predicting the most current version of the ml-HFI from
2019 GFCv1.7 imagery (Fig 1c). This 2019 map represents the most current prediction of the
human footprint index (previously most-updated for 2013 (Williams et al 2020)).
Changes in ml-HFI between 2000 and 2019 are visualized globally in Fig 2 with positive values
indicating a large increase in ml-HFI, and negative values indicating large decreases in the ml-
HFI (where “large” is defined as changes greater than 0.25). Patterns of change are consistent
with a steady expansion of human encroachment into low ml-HFI areas, as well as increasing
density of pressure in areas of higher ml-HFI. Inset panels in Fig 2 display the GFCv1.7 imagery
7
from 2000 (left panels) and 2019 (middle panels) for different forms of increasing human
pressure.
Attribution of the CNN’s Decisions using Layer-wise Relevance Propagation
While the human eye can easily see the changes in the images from one year to the next, we
leverage a neural network visualization technique called Layerwise Relevance Propagation
(LRP) (Bach et al 2015, Montavon et al 2017) to reveal the features that the neural network
deems most relevant for its calculation of the ml-HFI (right panels). This method highlights the
regions of the input, or the pixels of the 3 Landsat images, that were most relevant to the CNNs
output (i.e., prediction of the ml-HFI). This technique acts to “open the black box” for neural
network tasks and, importantly, provides intuitive confirmation or possible refutation of the
CNN’s decision-making process. In our case, we use LRP as a quality control step to check that
the CNN is paying attention to the correct features that would lead to high or low ml-HFI
predicted values.
LRP, applied separately to each prediction of the CNN, produces a heatmap displaying the
relevance of each pixel of each of the 3 input Landsat images (i.e., channels). We have computed
the LRP relevance heatmaps for selected predictions of the 2019 ml-HFI where human pressure
has substantially increased. The average relevance across the 3 input channels are displayed in
Fig. 2 and in Supp. Fig. 2 where warmer colors denote pixels that were most relevant to the ml-
HFI prediction of the CNN. For all of these examples, the CNN places the highest relevance on
features within the image that are clearly due to human activity as seen by eye in the 2019
Landsat imagery. This provides us with confidence that the CNN predictions are correct for the
right reasons.
Evaluation of CNN results
We assess the performance of the year-2000 ml-HFI relative to the year-2000 Williams-HFI by
examining both the accuracy (Supp. Fig. 5) and absolute error (Supp. Fig. 6). The ml-HFI values
are within an error tolerance of 0.1 of the Williams-HFI values for 73% of the terrestrial globe.
When the error tolerance is increased to 0.2, the percentage jumps to 93%. Mean absolute errors
as a function of the Williams-HFI values (binned in increments of 0.1) are shown in Supp. Fig. 6.
8
The median absolute error is well below 0.1 for most bins, and this is especially true for regions
of very low or very high ml-HFI. The training samples and testing samples generate similar error
distributions, which suggests that the CNN models were properly trained. Global maps of the
errors and absolute errors are displayed in Supp. Fig. 7-8.
To demonstrate the success of the ml-HFI in predicting human pressure from GFCv1.7 Landsat
imagery, Supp. Figs. 9-12 provide example outputs of the ml-HFI for urban, agricultural, mining,
and rural landscapes in the year-2000. In all of these cases, the ml-HFI compares well with the
Williams-HFI, although it is clear that the Williams-HFI emphasizes road networks and is
smoother than the ml-HFI which is predicted on a pixel-by-pixel basis.
We assess the skill in the ml-HFI values using the Williams-HFI values as “truth”; however, the
Williams-HFI values themselves are not ground truth, but are estimations that can also be
incorrect at times (although this is likely rare due to the extensive amount of work that goes into
developing this index). With that said, we have found multiple locations where the ml-HFI
predicts high human pressure but the Williams-HFI does not. Upon inspection of these cases we
believe the ml-HFI to be correct. Four such examples, provided in Supp. Fig. 13, highlight how
the ml-HFI can be useful for identifying regions of high human pressure that may have been
overlooked in the Williams-HFI.
There are also regions where we believe the ml-HFI struggles. The largest errors appear over
dry, rocky terrain devoid of human activity, for example, the rocky parts of the Saharan Desert,
the Australian desert, and the spine of the Andes (Supp. Fig. 7-8). In these areas, the ml-HFI
predicts human activity where none exists. We provide three such examples in Supp. Fig. 14.
However, ml-HFI errors over these problematic regions with low Williams-HFI are still
relatively small (Supp. Fig. 6-8), with the 90th percentile error being approximately 0.12.
Because the mI-HFI appears to struggle over rocky deserts and mountains where the Williams-
HFI is zero, we mask out regions where the Williams-HFI was less than 0.02 in 2000 (termed
“Wilderness” by (Williams et al 2020)) and the mI-HFI error was larger than 0.001 for our
change calculations. That is, we mask out changes where the 2000 mI-HFI initially struggled
with a zero prediction (i.e. no human activity at all).
9
Biodiversity gains despite development
Previous research underscores the transformative effects that human activities have on natural
ecosystems (Biggs et al 2009, Rocha et al 2018). Many of these studies indicate that heightened
human pressure on the landscape diminishes the capacity for those lands to support biodiversity.
We interrogate those assumptions using the ml-HFI for the years 2000 and 2019 with a
contemporary assessment of country-level progress toward achievement of SDG15 (Sachs et al
2020a). Globally, there are 119 countries experiencing increases in the ml-HFI. We find that 76
of these countries are experiencing either stagnating or decreasing progress toward SDG15 (not
shown). This supports the common belief that an increasing human footprint is not compatible
with the maintenance or improvement of terrestrial biodiversity. Contrary to this common belief,
however, we also find that 43 countries (Fig 2, green outlines) are experiencing substantial
increases in human pressure over a large fraction of their land surface (increases of 0.25 or larger
over 1% or more of their area) while at the same time “moderately improving” or actively “on
track to achieve” SDG15. In other words, 36% of countries experiencing increases in human
pressure, are also making progress toward SDG15.
These substantial changes identified by the ml-HFI have been further confirmed manually for
more than 2,000 locations across the 43 countries using the time lapse feature in Google Earth
Pro. Norway, Gabon, Libya and Sudan have been removed from the list as further inspection
shows that the CNN struggles in these locations. In Norway the presence of high latitude rocky
areas are mis-classified as experiencing change. In Gabon there is an apparent GFCv1.7 imagery
issue which appears to be localized specifically to Gabon’s interior. Both Libya and Sudan
contain large areas of rocky, remote deserts that are misclassified as experiencing substantial
human impact.
Our core finding, that the ml-HFI is increasing while at the same time biodiversity gains in some
countries are also increasing, remains nascent in the conservation and development discourse
(Sarkar 1999, Newbold et al 2016, Ellis et al 2012, Watson et al 2018). To better understand this
apparent coexistence, we summarize the types of ml-HFI changes across a subset of eight
countries along with specific SDG15 indicator progress in Fig 3. The ml-HFI reveals a wide
10
spectrum of human pressure on the landscape, including: deforestation for extensive mining
(Suriname, Guyana), spreading road networks (Estonia, Slovenia), urbanization (Uganda, The
Gambia), and expanding agriculture (Benin, Morocco). Moreover, there is no uniform pathway
for making progress across the SDG15 indicators for forest- (Curtis et al 2018), freshwater-
(Lenzen et al 2012, 2013), and threatened “red list” species conservation (Butchart et al 2007).
For example, Slovenia and Estonia are on track to achieve SDG15 in all three indicators. But for
most countries, it is a very mixed pathway across the three indicators. For example, Uganda may
be making progress on forest and freshwater conservation, while decreasing toward its “red list”
species target. Evidently, the types of increasing pressures revealed by the ml-HFI do not lead to
identical patterns of SDG15 progress. This suggests additional country-level policy information
may deepen our understanding and we provide three such case studies: Guyana, The Gambia,
and Morocco.
Cases of development processes
A core function of the ml-HFI is to monitor remote regions experiencing rapid change, as clearly
identified in Guyana - including considerable expansion of road networks, deforestation and
extensive mining in the forested interior (Fig 3, Supp. Fig. 15). Notwithstanding its accelerating
economic development, Guyana has the second-highest per capita forest cover in the world,
second only to its neighbor, Suriname (Government of the Co-operative Republic of Guyana
2019). Given the enormous potential for forest preservation, Guyana was one of the first
countries to implement REDD+ (Reduced Emissions from Deforestation and Forest
Degradation), partnering with Norway which provides financial resources to incentivize the
reduction in deforestation rates, and the establishment of protected areas in collaboration with
indigenous communities (Government of the Co-operative Republic of Guyana 2019). These
activities contribute to Guyana’s current progress toward its SDG15 goals and, if continued, will
lead to consistently improving trends in forest and freshwater habitat. Yet, the growing network
of roads and mining across the Guiana Shield revealed by the ml-HFI shows how quickly
development processes can encroach into wild and remote areas, and may threaten the continued
success of biodiversity conservation efforts.
Urbanization of agricultural areas is a key feature of development (Ellis and Ramankutty 2008),
11
and as evidenced by the ml-HFI results for The Gambia, can take place explosively over the
course of less than 20 years (Fig 3, Supp. Fig. 16). Elsewhere in The Gambia, the ml-HFI
reveals changes driven by deforestation from roads and agriculture. The Gambia’s Voluntary
National Review for SDG15 highlights the role that climate change has had on exacerbating
challenges to natural resource management, as well as the tension between preserving
biodiversity with food security for its population - which remains 50% rural (The Republic of
Gambia 2020). Nonetheless, The Gambia includes multiple protected forests, including
biodiverse areas near the mouth of the The Gambia River. The ml-HFI results for 2019 show the
lowest levels of human pressure in precisely these riparian locations, emphasizing the need for
continued protection of these forests.
Inspection of the ml-HFI shows that Morocco is experiencing development pressure in the form
of expanding peri-urban farmland, as well as road and energy infrastructure (Fig 3, Supp. Fig.
17). As a middle income country, with both semi-arid forests and desert ecosystems, Morocco
has focused on boundary demarcation of forested areas, enforcement of forest protection from
illegal activities, and expanding coverage of forest management plans (Royaume du Maroc
2020). Such efforts are representative of Morocco’s capacity for management and enforcement,
and suggest that advanced monitoring tools, like the ml-HFI, could contribute important and
actionable information. As with Guyana and The Gambia, development is still taking place and,
as clarified in Morocco’s Voluntary National Review for the SDGs, there are ongoing plans for
water and land development as well as forest conservation in Morocco (Royaume du Maroc
2020). The ml-HFI results for 2019 support the assessment that considerable potential exists for
conservation, especially in the remote mountains and arid ecosystems of the east and southern
reaches of Morocco.
Prospects for development monitoring
Combining the SDG15 progress with the 2019 present-day human footprint index shows that
progress toward biodiversity goals can be made despite increases in human pressure. These
increases occur in countries ranging in wealth, economic development, or reliance on specific
natural resources. For many countries that are aiming for resource intensive economic
development, monitoring and evaluation will be critical for tracking changes in near-real time.
12
Countries that are still actively developing may find the ml-HFI useful to identify strategies to
provide effective guardrails for staying on track for protecting biodiversity (e.g., Dominican
Republic, Republic of Congo), especially when increased human activity encroaches on pockets
of biodiversity. For wealthier or more developed countries with longstanding regulatory
structures for protected areas, the ml-HFI can help determine conservation policy efficacy and
ensure that additional increases in ml-HFI concentrate in already developed areas (e.g., Sweden,
Switzerland).
Conclusions
Although we detail compelling evidence of progress toward biodiversity in the face of increasing
human pressure for some countries, the more common pattern is that of increasing human
pressure at the expense of gains in biodiversity conservation. We do not single out any specific
countries that are not on track to achieve SDG15 (Sachs et al 2020a), but the changes in their ml-
HFI results reveal similar pressures to the cases detailed above - urban growth, agricultural
expansion, and resource extraction (mining and timber harvesting). While considerable efforts
are being made to protect intact ecosystems from encroaching or intensifying human activities,
growth in human footprint is likely outpacing field-based monitoring protocols to assess
environmental impacts and highlight that many current policies are insufficient to regulate
development (Curtis et al 2018). Capable of capturing near present changes to the land surface,
the ml-HFI could be an important tool for assisting countries that are flagging behind their
SDG15 targets to close the gap by 2030. We also acknowledge that this is version 1.0 of the ml-
HFI, and that with greater computational power, for instance, more data could be used for
training, with corresponding increases in accuracy. Furthermore, the use of innovative tools such
as Layerwise Relevance Propagation indicates that the machine learning-based approach is
correctly identifying human-made features and activities (e.g., transportation networks or land-
cover conversion) providing confidence in, and interpretability of, the CNN’s predictions. This
step is especially important if the ml-HFI is to be used for policy decisions as national and
international laws have begun to require explainable AI systems for decision-making (High
Level Expert Group on AI 2019, Phillips et al 2020).
That some countries can simultaneously be progressing toward sustainable development goals
while experiencing dramatic increases in human pressure confounds expectations. The near-
13
present, global ml-HFI provides an analytical tool for interpreting this unexpected coexistence.
Moving forward, science and policy must work to understand why and how this is possible. Such
activities will permit the development of better policies that will foster a thriving society and
environment in the Anthropocene.
14
References
Anon 2020 Time to revise the Sustainable Development Goals Nature 583 331–2
Bach S, Binder A, Montavon G, Klauschen F, Müller K-R and Samek W 2015 On Pixel-Wise
Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation
PLoS One 10 e0130140
Belote R T, Faurby S, Brennan A, Carter N H, Dietz M S, Hahn B, McShea W J and Gage J
2020 Mammal species composition reveals new insights into Earth’s remaining wilderness
Front. Ecol. Environ. 7 10399
Beyer H L, Venter O, Grantham H S and Watson J E M 2020 Substantial losses in ecoregion
intactness highlight urgency of globally coordinated action CONSERVATION LETTERS 13
1
Biggs R, Carpenter S R and Brock W A 2009 Turning back from the brink: detecting an
impending regime shift in time to avert it Proc. Natl. Acad. Sci. U. S. A. 106 826–31
Butchart S H M, Resit Akçakaya H, Chanson J, Baillie J E M, Collen B, Quader S, Turner W R,
Amin R, Stuart S N and Hilton-Taylor C 2007 Improvements to the Red List Index PLoS
One 2 e140
Curtis P G, Slay C M, Harris N L, Tyukavina A and Hansen M C 2018 Classifying drivers of
global forest loss Science 361 1108–11
Di Marco M, Venter O, Possingham H P and Watson J E M 2018 Changes in human footprint
drive changes in species extinction risk Nat. Commun. 9 4621
Ellis E C, Antill E C and Kreft H 2012 All is not loss: plant biodiversity in the anthropocene
PLoS One 7 e30535
Ellis E C and Ramankutty N 2008 Putting people in the map: anthropogenic biomes of the world
Front. Ecol. Environ. 6 439–47
Géron A 2019 Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow:
Concepts, Tools, and Techniques to Build Intelligent Systems (O’Reilly Media)
Government of the Co-operative Republic of Guyana 2019 Guyana First Voluntary National
Review of the Sustainable Development Goals (Ministry of Finance) Online:
https://sustainabledevelopment.un.org/content/documents/24297Guyana_VNR2019_FINAL
_REPORT_070819.pdf
Hansen M C, Potapov P V, Moore R, Hancher M, Turubanova S A, Tyukavina A, Thau D,
Stehman S V, Goetz S J, Loveland T R, Kommareddy A, Egorov A, Chini L, Justice C O
and Townshend J R G 2013 High-resolution global maps of 21st-century forest cover
change Science 342 850–3
15
High Level Expert Group on AI 2019 Ethics Guidelines for Trustworthy AI: High-Level Expert
Group on Artificial Intelligence (European Commission)
Hoffman F M, Larson J W, Mills R T, Brooks B-G J, Ganguly A R, Hargrove W W, Huang J,
Kumar J and Vatsavai R R 2011 Data Mining in Earth System Science (DMESS 2011)
Procedia Comput. Sci. 4 1450–5
Jean N, Burke M, Xie M, Davis W M, Lobell D B and Ermon S 2016 Combining satellite
imagery and machine learning to predict poverty Science 353 790–4
Krizhevsky A, Sutskever I and Hinton G E 2012 ImageNet Classification with Deep
Convolutional Neural Networks Advances in Neural Information Processing Systems 25 ed
F Pereira, C J C Burges, L Bottou and K Q Weinberger (Curran Associates, Inc.) pp 1097–
105
Kumar V, Jia X, Khandelwal A and Karpatne A 2019 Predicting land covers from satellite
images using temporal and spatial contexts US Patent Online:
https://patentimages.storage.googleapis.com/9d/8d/89/7c7e4fce8eae7e/US20190303703A1.
pdf
LeCun Y, Bengio Y and Hinton G 2015 Deep learning Nature 521 436–44
Lenzen M, Moran D, Bhaduri A, Kanemoto K, Bekchanov M, Geschke A and Foran B 2013
International trade of scarce water Ecol. Econ. 94 78–85
Lenzen M, Moran D, Kanemoto K, Foran B, Lobefaro L and Geschke A 2012 International trade
drives biodiversity threats in developing nations Nature 486 109–12
Montavon G, Lapuschkin S, Binder A, Samek W and Müller K-R 2017 Explaining nonlinear
classification decisions with deep Taylor decomposition Pattern Recognit. 65 211–22
Naidoo R and Fisher B 2020 Reset Sustainable Development Goals for a pandemic world Nature
583 198–201
Newbold T, Hudson L N, Arnell A P, Contu S, De Palma A, Ferrier S, Hill S L L, Hoskins A J,
Lysenko I, Phillips H R P, Burton V J, Chng C W T, Emerson S, Gao D, Pask-Hale G,
Hutton J, Jung M, Sanchez-Ortiz K, Simmons B I, Whitmee S, Zhang H, Scharlemann J P
W and Purvis A 2016 Has land use pushed terrestrial biodiversity beyond the planetary
boundary? A global assessment Science 353 288–91
Omeiza D 2019 Efficient Machine Learning for Large-Scale Urban Land-Use Forecasting in
Sub-Saharan Africa arXiv [cs.CV] Online: http://arxiv.org/abs/1908.00340
Pascual U, Balvanera P, Díaz S, Pataki G, Roth E, Stenseke M, Watson R T, Başak Dessane E,
Islar M, Kelemen E, Maris V, Quaas M, Subramanian S M, Wittmer H, Adlan A, Ahn S,
Al-Hafedh Y S, Amankwah E, Asah S T, Berry P, Bilgin A, Breslow S J, Bullock C,
Cáceres D, Daly-Hassen H, Figueroa E, Golden C D, Gómez-Baggethun E, González-
Jiménez D, Houdet J, Keune H, Kumar R, Ma K, May P H, Mead A, O’Farrell P, Pandit R,
16
Pengue W, Pichis-Madruga R, Popa F, Preston S, Pacheco-Balanza D, Saarikoski H,
Strassburg B B, van den Belt M, Verma M, Wickson F and Yagi N 2017 Valuing nature’s
contributions to people: the IPBES approach Current Opinion in Environmental
Sustainability 26-27 7–16
Phillips P J, Hahn C A, Fontana P C, Broniatowski D A and Przybocki M A 2020 Four
Principles of Explainable Artificial Intelligence Online:
http://dx.doi.org/10.6028/NIST.IR.8312-draft
Qiu C, Schmitt M, Geiß C, Chen T-H K and Zhu X X 2020 A framework for large-scale
mapping of human settlement extent from Sentinel-2 images via fully convolutional neural
networks ISPRS J. Photogramm. Remote Sens. 163 152–70
Riggio J, Baillie J E M, Brumby S, Ellis E, Kennedy C M, Oakleaf J R, Tait A, Tepe T,
Theobald D M, Venter O, Watson J E M and Jacobson A P 2020 Global human influence
maps reveal clear opportunities in conserving Earth’s remaining intact terrestrial ecosystems
Glob. Chang. Biol. Online: http://dx.doi.org/10.1111/gcb.15109
Rocha J C, Peterson G, Bodin Ö and Levin S 2018 Cascading regime shifts within and across
scales Science 362 1379–83
Royaume du Maroc 2020 Examen National Volontaire de la Mise en Oeuvre des Objectifs de
Développement Durable Online:
https://sustainabledevelopment.un.org/content/documents/26406VNR_2020_Morocco_Rep
ort_French.pdf
Ruckelshaus M H, Jackson S T, Mooney H A, Jacobs K L, Kassam K-A S, Arroyo M T K, Báldi
A, Bartuska A M, Boyd J, Joppa L N, Kovács-Hostyánszki A, Parsons J P, Scholes R J,
Shogren J F and Ouyang Z 2020 The IPBES Global Assessment: Pathways to Action
Trends Ecol. Evol. 35 407–14
Sachs J, Schmidt-Traub G, Kroll C, Lafortune G, Fuller G and Woelm F 2020a The Sustainable
Development Report 2020 - The Sustainable Development Goals and COVID-19
(Cambridge University Press) Online:
https://s3.amazonaws.com/sustainabledevelopment.report/2020/2020_sustainable_develop
ment_report.pdf
Sachs J, Schmidt-Traub G and Lafortune G 2020b Speaking truth to power about the SDGs
Nature 584 344
Sanderson E W, Jaiteh M, Levy M A, Redford K H, Wannebo A V and Woolmer G 2002 The
Human Footprint and the Last of the Wild: The human footprint is a global map of human
influence on the land surface, which suggests that human beings are stewards of nature,
whether we like it or not Bioscience 52 891–904
Sarkar S 1999 Wilderness preservation and biodiversity conservation—keeping divergent goals
distinct Bioscience 49 405–12
17
The Republic of Gambia 2020 The Gambia. Voluntary National Review. A report on the
progress of implementation of SDGs Online:
https://sustainabledevelopment.un.org/content/documents/263072020GambiaVNR.pdf
Venter O, Sanderson E W, Magrach A, Allan J R, Beher J, Jones K R, Possingham H P,
Laurance W F, Wood P, Fekete B M, Levy M A and Watson J E M 2016a Global terrestrial
Human Footprint maps for 1993 and 2009 Sci Data 3 160067
Venter O, Sanderson E W, Magrach A, Allan J R, Beher J, Jones K R, Possingham H P,
Laurance W F, Wood P, Fekete B M, Levy M A and Watson J E M 2016b Sixteen years of
change in the global terrestrial human footprint and implications for biodiversity
conservation Nat. Commun. 7 12558
Watson J E M, Evans T, Venter O, Williams B, Tulloch A, Stewart C, Thompson I, Ray J C,
Murray K, Salazar A, McAlpine C, Potapov P, Walston J, Robinson J G, Painter M, Wilkie
D, Filardi C, Laurance W F, Houghton R A, Maxwell S, Grantham H, Samper C, Wang S,
Laestadius L, Runting R K, Silva-Chávez G A, Ervin J and Lindenmayer D 2018 The
exceptional value of intact forest ecosystems Nat Ecol Evol 2 599–610
Williams B A, Venter O, Allan J R, Atkinson S C, Rehbein J A, Ward M S, Di Marco M,
Grantham H S, Ervin J, Goetz S, Hansen A J, Jantz P, Pillay R, Rodríguez-Buriticá S,
Supples C, Virnig A L S and Watson J E M 2020 Change in terrestrial human footprint
drives continued loss of intact ecosystems bioRxiv 2020.05.04.077818 Online:
https://www.biorxiv.org/content/10.1101/2020.05.04.077818v1
Xie M, Jean N, Burke M, Lobell D and Ermon S 2015 Transfer Learning from Deep Features for
Remote Sensing and Poverty Mapping arXiv [cs.CV] Online:
http://arxiv.org/abs/1510.00098
Yamashita R, Nishio M, Do R K G and Togashi K 2018 Convolutional neural networks: an
overview and application in radiology Insights Imaging 9 611–29
18
Acknowledgments
This research was funded, in part, by the United States National Aeronautics and Space
Administration (NASA) under grant #18-SLSCVC18-0006.
Author contributions
All authors devised the research plan. E.A.B. undertook the development of the convolutional
neural network, including design, training and validation. P.W.K. processed sustainable
development goal data and performed quality control on neural network model output. All
authors contributed to the writing of the manuscript.
Data Availability and Code Availability
The ml-HFI for 2000 and 2019 as well as the weights for the trained convolutional neural
networks will be made available to the community via the Mountain Scholar permanent data
repository with a permanent DOI and via github. All data used in this study is publically
available. Human Footprint Index data were downloaded from the github repository
https://github.com/scabecks/humanfootprint_2000-2013. Sustainable Development Goal data
were downloaded from the United Nations website https://dashboards.sdgindex.org/downloads.
Landsat imagery was downloaded from the Hansen Global Forest Change v1.7 repository
https://earthenginepartners.appspot.com/science-2013-global-forest/download_v1.7.html .
19
Figure 1: The human footprint index (HFI), a measurement of human pressure on the landscape
scaled between 0 and 1. (a) Year-2000 HFI as computed by Williams et al. 2020, (b) year-2000
HFI as computed by a machine learning method, (c) as in (b) but for year-2019.
20
Figure 2: Large changes in human pressure defined as the ml-HFI in 2019 minus that in 2000,
where red shading denotes changes larger than 0.25 and blue shading denotes changes smaller
than -0.25. Gray shading denotes the ml-HFI from 2000 for reference. Green outlined countries
are experiencing substantial increases in HFI and making progress toward SDG15. Inset panels
provide examples of increasing human pressure and the relevant features used by the CNN to
identify human activity. From left to right, each inset shows (left) year-2000 GFCv1.7 imagery,
(middle) year-2019 GFCv1.7 imagery and (right) features most relevant to the CNN for its year-
2019 prediction of the ml-HFI. The GFCv1.7 imagery is plotted in false color as its spectral
bands are outside of the visible spectrum.
21
Figure 3: Countries experiencing substantial changes in human pressure that are also making
progress toward SDG15. Percent indicates the fraction of each countries’ land surface area that is
experiencing substantial increases in human pressure, defined as a change of 0.25 on the 0-1
scale of the ml-HFI, over land surface area greater than or equal to 1% of the total country area.
Countries shown are a subset of the full list of 43 countries meeting these criteria detailed in
Supp Table 1. ml-HFI changes displayed here are degraded in resolution for easier visualization.
| 2021 | A machine-learning approach to human footprint index estimation with applications to sustainable development | 10.1101/2020.09.06.284414 | [
"Keys Patrick W.",
"Barnes Elizabeth A.",
"Carter Neil H."
] | creative-commons |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.